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                 fs_info = BTRFS_I(inode)->root->fs_info;
 683                 ret = repair_io_failure(fs_info, offset, PAGE_SIZE,
 684                                         fixup->logical, 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 void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1363                                          struct scrub_block *sblock,
1364                                          int is_metadata, int have_csum,
1365                                          const u8 *csum, u64 generation,
1366                                          u16 csum_size)
1367 {
1368         int page_num;
1369         u8 calculated_csum[BTRFS_CSUM_SIZE];
1370         u32 crc = ~(u32)0;
1371         void *mapped_buffer;
1372
1373         WARN_ON(!sblock->pagev[0]->page);
1374         if (is_metadata) {
1375                 struct btrfs_header *h;
1376
1377                 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1378                 h = (struct btrfs_header *)mapped_buffer;
1379
1380                 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h) ||
1381                     memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
1382                     memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1383                            BTRFS_UUID_SIZE)) {
1384                         sblock->header_error = 1;
1385                 } else if (generation != btrfs_stack_header_generation(h)) {
1386                         sblock->header_error = 1;
1387                         sblock->generation_error = 1;
1388                 }
1389                 csum = h->csum;
1390         } else {
1391                 if (!have_csum)
1392                         return;
1393
1394                 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1395         }
1396
1397         for (page_num = 0;;) {
1398                 if (page_num == 0 && is_metadata)
1399                         crc = btrfs_csum_data(
1400                                 ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1401                                 crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1402                 else
1403                         crc = btrfs_csum_data(mapped_buffer, crc, PAGE_SIZE);
1404
1405                 kunmap_atomic(mapped_buffer);
1406                 page_num++;
1407                 if (page_num >= sblock->page_count)
1408                         break;
1409                 WARN_ON(!sblock->pagev[page_num]->page);
1410
1411                 mapped_buffer = kmap_atomic(sblock->pagev[page_num]->page);
1412         }
1413
1414         btrfs_csum_final(crc, calculated_csum);
1415         if (memcmp(calculated_csum, csum, csum_size))
1416                 sblock->checksum_error = 1;
1417 }
1418
1419 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1420                                              struct scrub_block *sblock_good,
1421                                              int force_write)
1422 {
1423         int page_num;
1424         int ret = 0;
1425
1426         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1427                 int ret_sub;
1428
1429                 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1430                                                            sblock_good,
1431                                                            page_num,
1432                                                            force_write);
1433                 if (ret_sub)
1434                         ret = ret_sub;
1435         }
1436
1437         return ret;
1438 }
1439
1440 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1441                                             struct scrub_block *sblock_good,
1442                                             int page_num, int force_write)
1443 {
1444         struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1445         struct scrub_page *page_good = sblock_good->pagev[page_num];
1446
1447         BUG_ON(page_bad->page == NULL);
1448         BUG_ON(page_good->page == NULL);
1449         if (force_write || sblock_bad->header_error ||
1450             sblock_bad->checksum_error || page_bad->io_error) {
1451                 struct bio *bio;
1452                 int ret;
1453
1454                 if (!page_bad->dev->bdev) {
1455                         printk_ratelimited(KERN_WARNING "BTRFS: "
1456                                 "scrub_repair_page_from_good_copy(bdev == NULL) "
1457                                 "is unexpected!\n");
1458                         return -EIO;
1459                 }
1460
1461                 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1462                 if (!bio)
1463                         return -EIO;
1464                 bio->bi_bdev = page_bad->dev->bdev;
1465                 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1466
1467                 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1468                 if (PAGE_SIZE != ret) {
1469                         bio_put(bio);
1470                         return -EIO;
1471                 }
1472
1473                 if (btrfsic_submit_bio_wait(WRITE, bio)) {
1474                         btrfs_dev_stat_inc_and_print(page_bad->dev,
1475                                 BTRFS_DEV_STAT_WRITE_ERRS);
1476                         btrfs_dev_replace_stats_inc(
1477                                 &sblock_bad->sctx->dev_root->fs_info->
1478                                 dev_replace.num_write_errors);
1479                         bio_put(bio);
1480                         return -EIO;
1481                 }
1482                 bio_put(bio);
1483         }
1484
1485         return 0;
1486 }
1487
1488 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1489 {
1490         int page_num;
1491
1492         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1493                 int ret;
1494
1495                 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1496                 if (ret)
1497                         btrfs_dev_replace_stats_inc(
1498                                 &sblock->sctx->dev_root->fs_info->dev_replace.
1499                                 num_write_errors);
1500         }
1501 }
1502
1503 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1504                                            int page_num)
1505 {
1506         struct scrub_page *spage = sblock->pagev[page_num];
1507
1508         BUG_ON(spage->page == NULL);
1509         if (spage->io_error) {
1510                 void *mapped_buffer = kmap_atomic(spage->page);
1511
1512                 memset(mapped_buffer, 0, PAGE_CACHE_SIZE);
1513                 flush_dcache_page(spage->page);
1514                 kunmap_atomic(mapped_buffer);
1515         }
1516         return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1517 }
1518
1519 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1520                                     struct scrub_page *spage)
1521 {
1522         struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1523         struct scrub_bio *sbio;
1524         int ret;
1525
1526         mutex_lock(&wr_ctx->wr_lock);
1527 again:
1528         if (!wr_ctx->wr_curr_bio) {
1529                 wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1530                                               GFP_NOFS);
1531                 if (!wr_ctx->wr_curr_bio) {
1532                         mutex_unlock(&wr_ctx->wr_lock);
1533                         return -ENOMEM;
1534                 }
1535                 wr_ctx->wr_curr_bio->sctx = sctx;
1536                 wr_ctx->wr_curr_bio->page_count = 0;
1537         }
1538         sbio = wr_ctx->wr_curr_bio;
1539         if (sbio->page_count == 0) {
1540                 struct bio *bio;
1541
1542                 sbio->physical = spage->physical_for_dev_replace;
1543                 sbio->logical = spage->logical;
1544                 sbio->dev = wr_ctx->tgtdev;
1545                 bio = sbio->bio;
1546                 if (!bio) {
1547                         bio = btrfs_io_bio_alloc(GFP_NOFS, wr_ctx->pages_per_wr_bio);
1548                         if (!bio) {
1549                                 mutex_unlock(&wr_ctx->wr_lock);
1550                                 return -ENOMEM;
1551                         }
1552                         sbio->bio = bio;
1553                 }
1554
1555                 bio->bi_private = sbio;
1556                 bio->bi_end_io = scrub_wr_bio_end_io;
1557                 bio->bi_bdev = sbio->dev->bdev;
1558                 bio->bi_iter.bi_sector = sbio->physical >> 9;
1559                 sbio->err = 0;
1560         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1561                    spage->physical_for_dev_replace ||
1562                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1563                    spage->logical) {
1564                 scrub_wr_submit(sctx);
1565                 goto again;
1566         }
1567
1568         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1569         if (ret != PAGE_SIZE) {
1570                 if (sbio->page_count < 1) {
1571                         bio_put(sbio->bio);
1572                         sbio->bio = NULL;
1573                         mutex_unlock(&wr_ctx->wr_lock);
1574                         return -EIO;
1575                 }
1576                 scrub_wr_submit(sctx);
1577                 goto again;
1578         }
1579
1580         sbio->pagev[sbio->page_count] = spage;
1581         scrub_page_get(spage);
1582         sbio->page_count++;
1583         if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1584                 scrub_wr_submit(sctx);
1585         mutex_unlock(&wr_ctx->wr_lock);
1586
1587         return 0;
1588 }
1589
1590 static void scrub_wr_submit(struct scrub_ctx *sctx)
1591 {
1592         struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1593         struct scrub_bio *sbio;
1594
1595         if (!wr_ctx->wr_curr_bio)
1596                 return;
1597
1598         sbio = wr_ctx->wr_curr_bio;
1599         wr_ctx->wr_curr_bio = NULL;
1600         WARN_ON(!sbio->bio->bi_bdev);
1601         scrub_pending_bio_inc(sctx);
1602         /* process all writes in a single worker thread. Then the block layer
1603          * orders the requests before sending them to the driver which
1604          * doubled the write performance on spinning disks when measured
1605          * with Linux 3.5 */
1606         btrfsic_submit_bio(WRITE, sbio->bio);
1607 }
1608
1609 static void scrub_wr_bio_end_io(struct bio *bio, int err)
1610 {
1611         struct scrub_bio *sbio = bio->bi_private;
1612         struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
1613
1614         sbio->err = err;
1615         sbio->bio = bio;
1616
1617         btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
1618                          scrub_wr_bio_end_io_worker, NULL, NULL);
1619         btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1620 }
1621
1622 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1623 {
1624         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1625         struct scrub_ctx *sctx = sbio->sctx;
1626         int i;
1627
1628         WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1629         if (sbio->err) {
1630                 struct btrfs_dev_replace *dev_replace =
1631                         &sbio->sctx->dev_root->fs_info->dev_replace;
1632
1633                 for (i = 0; i < sbio->page_count; i++) {
1634                         struct scrub_page *spage = sbio->pagev[i];
1635
1636                         spage->io_error = 1;
1637                         btrfs_dev_replace_stats_inc(&dev_replace->
1638                                                     num_write_errors);
1639                 }
1640         }
1641
1642         for (i = 0; i < sbio->page_count; i++)
1643                 scrub_page_put(sbio->pagev[i]);
1644
1645         bio_put(sbio->bio);
1646         kfree(sbio);
1647         scrub_pending_bio_dec(sctx);
1648 }
1649
1650 static int scrub_checksum(struct scrub_block *sblock)
1651 {
1652         u64 flags;
1653         int ret;
1654
1655         WARN_ON(sblock->page_count < 1);
1656         flags = sblock->pagev[0]->flags;
1657         ret = 0;
1658         if (flags & BTRFS_EXTENT_FLAG_DATA)
1659                 ret = scrub_checksum_data(sblock);
1660         else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1661                 ret = scrub_checksum_tree_block(sblock);
1662         else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1663                 (void)scrub_checksum_super(sblock);
1664         else
1665                 WARN_ON(1);
1666         if (ret)
1667                 scrub_handle_errored_block(sblock);
1668
1669         return ret;
1670 }
1671
1672 static int scrub_checksum_data(struct scrub_block *sblock)
1673 {
1674         struct scrub_ctx *sctx = sblock->sctx;
1675         u8 csum[BTRFS_CSUM_SIZE];
1676         u8 *on_disk_csum;
1677         struct page *page;
1678         void *buffer;
1679         u32 crc = ~(u32)0;
1680         int fail = 0;
1681         u64 len;
1682         int index;
1683
1684         BUG_ON(sblock->page_count < 1);
1685         if (!sblock->pagev[0]->have_csum)
1686                 return 0;
1687
1688         on_disk_csum = sblock->pagev[0]->csum;
1689         page = sblock->pagev[0]->page;
1690         buffer = kmap_atomic(page);
1691
1692         len = sctx->sectorsize;
1693         index = 0;
1694         for (;;) {
1695                 u64 l = min_t(u64, len, PAGE_SIZE);
1696
1697                 crc = btrfs_csum_data(buffer, crc, l);
1698                 kunmap_atomic(buffer);
1699                 len -= l;
1700                 if (len == 0)
1701                         break;
1702                 index++;
1703                 BUG_ON(index >= sblock->page_count);
1704                 BUG_ON(!sblock->pagev[index]->page);
1705                 page = sblock->pagev[index]->page;
1706                 buffer = kmap_atomic(page);
1707         }
1708
1709         btrfs_csum_final(crc, csum);
1710         if (memcmp(csum, on_disk_csum, sctx->csum_size))
1711                 fail = 1;
1712
1713         return fail;
1714 }
1715
1716 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1717 {
1718         struct scrub_ctx *sctx = sblock->sctx;
1719         struct btrfs_header *h;
1720         struct btrfs_root *root = sctx->dev_root;
1721         struct btrfs_fs_info *fs_info = root->fs_info;
1722         u8 calculated_csum[BTRFS_CSUM_SIZE];
1723         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1724         struct page *page;
1725         void *mapped_buffer;
1726         u64 mapped_size;
1727         void *p;
1728         u32 crc = ~(u32)0;
1729         int fail = 0;
1730         int crc_fail = 0;
1731         u64 len;
1732         int index;
1733
1734         BUG_ON(sblock->page_count < 1);
1735         page = sblock->pagev[0]->page;
1736         mapped_buffer = kmap_atomic(page);
1737         h = (struct btrfs_header *)mapped_buffer;
1738         memcpy(on_disk_csum, h->csum, sctx->csum_size);
1739
1740         /*
1741          * we don't use the getter functions here, as we
1742          * a) don't have an extent buffer and
1743          * b) the page is already kmapped
1744          */
1745
1746         if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1747                 ++fail;
1748
1749         if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h))
1750                 ++fail;
1751
1752         if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1753                 ++fail;
1754
1755         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1756                    BTRFS_UUID_SIZE))
1757                 ++fail;
1758
1759         len = sctx->nodesize - BTRFS_CSUM_SIZE;
1760         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1761         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1762         index = 0;
1763         for (;;) {
1764                 u64 l = min_t(u64, len, mapped_size);
1765
1766                 crc = btrfs_csum_data(p, crc, l);
1767                 kunmap_atomic(mapped_buffer);
1768                 len -= l;
1769                 if (len == 0)
1770                         break;
1771                 index++;
1772                 BUG_ON(index >= sblock->page_count);
1773                 BUG_ON(!sblock->pagev[index]->page);
1774                 page = sblock->pagev[index]->page;
1775                 mapped_buffer = kmap_atomic(page);
1776                 mapped_size = PAGE_SIZE;
1777                 p = mapped_buffer;
1778         }
1779
1780         btrfs_csum_final(crc, calculated_csum);
1781         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1782                 ++crc_fail;
1783
1784         return fail || crc_fail;
1785 }
1786
1787 static int scrub_checksum_super(struct scrub_block *sblock)
1788 {
1789         struct btrfs_super_block *s;
1790         struct scrub_ctx *sctx = sblock->sctx;
1791         struct btrfs_root *root = sctx->dev_root;
1792         struct btrfs_fs_info *fs_info = root->fs_info;
1793         u8 calculated_csum[BTRFS_CSUM_SIZE];
1794         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1795         struct page *page;
1796         void *mapped_buffer;
1797         u64 mapped_size;
1798         void *p;
1799         u32 crc = ~(u32)0;
1800         int fail_gen = 0;
1801         int fail_cor = 0;
1802         u64 len;
1803         int index;
1804
1805         BUG_ON(sblock->page_count < 1);
1806         page = sblock->pagev[0]->page;
1807         mapped_buffer = kmap_atomic(page);
1808         s = (struct btrfs_super_block *)mapped_buffer;
1809         memcpy(on_disk_csum, s->csum, sctx->csum_size);
1810
1811         if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1812                 ++fail_cor;
1813
1814         if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1815                 ++fail_gen;
1816
1817         if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1818                 ++fail_cor;
1819
1820         len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1821         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1822         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1823         index = 0;
1824         for (;;) {
1825                 u64 l = min_t(u64, len, mapped_size);
1826
1827                 crc = btrfs_csum_data(p, crc, l);
1828                 kunmap_atomic(mapped_buffer);
1829                 len -= l;
1830                 if (len == 0)
1831                         break;
1832                 index++;
1833                 BUG_ON(index >= sblock->page_count);
1834                 BUG_ON(!sblock->pagev[index]->page);
1835                 page = sblock->pagev[index]->page;
1836                 mapped_buffer = kmap_atomic(page);
1837                 mapped_size = PAGE_SIZE;
1838                 p = mapped_buffer;
1839         }
1840
1841         btrfs_csum_final(crc, calculated_csum);
1842         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1843                 ++fail_cor;
1844
1845         if (fail_cor + fail_gen) {
1846                 /*
1847                  * if we find an error in a super block, we just report it.
1848                  * They will get written with the next transaction commit
1849                  * anyway
1850                  */
1851                 spin_lock(&sctx->stat_lock);
1852                 ++sctx->stat.super_errors;
1853                 spin_unlock(&sctx->stat_lock);
1854                 if (fail_cor)
1855                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1856                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1857                 else
1858                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1859                                 BTRFS_DEV_STAT_GENERATION_ERRS);
1860         }
1861
1862         return fail_cor + fail_gen;
1863 }
1864
1865 static void scrub_block_get(struct scrub_block *sblock)
1866 {
1867         atomic_inc(&sblock->ref_count);
1868 }
1869
1870 static void scrub_block_put(struct scrub_block *sblock)
1871 {
1872         if (atomic_dec_and_test(&sblock->ref_count)) {
1873                 int i;
1874
1875                 for (i = 0; i < sblock->page_count; i++)
1876                         scrub_page_put(sblock->pagev[i]);
1877                 kfree(sblock);
1878         }
1879 }
1880
1881 static void scrub_page_get(struct scrub_page *spage)
1882 {
1883         atomic_inc(&spage->ref_count);
1884 }
1885
1886 static void scrub_page_put(struct scrub_page *spage)
1887 {
1888         if (atomic_dec_and_test(&spage->ref_count)) {
1889                 if (spage->page)
1890                         __free_page(spage->page);
1891                 kfree(spage);
1892         }
1893 }
1894
1895 static void scrub_submit(struct scrub_ctx *sctx)
1896 {
1897         struct scrub_bio *sbio;
1898
1899         if (sctx->curr == -1)
1900                 return;
1901
1902         sbio = sctx->bios[sctx->curr];
1903         sctx->curr = -1;
1904         scrub_pending_bio_inc(sctx);
1905
1906         if (!sbio->bio->bi_bdev) {
1907                 /*
1908                  * this case should not happen. If btrfs_map_block() is
1909                  * wrong, it could happen for dev-replace operations on
1910                  * missing devices when no mirrors are available, but in
1911                  * this case it should already fail the mount.
1912                  * This case is handled correctly (but _very_ slowly).
1913                  */
1914                 printk_ratelimited(KERN_WARNING
1915                         "BTRFS: scrub_submit(bio bdev == NULL) is unexpected!\n");
1916                 bio_endio(sbio->bio, -EIO);
1917         } else {
1918                 btrfsic_submit_bio(READ, sbio->bio);
1919         }
1920 }
1921
1922 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
1923                                     struct scrub_page *spage)
1924 {
1925         struct scrub_block *sblock = spage->sblock;
1926         struct scrub_bio *sbio;
1927         int ret;
1928
1929 again:
1930         /*
1931          * grab a fresh bio or wait for one to become available
1932          */
1933         while (sctx->curr == -1) {
1934                 spin_lock(&sctx->list_lock);
1935                 sctx->curr = sctx->first_free;
1936                 if (sctx->curr != -1) {
1937                         sctx->first_free = sctx->bios[sctx->curr]->next_free;
1938                         sctx->bios[sctx->curr]->next_free = -1;
1939                         sctx->bios[sctx->curr]->page_count = 0;
1940                         spin_unlock(&sctx->list_lock);
1941                 } else {
1942                         spin_unlock(&sctx->list_lock);
1943                         wait_event(sctx->list_wait, sctx->first_free != -1);
1944                 }
1945         }
1946         sbio = sctx->bios[sctx->curr];
1947         if (sbio->page_count == 0) {
1948                 struct bio *bio;
1949
1950                 sbio->physical = spage->physical;
1951                 sbio->logical = spage->logical;
1952                 sbio->dev = spage->dev;
1953                 bio = sbio->bio;
1954                 if (!bio) {
1955                         bio = btrfs_io_bio_alloc(GFP_NOFS, sctx->pages_per_rd_bio);
1956                         if (!bio)
1957                                 return -ENOMEM;
1958                         sbio->bio = bio;
1959                 }
1960
1961                 bio->bi_private = sbio;
1962                 bio->bi_end_io = scrub_bio_end_io;
1963                 bio->bi_bdev = sbio->dev->bdev;
1964                 bio->bi_iter.bi_sector = sbio->physical >> 9;
1965                 sbio->err = 0;
1966         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1967                    spage->physical ||
1968                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1969                    spage->logical ||
1970                    sbio->dev != spage->dev) {
1971                 scrub_submit(sctx);
1972                 goto again;
1973         }
1974
1975         sbio->pagev[sbio->page_count] = spage;
1976         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1977         if (ret != PAGE_SIZE) {
1978                 if (sbio->page_count < 1) {
1979                         bio_put(sbio->bio);
1980                         sbio->bio = NULL;
1981                         return -EIO;
1982                 }
1983                 scrub_submit(sctx);
1984                 goto again;
1985         }
1986
1987         scrub_block_get(sblock); /* one for the page added to the bio */
1988         atomic_inc(&sblock->outstanding_pages);
1989         sbio->page_count++;
1990         if (sbio->page_count == sctx->pages_per_rd_bio)
1991                 scrub_submit(sctx);
1992
1993         return 0;
1994 }
1995
1996 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
1997                        u64 physical, struct btrfs_device *dev, u64 flags,
1998                        u64 gen, int mirror_num, u8 *csum, int force,
1999                        u64 physical_for_dev_replace)
2000 {
2001         struct scrub_block *sblock;
2002         int index;
2003
2004         sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
2005         if (!sblock) {
2006                 spin_lock(&sctx->stat_lock);
2007                 sctx->stat.malloc_errors++;
2008                 spin_unlock(&sctx->stat_lock);
2009                 return -ENOMEM;
2010         }
2011
2012         /* one ref inside this function, plus one for each page added to
2013          * a bio later on */
2014         atomic_set(&sblock->ref_count, 1);
2015         sblock->sctx = sctx;
2016         sblock->no_io_error_seen = 1;
2017
2018         for (index = 0; len > 0; index++) {
2019                 struct scrub_page *spage;
2020                 u64 l = min_t(u64, len, PAGE_SIZE);
2021
2022                 spage = kzalloc(sizeof(*spage), GFP_NOFS);
2023                 if (!spage) {
2024 leave_nomem:
2025                         spin_lock(&sctx->stat_lock);
2026                         sctx->stat.malloc_errors++;
2027                         spin_unlock(&sctx->stat_lock);
2028                         scrub_block_put(sblock);
2029                         return -ENOMEM;
2030                 }
2031                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2032                 scrub_page_get(spage);
2033                 sblock->pagev[index] = spage;
2034                 spage->sblock = sblock;
2035                 spage->dev = dev;
2036                 spage->flags = flags;
2037                 spage->generation = gen;
2038                 spage->logical = logical;
2039                 spage->physical = physical;
2040                 spage->physical_for_dev_replace = physical_for_dev_replace;
2041                 spage->mirror_num = mirror_num;
2042                 if (csum) {
2043                         spage->have_csum = 1;
2044                         memcpy(spage->csum, csum, sctx->csum_size);
2045                 } else {
2046                         spage->have_csum = 0;
2047                 }
2048                 sblock->page_count++;
2049                 spage->page = alloc_page(GFP_NOFS);
2050                 if (!spage->page)
2051                         goto leave_nomem;
2052                 len -= l;
2053                 logical += l;
2054                 physical += l;
2055                 physical_for_dev_replace += l;
2056         }
2057
2058         WARN_ON(sblock->page_count == 0);
2059         for (index = 0; index < sblock->page_count; index++) {
2060                 struct scrub_page *spage = sblock->pagev[index];
2061                 int ret;
2062
2063                 ret = scrub_add_page_to_rd_bio(sctx, spage);
2064                 if (ret) {
2065                         scrub_block_put(sblock);
2066                         return ret;
2067                 }
2068         }
2069
2070         if (force)
2071                 scrub_submit(sctx);
2072
2073         /* last one frees, either here or in bio completion for last page */
2074         scrub_block_put(sblock);
2075         return 0;
2076 }
2077
2078 static void scrub_bio_end_io(struct bio *bio, int err)
2079 {
2080         struct scrub_bio *sbio = bio->bi_private;
2081         struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2082
2083         sbio->err = err;
2084         sbio->bio = bio;
2085
2086         btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2087 }
2088
2089 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2090 {
2091         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2092         struct scrub_ctx *sctx = sbio->sctx;
2093         int i;
2094
2095         BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2096         if (sbio->err) {
2097                 for (i = 0; i < sbio->page_count; i++) {
2098                         struct scrub_page *spage = sbio->pagev[i];
2099
2100                         spage->io_error = 1;
2101                         spage->sblock->no_io_error_seen = 0;
2102                 }
2103         }
2104
2105         /* now complete the scrub_block items that have all pages completed */
2106         for (i = 0; i < sbio->page_count; i++) {
2107                 struct scrub_page *spage = sbio->pagev[i];
2108                 struct scrub_block *sblock = spage->sblock;
2109
2110                 if (atomic_dec_and_test(&sblock->outstanding_pages))
2111                         scrub_block_complete(sblock);
2112                 scrub_block_put(sblock);
2113         }
2114
2115         bio_put(sbio->bio);
2116         sbio->bio = NULL;
2117         spin_lock(&sctx->list_lock);
2118         sbio->next_free = sctx->first_free;
2119         sctx->first_free = sbio->index;
2120         spin_unlock(&sctx->list_lock);
2121
2122         if (sctx->is_dev_replace &&
2123             atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2124                 mutex_lock(&sctx->wr_ctx.wr_lock);
2125                 scrub_wr_submit(sctx);
2126                 mutex_unlock(&sctx->wr_ctx.wr_lock);
2127         }
2128
2129         scrub_pending_bio_dec(sctx);
2130 }
2131
2132 static void scrub_block_complete(struct scrub_block *sblock)
2133 {
2134         if (!sblock->no_io_error_seen) {
2135                 scrub_handle_errored_block(sblock);
2136         } else {
2137                 /*
2138                  * if has checksum error, write via repair mechanism in
2139                  * dev replace case, otherwise write here in dev replace
2140                  * case.
2141                  */
2142                 if (!scrub_checksum(sblock) && sblock->sctx->is_dev_replace)
2143                         scrub_write_block_to_dev_replace(sblock);
2144         }
2145 }
2146
2147 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u64 len,
2148                            u8 *csum)
2149 {
2150         struct btrfs_ordered_sum *sum = NULL;
2151         unsigned long index;
2152         unsigned long num_sectors;
2153
2154         while (!list_empty(&sctx->csum_list)) {
2155                 sum = list_first_entry(&sctx->csum_list,
2156                                        struct btrfs_ordered_sum, list);
2157                 if (sum->bytenr > logical)
2158                         return 0;
2159                 if (sum->bytenr + sum->len > logical)
2160                         break;
2161
2162                 ++sctx->stat.csum_discards;
2163                 list_del(&sum->list);
2164                 kfree(sum);
2165                 sum = NULL;
2166         }
2167         if (!sum)
2168                 return 0;
2169
2170         index = ((u32)(logical - sum->bytenr)) / sctx->sectorsize;
2171         num_sectors = sum->len / sctx->sectorsize;
2172         memcpy(csum, sum->sums + index, sctx->csum_size);
2173         if (index == num_sectors - 1) {
2174                 list_del(&sum->list);
2175                 kfree(sum);
2176         }
2177         return 1;
2178 }
2179
2180 /* scrub extent tries to collect up to 64 kB for each bio */
2181 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2182                         u64 physical, struct btrfs_device *dev, u64 flags,
2183                         u64 gen, int mirror_num, u64 physical_for_dev_replace)
2184 {
2185         int ret;
2186         u8 csum[BTRFS_CSUM_SIZE];
2187         u32 blocksize;
2188
2189         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2190                 blocksize = sctx->sectorsize;
2191                 spin_lock(&sctx->stat_lock);
2192                 sctx->stat.data_extents_scrubbed++;
2193                 sctx->stat.data_bytes_scrubbed += len;
2194                 spin_unlock(&sctx->stat_lock);
2195         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2196                 blocksize = sctx->nodesize;
2197                 spin_lock(&sctx->stat_lock);
2198                 sctx->stat.tree_extents_scrubbed++;
2199                 sctx->stat.tree_bytes_scrubbed += len;
2200                 spin_unlock(&sctx->stat_lock);
2201         } else {
2202                 blocksize = sctx->sectorsize;
2203                 WARN_ON(1);
2204         }
2205
2206         while (len) {
2207                 u64 l = min_t(u64, len, blocksize);
2208                 int have_csum = 0;
2209
2210                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2211                         /* push csums to sbio */
2212                         have_csum = scrub_find_csum(sctx, logical, l, csum);
2213                         if (have_csum == 0)
2214                                 ++sctx->stat.no_csum;
2215                         if (sctx->is_dev_replace && !have_csum) {
2216                                 ret = copy_nocow_pages(sctx, logical, l,
2217                                                        mirror_num,
2218                                                       physical_for_dev_replace);
2219                                 goto behind_scrub_pages;
2220                         }
2221                 }
2222                 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2223                                   mirror_num, have_csum ? csum : NULL, 0,
2224                                   physical_for_dev_replace);
2225 behind_scrub_pages:
2226                 if (ret)
2227                         return ret;
2228                 len -= l;
2229                 logical += l;
2230                 physical += l;
2231                 physical_for_dev_replace += l;
2232         }
2233         return 0;
2234 }
2235
2236 /*
2237  * Given a physical address, this will calculate it's
2238  * logical offset. if this is a parity stripe, it will return
2239  * the most left data stripe's logical offset.
2240  *
2241  * return 0 if it is a data stripe, 1 means parity stripe.
2242  */
2243 static int get_raid56_logic_offset(u64 physical, int num,
2244                                    struct map_lookup *map, u64 *offset)
2245 {
2246         int i;
2247         int j = 0;
2248         u64 stripe_nr;
2249         u64 last_offset;
2250         int stripe_index;
2251         int rot;
2252
2253         last_offset = (physical - map->stripes[num].physical) *
2254                       nr_data_stripes(map);
2255         *offset = last_offset;
2256         for (i = 0; i < nr_data_stripes(map); i++) {
2257                 *offset = last_offset + i * map->stripe_len;
2258
2259                 stripe_nr = *offset;
2260                 do_div(stripe_nr, map->stripe_len);
2261                 do_div(stripe_nr, nr_data_stripes(map));
2262
2263                 /* Work out the disk rotation on this stripe-set */
2264                 rot = do_div(stripe_nr, map->num_stripes);
2265                 /* calculate which stripe this data locates */
2266                 rot += i;
2267                 stripe_index = rot % map->num_stripes;
2268                 if (stripe_index == num)
2269                         return 0;
2270                 if (stripe_index < num)
2271                         j++;
2272         }
2273         *offset = last_offset + j * map->stripe_len;
2274         return 1;
2275 }
2276
2277 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2278                                            struct map_lookup *map,
2279                                            struct btrfs_device *scrub_dev,
2280                                            int num, u64 base, u64 length,
2281                                            int is_dev_replace)
2282 {
2283         struct btrfs_path *path;
2284         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2285         struct btrfs_root *root = fs_info->extent_root;
2286         struct btrfs_root *csum_root = fs_info->csum_root;
2287         struct btrfs_extent_item *extent;
2288         struct blk_plug plug;
2289         u64 flags;
2290         int ret;
2291         int slot;
2292         u64 nstripes;
2293         struct extent_buffer *l;
2294         struct btrfs_key key;
2295         u64 physical;
2296         u64 logical;
2297         u64 logic_end;
2298         u64 physical_end;
2299         u64 generation;
2300         int mirror_num;
2301         struct reada_control *reada1;
2302         struct reada_control *reada2;
2303         struct btrfs_key key_start;
2304         struct btrfs_key key_end;
2305         u64 increment = map->stripe_len;
2306         u64 offset;
2307         u64 extent_logical;
2308         u64 extent_physical;
2309         u64 extent_len;
2310         struct btrfs_device *extent_dev;
2311         int extent_mirror_num;
2312         int stop_loop = 0;
2313
2314         nstripes = length;
2315         physical = map->stripes[num].physical;
2316         offset = 0;
2317         do_div(nstripes, map->stripe_len);
2318         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
2319                 offset = map->stripe_len * num;
2320                 increment = map->stripe_len * map->num_stripes;
2321                 mirror_num = 1;
2322         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
2323                 int factor = map->num_stripes / map->sub_stripes;
2324                 offset = map->stripe_len * (num / map->sub_stripes);
2325                 increment = map->stripe_len * factor;
2326                 mirror_num = num % map->sub_stripes + 1;
2327         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
2328                 increment = map->stripe_len;
2329                 mirror_num = num % map->num_stripes + 1;
2330         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
2331                 increment = map->stripe_len;
2332                 mirror_num = num % map->num_stripes + 1;
2333         } else if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2334                                 BTRFS_BLOCK_GROUP_RAID6)) {
2335                 get_raid56_logic_offset(physical, num, map, &offset);
2336                 increment = map->stripe_len * nr_data_stripes(map);
2337                 mirror_num = 1;
2338         } else {
2339                 increment = map->stripe_len;
2340                 mirror_num = 1;
2341         }
2342
2343         path = btrfs_alloc_path();
2344         if (!path)
2345                 return -ENOMEM;
2346
2347         /*
2348          * work on commit root. The related disk blocks are static as
2349          * long as COW is applied. This means, it is save to rewrite
2350          * them to repair disk errors without any race conditions
2351          */
2352         path->search_commit_root = 1;
2353         path->skip_locking = 1;
2354
2355         /*
2356          * trigger the readahead for extent tree csum tree and wait for
2357          * completion. During readahead, the scrub is officially paused
2358          * to not hold off transaction commits
2359          */
2360         logical = base + offset;
2361         physical_end = physical + nstripes * map->stripe_len;
2362         if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2363                          BTRFS_BLOCK_GROUP_RAID6)) {
2364                 get_raid56_logic_offset(physical_end, num,
2365                                         map, &logic_end);
2366                 logic_end += base;
2367         } else {
2368                 logic_end = logical + increment * nstripes;
2369         }
2370         wait_event(sctx->list_wait,
2371                    atomic_read(&sctx->bios_in_flight) == 0);
2372         scrub_blocked_if_needed(fs_info);
2373
2374         /* FIXME it might be better to start readahead at commit root */
2375         key_start.objectid = logical;
2376         key_start.type = BTRFS_EXTENT_ITEM_KEY;
2377         key_start.offset = (u64)0;
2378         key_end.objectid = logic_end;
2379         key_end.type = BTRFS_METADATA_ITEM_KEY;
2380         key_end.offset = (u64)-1;
2381         reada1 = btrfs_reada_add(root, &key_start, &key_end);
2382
2383         key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2384         key_start.type = BTRFS_EXTENT_CSUM_KEY;
2385         key_start.offset = logical;
2386         key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2387         key_end.type = BTRFS_EXTENT_CSUM_KEY;
2388         key_end.offset = logic_end;
2389         reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
2390
2391         if (!IS_ERR(reada1))
2392                 btrfs_reada_wait(reada1);
2393         if (!IS_ERR(reada2))
2394                 btrfs_reada_wait(reada2);
2395
2396
2397         /*
2398          * collect all data csums for the stripe to avoid seeking during
2399          * the scrub. This might currently (crc32) end up to be about 1MB
2400          */
2401         blk_start_plug(&plug);
2402
2403         /*
2404          * now find all extents for each stripe and scrub them
2405          */
2406         ret = 0;
2407         while (physical < physical_end) {
2408                 /* for raid56, we skip parity stripe */
2409                 if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2410                                 BTRFS_BLOCK_GROUP_RAID6)) {
2411                         ret = get_raid56_logic_offset(physical, num,
2412                                         map, &logical);
2413                         logical += base;
2414                         if (ret)
2415                                 goto skip;
2416                 }
2417                 /*
2418                  * canceled?
2419                  */
2420                 if (atomic_read(&fs_info->scrub_cancel_req) ||
2421                     atomic_read(&sctx->cancel_req)) {
2422                         ret = -ECANCELED;
2423                         goto out;
2424                 }
2425                 /*
2426                  * check to see if we have to pause
2427                  */
2428                 if (atomic_read(&fs_info->scrub_pause_req)) {
2429                         /* push queued extents */
2430                         atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2431                         scrub_submit(sctx);
2432                         mutex_lock(&sctx->wr_ctx.wr_lock);
2433                         scrub_wr_submit(sctx);
2434                         mutex_unlock(&sctx->wr_ctx.wr_lock);
2435                         wait_event(sctx->list_wait,
2436                                    atomic_read(&sctx->bios_in_flight) == 0);
2437                         atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2438                         scrub_blocked_if_needed(fs_info);
2439                 }
2440
2441                 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2442                         key.type = BTRFS_METADATA_ITEM_KEY;
2443                 else
2444                         key.type = BTRFS_EXTENT_ITEM_KEY;
2445                 key.objectid = logical;
2446                 key.offset = (u64)-1;
2447
2448                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2449                 if (ret < 0)
2450                         goto out;
2451
2452                 if (ret > 0) {
2453                         ret = btrfs_previous_extent_item(root, path, 0);
2454                         if (ret < 0)
2455                                 goto out;
2456                         if (ret > 0) {
2457                                 /* there's no smaller item, so stick with the
2458                                  * larger one */
2459                                 btrfs_release_path(path);
2460                                 ret = btrfs_search_slot(NULL, root, &key,
2461                                                         path, 0, 0);
2462                                 if (ret < 0)
2463                                         goto out;
2464                         }
2465                 }
2466
2467                 stop_loop = 0;
2468                 while (1) {
2469                         u64 bytes;
2470
2471                         l = path->nodes[0];
2472                         slot = path->slots[0];
2473                         if (slot >= btrfs_header_nritems(l)) {
2474                                 ret = btrfs_next_leaf(root, path);
2475                                 if (ret == 0)
2476                                         continue;
2477                                 if (ret < 0)
2478                                         goto out;
2479
2480                                 stop_loop = 1;
2481                                 break;
2482                         }
2483                         btrfs_item_key_to_cpu(l, &key, slot);
2484
2485                         if (key.type == BTRFS_METADATA_ITEM_KEY)
2486                                 bytes = root->nodesize;
2487                         else
2488                                 bytes = key.offset;
2489
2490                         if (key.objectid + bytes <= logical)
2491                                 goto next;
2492
2493                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2494                             key.type != BTRFS_METADATA_ITEM_KEY)
2495                                 goto next;
2496
2497                         if (key.objectid >= logical + map->stripe_len) {
2498                                 /* out of this device extent */
2499                                 if (key.objectid >= logic_end)
2500                                         stop_loop = 1;
2501                                 break;
2502                         }
2503
2504                         extent = btrfs_item_ptr(l, slot,
2505                                                 struct btrfs_extent_item);
2506                         flags = btrfs_extent_flags(l, extent);
2507                         generation = btrfs_extent_generation(l, extent);
2508
2509                         if (key.objectid < logical &&
2510                             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
2511                                 btrfs_err(fs_info,
2512                                            "scrub: tree block %llu spanning "
2513                                            "stripes, ignored. logical=%llu",
2514                                        key.objectid, logical);
2515                                 goto next;
2516                         }
2517
2518 again:
2519                         extent_logical = key.objectid;
2520                         extent_len = bytes;
2521
2522                         /*
2523                          * trim extent to this stripe
2524                          */
2525                         if (extent_logical < logical) {
2526                                 extent_len -= logical - extent_logical;
2527                                 extent_logical = logical;
2528                         }
2529                         if (extent_logical + extent_len >
2530                             logical + map->stripe_len) {
2531                                 extent_len = logical + map->stripe_len -
2532                                              extent_logical;
2533                         }
2534
2535                         extent_physical = extent_logical - logical + physical;
2536                         extent_dev = scrub_dev;
2537                         extent_mirror_num = mirror_num;
2538                         if (is_dev_replace)
2539                                 scrub_remap_extent(fs_info, extent_logical,
2540                                                    extent_len, &extent_physical,
2541                                                    &extent_dev,
2542                                                    &extent_mirror_num);
2543
2544                         ret = btrfs_lookup_csums_range(csum_root, logical,
2545                                                 logical + map->stripe_len - 1,
2546                                                 &sctx->csum_list, 1);
2547                         if (ret)
2548                                 goto out;
2549
2550                         ret = scrub_extent(sctx, extent_logical, extent_len,
2551                                            extent_physical, extent_dev, flags,
2552                                            generation, extent_mirror_num,
2553                                            extent_logical - logical + physical);
2554                         if (ret)
2555                                 goto out;
2556
2557                         scrub_free_csums(sctx);
2558                         if (extent_logical + extent_len <
2559                             key.objectid + bytes) {
2560                                 if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2561                                         BTRFS_BLOCK_GROUP_RAID6)) {
2562                                         /*
2563                                          * loop until we find next data stripe
2564                                          * or we have finished all stripes.
2565                                          */
2566                                         do {
2567                                                 physical += map->stripe_len;
2568                                                 ret = get_raid56_logic_offset(
2569                                                                 physical, num,
2570                                                                 map, &logical);
2571                                                 logical += base;
2572                                         } while (physical < physical_end && ret);
2573                                 } else {
2574                                         physical += map->stripe_len;
2575                                         logical += increment;
2576                                 }
2577                                 if (logical < key.objectid + bytes) {
2578                                         cond_resched();
2579                                         goto again;
2580                                 }
2581
2582                                 if (physical >= physical_end) {
2583                                         stop_loop = 1;
2584                                         break;
2585                                 }
2586                         }
2587 next:
2588                         path->slots[0]++;
2589                 }
2590                 btrfs_release_path(path);
2591 skip:
2592                 logical += increment;
2593                 physical += map->stripe_len;
2594                 spin_lock(&sctx->stat_lock);
2595                 if (stop_loop)
2596                         sctx->stat.last_physical = map->stripes[num].physical +
2597                                                    length;
2598                 else
2599                         sctx->stat.last_physical = physical;
2600                 spin_unlock(&sctx->stat_lock);
2601                 if (stop_loop)
2602                         break;
2603         }
2604 out:
2605         /* push queued extents */
2606         scrub_submit(sctx);
2607         mutex_lock(&sctx->wr_ctx.wr_lock);
2608         scrub_wr_submit(sctx);
2609         mutex_unlock(&sctx->wr_ctx.wr_lock);
2610
2611         blk_finish_plug(&plug);
2612         btrfs_free_path(path);
2613         return ret < 0 ? ret : 0;
2614 }
2615
2616 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2617                                           struct btrfs_device *scrub_dev,
2618                                           u64 chunk_tree, u64 chunk_objectid,
2619                                           u64 chunk_offset, u64 length,
2620                                           u64 dev_offset, int is_dev_replace)
2621 {
2622         struct btrfs_mapping_tree *map_tree =
2623                 &sctx->dev_root->fs_info->mapping_tree;
2624         struct map_lookup *map;
2625         struct extent_map *em;
2626         int i;
2627         int ret = 0;
2628
2629         read_lock(&map_tree->map_tree.lock);
2630         em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2631         read_unlock(&map_tree->map_tree.lock);
2632
2633         if (!em)
2634                 return -EINVAL;
2635
2636         map = (struct map_lookup *)em->bdev;
2637         if (em->start != chunk_offset)
2638                 goto out;
2639
2640         if (em->len < length)
2641                 goto out;
2642
2643         for (i = 0; i < map->num_stripes; ++i) {
2644                 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2645                     map->stripes[i].physical == dev_offset) {
2646                         ret = scrub_stripe(sctx, map, scrub_dev, i,
2647                                            chunk_offset, length,
2648                                            is_dev_replace);
2649                         if (ret)
2650                                 goto out;
2651                 }
2652         }
2653 out:
2654         free_extent_map(em);
2655
2656         return ret;
2657 }
2658
2659 static noinline_for_stack
2660 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2661                            struct btrfs_device *scrub_dev, u64 start, u64 end,
2662                            int is_dev_replace)
2663 {
2664         struct btrfs_dev_extent *dev_extent = NULL;
2665         struct btrfs_path *path;
2666         struct btrfs_root *root = sctx->dev_root;
2667         struct btrfs_fs_info *fs_info = root->fs_info;
2668         u64 length;
2669         u64 chunk_tree;
2670         u64 chunk_objectid;
2671         u64 chunk_offset;
2672         int ret;
2673         int slot;
2674         struct extent_buffer *l;
2675         struct btrfs_key key;
2676         struct btrfs_key found_key;
2677         struct btrfs_block_group_cache *cache;
2678         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2679
2680         path = btrfs_alloc_path();
2681         if (!path)
2682                 return -ENOMEM;
2683
2684         path->reada = 2;
2685         path->search_commit_root = 1;
2686         path->skip_locking = 1;
2687
2688         key.objectid = scrub_dev->devid;
2689         key.offset = 0ull;
2690         key.type = BTRFS_DEV_EXTENT_KEY;
2691
2692         while (1) {
2693                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2694                 if (ret < 0)
2695                         break;
2696                 if (ret > 0) {
2697                         if (path->slots[0] >=
2698                             btrfs_header_nritems(path->nodes[0])) {
2699                                 ret = btrfs_next_leaf(root, path);
2700                                 if (ret)
2701                                         break;
2702                         }
2703                 }
2704
2705                 l = path->nodes[0];
2706                 slot = path->slots[0];
2707
2708                 btrfs_item_key_to_cpu(l, &found_key, slot);
2709
2710                 if (found_key.objectid != scrub_dev->devid)
2711                         break;
2712
2713                 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
2714                         break;
2715
2716                 if (found_key.offset >= end)
2717                         break;
2718
2719                 if (found_key.offset < key.offset)
2720                         break;
2721
2722                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2723                 length = btrfs_dev_extent_length(l, dev_extent);
2724
2725                 if (found_key.offset + length <= start)
2726                         goto skip;
2727
2728                 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2729                 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2730                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2731
2732                 /*
2733                  * get a reference on the corresponding block group to prevent
2734                  * the chunk from going away while we scrub it
2735                  */
2736                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2737
2738                 /* some chunks are removed but not committed to disk yet,
2739                  * continue scrubbing */
2740                 if (!cache)
2741                         goto skip;
2742
2743                 dev_replace->cursor_right = found_key.offset + length;
2744                 dev_replace->cursor_left = found_key.offset;
2745                 dev_replace->item_needs_writeback = 1;
2746                 ret = scrub_chunk(sctx, scrub_dev, chunk_tree, chunk_objectid,
2747                                   chunk_offset, length, found_key.offset,
2748                                   is_dev_replace);
2749
2750                 /*
2751                  * flush, submit all pending read and write bios, afterwards
2752                  * wait for them.
2753                  * Note that in the dev replace case, a read request causes
2754                  * write requests that are submitted in the read completion
2755                  * worker. Therefore in the current situation, it is required
2756                  * that all write requests are flushed, so that all read and
2757                  * write requests are really completed when bios_in_flight
2758                  * changes to 0.
2759                  */
2760                 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2761                 scrub_submit(sctx);
2762                 mutex_lock(&sctx->wr_ctx.wr_lock);
2763                 scrub_wr_submit(sctx);
2764                 mutex_unlock(&sctx->wr_ctx.wr_lock);
2765
2766                 wait_event(sctx->list_wait,
2767                            atomic_read(&sctx->bios_in_flight) == 0);
2768                 atomic_inc(&fs_info->scrubs_paused);
2769                 wake_up(&fs_info->scrub_pause_wait);
2770
2771                 /*
2772                  * must be called before we decrease @scrub_paused.
2773                  * make sure we don't block transaction commit while
2774                  * we are waiting pending workers finished.
2775                  */
2776                 wait_event(sctx->list_wait,
2777                            atomic_read(&sctx->workers_pending) == 0);
2778                 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2779
2780                 mutex_lock(&fs_info->scrub_lock);
2781                 __scrub_blocked_if_needed(fs_info);
2782                 atomic_dec(&fs_info->scrubs_paused);
2783                 mutex_unlock(&fs_info->scrub_lock);
2784                 wake_up(&fs_info->scrub_pause_wait);
2785
2786                 btrfs_put_block_group(cache);
2787                 if (ret)
2788                         break;
2789                 if (is_dev_replace &&
2790                     atomic64_read(&dev_replace->num_write_errors) > 0) {
2791                         ret = -EIO;
2792                         break;
2793                 }
2794                 if (sctx->stat.malloc_errors > 0) {
2795                         ret = -ENOMEM;
2796                         break;
2797                 }
2798
2799                 dev_replace->cursor_left = dev_replace->cursor_right;
2800                 dev_replace->item_needs_writeback = 1;
2801 skip:
2802                 key.offset = found_key.offset + length;
2803                 btrfs_release_path(path);
2804         }
2805
2806         btrfs_free_path(path);
2807
2808         /*
2809          * ret can still be 1 from search_slot or next_leaf,
2810          * that's not an error
2811          */
2812         return ret < 0 ? ret : 0;
2813 }
2814
2815 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2816                                            struct btrfs_device *scrub_dev)
2817 {
2818         int     i;
2819         u64     bytenr;
2820         u64     gen;
2821         int     ret;
2822         struct btrfs_root *root = sctx->dev_root;
2823
2824         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
2825                 return -EIO;
2826
2827         gen = root->fs_info->last_trans_committed;
2828
2829         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2830                 bytenr = btrfs_sb_offset(i);
2831                 if (bytenr + BTRFS_SUPER_INFO_SIZE > scrub_dev->total_bytes)
2832                         break;
2833
2834                 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2835                                   scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
2836                                   NULL, 1, bytenr);
2837                 if (ret)
2838                         return ret;
2839         }
2840         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2841
2842         return 0;
2843 }
2844
2845 /*
2846  * get a reference count on fs_info->scrub_workers. start worker if necessary
2847  */
2848 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
2849                                                 int is_dev_replace)
2850 {
2851         int ret = 0;
2852         int flags = WQ_FREEZABLE | WQ_UNBOUND;
2853         int max_active = fs_info->thread_pool_size;
2854
2855         if (fs_info->scrub_workers_refcnt == 0) {
2856                 if (is_dev_replace)
2857                         fs_info->scrub_workers =
2858                                 btrfs_alloc_workqueue("btrfs-scrub", flags,
2859                                                       1, 4);
2860                 else
2861                         fs_info->scrub_workers =
2862                                 btrfs_alloc_workqueue("btrfs-scrub", flags,
2863                                                       max_active, 4);
2864                 if (!fs_info->scrub_workers) {
2865                         ret = -ENOMEM;
2866                         goto out;
2867                 }
2868                 fs_info->scrub_wr_completion_workers =
2869                         btrfs_alloc_workqueue("btrfs-scrubwrc", flags,
2870                                               max_active, 2);
2871                 if (!fs_info->scrub_wr_completion_workers) {
2872                         ret = -ENOMEM;
2873                         goto out;
2874                 }
2875                 fs_info->scrub_nocow_workers =
2876                         btrfs_alloc_workqueue("btrfs-scrubnc", flags, 1, 0);
2877                 if (!fs_info->scrub_nocow_workers) {
2878                         ret = -ENOMEM;
2879                         goto out;
2880                 }
2881         }
2882         ++fs_info->scrub_workers_refcnt;
2883 out:
2884         return ret;
2885 }
2886
2887 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
2888 {
2889         if (--fs_info->scrub_workers_refcnt == 0) {
2890                 btrfs_destroy_workqueue(fs_info->scrub_workers);
2891                 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
2892                 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
2893         }
2894         WARN_ON(fs_info->scrub_workers_refcnt < 0);
2895 }
2896
2897 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2898                     u64 end, struct btrfs_scrub_progress *progress,
2899                     int readonly, int is_dev_replace)
2900 {
2901         struct scrub_ctx *sctx;
2902         int ret;
2903         struct btrfs_device *dev;
2904         struct rcu_string *name;
2905
2906         if (btrfs_fs_closing(fs_info))
2907                 return -EINVAL;
2908
2909         if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) {
2910                 /*
2911                  * in this case scrub is unable to calculate the checksum
2912                  * the way scrub is implemented. Do not handle this
2913                  * situation at all because it won't ever happen.
2914                  */
2915                 btrfs_err(fs_info,
2916                            "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
2917                        fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN);
2918                 return -EINVAL;
2919         }
2920
2921         if (fs_info->chunk_root->sectorsize != PAGE_SIZE) {
2922                 /* not supported for data w/o checksums */
2923                 btrfs_err(fs_info,
2924                            "scrub: size assumption sectorsize != PAGE_SIZE "
2925                            "(%d != %lu) fails",
2926                        fs_info->chunk_root->sectorsize, PAGE_SIZE);
2927                 return -EINVAL;
2928         }
2929
2930         if (fs_info->chunk_root->nodesize >
2931             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
2932             fs_info->chunk_root->sectorsize >
2933             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
2934                 /*
2935                  * would exhaust the array bounds of pagev member in
2936                  * struct scrub_block
2937                  */
2938                 btrfs_err(fs_info, "scrub: size assumption nodesize and sectorsize "
2939                            "<= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
2940                        fs_info->chunk_root->nodesize,
2941                        SCRUB_MAX_PAGES_PER_BLOCK,
2942                        fs_info->chunk_root->sectorsize,
2943                        SCRUB_MAX_PAGES_PER_BLOCK);
2944                 return -EINVAL;
2945         }
2946
2947
2948         mutex_lock(&fs_info->fs_devices->device_list_mutex);
2949         dev = btrfs_find_device(fs_info, devid, NULL, NULL);
2950         if (!dev || (dev->missing && !is_dev_replace)) {
2951                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2952                 return -ENODEV;
2953         }
2954
2955         if (!is_dev_replace && !readonly && !dev->writeable) {
2956                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2957                 rcu_read_lock();
2958                 name = rcu_dereference(dev->name);
2959                 btrfs_err(fs_info, "scrub: device %s is not writable",
2960                           name->str);
2961                 rcu_read_unlock();
2962                 return -EROFS;
2963         }
2964
2965         mutex_lock(&fs_info->scrub_lock);
2966         if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
2967                 mutex_unlock(&fs_info->scrub_lock);
2968                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2969                 return -EIO;
2970         }
2971
2972         btrfs_dev_replace_lock(&fs_info->dev_replace);
2973         if (dev->scrub_device ||
2974             (!is_dev_replace &&
2975              btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2976                 btrfs_dev_replace_unlock(&fs_info->dev_replace);
2977                 mutex_unlock(&fs_info->scrub_lock);
2978                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2979                 return -EINPROGRESS;
2980         }
2981         btrfs_dev_replace_unlock(&fs_info->dev_replace);
2982
2983         ret = scrub_workers_get(fs_info, is_dev_replace);
2984         if (ret) {
2985                 mutex_unlock(&fs_info->scrub_lock);
2986                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2987                 return ret;
2988         }
2989
2990         sctx = scrub_setup_ctx(dev, is_dev_replace);
2991         if (IS_ERR(sctx)) {
2992                 mutex_unlock(&fs_info->scrub_lock);
2993                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2994                 scrub_workers_put(fs_info);
2995                 return PTR_ERR(sctx);
2996         }
2997         sctx->readonly = readonly;
2998         dev->scrub_device = sctx;
2999         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3000
3001         /*
3002          * checking @scrub_pause_req here, we can avoid
3003          * race between committing transaction and scrubbing.
3004          */
3005         __scrub_blocked_if_needed(fs_info);
3006         atomic_inc(&fs_info->scrubs_running);
3007         mutex_unlock(&fs_info->scrub_lock);
3008
3009         if (!is_dev_replace) {
3010                 /*
3011                  * by holding device list mutex, we can
3012                  * kick off writing super in log tree sync.
3013                  */
3014                 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3015                 ret = scrub_supers(sctx, dev);
3016                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3017         }
3018
3019         if (!ret)
3020                 ret = scrub_enumerate_chunks(sctx, dev, start, end,
3021                                              is_dev_replace);
3022
3023         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3024         atomic_dec(&fs_info->scrubs_running);
3025         wake_up(&fs_info->scrub_pause_wait);
3026
3027         wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3028
3029         if (progress)
3030                 memcpy(progress, &sctx->stat, sizeof(*progress));
3031
3032         mutex_lock(&fs_info->scrub_lock);
3033         dev->scrub_device = NULL;
3034         scrub_workers_put(fs_info);
3035         mutex_unlock(&fs_info->scrub_lock);
3036
3037         scrub_free_ctx(sctx);
3038
3039         return ret;
3040 }
3041
3042 void btrfs_scrub_pause(struct btrfs_root *root)
3043 {
3044         struct btrfs_fs_info *fs_info = root->fs_info;
3045
3046         mutex_lock(&fs_info->scrub_lock);
3047         atomic_inc(&fs_info->scrub_pause_req);
3048         while (atomic_read(&fs_info->scrubs_paused) !=
3049                atomic_read(&fs_info->scrubs_running)) {
3050                 mutex_unlock(&fs_info->scrub_lock);
3051                 wait_event(fs_info->scrub_pause_wait,
3052                            atomic_read(&fs_info->scrubs_paused) ==
3053                            atomic_read(&fs_info->scrubs_running));
3054                 mutex_lock(&fs_info->scrub_lock);
3055         }
3056         mutex_unlock(&fs_info->scrub_lock);
3057 }
3058
3059 void btrfs_scrub_continue(struct btrfs_root *root)
3060 {
3061         struct btrfs_fs_info *fs_info = root->fs_info;
3062
3063         atomic_dec(&fs_info->scrub_pause_req);
3064         wake_up(&fs_info->scrub_pause_wait);
3065 }
3066
3067 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3068 {
3069         mutex_lock(&fs_info->scrub_lock);
3070         if (!atomic_read(&fs_info->scrubs_running)) {
3071                 mutex_unlock(&fs_info->scrub_lock);
3072                 return -ENOTCONN;
3073         }
3074
3075         atomic_inc(&fs_info->scrub_cancel_req);
3076         while (atomic_read(&fs_info->scrubs_running)) {
3077                 mutex_unlock(&fs_info->scrub_lock);
3078                 wait_event(fs_info->scrub_pause_wait,
3079                            atomic_read(&fs_info->scrubs_running) == 0);
3080                 mutex_lock(&fs_info->scrub_lock);
3081         }
3082         atomic_dec(&fs_info->scrub_cancel_req);
3083         mutex_unlock(&fs_info->scrub_lock);
3084
3085         return 0;
3086 }
3087
3088 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
3089                            struct btrfs_device *dev)
3090 {
3091         struct scrub_ctx *sctx;
3092
3093         mutex_lock(&fs_info->scrub_lock);
3094         sctx = dev->scrub_device;
3095         if (!sctx) {
3096                 mutex_unlock(&fs_info->scrub_lock);
3097                 return -ENOTCONN;
3098         }
3099         atomic_inc(&sctx->cancel_req);
3100         while (dev->scrub_device) {
3101                 mutex_unlock(&fs_info->scrub_lock);
3102                 wait_event(fs_info->scrub_pause_wait,
3103                            dev->scrub_device == NULL);
3104                 mutex_lock(&fs_info->scrub_lock);
3105         }
3106         mutex_unlock(&fs_info->scrub_lock);
3107
3108         return 0;
3109 }
3110
3111 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
3112                          struct btrfs_scrub_progress *progress)
3113 {
3114         struct btrfs_device *dev;
3115         struct scrub_ctx *sctx = NULL;
3116
3117         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
3118         dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
3119         if (dev)
3120                 sctx = dev->scrub_device;
3121         if (sctx)
3122                 memcpy(progress, &sctx->stat, sizeof(*progress));
3123         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
3124
3125         return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3126 }
3127
3128 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
3129                                u64 extent_logical, u64 extent_len,
3130                                u64 *extent_physical,
3131                                struct btrfs_device **extent_dev,
3132                                int *extent_mirror_num)
3133 {
3134         u64 mapped_length;
3135         struct btrfs_bio *bbio = NULL;
3136         int ret;
3137
3138         mapped_length = extent_len;
3139         ret = btrfs_map_block(fs_info, READ, extent_logical,
3140                               &mapped_length, &bbio, 0);
3141         if (ret || !bbio || mapped_length < extent_len ||
3142             !bbio->stripes[0].dev->bdev) {
3143                 kfree(bbio);
3144                 return;
3145         }
3146
3147         *extent_physical = bbio->stripes[0].physical;
3148         *extent_mirror_num = bbio->mirror_num;
3149         *extent_dev = bbio->stripes[0].dev;
3150         kfree(bbio);
3151 }
3152
3153 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
3154                               struct scrub_wr_ctx *wr_ctx,
3155                               struct btrfs_fs_info *fs_info,
3156                               struct btrfs_device *dev,
3157                               int is_dev_replace)
3158 {
3159         WARN_ON(wr_ctx->wr_curr_bio != NULL);
3160
3161         mutex_init(&wr_ctx->wr_lock);
3162         wr_ctx->wr_curr_bio = NULL;
3163         if (!is_dev_replace)
3164                 return 0;
3165
3166         WARN_ON(!dev->bdev);
3167         wr_ctx->pages_per_wr_bio = min_t(int, SCRUB_PAGES_PER_WR_BIO,
3168                                          bio_get_nr_vecs(dev->bdev));
3169         wr_ctx->tgtdev = dev;
3170         atomic_set(&wr_ctx->flush_all_writes, 0);
3171         return 0;
3172 }
3173
3174 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
3175 {
3176         mutex_lock(&wr_ctx->wr_lock);
3177         kfree(wr_ctx->wr_curr_bio);
3178         wr_ctx->wr_curr_bio = NULL;
3179         mutex_unlock(&wr_ctx->wr_lock);
3180 }
3181
3182 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
3183                             int mirror_num, u64 physical_for_dev_replace)
3184 {
3185         struct scrub_copy_nocow_ctx *nocow_ctx;
3186         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3187
3188         nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
3189         if (!nocow_ctx) {
3190                 spin_lock(&sctx->stat_lock);
3191                 sctx->stat.malloc_errors++;
3192                 spin_unlock(&sctx->stat_lock);
3193                 return -ENOMEM;
3194         }
3195
3196         scrub_pending_trans_workers_inc(sctx);
3197
3198         nocow_ctx->sctx = sctx;
3199         nocow_ctx->logical = logical;
3200         nocow_ctx->len = len;
3201         nocow_ctx->mirror_num = mirror_num;
3202         nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
3203         btrfs_init_work(&nocow_ctx->work, btrfs_scrubnc_helper,
3204                         copy_nocow_pages_worker, NULL, NULL);
3205         INIT_LIST_HEAD(&nocow_ctx->inodes);
3206         btrfs_queue_work(fs_info->scrub_nocow_workers,
3207                          &nocow_ctx->work);
3208
3209         return 0;
3210 }
3211
3212 static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
3213 {
3214         struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
3215         struct scrub_nocow_inode *nocow_inode;
3216
3217         nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
3218         if (!nocow_inode)
3219                 return -ENOMEM;
3220         nocow_inode->inum = inum;
3221         nocow_inode->offset = offset;
3222         nocow_inode->root = root;
3223         list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
3224         return 0;
3225 }
3226
3227 #define COPY_COMPLETE 1
3228
3229 static void copy_nocow_pages_worker(struct btrfs_work *work)
3230 {
3231         struct scrub_copy_nocow_ctx *nocow_ctx =
3232                 container_of(work, struct scrub_copy_nocow_ctx, work);
3233         struct scrub_ctx *sctx = nocow_ctx->sctx;
3234         u64 logical = nocow_ctx->logical;
3235         u64 len = nocow_ctx->len;
3236         int mirror_num = nocow_ctx->mirror_num;
3237         u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3238         int ret;
3239         struct btrfs_trans_handle *trans = NULL;
3240         struct btrfs_fs_info *fs_info;
3241         struct btrfs_path *path;
3242         struct btrfs_root *root;
3243         int not_written = 0;
3244
3245         fs_info = sctx->dev_root->fs_info;
3246         root = fs_info->extent_root;
3247
3248         path = btrfs_alloc_path();
3249         if (!path) {
3250                 spin_lock(&sctx->stat_lock);
3251                 sctx->stat.malloc_errors++;
3252                 spin_unlock(&sctx->stat_lock);
3253                 not_written = 1;
3254                 goto out;
3255         }
3256
3257         trans = btrfs_join_transaction(root);
3258         if (IS_ERR(trans)) {
3259                 not_written = 1;
3260                 goto out;
3261         }
3262
3263         ret = iterate_inodes_from_logical(logical, fs_info, path,
3264                                           record_inode_for_nocow, nocow_ctx);
3265         if (ret != 0 && ret != -ENOENT) {
3266                 btrfs_warn(fs_info, "iterate_inodes_from_logical() failed: log %llu, "
3267                         "phys %llu, len %llu, mir %u, ret %d",
3268                         logical, physical_for_dev_replace, len, mirror_num,
3269                         ret);
3270                 not_written = 1;
3271                 goto out;
3272         }
3273
3274         btrfs_end_transaction(trans, root);
3275         trans = NULL;
3276         while (!list_empty(&nocow_ctx->inodes)) {
3277                 struct scrub_nocow_inode *entry;
3278                 entry = list_first_entry(&nocow_ctx->inodes,
3279                                          struct scrub_nocow_inode,
3280                                          list);
3281                 list_del_init(&entry->list);
3282                 ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
3283                                                  entry->root, nocow_ctx);
3284                 kfree(entry);
3285                 if (ret == COPY_COMPLETE) {
3286                         ret = 0;
3287                         break;
3288                 } else if (ret) {
3289                         break;
3290                 }
3291         }
3292 out:
3293         while (!list_empty(&nocow_ctx->inodes)) {
3294                 struct scrub_nocow_inode *entry;
3295                 entry = list_first_entry(&nocow_ctx->inodes,
3296                                          struct scrub_nocow_inode,
3297                                          list);
3298                 list_del_init(&entry->list);
3299                 kfree(entry);
3300         }
3301         if (trans && !IS_ERR(trans))
3302                 btrfs_end_transaction(trans, root);
3303         if (not_written)
3304                 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
3305                                             num_uncorrectable_read_errors);
3306
3307         btrfs_free_path(path);
3308         kfree(nocow_ctx);
3309
3310         scrub_pending_trans_workers_dec(sctx);
3311 }
3312
3313 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
3314                                       struct scrub_copy_nocow_ctx *nocow_ctx)
3315 {
3316         struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
3317         struct btrfs_key key;
3318         struct inode *inode;
3319         struct page *page;
3320         struct btrfs_root *local_root;
3321         struct btrfs_ordered_extent *ordered;
3322         struct extent_map *em;
3323         struct extent_state *cached_state = NULL;
3324         struct extent_io_tree *io_tree;
3325         u64 physical_for_dev_replace;
3326         u64 len = nocow_ctx->len;
3327         u64 lockstart = offset, lockend = offset + len - 1;
3328         unsigned long index;
3329         int srcu_index;
3330         int ret = 0;
3331         int err = 0;
3332
3333         key.objectid = root;
3334         key.type = BTRFS_ROOT_ITEM_KEY;
3335         key.offset = (u64)-1;
3336
3337         srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
3338
3339         local_root = btrfs_read_fs_root_no_name(fs_info, &key);
3340         if (IS_ERR(local_root)) {
3341                 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3342                 return PTR_ERR(local_root);
3343         }
3344
3345         key.type = BTRFS_INODE_ITEM_KEY;
3346         key.objectid = inum;
3347         key.offset = 0;
3348         inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
3349         srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3350         if (IS_ERR(inode))
3351                 return PTR_ERR(inode);
3352
3353         /* Avoid truncate/dio/punch hole.. */
3354         mutex_lock(&inode->i_mutex);
3355         inode_dio_wait(inode);
3356
3357         physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3358         io_tree = &BTRFS_I(inode)->io_tree;
3359
3360         lock_extent_bits(io_tree, lockstart, lockend, 0, &cached_state);
3361         ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
3362         if (ordered) {
3363                 btrfs_put_ordered_extent(ordered);
3364                 goto out_unlock;
3365         }
3366
3367         em = btrfs_get_extent(inode, NULL, 0, lockstart, len, 0);
3368         if (IS_ERR(em)) {
3369                 ret = PTR_ERR(em);
3370                 goto out_unlock;
3371         }
3372
3373         /*
3374          * This extent does not actually cover the logical extent anymore,
3375          * move on to the next inode.
3376          */
3377         if (em->block_start > nocow_ctx->logical ||
3378             em->block_start + em->block_len < nocow_ctx->logical + len) {
3379                 free_extent_map(em);
3380                 goto out_unlock;
3381         }
3382         free_extent_map(em);
3383
3384         while (len >= PAGE_CACHE_SIZE) {
3385                 index = offset >> PAGE_CACHE_SHIFT;
3386 again:
3387                 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
3388                 if (!page) {
3389                         btrfs_err(fs_info, "find_or_create_page() failed");
3390                         ret = -ENOMEM;
3391                         goto out;
3392                 }
3393
3394                 if (PageUptodate(page)) {
3395                         if (PageDirty(page))
3396                                 goto next_page;
3397                 } else {
3398                         ClearPageError(page);
3399                         err = extent_read_full_page_nolock(io_tree, page,
3400                                                            btrfs_get_extent,
3401                                                            nocow_ctx->mirror_num);
3402                         if (err) {
3403                                 ret = err;
3404                                 goto next_page;
3405                         }
3406
3407                         lock_page(page);
3408                         /*
3409                          * If the page has been remove from the page cache,
3410                          * the data on it is meaningless, because it may be
3411                          * old one, the new data may be written into the new
3412                          * page in the page cache.
3413                          */
3414                         if (page->mapping != inode->i_mapping) {
3415                                 unlock_page(page);
3416                                 page_cache_release(page);
3417                                 goto again;
3418                         }
3419                         if (!PageUptodate(page)) {
3420                                 ret = -EIO;
3421                                 goto next_page;
3422                         }
3423                 }
3424                 err = write_page_nocow(nocow_ctx->sctx,
3425                                        physical_for_dev_replace, page);
3426                 if (err)
3427                         ret = err;
3428 next_page:
3429                 unlock_page(page);
3430                 page_cache_release(page);
3431
3432                 if (ret)
3433                         break;
3434
3435                 offset += PAGE_CACHE_SIZE;
3436                 physical_for_dev_replace += PAGE_CACHE_SIZE;
3437                 len -= PAGE_CACHE_SIZE;
3438         }
3439         ret = COPY_COMPLETE;
3440 out_unlock:
3441         unlock_extent_cached(io_tree, lockstart, lockend, &cached_state,
3442                              GFP_NOFS);
3443 out:
3444         mutex_unlock(&inode->i_mutex);
3445         iput(inode);
3446         return ret;
3447 }
3448
3449 static int write_page_nocow(struct scrub_ctx *sctx,
3450                             u64 physical_for_dev_replace, struct page *page)
3451 {
3452         struct bio *bio;
3453         struct btrfs_device *dev;
3454         int ret;
3455
3456         dev = sctx->wr_ctx.tgtdev;
3457         if (!dev)
3458                 return -EIO;
3459         if (!dev->bdev) {
3460                 printk_ratelimited(KERN_WARNING
3461                         "BTRFS: scrub write_page_nocow(bdev == NULL) is unexpected!\n");
3462                 return -EIO;
3463         }
3464         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
3465         if (!bio) {
3466                 spin_lock(&sctx->stat_lock);
3467                 sctx->stat.malloc_errors++;
3468                 spin_unlock(&sctx->stat_lock);
3469                 return -ENOMEM;
3470         }
3471         bio->bi_iter.bi_size = 0;
3472         bio->bi_iter.bi_sector = physical_for_dev_replace >> 9;
3473         bio->bi_bdev = dev->bdev;
3474         ret = bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
3475         if (ret != PAGE_CACHE_SIZE) {
3476 leave_with_eio:
3477                 bio_put(bio);
3478                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
3479                 return -EIO;
3480         }
3481
3482         if (btrfsic_submit_bio_wait(WRITE_SYNC, bio))
3483                 goto leave_with_eio;
3484
3485         bio_put(bio);
3486         return 0;
3487 }