fs/btrfs/scrub.c

   1 /*
   2  * Copyright (C) 2011, 2012 STRATO.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/blkdev.h>
  20 #include <linux/ratelimit.h>
  21 #include "ctree.h"
  22 #include "volumes.h"
  23 #include "disk-io.h"
  24 #include "ordered-data.h"
  25 #include "transaction.h"
  26 #include "backref.h"
  27 #include "extent_io.h"
  28 #include "dev-replace.h"
  29 #include "check-integrity.h"
  30 #include "rcu-string.h"
  31 #include "raid56.h"
  32
  33 /*
  34  * This is only the first step towards a full-features scrub. It reads all
  35  * extent and super block and verifies the checksums. In case a bad checksum
  36  * is found or the extent cannot be read, good data will be written back if
  37  * any can be found.
  38  *
  39  * Future enhancements:
  40  *  - In case an unrepairable extent is encountered, track which files are
  41  *    affected and report them
  42  *  - track and record media errors, throw out bad devices
  43  *  - add a mode to also read unallocated space
  44  */
  45
  46 struct scrub_block;
  47 struct scrub_ctx;
  48
  49 /*
  50  * the following three values only influence the performance.
  51  * The last one configures the number of parallel and outstanding I/O
  52  * operations. The first two values configure an upper limit for the number
  53  * of (dynamically allocated) pages that are added to a bio.
  54  */
  55 #define SCRUB_PAGES_PER_RD_BIO  32      /* 128k per bio */
  56 #define SCRUB_PAGES_PER_WR_BIO  32      /* 128k per bio */
  57 #define SCRUB_BIOS_PER_SCTX     64      /* 8MB per device in flight */
  58
  59 /*
  60  * the following value times PAGE_SIZE needs to be large enough to match the
  61  * largest node/leaf/sector size that shall be supported.
  62  * Values larger than BTRFS_STRIPE_LEN are not supported.
  63  */
  64 #define SCRUB_MAX_PAGES_PER_BLOCK       16      /* 64k per node/leaf/sector */
  65
  66 struct scrub_page {
  67         struct scrub_block      *sblock;
  68         struct page             *page;
  69         struct btrfs_device     *dev;
  70         u64                     flags;  /* extent flags */
  71         u64                     generation;
  72         u64                     logical;
  73         u64                     physical;
  74         u64                     physical_for_dev_replace;
  75         atomic_t                ref_count;
  76         struct {
  77                 unsigned int    mirror_num:8;
  78                 unsigned int    have_csum:1;
  79                 unsigned int    io_error:1;
  80         };
  81         u8                      csum[BTRFS_CSUM_SIZE];
  82 };
  83
  84 struct scrub_bio {
  85         int                     index;
  86         struct scrub_ctx        *sctx;
  87         struct btrfs_device     *dev;
  88         struct bio              *bio;
  89         int                     err;
  90         u64                     logical;
  91         u64                     physical;
  92 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
  93         struct scrub_page       *pagev[SCRUB_PAGES_PER_WR_BIO];
  94 #else
  95         struct scrub_page       *pagev[SCRUB_PAGES_PER_RD_BIO];
  96 #endif
  97         int                     page_count;
  98         int                     next_free;
  99         struct btrfs_work       work;
 100 };
 101
 102 struct scrub_block {
 103         struct scrub_page       *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
 104         int                     page_count;
 105         atomic_t                outstanding_pages;
 106         atomic_t                ref_count; /* free mem on transition to zero */
 107         struct scrub_ctx        *sctx;
 108         struct {
 109                 unsigned int    header_error:1;
 110                 unsigned int    checksum_error:1;
 111                 unsigned int    no_io_error_seen:1;
 112                 unsigned int    generation_error:1; /* also sets header_error */
 113         };
 114 };
 115
 116 struct scrub_wr_ctx {
 117         struct scrub_bio *wr_curr_bio;
 118         struct btrfs_device *tgtdev;
 119         int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
 120         atomic_t flush_all_writes;
 121         struct mutex wr_lock;
 122 };
 123
 124 struct scrub_ctx {
 125         struct scrub_bio        *bios[SCRUB_BIOS_PER_SCTX];
 126         struct btrfs_root       *dev_root;
 127         int                     first_free;
 128         int                     curr;
 129         atomic_t                bios_in_flight;
 130         atomic_t                workers_pending;
 131         spinlock_t              list_lock;
 132         wait_queue_head_t       list_wait;
 133         u16                     csum_size;
 134         struct list_head        csum_list;
 135         atomic_t                cancel_req;
 136         int                     readonly;
 137         int                     pages_per_rd_bio;
 138         u32                     sectorsize;
 139         u32                     nodesize;
 140         u32                     leafsize;
 141
 142         int                     is_dev_replace;
 143         struct scrub_wr_ctx     wr_ctx;
 144
 145         /*
 146          * statistics
 147          */
 148         struct btrfs_scrub_progress stat;
 149         spinlock_t              stat_lock;
 150 };
 151
 152 struct scrub_fixup_nodatasum {
 153         struct scrub_ctx        *sctx;
 154         struct btrfs_device     *dev;
 155         u64                     logical;
 156         struct btrfs_root       *root;
 157         struct btrfs_work       work;
 158         int                     mirror_num;
 159 };
 160
 161 struct scrub_nocow_inode {
 162         u64                     inum;
 163         u64                     offset;
 164         u64                     root;
 165         struct list_head        list;
 166 };
 167
 168 struct scrub_copy_nocow_ctx {
 169         struct scrub_ctx        *sctx;
 170         u64                     logical;
 171         u64                     len;
 172         int                     mirror_num;
 173         u64                     physical_for_dev_replace;
 174         struct list_head        inodes;
 175         struct btrfs_work       work;
 176 };
 177
 178 struct scrub_warning {
 179         struct btrfs_path       *path;
 180         u64                     extent_item_size;
 181         char                    *scratch_buf;
 182         char                    *msg_buf;
 183         const char              *errstr;
 184         sector_t                sector;
 185         u64                     logical;
 186         struct btrfs_device     *dev;
 187         int                     msg_bufsize;
 188         int                     scratch_bufsize;
 189 };
 190
 191
 192 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
 193 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
 194 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
 195 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
 196 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
 197 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
 198                                      struct btrfs_fs_info *fs_info,
 199                                      struct scrub_block *original_sblock,
 200                                      u64 length, u64 logical,
 201                                      struct scrub_block *sblocks_for_recheck);
 202 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
 203                                 struct scrub_block *sblock, int is_metadata,
 204                                 int have_csum, u8 *csum, u64 generation,
 205                                 u16 csum_size);
 206 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
 207                                          struct scrub_block *sblock,
 208                                          int is_metadata, int have_csum,
 209                                          const u8 *csum, u64 generation,
 210                                          u16 csum_size);
 211 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
 212                                              struct scrub_block *sblock_good,
 213                                              int force_write);
 214 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
 215                                             struct scrub_block *sblock_good,
 216                                             int page_num, int force_write);
 217 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
 218 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
 219                                            int page_num);
 220 static int scrub_checksum_data(struct scrub_block *sblock);
 221 static int scrub_checksum_tree_block(struct scrub_block *sblock);
 222 static int scrub_checksum_super(struct scrub_block *sblock);
 223 static void scrub_block_get(struct scrub_block *sblock);
 224 static void scrub_block_put(struct scrub_block *sblock);
 225 static void scrub_page_get(struct scrub_page *spage);
 226 static void scrub_page_put(struct scrub_page *spage);
 227 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
 228                                     struct scrub_page *spage);
 229 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
 230                        u64 physical, struct btrfs_device *dev, u64 flags,
 231                        u64 gen, int mirror_num, u8 *csum, int force,
 232                        u64 physical_for_dev_replace);
 233 static void scrub_bio_end_io(struct bio *bio, int err);
 234 static void scrub_bio_end_io_worker(struct btrfs_work *work);
 235 static void scrub_block_complete(struct scrub_block *sblock);
 236 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
 237                                u64 extent_logical, u64 extent_len,
 238                                u64 *extent_physical,
 239                                struct btrfs_device **extent_dev,
 240                                int *extent_mirror_num);
 241 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
 242                               struct scrub_wr_ctx *wr_ctx,
 243                               struct btrfs_fs_info *fs_info,
 244                               struct btrfs_device *dev,
 245                               int is_dev_replace);
 246 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx);
 247 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
 248                                     struct scrub_page *spage);
 249 static void scrub_wr_submit(struct scrub_ctx *sctx);
 250 static void scrub_wr_bio_end_io(struct bio *bio, int err);
 251 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
 252 static int write_page_nocow(struct scrub_ctx *sctx,
 253                             u64 physical_for_dev_replace, struct page *page);
 254 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
 255                                       struct scrub_copy_nocow_ctx *ctx);
 256 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
 257                             int mirror_num, u64 physical_for_dev_replace);
 258 static void copy_nocow_pages_worker(struct btrfs_work *work);
 259
 260
 261 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
 262 {
 263         atomic_inc(&sctx->bios_in_flight);
 264 }
 265
 266 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
 267 {
 268         atomic_dec(&sctx->bios_in_flight);
 269         wake_up(&sctx->list_wait);
 270 }
 271
 272 /*
 273  * used for workers that require transaction commits (i.e., for the
 274  * NOCOW case)
 275  */
 276 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
 277 {
 278         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
 279
 280         /*
 281          * increment scrubs_running to prevent cancel requests from
 282          * completing as long as a worker is running. we must also
 283          * increment scrubs_paused to prevent deadlocking on pause
 284          * requests used for transactions commits (as the worker uses a
 285          * transaction context). it is safe to regard the worker
 286          * as paused for all matters practical. effectively, we only
 287          * avoid cancellation requests from completing.
 288          */
 289         mutex_lock(&fs_info->scrub_lock);
 290         atomic_inc(&fs_info->scrubs_running);
 291         atomic_inc(&fs_info->scrubs_paused);
 292         mutex_unlock(&fs_info->scrub_lock);
 293         atomic_inc(&sctx->workers_pending);
 294 }
 295
 296 /* used for workers that require transaction commits */
 297 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
 298 {
 299         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
 300
 301         /*
 302          * see scrub_pending_trans_workers_inc() why we're pretending
 303          * to be paused in the scrub counters
 304          */
 305         mutex_lock(&fs_info->scrub_lock);
 306         atomic_dec(&fs_info->scrubs_running);
 307         atomic_dec(&fs_info->scrubs_paused);
 308         mutex_unlock(&fs_info->scrub_lock);
 309         atomic_dec(&sctx->workers_pending);
 310         wake_up(&fs_info->scrub_pause_wait);
 311         wake_up(&sctx->list_wait);
 312 }
 313
 314 static void scrub_free_csums(struct scrub_ctx *sctx)
 315 {
 316         while (!list_empty(&sctx->csum_list)) {
 317                 struct btrfs_ordered_sum *sum;
 318                 sum = list_first_entry(&sctx->csum_list,
 319                                        struct btrfs_ordered_sum, list);
 320                 list_del(&sum->list);
 321                 kfree(sum);
 322         }
 323 }
 324
 325 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
 326 {
 327         int i;
 328
 329         if (!sctx)
 330                 return;
 331
 332         scrub_free_wr_ctx(&sctx->wr_ctx);
 333
 334         /* this can happen when scrub is cancelled */
 335         if (sctx->curr != -1) {
 336                 struct scrub_bio *sbio = sctx->bios[sctx->curr];
 337
 338                 for (i = 0; i < sbio->page_count; i++) {
 339                         WARN_ON(!sbio->pagev[i]->page);
 340                         scrub_block_put(sbio->pagev[i]->sblock);
 341                 }
 342                 bio_put(sbio->bio);
 343         }
 344
 345         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 346                 struct scrub_bio *sbio = sctx->bios[i];
 347
 348                 if (!sbio)
 349                         break;
 350                 kfree(sbio);
 351         }
 352
 353         scrub_free_csums(sctx);
 354         kfree(sctx);
 355 }
 356
 357 static noinline_for_stack
 358 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
 359 {
 360         struct scrub_ctx *sctx;
 361         int             i;
 362         struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
 363         int pages_per_rd_bio;
 364         int ret;
 365
 366         /*
 367          * the setting of pages_per_rd_bio is correct for scrub but might
 368          * be wrong for the dev_replace code where we might read from
 369          * different devices in the initial huge bios. However, that
 370          * code is able to correctly handle the case when adding a page
 371          * to a bio fails.
 372          */
 373         if (dev->bdev)
 374                 pages_per_rd_bio = min_t(int, SCRUB_PAGES_PER_RD_BIO,
 375                                          bio_get_nr_vecs(dev->bdev));
 376         else
 377                 pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
 378         sctx = kzalloc(sizeof(*sctx), GFP_NOFS);
 379         if (!sctx)
 380                 goto nomem;
 381         sctx->is_dev_replace = is_dev_replace;
 382         sctx->pages_per_rd_bio = pages_per_rd_bio;
 383         sctx->curr = -1;
 384         sctx->dev_root = dev->dev_root;
 385         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 386                 struct scrub_bio *sbio;
 387
 388                 sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
 389                 if (!sbio)
 390                         goto nomem;
 391                 sctx->bios[i] = sbio;
 392
 393                 sbio->index = i;
 394                 sbio->sctx = sctx;
 395                 sbio->page_count = 0;
 396                 sbio->work.func = scrub_bio_end_io_worker;
 397
 398                 if (i != SCRUB_BIOS_PER_SCTX - 1)
 399                         sctx->bios[i]->next_free = i + 1;
 400                 else
 401                         sctx->bios[i]->next_free = -1;
 402         }
 403         sctx->first_free = 0;
 404         sctx->nodesize = dev->dev_root->nodesize;
 405         sctx->leafsize = dev->dev_root->leafsize;
 406         sctx->sectorsize = dev->dev_root->sectorsize;
 407         atomic_set(&sctx->bios_in_flight, 0);
 408         atomic_set(&sctx->workers_pending, 0);
 409         atomic_set(&sctx->cancel_req, 0);
 410         sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
 411         INIT_LIST_HEAD(&sctx->csum_list);
 412
 413         spin_lock_init(&sctx->list_lock);
 414         spin_lock_init(&sctx->stat_lock);
 415         init_waitqueue_head(&sctx->list_wait);
 416
 417         ret = scrub_setup_wr_ctx(sctx, &sctx->wr_ctx, fs_info,
 418                                  fs_info->dev_replace.tgtdev, is_dev_replace);
 419         if (ret) {
 420                 scrub_free_ctx(sctx);
 421                 return ERR_PTR(ret);
 422         }
 423         return sctx;
 424
 425 nomem:
 426         scrub_free_ctx(sctx);
 427         return ERR_PTR(-ENOMEM);
 428 }
 429
 430 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
 431                                      void *warn_ctx)
 432 {
 433         u64 isize;
 434         u32 nlink;
 435         int ret;
 436         int i;
 437         struct extent_buffer *eb;
 438         struct btrfs_inode_item *inode_item;
 439         struct scrub_warning *swarn = warn_ctx;
 440         struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
 441         struct inode_fs_paths *ipath = NULL;
 442         struct btrfs_root *local_root;
 443         struct btrfs_key root_key;
 444
 445         root_key.objectid = root;
 446         root_key.type = BTRFS_ROOT_ITEM_KEY;
 447         root_key.offset = (u64)-1;
 448         local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
 449         if (IS_ERR(local_root)) {
 450                 ret = PTR_ERR(local_root);
 451                 goto err;
 452         }
 453
 454         ret = inode_item_info(inum, 0, local_root, swarn->path);
 455         if (ret) {
 456                 btrfs_release_path(swarn->path);
 457                 goto err;
 458         }
 459
 460         eb = swarn->path->nodes[0];
 461         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 462                                         struct btrfs_inode_item);
 463         isize = btrfs_inode_size(eb, inode_item);
 464         nlink = btrfs_inode_nlink(eb, inode_item);
 465         btrfs_release_path(swarn->path);
 466
 467         ipath = init_ipath(4096, local_root, swarn->path);
 468         if (IS_ERR(ipath)) {
 469                 ret = PTR_ERR(ipath);
 470                 ipath = NULL;
 471                 goto err;
 472         }
 473         ret = paths_from_inode(inum, ipath);
 474
 475         if (ret < 0)
 476                 goto err;
 477
 478         /*
 479          * we deliberately ignore the bit ipath might have been too small to
 480          * hold all of the paths here
 481          */
 482         for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 483                 printk_in_rcu(KERN_WARNING "btrfs: %s at logical %llu on dev "
 484                         "%s, sector %llu, root %llu, inode %llu, offset %llu, "
 485                         "length %llu, links %u (path: %s)\n", swarn->errstr,
 486                         swarn->logical, rcu_str_deref(swarn->dev->name),
 487                         (unsigned long long)swarn->sector, root, inum, offset,
 488                         min(isize - offset, (u64)PAGE_SIZE), nlink,
 489                         (char *)(unsigned long)ipath->fspath->val[i]);
 490
 491         free_ipath(ipath);
 492         return 0;
 493
 494 err:
 495         printk_in_rcu(KERN_WARNING "btrfs: %s at logical %llu on dev "
 496                 "%s, sector %llu, root %llu, inode %llu, offset %llu: path "
 497                 "resolving failed with ret=%d\n", swarn->errstr,
 498                 swarn->logical, rcu_str_deref(swarn->dev->name),
 499                 (unsigned long long)swarn->sector, root, inum, offset, ret);
 500
 501         free_ipath(ipath);
 502         return 0;
 503 }
 504
 505 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
 506 {
 507         struct btrfs_device *dev;
 508         struct btrfs_fs_info *fs_info;
 509         struct btrfs_path *path;
 510         struct btrfs_key found_key;
 511         struct extent_buffer *eb;
 512         struct btrfs_extent_item *ei;
 513         struct scrub_warning swarn;
 514         unsigned long ptr = 0;
 515         u64 extent_item_pos;
 516         u64 flags = 0;
 517         u64 ref_root;
 518         u32 item_size;
 519         u8 ref_level;
 520         const int bufsize = 4096;
 521         int ret;
 522
 523         WARN_ON(sblock->page_count < 1);
 524         dev = sblock->pagev[0]->dev;
 525         fs_info = sblock->sctx->dev_root->fs_info;
 526
 527         path = btrfs_alloc_path();
 528
 529         swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
 530         swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
 531         swarn.sector = (sblock->pagev[0]->physical) >> 9;
 532         swarn.logical = sblock->pagev[0]->logical;
 533         swarn.errstr = errstr;
 534         swarn.dev = NULL;
 535         swarn.msg_bufsize = bufsize;
 536         swarn.scratch_bufsize = bufsize;
 537
 538         if (!path || !swarn.scratch_buf || !swarn.msg_buf)
 539                 goto out;
 540
 541         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
 542                                   &flags);
 543         if (ret < 0)
 544                 goto out;
 545
 546         extent_item_pos = swarn.logical - found_key.objectid;
 547         swarn.extent_item_size = found_key.offset;
 548
 549         eb = path->nodes[0];
 550         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 551         item_size = btrfs_item_size_nr(eb, path->slots[0]);
 552
 553         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 554                 do {
 555                         ret = tree_backref_for_extent(&ptr, eb, ei, item_size,
 556                                                         &ref_root, &ref_level);
 557                         printk_in_rcu(KERN_WARNING
 558                                 "btrfs: %s at logical %llu on dev %s, "
 559                                 "sector %llu: metadata %s (level %d) in tree "
 560                                 "%llu\n", errstr, swarn.logical,
 561                                 rcu_str_deref(dev->name),
 562                                 (unsigned long long)swarn.sector,
 563                                 ref_level ? "node" : "leaf",
 564                                 ret < 0 ? -1 : ref_level,
 565                                 ret < 0 ? -1 : ref_root);
 566                 } while (ret != 1);
 567                 btrfs_release_path(path);
 568         } else {
 569                 btrfs_release_path(path);
 570                 swarn.path = path;
 571                 swarn.dev = dev;
 572                 iterate_extent_inodes(fs_info, found_key.objectid,
 573                                         extent_item_pos, 1,
 574                                         scrub_print_warning_inode, &swarn);
 575         }
 576
 577 out:
 578         btrfs_free_path(path);
 579         kfree(swarn.scratch_buf);
 580         kfree(swarn.msg_buf);
 581 }
 582
 583 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
 584 {
 585         struct page *page = NULL;
 586         unsigned long index;
 587         struct scrub_fixup_nodatasum *fixup = fixup_ctx;
 588         int ret;
 589         int corrected = 0;
 590         struct btrfs_key key;
 591         struct inode *inode = NULL;
 592         struct btrfs_fs_info *fs_info;
 593         u64 end = offset + PAGE_SIZE - 1;
 594         struct btrfs_root *local_root;
 595         int srcu_index;
 596
 597         key.objectid = root;
 598         key.type = BTRFS_ROOT_ITEM_KEY;
 599         key.offset = (u64)-1;
 600
 601         fs_info = fixup->root->fs_info;
 602         srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
 603
 604         local_root = btrfs_read_fs_root_no_name(fs_info, &key);
 605         if (IS_ERR(local_root)) {
 606                 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
 607                 return PTR_ERR(local_root);
 608         }
 609
 610         key.type = BTRFS_INODE_ITEM_KEY;
 611         key.objectid = inum;
 612         key.offset = 0;
 613         inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
 614         srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
 615         if (IS_ERR(inode))
 616                 return PTR_ERR(inode);
 617
 618         index = offset >> PAGE_CACHE_SHIFT;
 619
 620         page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
 621         if (!page) {
 622                 ret = -ENOMEM;
 623                 goto out;
 624         }
 625
 626         if (PageUptodate(page)) {
 627                 if (PageDirty(page)) {
 628                         /*
 629                          * we need to write the data to the defect sector. the
 630                          * data that was in that sector is not in memory,
 631                          * because the page was modified. we must not write the
 632                          * modified page to that sector.
 633                          *
 634                          * TODO: what could be done here: wait for the delalloc
 635                          *       runner to write out that page (might involve
 636                          *       COW) and see whether the sector is still
 637                          *       referenced afterwards.
 638                          *
 639                          * For the meantime, we'll treat this error
 640                          * incorrectable, although there is a chance that a
 641                          * later scrub will find the bad sector again and that
 642                          * there's no dirty page in memory, then.
 643                          */
 644                         ret = -EIO;
 645                         goto out;
 646                 }
 647                 fs_info = BTRFS_I(inode)->root->fs_info;
 648                 ret = repair_io_failure(fs_info, offset, PAGE_SIZE,
 649                                         fixup->logical, page,
 650                                         fixup->mirror_num);
 651                 unlock_page(page);
 652                 corrected = !ret;
 653         } else {
 654                 /*
 655                  * we need to get good data first. the general readpage path
 656                  * will call repair_io_failure for us, we just have to make
 657                  * sure we read the bad mirror.
 658                  */
 659                 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 660                                         EXTENT_DAMAGED, GFP_NOFS);
 661                 if (ret) {
 662                         /* set_extent_bits should give proper error */
 663                         WARN_ON(ret > 0);
 664                         if (ret > 0)
 665                                 ret = -EFAULT;
 666                         goto out;
 667                 }
 668
 669                 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
 670                                                 btrfs_get_extent,
 671                                                 fixup->mirror_num);
 672                 wait_on_page_locked(page);
 673
 674                 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
 675                                                 end, EXTENT_DAMAGED, 0, NULL);
 676                 if (!corrected)
 677                         clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 678                                                 EXTENT_DAMAGED, GFP_NOFS);
 679         }
 680
 681 out:
 682         if (page)
 683                 put_page(page);
 684         if (inode)
 685                 iput(inode);
 686
 687         if (ret < 0)
 688                 return ret;
 689
 690         if (ret == 0 && corrected) {
 691                 /*
 692                  * we only need to call readpage for one of the inodes belonging
 693                  * to this extent. so make iterate_extent_inodes stop
 694                  */
 695                 return 1;
 696         }
 697
 698         return -EIO;
 699 }
 700
 701 static void scrub_fixup_nodatasum(struct btrfs_work *work)
 702 {
 703         int ret;
 704         struct scrub_fixup_nodatasum *fixup;
 705         struct scrub_ctx *sctx;
 706         struct btrfs_trans_handle *trans = NULL;
 707         struct btrfs_fs_info *fs_info;
 708         struct btrfs_path *path;
 709         int uncorrectable = 0;
 710
 711         fixup = container_of(work, struct scrub_fixup_nodatasum, work);
 712         sctx = fixup->sctx;
 713         fs_info = fixup->root->fs_info;
 714
 715         path = btrfs_alloc_path();
 716         if (!path) {
 717                 spin_lock(&sctx->stat_lock);
 718                 ++sctx->stat.malloc_errors;
 719                 spin_unlock(&sctx->stat_lock);
 720                 uncorrectable = 1;
 721                 goto out;
 722         }
 723
 724         trans = btrfs_join_transaction(fixup->root);
 725         if (IS_ERR(trans)) {
 726                 uncorrectable = 1;
 727                 goto out;
 728         }
 729
 730         /*
 731          * the idea is to trigger a regular read through the standard path. we
 732          * read a page from the (failed) logical address by specifying the
 733          * corresponding copynum of the failed sector. thus, that readpage is
 734          * expected to fail.
 735          * that is the point where on-the-fly error correction will kick in
 736          * (once it's finished) and rewrite the failed sector if a good copy
 737          * can be found.
 738          */
 739         ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
 740                                                 path, scrub_fixup_readpage,
 741                                                 fixup);
 742         if (ret < 0) {
 743                 uncorrectable = 1;
 744                 goto out;
 745         }
 746         WARN_ON(ret != 1);
 747
 748         spin_lock(&sctx->stat_lock);
 749         ++sctx->stat.corrected_errors;
 750         spin_unlock(&sctx->stat_lock);
 751
 752 out:
 753         if (trans && !IS_ERR(trans))
 754                 btrfs_end_transaction(trans, fixup->root);
 755         if (uncorrectable) {
 756                 spin_lock(&sctx->stat_lock);
 757                 ++sctx->stat.uncorrectable_errors;
 758                 spin_unlock(&sctx->stat_lock);
 759                 btrfs_dev_replace_stats_inc(
 760                         &sctx->dev_root->fs_info->dev_replace.
 761                         num_uncorrectable_read_errors);
 762                 printk_ratelimited_in_rcu(KERN_ERR
 763                         "btrfs: unable to fixup (nodatasum) error at logical %llu on dev %s\n",
 764                         fixup->logical, rcu_str_deref(fixup->dev->name));
 765         }
 766
 767         btrfs_free_path(path);
 768         kfree(fixup);
 769
 770         scrub_pending_trans_workers_dec(sctx);
 771 }
 772
 773 /*
 774  * scrub_handle_errored_block gets called when either verification of the
 775  * pages failed or the bio failed to read, e.g. with EIO. In the latter
 776  * case, this function handles all pages in the bio, even though only one
 777  * may be bad.
 778  * The goal of this function is to repair the errored block by using the
 779  * contents of one of the mirrors.
 780  */
 781 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
 782 {
 783         struct scrub_ctx *sctx = sblock_to_check->sctx;
 784         struct btrfs_device *dev;
 785         struct btrfs_fs_info *fs_info;
 786         u64 length;
 787         u64 logical;
 788         u64 generation;
 789         unsigned int failed_mirror_index;
 790         unsigned int is_metadata;
 791         unsigned int have_csum;
 792         u8 *csum;
 793         struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
 794         struct scrub_block *sblock_bad;
 795         int ret;
 796         int mirror_index;
 797         int page_num;
 798         int success;
 799         static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
 800                                       DEFAULT_RATELIMIT_BURST);
 801
 802         BUG_ON(sblock_to_check->page_count < 1);
 803         fs_info = sctx->dev_root->fs_info;
 804         if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
 805                 /*
 806                  * if we find an error in a super block, we just report it.
 807                  * They will get written with the next transaction commit
 808                  * anyway
 809                  */
 810                 spin_lock(&sctx->stat_lock);
 811                 ++sctx->stat.super_errors;
 812                 spin_unlock(&sctx->stat_lock);
 813                 return 0;
 814         }
 815         length = sblock_to_check->page_count * PAGE_SIZE;
 816         logical = sblock_to_check->pagev[0]->logical;
 817         generation = sblock_to_check->pagev[0]->generation;
 818         BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
 819         failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
 820         is_metadata = !(sblock_to_check->pagev[0]->flags &
 821                         BTRFS_EXTENT_FLAG_DATA);
 822         have_csum = sblock_to_check->pagev[0]->have_csum;
 823         csum = sblock_to_check->pagev[0]->csum;
 824         dev = sblock_to_check->pagev[0]->dev;
 825
 826         if (sctx->is_dev_replace && !is_metadata && !have_csum) {
 827                 sblocks_for_recheck = NULL;
 828                 goto nodatasum_case;
 829         }
 830
 831         /*
 832          * read all mirrors one after the other. This includes to
 833          * re-read the extent or metadata block that failed (that was
 834          * the cause that this fixup code is called) another time,
 835          * page by page this time in order to know which pages
 836          * caused I/O errors and which ones are good (for all mirrors).
 837          * It is the goal to handle the situation when more than one
 838          * mirror contains I/O errors, but the errors do not
 839          * overlap, i.e. the data can be repaired by selecting the
 840          * pages from those mirrors without I/O error on the
 841          * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
 842          * would be that mirror #1 has an I/O error on the first page,
 843          * the second page is good, and mirror #2 has an I/O error on
 844          * the second page, but the first page is good.
 845          * Then the first page of the first mirror can be repaired by
 846          * taking the first page of the second mirror, and the
 847          * second page of the second mirror can be repaired by
 848          * copying the contents of the 2nd page of the 1st mirror.
 849          * One more note: if the pages of one mirror contain I/O
 850          * errors, the checksum cannot be verified. In order to get
 851          * the best data for repairing, the first attempt is to find
 852          * a mirror without I/O errors and with a validated checksum.
 853          * Only if this is not possible, the pages are picked from
 854          * mirrors with I/O errors without considering the checksum.
 855          * If the latter is the case, at the end, the checksum of the
 856          * repaired area is verified in order to correctly maintain
 857          * the statistics.
 858          */
 859
 860         sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
 861                                      sizeof(*sblocks_for_recheck),
 862                                      GFP_NOFS);
 863         if (!sblocks_for_recheck) {
 864                 spin_lock(&sctx->stat_lock);
 865                 sctx->stat.malloc_errors++;
 866                 sctx->stat.read_errors++;
 867                 sctx->stat.uncorrectable_errors++;
 868                 spin_unlock(&sctx->stat_lock);
 869                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 870                 goto out;
 871         }
 872
 873         /* setup the context, map the logical blocks and alloc the pages */
 874         ret = scrub_setup_recheck_block(sctx, fs_info, sblock_to_check, length,
 875                                         logical, sblocks_for_recheck);
 876         if (ret) {
 877                 spin_lock(&sctx->stat_lock);
 878                 sctx->stat.read_errors++;
 879                 sctx->stat.uncorrectable_errors++;
 880                 spin_unlock(&sctx->stat_lock);
 881                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 882                 goto out;
 883         }
 884         BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
 885         sblock_bad = sblocks_for_recheck + failed_mirror_index;
 886
 887         /* build and submit the bios for the failed mirror, check checksums */
 888         scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
 889                             csum, generation, sctx->csum_size);
 890
 891         if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
 892             sblock_bad->no_io_error_seen) {
 893                 /*
 894                  * the error disappeared after reading page by page, or
 895                  * the area was part of a huge bio and other parts of the
 896                  * bio caused I/O errors, or the block layer merged several
 897                  * read requests into one and the error is caused by a
 898                  * different bio (usually one of the two latter cases is
 899                  * the cause)
 900                  */
 901                 spin_lock(&sctx->stat_lock);
 902                 sctx->stat.unverified_errors++;
 903                 spin_unlock(&sctx->stat_lock);
 904
 905                 if (sctx->is_dev_replace)
 906                         scrub_write_block_to_dev_replace(sblock_bad);
 907                 goto out;
 908         }
 909
 910         if (!sblock_bad->no_io_error_seen) {
 911                 spin_lock(&sctx->stat_lock);
 912                 sctx->stat.read_errors++;
 913                 spin_unlock(&sctx->stat_lock);
 914                 if (__ratelimit(&_rs))
 915                         scrub_print_warning("i/o error", sblock_to_check);
 916                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 917         } else if (sblock_bad->checksum_error) {
 918                 spin_lock(&sctx->stat_lock);
 919                 sctx->stat.csum_errors++;
 920                 spin_unlock(&sctx->stat_lock);
 921                 if (__ratelimit(&_rs))
 922                         scrub_print_warning("checksum error", sblock_to_check);
 923                 btrfs_dev_stat_inc_and_print(dev,
 924                                              BTRFS_DEV_STAT_CORRUPTION_ERRS);
 925         } else if (sblock_bad->header_error) {
 926                 spin_lock(&sctx->stat_lock);
 927                 sctx->stat.verify_errors++;
 928                 spin_unlock(&sctx->stat_lock);
 929                 if (__ratelimit(&_rs))
 930                         scrub_print_warning("checksum/header error",
 931                                             sblock_to_check);
 932                 if (sblock_bad->generation_error)
 933                         btrfs_dev_stat_inc_and_print(dev,
 934                                 BTRFS_DEV_STAT_GENERATION_ERRS);
 935                 else
 936                         btrfs_dev_stat_inc_and_print(dev,
 937                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
 938         }
 939
 940         if (sctx->readonly) {
 941                 ASSERT(!sctx->is_dev_replace);
 942                 goto out;
 943         }
 944
 945         if (!is_metadata && !have_csum) {
 946                 struct scrub_fixup_nodatasum *fixup_nodatasum;
 947
 948 nodatasum_case:
 949                 WARN_ON(sctx->is_dev_replace);
 950
 951                 /*
 952                  * !is_metadata and !have_csum, this means that the data
 953                  * might not be COW'ed, that it might be modified
 954                  * concurrently. The general strategy to work on the
 955                  * commit root does not help in the case when COW is not
 956                  * used.
 957                  */
 958                 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
 959                 if (!fixup_nodatasum)
 960                         goto did_not_correct_error;
 961                 fixup_nodatasum->sctx = sctx;
 962                 fixup_nodatasum->dev = dev;
 963                 fixup_nodatasum->logical = logical;
 964                 fixup_nodatasum->root = fs_info->extent_root;
 965                 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
 966                 scrub_pending_trans_workers_inc(sctx);
 967                 fixup_nodatasum->work.func = scrub_fixup_nodatasum;
 968                 btrfs_queue_worker(&fs_info->scrub_workers,
 969                                    &fixup_nodatasum->work);
 970                 goto out;
 971         }
 972
 973         /*
 974          * now build and submit the bios for the other mirrors, check
 975          * checksums.
 976          * First try to pick the mirror which is completely without I/O
 977          * errors and also does not have a checksum error.
 978          * If one is found, and if a checksum is present, the full block
 979          * that is known to contain an error is rewritten. Afterwards
 980          * the block is known to be corrected.
 981          * If a mirror is found which is completely correct, and no
 982          * checksum is present, only those pages are rewritten that had
 983          * an I/O error in the block to be repaired, since it cannot be
 984          * determined, which copy of the other pages is better (and it
 985          * could happen otherwise that a correct page would be
 986          * overwritten by a bad one).
 987          */
 988         for (mirror_index = 0;
 989              mirror_index < BTRFS_MAX_MIRRORS &&
 990              sblocks_for_recheck[mirror_index].page_count > 0;
 991              mirror_index++) {
 992                 struct scrub_block *sblock_other;
 993
 994                 if (mirror_index == failed_mirror_index)
 995                         continue;
 996                 sblock_other = sblocks_for_recheck + mirror_index;
 997
 998                 /* build and submit the bios, check checksums */
 999                 scrub_recheck_block(fs_info, sblock_other, is_metadata,
1000                                     have_csum, csum, generation,
1001                                     sctx->csum_size);
1002
1003                 if (!sblock_other->header_error &&
1004                     !sblock_other->checksum_error &&
1005                     sblock_other->no_io_error_seen) {
1006                         if (sctx->is_dev_replace) {
1007                                 scrub_write_block_to_dev_replace(sblock_other);
1008                         } else {
1009                                 int force_write = is_metadata || have_csum;
1010
1011                                 ret = scrub_repair_block_from_good_copy(
1012                                                 sblock_bad, sblock_other,
1013                                                 force_write);
1014                         }
1015                         if (0 == ret)
1016                                 goto corrected_error;
1017                 }
1018         }
1019
1020         /*
1021          * for dev_replace, pick good pages and write to the target device.
1022          */
1023         if (sctx->is_dev_replace) {
1024                 success = 1;
1025                 for (page_num = 0; page_num < sblock_bad->page_count;
1026                      page_num++) {
1027                         int sub_success;
1028
1029                         sub_success = 0;
1030                         for (mirror_index = 0;
1031                              mirror_index < BTRFS_MAX_MIRRORS &&
1032                              sblocks_for_recheck[mirror_index].page_count > 0;
1033                              mirror_index++) {
1034                                 struct scrub_block *sblock_other =
1035                                         sblocks_for_recheck + mirror_index;
1036                                 struct scrub_page *page_other =
1037                                         sblock_other->pagev[page_num];
1038
1039                                 if (!page_other->io_error) {
1040                                         ret = scrub_write_page_to_dev_replace(
1041                                                         sblock_other, page_num);
1042                                         if (ret == 0) {
1043                                                 /* succeeded for this page */
1044                                                 sub_success = 1;
1045                                                 break;
1046                                         } else {
1047                                                 btrfs_dev_replace_stats_inc(
1048                                                         &sctx->dev_root->
1049                                                         fs_info->dev_replace.
1050                                                         num_write_errors);
1051                                         }
1052                                 }
1053                         }
1054
1055                         if (!sub_success) {
1056                                 /*
1057                                  * did not find a mirror to fetch the page
1058                                  * from. scrub_write_page_to_dev_replace()
1059                                  * handles this case (page->io_error), by
1060                                  * filling the block with zeros before
1061                                  * submitting the write request
1062                                  */
1063                                 success = 0;
1064                                 ret = scrub_write_page_to_dev_replace(
1065                                                 sblock_bad, page_num);
1066                                 if (ret)
1067                                         btrfs_dev_replace_stats_inc(
1068                                                 &sctx->dev_root->fs_info->
1069                                                 dev_replace.num_write_errors);
1070                         }
1071                 }
1072
1073                 goto out;
1074         }
1075
1076         /*
1077          * for regular scrub, repair those pages that are errored.
1078          * In case of I/O errors in the area that is supposed to be
1079          * repaired, continue by picking good copies of those pages.
1080          * Select the good pages from mirrors to rewrite bad pages from
1081          * the area to fix. Afterwards verify the checksum of the block
1082          * that is supposed to be repaired. This verification step is
1083          * only done for the purpose of statistic counting and for the
1084          * final scrub report, whether errors remain.
1085          * A perfect algorithm could make use of the checksum and try
1086          * all possible combinations of pages from the different mirrors
1087          * until the checksum verification succeeds. For example, when
1088          * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1089          * of mirror #2 is readable but the final checksum test fails,
1090          * then the 2nd page of mirror #3 could be tried, whether now
1091          * the final checksum succeedes. But this would be a rare
1092          * exception and is therefore not implemented. At least it is
1093          * avoided that the good copy is overwritten.
1094          * A more useful improvement would be to pick the sectors
1095          * without I/O error based on sector sizes (512 bytes on legacy
1096          * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1097          * mirror could be repaired by taking 512 byte of a different
1098          * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1099          * area are unreadable.
1100          */
1101
1102         /* can only fix I/O errors from here on */
1103         if (sblock_bad->no_io_error_seen)
1104                 goto did_not_correct_error;
1105
1106         success = 1;
1107         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1108                 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1109
1110                 if (!page_bad->io_error)
1111                         continue;
1112
1113                 for (mirror_index = 0;
1114                      mirror_index < BTRFS_MAX_MIRRORS &&
1115                      sblocks_for_recheck[mirror_index].page_count > 0;
1116                      mirror_index++) {
1117                         struct scrub_block *sblock_other = sblocks_for_recheck +
1118                                                            mirror_index;
1119                         struct scrub_page *page_other = sblock_other->pagev[
1120                                                         page_num];
1121
1122                         if (!page_other->io_error) {
1123                                 ret = scrub_repair_page_from_good_copy(
1124                                         sblock_bad, sblock_other, page_num, 0);
1125                                 if (0 == ret) {
1126                                         page_bad->io_error = 0;
1127                                         break; /* succeeded for this page */
1128                                 }
1129                         }
1130                 }
1131
1132                 if (page_bad->io_error) {
1133                         /* did not find a mirror to copy the page from */
1134                         success = 0;
1135                 }
1136         }
1137
1138         if (success) {
1139                 if (is_metadata || have_csum) {
1140                         /*
1141                          * need to verify the checksum now that all
1142                          * sectors on disk are repaired (the write
1143                          * request for data to be repaired is on its way).
1144                          * Just be lazy and use scrub_recheck_block()
1145                          * which re-reads the data before the checksum
1146                          * is verified, but most likely the data comes out
1147                          * of the page cache.
1148                          */
1149                         scrub_recheck_block(fs_info, sblock_bad,
1150                                             is_metadata, have_csum, csum,
1151                                             generation, sctx->csum_size);
1152                         if (!sblock_bad->header_error &&
1153                             !sblock_bad->checksum_error &&
1154                             sblock_bad->no_io_error_seen)
1155                                 goto corrected_error;
1156                         else
1157                                 goto did_not_correct_error;
1158                 } else {
1159 corrected_error:
1160                         spin_lock(&sctx->stat_lock);
1161                         sctx->stat.corrected_errors++;
1162                         spin_unlock(&sctx->stat_lock);
1163                         printk_ratelimited_in_rcu(KERN_ERR
1164                                 "btrfs: fixed up error at logical %llu on dev %s\n",
1165                                 logical, rcu_str_deref(dev->name));
1166                 }
1167         } else {
1168 did_not_correct_error:
1169                 spin_lock(&sctx->stat_lock);
1170                 sctx->stat.uncorrectable_errors++;
1171                 spin_unlock(&sctx->stat_lock);
1172                 printk_ratelimited_in_rcu(KERN_ERR
1173                         "btrfs: unable to fixup (regular) error at logical %llu on dev %s\n",
1174                         logical, rcu_str_deref(dev->name));
1175         }
1176
1177 out:
1178         if (sblocks_for_recheck) {
1179                 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1180                      mirror_index++) {
1181                         struct scrub_block *sblock = sblocks_for_recheck +
1182                                                      mirror_index;
1183                         int page_index;
1184
1185                         for (page_index = 0; page_index < sblock->page_count;
1186                              page_index++) {
1187                                 sblock->pagev[page_index]->sblock = NULL;
1188                                 scrub_page_put(sblock->pagev[page_index]);
1189                         }
1190                 }
1191                 kfree(sblocks_for_recheck);
1192         }
1193
1194         return 0;
1195 }
1196
1197 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
1198                                      struct btrfs_fs_info *fs_info,
1199                                      struct scrub_block *original_sblock,
1200                                      u64 length, u64 logical,
1201                                      struct scrub_block *sblocks_for_recheck)
1202 {
1203         int page_index;
1204         int mirror_index;
1205         int ret;
1206
1207         /*
1208          * note: the two members ref_count and outstanding_pages
1209          * are not used (and not set) in the blocks that are used for
1210          * the recheck procedure
1211          */
1212
1213         page_index = 0;
1214         while (length > 0) {
1215                 u64 sublen = min_t(u64, length, PAGE_SIZE);
1216                 u64 mapped_length = sublen;
1217                 struct btrfs_bio *bbio = NULL;
1218
1219                 /*
1220                  * with a length of PAGE_SIZE, each returned stripe
1221                  * represents one mirror
1222                  */
1223                 ret = btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS, logical,
1224                                       &mapped_length, &bbio, 0);
1225                 if (ret || !bbio || mapped_length < sublen) {
1226                         kfree(bbio);
1227                         return -EIO;
1228                 }
1229
1230                 BUG_ON(page_index >= SCRUB_PAGES_PER_RD_BIO);
1231                 for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
1232                      mirror_index++) {
1233                         struct scrub_block *sblock;
1234                         struct scrub_page *page;
1235
1236                         if (mirror_index >= BTRFS_MAX_MIRRORS)
1237                                 continue;
1238
1239                         sblock = sblocks_for_recheck + mirror_index;
1240                         sblock->sctx = sctx;
1241                         page = kzalloc(sizeof(*page), GFP_NOFS);
1242                         if (!page) {
1243 leave_nomem:
1244                                 spin_lock(&sctx->stat_lock);
1245                                 sctx->stat.malloc_errors++;
1246                                 spin_unlock(&sctx->stat_lock);
1247                                 kfree(bbio);
1248                                 return -ENOMEM;
1249                         }
1250                         scrub_page_get(page);
1251                         sblock->pagev[page_index] = page;
1252                         page->logical = logical;
1253                         page->physical = bbio->stripes[mirror_index].physical;
1254                         BUG_ON(page_index >= original_sblock->page_count);
1255                         page->physical_for_dev_replace =
1256                                 original_sblock->pagev[page_index]->
1257                                 physical_for_dev_replace;
1258                         /* for missing devices, dev->bdev is NULL */
1259                         page->dev = bbio->stripes[mirror_index].dev;
1260                         page->mirror_num = mirror_index + 1;
1261                         sblock->page_count++;
1262                         page->page = alloc_page(GFP_NOFS);
1263                         if (!page->page)
1264                                 goto leave_nomem;
1265                 }
1266                 kfree(bbio);
1267                 length -= sublen;
1268                 logical += sublen;
1269                 page_index++;
1270         }
1271
1272         return 0;
1273 }
1274
1275 /*
1276  * this function will check the on disk data for checksum errors, header
1277  * errors and read I/O errors. If any I/O errors happen, the exact pages
1278  * which are errored are marked as being bad. The goal is to enable scrub
1279  * to take those pages that are not errored from all the mirrors so that
1280  * the pages that are errored in the just handled mirror can be repaired.
1281  */
1282 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1283                                 struct scrub_block *sblock, int is_metadata,
1284                                 int have_csum, u8 *csum, u64 generation,
1285                                 u16 csum_size)
1286 {
1287         int page_num;
1288
1289         sblock->no_io_error_seen = 1;
1290         sblock->header_error = 0;
1291         sblock->checksum_error = 0;
1292
1293         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1294                 struct bio *bio;
1295                 struct scrub_page *page = sblock->pagev[page_num];
1296
1297                 if (page->dev->bdev == NULL) {
1298                         page->io_error = 1;
1299                         sblock->no_io_error_seen = 0;
1300                         continue;
1301                 }
1302
1303                 WARN_ON(!page->page);
1304                 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1305                 if (!bio) {
1306                         page->io_error = 1;
1307                         sblock->no_io_error_seen = 0;
1308                         continue;
1309                 }
1310                 bio->bi_bdev = page->dev->bdev;
1311                 bio->bi_sector = page->physical >> 9;
1312
1313                 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1314                 if (btrfsic_submit_bio_wait(READ, bio))
1315                         sblock->no_io_error_seen = 0;
1316
1317                 bio_put(bio);
1318         }
1319
1320         if (sblock->no_io_error_seen)
1321                 scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1322                                              have_csum, csum, generation,
1323                                              csum_size);
1324
1325         return;
1326 }
1327
1328 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1329                                          struct scrub_block *sblock,
1330                                          int is_metadata, int have_csum,
1331                                          const u8 *csum, u64 generation,
1332                                          u16 csum_size)
1333 {
1334         int page_num;
1335         u8 calculated_csum[BTRFS_CSUM_SIZE];
1336         u32 crc = ~(u32)0;
1337         void *mapped_buffer;
1338
1339         WARN_ON(!sblock->pagev[0]->page);
1340         if (is_metadata) {
1341                 struct btrfs_header *h;
1342
1343                 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1344                 h = (struct btrfs_header *)mapped_buffer;
1345
1346                 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h) ||
1347                     memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
1348                     memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1349                            BTRFS_UUID_SIZE)) {
1350                         sblock->header_error = 1;
1351                 } else if (generation != btrfs_stack_header_generation(h)) {
1352                         sblock->header_error = 1;
1353                         sblock->generation_error = 1;
1354                 }
1355                 csum = h->csum;
1356         } else {
1357                 if (!have_csum)
1358                         return;
1359
1360                 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1361         }
1362
1363         for (page_num = 0;;) {
1364                 if (page_num == 0 && is_metadata)
1365                         crc = btrfs_csum_data(
1366                                 ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1367                                 crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1368                 else
1369                         crc = btrfs_csum_data(mapped_buffer, crc, PAGE_SIZE);
1370
1371                 kunmap_atomic(mapped_buffer);
1372                 page_num++;
1373                 if (page_num >= sblock->page_count)
1374                         break;
1375                 WARN_ON(!sblock->pagev[page_num]->page);
1376
1377                 mapped_buffer = kmap_atomic(sblock->pagev[page_num]->page);
1378         }
1379
1380         btrfs_csum_final(crc, calculated_csum);
1381         if (memcmp(calculated_csum, csum, csum_size))
1382                 sblock->checksum_error = 1;
1383 }
1384
1385 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1386                                              struct scrub_block *sblock_good,
1387                                              int force_write)
1388 {
1389         int page_num;
1390         int ret = 0;
1391
1392         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1393                 int ret_sub;
1394
1395                 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1396                                                            sblock_good,
1397                                                            page_num,
1398                                                            force_write);
1399                 if (ret_sub)
1400                         ret = ret_sub;
1401         }
1402
1403         return ret;
1404 }
1405
1406 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1407                                             struct scrub_block *sblock_good,
1408                                             int page_num, int force_write)
1409 {
1410         struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1411         struct scrub_page *page_good = sblock_good->pagev[page_num];
1412
1413         BUG_ON(page_bad->page == NULL);
1414         BUG_ON(page_good->page == NULL);
1415         if (force_write || sblock_bad->header_error ||
1416             sblock_bad->checksum_error || page_bad->io_error) {
1417                 struct bio *bio;
1418                 int ret;
1419
1420                 if (!page_bad->dev->bdev) {
1421                         printk_ratelimited(KERN_WARNING
1422                                 "btrfs: scrub_repair_page_from_good_copy(bdev == NULL) is unexpected!\n");
1423                         return -EIO;
1424                 }
1425
1426                 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1427                 if (!bio)
1428                         return -EIO;
1429                 bio->bi_bdev = page_bad->dev->bdev;
1430                 bio->bi_sector = page_bad->physical >> 9;
1431
1432                 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1433                 if (PAGE_SIZE != ret) {
1434                         bio_put(bio);
1435                         return -EIO;
1436                 }
1437
1438                 if (btrfsic_submit_bio_wait(WRITE, bio)) {
1439                         btrfs_dev_stat_inc_and_print(page_bad->dev,
1440                                 BTRFS_DEV_STAT_WRITE_ERRS);
1441                         btrfs_dev_replace_stats_inc(
1442                                 &sblock_bad->sctx->dev_root->fs_info->
1443                                 dev_replace.num_write_errors);
1444                         bio_put(bio);
1445                         return -EIO;
1446                 }
1447                 bio_put(bio);
1448         }
1449
1450         return 0;
1451 }
1452
1453 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1454 {
1455         int page_num;
1456
1457         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1458                 int ret;
1459
1460                 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1461                 if (ret)
1462                         btrfs_dev_replace_stats_inc(
1463                                 &sblock->sctx->dev_root->fs_info->dev_replace.
1464                                 num_write_errors);
1465         }
1466 }
1467
1468 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1469                                            int page_num)
1470 {
1471         struct scrub_page *spage = sblock->pagev[page_num];
1472
1473         BUG_ON(spage->page == NULL);
1474         if (spage->io_error) {
1475                 void *mapped_buffer = kmap_atomic(spage->page);
1476
1477                 memset(mapped_buffer, 0, PAGE_CACHE_SIZE);
1478                 flush_dcache_page(spage->page);
1479                 kunmap_atomic(mapped_buffer);
1480         }
1481         return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1482 }
1483
1484 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1485                                     struct scrub_page *spage)
1486 {
1487         struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1488         struct scrub_bio *sbio;
1489         int ret;
1490
1491         mutex_lock(&wr_ctx->wr_lock);
1492 again:
1493         if (!wr_ctx->wr_curr_bio) {
1494                 wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1495                                               GFP_NOFS);
1496                 if (!wr_ctx->wr_curr_bio) {
1497                         mutex_unlock(&wr_ctx->wr_lock);
1498                         return -ENOMEM;
1499                 }
1500                 wr_ctx->wr_curr_bio->sctx = sctx;
1501                 wr_ctx->wr_curr_bio->page_count = 0;
1502         }
1503         sbio = wr_ctx->wr_curr_bio;
1504         if (sbio->page_count == 0) {
1505                 struct bio *bio;
1506
1507                 sbio->physical = spage->physical_for_dev_replace;
1508                 sbio->logical = spage->logical;
1509                 sbio->dev = wr_ctx->tgtdev;
1510                 bio = sbio->bio;
1511                 if (!bio) {
1512                         bio = btrfs_io_bio_alloc(GFP_NOFS, wr_ctx->pages_per_wr_bio);
1513                         if (!bio) {
1514                                 mutex_unlock(&wr_ctx->wr_lock);
1515                                 return -ENOMEM;
1516                         }
1517                         sbio->bio = bio;
1518                 }
1519
1520                 bio->bi_private = sbio;
1521                 bio->bi_end_io = scrub_wr_bio_end_io;
1522                 bio->bi_bdev = sbio->dev->bdev;
1523                 bio->bi_sector = sbio->physical >> 9;
1524                 sbio->err = 0;
1525         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1526                    spage->physical_for_dev_replace ||
1527                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1528                    spage->logical) {
1529                 scrub_wr_submit(sctx);
1530                 goto again;
1531         }
1532
1533         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1534         if (ret != PAGE_SIZE) {
1535                 if (sbio->page_count < 1) {
1536                         bio_put(sbio->bio);
1537                         sbio->bio = NULL;
1538                         mutex_unlock(&wr_ctx->wr_lock);
1539                         return -EIO;
1540                 }
1541                 scrub_wr_submit(sctx);
1542                 goto again;
1543         }
1544
1545         sbio->pagev[sbio->page_count] = spage;
1546         scrub_page_get(spage);
1547         sbio->page_count++;
1548         if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1549                 scrub_wr_submit(sctx);
1550         mutex_unlock(&wr_ctx->wr_lock);
1551
1552         return 0;
1553 }
1554
1555 static void scrub_wr_submit(struct scrub_ctx *sctx)
1556 {
1557         struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1558         struct scrub_bio *sbio;
1559
1560         if (!wr_ctx->wr_curr_bio)
1561                 return;
1562
1563         sbio = wr_ctx->wr_curr_bio;
1564         wr_ctx->wr_curr_bio = NULL;
1565         WARN_ON(!sbio->bio->bi_bdev);
1566         scrub_pending_bio_inc(sctx);
1567         /* process all writes in a single worker thread. Then the block layer
1568          * orders the requests before sending them to the driver which
1569          * doubled the write performance on spinning disks when measured
1570          * with Linux 3.5 */
1571         btrfsic_submit_bio(WRITE, sbio->bio);
1572 }
1573
1574 static void scrub_wr_bio_end_io(struct bio *bio, int err)
1575 {
1576         struct scrub_bio *sbio = bio->bi_private;
1577         struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
1578
1579         sbio->err = err;
1580         sbio->bio = bio;
1581
1582         sbio->work.func = scrub_wr_bio_end_io_worker;
1583         btrfs_queue_worker(&fs_info->scrub_wr_completion_workers, &sbio->work);
1584 }
1585
1586 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1587 {
1588         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1589         struct scrub_ctx *sctx = sbio->sctx;
1590         int i;
1591
1592         WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1593         if (sbio->err) {
1594                 struct btrfs_dev_replace *dev_replace =
1595                         &sbio->sctx->dev_root->fs_info->dev_replace;
1596
1597                 for (i = 0; i < sbio->page_count; i++) {
1598                         struct scrub_page *spage = sbio->pagev[i];
1599
1600                         spage->io_error = 1;
1601                         btrfs_dev_replace_stats_inc(&dev_replace->
1602                                                     num_write_errors);
1603                 }
1604         }
1605
1606         for (i = 0; i < sbio->page_count; i++)
1607                 scrub_page_put(sbio->pagev[i]);
1608
1609         bio_put(sbio->bio);
1610         kfree(sbio);
1611         scrub_pending_bio_dec(sctx);
1612 }
1613
1614 static int scrub_checksum(struct scrub_block *sblock)
1615 {
1616         u64 flags;
1617         int ret;
1618
1619         WARN_ON(sblock->page_count < 1);
1620         flags = sblock->pagev[0]->flags;
1621         ret = 0;
1622         if (flags & BTRFS_EXTENT_FLAG_DATA)
1623                 ret = scrub_checksum_data(sblock);
1624         else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1625                 ret = scrub_checksum_tree_block(sblock);
1626         else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1627                 (void)scrub_checksum_super(sblock);
1628         else
1629                 WARN_ON(1);
1630         if (ret)
1631                 scrub_handle_errored_block(sblock);
1632
1633         return ret;
1634 }
1635
1636 static int scrub_checksum_data(struct scrub_block *sblock)
1637 {
1638         struct scrub_ctx *sctx = sblock->sctx;
1639         u8 csum[BTRFS_CSUM_SIZE];
1640         u8 *on_disk_csum;
1641         struct page *page;
1642         void *buffer;
1643         u32 crc = ~(u32)0;
1644         int fail = 0;
1645         u64 len;
1646         int index;
1647
1648         BUG_ON(sblock->page_count < 1);
1649         if (!sblock->pagev[0]->have_csum)
1650                 return 0;
1651
1652         on_disk_csum = sblock->pagev[0]->csum;
1653         page = sblock->pagev[0]->page;
1654         buffer = kmap_atomic(page);
1655
1656         len = sctx->sectorsize;
1657         index = 0;
1658         for (;;) {
1659                 u64 l = min_t(u64, len, PAGE_SIZE);
1660
1661                 crc = btrfs_csum_data(buffer, crc, l);
1662                 kunmap_atomic(buffer);
1663                 len -= l;
1664                 if (len == 0)
1665                         break;
1666                 index++;
1667                 BUG_ON(index >= sblock->page_count);
1668                 BUG_ON(!sblock->pagev[index]->page);
1669                 page = sblock->pagev[index]->page;
1670                 buffer = kmap_atomic(page);
1671         }
1672
1673         btrfs_csum_final(crc, csum);
1674         if (memcmp(csum, on_disk_csum, sctx->csum_size))
1675                 fail = 1;
1676
1677         return fail;
1678 }
1679
1680 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1681 {
1682         struct scrub_ctx *sctx = sblock->sctx;
1683         struct btrfs_header *h;
1684         struct btrfs_root *root = sctx->dev_root;
1685         struct btrfs_fs_info *fs_info = root->fs_info;
1686         u8 calculated_csum[BTRFS_CSUM_SIZE];
1687         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1688         struct page *page;
1689         void *mapped_buffer;
1690         u64 mapped_size;
1691         void *p;
1692         u32 crc = ~(u32)0;
1693         int fail = 0;
1694         int crc_fail = 0;
1695         u64 len;
1696         int index;
1697
1698         BUG_ON(sblock->page_count < 1);
1699         page = sblock->pagev[0]->page;
1700         mapped_buffer = kmap_atomic(page);
1701         h = (struct btrfs_header *)mapped_buffer;
1702         memcpy(on_disk_csum, h->csum, sctx->csum_size);
1703
1704         /*
1705          * we don't use the getter functions here, as we
1706          * a) don't have an extent buffer and
1707          * b) the page is already kmapped
1708          */
1709
1710         if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1711                 ++fail;
1712
1713         if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h))
1714                 ++fail;
1715
1716         if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1717                 ++fail;
1718
1719         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1720                    BTRFS_UUID_SIZE))
1721                 ++fail;
1722
1723         WARN_ON(sctx->nodesize != sctx->leafsize);
1724         len = sctx->nodesize - BTRFS_CSUM_SIZE;
1725         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1726         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1727         index = 0;
1728         for (;;) {
1729                 u64 l = min_t(u64, len, mapped_size);
1730
1731                 crc = btrfs_csum_data(p, crc, l);
1732                 kunmap_atomic(mapped_buffer);
1733                 len -= l;
1734                 if (len == 0)
1735                         break;
1736                 index++;
1737                 BUG_ON(index >= sblock->page_count);
1738                 BUG_ON(!sblock->pagev[index]->page);
1739                 page = sblock->pagev[index]->page;
1740                 mapped_buffer = kmap_atomic(page);
1741                 mapped_size = PAGE_SIZE;
1742                 p = mapped_buffer;
1743         }
1744
1745         btrfs_csum_final(crc, calculated_csum);
1746         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1747                 ++crc_fail;
1748
1749         return fail || crc_fail;
1750 }
1751
1752 static int scrub_checksum_super(struct scrub_block *sblock)
1753 {
1754         struct btrfs_super_block *s;
1755         struct scrub_ctx *sctx = sblock->sctx;
1756         struct btrfs_root *root = sctx->dev_root;
1757         struct btrfs_fs_info *fs_info = root->fs_info;
1758         u8 calculated_csum[BTRFS_CSUM_SIZE];
1759         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1760         struct page *page;
1761         void *mapped_buffer;
1762         u64 mapped_size;
1763         void *p;
1764         u32 crc = ~(u32)0;
1765         int fail_gen = 0;
1766         int fail_cor = 0;
1767         u64 len;
1768         int index;
1769
1770         BUG_ON(sblock->page_count < 1);
1771         page = sblock->pagev[0]->page;
1772         mapped_buffer = kmap_atomic(page);
1773         s = (struct btrfs_super_block *)mapped_buffer;
1774         memcpy(on_disk_csum, s->csum, sctx->csum_size);
1775
1776         if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1777                 ++fail_cor;
1778
1779         if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1780                 ++fail_gen;
1781
1782         if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1783                 ++fail_cor;
1784
1785         len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1786         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1787         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1788         index = 0;
1789         for (;;) {
1790                 u64 l = min_t(u64, len, mapped_size);
1791
1792                 crc = btrfs_csum_data(p, crc, l);
1793                 kunmap_atomic(mapped_buffer);
1794                 len -= l;
1795                 if (len == 0)
1796                         break;
1797                 index++;
1798                 BUG_ON(index >= sblock->page_count);
1799                 BUG_ON(!sblock->pagev[index]->page);
1800                 page = sblock->pagev[index]->page;
1801                 mapped_buffer = kmap_atomic(page);
1802                 mapped_size = PAGE_SIZE;
1803                 p = mapped_buffer;
1804         }
1805
1806         btrfs_csum_final(crc, calculated_csum);
1807         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1808                 ++fail_cor;
1809
1810         if (fail_cor + fail_gen) {
1811                 /*
1812                  * if we find an error in a super block, we just report it.
1813                  * They will get written with the next transaction commit
1814                  * anyway
1815                  */
1816                 spin_lock(&sctx->stat_lock);
1817                 ++sctx->stat.super_errors;
1818                 spin_unlock(&sctx->stat_lock);
1819                 if (fail_cor)
1820                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1821                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1822                 else
1823                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1824                                 BTRFS_DEV_STAT_GENERATION_ERRS);
1825         }
1826
1827         return fail_cor + fail_gen;
1828 }
1829
1830 static void scrub_block_get(struct scrub_block *sblock)
1831 {
1832         atomic_inc(&sblock->ref_count);
1833 }
1834
1835 static void scrub_block_put(struct scrub_block *sblock)
1836 {
1837         if (atomic_dec_and_test(&sblock->ref_count)) {
1838                 int i;
1839
1840                 for (i = 0; i < sblock->page_count; i++)
1841                         scrub_page_put(sblock->pagev[i]);
1842                 kfree(sblock);
1843         }
1844 }
1845
1846 static void scrub_page_get(struct scrub_page *spage)
1847 {
1848         atomic_inc(&spage->ref_count);
1849 }
1850
1851 static void scrub_page_put(struct scrub_page *spage)
1852 {
1853         if (atomic_dec_and_test(&spage->ref_count)) {
1854                 if (spage->page)
1855                         __free_page(spage->page);
1856                 kfree(spage);
1857         }
1858 }
1859
1860 static void scrub_submit(struct scrub_ctx *sctx)
1861 {
1862         struct scrub_bio *sbio;
1863
1864         if (sctx->curr == -1)
1865                 return;
1866
1867         sbio = sctx->bios[sctx->curr];
1868         sctx->curr = -1;
1869         scrub_pending_bio_inc(sctx);
1870
1871         if (!sbio->bio->bi_bdev) {
1872                 /*
1873                  * this case should not happen. If btrfs_map_block() is
1874                  * wrong, it could happen for dev-replace operations on
1875                  * missing devices when no mirrors are available, but in
1876                  * this case it should already fail the mount.
1877                  * This case is handled correctly (but _very_ slowly).
1878                  */
1879                 printk_ratelimited(KERN_WARNING
1880                         "btrfs: scrub_submit(bio bdev == NULL) is unexpected!\n");
1881                 bio_endio(sbio->bio, -EIO);
1882         } else {
1883                 btrfsic_submit_bio(READ, sbio->bio);
1884         }
1885 }
1886
1887 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
1888                                     struct scrub_page *spage)
1889 {
1890         struct scrub_block *sblock = spage->sblock;
1891         struct scrub_bio *sbio;
1892         int ret;
1893
1894 again:
1895         /*
1896          * grab a fresh bio or wait for one to become available
1897          */
1898         while (sctx->curr == -1) {
1899                 spin_lock(&sctx->list_lock);
1900                 sctx->curr = sctx->first_free;
1901                 if (sctx->curr != -1) {
1902                         sctx->first_free = sctx->bios[sctx->curr]->next_free;
1903                         sctx->bios[sctx->curr]->next_free = -1;
1904                         sctx->bios[sctx->curr]->page_count = 0;
1905                         spin_unlock(&sctx->list_lock);
1906                 } else {
1907                         spin_unlock(&sctx->list_lock);
1908                         wait_event(sctx->list_wait, sctx->first_free != -1);
1909                 }
1910         }
1911         sbio = sctx->bios[sctx->curr];
1912         if (sbio->page_count == 0) {
1913                 struct bio *bio;
1914
1915                 sbio->physical = spage->physical;
1916                 sbio->logical = spage->logical;
1917                 sbio->dev = spage->dev;
1918                 bio = sbio->bio;
1919                 if (!bio) {
1920                         bio = btrfs_io_bio_alloc(GFP_NOFS, sctx->pages_per_rd_bio);
1921                         if (!bio)
1922                                 return -ENOMEM;
1923                         sbio->bio = bio;
1924                 }
1925
1926                 bio->bi_private = sbio;
1927                 bio->bi_end_io = scrub_bio_end_io;
1928                 bio->bi_bdev = sbio->dev->bdev;
1929                 bio->bi_sector = sbio->physical >> 9;
1930                 sbio->err = 0;
1931         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1932                    spage->physical ||
1933                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1934                    spage->logical ||
1935                    sbio->dev != spage->dev) {
1936                 scrub_submit(sctx);
1937                 goto again;
1938         }
1939
1940         sbio->pagev[sbio->page_count] = spage;
1941         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1942         if (ret != PAGE_SIZE) {
1943                 if (sbio->page_count < 1) {
1944                         bio_put(sbio->bio);
1945                         sbio->bio = NULL;
1946                         return -EIO;
1947                 }
1948                 scrub_submit(sctx);
1949                 goto again;
1950         }
1951
1952         scrub_block_get(sblock); /* one for the page added to the bio */
1953         atomic_inc(&sblock->outstanding_pages);
1954         sbio->page_count++;
1955         if (sbio->page_count == sctx->pages_per_rd_bio)
1956                 scrub_submit(sctx);
1957
1958         return 0;
1959 }
1960
1961 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
1962                        u64 physical, struct btrfs_device *dev, u64 flags,
1963                        u64 gen, int mirror_num, u8 *csum, int force,
1964                        u64 physical_for_dev_replace)
1965 {
1966         struct scrub_block *sblock;
1967         int index;
1968
1969         sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
1970         if (!sblock) {
1971                 spin_lock(&sctx->stat_lock);
1972                 sctx->stat.malloc_errors++;
1973                 spin_unlock(&sctx->stat_lock);
1974                 return -ENOMEM;
1975         }
1976
1977         /* one ref inside this function, plus one for each page added to
1978          * a bio later on */
1979         atomic_set(&sblock->ref_count, 1);
1980         sblock->sctx = sctx;
1981         sblock->no_io_error_seen = 1;
1982
1983         for (index = 0; len > 0; index++) {
1984                 struct scrub_page *spage;
1985                 u64 l = min_t(u64, len, PAGE_SIZE);
1986
1987                 spage = kzalloc(sizeof(*spage), GFP_NOFS);
1988                 if (!spage) {
1989 leave_nomem:
1990                         spin_lock(&sctx->stat_lock);
1991                         sctx->stat.malloc_errors++;
1992                         spin_unlock(&sctx->stat_lock);
1993                         scrub_block_put(sblock);
1994                         return -ENOMEM;
1995                 }
1996                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
1997                 scrub_page_get(spage);
1998                 sblock->pagev[index] = spage;
1999                 spage->sblock = sblock;
2000                 spage->dev = dev;
2001                 spage->flags = flags;
2002                 spage->generation = gen;
2003                 spage->logical = logical;
2004                 spage->physical = physical;
2005                 spage->physical_for_dev_replace = physical_for_dev_replace;
2006                 spage->mirror_num = mirror_num;
2007                 if (csum) {
2008                         spage->have_csum = 1;
2009                         memcpy(spage->csum, csum, sctx->csum_size);
2010                 } else {
2011                         spage->have_csum = 0;
2012                 }
2013                 sblock->page_count++;
2014                 spage->page = alloc_page(GFP_NOFS);
2015                 if (!spage->page)
2016                         goto leave_nomem;
2017                 len -= l;
2018                 logical += l;
2019                 physical += l;
2020                 physical_for_dev_replace += l;
2021         }
2022
2023         WARN_ON(sblock->page_count == 0);
2024         for (index = 0; index < sblock->page_count; index++) {
2025                 struct scrub_page *spage = sblock->pagev[index];
2026                 int ret;
2027
2028                 ret = scrub_add_page_to_rd_bio(sctx, spage);
2029                 if (ret) {
2030                         scrub_block_put(sblock);
2031                         return ret;
2032                 }
2033         }
2034
2035         if (force)
2036                 scrub_submit(sctx);
2037
2038         /* last one frees, either here or in bio completion for last page */
2039         scrub_block_put(sblock);
2040         return 0;
2041 }
2042
2043 static void scrub_bio_end_io(struct bio *bio, int err)
2044 {
2045         struct scrub_bio *sbio = bio->bi_private;
2046         struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2047
2048         sbio->err = err;
2049         sbio->bio = bio;
2050
2051         btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
2052 }
2053
2054 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2055 {
2056         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2057         struct scrub_ctx *sctx = sbio->sctx;
2058         int i;
2059
2060         BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2061         if (sbio->err) {
2062                 for (i = 0; i < sbio->page_count; i++) {
2063                         struct scrub_page *spage = sbio->pagev[i];
2064
2065                         spage->io_error = 1;
2066                         spage->sblock->no_io_error_seen = 0;
2067                 }
2068         }
2069
2070         /* now complete the scrub_block items that have all pages completed */
2071         for (i = 0; i < sbio->page_count; i++) {
2072                 struct scrub_page *spage = sbio->pagev[i];
2073                 struct scrub_block *sblock = spage->sblock;
2074
2075                 if (atomic_dec_and_test(&sblock->outstanding_pages))
2076                         scrub_block_complete(sblock);
2077                 scrub_block_put(sblock);
2078         }
2079
2080         bio_put(sbio->bio);
2081         sbio->bio = NULL;
2082         spin_lock(&sctx->list_lock);
2083         sbio->next_free = sctx->first_free;
2084         sctx->first_free = sbio->index;
2085         spin_unlock(&sctx->list_lock);
2086
2087         if (sctx->is_dev_replace &&
2088             atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2089                 mutex_lock(&sctx->wr_ctx.wr_lock);
2090                 scrub_wr_submit(sctx);
2091                 mutex_unlock(&sctx->wr_ctx.wr_lock);
2092         }
2093
2094         scrub_pending_bio_dec(sctx);
2095 }
2096
2097 static void scrub_block_complete(struct scrub_block *sblock)
2098 {
2099         if (!sblock->no_io_error_seen) {
2100                 scrub_handle_errored_block(sblock);
2101         } else {
2102                 /*
2103                  * if has checksum error, write via repair mechanism in
2104                  * dev replace case, otherwise write here in dev replace
2105                  * case.
2106                  */
2107                 if (!scrub_checksum(sblock) && sblock->sctx->is_dev_replace)
2108                         scrub_write_block_to_dev_replace(sblock);
2109         }
2110 }
2111
2112 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u64 len,
2113                            u8 *csum)
2114 {
2115         struct btrfs_ordered_sum *sum = NULL;
2116         unsigned long index;
2117         unsigned long num_sectors;
2118
2119         while (!list_empty(&sctx->csum_list)) {
2120                 sum = list_first_entry(&sctx->csum_list,
2121                                        struct btrfs_ordered_sum, list);
2122                 if (sum->bytenr > logical)
2123                         return 0;
2124                 if (sum->bytenr + sum->len > logical)
2125                         break;
2126
2127                 ++sctx->stat.csum_discards;
2128                 list_del(&sum->list);
2129                 kfree(sum);
2130                 sum = NULL;
2131         }
2132         if (!sum)
2133                 return 0;
2134
2135         index = ((u32)(logical - sum->bytenr)) / sctx->sectorsize;
2136         num_sectors = sum->len / sctx->sectorsize;
2137         memcpy(csum, sum->sums + index, sctx->csum_size);
2138         if (index == num_sectors - 1) {
2139                 list_del(&sum->list);
2140                 kfree(sum);
2141         }
2142         return 1;
2143 }
2144
2145 /* scrub extent tries to collect up to 64 kB for each bio */
2146 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2147                         u64 physical, struct btrfs_device *dev, u64 flags,
2148                         u64 gen, int mirror_num, u64 physical_for_dev_replace)
2149 {
2150         int ret;
2151         u8 csum[BTRFS_CSUM_SIZE];
2152         u32 blocksize;
2153
2154         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2155                 blocksize = sctx->sectorsize;
2156                 spin_lock(&sctx->stat_lock);
2157                 sctx->stat.data_extents_scrubbed++;
2158                 sctx->stat.data_bytes_scrubbed += len;
2159                 spin_unlock(&sctx->stat_lock);
2160         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2161                 WARN_ON(sctx->nodesize != sctx->leafsize);
2162                 blocksize = sctx->nodesize;
2163                 spin_lock(&sctx->stat_lock);
2164                 sctx->stat.tree_extents_scrubbed++;
2165                 sctx->stat.tree_bytes_scrubbed += len;
2166                 spin_unlock(&sctx->stat_lock);
2167         } else {
2168                 blocksize = sctx->sectorsize;
2169                 WARN_ON(1);
2170         }
2171
2172         while (len) {
2173                 u64 l = min_t(u64, len, blocksize);
2174                 int have_csum = 0;
2175
2176                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2177                         /* push csums to sbio */
2178                         have_csum = scrub_find_csum(sctx, logical, l, csum);
2179                         if (have_csum == 0)
2180                                 ++sctx->stat.no_csum;
2181                         if (sctx->is_dev_replace && !have_csum) {
2182                                 ret = copy_nocow_pages(sctx, logical, l,
2183                                                        mirror_num,
2184                                                       physical_for_dev_replace);
2185                                 goto behind_scrub_pages;
2186                         }
2187                 }
2188                 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2189                                   mirror_num, have_csum ? csum : NULL, 0,
2190                                   physical_for_dev_replace);
2191 behind_scrub_pages:
2192                 if (ret)
2193                         return ret;
2194                 len -= l;
2195                 logical += l;
2196                 physical += l;
2197                 physical_for_dev_replace += l;
2198         }
2199         return 0;
2200 }
2201
2202 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2203                                            struct map_lookup *map,
2204                                            struct btrfs_device *scrub_dev,
2205                                            int num, u64 base, u64 length,
2206                                            int is_dev_replace)
2207 {
2208         struct btrfs_path *path;
2209         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2210         struct btrfs_root *root = fs_info->extent_root;
2211         struct btrfs_root *csum_root = fs_info->csum_root;
2212         struct btrfs_extent_item *extent;
2213         struct blk_plug plug;
2214         u64 flags;
2215         int ret;
2216         int slot;
2217         u64 nstripes;
2218         struct extent_buffer *l;
2219         struct btrfs_key key;
2220         u64 physical;
2221         u64 logical;
2222         u64 logic_end;
2223         u64 generation;
2224         int mirror_num;
2225         struct reada_control *reada1;
2226         struct reada_control *reada2;
2227         struct btrfs_key key_start;
2228         struct btrfs_key key_end;
2229         u64 increment = map->stripe_len;
2230         u64 offset;
2231         u64 extent_logical;
2232         u64 extent_physical;
2233         u64 extent_len;
2234         struct btrfs_device *extent_dev;
2235         int extent_mirror_num;
2236         int stop_loop;
2237
2238         if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2239                          BTRFS_BLOCK_GROUP_RAID6)) {
2240                 if (num >= nr_data_stripes(map)) {
2241                         return 0;
2242                 }
2243         }
2244
2245         nstripes = length;
2246         offset = 0;
2247         do_div(nstripes, map->stripe_len);
2248         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
2249                 offset = map->stripe_len * num;
2250                 increment = map->stripe_len * map->num_stripes;
2251                 mirror_num = 1;
2252         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
2253                 int factor = map->num_stripes / map->sub_stripes;
2254                 offset = map->stripe_len * (num / map->sub_stripes);
2255                 increment = map->stripe_len * factor;
2256                 mirror_num = num % map->sub_stripes + 1;
2257         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
2258                 increment = map->stripe_len;
2259                 mirror_num = num % map->num_stripes + 1;
2260         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
2261                 increment = map->stripe_len;
2262                 mirror_num = num % map->num_stripes + 1;
2263         } else {
2264                 increment = map->stripe_len;
2265                 mirror_num = 1;
2266         }
2267
2268         path = btrfs_alloc_path();
2269         if (!path)
2270                 return -ENOMEM;
2271
2272         /*
2273          * work on commit root. The related disk blocks are static as
2274          * long as COW is applied. This means, it is save to rewrite
2275          * them to repair disk errors without any race conditions
2276          */
2277         path->search_commit_root = 1;
2278         path->skip_locking = 1;
2279
2280         /*
2281          * trigger the readahead for extent tree csum tree and wait for
2282          * completion. During readahead, the scrub is officially paused
2283          * to not hold off transaction commits
2284          */
2285         logical = base + offset;
2286
2287         wait_event(sctx->list_wait,
2288                    atomic_read(&sctx->bios_in_flight) == 0);
2289         atomic_inc(&fs_info->scrubs_paused);
2290         wake_up(&fs_info->scrub_pause_wait);
2291
2292         /* FIXME it might be better to start readahead at commit root */
2293         key_start.objectid = logical;
2294         key_start.type = BTRFS_EXTENT_ITEM_KEY;
2295         key_start.offset = (u64)0;
2296         key_end.objectid = base + offset + nstripes * increment;
2297         key_end.type = BTRFS_METADATA_ITEM_KEY;
2298         key_end.offset = (u64)-1;
2299         reada1 = btrfs_reada_add(root, &key_start, &key_end);
2300
2301         key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2302         key_start.type = BTRFS_EXTENT_CSUM_KEY;
2303         key_start.offset = logical;
2304         key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2305         key_end.type = BTRFS_EXTENT_CSUM_KEY;
2306         key_end.offset = base + offset + nstripes * increment;
2307         reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
2308
2309         if (!IS_ERR(reada1))
2310                 btrfs_reada_wait(reada1);
2311         if (!IS_ERR(reada2))
2312                 btrfs_reada_wait(reada2);
2313
2314         mutex_lock(&fs_info->scrub_lock);
2315         while (atomic_read(&fs_info->scrub_pause_req)) {
2316                 mutex_unlock(&fs_info->scrub_lock);
2317                 wait_event(fs_info->scrub_pause_wait,
2318                    atomic_read(&fs_info->scrub_pause_req) == 0);
2319                 mutex_lock(&fs_info->scrub_lock);
2320         }
2321         atomic_dec(&fs_info->scrubs_paused);
2322         mutex_unlock(&fs_info->scrub_lock);
2323         wake_up(&fs_info->scrub_pause_wait);
2324
2325         /*
2326          * collect all data csums for the stripe to avoid seeking during
2327          * the scrub. This might currently (crc32) end up to be about 1MB
2328          */
2329         blk_start_plug(&plug);
2330
2331         /*
2332          * now find all extents for each stripe and scrub them
2333          */
2334         logical = base + offset;
2335         physical = map->stripes[num].physical;
2336         logic_end = logical + increment * nstripes;
2337         ret = 0;
2338         while (logical < logic_end) {
2339                 /*
2340                  * canceled?
2341                  */
2342                 if (atomic_read(&fs_info->scrub_cancel_req) ||
2343                     atomic_read(&sctx->cancel_req)) {
2344                         ret = -ECANCELED;
2345                         goto out;
2346                 }
2347                 /*
2348                  * check to see if we have to pause
2349                  */
2350                 if (atomic_read(&fs_info->scrub_pause_req)) {
2351                         /* push queued extents */
2352                         atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2353                         scrub_submit(sctx);
2354                         mutex_lock(&sctx->wr_ctx.wr_lock);
2355                         scrub_wr_submit(sctx);
2356                         mutex_unlock(&sctx->wr_ctx.wr_lock);
2357                         wait_event(sctx->list_wait,
2358                                    atomic_read(&sctx->bios_in_flight) == 0);
2359                         atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2360                         atomic_inc(&fs_info->scrubs_paused);
2361                         wake_up(&fs_info->scrub_pause_wait);
2362                         mutex_lock(&fs_info->scrub_lock);
2363                         while (atomic_read(&fs_info->scrub_pause_req)) {
2364                                 mutex_unlock(&fs_info->scrub_lock);
2365                                 wait_event(fs_info->scrub_pause_wait,
2366                                    atomic_read(&fs_info->scrub_pause_req) == 0);
2367                                 mutex_lock(&fs_info->scrub_lock);
2368                         }
2369                         atomic_dec(&fs_info->scrubs_paused);
2370                         mutex_unlock(&fs_info->scrub_lock);
2371                         wake_up(&fs_info->scrub_pause_wait);
2372                 }
2373
2374                 key.objectid = logical;
2375                 key.type = BTRFS_EXTENT_ITEM_KEY;
2376                 key.offset = (u64)-1;
2377
2378                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2379                 if (ret < 0)
2380                         goto out;
2381
2382                 if (ret > 0) {
2383                         ret = btrfs_previous_item(root, path, 0,
2384                                                   BTRFS_EXTENT_ITEM_KEY);
2385                         if (ret < 0)
2386                                 goto out;
2387                         if (ret > 0) {
2388                                 /* there's no smaller item, so stick with the
2389                                  * larger one */
2390                                 btrfs_release_path(path);
2391                                 ret = btrfs_search_slot(NULL, root, &key,
2392                                                         path, 0, 0);
2393                                 if (ret < 0)
2394                                         goto out;
2395                         }
2396                 }
2397
2398                 stop_loop = 0;
2399                 while (1) {
2400                         u64 bytes;
2401
2402                         l = path->nodes[0];
2403                         slot = path->slots[0];
2404                         if (slot >= btrfs_header_nritems(l)) {
2405                                 ret = btrfs_next_leaf(root, path);
2406                                 if (ret == 0)
2407                                         continue;
2408                                 if (ret < 0)
2409                                         goto out;
2410
2411                                 stop_loop = 1;
2412                                 break;
2413                         }
2414                         btrfs_item_key_to_cpu(l, &key, slot);
2415
2416                         if (key.type == BTRFS_METADATA_ITEM_KEY)
2417                                 bytes = root->leafsize;
2418                         else
2419                                 bytes = key.offset;
2420
2421                         if (key.objectid + bytes <= logical)
2422                                 goto next;
2423
2424                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2425                             key.type != BTRFS_METADATA_ITEM_KEY)
2426                                 goto next;
2427
2428                         if (key.objectid >= logical + map->stripe_len) {
2429                                 /* out of this device extent */
2430                                 if (key.objectid >= logic_end)
2431                                         stop_loop = 1;
2432                                 break;
2433                         }
2434
2435                         extent = btrfs_item_ptr(l, slot,
2436                                                 struct btrfs_extent_item);
2437                         flags = btrfs_extent_flags(l, extent);
2438                         generation = btrfs_extent_generation(l, extent);
2439
2440                         if (key.objectid < logical &&
2441                             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
2442                                 printk(KERN_ERR
2443                                        "btrfs scrub: tree block %llu spanning "
2444                                        "stripes, ignored. logical=%llu\n",
2445                                        key.objectid, logical);
2446                                 goto next;
2447                         }
2448
2449 again:
2450                         extent_logical = key.objectid;
2451                         extent_len = bytes;
2452
2453                         /*
2454                          * trim extent to this stripe
2455                          */
2456                         if (extent_logical < logical) {
2457                                 extent_len -= logical - extent_logical;
2458                                 extent_logical = logical;
2459                         }
2460                         if (extent_logical + extent_len >
2461                             logical + map->stripe_len) {
2462                                 extent_len = logical + map->stripe_len -
2463                                              extent_logical;
2464                         }
2465
2466                         extent_physical = extent_logical - logical + physical;
2467                         extent_dev = scrub_dev;
2468                         extent_mirror_num = mirror_num;
2469                         if (is_dev_replace)
2470                                 scrub_remap_extent(fs_info, extent_logical,
2471                                                    extent_len, &extent_physical,
2472                                                    &extent_dev,
2473                                                    &extent_mirror_num);
2474
2475                         ret = btrfs_lookup_csums_range(csum_root, logical,
2476                                                 logical + map->stripe_len - 1,
2477                                                 &sctx->csum_list, 1);
2478                         if (ret)
2479                                 goto out;
2480
2481                         ret = scrub_extent(sctx, extent_logical, extent_len,
2482                                            extent_physical, extent_dev, flags,
2483                                            generation, extent_mirror_num,
2484                                            extent_logical - logical + physical);
2485                         if (ret)
2486                                 goto out;
2487
2488                         scrub_free_csums(sctx);
2489                         if (extent_logical + extent_len <
2490                             key.objectid + bytes) {
2491                                 logical += increment;
2492                                 physical += map->stripe_len;
2493
2494                                 if (logical < key.objectid + bytes) {
2495                                         cond_resched();
2496                                         goto again;
2497                                 }
2498
2499                                 if (logical >= logic_end) {
2500                                         stop_loop = 1;
2501                                         break;
2502                                 }
2503                         }
2504 next:
2505                         path->slots[0]++;
2506                 }
2507                 btrfs_release_path(path);
2508                 logical += increment;
2509                 physical += map->stripe_len;
2510                 spin_lock(&sctx->stat_lock);
2511                 if (stop_loop)
2512                         sctx->stat.last_physical = map->stripes[num].physical +
2513                                                    length;
2514                 else
2515                         sctx->stat.last_physical = physical;
2516                 spin_unlock(&sctx->stat_lock);
2517                 if (stop_loop)
2518                         break;
2519         }
2520 out:
2521         /* push queued extents */
2522         scrub_submit(sctx);
2523         mutex_lock(&sctx->wr_ctx.wr_lock);
2524         scrub_wr_submit(sctx);
2525         mutex_unlock(&sctx->wr_ctx.wr_lock);
2526
2527         blk_finish_plug(&plug);
2528         btrfs_free_path(path);
2529         return ret < 0 ? ret : 0;
2530 }
2531
2532 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2533                                           struct btrfs_device *scrub_dev,
2534                                           u64 chunk_tree, u64 chunk_objectid,
2535                                           u64 chunk_offset, u64 length,
2536                                           u64 dev_offset, int is_dev_replace)
2537 {
2538         struct btrfs_mapping_tree *map_tree =
2539                 &sctx->dev_root->fs_info->mapping_tree;
2540         struct map_lookup *map;
2541         struct extent_map *em;
2542         int i;
2543         int ret = 0;
2544
2545         read_lock(&map_tree->map_tree.lock);
2546         em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2547         read_unlock(&map_tree->map_tree.lock);
2548
2549         if (!em)
2550                 return -EINVAL;
2551
2552         map = (struct map_lookup *)em->bdev;
2553         if (em->start != chunk_offset)
2554                 goto out;
2555
2556         if (em->len < length)
2557                 goto out;
2558
2559         for (i = 0; i < map->num_stripes; ++i) {
2560                 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2561                     map->stripes[i].physical == dev_offset) {
2562                         ret = scrub_stripe(sctx, map, scrub_dev, i,
2563                                            chunk_offset, length,
2564                                            is_dev_replace);
2565                         if (ret)
2566                                 goto out;
2567                 }
2568         }
2569 out:
2570         free_extent_map(em);
2571
2572         return ret;
2573 }
2574
2575 static noinline_for_stack
2576 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2577                            struct btrfs_device *scrub_dev, u64 start, u64 end,
2578                            int is_dev_replace)
2579 {
2580         struct btrfs_dev_extent *dev_extent = NULL;
2581         struct btrfs_path *path;
2582         struct btrfs_root *root = sctx->dev_root;
2583         struct btrfs_fs_info *fs_info = root->fs_info;
2584         u64 length;
2585         u64 chunk_tree;
2586         u64 chunk_objectid;
2587         u64 chunk_offset;
2588         int ret;
2589         int slot;
2590         struct extent_buffer *l;
2591         struct btrfs_key key;
2592         struct btrfs_key found_key;
2593         struct btrfs_block_group_cache *cache;
2594         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2595
2596         path = btrfs_alloc_path();
2597         if (!path)
2598                 return -ENOMEM;
2599
2600         path->reada = 2;
2601         path->search_commit_root = 1;
2602         path->skip_locking = 1;
2603
2604         key.objectid = scrub_dev->devid;
2605         key.offset = 0ull;
2606         key.type = BTRFS_DEV_EXTENT_KEY;
2607
2608         while (1) {
2609                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2610                 if (ret < 0)
2611                         break;
2612                 if (ret > 0) {
2613                         if (path->slots[0] >=
2614                             btrfs_header_nritems(path->nodes[0])) {
2615                                 ret = btrfs_next_leaf(root, path);
2616                                 if (ret)
2617                                         break;
2618                         }
2619                 }
2620
2621                 l = path->nodes[0];
2622                 slot = path->slots[0];
2623
2624                 btrfs_item_key_to_cpu(l, &found_key, slot);
2625
2626                 if (found_key.objectid != scrub_dev->devid)
2627                         break;
2628
2629                 if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
2630                         break;
2631
2632                 if (found_key.offset >= end)
2633                         break;
2634
2635                 if (found_key.offset < key.offset)
2636                         break;
2637
2638                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2639                 length = btrfs_dev_extent_length(l, dev_extent);
2640
2641                 if (found_key.offset + length <= start) {
2642                         key.offset = found_key.offset + length;
2643                         btrfs_release_path(path);
2644                         continue;
2645                 }
2646
2647                 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2648                 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2649                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2650
2651                 /*
2652                  * get a reference on the corresponding block group to prevent
2653                  * the chunk from going away while we scrub it
2654                  */
2655                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2656                 if (!cache) {
2657                         ret = -ENOENT;
2658                         break;
2659                 }
2660                 dev_replace->cursor_right = found_key.offset + length;
2661                 dev_replace->cursor_left = found_key.offset;
2662                 dev_replace->item_needs_writeback = 1;
2663                 ret = scrub_chunk(sctx, scrub_dev, chunk_tree, chunk_objectid,
2664                                   chunk_offset, length, found_key.offset,
2665                                   is_dev_replace);
2666
2667                 /*
2668                  * flush, submit all pending read and write bios, afterwards
2669                  * wait for them.
2670                  * Note that in the dev replace case, a read request causes
2671                  * write requests that are submitted in the read completion
2672                  * worker. Therefore in the current situation, it is required
2673                  * that all write requests are flushed, so that all read and
2674                  * write requests are really completed when bios_in_flight
2675                  * changes to 0.
2676                  */
2677                 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2678                 scrub_submit(sctx);
2679                 mutex_lock(&sctx->wr_ctx.wr_lock);
2680                 scrub_wr_submit(sctx);
2681                 mutex_unlock(&sctx->wr_ctx.wr_lock);
2682
2683                 wait_event(sctx->list_wait,
2684                            atomic_read(&sctx->bios_in_flight) == 0);
2685                 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2686                 atomic_inc(&fs_info->scrubs_paused);
2687                 wake_up(&fs_info->scrub_pause_wait);
2688                 wait_event(sctx->list_wait,
2689                            atomic_read(&sctx->workers_pending) == 0);
2690
2691                 mutex_lock(&fs_info->scrub_lock);
2692                 while (atomic_read(&fs_info->scrub_pause_req)) {
2693                         mutex_unlock(&fs_info->scrub_lock);
2694                         wait_event(fs_info->scrub_pause_wait,
2695                            atomic_read(&fs_info->scrub_pause_req) == 0);
2696                         mutex_lock(&fs_info->scrub_lock);
2697                 }
2698                 atomic_dec(&fs_info->scrubs_paused);
2699                 mutex_unlock(&fs_info->scrub_lock);
2700                 wake_up(&fs_info->scrub_pause_wait);
2701
2702                 btrfs_put_block_group(cache);
2703                 if (ret)
2704                         break;
2705                 if (is_dev_replace &&
2706                     atomic64_read(&dev_replace->num_write_errors) > 0) {
2707                         ret = -EIO;
2708                         break;
2709                 }
2710                 if (sctx->stat.malloc_errors > 0) {
2711                         ret = -ENOMEM;
2712                         break;
2713                 }
2714
2715                 dev_replace->cursor_left = dev_replace->cursor_right;
2716                 dev_replace->item_needs_writeback = 1;
2717
2718                 key.offset = found_key.offset + length;
2719                 btrfs_release_path(path);
2720         }
2721
2722         btrfs_free_path(path);
2723
2724         /*
2725          * ret can still be 1 from search_slot or next_leaf,
2726          * that's not an error
2727          */
2728         return ret < 0 ? ret : 0;
2729 }
2730
2731 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2732                                            struct btrfs_device *scrub_dev)
2733 {
2734         int     i;
2735         u64     bytenr;
2736         u64     gen;
2737         int     ret;
2738         struct btrfs_root *root = sctx->dev_root;
2739
2740         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
2741                 return -EIO;
2742
2743         gen = root->fs_info->last_trans_committed;
2744
2745         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2746                 bytenr = btrfs_sb_offset(i);
2747                 if (bytenr + BTRFS_SUPER_INFO_SIZE > scrub_dev->total_bytes)
2748                         break;
2749
2750                 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2751                                   scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
2752                                   NULL, 1, bytenr);
2753                 if (ret)
2754                         return ret;
2755         }
2756         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2757
2758         return 0;
2759 }
2760
2761 /*
2762  * get a reference count on fs_info->scrub_workers. start worker if necessary
2763  */
2764 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
2765                                                 int is_dev_replace)
2766 {
2767         int ret = 0;
2768
2769         if (fs_info->scrub_workers_refcnt == 0) {
2770                 if (is_dev_replace)
2771                         btrfs_init_workers(&fs_info->scrub_workers, "scrub", 1,
2772                                         &fs_info->generic_worker);
2773                 else
2774                         btrfs_init_workers(&fs_info->scrub_workers, "scrub",
2775                                         fs_info->thread_pool_size,
2776                                         &fs_info->generic_worker);
2777                 fs_info->scrub_workers.idle_thresh = 4;
2778                 ret = btrfs_start_workers(&fs_info->scrub_workers);
2779                 if (ret)
2780                         goto out;
2781                 btrfs_init_workers(&fs_info->scrub_wr_completion_workers,
2782                                    "scrubwrc",
2783                                    fs_info->thread_pool_size,
2784                                    &fs_info->generic_worker);
2785                 fs_info->scrub_wr_completion_workers.idle_thresh = 2;
2786                 ret = btrfs_start_workers(
2787                                 &fs_info->scrub_wr_completion_workers);
2788                 if (ret)
2789                         goto out;
2790                 btrfs_init_workers(&fs_info->scrub_nocow_workers, "scrubnc", 1,
2791                                    &fs_info->generic_worker);
2792                 ret = btrfs_start_workers(&fs_info->scrub_nocow_workers);
2793                 if (ret)
2794                         goto out;
2795         }
2796         ++fs_info->scrub_workers_refcnt;
2797 out:
2798         return ret;
2799 }
2800
2801 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
2802 {
2803         if (--fs_info->scrub_workers_refcnt == 0) {
2804                 btrfs_stop_workers(&fs_info->scrub_workers);
2805                 btrfs_stop_workers(&fs_info->scrub_wr_completion_workers);
2806                 btrfs_stop_workers(&fs_info->scrub_nocow_workers);
2807         }
2808         WARN_ON(fs_info->scrub_workers_refcnt < 0);
2809 }
2810
2811 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2812                     u64 end, struct btrfs_scrub_progress *progress,
2813                     int readonly, int is_dev_replace)
2814 {
2815         struct scrub_ctx *sctx;
2816         int ret;
2817         struct btrfs_device *dev;
2818
2819         if (btrfs_fs_closing(fs_info))
2820                 return -EINVAL;
2821
2822         /*
2823          * check some assumptions
2824          */
2825         if (fs_info->chunk_root->nodesize != fs_info->chunk_root->leafsize) {
2826                 printk(KERN_ERR
2827                        "btrfs_scrub: size assumption nodesize == leafsize (%d == %d) fails\n",
2828                        fs_info->chunk_root->nodesize,
2829                        fs_info->chunk_root->leafsize);
2830                 return -EINVAL;
2831         }
2832
2833         if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) {
2834                 /*
2835                  * in this case scrub is unable to calculate the checksum
2836                  * the way scrub is implemented. Do not handle this
2837                  * situation at all because it won't ever happen.
2838                  */
2839                 printk(KERN_ERR
2840                        "btrfs_scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails\n",
2841                        fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN);
2842                 return -EINVAL;
2843         }
2844
2845         if (fs_info->chunk_root->sectorsize != PAGE_SIZE) {
2846                 /* not supported for data w/o checksums */
2847                 printk(KERN_ERR
2848                        "btrfs_scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails\n",
2849                        fs_info->chunk_root->sectorsize, PAGE_SIZE);
2850                 return -EINVAL;
2851         }
2852
2853         if (fs_info->chunk_root->nodesize >
2854             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
2855             fs_info->chunk_root->sectorsize >
2856             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
2857                 /*
2858                  * would exhaust the array bounds of pagev member in
2859                  * struct scrub_block
2860                  */
2861                 pr_err("btrfs_scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails\n",
2862                        fs_info->chunk_root->nodesize,
2863                        SCRUB_MAX_PAGES_PER_BLOCK,
2864                        fs_info->chunk_root->sectorsize,
2865                        SCRUB_MAX_PAGES_PER_BLOCK);
2866                 return -EINVAL;
2867         }
2868
2869
2870         mutex_lock(&fs_info->fs_devices->device_list_mutex);
2871         dev = btrfs_find_device(fs_info, devid, NULL, NULL);
2872         if (!dev || (dev->missing && !is_dev_replace)) {
2873                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2874                 return -ENODEV;
2875         }
2876
2877         mutex_lock(&fs_info->scrub_lock);
2878         if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
2879                 mutex_unlock(&fs_info->scrub_lock);
2880                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2881                 return -EIO;
2882         }
2883
2884         btrfs_dev_replace_lock(&fs_info->dev_replace);
2885         if (dev->scrub_device ||
2886             (!is_dev_replace &&
2887              btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2888                 btrfs_dev_replace_unlock(&fs_info->dev_replace);
2889                 mutex_unlock(&fs_info->scrub_lock);
2890                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2891                 return -EINPROGRESS;
2892         }
2893         btrfs_dev_replace_unlock(&fs_info->dev_replace);
2894
2895         ret = scrub_workers_get(fs_info, is_dev_replace);
2896         if (ret) {
2897                 mutex_unlock(&fs_info->scrub_lock);
2898                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2899                 return ret;
2900         }
2901
2902         sctx = scrub_setup_ctx(dev, is_dev_replace);
2903         if (IS_ERR(sctx)) {
2904                 mutex_unlock(&fs_info->scrub_lock);
2905                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2906                 scrub_workers_put(fs_info);
2907                 return PTR_ERR(sctx);
2908         }
2909         sctx->readonly = readonly;
2910         dev->scrub_device = sctx;
2911
2912         atomic_inc(&fs_info->scrubs_running);
2913         mutex_unlock(&fs_info->scrub_lock);
2914
2915         if (!is_dev_replace) {
2916                 /*
2917                  * by holding device list mutex, we can
2918                  * kick off writing super in log tree sync.
2919                  */
2920                 ret = scrub_supers(sctx, dev);
2921         }
2922         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2923
2924         if (!ret)
2925                 ret = scrub_enumerate_chunks(sctx, dev, start, end,
2926                                              is_dev_replace);
2927
2928         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2929         atomic_dec(&fs_info->scrubs_running);
2930         wake_up(&fs_info->scrub_pause_wait);
2931
2932         wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
2933
2934         if (progress)
2935                 memcpy(progress, &sctx->stat, sizeof(*progress));
2936
2937         mutex_lock(&fs_info->scrub_lock);
2938         dev->scrub_device = NULL;
2939         scrub_workers_put(fs_info);
2940         mutex_unlock(&fs_info->scrub_lock);
2941
2942         scrub_free_ctx(sctx);
2943
2944         return ret;
2945 }
2946
2947 void btrfs_scrub_pause(struct btrfs_root *root)
2948 {
2949         struct btrfs_fs_info *fs_info = root->fs_info;
2950
2951         mutex_lock(&fs_info->scrub_lock);
2952         atomic_inc(&fs_info->scrub_pause_req);
2953         while (atomic_read(&fs_info->scrubs_paused) !=
2954                atomic_read(&fs_info->scrubs_running)) {
2955                 mutex_unlock(&fs_info->scrub_lock);
2956                 wait_event(fs_info->scrub_pause_wait,
2957                            atomic_read(&fs_info->scrubs_paused) ==
2958                            atomic_read(&fs_info->scrubs_running));
2959                 mutex_lock(&fs_info->scrub_lock);
2960         }
2961         mutex_unlock(&fs_info->scrub_lock);
2962 }
2963
2964 void btrfs_scrub_continue(struct btrfs_root *root)
2965 {
2966         struct btrfs_fs_info *fs_info = root->fs_info;
2967
2968         atomic_dec(&fs_info->scrub_pause_req);
2969         wake_up(&fs_info->scrub_pause_wait);
2970 }
2971
2972 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
2973 {
2974         mutex_lock(&fs_info->scrub_lock);
2975         if (!atomic_read(&fs_info->scrubs_running)) {
2976                 mutex_unlock(&fs_info->scrub_lock);
2977                 return -ENOTCONN;
2978         }
2979
2980         atomic_inc(&fs_info->scrub_cancel_req);
2981         while (atomic_read(&fs_info->scrubs_running)) {
2982                 mutex_unlock(&fs_info->scrub_lock);
2983                 wait_event(fs_info->scrub_pause_wait,
2984                            atomic_read(&fs_info->scrubs_running) == 0);
2985                 mutex_lock(&fs_info->scrub_lock);
2986         }
2987         atomic_dec(&fs_info->scrub_cancel_req);
2988         mutex_unlock(&fs_info->scrub_lock);
2989
2990         return 0;
2991 }
2992
2993 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
2994                            struct btrfs_device *dev)
2995 {
2996         struct scrub_ctx *sctx;
2997
2998         mutex_lock(&fs_info->scrub_lock);
2999         sctx = dev->scrub_device;
3000         if (!sctx) {
3001                 mutex_unlock(&fs_info->scrub_lock);
3002                 return -ENOTCONN;
3003         }
3004         atomic_inc(&sctx->cancel_req);
3005         while (dev->scrub_device) {
3006                 mutex_unlock(&fs_info->scrub_lock);
3007                 wait_event(fs_info->scrub_pause_wait,
3008                            dev->scrub_device == NULL);
3009                 mutex_lock(&fs_info->scrub_lock);
3010         }
3011         mutex_unlock(&fs_info->scrub_lock);
3012
3013         return 0;
3014 }
3015
3016 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
3017                          struct btrfs_scrub_progress *progress)
3018 {
3019         struct btrfs_device *dev;
3020         struct scrub_ctx *sctx = NULL;
3021
3022         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
3023         dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
3024         if (dev)
3025                 sctx = dev->scrub_device;
3026         if (sctx)
3027                 memcpy(progress, &sctx->stat, sizeof(*progress));
3028         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
3029
3030         return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3031 }
3032
3033 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
3034                                u64 extent_logical, u64 extent_len,
3035                                u64 *extent_physical,
3036                                struct btrfs_device **extent_dev,
3037                                int *extent_mirror_num)
3038 {
3039         u64 mapped_length;
3040         struct btrfs_bio *bbio = NULL;
3041         int ret;
3042
3043         mapped_length = extent_len;
3044         ret = btrfs_map_block(fs_info, READ, extent_logical,
3045                               &mapped_length, &bbio, 0);
3046         if (ret || !bbio || mapped_length < extent_len ||
3047             !bbio->stripes[0].dev->bdev) {
3048                 kfree(bbio);
3049                 return;
3050         }
3051
3052         *extent_physical = bbio->stripes[0].physical;
3053         *extent_mirror_num = bbio->mirror_num;
3054         *extent_dev = bbio->stripes[0].dev;
3055         kfree(bbio);
3056 }
3057
3058 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
3059                               struct scrub_wr_ctx *wr_ctx,
3060                               struct btrfs_fs_info *fs_info,
3061                               struct btrfs_device *dev,
3062                               int is_dev_replace)
3063 {
3064         WARN_ON(wr_ctx->wr_curr_bio != NULL);
3065
3066         mutex_init(&wr_ctx->wr_lock);
3067         wr_ctx->wr_curr_bio = NULL;
3068         if (!is_dev_replace)
3069                 return 0;
3070
3071         WARN_ON(!dev->bdev);
3072         wr_ctx->pages_per_wr_bio = min_t(int, SCRUB_PAGES_PER_WR_BIO,
3073                                          bio_get_nr_vecs(dev->bdev));
3074         wr_ctx->tgtdev = dev;
3075         atomic_set(&wr_ctx->flush_all_writes, 0);
3076         return 0;
3077 }
3078
3079 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
3080 {
3081         mutex_lock(&wr_ctx->wr_lock);
3082         kfree(wr_ctx->wr_curr_bio);
3083         wr_ctx->wr_curr_bio = NULL;
3084         mutex_unlock(&wr_ctx->wr_lock);
3085 }
3086
3087 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
3088                             int mirror_num, u64 physical_for_dev_replace)
3089 {
3090         struct scrub_copy_nocow_ctx *nocow_ctx;
3091         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3092
3093         nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
3094         if (!nocow_ctx) {
3095                 spin_lock(&sctx->stat_lock);
3096                 sctx->stat.malloc_errors++;
3097                 spin_unlock(&sctx->stat_lock);
3098                 return -ENOMEM;
3099         }
3100
3101         scrub_pending_trans_workers_inc(sctx);
3102
3103         nocow_ctx->sctx = sctx;
3104         nocow_ctx->logical = logical;
3105         nocow_ctx->len = len;
3106         nocow_ctx->mirror_num = mirror_num;
3107         nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
3108         nocow_ctx->work.func = copy_nocow_pages_worker;
3109         INIT_LIST_HEAD(&nocow_ctx->inodes);
3110         btrfs_queue_worker(&fs_info->scrub_nocow_workers,
3111                            &nocow_ctx->work);
3112
3113         return 0;
3114 }
3115
3116 static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
3117 {
3118         struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
3119         struct scrub_nocow_inode *nocow_inode;
3120
3121         nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
3122         if (!nocow_inode)
3123                 return -ENOMEM;
3124         nocow_inode->inum = inum;
3125         nocow_inode->offset = offset;
3126         nocow_inode->root = root;
3127         list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
3128         return 0;
3129 }
3130
3131 #define COPY_COMPLETE 1
3132
3133 static void copy_nocow_pages_worker(struct btrfs_work *work)
3134 {
3135         struct scrub_copy_nocow_ctx *nocow_ctx =
3136                 container_of(work, struct scrub_copy_nocow_ctx, work);
3137         struct scrub_ctx *sctx = nocow_ctx->sctx;
3138         u64 logical = nocow_ctx->logical;
3139         u64 len = nocow_ctx->len;
3140         int mirror_num = nocow_ctx->mirror_num;
3141         u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3142         int ret;
3143         struct btrfs_trans_handle *trans = NULL;
3144         struct btrfs_fs_info *fs_info;
3145         struct btrfs_path *path;
3146         struct btrfs_root *root;
3147         int not_written = 0;
3148
3149         fs_info = sctx->dev_root->fs_info;
3150         root = fs_info->extent_root;
3151
3152         path = btrfs_alloc_path();
3153         if (!path) {
3154                 spin_lock(&sctx->stat_lock);
3155                 sctx->stat.malloc_errors++;
3156                 spin_unlock(&sctx->stat_lock);
3157                 not_written = 1;
3158                 goto out;
3159         }
3160
3161         trans = btrfs_join_transaction(root);
3162         if (IS_ERR(trans)) {
3163                 not_written = 1;
3164                 goto out;
3165         }
3166
3167         ret = iterate_inodes_from_logical(logical, fs_info, path,
3168                                           record_inode_for_nocow, nocow_ctx);
3169         if (ret != 0 && ret != -ENOENT) {
3170                 pr_warn("iterate_inodes_from_logical() failed: log %llu, phys %llu, len %llu, mir %u, ret %d\n",
3171                         logical, physical_for_dev_replace, len, mirror_num,
3172                         ret);
3173                 not_written = 1;
3174                 goto out;
3175         }
3176
3177         btrfs_end_transaction(trans, root);
3178         trans = NULL;
3179         while (!list_empty(&nocow_ctx->inodes)) {
3180                 struct scrub_nocow_inode *entry;
3181                 entry = list_first_entry(&nocow_ctx->inodes,
3182                                          struct scrub_nocow_inode,
3183                                          list);
3184                 list_del_init(&entry->list);
3185                 ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
3186                                                  entry->root, nocow_ctx);
3187                 kfree(entry);
3188                 if (ret == COPY_COMPLETE) {
3189                         ret = 0;
3190                         break;
3191                 } else if (ret) {
3192                         break;
3193                 }
3194         }
3195 out:
3196         while (!list_empty(&nocow_ctx->inodes)) {
3197                 struct scrub_nocow_inode *entry;
3198                 entry = list_first_entry(&nocow_ctx->inodes,
3199                                          struct scrub_nocow_inode,
3200                                          list);
3201                 list_del_init(&entry->list);
3202                 kfree(entry);
3203         }
3204         if (trans && !IS_ERR(trans))
3205                 btrfs_end_transaction(trans, root);
3206         if (not_written)
3207                 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
3208                                             num_uncorrectable_read_errors);
3209
3210         btrfs_free_path(path);
3211         kfree(nocow_ctx);
3212
3213         scrub_pending_trans_workers_dec(sctx);
3214 }
3215
3216 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
3217                                       struct scrub_copy_nocow_ctx *nocow_ctx)
3218 {
3219         struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
3220         struct btrfs_key key;
3221         struct inode *inode;
3222         struct page *page;
3223         struct btrfs_root *local_root;
3224         struct btrfs_ordered_extent *ordered;
3225         struct extent_map *em;
3226         struct extent_state *cached_state = NULL;
3227         struct extent_io_tree *io_tree;
3228         u64 physical_for_dev_replace;
3229         u64 len = nocow_ctx->len;
3230         u64 lockstart = offset, lockend = offset + len - 1;
3231         unsigned long index;
3232         int srcu_index;
3233         int ret = 0;
3234         int err = 0;
3235
3236         key.objectid = root;
3237         key.type = BTRFS_ROOT_ITEM_KEY;
3238         key.offset = (u64)-1;
3239
3240         srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
3241
3242         local_root = btrfs_read_fs_root_no_name(fs_info, &key);
3243         if (IS_ERR(local_root)) {
3244                 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3245                 return PTR_ERR(local_root);
3246         }
3247
3248         key.type = BTRFS_INODE_ITEM_KEY;
3249         key.objectid = inum;
3250         key.offset = 0;
3251         inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
3252         srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3253         if (IS_ERR(inode))
3254                 return PTR_ERR(inode);
3255
3256         /* Avoid truncate/dio/punch hole.. */
3257         mutex_lock(&inode->i_mutex);
3258         inode_dio_wait(inode);
3259
3260         physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3261         io_tree = &BTRFS_I(inode)->io_tree;
3262
3263         lock_extent_bits(io_tree, lockstart, lockend, 0, &cached_state);
3264         ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
3265         if (ordered) {
3266                 btrfs_put_ordered_extent(ordered);
3267                 goto out_unlock;
3268         }
3269
3270         em = btrfs_get_extent(inode, NULL, 0, lockstart, len, 0);
3271         if (IS_ERR(em)) {
3272                 ret = PTR_ERR(em);
3273                 goto out_unlock;
3274         }
3275
3276         /*
3277          * This extent does not actually cover the logical extent anymore,
3278          * move on to the next inode.
3279          */
3280         if (em->block_start > nocow_ctx->logical ||
3281             em->block_start + em->block_len < nocow_ctx->logical + len) {
3282                 free_extent_map(em);
3283                 goto out_unlock;
3284         }
3285         free_extent_map(em);
3286
3287         while (len >= PAGE_CACHE_SIZE) {
3288                 index = offset >> PAGE_CACHE_SHIFT;
3289 again:
3290                 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
3291                 if (!page) {
3292                         pr_err("find_or_create_page() failed\n");
3293                         ret = -ENOMEM;
3294                         goto out;
3295                 }
3296
3297                 if (PageUptodate(page)) {
3298                         if (PageDirty(page))
3299                                 goto next_page;
3300                 } else {
3301                         ClearPageError(page);
3302                         err = extent_read_full_page_nolock(io_tree, page,
3303                                                            btrfs_get_extent,
3304                                                            nocow_ctx->mirror_num);
3305                         if (err) {
3306                                 ret = err;
3307                                 goto next_page;
3308                         }
3309
3310                         lock_page(page);
3311                         /*
3312                          * If the page has been remove from the page cache,
3313                          * the data on it is meaningless, because it may be
3314                          * old one, the new data may be written into the new
3315                          * page in the page cache.
3316                          */
3317                         if (page->mapping != inode->i_mapping) {
3318                                 unlock_page(page);
3319                                 page_cache_release(page);
3320                                 goto again;
3321                         }
3322                         if (!PageUptodate(page)) {
3323                                 ret = -EIO;
3324                                 goto next_page;
3325                         }
3326                 }
3327                 err = write_page_nocow(nocow_ctx->sctx,
3328                                        physical_for_dev_replace, page);
3329                 if (err)
3330                         ret = err;
3331 next_page:
3332                 unlock_page(page);
3333                 page_cache_release(page);
3334
3335                 if (ret)
3336                         break;
3337
3338                 offset += PAGE_CACHE_SIZE;
3339                 physical_for_dev_replace += PAGE_CACHE_SIZE;
3340                 len -= PAGE_CACHE_SIZE;
3341         }
3342         ret = COPY_COMPLETE;
3343 out_unlock:
3344         unlock_extent_cached(io_tree, lockstart, lockend, &cached_state,
3345                              GFP_NOFS);
3346 out:
3347         mutex_unlock(&inode->i_mutex);
3348         iput(inode);
3349         return ret;
3350 }
3351
3352 static int write_page_nocow(struct scrub_ctx *sctx,
3353                             u64 physical_for_dev_replace, struct page *page)
3354 {
3355         struct bio *bio;
3356         struct btrfs_device *dev;
3357         int ret;
3358
3359         dev = sctx->wr_ctx.tgtdev;
3360         if (!dev)
3361                 return -EIO;
3362         if (!dev->bdev) {
3363                 printk_ratelimited(KERN_WARNING
3364                         "btrfs: scrub write_page_nocow(bdev == NULL) is unexpected!\n");
3365                 return -EIO;
3366         }
3367         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
3368         if (!bio) {
3369                 spin_lock(&sctx->stat_lock);
3370                 sctx->stat.malloc_errors++;
3371                 spin_unlock(&sctx->stat_lock);
3372                 return -ENOMEM;
3373         }
3374         bio->bi_size = 0;
3375         bio->bi_sector = physical_for_dev_replace >> 9;
3376         bio->bi_bdev = dev->bdev;
3377         ret = bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
3378         if (ret != PAGE_CACHE_SIZE) {
3379 leave_with_eio:
3380                 bio_put(bio);
3381                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
3382                 return -EIO;
3383         }
3384
3385         if (btrfsic_submit_bio_wait(WRITE_SYNC, bio))
3386                 goto leave_with_eio;
3387
3388         bio_put(bio);
3389         return 0;
3390 }