fs/btrfs/scrub.c

   1 /*
   2  * Copyright (C) 2011 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 "check-integrity.h"
  29
  30 /*
  31  * This is only the first step towards a full-features scrub. It reads all
  32  * extent and super block and verifies the checksums. In case a bad checksum
  33  * is found or the extent cannot be read, good data will be written back if
  34  * any can be found.
  35  *
  36  * Future enhancements:
  37  *  - In case an unrepairable extent is encountered, track which files are
  38  *    affected and report them
  39  *  - track and record media errors, throw out bad devices
  40  *  - add a mode to also read unallocated space
  41  */
  42
  43 struct scrub_block;
  44 struct scrub_dev;
  45
  46 #define SCRUB_PAGES_PER_BIO     16      /* 64k per bio */
  47 #define SCRUB_BIOS_PER_DEV      16      /* 1 MB per device in flight */
  48 #define SCRUB_MAX_PAGES_PER_BLOCK       16      /* 64k per node/leaf/sector */
  49
  50 struct scrub_page {
  51         struct scrub_block      *sblock;
  52         struct page             *page;
  53         struct block_device     *bdev;
  54         u64                     flags;  /* extent flags */
  55         u64                     generation;
  56         u64                     logical;
  57         u64                     physical;
  58         struct {
  59                 unsigned int    mirror_num:8;
  60                 unsigned int    have_csum:1;
  61                 unsigned int    io_error:1;
  62         };
  63         u8                      csum[BTRFS_CSUM_SIZE];
  64 };
  65
  66 struct scrub_bio {
  67         int                     index;
  68         struct scrub_dev        *sdev;
  69         struct bio              *bio;
  70         int                     err;
  71         u64                     logical;
  72         u64                     physical;
  73         struct scrub_page       *pagev[SCRUB_PAGES_PER_BIO];
  74         int                     page_count;
  75         int                     next_free;
  76         struct btrfs_work       work;
  77 };
  78
  79 struct scrub_block {
  80         struct scrub_page       pagev[SCRUB_MAX_PAGES_PER_BLOCK];
  81         int                     page_count;
  82         atomic_t                outstanding_pages;
  83         atomic_t                ref_count; /* free mem on transition to zero */
  84         struct scrub_dev        *sdev;
  85         struct {
  86                 unsigned int    header_error:1;
  87                 unsigned int    checksum_error:1;
  88                 unsigned int    no_io_error_seen:1;
  89         };
  90 };
  91
  92 struct scrub_dev {
  93         struct scrub_bio        *bios[SCRUB_BIOS_PER_DEV];
  94         struct btrfs_device     *dev;
  95         int                     first_free;
  96         int                     curr;
  97         atomic_t                in_flight;
  98         atomic_t                fixup_cnt;
  99         spinlock_t              list_lock;
 100         wait_queue_head_t       list_wait;
 101         u16                     csum_size;
 102         struct list_head        csum_list;
 103         atomic_t                cancel_req;
 104         int                     readonly;
 105         int                     pages_per_bio; /* <= SCRUB_PAGES_PER_BIO */
 106         u32                     sectorsize;
 107         u32                     nodesize;
 108         u32                     leafsize;
 109         /*
 110          * statistics
 111          */
 112         struct btrfs_scrub_progress stat;
 113         spinlock_t              stat_lock;
 114 };
 115
 116 struct scrub_fixup_nodatasum {
 117         struct scrub_dev        *sdev;
 118         u64                     logical;
 119         struct btrfs_root       *root;
 120         struct btrfs_work       work;
 121         int                     mirror_num;
 122 };
 123
 124 struct scrub_warning {
 125         struct btrfs_path       *path;
 126         u64                     extent_item_size;
 127         char                    *scratch_buf;
 128         char                    *msg_buf;
 129         const char              *errstr;
 130         sector_t                sector;
 131         u64                     logical;
 132         struct btrfs_device     *dev;
 133         int                     msg_bufsize;
 134         int                     scratch_bufsize;
 135 };
 136
 137
 138 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
 139 static int scrub_setup_recheck_block(struct scrub_dev *sdev,
 140                                      struct btrfs_mapping_tree *map_tree,
 141                                      u64 length, u64 logical,
 142                                      struct scrub_block *sblock);
 143 static int scrub_recheck_block(struct btrfs_fs_info *fs_info,
 144                                struct scrub_block *sblock, int is_metadata,
 145                                int have_csum, u8 *csum, u64 generation,
 146                                u16 csum_size);
 147 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
 148                                          struct scrub_block *sblock,
 149                                          int is_metadata, int have_csum,
 150                                          const u8 *csum, u64 generation,
 151                                          u16 csum_size);
 152 static void scrub_complete_bio_end_io(struct bio *bio, int err);
 153 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
 154                                              struct scrub_block *sblock_good,
 155                                              int force_write);
 156 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
 157                                             struct scrub_block *sblock_good,
 158                                             int page_num, int force_write);
 159 static int scrub_checksum_data(struct scrub_block *sblock);
 160 static int scrub_checksum_tree_block(struct scrub_block *sblock);
 161 static int scrub_checksum_super(struct scrub_block *sblock);
 162 static void scrub_block_get(struct scrub_block *sblock);
 163 static void scrub_block_put(struct scrub_block *sblock);
 164 static int scrub_add_page_to_bio(struct scrub_dev *sdev,
 165                                  struct scrub_page *spage);
 166 static int scrub_pages(struct scrub_dev *sdev, u64 logical, u64 len,
 167                        u64 physical, u64 flags, u64 gen, int mirror_num,
 168                        u8 *csum, int force);
 169 static void scrub_bio_end_io(struct bio *bio, int err);
 170 static void scrub_bio_end_io_worker(struct btrfs_work *work);
 171 static void scrub_block_complete(struct scrub_block *sblock);
 172
 173
 174 static void scrub_free_csums(struct scrub_dev *sdev)
 175 {
 176         while (!list_empty(&sdev->csum_list)) {
 177                 struct btrfs_ordered_sum *sum;
 178                 sum = list_first_entry(&sdev->csum_list,
 179                                        struct btrfs_ordered_sum, list);
 180                 list_del(&sum->list);
 181                 kfree(sum);
 182         }
 183 }
 184
 185 static noinline_for_stack void scrub_free_dev(struct scrub_dev *sdev)
 186 {
 187         int i;
 188
 189         if (!sdev)
 190                 return;
 191
 192         /* this can happen when scrub is cancelled */
 193         if (sdev->curr != -1) {
 194                 struct scrub_bio *sbio = sdev->bios[sdev->curr];
 195
 196                 for (i = 0; i < sbio->page_count; i++) {
 197                         BUG_ON(!sbio->pagev[i]);
 198                         BUG_ON(!sbio->pagev[i]->page);
 199                         scrub_block_put(sbio->pagev[i]->sblock);
 200                 }
 201                 bio_put(sbio->bio);
 202         }
 203
 204         for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
 205                 struct scrub_bio *sbio = sdev->bios[i];
 206
 207                 if (!sbio)
 208                         break;
 209                 kfree(sbio);
 210         }
 211
 212         scrub_free_csums(sdev);
 213         kfree(sdev);
 214 }
 215
 216 static noinline_for_stack
 217 struct scrub_dev *scrub_setup_dev(struct btrfs_device *dev)
 218 {
 219         struct scrub_dev *sdev;
 220         int             i;
 221         struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
 222         int pages_per_bio;
 223
 224         pages_per_bio = min_t(int, SCRUB_PAGES_PER_BIO,
 225                               bio_get_nr_vecs(dev->bdev));
 226         sdev = kzalloc(sizeof(*sdev), GFP_NOFS);
 227         if (!sdev)
 228                 goto nomem;
 229         sdev->dev = dev;
 230         sdev->pages_per_bio = pages_per_bio;
 231         sdev->curr = -1;
 232         for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
 233                 struct scrub_bio *sbio;
 234
 235                 sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
 236                 if (!sbio)
 237                         goto nomem;
 238                 sdev->bios[i] = sbio;
 239
 240                 sbio->index = i;
 241                 sbio->sdev = sdev;
 242                 sbio->page_count = 0;
 243                 sbio->work.func = scrub_bio_end_io_worker;
 244
 245                 if (i != SCRUB_BIOS_PER_DEV-1)
 246                         sdev->bios[i]->next_free = i + 1;
 247                 else
 248                         sdev->bios[i]->next_free = -1;
 249         }
 250         sdev->first_free = 0;
 251         sdev->nodesize = dev->dev_root->nodesize;
 252         sdev->leafsize = dev->dev_root->leafsize;
 253         sdev->sectorsize = dev->dev_root->sectorsize;
 254         atomic_set(&sdev->in_flight, 0);
 255         atomic_set(&sdev->fixup_cnt, 0);
 256         atomic_set(&sdev->cancel_req, 0);
 257         sdev->csum_size = btrfs_super_csum_size(fs_info->super_copy);
 258         INIT_LIST_HEAD(&sdev->csum_list);
 259
 260         spin_lock_init(&sdev->list_lock);
 261         spin_lock_init(&sdev->stat_lock);
 262         init_waitqueue_head(&sdev->list_wait);
 263         return sdev;
 264
 265 nomem:
 266         scrub_free_dev(sdev);
 267         return ERR_PTR(-ENOMEM);
 268 }
 269
 270 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root, void *ctx)
 271 {
 272         u64 isize;
 273         u32 nlink;
 274         int ret;
 275         int i;
 276         struct extent_buffer *eb;
 277         struct btrfs_inode_item *inode_item;
 278         struct scrub_warning *swarn = ctx;
 279         struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
 280         struct inode_fs_paths *ipath = NULL;
 281         struct btrfs_root *local_root;
 282         struct btrfs_key root_key;
 283
 284         root_key.objectid = root;
 285         root_key.type = BTRFS_ROOT_ITEM_KEY;
 286         root_key.offset = (u64)-1;
 287         local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
 288         if (IS_ERR(local_root)) {
 289                 ret = PTR_ERR(local_root);
 290                 goto err;
 291         }
 292
 293         ret = inode_item_info(inum, 0, local_root, swarn->path);
 294         if (ret) {
 295                 btrfs_release_path(swarn->path);
 296                 goto err;
 297         }
 298
 299         eb = swarn->path->nodes[0];
 300         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 301                                         struct btrfs_inode_item);
 302         isize = btrfs_inode_size(eb, inode_item);
 303         nlink = btrfs_inode_nlink(eb, inode_item);
 304         btrfs_release_path(swarn->path);
 305
 306         ipath = init_ipath(4096, local_root, swarn->path);
 307         if (IS_ERR(ipath)) {
 308                 ret = PTR_ERR(ipath);
 309                 ipath = NULL;
 310                 goto err;
 311         }
 312         ret = paths_from_inode(inum, ipath);
 313
 314         if (ret < 0)
 315                 goto err;
 316
 317         /*
 318          * we deliberately ignore the bit ipath might have been too small to
 319          * hold all of the paths here
 320          */
 321         for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 322                 printk(KERN_WARNING "btrfs: %s at logical %llu on dev "
 323                         "%s, sector %llu, root %llu, inode %llu, offset %llu, "
 324                         "length %llu, links %u (path: %s)\n", swarn->errstr,
 325                         swarn->logical, swarn->dev->name,
 326                         (unsigned long long)swarn->sector, root, inum, offset,
 327                         min(isize - offset, (u64)PAGE_SIZE), nlink,
 328                         (char *)(unsigned long)ipath->fspath->val[i]);
 329
 330         free_ipath(ipath);
 331         return 0;
 332
 333 err:
 334         printk(KERN_WARNING "btrfs: %s at logical %llu on dev "
 335                 "%s, sector %llu, root %llu, inode %llu, offset %llu: path "
 336                 "resolving failed with ret=%d\n", swarn->errstr,
 337                 swarn->logical, swarn->dev->name,
 338                 (unsigned long long)swarn->sector, root, inum, offset, ret);
 339
 340         free_ipath(ipath);
 341         return 0;
 342 }
 343
 344 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
 345 {
 346         struct btrfs_device *dev = sblock->sdev->dev;
 347         struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
 348         struct btrfs_path *path;
 349         struct btrfs_key found_key;
 350         struct extent_buffer *eb;
 351         struct btrfs_extent_item *ei;
 352         struct scrub_warning swarn;
 353         u32 item_size;
 354         int ret;
 355         u64 ref_root;
 356         u8 ref_level;
 357         unsigned long ptr = 0;
 358         const int bufsize = 4096;
 359         u64 extent_item_pos;
 360
 361         path = btrfs_alloc_path();
 362
 363         swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
 364         swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
 365         BUG_ON(sblock->page_count < 1);
 366         swarn.sector = (sblock->pagev[0].physical) >> 9;
 367         swarn.logical = sblock->pagev[0].logical;
 368         swarn.errstr = errstr;
 369         swarn.dev = dev;
 370         swarn.msg_bufsize = bufsize;
 371         swarn.scratch_bufsize = bufsize;
 372
 373         if (!path || !swarn.scratch_buf || !swarn.msg_buf)
 374                 goto out;
 375
 376         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key);
 377         if (ret < 0)
 378                 goto out;
 379
 380         extent_item_pos = swarn.logical - found_key.objectid;
 381         swarn.extent_item_size = found_key.offset;
 382
 383         eb = path->nodes[0];
 384         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 385         item_size = btrfs_item_size_nr(eb, path->slots[0]);
 386         btrfs_release_path(path);
 387
 388         if (ret & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 389                 do {
 390                         ret = tree_backref_for_extent(&ptr, eb, ei, item_size,
 391                                                         &ref_root, &ref_level);
 392                         printk(KERN_WARNING
 393                                 "btrfs: %s at logical %llu on dev %s, "
 394                                 "sector %llu: metadata %s (level %d) in tree "
 395                                 "%llu\n", errstr, swarn.logical, dev->name,
 396                                 (unsigned long long)swarn.sector,
 397                                 ref_level ? "node" : "leaf",
 398                                 ret < 0 ? -1 : ref_level,
 399                                 ret < 0 ? -1 : ref_root);
 400                 } while (ret != 1);
 401         } else {
 402                 swarn.path = path;
 403                 iterate_extent_inodes(fs_info, found_key.objectid,
 404                                         extent_item_pos, 1,
 405                                         scrub_print_warning_inode, &swarn);
 406         }
 407
 408 out:
 409         btrfs_free_path(path);
 410         kfree(swarn.scratch_buf);
 411         kfree(swarn.msg_buf);
 412 }
 413
 414 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *ctx)
 415 {
 416         struct page *page = NULL;
 417         unsigned long index;
 418         struct scrub_fixup_nodatasum *fixup = ctx;
 419         int ret;
 420         int corrected = 0;
 421         struct btrfs_key key;
 422         struct inode *inode = NULL;
 423         u64 end = offset + PAGE_SIZE - 1;
 424         struct btrfs_root *local_root;
 425
 426         key.objectid = root;
 427         key.type = BTRFS_ROOT_ITEM_KEY;
 428         key.offset = (u64)-1;
 429         local_root = btrfs_read_fs_root_no_name(fixup->root->fs_info, &key);
 430         if (IS_ERR(local_root))
 431                 return PTR_ERR(local_root);
 432
 433         key.type = BTRFS_INODE_ITEM_KEY;
 434         key.objectid = inum;
 435         key.offset = 0;
 436         inode = btrfs_iget(fixup->root->fs_info->sb, &key, local_root, NULL);
 437         if (IS_ERR(inode))
 438                 return PTR_ERR(inode);
 439
 440         index = offset >> PAGE_CACHE_SHIFT;
 441
 442         page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
 443         if (!page) {
 444                 ret = -ENOMEM;
 445                 goto out;
 446         }
 447
 448         if (PageUptodate(page)) {
 449                 struct btrfs_mapping_tree *map_tree;
 450                 if (PageDirty(page)) {
 451                         /*
 452                          * we need to write the data to the defect sector. the
 453                          * data that was in that sector is not in memory,
 454                          * because the page was modified. we must not write the
 455                          * modified page to that sector.
 456                          *
 457                          * TODO: what could be done here: wait for the delalloc
 458                          *       runner to write out that page (might involve
 459                          *       COW) and see whether the sector is still
 460                          *       referenced afterwards.
 461                          *
 462                          * For the meantime, we'll treat this error
 463                          * incorrectable, although there is a chance that a
 464                          * later scrub will find the bad sector again and that
 465                          * there's no dirty page in memory, then.
 466                          */
 467                         ret = -EIO;
 468                         goto out;
 469                 }
 470                 map_tree = &BTRFS_I(inode)->root->fs_info->mapping_tree;
 471                 ret = repair_io_failure(map_tree, offset, PAGE_SIZE,
 472                                         fixup->logical, page,
 473                                         fixup->mirror_num);
 474                 unlock_page(page);
 475                 corrected = !ret;
 476         } else {
 477                 /*
 478                  * we need to get good data first. the general readpage path
 479                  * will call repair_io_failure for us, we just have to make
 480                  * sure we read the bad mirror.
 481                  */
 482                 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 483                                         EXTENT_DAMAGED, GFP_NOFS);
 484                 if (ret) {
 485                         /* set_extent_bits should give proper error */
 486                         WARN_ON(ret > 0);
 487                         if (ret > 0)
 488                                 ret = -EFAULT;
 489                         goto out;
 490                 }
 491
 492                 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
 493                                                 btrfs_get_extent,
 494                                                 fixup->mirror_num);
 495                 wait_on_page_locked(page);
 496
 497                 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
 498                                                 end, EXTENT_DAMAGED, 0, NULL);
 499                 if (!corrected)
 500                         clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 501                                                 EXTENT_DAMAGED, GFP_NOFS);
 502         }
 503
 504 out:
 505         if (page)
 506                 put_page(page);
 507         if (inode)
 508                 iput(inode);
 509
 510         if (ret < 0)
 511                 return ret;
 512
 513         if (ret == 0 && corrected) {
 514                 /*
 515                  * we only need to call readpage for one of the inodes belonging
 516                  * to this extent. so make iterate_extent_inodes stop
 517                  */
 518                 return 1;
 519         }
 520
 521         return -EIO;
 522 }
 523
 524 static void scrub_fixup_nodatasum(struct btrfs_work *work)
 525 {
 526         int ret;
 527         struct scrub_fixup_nodatasum *fixup;
 528         struct scrub_dev *sdev;
 529         struct btrfs_trans_handle *trans = NULL;
 530         struct btrfs_fs_info *fs_info;
 531         struct btrfs_path *path;
 532         int uncorrectable = 0;
 533
 534         fixup = container_of(work, struct scrub_fixup_nodatasum, work);
 535         sdev = fixup->sdev;
 536         fs_info = fixup->root->fs_info;
 537
 538         path = btrfs_alloc_path();
 539         if (!path) {
 540                 spin_lock(&sdev->stat_lock);
 541                 ++sdev->stat.malloc_errors;
 542                 spin_unlock(&sdev->stat_lock);
 543                 uncorrectable = 1;
 544                 goto out;
 545         }
 546
 547         trans = btrfs_join_transaction(fixup->root);
 548         if (IS_ERR(trans)) {
 549                 uncorrectable = 1;
 550                 goto out;
 551         }
 552
 553         /*
 554          * the idea is to trigger a regular read through the standard path. we
 555          * read a page from the (failed) logical address by specifying the
 556          * corresponding copynum of the failed sector. thus, that readpage is
 557          * expected to fail.
 558          * that is the point where on-the-fly error correction will kick in
 559          * (once it's finished) and rewrite the failed sector if a good copy
 560          * can be found.
 561          */
 562         ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
 563                                                 path, scrub_fixup_readpage,
 564                                                 fixup);
 565         if (ret < 0) {
 566                 uncorrectable = 1;
 567                 goto out;
 568         }
 569         WARN_ON(ret != 1);
 570
 571         spin_lock(&sdev->stat_lock);
 572         ++sdev->stat.corrected_errors;
 573         spin_unlock(&sdev->stat_lock);
 574
 575 out:
 576         if (trans && !IS_ERR(trans))
 577                 btrfs_end_transaction(trans, fixup->root);
 578         if (uncorrectable) {
 579                 spin_lock(&sdev->stat_lock);
 580                 ++sdev->stat.uncorrectable_errors;
 581                 spin_unlock(&sdev->stat_lock);
 582                 printk_ratelimited(KERN_ERR
 583                         "btrfs: unable to fixup (nodatasum) error at logical %llu on dev %s\n",
 584                         (unsigned long long)fixup->logical, sdev->dev->name);
 585         }
 586
 587         btrfs_free_path(path);
 588         kfree(fixup);
 589
 590         /* see caller why we're pretending to be paused in the scrub counters */
 591         mutex_lock(&fs_info->scrub_lock);
 592         atomic_dec(&fs_info->scrubs_running);
 593         atomic_dec(&fs_info->scrubs_paused);
 594         mutex_unlock(&fs_info->scrub_lock);
 595         atomic_dec(&sdev->fixup_cnt);
 596         wake_up(&fs_info->scrub_pause_wait);
 597         wake_up(&sdev->list_wait);
 598 }
 599
 600 /*
 601  * scrub_handle_errored_block gets called when either verification of the
 602  * pages failed or the bio failed to read, e.g. with EIO. In the latter
 603  * case, this function handles all pages in the bio, even though only one
 604  * may be bad.
 605  * The goal of this function is to repair the errored block by using the
 606  * contents of one of the mirrors.
 607  */
 608 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
 609 {
 610         struct scrub_dev *sdev = sblock_to_check->sdev;
 611         struct btrfs_fs_info *fs_info;
 612         u64 length;
 613         u64 logical;
 614         u64 generation;
 615         unsigned int failed_mirror_index;
 616         unsigned int is_metadata;
 617         unsigned int have_csum;
 618         u8 *csum;
 619         struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
 620         struct scrub_block *sblock_bad;
 621         int ret;
 622         int mirror_index;
 623         int page_num;
 624         int success;
 625         static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
 626                                       DEFAULT_RATELIMIT_BURST);
 627
 628         BUG_ON(sblock_to_check->page_count < 1);
 629         fs_info = sdev->dev->dev_root->fs_info;
 630         length = sblock_to_check->page_count * PAGE_SIZE;
 631         logical = sblock_to_check->pagev[0].logical;
 632         generation = sblock_to_check->pagev[0].generation;
 633         BUG_ON(sblock_to_check->pagev[0].mirror_num < 1);
 634         failed_mirror_index = sblock_to_check->pagev[0].mirror_num - 1;
 635         is_metadata = !(sblock_to_check->pagev[0].flags &
 636                         BTRFS_EXTENT_FLAG_DATA);
 637         have_csum = sblock_to_check->pagev[0].have_csum;
 638         csum = sblock_to_check->pagev[0].csum;
 639
 640         /*
 641          * read all mirrors one after the other. This includes to
 642          * re-read the extent or metadata block that failed (that was
 643          * the cause that this fixup code is called) another time,
 644          * page by page this time in order to know which pages
 645          * caused I/O errors and which ones are good (for all mirrors).
 646          * It is the goal to handle the situation when more than one
 647          * mirror contains I/O errors, but the errors do not
 648          * overlap, i.e. the data can be repaired by selecting the
 649          * pages from those mirrors without I/O error on the
 650          * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
 651          * would be that mirror #1 has an I/O error on the first page,
 652          * the second page is good, and mirror #2 has an I/O error on
 653          * the second page, but the first page is good.
 654          * Then the first page of the first mirror can be repaired by
 655          * taking the first page of the second mirror, and the
 656          * second page of the second mirror can be repaired by
 657          * copying the contents of the 2nd page of the 1st mirror.
 658          * One more note: if the pages of one mirror contain I/O
 659          * errors, the checksum cannot be verified. In order to get
 660          * the best data for repairing, the first attempt is to find
 661          * a mirror without I/O errors and with a validated checksum.
 662          * Only if this is not possible, the pages are picked from
 663          * mirrors with I/O errors without considering the checksum.
 664          * If the latter is the case, at the end, the checksum of the
 665          * repaired area is verified in order to correctly maintain
 666          * the statistics.
 667          */
 668
 669         sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
 670                                      sizeof(*sblocks_for_recheck),
 671                                      GFP_NOFS);
 672         if (!sblocks_for_recheck) {
 673                 spin_lock(&sdev->stat_lock);
 674                 sdev->stat.malloc_errors++;
 675                 sdev->stat.read_errors++;
 676                 sdev->stat.uncorrectable_errors++;
 677                 spin_unlock(&sdev->stat_lock);
 678                 goto out;
 679         }
 680
 681         /* setup the context, map the logical blocks and alloc the pages */
 682         ret = scrub_setup_recheck_block(sdev, &fs_info->mapping_tree, length,
 683                                         logical, sblocks_for_recheck);
 684         if (ret) {
 685                 spin_lock(&sdev->stat_lock);
 686                 sdev->stat.read_errors++;
 687                 sdev->stat.uncorrectable_errors++;
 688                 spin_unlock(&sdev->stat_lock);
 689                 goto out;
 690         }
 691         BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
 692         sblock_bad = sblocks_for_recheck + failed_mirror_index;
 693
 694         /* build and submit the bios for the failed mirror, check checksums */
 695         ret = scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
 696                                   csum, generation, sdev->csum_size);
 697         if (ret) {
 698                 spin_lock(&sdev->stat_lock);
 699                 sdev->stat.read_errors++;
 700                 sdev->stat.uncorrectable_errors++;
 701                 spin_unlock(&sdev->stat_lock);
 702                 goto out;
 703         }
 704
 705         if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
 706             sblock_bad->no_io_error_seen) {
 707                 /*
 708                  * the error disappeared after reading page by page, or
 709                  * the area was part of a huge bio and other parts of the
 710                  * bio caused I/O errors, or the block layer merged several
 711                  * read requests into one and the error is caused by a
 712                  * different bio (usually one of the two latter cases is
 713                  * the cause)
 714                  */
 715                 spin_lock(&sdev->stat_lock);
 716                 sdev->stat.unverified_errors++;
 717                 spin_unlock(&sdev->stat_lock);
 718
 719                 goto out;
 720         }
 721
 722         if (!sblock_bad->no_io_error_seen) {
 723                 spin_lock(&sdev->stat_lock);
 724                 sdev->stat.read_errors++;
 725                 spin_unlock(&sdev->stat_lock);
 726                 if (__ratelimit(&_rs))
 727                         scrub_print_warning("i/o error", sblock_to_check);
 728         } else if (sblock_bad->checksum_error) {
 729                 spin_lock(&sdev->stat_lock);
 730                 sdev->stat.csum_errors++;
 731                 spin_unlock(&sdev->stat_lock);
 732                 if (__ratelimit(&_rs))
 733                         scrub_print_warning("checksum error", sblock_to_check);
 734         } else if (sblock_bad->header_error) {
 735                 spin_lock(&sdev->stat_lock);
 736                 sdev->stat.verify_errors++;
 737                 spin_unlock(&sdev->stat_lock);
 738                 if (__ratelimit(&_rs))
 739                         scrub_print_warning("checksum/header error",
 740                                             sblock_to_check);
 741         }
 742
 743         if (sdev->readonly)
 744                 goto did_not_correct_error;
 745
 746         if (!is_metadata && !have_csum) {
 747                 struct scrub_fixup_nodatasum *fixup_nodatasum;
 748
 749                 /*
 750                  * !is_metadata and !have_csum, this means that the data
 751                  * might not be COW'ed, that it might be modified
 752                  * concurrently. The general strategy to work on the
 753                  * commit root does not help in the case when COW is not
 754                  * used.
 755                  */
 756                 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
 757                 if (!fixup_nodatasum)
 758                         goto did_not_correct_error;
 759                 fixup_nodatasum->sdev = sdev;
 760                 fixup_nodatasum->logical = logical;
 761                 fixup_nodatasum->root = fs_info->extent_root;
 762                 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
 763                 /*
 764                  * increment scrubs_running to prevent cancel requests from
 765                  * completing as long as a fixup worker is running. we must also
 766                  * increment scrubs_paused to prevent deadlocking on pause
 767                  * requests used for transactions commits (as the worker uses a
 768                  * transaction context). it is safe to regard the fixup worker
 769                  * as paused for all matters practical. effectively, we only
 770                  * avoid cancellation requests from completing.
 771                  */
 772                 mutex_lock(&fs_info->scrub_lock);
 773                 atomic_inc(&fs_info->scrubs_running);
 774                 atomic_inc(&fs_info->scrubs_paused);
 775                 mutex_unlock(&fs_info->scrub_lock);
 776                 atomic_inc(&sdev->fixup_cnt);
 777                 fixup_nodatasum->work.func = scrub_fixup_nodatasum;
 778                 btrfs_queue_worker(&fs_info->scrub_workers,
 779                                    &fixup_nodatasum->work);
 780                 goto out;
 781         }
 782
 783         /*
 784          * now build and submit the bios for the other mirrors, check
 785          * checksums
 786          */
 787         for (mirror_index = 0;
 788              mirror_index < BTRFS_MAX_MIRRORS &&
 789              sblocks_for_recheck[mirror_index].page_count > 0;
 790              mirror_index++) {
 791                 if (mirror_index == failed_mirror_index)
 792                         continue;
 793
 794                 /* build and submit the bios, check checksums */
 795                 ret = scrub_recheck_block(fs_info,
 796                                           sblocks_for_recheck + mirror_index,
 797                                           is_metadata, have_csum, csum,
 798                                           generation, sdev->csum_size);
 799                 if (ret)
 800                         goto did_not_correct_error;
 801         }
 802
 803         /*
 804          * first try to pick the mirror which is completely without I/O
 805          * errors and also does not have a checksum error.
 806          * If one is found, and if a checksum is present, the full block
 807          * that is known to contain an error is rewritten. Afterwards
 808          * the block is known to be corrected.
 809          * If a mirror is found which is completely correct, and no
 810          * checksum is present, only those pages are rewritten that had
 811          * an I/O error in the block to be repaired, since it cannot be
 812          * determined, which copy of the other pages is better (and it
 813          * could happen otherwise that a correct page would be
 814          * overwritten by a bad one).
 815          */
 816         for (mirror_index = 0;
 817              mirror_index < BTRFS_MAX_MIRRORS &&
 818              sblocks_for_recheck[mirror_index].page_count > 0;
 819              mirror_index++) {
 820                 struct scrub_block *sblock_other = sblocks_for_recheck +
 821                                                    mirror_index;
 822
 823                 if (!sblock_other->header_error &&
 824                     !sblock_other->checksum_error &&
 825                     sblock_other->no_io_error_seen) {
 826                         int force_write = is_metadata || have_csum;
 827
 828                         ret = scrub_repair_block_from_good_copy(sblock_bad,
 829                                                                 sblock_other,
 830                                                                 force_write);
 831                         if (0 == ret)
 832                                 goto corrected_error;
 833                 }
 834         }
 835
 836         /*
 837          * in case of I/O errors in the area that is supposed to be
 838          * repaired, continue by picking good copies of those pages.
 839          * Select the good pages from mirrors to rewrite bad pages from
 840          * the area to fix. Afterwards verify the checksum of the block
 841          * that is supposed to be repaired. This verification step is
 842          * only done for the purpose of statistic counting and for the
 843          * final scrub report, whether errors remain.
 844          * A perfect algorithm could make use of the checksum and try
 845          * all possible combinations of pages from the different mirrors
 846          * until the checksum verification succeeds. For example, when
 847          * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
 848          * of mirror #2 is readable but the final checksum test fails,
 849          * then the 2nd page of mirror #3 could be tried, whether now
 850          * the final checksum succeedes. But this would be a rare
 851          * exception and is therefore not implemented. At least it is
 852          * avoided that the good copy is overwritten.
 853          * A more useful improvement would be to pick the sectors
 854          * without I/O error based on sector sizes (512 bytes on legacy
 855          * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
 856          * mirror could be repaired by taking 512 byte of a different
 857          * mirror, even if other 512 byte sectors in the same PAGE_SIZE
 858          * area are unreadable.
 859          */
 860
 861         /* can only fix I/O errors from here on */
 862         if (sblock_bad->no_io_error_seen)
 863                 goto did_not_correct_error;
 864
 865         success = 1;
 866         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
 867                 struct scrub_page *page_bad = sblock_bad->pagev + page_num;
 868
 869                 if (!page_bad->io_error)
 870                         continue;
 871
 872                 for (mirror_index = 0;
 873                      mirror_index < BTRFS_MAX_MIRRORS &&
 874                      sblocks_for_recheck[mirror_index].page_count > 0;
 875                      mirror_index++) {
 876                         struct scrub_block *sblock_other = sblocks_for_recheck +
 877                                                            mirror_index;
 878                         struct scrub_page *page_other = sblock_other->pagev +
 879                                                         page_num;
 880
 881                         if (!page_other->io_error) {
 882                                 ret = scrub_repair_page_from_good_copy(
 883                                         sblock_bad, sblock_other, page_num, 0);
 884                                 if (0 == ret) {
 885                                         page_bad->io_error = 0;
 886                                         break; /* succeeded for this page */
 887                                 }
 888                         }
 889                 }
 890
 891                 if (page_bad->io_error) {
 892                         /* did not find a mirror to copy the page from */
 893                         success = 0;
 894                 }
 895         }
 896
 897         if (success) {
 898                 if (is_metadata || have_csum) {
 899                         /*
 900                          * need to verify the checksum now that all
 901                          * sectors on disk are repaired (the write
 902                          * request for data to be repaired is on its way).
 903                          * Just be lazy and use scrub_recheck_block()
 904                          * which re-reads the data before the checksum
 905                          * is verified, but most likely the data comes out
 906                          * of the page cache.
 907                          */
 908                         ret = scrub_recheck_block(fs_info, sblock_bad,
 909                                                   is_metadata, have_csum, csum,
 910                                                   generation, sdev->csum_size);
 911                         if (!ret && !sblock_bad->header_error &&
 912                             !sblock_bad->checksum_error &&
 913                             sblock_bad->no_io_error_seen)
 914                                 goto corrected_error;
 915                         else
 916                                 goto did_not_correct_error;
 917                 } else {
 918 corrected_error:
 919                         spin_lock(&sdev->stat_lock);
 920                         sdev->stat.corrected_errors++;
 921                         spin_unlock(&sdev->stat_lock);
 922                         printk_ratelimited(KERN_ERR
 923                                 "btrfs: fixed up error at logical %llu on dev %s\n",
 924                                 (unsigned long long)logical, sdev->dev->name);
 925                 }
 926         } else {
 927 did_not_correct_error:
 928                 spin_lock(&sdev->stat_lock);
 929                 sdev->stat.uncorrectable_errors++;
 930                 spin_unlock(&sdev->stat_lock);
 931                 printk_ratelimited(KERN_ERR
 932                         "btrfs: unable to fixup (regular) error at logical %llu on dev %s\n",
 933                         (unsigned long long)logical, sdev->dev->name);
 934         }
 935
 936 out:
 937         if (sblocks_for_recheck) {
 938                 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
 939                      mirror_index++) {
 940                         struct scrub_block *sblock = sblocks_for_recheck +
 941                                                      mirror_index;
 942                         int page_index;
 943
 944                         for (page_index = 0; page_index < SCRUB_PAGES_PER_BIO;
 945                              page_index++)
 946                                 if (sblock->pagev[page_index].page)
 947                                         __free_page(
 948                                                 sblock->pagev[page_index].page);
 949                 }
 950                 kfree(sblocks_for_recheck);
 951         }
 952
 953         return 0;
 954 }
 955
 956 static int scrub_setup_recheck_block(struct scrub_dev *sdev,
 957                                      struct btrfs_mapping_tree *map_tree,
 958                                      u64 length, u64 logical,
 959                                      struct scrub_block *sblocks_for_recheck)
 960 {
 961         int page_index;
 962         int mirror_index;
 963         int ret;
 964
 965         /*
 966          * note: the three members sdev, ref_count and outstanding_pages
 967          * are not used (and not set) in the blocks that are used for
 968          * the recheck procedure
 969          */
 970
 971         page_index = 0;
 972         while (length > 0) {
 973                 u64 sublen = min_t(u64, length, PAGE_SIZE);
 974                 u64 mapped_length = sublen;
 975                 struct btrfs_bio *bbio = NULL;
 976
 977                 /*
 978                  * with a length of PAGE_SIZE, each returned stripe
 979                  * represents one mirror
 980                  */
 981                 ret = btrfs_map_block(map_tree, WRITE, logical, &mapped_length,
 982                                       &bbio, 0);
 983                 if (ret || !bbio || mapped_length < sublen) {
 984                         kfree(bbio);
 985                         return -EIO;
 986                 }
 987
 988                 BUG_ON(page_index >= SCRUB_PAGES_PER_BIO);
 989                 for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
 990                      mirror_index++) {
 991                         struct scrub_block *sblock;
 992                         struct scrub_page *page;
 993
 994                         if (mirror_index >= BTRFS_MAX_MIRRORS)
 995                                 continue;
 996
 997                         sblock = sblocks_for_recheck + mirror_index;
 998                         page = sblock->pagev + page_index;
 999                         page->logical = logical;
1000                         page->physical = bbio->stripes[mirror_index].physical;
1001                         page->bdev = bbio->stripes[mirror_index].dev->bdev;
1002                         page->mirror_num = mirror_index + 1;
1003                         page->page = alloc_page(GFP_NOFS);
1004                         if (!page->page) {
1005                                 spin_lock(&sdev->stat_lock);
1006                                 sdev->stat.malloc_errors++;
1007                                 spin_unlock(&sdev->stat_lock);
1008                                 return -ENOMEM;
1009                         }
1010                         sblock->page_count++;
1011                 }
1012                 kfree(bbio);
1013                 length -= sublen;
1014                 logical += sublen;
1015                 page_index++;
1016         }
1017
1018         return 0;
1019 }
1020
1021 /*
1022  * this function will check the on disk data for checksum errors, header
1023  * errors and read I/O errors. If any I/O errors happen, the exact pages
1024  * which are errored are marked as being bad. The goal is to enable scrub
1025  * to take those pages that are not errored from all the mirrors so that
1026  * the pages that are errored in the just handled mirror can be repaired.
1027  */
1028 static int scrub_recheck_block(struct btrfs_fs_info *fs_info,
1029                                struct scrub_block *sblock, int is_metadata,
1030                                int have_csum, u8 *csum, u64 generation,
1031                                u16 csum_size)
1032 {
1033         int page_num;
1034
1035         sblock->no_io_error_seen = 1;
1036         sblock->header_error = 0;
1037         sblock->checksum_error = 0;
1038
1039         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1040                 struct bio *bio;
1041                 int ret;
1042                 struct scrub_page *page = sblock->pagev + page_num;
1043                 DECLARE_COMPLETION_ONSTACK(complete);
1044
1045                 BUG_ON(!page->page);
1046                 bio = bio_alloc(GFP_NOFS, 1);
1047                 if (!bio)
1048                         return -EIO;
1049                 bio->bi_bdev = page->bdev;
1050                 bio->bi_sector = page->physical >> 9;
1051                 bio->bi_end_io = scrub_complete_bio_end_io;
1052                 bio->bi_private = &complete;
1053
1054                 ret = bio_add_page(bio, page->page, PAGE_SIZE, 0);
1055                 if (PAGE_SIZE != ret) {
1056                         bio_put(bio);
1057                         return -EIO;
1058                 }
1059                 btrfsic_submit_bio(READ, bio);
1060
1061                 /* this will also unplug the queue */
1062                 wait_for_completion(&complete);
1063
1064                 page->io_error = !test_bit(BIO_UPTODATE, &bio->bi_flags);
1065                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1066                         sblock->no_io_error_seen = 0;
1067                 bio_put(bio);
1068         }
1069
1070         if (sblock->no_io_error_seen)
1071                 scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1072                                              have_csum, csum, generation,
1073                                              csum_size);
1074
1075         return 0;
1076 }
1077
1078 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1079                                          struct scrub_block *sblock,
1080                                          int is_metadata, int have_csum,
1081                                          const u8 *csum, u64 generation,
1082                                          u16 csum_size)
1083 {
1084         int page_num;
1085         u8 calculated_csum[BTRFS_CSUM_SIZE];
1086         u32 crc = ~(u32)0;
1087         struct btrfs_root *root = fs_info->extent_root;
1088         void *mapped_buffer;
1089
1090         BUG_ON(!sblock->pagev[0].page);
1091         if (is_metadata) {
1092                 struct btrfs_header *h;
1093
1094                 mapped_buffer = kmap_atomic(sblock->pagev[0].page);
1095                 h = (struct btrfs_header *)mapped_buffer;
1096
1097                 if (sblock->pagev[0].logical != le64_to_cpu(h->bytenr) ||
1098                     generation != le64_to_cpu(h->generation) ||
1099                     memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
1100                     memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1101                            BTRFS_UUID_SIZE))
1102                         sblock->header_error = 1;
1103                 csum = h->csum;
1104         } else {
1105                 if (!have_csum)
1106                         return;
1107
1108                 mapped_buffer = kmap_atomic(sblock->pagev[0].page);
1109         }
1110
1111         for (page_num = 0;;) {
1112                 if (page_num == 0 && is_metadata)
1113                         crc = btrfs_csum_data(root,
1114                                 ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1115                                 crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1116                 else
1117                         crc = btrfs_csum_data(root, mapped_buffer, crc,
1118                                               PAGE_SIZE);
1119
1120                 kunmap_atomic(mapped_buffer);
1121                 page_num++;
1122                 if (page_num >= sblock->page_count)
1123                         break;
1124                 BUG_ON(!sblock->pagev[page_num].page);
1125
1126                 mapped_buffer = kmap_atomic(sblock->pagev[page_num].page);
1127         }
1128
1129         btrfs_csum_final(crc, calculated_csum);
1130         if (memcmp(calculated_csum, csum, csum_size))
1131                 sblock->checksum_error = 1;
1132 }
1133
1134 static void scrub_complete_bio_end_io(struct bio *bio, int err)
1135 {
1136         complete((struct completion *)bio->bi_private);
1137 }
1138
1139 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1140                                              struct scrub_block *sblock_good,
1141                                              int force_write)
1142 {
1143         int page_num;
1144         int ret = 0;
1145
1146         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1147                 int ret_sub;
1148
1149                 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1150                                                            sblock_good,
1151                                                            page_num,
1152                                                            force_write);
1153                 if (ret_sub)
1154                         ret = ret_sub;
1155         }
1156
1157         return ret;
1158 }
1159
1160 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1161                                             struct scrub_block *sblock_good,
1162                                             int page_num, int force_write)
1163 {
1164         struct scrub_page *page_bad = sblock_bad->pagev + page_num;
1165         struct scrub_page *page_good = sblock_good->pagev + page_num;
1166
1167         BUG_ON(sblock_bad->pagev[page_num].page == NULL);
1168         BUG_ON(sblock_good->pagev[page_num].page == NULL);
1169         if (force_write || sblock_bad->header_error ||
1170             sblock_bad->checksum_error || page_bad->io_error) {
1171                 struct bio *bio;
1172                 int ret;
1173                 DECLARE_COMPLETION_ONSTACK(complete);
1174
1175                 bio = bio_alloc(GFP_NOFS, 1);
1176                 if (!bio)
1177                         return -EIO;
1178                 bio->bi_bdev = page_bad->bdev;
1179                 bio->bi_sector = page_bad->physical >> 9;
1180                 bio->bi_end_io = scrub_complete_bio_end_io;
1181                 bio->bi_private = &complete;
1182
1183                 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1184                 if (PAGE_SIZE != ret) {
1185                         bio_put(bio);
1186                         return -EIO;
1187                 }
1188                 btrfsic_submit_bio(WRITE, bio);
1189
1190                 /* this will also unplug the queue */
1191                 wait_for_completion(&complete);
1192                 bio_put(bio);
1193         }
1194
1195         return 0;
1196 }
1197
1198 static void scrub_checksum(struct scrub_block *sblock)
1199 {
1200         u64 flags;
1201         int ret;
1202
1203         BUG_ON(sblock->page_count < 1);
1204         flags = sblock->pagev[0].flags;
1205         ret = 0;
1206         if (flags & BTRFS_EXTENT_FLAG_DATA)
1207                 ret = scrub_checksum_data(sblock);
1208         else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1209                 ret = scrub_checksum_tree_block(sblock);
1210         else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1211                 (void)scrub_checksum_super(sblock);
1212         else
1213                 WARN_ON(1);
1214         if (ret)
1215                 scrub_handle_errored_block(sblock);
1216 }
1217
1218 static int scrub_checksum_data(struct scrub_block *sblock)
1219 {
1220         struct scrub_dev *sdev = sblock->sdev;
1221         u8 csum[BTRFS_CSUM_SIZE];
1222         u8 *on_disk_csum;
1223         struct page *page;
1224         void *buffer;
1225         u32 crc = ~(u32)0;
1226         int fail = 0;
1227         struct btrfs_root *root = sdev->dev->dev_root;
1228         u64 len;
1229         int index;
1230
1231         BUG_ON(sblock->page_count < 1);
1232         if (!sblock->pagev[0].have_csum)
1233                 return 0;
1234
1235         on_disk_csum = sblock->pagev[0].csum;
1236         page = sblock->pagev[0].page;
1237         buffer = kmap_atomic(page);
1238
1239         len = sdev->sectorsize;
1240         index = 0;
1241         for (;;) {
1242                 u64 l = min_t(u64, len, PAGE_SIZE);
1243
1244                 crc = btrfs_csum_data(root, buffer, crc, l);
1245                 kunmap_atomic(buffer);
1246                 len -= l;
1247                 if (len == 0)
1248                         break;
1249                 index++;
1250                 BUG_ON(index >= sblock->page_count);
1251                 BUG_ON(!sblock->pagev[index].page);
1252                 page = sblock->pagev[index].page;
1253                 buffer = kmap_atomic(page);
1254         }
1255
1256         btrfs_csum_final(crc, csum);
1257         if (memcmp(csum, on_disk_csum, sdev->csum_size))
1258                 fail = 1;
1259
1260         return fail;
1261 }
1262
1263 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1264 {
1265         struct scrub_dev *sdev = sblock->sdev;
1266         struct btrfs_header *h;
1267         struct btrfs_root *root = sdev->dev->dev_root;
1268         struct btrfs_fs_info *fs_info = root->fs_info;
1269         u8 calculated_csum[BTRFS_CSUM_SIZE];
1270         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1271         struct page *page;
1272         void *mapped_buffer;
1273         u64 mapped_size;
1274         void *p;
1275         u32 crc = ~(u32)0;
1276         int fail = 0;
1277         int crc_fail = 0;
1278         u64 len;
1279         int index;
1280
1281         BUG_ON(sblock->page_count < 1);
1282         page = sblock->pagev[0].page;
1283         mapped_buffer = kmap_atomic(page);
1284         h = (struct btrfs_header *)mapped_buffer;
1285         memcpy(on_disk_csum, h->csum, sdev->csum_size);
1286
1287         /*
1288          * we don't use the getter functions here, as we
1289          * a) don't have an extent buffer and
1290          * b) the page is already kmapped
1291          */
1292
1293         if (sblock->pagev[0].logical != le64_to_cpu(h->bytenr))
1294                 ++fail;
1295
1296         if (sblock->pagev[0].generation != le64_to_cpu(h->generation))
1297                 ++fail;
1298
1299         if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1300                 ++fail;
1301
1302         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1303                    BTRFS_UUID_SIZE))
1304                 ++fail;
1305
1306         BUG_ON(sdev->nodesize != sdev->leafsize);
1307         len = sdev->nodesize - BTRFS_CSUM_SIZE;
1308         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1309         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1310         index = 0;
1311         for (;;) {
1312                 u64 l = min_t(u64, len, mapped_size);
1313
1314                 crc = btrfs_csum_data(root, p, crc, l);
1315                 kunmap_atomic(mapped_buffer);
1316                 len -= l;
1317                 if (len == 0)
1318                         break;
1319                 index++;
1320                 BUG_ON(index >= sblock->page_count);
1321                 BUG_ON(!sblock->pagev[index].page);
1322                 page = sblock->pagev[index].page;
1323                 mapped_buffer = kmap_atomic(page);
1324                 mapped_size = PAGE_SIZE;
1325                 p = mapped_buffer;
1326         }
1327
1328         btrfs_csum_final(crc, calculated_csum);
1329         if (memcmp(calculated_csum, on_disk_csum, sdev->csum_size))
1330                 ++crc_fail;
1331
1332         return fail || crc_fail;
1333 }
1334
1335 static int scrub_checksum_super(struct scrub_block *sblock)
1336 {
1337         struct btrfs_super_block *s;
1338         struct scrub_dev *sdev = sblock->sdev;
1339         struct btrfs_root *root = sdev->dev->dev_root;
1340         struct btrfs_fs_info *fs_info = root->fs_info;
1341         u8 calculated_csum[BTRFS_CSUM_SIZE];
1342         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1343         struct page *page;
1344         void *mapped_buffer;
1345         u64 mapped_size;
1346         void *p;
1347         u32 crc = ~(u32)0;
1348         int fail = 0;
1349         u64 len;
1350         int index;
1351
1352         BUG_ON(sblock->page_count < 1);
1353         page = sblock->pagev[0].page;
1354         mapped_buffer = kmap_atomic(page);
1355         s = (struct btrfs_super_block *)mapped_buffer;
1356         memcpy(on_disk_csum, s->csum, sdev->csum_size);
1357
1358         if (sblock->pagev[0].logical != le64_to_cpu(s->bytenr))
1359                 ++fail;
1360
1361         if (sblock->pagev[0].generation != le64_to_cpu(s->generation))
1362                 ++fail;
1363
1364         if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1365                 ++fail;
1366
1367         len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1368         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1369         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1370         index = 0;
1371         for (;;) {
1372                 u64 l = min_t(u64, len, mapped_size);
1373
1374                 crc = btrfs_csum_data(root, p, crc, l);
1375                 kunmap_atomic(mapped_buffer);
1376                 len -= l;
1377                 if (len == 0)
1378                         break;
1379                 index++;
1380                 BUG_ON(index >= sblock->page_count);
1381                 BUG_ON(!sblock->pagev[index].page);
1382                 page = sblock->pagev[index].page;
1383                 mapped_buffer = kmap_atomic(page);
1384                 mapped_size = PAGE_SIZE;
1385                 p = mapped_buffer;
1386         }
1387
1388         btrfs_csum_final(crc, calculated_csum);
1389         if (memcmp(calculated_csum, on_disk_csum, sdev->csum_size))
1390                 ++fail;
1391
1392         if (fail) {
1393                 /*
1394                  * if we find an error in a super block, we just report it.
1395                  * They will get written with the next transaction commit
1396                  * anyway
1397                  */
1398                 spin_lock(&sdev->stat_lock);
1399                 ++sdev->stat.super_errors;
1400                 spin_unlock(&sdev->stat_lock);
1401         }
1402
1403         return fail;
1404 }
1405
1406 static void scrub_block_get(struct scrub_block *sblock)
1407 {
1408         atomic_inc(&sblock->ref_count);
1409 }
1410
1411 static void scrub_block_put(struct scrub_block *sblock)
1412 {
1413         if (atomic_dec_and_test(&sblock->ref_count)) {
1414                 int i;
1415
1416                 for (i = 0; i < sblock->page_count; i++)
1417                         if (sblock->pagev[i].page)
1418                                 __free_page(sblock->pagev[i].page);
1419                 kfree(sblock);
1420         }
1421 }
1422
1423 static void scrub_submit(struct scrub_dev *sdev)
1424 {
1425         struct scrub_bio *sbio;
1426
1427         if (sdev->curr == -1)
1428                 return;
1429
1430         sbio = sdev->bios[sdev->curr];
1431         sdev->curr = -1;
1432         atomic_inc(&sdev->in_flight);
1433
1434         btrfsic_submit_bio(READ, sbio->bio);
1435 }
1436
1437 static int scrub_add_page_to_bio(struct scrub_dev *sdev,
1438                                  struct scrub_page *spage)
1439 {
1440         struct scrub_block *sblock = spage->sblock;
1441         struct scrub_bio *sbio;
1442         int ret;
1443
1444 again:
1445         /*
1446          * grab a fresh bio or wait for one to become available
1447          */
1448         while (sdev->curr == -1) {
1449                 spin_lock(&sdev->list_lock);
1450                 sdev->curr = sdev->first_free;
1451                 if (sdev->curr != -1) {
1452                         sdev->first_free = sdev->bios[sdev->curr]->next_free;
1453                         sdev->bios[sdev->curr]->next_free = -1;
1454                         sdev->bios[sdev->curr]->page_count = 0;
1455                         spin_unlock(&sdev->list_lock);
1456                 } else {
1457                         spin_unlock(&sdev->list_lock);
1458                         wait_event(sdev->list_wait, sdev->first_free != -1);
1459                 }
1460         }
1461         sbio = sdev->bios[sdev->curr];
1462         if (sbio->page_count == 0) {
1463                 struct bio *bio;
1464
1465                 sbio->physical = spage->physical;
1466                 sbio->logical = spage->logical;
1467                 bio = sbio->bio;
1468                 if (!bio) {
1469                         bio = bio_alloc(GFP_NOFS, sdev->pages_per_bio);
1470                         if (!bio)
1471                                 return -ENOMEM;
1472                         sbio->bio = bio;
1473                 }
1474
1475                 bio->bi_private = sbio;
1476                 bio->bi_end_io = scrub_bio_end_io;
1477                 bio->bi_bdev = sdev->dev->bdev;
1478                 bio->bi_sector = spage->physical >> 9;
1479                 sbio->err = 0;
1480         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1481                    spage->physical ||
1482                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1483                    spage->logical) {
1484                 scrub_submit(sdev);
1485                 goto again;
1486         }
1487
1488         sbio->pagev[sbio->page_count] = spage;
1489         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1490         if (ret != PAGE_SIZE) {
1491                 if (sbio->page_count < 1) {
1492                         bio_put(sbio->bio);
1493                         sbio->bio = NULL;
1494                         return -EIO;
1495                 }
1496                 scrub_submit(sdev);
1497                 goto again;
1498         }
1499
1500         scrub_block_get(sblock); /* one for the added page */
1501         atomic_inc(&sblock->outstanding_pages);
1502         sbio->page_count++;
1503         if (sbio->page_count == sdev->pages_per_bio)
1504                 scrub_submit(sdev);
1505
1506         return 0;
1507 }
1508
1509 static int scrub_pages(struct scrub_dev *sdev, u64 logical, u64 len,
1510                        u64 physical, u64 flags, u64 gen, int mirror_num,
1511                        u8 *csum, int force)
1512 {
1513         struct scrub_block *sblock;
1514         int index;
1515
1516         sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
1517         if (!sblock) {
1518                 spin_lock(&sdev->stat_lock);
1519                 sdev->stat.malloc_errors++;
1520                 spin_unlock(&sdev->stat_lock);
1521                 return -ENOMEM;
1522         }
1523
1524         /* one ref inside this function, plus one for each page later on */
1525         atomic_set(&sblock->ref_count, 1);
1526         sblock->sdev = sdev;
1527         sblock->no_io_error_seen = 1;
1528
1529         for (index = 0; len > 0; index++) {
1530                 struct scrub_page *spage = sblock->pagev + index;
1531                 u64 l = min_t(u64, len, PAGE_SIZE);
1532
1533                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
1534                 spage->page = alloc_page(GFP_NOFS);
1535                 if (!spage->page) {
1536                         spin_lock(&sdev->stat_lock);
1537                         sdev->stat.malloc_errors++;
1538                         spin_unlock(&sdev->stat_lock);
1539                         while (index > 0) {
1540                                 index--;
1541                                 __free_page(sblock->pagev[index].page);
1542                         }
1543                         kfree(sblock);
1544                         return -ENOMEM;
1545                 }
1546                 spage->sblock = sblock;
1547                 spage->bdev = sdev->dev->bdev;
1548                 spage->flags = flags;
1549                 spage->generation = gen;
1550                 spage->logical = logical;
1551                 spage->physical = physical;
1552                 spage->mirror_num = mirror_num;
1553                 if (csum) {
1554                         spage->have_csum = 1;
1555                         memcpy(spage->csum, csum, sdev->csum_size);
1556                 } else {
1557                         spage->have_csum = 0;
1558                 }
1559                 sblock->page_count++;
1560                 len -= l;
1561                 logical += l;
1562                 physical += l;
1563         }
1564
1565         BUG_ON(sblock->page_count == 0);
1566         for (index = 0; index < sblock->page_count; index++) {
1567                 struct scrub_page *spage = sblock->pagev + index;
1568                 int ret;
1569
1570                 ret = scrub_add_page_to_bio(sdev, spage);
1571                 if (ret) {
1572                         scrub_block_put(sblock);
1573                         return ret;
1574                 }
1575         }
1576
1577         if (force)
1578                 scrub_submit(sdev);
1579
1580         /* last one frees, either here or in bio completion for last page */
1581         scrub_block_put(sblock);
1582         return 0;
1583 }
1584
1585 static void scrub_bio_end_io(struct bio *bio, int err)
1586 {
1587         struct scrub_bio *sbio = bio->bi_private;
1588         struct scrub_dev *sdev = sbio->sdev;
1589         struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
1590
1591         sbio->err = err;
1592         sbio->bio = bio;
1593
1594         btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
1595 }
1596
1597 static void scrub_bio_end_io_worker(struct btrfs_work *work)
1598 {
1599         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1600         struct scrub_dev *sdev = sbio->sdev;
1601         int i;
1602
1603         BUG_ON(sbio->page_count > SCRUB_PAGES_PER_BIO);
1604         if (sbio->err) {
1605                 for (i = 0; i < sbio->page_count; i++) {
1606                         struct scrub_page *spage = sbio->pagev[i];
1607
1608                         spage->io_error = 1;
1609                         spage->sblock->no_io_error_seen = 0;
1610                 }
1611         }
1612
1613         /* now complete the scrub_block items that have all pages completed */
1614         for (i = 0; i < sbio->page_count; i++) {
1615                 struct scrub_page *spage = sbio->pagev[i];
1616                 struct scrub_block *sblock = spage->sblock;
1617
1618                 if (atomic_dec_and_test(&sblock->outstanding_pages))
1619                         scrub_block_complete(sblock);
1620                 scrub_block_put(sblock);
1621         }
1622
1623         if (sbio->err) {
1624                 /* what is this good for??? */
1625                 sbio->bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1626                 sbio->bio->bi_flags |= 1 << BIO_UPTODATE;
1627                 sbio->bio->bi_phys_segments = 0;
1628                 sbio->bio->bi_idx = 0;
1629
1630                 for (i = 0; i < sbio->page_count; i++) {
1631                         struct bio_vec *bi;
1632                         bi = &sbio->bio->bi_io_vec[i];
1633                         bi->bv_offset = 0;
1634                         bi->bv_len = PAGE_SIZE;
1635                 }
1636         }
1637
1638         bio_put(sbio->bio);
1639         sbio->bio = NULL;
1640         spin_lock(&sdev->list_lock);
1641         sbio->next_free = sdev->first_free;
1642         sdev->first_free = sbio->index;
1643         spin_unlock(&sdev->list_lock);
1644         atomic_dec(&sdev->in_flight);
1645         wake_up(&sdev->list_wait);
1646 }
1647
1648 static void scrub_block_complete(struct scrub_block *sblock)
1649 {
1650         if (!sblock->no_io_error_seen)
1651                 scrub_handle_errored_block(sblock);
1652         else
1653                 scrub_checksum(sblock);
1654 }
1655
1656 static int scrub_find_csum(struct scrub_dev *sdev, u64 logical, u64 len,
1657                            u8 *csum)
1658 {
1659         struct btrfs_ordered_sum *sum = NULL;
1660         int ret = 0;
1661         unsigned long i;
1662         unsigned long num_sectors;
1663
1664         while (!list_empty(&sdev->csum_list)) {
1665                 sum = list_first_entry(&sdev->csum_list,
1666                                        struct btrfs_ordered_sum, list);
1667                 if (sum->bytenr > logical)
1668                         return 0;
1669                 if (sum->bytenr + sum->len > logical)
1670                         break;
1671
1672                 ++sdev->stat.csum_discards;
1673                 list_del(&sum->list);
1674                 kfree(sum);
1675                 sum = NULL;
1676         }
1677         if (!sum)
1678                 return 0;
1679
1680         num_sectors = sum->len / sdev->sectorsize;
1681         for (i = 0; i < num_sectors; ++i) {
1682                 if (sum->sums[i].bytenr == logical) {
1683                         memcpy(csum, &sum->sums[i].sum, sdev->csum_size);
1684                         ret = 1;
1685                         break;
1686                 }
1687         }
1688         if (ret && i == num_sectors - 1) {
1689                 list_del(&sum->list);
1690                 kfree(sum);
1691         }
1692         return ret;
1693 }
1694
1695 /* scrub extent tries to collect up to 64 kB for each bio */
1696 static int scrub_extent(struct scrub_dev *sdev, u64 logical, u64 len,
1697                         u64 physical, u64 flags, u64 gen, int mirror_num)
1698 {
1699         int ret;
1700         u8 csum[BTRFS_CSUM_SIZE];
1701         u32 blocksize;
1702
1703         if (flags & BTRFS_EXTENT_FLAG_DATA) {
1704                 blocksize = sdev->sectorsize;
1705                 spin_lock(&sdev->stat_lock);
1706                 sdev->stat.data_extents_scrubbed++;
1707                 sdev->stat.data_bytes_scrubbed += len;
1708                 spin_unlock(&sdev->stat_lock);
1709         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1710                 BUG_ON(sdev->nodesize != sdev->leafsize);
1711                 blocksize = sdev->nodesize;
1712                 spin_lock(&sdev->stat_lock);
1713                 sdev->stat.tree_extents_scrubbed++;
1714                 sdev->stat.tree_bytes_scrubbed += len;
1715                 spin_unlock(&sdev->stat_lock);
1716         } else {
1717                 blocksize = sdev->sectorsize;
1718                 BUG_ON(1);
1719         }
1720
1721         while (len) {
1722                 u64 l = min_t(u64, len, blocksize);
1723                 int have_csum = 0;
1724
1725                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
1726                         /* push csums to sbio */
1727                         have_csum = scrub_find_csum(sdev, logical, l, csum);
1728                         if (have_csum == 0)
1729                                 ++sdev->stat.no_csum;
1730                 }
1731                 ret = scrub_pages(sdev, logical, l, physical, flags, gen,
1732                                   mirror_num, have_csum ? csum : NULL, 0);
1733                 if (ret)
1734                         return ret;
1735                 len -= l;
1736                 logical += l;
1737                 physical += l;
1738         }
1739         return 0;
1740 }
1741
1742 static noinline_for_stack int scrub_stripe(struct scrub_dev *sdev,
1743         struct map_lookup *map, int num, u64 base, u64 length)
1744 {
1745         struct btrfs_path *path;
1746         struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
1747         struct btrfs_root *root = fs_info->extent_root;
1748         struct btrfs_root *csum_root = fs_info->csum_root;
1749         struct btrfs_extent_item *extent;
1750         struct blk_plug plug;
1751         u64 flags;
1752         int ret;
1753         int slot;
1754         int i;
1755         u64 nstripes;
1756         struct extent_buffer *l;
1757         struct btrfs_key key;
1758         u64 physical;
1759         u64 logical;
1760         u64 generation;
1761         int mirror_num;
1762         struct reada_control *reada1;
1763         struct reada_control *reada2;
1764         struct btrfs_key key_start;
1765         struct btrfs_key key_end;
1766
1767         u64 increment = map->stripe_len;
1768         u64 offset;
1769
1770         nstripes = length;
1771         offset = 0;
1772         do_div(nstripes, map->stripe_len);
1773         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
1774                 offset = map->stripe_len * num;
1775                 increment = map->stripe_len * map->num_stripes;
1776                 mirror_num = 1;
1777         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
1778                 int factor = map->num_stripes / map->sub_stripes;
1779                 offset = map->stripe_len * (num / map->sub_stripes);
1780                 increment = map->stripe_len * factor;
1781                 mirror_num = num % map->sub_stripes + 1;
1782         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
1783                 increment = map->stripe_len;
1784                 mirror_num = num % map->num_stripes + 1;
1785         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
1786                 increment = map->stripe_len;
1787                 mirror_num = num % map->num_stripes + 1;
1788         } else {
1789                 increment = map->stripe_len;
1790                 mirror_num = 1;
1791         }
1792
1793         path = btrfs_alloc_path();
1794         if (!path)
1795                 return -ENOMEM;
1796
1797         /*
1798          * work on commit root. The related disk blocks are static as
1799          * long as COW is applied. This means, it is save to rewrite
1800          * them to repair disk errors without any race conditions
1801          */
1802         path->search_commit_root = 1;
1803         path->skip_locking = 1;
1804
1805         /*
1806          * trigger the readahead for extent tree csum tree and wait for
1807          * completion. During readahead, the scrub is officially paused
1808          * to not hold off transaction commits
1809          */
1810         logical = base + offset;
1811
1812         wait_event(sdev->list_wait,
1813                    atomic_read(&sdev->in_flight) == 0);
1814         atomic_inc(&fs_info->scrubs_paused);
1815         wake_up(&fs_info->scrub_pause_wait);
1816
1817         /* FIXME it might be better to start readahead at commit root */
1818         key_start.objectid = logical;
1819         key_start.type = BTRFS_EXTENT_ITEM_KEY;
1820         key_start.offset = (u64)0;
1821         key_end.objectid = base + offset + nstripes * increment;
1822         key_end.type = BTRFS_EXTENT_ITEM_KEY;
1823         key_end.offset = (u64)0;
1824         reada1 = btrfs_reada_add(root, &key_start, &key_end);
1825
1826         key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
1827         key_start.type = BTRFS_EXTENT_CSUM_KEY;
1828         key_start.offset = logical;
1829         key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
1830         key_end.type = BTRFS_EXTENT_CSUM_KEY;
1831         key_end.offset = base + offset + nstripes * increment;
1832         reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
1833
1834         if (!IS_ERR(reada1))
1835                 btrfs_reada_wait(reada1);
1836         if (!IS_ERR(reada2))
1837                 btrfs_reada_wait(reada2);
1838
1839         mutex_lock(&fs_info->scrub_lock);
1840         while (atomic_read(&fs_info->scrub_pause_req)) {
1841                 mutex_unlock(&fs_info->scrub_lock);
1842                 wait_event(fs_info->scrub_pause_wait,
1843                    atomic_read(&fs_info->scrub_pause_req) == 0);
1844                 mutex_lock(&fs_info->scrub_lock);
1845         }
1846         atomic_dec(&fs_info->scrubs_paused);
1847         mutex_unlock(&fs_info->scrub_lock);
1848         wake_up(&fs_info->scrub_pause_wait);
1849
1850         /*
1851          * collect all data csums for the stripe to avoid seeking during
1852          * the scrub. This might currently (crc32) end up to be about 1MB
1853          */
1854         blk_start_plug(&plug);
1855
1856         /*
1857          * now find all extents for each stripe and scrub them
1858          */
1859         logical = base + offset;
1860         physical = map->stripes[num].physical;
1861         ret = 0;
1862         for (i = 0; i < nstripes; ++i) {
1863                 /*
1864                  * canceled?
1865                  */
1866                 if (atomic_read(&fs_info->scrub_cancel_req) ||
1867                     atomic_read(&sdev->cancel_req)) {
1868                         ret = -ECANCELED;
1869                         goto out;
1870                 }
1871                 /*
1872                  * check to see if we have to pause
1873                  */
1874                 if (atomic_read(&fs_info->scrub_pause_req)) {
1875                         /* push queued extents */
1876                         scrub_submit(sdev);
1877                         wait_event(sdev->list_wait,
1878                                    atomic_read(&sdev->in_flight) == 0);
1879                         atomic_inc(&fs_info->scrubs_paused);
1880                         wake_up(&fs_info->scrub_pause_wait);
1881                         mutex_lock(&fs_info->scrub_lock);
1882                         while (atomic_read(&fs_info->scrub_pause_req)) {
1883                                 mutex_unlock(&fs_info->scrub_lock);
1884                                 wait_event(fs_info->scrub_pause_wait,
1885                                    atomic_read(&fs_info->scrub_pause_req) == 0);
1886                                 mutex_lock(&fs_info->scrub_lock);
1887                         }
1888                         atomic_dec(&fs_info->scrubs_paused);
1889                         mutex_unlock(&fs_info->scrub_lock);
1890                         wake_up(&fs_info->scrub_pause_wait);
1891                 }
1892
1893                 ret = btrfs_lookup_csums_range(csum_root, logical,
1894                                                logical + map->stripe_len - 1,
1895                                                &sdev->csum_list, 1);
1896                 if (ret)
1897                         goto out;
1898
1899                 key.objectid = logical;
1900                 key.type = BTRFS_EXTENT_ITEM_KEY;
1901                 key.offset = (u64)0;
1902
1903                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1904                 if (ret < 0)
1905                         goto out;
1906                 if (ret > 0) {
1907                         ret = btrfs_previous_item(root, path, 0,
1908                                                   BTRFS_EXTENT_ITEM_KEY);
1909                         if (ret < 0)
1910                                 goto out;
1911                         if (ret > 0) {
1912                                 /* there's no smaller item, so stick with the
1913                                  * larger one */
1914                                 btrfs_release_path(path);
1915                                 ret = btrfs_search_slot(NULL, root, &key,
1916                                                         path, 0, 0);
1917                                 if (ret < 0)
1918                                         goto out;
1919                         }
1920                 }
1921
1922                 while (1) {
1923                         l = path->nodes[0];
1924                         slot = path->slots[0];
1925                         if (slot >= btrfs_header_nritems(l)) {
1926                                 ret = btrfs_next_leaf(root, path);
1927                                 if (ret == 0)
1928                                         continue;
1929                                 if (ret < 0)
1930                                         goto out;
1931
1932                                 break;
1933                         }
1934                         btrfs_item_key_to_cpu(l, &key, slot);
1935
1936                         if (key.objectid + key.offset <= logical)
1937                                 goto next;
1938
1939                         if (key.objectid >= logical + map->stripe_len)
1940                                 break;
1941
1942                         if (btrfs_key_type(&key) != BTRFS_EXTENT_ITEM_KEY)
1943                                 goto next;
1944
1945                         extent = btrfs_item_ptr(l, slot,
1946                                                 struct btrfs_extent_item);
1947                         flags = btrfs_extent_flags(l, extent);
1948                         generation = btrfs_extent_generation(l, extent);
1949
1950                         if (key.objectid < logical &&
1951                             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
1952                                 printk(KERN_ERR
1953                                        "btrfs scrub: tree block %llu spanning "
1954                                        "stripes, ignored. logical=%llu\n",
1955                                        (unsigned long long)key.objectid,
1956                                        (unsigned long long)logical);
1957                                 goto next;
1958                         }
1959
1960                         /*
1961                          * trim extent to this stripe
1962                          */
1963                         if (key.objectid < logical) {
1964                                 key.offset -= logical - key.objectid;
1965                                 key.objectid = logical;
1966                         }
1967                         if (key.objectid + key.offset >
1968                             logical + map->stripe_len) {
1969                                 key.offset = logical + map->stripe_len -
1970                                              key.objectid;
1971                         }
1972
1973                         ret = scrub_extent(sdev, key.objectid, key.offset,
1974                                            key.objectid - logical + physical,
1975                                            flags, generation, mirror_num);
1976                         if (ret)
1977                                 goto out;
1978
1979 next:
1980                         path->slots[0]++;
1981                 }
1982                 btrfs_release_path(path);
1983                 logical += increment;
1984                 physical += map->stripe_len;
1985                 spin_lock(&sdev->stat_lock);
1986                 sdev->stat.last_physical = physical;
1987                 spin_unlock(&sdev->stat_lock);
1988         }
1989         /* push queued extents */
1990         scrub_submit(sdev);
1991
1992 out:
1993         blk_finish_plug(&plug);
1994         btrfs_free_path(path);
1995         return ret < 0 ? ret : 0;
1996 }
1997
1998 static noinline_for_stack int scrub_chunk(struct scrub_dev *sdev,
1999         u64 chunk_tree, u64 chunk_objectid, u64 chunk_offset, u64 length,
2000         u64 dev_offset)
2001 {
2002         struct btrfs_mapping_tree *map_tree =
2003                 &sdev->dev->dev_root->fs_info->mapping_tree;
2004         struct map_lookup *map;
2005         struct extent_map *em;
2006         int i;
2007         int ret = -EINVAL;
2008
2009         read_lock(&map_tree->map_tree.lock);
2010         em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2011         read_unlock(&map_tree->map_tree.lock);
2012
2013         if (!em)
2014                 return -EINVAL;
2015
2016         map = (struct map_lookup *)em->bdev;
2017         if (em->start != chunk_offset)
2018                 goto out;
2019
2020         if (em->len < length)
2021                 goto out;
2022
2023         for (i = 0; i < map->num_stripes; ++i) {
2024                 if (map->stripes[i].dev == sdev->dev &&
2025                     map->stripes[i].physical == dev_offset) {
2026                         ret = scrub_stripe(sdev, map, i, chunk_offset, length);
2027                         if (ret)
2028                                 goto out;
2029                 }
2030         }
2031 out:
2032         free_extent_map(em);
2033
2034         return ret;
2035 }
2036
2037 static noinline_for_stack
2038 int scrub_enumerate_chunks(struct scrub_dev *sdev, u64 start, u64 end)
2039 {
2040         struct btrfs_dev_extent *dev_extent = NULL;
2041         struct btrfs_path *path;
2042         struct btrfs_root *root = sdev->dev->dev_root;
2043         struct btrfs_fs_info *fs_info = root->fs_info;
2044         u64 length;
2045         u64 chunk_tree;
2046         u64 chunk_objectid;
2047         u64 chunk_offset;
2048         int ret;
2049         int slot;
2050         struct extent_buffer *l;
2051         struct btrfs_key key;
2052         struct btrfs_key found_key;
2053         struct btrfs_block_group_cache *cache;
2054
2055         path = btrfs_alloc_path();
2056         if (!path)
2057                 return -ENOMEM;
2058
2059         path->reada = 2;
2060         path->search_commit_root = 1;
2061         path->skip_locking = 1;
2062
2063         key.objectid = sdev->dev->devid;
2064         key.offset = 0ull;
2065         key.type = BTRFS_DEV_EXTENT_KEY;
2066
2067
2068         while (1) {
2069                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2070                 if (ret < 0)
2071                         break;
2072                 if (ret > 0) {
2073                         if (path->slots[0] >=
2074                             btrfs_header_nritems(path->nodes[0])) {
2075                                 ret = btrfs_next_leaf(root, path);
2076                                 if (ret)
2077                                         break;
2078                         }
2079                 }
2080
2081                 l = path->nodes[0];
2082                 slot = path->slots[0];
2083
2084                 btrfs_item_key_to_cpu(l, &found_key, slot);
2085
2086                 if (found_key.objectid != sdev->dev->devid)
2087                         break;
2088
2089                 if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
2090                         break;
2091
2092                 if (found_key.offset >= end)
2093                         break;
2094
2095                 if (found_key.offset < key.offset)
2096                         break;
2097
2098                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2099                 length = btrfs_dev_extent_length(l, dev_extent);
2100
2101                 if (found_key.offset + length <= start) {
2102                         key.offset = found_key.offset + length;
2103                         btrfs_release_path(path);
2104                         continue;
2105                 }
2106
2107                 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2108                 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2109                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2110
2111                 /*
2112                  * get a reference on the corresponding block group to prevent
2113                  * the chunk from going away while we scrub it
2114                  */
2115                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2116                 if (!cache) {
2117                         ret = -ENOENT;
2118                         break;
2119                 }
2120                 ret = scrub_chunk(sdev, chunk_tree, chunk_objectid,
2121                                   chunk_offset, length, found_key.offset);
2122                 btrfs_put_block_group(cache);
2123                 if (ret)
2124                         break;
2125
2126                 key.offset = found_key.offset + length;
2127                 btrfs_release_path(path);
2128         }
2129
2130         btrfs_free_path(path);
2131
2132         /*
2133          * ret can still be 1 from search_slot or next_leaf,
2134          * that's not an error
2135          */
2136         return ret < 0 ? ret : 0;
2137 }
2138
2139 static noinline_for_stack int scrub_supers(struct scrub_dev *sdev)
2140 {
2141         int     i;
2142         u64     bytenr;
2143         u64     gen;
2144         int     ret;
2145         struct btrfs_device *device = sdev->dev;
2146         struct btrfs_root *root = device->dev_root;
2147
2148         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
2149                 return -EIO;
2150
2151         gen = root->fs_info->last_trans_committed;
2152
2153         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2154                 bytenr = btrfs_sb_offset(i);
2155                 if (bytenr + BTRFS_SUPER_INFO_SIZE > device->total_bytes)
2156                         break;
2157
2158                 ret = scrub_pages(sdev, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2159                                      BTRFS_EXTENT_FLAG_SUPER, gen, i, NULL, 1);
2160                 if (ret)
2161                         return ret;
2162         }
2163         wait_event(sdev->list_wait, atomic_read(&sdev->in_flight) == 0);
2164
2165         return 0;
2166 }
2167
2168 /*
2169  * get a reference count on fs_info->scrub_workers. start worker if necessary
2170  */
2171 static noinline_for_stack int scrub_workers_get(struct btrfs_root *root)
2172 {
2173         struct btrfs_fs_info *fs_info = root->fs_info;
2174         int ret = 0;
2175
2176         mutex_lock(&fs_info->scrub_lock);
2177         if (fs_info->scrub_workers_refcnt == 0) {
2178                 btrfs_init_workers(&fs_info->scrub_workers, "scrub",
2179                            fs_info->thread_pool_size, &fs_info->generic_worker);
2180                 fs_info->scrub_workers.idle_thresh = 4;
2181                 ret = btrfs_start_workers(&fs_info->scrub_workers);
2182                 if (ret)
2183                         goto out;
2184         }
2185         ++fs_info->scrub_workers_refcnt;
2186 out:
2187         mutex_unlock(&fs_info->scrub_lock);
2188
2189         return ret;
2190 }
2191
2192 static noinline_for_stack void scrub_workers_put(struct btrfs_root *root)
2193 {
2194         struct btrfs_fs_info *fs_info = root->fs_info;
2195
2196         mutex_lock(&fs_info->scrub_lock);
2197         if (--fs_info->scrub_workers_refcnt == 0)
2198                 btrfs_stop_workers(&fs_info->scrub_workers);
2199         WARN_ON(fs_info->scrub_workers_refcnt < 0);
2200         mutex_unlock(&fs_info->scrub_lock);
2201 }
2202
2203
2204 int btrfs_scrub_dev(struct btrfs_root *root, u64 devid, u64 start, u64 end,
2205                     struct btrfs_scrub_progress *progress, int readonly)
2206 {
2207         struct scrub_dev *sdev;
2208         struct btrfs_fs_info *fs_info = root->fs_info;
2209         int ret;
2210         struct btrfs_device *dev;
2211
2212         if (btrfs_fs_closing(root->fs_info))
2213                 return -EINVAL;
2214
2215         /*
2216          * check some assumptions
2217          */
2218         if (root->nodesize != root->leafsize) {
2219                 printk(KERN_ERR
2220                        "btrfs_scrub: size assumption nodesize == leafsize (%d == %d) fails\n",
2221                        root->nodesize, root->leafsize);
2222                 return -EINVAL;
2223         }
2224
2225         if (root->nodesize > BTRFS_STRIPE_LEN) {
2226                 /*
2227                  * in this case scrub is unable to calculate the checksum
2228                  * the way scrub is implemented. Do not handle this
2229                  * situation at all because it won't ever happen.
2230                  */
2231                 printk(KERN_ERR
2232                        "btrfs_scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails\n",
2233                        root->nodesize, BTRFS_STRIPE_LEN);
2234                 return -EINVAL;
2235         }
2236
2237         if (root->sectorsize != PAGE_SIZE) {
2238                 /* not supported for data w/o checksums */
2239                 printk(KERN_ERR
2240                        "btrfs_scrub: size assumption sectorsize != PAGE_SIZE (%d != %lld) fails\n",
2241                        root->sectorsize, (unsigned long long)PAGE_SIZE);
2242                 return -EINVAL;
2243         }
2244
2245         ret = scrub_workers_get(root);
2246         if (ret)
2247                 return ret;
2248
2249         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
2250         dev = btrfs_find_device(root, devid, NULL, NULL);
2251         if (!dev || dev->missing) {
2252                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2253                 scrub_workers_put(root);
2254                 return -ENODEV;
2255         }
2256         mutex_lock(&fs_info->scrub_lock);
2257
2258         if (!dev->in_fs_metadata) {
2259                 mutex_unlock(&fs_info->scrub_lock);
2260                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2261                 scrub_workers_put(root);
2262                 return -ENODEV;
2263         }
2264
2265         if (dev->scrub_device) {
2266                 mutex_unlock(&fs_info->scrub_lock);
2267                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2268                 scrub_workers_put(root);
2269                 return -EINPROGRESS;
2270         }
2271         sdev = scrub_setup_dev(dev);
2272         if (IS_ERR(sdev)) {
2273                 mutex_unlock(&fs_info->scrub_lock);
2274                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2275                 scrub_workers_put(root);
2276                 return PTR_ERR(sdev);
2277         }
2278         sdev->readonly = readonly;
2279         dev->scrub_device = sdev;
2280
2281         atomic_inc(&fs_info->scrubs_running);
2282         mutex_unlock(&fs_info->scrub_lock);
2283         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2284
2285         down_read(&fs_info->scrub_super_lock);
2286         ret = scrub_supers(sdev);
2287         up_read(&fs_info->scrub_super_lock);
2288
2289         if (!ret)
2290                 ret = scrub_enumerate_chunks(sdev, start, end);
2291
2292         wait_event(sdev->list_wait, atomic_read(&sdev->in_flight) == 0);
2293         atomic_dec(&fs_info->scrubs_running);
2294         wake_up(&fs_info->scrub_pause_wait);
2295
2296         wait_event(sdev->list_wait, atomic_read(&sdev->fixup_cnt) == 0);
2297
2298         if (progress)
2299                 memcpy(progress, &sdev->stat, sizeof(*progress));
2300
2301         mutex_lock(&fs_info->scrub_lock);
2302         dev->scrub_device = NULL;
2303         mutex_unlock(&fs_info->scrub_lock);
2304
2305         scrub_free_dev(sdev);
2306         scrub_workers_put(root);
2307
2308         return ret;
2309 }
2310
2311 void btrfs_scrub_pause(struct btrfs_root *root)
2312 {
2313         struct btrfs_fs_info *fs_info = root->fs_info;
2314
2315         mutex_lock(&fs_info->scrub_lock);
2316         atomic_inc(&fs_info->scrub_pause_req);
2317         while (atomic_read(&fs_info->scrubs_paused) !=
2318                atomic_read(&fs_info->scrubs_running)) {
2319                 mutex_unlock(&fs_info->scrub_lock);
2320                 wait_event(fs_info->scrub_pause_wait,
2321                            atomic_read(&fs_info->scrubs_paused) ==
2322                            atomic_read(&fs_info->scrubs_running));
2323                 mutex_lock(&fs_info->scrub_lock);
2324         }
2325         mutex_unlock(&fs_info->scrub_lock);
2326 }
2327
2328 void btrfs_scrub_continue(struct btrfs_root *root)
2329 {
2330         struct btrfs_fs_info *fs_info = root->fs_info;
2331
2332         atomic_dec(&fs_info->scrub_pause_req);
2333         wake_up(&fs_info->scrub_pause_wait);
2334 }
2335
2336 void btrfs_scrub_pause_super(struct btrfs_root *root)
2337 {
2338         down_write(&root->fs_info->scrub_super_lock);
2339 }
2340
2341 void btrfs_scrub_continue_super(struct btrfs_root *root)
2342 {
2343         up_write(&root->fs_info->scrub_super_lock);
2344 }
2345
2346 int __btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
2347 {
2348
2349         mutex_lock(&fs_info->scrub_lock);
2350         if (!atomic_read(&fs_info->scrubs_running)) {
2351                 mutex_unlock(&fs_info->scrub_lock);
2352                 return -ENOTCONN;
2353         }
2354
2355         atomic_inc(&fs_info->scrub_cancel_req);
2356         while (atomic_read(&fs_info->scrubs_running)) {
2357                 mutex_unlock(&fs_info->scrub_lock);
2358                 wait_event(fs_info->scrub_pause_wait,
2359                            atomic_read(&fs_info->scrubs_running) == 0);
2360                 mutex_lock(&fs_info->scrub_lock);
2361         }
2362         atomic_dec(&fs_info->scrub_cancel_req);
2363         mutex_unlock(&fs_info->scrub_lock);
2364
2365         return 0;
2366 }
2367
2368 int btrfs_scrub_cancel(struct btrfs_root *root)
2369 {
2370         return __btrfs_scrub_cancel(root->fs_info);
2371 }
2372
2373 int btrfs_scrub_cancel_dev(struct btrfs_root *root, struct btrfs_device *dev)
2374 {
2375         struct btrfs_fs_info *fs_info = root->fs_info;
2376         struct scrub_dev *sdev;
2377
2378         mutex_lock(&fs_info->scrub_lock);
2379         sdev = dev->scrub_device;
2380         if (!sdev) {
2381                 mutex_unlock(&fs_info->scrub_lock);
2382                 return -ENOTCONN;
2383         }
2384         atomic_inc(&sdev->cancel_req);
2385         while (dev->scrub_device) {
2386                 mutex_unlock(&fs_info->scrub_lock);
2387                 wait_event(fs_info->scrub_pause_wait,
2388                            dev->scrub_device == NULL);
2389                 mutex_lock(&fs_info->scrub_lock);
2390         }
2391         mutex_unlock(&fs_info->scrub_lock);
2392
2393         return 0;
2394 }
2395
2396 int btrfs_scrub_cancel_devid(struct btrfs_root *root, u64 devid)
2397 {
2398         struct btrfs_fs_info *fs_info = root->fs_info;
2399         struct btrfs_device *dev;
2400         int ret;
2401
2402         /*
2403          * we have to hold the device_list_mutex here so the device
2404          * does not go away in cancel_dev. FIXME: find a better solution
2405          */
2406         mutex_lock(&fs_info->fs_devices->device_list_mutex);
2407         dev = btrfs_find_device(root, devid, NULL, NULL);
2408         if (!dev) {
2409                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2410                 return -ENODEV;
2411         }
2412         ret = btrfs_scrub_cancel_dev(root, dev);
2413         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2414
2415         return ret;
2416 }
2417
2418 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
2419                          struct btrfs_scrub_progress *progress)
2420 {
2421         struct btrfs_device *dev;
2422         struct scrub_dev *sdev = NULL;
2423
2424         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
2425         dev = btrfs_find_device(root, devid, NULL, NULL);
2426         if (dev)
2427                 sdev = dev->scrub_device;
2428         if (sdev)
2429                 memcpy(progress, &sdev->stat, sizeof(*progress));
2430         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2431
2432         return dev ? (sdev ? 0 : -ENOTCONN) : -ENODEV;
2433 }