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_dev;
  44
  45 #define SCRUB_PAGES_PER_BIO     16      /* 64k per bio */
  46 #define SCRUB_BIOS_PER_DEV      16      /* 1 MB per device in flight */
  47
  48 struct scrub_page {
  49         u64                     flags;  /* extent flags */
  50         u64                     generation;
  51         int                     mirror_num;
  52         int                     have_csum;
  53         u8                      csum[BTRFS_CSUM_SIZE];
  54 };
  55
  56 struct scrub_bio {
  57         int                     index;
  58         struct scrub_dev        *sdev;
  59         struct bio              *bio;
  60         int                     err;
  61         u64                     logical;
  62         u64                     physical;
  63         struct scrub_page       spag[SCRUB_PAGES_PER_BIO];
  64         u64                     count;
  65         int                     next_free;
  66         struct btrfs_work       work;
  67 };
  68
  69 struct scrub_dev {
  70         struct scrub_bio        *bios[SCRUB_BIOS_PER_DEV];
  71         struct btrfs_device     *dev;
  72         int                     first_free;
  73         int                     curr;
  74         atomic_t                in_flight;
  75         atomic_t                fixup_cnt;
  76         spinlock_t              list_lock;
  77         wait_queue_head_t       list_wait;
  78         u16                     csum_size;
  79         struct list_head        csum_list;
  80         atomic_t                cancel_req;
  81         int                     readonly;
  82         /*
  83          * statistics
  84          */
  85         struct btrfs_scrub_progress stat;
  86         spinlock_t              stat_lock;
  87 };
  88
  89 struct scrub_fixup_nodatasum {
  90         struct scrub_dev        *sdev;
  91         u64                     logical;
  92         struct btrfs_root       *root;
  93         struct btrfs_work       work;
  94         int                     mirror_num;
  95 };
  96
  97 struct scrub_warning {
  98         struct btrfs_path       *path;
  99         u64                     extent_item_size;
 100         char                    *scratch_buf;
 101         char                    *msg_buf;
 102         const char              *errstr;
 103         sector_t                sector;
 104         u64                     logical;
 105         struct btrfs_device     *dev;
 106         int                     msg_bufsize;
 107         int                     scratch_bufsize;
 108 };
 109
 110 static void scrub_bio_end_io(struct bio *bio, int err);
 111 static void scrub_checksum(struct btrfs_work *work);
 112 static int scrub_checksum_data(struct scrub_dev *sdev,
 113                                struct scrub_page *spag, void *buffer);
 114 static int scrub_checksum_tree_block(struct scrub_dev *sdev,
 115                                      struct scrub_page *spag, u64 logical,
 116                                      void *buffer);
 117 static int scrub_checksum_super(struct scrub_bio *sbio, void *buffer);
 118 static int scrub_fixup_check(struct scrub_bio *sbio, int ix);
 119 static void scrub_fixup_end_io(struct bio *bio, int err);
 120 static int scrub_fixup_io(int rw, struct block_device *bdev, sector_t sector,
 121                           struct page *page);
 122 static void scrub_fixup(struct scrub_bio *sbio, int ix);
 123
 124
 125 static void scrub_free_csums(struct scrub_dev *sdev)
 126 {
 127         while (!list_empty(&sdev->csum_list)) {
 128                 struct btrfs_ordered_sum *sum;
 129                 sum = list_first_entry(&sdev->csum_list,
 130                                        struct btrfs_ordered_sum, list);
 131                 list_del(&sum->list);
 132                 kfree(sum);
 133         }
 134 }
 135
 136 static void scrub_free_bio(struct bio *bio)
 137 {
 138         int i;
 139         struct page *last_page = NULL;
 140
 141         if (!bio)
 142                 return;
 143
 144         for (i = 0; i < bio->bi_vcnt; ++i) {
 145                 if (bio->bi_io_vec[i].bv_page == last_page)
 146                         continue;
 147                 last_page = bio->bi_io_vec[i].bv_page;
 148                 __free_page(last_page);
 149         }
 150         bio_put(bio);
 151 }
 152
 153 static noinline_for_stack void scrub_free_dev(struct scrub_dev *sdev)
 154 {
 155         int i;
 156
 157         if (!sdev)
 158                 return;
 159
 160         for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
 161                 struct scrub_bio *sbio = sdev->bios[i];
 162
 163                 if (!sbio)
 164                         break;
 165
 166                 scrub_free_bio(sbio->bio);
 167                 kfree(sbio);
 168         }
 169
 170         scrub_free_csums(sdev);
 171         kfree(sdev);
 172 }
 173
 174 static noinline_for_stack
 175 struct scrub_dev *scrub_setup_dev(struct btrfs_device *dev)
 176 {
 177         struct scrub_dev *sdev;
 178         int             i;
 179         struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
 180
 181         sdev = kzalloc(sizeof(*sdev), GFP_NOFS);
 182         if (!sdev)
 183                 goto nomem;
 184         sdev->dev = dev;
 185         for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
 186                 struct scrub_bio *sbio;
 187
 188                 sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
 189                 if (!sbio)
 190                         goto nomem;
 191                 sdev->bios[i] = sbio;
 192
 193                 sbio->index = i;
 194                 sbio->sdev = sdev;
 195                 sbio->count = 0;
 196                 sbio->work.func = scrub_checksum;
 197
 198                 if (i != SCRUB_BIOS_PER_DEV-1)
 199                         sdev->bios[i]->next_free = i + 1;
 200                 else
 201                         sdev->bios[i]->next_free = -1;
 202         }
 203         sdev->first_free = 0;
 204         sdev->curr = -1;
 205         atomic_set(&sdev->in_flight, 0);
 206         atomic_set(&sdev->fixup_cnt, 0);
 207         atomic_set(&sdev->cancel_req, 0);
 208         sdev->csum_size = btrfs_super_csum_size(fs_info->super_copy);
 209         INIT_LIST_HEAD(&sdev->csum_list);
 210
 211         spin_lock_init(&sdev->list_lock);
 212         spin_lock_init(&sdev->stat_lock);
 213         init_waitqueue_head(&sdev->list_wait);
 214         return sdev;
 215
 216 nomem:
 217         scrub_free_dev(sdev);
 218         return ERR_PTR(-ENOMEM);
 219 }
 220
 221 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root, void *ctx)
 222 {
 223         u64 isize;
 224         u32 nlink;
 225         int ret;
 226         int i;
 227         struct extent_buffer *eb;
 228         struct btrfs_inode_item *inode_item;
 229         struct scrub_warning *swarn = ctx;
 230         struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
 231         struct inode_fs_paths *ipath = NULL;
 232         struct btrfs_root *local_root;
 233         struct btrfs_key root_key;
 234
 235         root_key.objectid = root;
 236         root_key.type = BTRFS_ROOT_ITEM_KEY;
 237         root_key.offset = (u64)-1;
 238         local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
 239         if (IS_ERR(local_root)) {
 240                 ret = PTR_ERR(local_root);
 241                 goto err;
 242         }
 243
 244         ret = inode_item_info(inum, 0, local_root, swarn->path);
 245         if (ret) {
 246                 btrfs_release_path(swarn->path);
 247                 goto err;
 248         }
 249
 250         eb = swarn->path->nodes[0];
 251         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 252                                         struct btrfs_inode_item);
 253         isize = btrfs_inode_size(eb, inode_item);
 254         nlink = btrfs_inode_nlink(eb, inode_item);
 255         btrfs_release_path(swarn->path);
 256
 257         ipath = init_ipath(4096, local_root, swarn->path);
 258         if (IS_ERR(ipath)) {
 259                 ret = PTR_ERR(ipath);
 260                 ipath = NULL;
 261                 goto err;
 262         }
 263         ret = paths_from_inode(inum, ipath);
 264
 265         if (ret < 0)
 266                 goto err;
 267
 268         /*
 269          * we deliberately ignore the bit ipath might have been too small to
 270          * hold all of the paths here
 271          */
 272         for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 273                 printk(KERN_WARNING "btrfs: %s at logical %llu on dev "
 274                         "%s, sector %llu, root %llu, inode %llu, offset %llu, "
 275                         "length %llu, links %u (path: %s)\n", swarn->errstr,
 276                         swarn->logical, swarn->dev->name,
 277                         (unsigned long long)swarn->sector, root, inum, offset,
 278                         min(isize - offset, (u64)PAGE_SIZE), nlink,
 279                         (char *)(unsigned long)ipath->fspath->val[i]);
 280
 281         free_ipath(ipath);
 282         return 0;
 283
 284 err:
 285         printk(KERN_WARNING "btrfs: %s at logical %llu on dev "
 286                 "%s, sector %llu, root %llu, inode %llu, offset %llu: path "
 287                 "resolving failed with ret=%d\n", swarn->errstr,
 288                 swarn->logical, swarn->dev->name,
 289                 (unsigned long long)swarn->sector, root, inum, offset, ret);
 290
 291         free_ipath(ipath);
 292         return 0;
 293 }
 294
 295 static void scrub_print_warning(const char *errstr, struct scrub_bio *sbio,
 296                                 int ix)
 297 {
 298         struct btrfs_device *dev = sbio->sdev->dev;
 299         struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
 300         struct btrfs_path *path;
 301         struct btrfs_key found_key;
 302         struct extent_buffer *eb;
 303         struct btrfs_extent_item *ei;
 304         struct scrub_warning swarn;
 305         u32 item_size;
 306         int ret;
 307         u64 ref_root;
 308         u8 ref_level;
 309         unsigned long ptr = 0;
 310         const int bufsize = 4096;
 311         u64 extent_item_pos;
 312
 313         path = btrfs_alloc_path();
 314
 315         swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
 316         swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
 317         swarn.sector = (sbio->physical + ix * PAGE_SIZE) >> 9;
 318         swarn.logical = sbio->logical + ix * PAGE_SIZE;
 319         swarn.errstr = errstr;
 320         swarn.dev = dev;
 321         swarn.msg_bufsize = bufsize;
 322         swarn.scratch_bufsize = bufsize;
 323
 324         if (!path || !swarn.scratch_buf || !swarn.msg_buf)
 325                 goto out;
 326
 327         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key);
 328         if (ret < 0)
 329                 goto out;
 330
 331         extent_item_pos = swarn.logical - found_key.objectid;
 332         swarn.extent_item_size = found_key.offset;
 333
 334         eb = path->nodes[0];
 335         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 336         item_size = btrfs_item_size_nr(eb, path->slots[0]);
 337         btrfs_release_path(path);
 338
 339         if (ret & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 340                 do {
 341                         ret = tree_backref_for_extent(&ptr, eb, ei, item_size,
 342                                                         &ref_root, &ref_level);
 343                         printk(KERN_WARNING
 344                                 "btrfs: %s at logical %llu on dev %s, "
 345                                 "sector %llu: metadata %s (level %d) in tree "
 346                                 "%llu\n", errstr, swarn.logical, dev->name,
 347                                 (unsigned long long)swarn.sector,
 348                                 ref_level ? "node" : "leaf",
 349                                 ret < 0 ? -1 : ref_level,
 350                                 ret < 0 ? -1 : ref_root);
 351                 } while (ret != 1);
 352         } else {
 353                 swarn.path = path;
 354                 iterate_extent_inodes(fs_info, path, found_key.objectid,
 355                                         extent_item_pos,
 356                                         scrub_print_warning_inode, &swarn);
 357         }
 358
 359 out:
 360         btrfs_free_path(path);
 361         kfree(swarn.scratch_buf);
 362         kfree(swarn.msg_buf);
 363 }
 364
 365 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *ctx)
 366 {
 367         struct page *page = NULL;
 368         unsigned long index;
 369         struct scrub_fixup_nodatasum *fixup = ctx;
 370         int ret;
 371         int corrected = 0;
 372         struct btrfs_key key;
 373         struct inode *inode = NULL;
 374         u64 end = offset + PAGE_SIZE - 1;
 375         struct btrfs_root *local_root;
 376
 377         key.objectid = root;
 378         key.type = BTRFS_ROOT_ITEM_KEY;
 379         key.offset = (u64)-1;
 380         local_root = btrfs_read_fs_root_no_name(fixup->root->fs_info, &key);
 381         if (IS_ERR(local_root))
 382                 return PTR_ERR(local_root);
 383
 384         key.type = BTRFS_INODE_ITEM_KEY;
 385         key.objectid = inum;
 386         key.offset = 0;
 387         inode = btrfs_iget(fixup->root->fs_info->sb, &key, local_root, NULL);
 388         if (IS_ERR(inode))
 389                 return PTR_ERR(inode);
 390
 391         index = offset >> PAGE_CACHE_SHIFT;
 392
 393         page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
 394         if (!page) {
 395                 ret = -ENOMEM;
 396                 goto out;
 397         }
 398
 399         if (PageUptodate(page)) {
 400                 struct btrfs_mapping_tree *map_tree;
 401                 if (PageDirty(page)) {
 402                         /*
 403                          * we need to write the data to the defect sector. the
 404                          * data that was in that sector is not in memory,
 405                          * because the page was modified. we must not write the
 406                          * modified page to that sector.
 407                          *
 408                          * TODO: what could be done here: wait for the delalloc
 409                          *       runner to write out that page (might involve
 410                          *       COW) and see whether the sector is still
 411                          *       referenced afterwards.
 412                          *
 413                          * For the meantime, we'll treat this error
 414                          * incorrectable, although there is a chance that a
 415                          * later scrub will find the bad sector again and that
 416                          * there's no dirty page in memory, then.
 417                          */
 418                         ret = -EIO;
 419                         goto out;
 420                 }
 421                 map_tree = &BTRFS_I(inode)->root->fs_info->mapping_tree;
 422                 ret = repair_io_failure(map_tree, offset, PAGE_SIZE,
 423                                         fixup->logical, page,
 424                                         fixup->mirror_num);
 425                 unlock_page(page);
 426                 corrected = !ret;
 427         } else {
 428                 /*
 429                  * we need to get good data first. the general readpage path
 430                  * will call repair_io_failure for us, we just have to make
 431                  * sure we read the bad mirror.
 432                  */
 433                 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 434                                         EXTENT_DAMAGED, GFP_NOFS);
 435                 if (ret) {
 436                         /* set_extent_bits should give proper error */
 437                         WARN_ON(ret > 0);
 438                         if (ret > 0)
 439                                 ret = -EFAULT;
 440                         goto out;
 441                 }
 442
 443                 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
 444                                                 btrfs_get_extent,
 445                                                 fixup->mirror_num);
 446                 wait_on_page_locked(page);
 447
 448                 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
 449                                                 end, EXTENT_DAMAGED, 0, NULL);
 450                 if (!corrected)
 451                         clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 452                                                 EXTENT_DAMAGED, GFP_NOFS);
 453         }
 454
 455 out:
 456         if (page)
 457                 put_page(page);
 458         if (inode)
 459                 iput(inode);
 460
 461         if (ret < 0)
 462                 return ret;
 463
 464         if (ret == 0 && corrected) {
 465                 /*
 466                  * we only need to call readpage for one of the inodes belonging
 467                  * to this extent. so make iterate_extent_inodes stop
 468                  */
 469                 return 1;
 470         }
 471
 472         return -EIO;
 473 }
 474
 475 static void scrub_fixup_nodatasum(struct btrfs_work *work)
 476 {
 477         int ret;
 478         struct scrub_fixup_nodatasum *fixup;
 479         struct scrub_dev *sdev;
 480         struct btrfs_trans_handle *trans = NULL;
 481         struct btrfs_fs_info *fs_info;
 482         struct btrfs_path *path;
 483         int uncorrectable = 0;
 484
 485         fixup = container_of(work, struct scrub_fixup_nodatasum, work);
 486         sdev = fixup->sdev;
 487         fs_info = fixup->root->fs_info;
 488
 489         path = btrfs_alloc_path();
 490         if (!path) {
 491                 spin_lock(&sdev->stat_lock);
 492                 ++sdev->stat.malloc_errors;
 493                 spin_unlock(&sdev->stat_lock);
 494                 uncorrectable = 1;
 495                 goto out;
 496         }
 497
 498         trans = btrfs_join_transaction(fixup->root);
 499         if (IS_ERR(trans)) {
 500                 uncorrectable = 1;
 501                 goto out;
 502         }
 503
 504         /*
 505          * the idea is to trigger a regular read through the standard path. we
 506          * read a page from the (failed) logical address by specifying the
 507          * corresponding copynum of the failed sector. thus, that readpage is
 508          * expected to fail.
 509          * that is the point where on-the-fly error correction will kick in
 510          * (once it's finished) and rewrite the failed sector if a good copy
 511          * can be found.
 512          */
 513         ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
 514                                                 path, scrub_fixup_readpage,
 515                                                 fixup);
 516         if (ret < 0) {
 517                 uncorrectable = 1;
 518                 goto out;
 519         }
 520         WARN_ON(ret != 1);
 521
 522         spin_lock(&sdev->stat_lock);
 523         ++sdev->stat.corrected_errors;
 524         spin_unlock(&sdev->stat_lock);
 525
 526 out:
 527         if (trans && !IS_ERR(trans))
 528                 btrfs_end_transaction(trans, fixup->root);
 529         if (uncorrectable) {
 530                 spin_lock(&sdev->stat_lock);
 531                 ++sdev->stat.uncorrectable_errors;
 532                 spin_unlock(&sdev->stat_lock);
 533                 printk_ratelimited(KERN_ERR "btrfs: unable to fixup "
 534                                         "(nodatasum) error at logical %llu\n",
 535                                         fixup->logical);
 536         }
 537
 538         btrfs_free_path(path);
 539         kfree(fixup);
 540
 541         /* see caller why we're pretending to be paused in the scrub counters */
 542         mutex_lock(&fs_info->scrub_lock);
 543         atomic_dec(&fs_info->scrubs_running);
 544         atomic_dec(&fs_info->scrubs_paused);
 545         mutex_unlock(&fs_info->scrub_lock);
 546         atomic_dec(&sdev->fixup_cnt);
 547         wake_up(&fs_info->scrub_pause_wait);
 548         wake_up(&sdev->list_wait);
 549 }
 550
 551 /*
 552  * scrub_recheck_error gets called when either verification of the page
 553  * failed or the bio failed to read, e.g. with EIO. In the latter case,
 554  * recheck_error gets called for every page in the bio, even though only
 555  * one may be bad
 556  */
 557 static int scrub_recheck_error(struct scrub_bio *sbio, int ix)
 558 {
 559         struct scrub_dev *sdev = sbio->sdev;
 560         u64 sector = (sbio->physical + ix * PAGE_SIZE) >> 9;
 561         static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
 562                                         DEFAULT_RATELIMIT_BURST);
 563
 564         if (sbio->err) {
 565                 if (scrub_fixup_io(READ, sbio->sdev->dev->bdev, sector,
 566                                    sbio->bio->bi_io_vec[ix].bv_page) == 0) {
 567                         if (scrub_fixup_check(sbio, ix) == 0)
 568                                 return 0;
 569                 }
 570                 if (__ratelimit(&_rs))
 571                         scrub_print_warning("i/o error", sbio, ix);
 572         } else {
 573                 if (__ratelimit(&_rs))
 574                         scrub_print_warning("checksum error", sbio, ix);
 575         }
 576
 577         spin_lock(&sdev->stat_lock);
 578         ++sdev->stat.read_errors;
 579         spin_unlock(&sdev->stat_lock);
 580
 581         scrub_fixup(sbio, ix);
 582         return 1;
 583 }
 584
 585 static int scrub_fixup_check(struct scrub_bio *sbio, int ix)
 586 {
 587         int ret = 1;
 588         struct page *page;
 589         void *buffer;
 590         u64 flags = sbio->spag[ix].flags;
 591
 592         page = sbio->bio->bi_io_vec[ix].bv_page;
 593         buffer = kmap_atomic(page, KM_USER0);
 594         if (flags & BTRFS_EXTENT_FLAG_DATA) {
 595                 ret = scrub_checksum_data(sbio->sdev,
 596                                           sbio->spag + ix, buffer);
 597         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 598                 ret = scrub_checksum_tree_block(sbio->sdev,
 599                                                 sbio->spag + ix,
 600                                                 sbio->logical + ix * PAGE_SIZE,
 601                                                 buffer);
 602         } else {
 603                 WARN_ON(1);
 604         }
 605         kunmap_atomic(buffer, KM_USER0);
 606
 607         return ret;
 608 }
 609
 610 static void scrub_fixup_end_io(struct bio *bio, int err)
 611 {
 612         complete((struct completion *)bio->bi_private);
 613 }
 614
 615 static void scrub_fixup(struct scrub_bio *sbio, int ix)
 616 {
 617         struct scrub_dev *sdev = sbio->sdev;
 618         struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
 619         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
 620         struct btrfs_bio *bbio = NULL;
 621         struct scrub_fixup_nodatasum *fixup;
 622         u64 logical = sbio->logical + ix * PAGE_SIZE;
 623         u64 length;
 624         int i;
 625         int ret;
 626         DECLARE_COMPLETION_ONSTACK(complete);
 627
 628         if ((sbio->spag[ix].flags & BTRFS_EXTENT_FLAG_DATA) &&
 629             (sbio->spag[ix].have_csum == 0)) {
 630                 fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
 631                 if (!fixup)
 632                         goto uncorrectable;
 633                 fixup->sdev = sdev;
 634                 fixup->logical = logical;
 635                 fixup->root = fs_info->extent_root;
 636                 fixup->mirror_num = sbio->spag[ix].mirror_num;
 637                 /*
 638                  * increment scrubs_running to prevent cancel requests from
 639                  * completing as long as a fixup worker is running. we must also
 640                  * increment scrubs_paused to prevent deadlocking on pause
 641                  * requests used for transactions commits (as the worker uses a
 642                  * transaction context). it is safe to regard the fixup worker
 643                  * as paused for all matters practical. effectively, we only
 644                  * avoid cancellation requests from completing.
 645                  */
 646                 mutex_lock(&fs_info->scrub_lock);
 647                 atomic_inc(&fs_info->scrubs_running);
 648                 atomic_inc(&fs_info->scrubs_paused);
 649                 mutex_unlock(&fs_info->scrub_lock);
 650                 atomic_inc(&sdev->fixup_cnt);
 651                 fixup->work.func = scrub_fixup_nodatasum;
 652                 btrfs_queue_worker(&fs_info->scrub_workers, &fixup->work);
 653                 return;
 654         }
 655
 656         length = PAGE_SIZE;
 657         ret = btrfs_map_block(map_tree, REQ_WRITE, logical, &length,
 658                               &bbio, 0);
 659         if (ret || !bbio || length < PAGE_SIZE) {
 660                 printk(KERN_ERR
 661                        "scrub_fixup: btrfs_map_block failed us for %llu\n",
 662                        (unsigned long long)logical);
 663                 WARN_ON(1);
 664                 kfree(bbio);
 665                 return;
 666         }
 667
 668         if (bbio->num_stripes == 1)
 669                 /* there aren't any replicas */
 670                 goto uncorrectable;
 671
 672         /*
 673          * first find a good copy
 674          */
 675         for (i = 0; i < bbio->num_stripes; ++i) {
 676                 if (i + 1 == sbio->spag[ix].mirror_num)
 677                         continue;
 678
 679                 if (scrub_fixup_io(READ, bbio->stripes[i].dev->bdev,
 680                                    bbio->stripes[i].physical >> 9,
 681                                    sbio->bio->bi_io_vec[ix].bv_page)) {
 682                         /* I/O-error, this is not a good copy */
 683                         continue;
 684                 }
 685
 686                 if (scrub_fixup_check(sbio, ix) == 0)
 687                         break;
 688         }
 689         if (i == bbio->num_stripes)
 690                 goto uncorrectable;
 691
 692         if (!sdev->readonly) {
 693                 /*
 694                  * bi_io_vec[ix].bv_page now contains good data, write it back
 695                  */
 696                 if (scrub_fixup_io(WRITE, sdev->dev->bdev,
 697                                    (sbio->physical + ix * PAGE_SIZE) >> 9,
 698                                    sbio->bio->bi_io_vec[ix].bv_page)) {
 699                         /* I/O-error, writeback failed, give up */
 700                         goto uncorrectable;
 701                 }
 702         }
 703
 704         kfree(bbio);
 705         spin_lock(&sdev->stat_lock);
 706         ++sdev->stat.corrected_errors;
 707         spin_unlock(&sdev->stat_lock);
 708
 709         printk_ratelimited(KERN_ERR "btrfs: fixed up error at logical %llu\n",
 710                                (unsigned long long)logical);
 711         return;
 712
 713 uncorrectable:
 714         kfree(bbio);
 715         spin_lock(&sdev->stat_lock);
 716         ++sdev->stat.uncorrectable_errors;
 717         spin_unlock(&sdev->stat_lock);
 718
 719         printk_ratelimited(KERN_ERR "btrfs: unable to fixup (regular) error at "
 720                                 "logical %llu\n", (unsigned long long)logical);
 721 }
 722
 723 static int scrub_fixup_io(int rw, struct block_device *bdev, sector_t sector,
 724                          struct page *page)
 725 {
 726         struct bio *bio = NULL;
 727         int ret;
 728         DECLARE_COMPLETION_ONSTACK(complete);
 729
 730         bio = bio_alloc(GFP_NOFS, 1);
 731         bio->bi_bdev = bdev;
 732         bio->bi_sector = sector;
 733         bio_add_page(bio, page, PAGE_SIZE, 0);
 734         bio->bi_end_io = scrub_fixup_end_io;
 735         bio->bi_private = &complete;
 736         btrfsic_submit_bio(rw, bio);
 737
 738         /* this will also unplug the queue */
 739         wait_for_completion(&complete);
 740
 741         ret = !test_bit(BIO_UPTODATE, &bio->bi_flags);
 742         bio_put(bio);
 743         return ret;
 744 }
 745
 746 static void scrub_bio_end_io(struct bio *bio, int err)
 747 {
 748         struct scrub_bio *sbio = bio->bi_private;
 749         struct scrub_dev *sdev = sbio->sdev;
 750         struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
 751
 752         sbio->err = err;
 753         sbio->bio = bio;
 754
 755         btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
 756 }
 757
 758 static void scrub_checksum(struct btrfs_work *work)
 759 {
 760         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
 761         struct scrub_dev *sdev = sbio->sdev;
 762         struct page *page;
 763         void *buffer;
 764         int i;
 765         u64 flags;
 766         u64 logical;
 767         int ret;
 768
 769         if (sbio->err) {
 770                 ret = 0;
 771                 for (i = 0; i < sbio->count; ++i)
 772                         ret |= scrub_recheck_error(sbio, i);
 773                 if (!ret) {
 774                         spin_lock(&sdev->stat_lock);
 775                         ++sdev->stat.unverified_errors;
 776                         spin_unlock(&sdev->stat_lock);
 777                 }
 778
 779                 sbio->bio->bi_flags &= ~(BIO_POOL_MASK - 1);
 780                 sbio->bio->bi_flags |= 1 << BIO_UPTODATE;
 781                 sbio->bio->bi_phys_segments = 0;
 782                 sbio->bio->bi_idx = 0;
 783
 784                 for (i = 0; i < sbio->count; i++) {
 785                         struct bio_vec *bi;
 786                         bi = &sbio->bio->bi_io_vec[i];
 787                         bi->bv_offset = 0;
 788                         bi->bv_len = PAGE_SIZE;
 789                 }
 790                 goto out;
 791         }
 792         for (i = 0; i < sbio->count; ++i) {
 793                 page = sbio->bio->bi_io_vec[i].bv_page;
 794                 buffer = kmap_atomic(page, KM_USER0);
 795                 flags = sbio->spag[i].flags;
 796                 logical = sbio->logical + i * PAGE_SIZE;
 797                 ret = 0;
 798                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
 799                         ret = scrub_checksum_data(sdev, sbio->spag + i, buffer);
 800                 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 801                         ret = scrub_checksum_tree_block(sdev, sbio->spag + i,
 802                                                         logical, buffer);
 803                 } else if (flags & BTRFS_EXTENT_FLAG_SUPER) {
 804                         BUG_ON(i);
 805                         (void)scrub_checksum_super(sbio, buffer);
 806                 } else {
 807                         WARN_ON(1);
 808                 }
 809                 kunmap_atomic(buffer, KM_USER0);
 810                 if (ret) {
 811                         ret = scrub_recheck_error(sbio, i);
 812                         if (!ret) {
 813                                 spin_lock(&sdev->stat_lock);
 814                                 ++sdev->stat.unverified_errors;
 815                                 spin_unlock(&sdev->stat_lock);
 816                         }
 817                 }
 818         }
 819
 820 out:
 821         scrub_free_bio(sbio->bio);
 822         sbio->bio = NULL;
 823         spin_lock(&sdev->list_lock);
 824         sbio->next_free = sdev->first_free;
 825         sdev->first_free = sbio->index;
 826         spin_unlock(&sdev->list_lock);
 827         atomic_dec(&sdev->in_flight);
 828         wake_up(&sdev->list_wait);
 829 }
 830
 831 static int scrub_checksum_data(struct scrub_dev *sdev,
 832                                struct scrub_page *spag, void *buffer)
 833 {
 834         u8 csum[BTRFS_CSUM_SIZE];
 835         u32 crc = ~(u32)0;
 836         int fail = 0;
 837         struct btrfs_root *root = sdev->dev->dev_root;
 838
 839         if (!spag->have_csum)
 840                 return 0;
 841
 842         crc = btrfs_csum_data(root, buffer, crc, PAGE_SIZE);
 843         btrfs_csum_final(crc, csum);
 844         if (memcmp(csum, spag->csum, sdev->csum_size))
 845                 fail = 1;
 846
 847         spin_lock(&sdev->stat_lock);
 848         ++sdev->stat.data_extents_scrubbed;
 849         sdev->stat.data_bytes_scrubbed += PAGE_SIZE;
 850         if (fail)
 851                 ++sdev->stat.csum_errors;
 852         spin_unlock(&sdev->stat_lock);
 853
 854         return fail;
 855 }
 856
 857 static int scrub_checksum_tree_block(struct scrub_dev *sdev,
 858                                      struct scrub_page *spag, u64 logical,
 859                                      void *buffer)
 860 {
 861         struct btrfs_header *h;
 862         struct btrfs_root *root = sdev->dev->dev_root;
 863         struct btrfs_fs_info *fs_info = root->fs_info;
 864         u8 csum[BTRFS_CSUM_SIZE];
 865         u32 crc = ~(u32)0;
 866         int fail = 0;
 867         int crc_fail = 0;
 868
 869         /*
 870          * we don't use the getter functions here, as we
 871          * a) don't have an extent buffer and
 872          * b) the page is already kmapped
 873          */
 874         h = (struct btrfs_header *)buffer;
 875
 876         if (logical != le64_to_cpu(h->bytenr))
 877                 ++fail;
 878
 879         if (spag->generation != le64_to_cpu(h->generation))
 880                 ++fail;
 881
 882         if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
 883                 ++fail;
 884
 885         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
 886                    BTRFS_UUID_SIZE))
 887                 ++fail;
 888
 889         crc = btrfs_csum_data(root, buffer + BTRFS_CSUM_SIZE, crc,
 890                               PAGE_SIZE - BTRFS_CSUM_SIZE);
 891         btrfs_csum_final(crc, csum);
 892         if (memcmp(csum, h->csum, sdev->csum_size))
 893                 ++crc_fail;
 894
 895         spin_lock(&sdev->stat_lock);
 896         ++sdev->stat.tree_extents_scrubbed;
 897         sdev->stat.tree_bytes_scrubbed += PAGE_SIZE;
 898         if (crc_fail)
 899                 ++sdev->stat.csum_errors;
 900         if (fail)
 901                 ++sdev->stat.verify_errors;
 902         spin_unlock(&sdev->stat_lock);
 903
 904         return fail || crc_fail;
 905 }
 906
 907 static int scrub_checksum_super(struct scrub_bio *sbio, void *buffer)
 908 {
 909         struct btrfs_super_block *s;
 910         u64 logical;
 911         struct scrub_dev *sdev = sbio->sdev;
 912         struct btrfs_root *root = sdev->dev->dev_root;
 913         struct btrfs_fs_info *fs_info = root->fs_info;
 914         u8 csum[BTRFS_CSUM_SIZE];
 915         u32 crc = ~(u32)0;
 916         int fail = 0;
 917
 918         s = (struct btrfs_super_block *)buffer;
 919         logical = sbio->logical;
 920
 921         if (logical != le64_to_cpu(s->bytenr))
 922                 ++fail;
 923
 924         if (sbio->spag[0].generation != le64_to_cpu(s->generation))
 925                 ++fail;
 926
 927         if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
 928                 ++fail;
 929
 930         crc = btrfs_csum_data(root, buffer + BTRFS_CSUM_SIZE, crc,
 931                               PAGE_SIZE - BTRFS_CSUM_SIZE);
 932         btrfs_csum_final(crc, csum);
 933         if (memcmp(csum, s->csum, sbio->sdev->csum_size))
 934                 ++fail;
 935
 936         if (fail) {
 937                 /*
 938                  * if we find an error in a super block, we just report it.
 939                  * They will get written with the next transaction commit
 940                  * anyway
 941                  */
 942                 spin_lock(&sdev->stat_lock);
 943                 ++sdev->stat.super_errors;
 944                 spin_unlock(&sdev->stat_lock);
 945         }
 946
 947         return fail;
 948 }
 949
 950 static void scrub_submit(struct scrub_dev *sdev)
 951 {
 952         struct scrub_bio *sbio;
 953
 954         if (sdev->curr == -1)
 955                 return;
 956
 957         sbio = sdev->bios[sdev->curr];
 958         sbio->err = 0;
 959         sdev->curr = -1;
 960         atomic_inc(&sdev->in_flight);
 961
 962         btrfsic_submit_bio(READ, sbio->bio);
 963 }
 964
 965 static int scrub_page(struct scrub_dev *sdev, u64 logical, u64 len,
 966                       u64 physical, u64 flags, u64 gen, int mirror_num,
 967                       u8 *csum, int force)
 968 {
 969         struct scrub_bio *sbio;
 970         struct page *page;
 971         int ret;
 972
 973 again:
 974         /*
 975          * grab a fresh bio or wait for one to become available
 976          */
 977         while (sdev->curr == -1) {
 978                 spin_lock(&sdev->list_lock);
 979                 sdev->curr = sdev->first_free;
 980                 if (sdev->curr != -1) {
 981                         sdev->first_free = sdev->bios[sdev->curr]->next_free;
 982                         sdev->bios[sdev->curr]->next_free = -1;
 983                         sdev->bios[sdev->curr]->count = 0;
 984                         spin_unlock(&sdev->list_lock);
 985                 } else {
 986                         spin_unlock(&sdev->list_lock);
 987                         wait_event(sdev->list_wait, sdev->first_free != -1);
 988                 }
 989         }
 990         sbio = sdev->bios[sdev->curr];
 991         if (sbio->count == 0) {
 992                 struct bio *bio;
 993
 994                 sbio->physical = physical;
 995                 sbio->logical = logical;
 996                 bio = bio_alloc(GFP_NOFS, SCRUB_PAGES_PER_BIO);
 997                 if (!bio)
 998                         return -ENOMEM;
 999
1000                 bio->bi_private = sbio;
1001                 bio->bi_end_io = scrub_bio_end_io;
1002                 bio->bi_bdev = sdev->dev->bdev;
1003                 bio->bi_sector = sbio->physical >> 9;
1004                 sbio->err = 0;
1005                 sbio->bio = bio;
1006         } else if (sbio->physical + sbio->count * PAGE_SIZE != physical ||
1007                    sbio->logical + sbio->count * PAGE_SIZE != logical) {
1008                 scrub_submit(sdev);
1009                 goto again;
1010         }
1011         sbio->spag[sbio->count].flags = flags;
1012         sbio->spag[sbio->count].generation = gen;
1013         sbio->spag[sbio->count].have_csum = 0;
1014         sbio->spag[sbio->count].mirror_num = mirror_num;
1015
1016         page = alloc_page(GFP_NOFS);
1017         if (!page)
1018                 return -ENOMEM;
1019
1020         ret = bio_add_page(sbio->bio, page, PAGE_SIZE, 0);
1021         if (!ret) {
1022                 __free_page(page);
1023                 scrub_submit(sdev);
1024                 goto again;
1025         }
1026
1027         if (csum) {
1028                 sbio->spag[sbio->count].have_csum = 1;
1029                 memcpy(sbio->spag[sbio->count].csum, csum, sdev->csum_size);
1030         }
1031         ++sbio->count;
1032         if (sbio->count == SCRUB_PAGES_PER_BIO || force)
1033                 scrub_submit(sdev);
1034
1035         return 0;
1036 }
1037
1038 static int scrub_find_csum(struct scrub_dev *sdev, u64 logical, u64 len,
1039                            u8 *csum)
1040 {
1041         struct btrfs_ordered_sum *sum = NULL;
1042         int ret = 0;
1043         unsigned long i;
1044         unsigned long num_sectors;
1045         u32 sectorsize = sdev->dev->dev_root->sectorsize;
1046
1047         while (!list_empty(&sdev->csum_list)) {
1048                 sum = list_first_entry(&sdev->csum_list,
1049                                        struct btrfs_ordered_sum, list);
1050                 if (sum->bytenr > logical)
1051                         return 0;
1052                 if (sum->bytenr + sum->len > logical)
1053                         break;
1054
1055                 ++sdev->stat.csum_discards;
1056                 list_del(&sum->list);
1057                 kfree(sum);
1058                 sum = NULL;
1059         }
1060         if (!sum)
1061                 return 0;
1062
1063         num_sectors = sum->len / sectorsize;
1064         for (i = 0; i < num_sectors; ++i) {
1065                 if (sum->sums[i].bytenr == logical) {
1066                         memcpy(csum, &sum->sums[i].sum, sdev->csum_size);
1067                         ret = 1;
1068                         break;
1069                 }
1070         }
1071         if (ret && i == num_sectors - 1) {
1072                 list_del(&sum->list);
1073                 kfree(sum);
1074         }
1075         return ret;
1076 }
1077
1078 /* scrub extent tries to collect up to 64 kB for each bio */
1079 static int scrub_extent(struct scrub_dev *sdev, u64 logical, u64 len,
1080                         u64 physical, u64 flags, u64 gen, int mirror_num)
1081 {
1082         int ret;
1083         u8 csum[BTRFS_CSUM_SIZE];
1084
1085         while (len) {
1086                 u64 l = min_t(u64, len, PAGE_SIZE);
1087                 int have_csum = 0;
1088
1089                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
1090                         /* push csums to sbio */
1091                         have_csum = scrub_find_csum(sdev, logical, l, csum);
1092                         if (have_csum == 0)
1093                                 ++sdev->stat.no_csum;
1094                 }
1095                 ret = scrub_page(sdev, logical, l, physical, flags, gen,
1096                                  mirror_num, have_csum ? csum : NULL, 0);
1097                 if (ret)
1098                         return ret;
1099                 len -= l;
1100                 logical += l;
1101                 physical += l;
1102         }
1103         return 0;
1104 }
1105
1106 static noinline_for_stack int scrub_stripe(struct scrub_dev *sdev,
1107         struct map_lookup *map, int num, u64 base, u64 length)
1108 {
1109         struct btrfs_path *path;
1110         struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
1111         struct btrfs_root *root = fs_info->extent_root;
1112         struct btrfs_root *csum_root = fs_info->csum_root;
1113         struct btrfs_extent_item *extent;
1114         struct blk_plug plug;
1115         u64 flags;
1116         int ret;
1117         int slot;
1118         int i;
1119         u64 nstripes;
1120         struct extent_buffer *l;
1121         struct btrfs_key key;
1122         u64 physical;
1123         u64 logical;
1124         u64 generation;
1125         int mirror_num;
1126         struct reada_control *reada1;
1127         struct reada_control *reada2;
1128         struct btrfs_key key_start;
1129         struct btrfs_key key_end;
1130
1131         u64 increment = map->stripe_len;
1132         u64 offset;
1133
1134         nstripes = length;
1135         offset = 0;
1136         do_div(nstripes, map->stripe_len);
1137         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
1138                 offset = map->stripe_len * num;
1139                 increment = map->stripe_len * map->num_stripes;
1140                 mirror_num = 1;
1141         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
1142                 int factor = map->num_stripes / map->sub_stripes;
1143                 offset = map->stripe_len * (num / map->sub_stripes);
1144                 increment = map->stripe_len * factor;
1145                 mirror_num = num % map->sub_stripes + 1;
1146         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
1147                 increment = map->stripe_len;
1148                 mirror_num = num % map->num_stripes + 1;
1149         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
1150                 increment = map->stripe_len;
1151                 mirror_num = num % map->num_stripes + 1;
1152         } else {
1153                 increment = map->stripe_len;
1154                 mirror_num = 1;
1155         }
1156
1157         path = btrfs_alloc_path();
1158         if (!path)
1159                 return -ENOMEM;
1160
1161         path->search_commit_root = 1;
1162         path->skip_locking = 1;
1163
1164         /*
1165          * trigger the readahead for extent tree csum tree and wait for
1166          * completion. During readahead, the scrub is officially paused
1167          * to not hold off transaction commits
1168          */
1169         logical = base + offset;
1170
1171         wait_event(sdev->list_wait,
1172                    atomic_read(&sdev->in_flight) == 0);
1173         atomic_inc(&fs_info->scrubs_paused);
1174         wake_up(&fs_info->scrub_pause_wait);
1175
1176         /* FIXME it might be better to start readahead at commit root */
1177         key_start.objectid = logical;
1178         key_start.type = BTRFS_EXTENT_ITEM_KEY;
1179         key_start.offset = (u64)0;
1180         key_end.objectid = base + offset + nstripes * increment;
1181         key_end.type = BTRFS_EXTENT_ITEM_KEY;
1182         key_end.offset = (u64)0;
1183         reada1 = btrfs_reada_add(root, &key_start, &key_end);
1184
1185         key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
1186         key_start.type = BTRFS_EXTENT_CSUM_KEY;
1187         key_start.offset = logical;
1188         key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
1189         key_end.type = BTRFS_EXTENT_CSUM_KEY;
1190         key_end.offset = base + offset + nstripes * increment;
1191         reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
1192
1193         if (!IS_ERR(reada1))
1194                 btrfs_reada_wait(reada1);
1195         if (!IS_ERR(reada2))
1196                 btrfs_reada_wait(reada2);
1197
1198         mutex_lock(&fs_info->scrub_lock);
1199         while (atomic_read(&fs_info->scrub_pause_req)) {
1200                 mutex_unlock(&fs_info->scrub_lock);
1201                 wait_event(fs_info->scrub_pause_wait,
1202                    atomic_read(&fs_info->scrub_pause_req) == 0);
1203                 mutex_lock(&fs_info->scrub_lock);
1204         }
1205         atomic_dec(&fs_info->scrubs_paused);
1206         mutex_unlock(&fs_info->scrub_lock);
1207         wake_up(&fs_info->scrub_pause_wait);
1208
1209         /*
1210          * collect all data csums for the stripe to avoid seeking during
1211          * the scrub. This might currently (crc32) end up to be about 1MB
1212          */
1213         blk_start_plug(&plug);
1214
1215         /*
1216          * now find all extents for each stripe and scrub them
1217          */
1218         logical = base + offset;
1219         physical = map->stripes[num].physical;
1220         ret = 0;
1221         for (i = 0; i < nstripes; ++i) {
1222                 /*
1223                  * canceled?
1224                  */
1225                 if (atomic_read(&fs_info->scrub_cancel_req) ||
1226                     atomic_read(&sdev->cancel_req)) {
1227                         ret = -ECANCELED;
1228                         goto out;
1229                 }
1230                 /*
1231                  * check to see if we have to pause
1232                  */
1233                 if (atomic_read(&fs_info->scrub_pause_req)) {
1234                         /* push queued extents */
1235                         scrub_submit(sdev);
1236                         wait_event(sdev->list_wait,
1237                                    atomic_read(&sdev->in_flight) == 0);
1238                         atomic_inc(&fs_info->scrubs_paused);
1239                         wake_up(&fs_info->scrub_pause_wait);
1240                         mutex_lock(&fs_info->scrub_lock);
1241                         while (atomic_read(&fs_info->scrub_pause_req)) {
1242                                 mutex_unlock(&fs_info->scrub_lock);
1243                                 wait_event(fs_info->scrub_pause_wait,
1244                                    atomic_read(&fs_info->scrub_pause_req) == 0);
1245                                 mutex_lock(&fs_info->scrub_lock);
1246                         }
1247                         atomic_dec(&fs_info->scrubs_paused);
1248                         mutex_unlock(&fs_info->scrub_lock);
1249                         wake_up(&fs_info->scrub_pause_wait);
1250                 }
1251
1252                 ret = btrfs_lookup_csums_range(csum_root, logical,
1253                                                logical + map->stripe_len - 1,
1254                                                &sdev->csum_list, 1);
1255                 if (ret)
1256                         goto out;
1257
1258                 key.objectid = logical;
1259                 key.type = BTRFS_EXTENT_ITEM_KEY;
1260                 key.offset = (u64)0;
1261
1262                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1263                 if (ret < 0)
1264                         goto out;
1265                 if (ret > 0) {
1266                         ret = btrfs_previous_item(root, path, 0,
1267                                                   BTRFS_EXTENT_ITEM_KEY);
1268                         if (ret < 0)
1269                                 goto out;
1270                         if (ret > 0) {
1271                                 /* there's no smaller item, so stick with the
1272                                  * larger one */
1273                                 btrfs_release_path(path);
1274                                 ret = btrfs_search_slot(NULL, root, &key,
1275                                                         path, 0, 0);
1276                                 if (ret < 0)
1277                                         goto out;
1278                         }
1279                 }
1280
1281                 while (1) {
1282                         l = path->nodes[0];
1283                         slot = path->slots[0];
1284                         if (slot >= btrfs_header_nritems(l)) {
1285                                 ret = btrfs_next_leaf(root, path);
1286                                 if (ret == 0)
1287                                         continue;
1288                                 if (ret < 0)
1289                                         goto out;
1290
1291                                 break;
1292                         }
1293                         btrfs_item_key_to_cpu(l, &key, slot);
1294
1295                         if (key.objectid + key.offset <= logical)
1296                                 goto next;
1297
1298                         if (key.objectid >= logical + map->stripe_len)
1299                                 break;
1300
1301                         if (btrfs_key_type(&key) != BTRFS_EXTENT_ITEM_KEY)
1302                                 goto next;
1303
1304                         extent = btrfs_item_ptr(l, slot,
1305                                                 struct btrfs_extent_item);
1306                         flags = btrfs_extent_flags(l, extent);
1307                         generation = btrfs_extent_generation(l, extent);
1308
1309                         if (key.objectid < logical &&
1310                             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
1311                                 printk(KERN_ERR
1312                                        "btrfs scrub: tree block %llu spanning "
1313                                        "stripes, ignored. logical=%llu\n",
1314                                        (unsigned long long)key.objectid,
1315                                        (unsigned long long)logical);
1316                                 goto next;
1317                         }
1318
1319                         /*
1320                          * trim extent to this stripe
1321                          */
1322                         if (key.objectid < logical) {
1323                                 key.offset -= logical - key.objectid;
1324                                 key.objectid = logical;
1325                         }
1326                         if (key.objectid + key.offset >
1327                             logical + map->stripe_len) {
1328                                 key.offset = logical + map->stripe_len -
1329                                              key.objectid;
1330                         }
1331
1332                         ret = scrub_extent(sdev, key.objectid, key.offset,
1333                                            key.objectid - logical + physical,
1334                                            flags, generation, mirror_num);
1335                         if (ret)
1336                                 goto out;
1337
1338 next:
1339                         path->slots[0]++;
1340                 }
1341                 btrfs_release_path(path);
1342                 logical += increment;
1343                 physical += map->stripe_len;
1344                 spin_lock(&sdev->stat_lock);
1345                 sdev->stat.last_physical = physical;
1346                 spin_unlock(&sdev->stat_lock);
1347         }
1348         /* push queued extents */
1349         scrub_submit(sdev);
1350
1351 out:
1352         blk_finish_plug(&plug);
1353         btrfs_free_path(path);
1354         return ret < 0 ? ret : 0;
1355 }
1356
1357 static noinline_for_stack int scrub_chunk(struct scrub_dev *sdev,
1358         u64 chunk_tree, u64 chunk_objectid, u64 chunk_offset, u64 length,
1359         u64 dev_offset)
1360 {
1361         struct btrfs_mapping_tree *map_tree =
1362                 &sdev->dev->dev_root->fs_info->mapping_tree;
1363         struct map_lookup *map;
1364         struct extent_map *em;
1365         int i;
1366         int ret = -EINVAL;
1367
1368         read_lock(&map_tree->map_tree.lock);
1369         em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
1370         read_unlock(&map_tree->map_tree.lock);
1371
1372         if (!em)
1373                 return -EINVAL;
1374
1375         map = (struct map_lookup *)em->bdev;
1376         if (em->start != chunk_offset)
1377                 goto out;
1378
1379         if (em->len < length)
1380                 goto out;
1381
1382         for (i = 0; i < map->num_stripes; ++i) {
1383                 if (map->stripes[i].dev == sdev->dev &&
1384                     map->stripes[i].physical == dev_offset) {
1385                         ret = scrub_stripe(sdev, map, i, chunk_offset, length);
1386                         if (ret)
1387                                 goto out;
1388                 }
1389         }
1390 out:
1391         free_extent_map(em);
1392
1393         return ret;
1394 }
1395
1396 static noinline_for_stack
1397 int scrub_enumerate_chunks(struct scrub_dev *sdev, u64 start, u64 end)
1398 {
1399         struct btrfs_dev_extent *dev_extent = NULL;
1400         struct btrfs_path *path;
1401         struct btrfs_root *root = sdev->dev->dev_root;
1402         struct btrfs_fs_info *fs_info = root->fs_info;
1403         u64 length;
1404         u64 chunk_tree;
1405         u64 chunk_objectid;
1406         u64 chunk_offset;
1407         int ret;
1408         int slot;
1409         struct extent_buffer *l;
1410         struct btrfs_key key;
1411         struct btrfs_key found_key;
1412         struct btrfs_block_group_cache *cache;
1413
1414         path = btrfs_alloc_path();
1415         if (!path)
1416                 return -ENOMEM;
1417
1418         path->reada = 2;
1419         path->search_commit_root = 1;
1420         path->skip_locking = 1;
1421
1422         key.objectid = sdev->dev->devid;
1423         key.offset = 0ull;
1424         key.type = BTRFS_DEV_EXTENT_KEY;
1425
1426
1427         while (1) {
1428                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1429                 if (ret < 0)
1430                         break;
1431                 if (ret > 0) {
1432                         if (path->slots[0] >=
1433                             btrfs_header_nritems(path->nodes[0])) {
1434                                 ret = btrfs_next_leaf(root, path);
1435                                 if (ret)
1436                                         break;
1437                         }
1438                 }
1439
1440                 l = path->nodes[0];
1441                 slot = path->slots[0];
1442
1443                 btrfs_item_key_to_cpu(l, &found_key, slot);
1444
1445                 if (found_key.objectid != sdev->dev->devid)
1446                         break;
1447
1448                 if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
1449                         break;
1450
1451                 if (found_key.offset >= end)
1452                         break;
1453
1454                 if (found_key.offset < key.offset)
1455                         break;
1456
1457                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
1458                 length = btrfs_dev_extent_length(l, dev_extent);
1459
1460                 if (found_key.offset + length <= start) {
1461                         key.offset = found_key.offset + length;
1462                         btrfs_release_path(path);
1463                         continue;
1464                 }
1465
1466                 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
1467                 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
1468                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
1469
1470                 /*
1471                  * get a reference on the corresponding block group to prevent
1472                  * the chunk from going away while we scrub it
1473                  */
1474                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
1475                 if (!cache) {
1476                         ret = -ENOENT;
1477                         break;
1478                 }
1479                 ret = scrub_chunk(sdev, chunk_tree, chunk_objectid,
1480                                   chunk_offset, length, found_key.offset);
1481                 btrfs_put_block_group(cache);
1482                 if (ret)
1483                         break;
1484
1485                 key.offset = found_key.offset + length;
1486                 btrfs_release_path(path);
1487         }
1488
1489         btrfs_free_path(path);
1490
1491         /*
1492          * ret can still be 1 from search_slot or next_leaf,
1493          * that's not an error
1494          */
1495         return ret < 0 ? ret : 0;
1496 }
1497
1498 static noinline_for_stack int scrub_supers(struct scrub_dev *sdev)
1499 {
1500         int     i;
1501         u64     bytenr;
1502         u64     gen;
1503         int     ret;
1504         struct btrfs_device *device = sdev->dev;
1505         struct btrfs_root *root = device->dev_root;
1506
1507         gen = root->fs_info->last_trans_committed;
1508
1509         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1510                 bytenr = btrfs_sb_offset(i);
1511                 if (bytenr + BTRFS_SUPER_INFO_SIZE > device->total_bytes)
1512                         break;
1513
1514                 ret = scrub_page(sdev, bytenr, PAGE_SIZE, bytenr,
1515                                  BTRFS_EXTENT_FLAG_SUPER, gen, i, NULL, 1);
1516                 if (ret)
1517                         return ret;
1518         }
1519         wait_event(sdev->list_wait, atomic_read(&sdev->in_flight) == 0);
1520
1521         return 0;
1522 }
1523
1524 /*
1525  * get a reference count on fs_info->scrub_workers. start worker if necessary
1526  */
1527 static noinline_for_stack int scrub_workers_get(struct btrfs_root *root)
1528 {
1529         struct btrfs_fs_info *fs_info = root->fs_info;
1530         int ret = 0;
1531
1532         mutex_lock(&fs_info->scrub_lock);
1533         if (fs_info->scrub_workers_refcnt == 0) {
1534                 btrfs_init_workers(&fs_info->scrub_workers, "scrub",
1535                            fs_info->thread_pool_size, &fs_info->generic_worker);
1536                 fs_info->scrub_workers.idle_thresh = 4;
1537                 ret = btrfs_start_workers(&fs_info->scrub_workers);
1538                 if (ret)
1539                         goto out;
1540         }
1541         ++fs_info->scrub_workers_refcnt;
1542 out:
1543         mutex_unlock(&fs_info->scrub_lock);
1544
1545         return ret;
1546 }
1547
1548 static noinline_for_stack void scrub_workers_put(struct btrfs_root *root)
1549 {
1550         struct btrfs_fs_info *fs_info = root->fs_info;
1551
1552         mutex_lock(&fs_info->scrub_lock);
1553         if (--fs_info->scrub_workers_refcnt == 0)
1554                 btrfs_stop_workers(&fs_info->scrub_workers);
1555         WARN_ON(fs_info->scrub_workers_refcnt < 0);
1556         mutex_unlock(&fs_info->scrub_lock);
1557 }
1558
1559
1560 int btrfs_scrub_dev(struct btrfs_root *root, u64 devid, u64 start, u64 end,
1561                     struct btrfs_scrub_progress *progress, int readonly)
1562 {
1563         struct scrub_dev *sdev;
1564         struct btrfs_fs_info *fs_info = root->fs_info;
1565         int ret;
1566         struct btrfs_device *dev;
1567
1568         if (btrfs_fs_closing(root->fs_info))
1569                 return -EINVAL;
1570
1571         /*
1572          * check some assumptions
1573          */
1574         if (root->sectorsize != PAGE_SIZE ||
1575             root->sectorsize != root->leafsize ||
1576             root->sectorsize != root->nodesize) {
1577                 printk(KERN_ERR "btrfs_scrub: size assumptions fail\n");
1578                 return -EINVAL;
1579         }
1580
1581         ret = scrub_workers_get(root);
1582         if (ret)
1583                 return ret;
1584
1585         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
1586         dev = btrfs_find_device(root, devid, NULL, NULL);
1587         if (!dev || dev->missing) {
1588                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1589                 scrub_workers_put(root);
1590                 return -ENODEV;
1591         }
1592         mutex_lock(&fs_info->scrub_lock);
1593
1594         if (!dev->in_fs_metadata) {
1595                 mutex_unlock(&fs_info->scrub_lock);
1596                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1597                 scrub_workers_put(root);
1598                 return -ENODEV;
1599         }
1600
1601         if (dev->scrub_device) {
1602                 mutex_unlock(&fs_info->scrub_lock);
1603                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1604                 scrub_workers_put(root);
1605                 return -EINPROGRESS;
1606         }
1607         sdev = scrub_setup_dev(dev);
1608         if (IS_ERR(sdev)) {
1609                 mutex_unlock(&fs_info->scrub_lock);
1610                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1611                 scrub_workers_put(root);
1612                 return PTR_ERR(sdev);
1613         }
1614         sdev->readonly = readonly;
1615         dev->scrub_device = sdev;
1616
1617         atomic_inc(&fs_info->scrubs_running);
1618         mutex_unlock(&fs_info->scrub_lock);
1619         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1620
1621         down_read(&fs_info->scrub_super_lock);
1622         ret = scrub_supers(sdev);
1623         up_read(&fs_info->scrub_super_lock);
1624
1625         if (!ret)
1626                 ret = scrub_enumerate_chunks(sdev, start, end);
1627
1628         wait_event(sdev->list_wait, atomic_read(&sdev->in_flight) == 0);
1629         atomic_dec(&fs_info->scrubs_running);
1630         wake_up(&fs_info->scrub_pause_wait);
1631
1632         wait_event(sdev->list_wait, atomic_read(&sdev->fixup_cnt) == 0);
1633
1634         if (progress)
1635                 memcpy(progress, &sdev->stat, sizeof(*progress));
1636
1637         mutex_lock(&fs_info->scrub_lock);
1638         dev->scrub_device = NULL;
1639         mutex_unlock(&fs_info->scrub_lock);
1640
1641         scrub_free_dev(sdev);
1642         scrub_workers_put(root);
1643
1644         return ret;
1645 }
1646
1647 int btrfs_scrub_pause(struct btrfs_root *root)
1648 {
1649         struct btrfs_fs_info *fs_info = root->fs_info;
1650
1651         mutex_lock(&fs_info->scrub_lock);
1652         atomic_inc(&fs_info->scrub_pause_req);
1653         while (atomic_read(&fs_info->scrubs_paused) !=
1654                atomic_read(&fs_info->scrubs_running)) {
1655                 mutex_unlock(&fs_info->scrub_lock);
1656                 wait_event(fs_info->scrub_pause_wait,
1657                            atomic_read(&fs_info->scrubs_paused) ==
1658                            atomic_read(&fs_info->scrubs_running));
1659                 mutex_lock(&fs_info->scrub_lock);
1660         }
1661         mutex_unlock(&fs_info->scrub_lock);
1662
1663         return 0;
1664 }
1665
1666 int btrfs_scrub_continue(struct btrfs_root *root)
1667 {
1668         struct btrfs_fs_info *fs_info = root->fs_info;
1669
1670         atomic_dec(&fs_info->scrub_pause_req);
1671         wake_up(&fs_info->scrub_pause_wait);
1672         return 0;
1673 }
1674
1675 int btrfs_scrub_pause_super(struct btrfs_root *root)
1676 {
1677         down_write(&root->fs_info->scrub_super_lock);
1678         return 0;
1679 }
1680
1681 int btrfs_scrub_continue_super(struct btrfs_root *root)
1682 {
1683         up_write(&root->fs_info->scrub_super_lock);
1684         return 0;
1685 }
1686
1687 int btrfs_scrub_cancel(struct btrfs_root *root)
1688 {
1689         struct btrfs_fs_info *fs_info = root->fs_info;
1690
1691         mutex_lock(&fs_info->scrub_lock);
1692         if (!atomic_read(&fs_info->scrubs_running)) {
1693                 mutex_unlock(&fs_info->scrub_lock);
1694                 return -ENOTCONN;
1695         }
1696
1697         atomic_inc(&fs_info->scrub_cancel_req);
1698         while (atomic_read(&fs_info->scrubs_running)) {
1699                 mutex_unlock(&fs_info->scrub_lock);
1700                 wait_event(fs_info->scrub_pause_wait,
1701                            atomic_read(&fs_info->scrubs_running) == 0);
1702                 mutex_lock(&fs_info->scrub_lock);
1703         }
1704         atomic_dec(&fs_info->scrub_cancel_req);
1705         mutex_unlock(&fs_info->scrub_lock);
1706
1707         return 0;
1708 }
1709
1710 int btrfs_scrub_cancel_dev(struct btrfs_root *root, struct btrfs_device *dev)
1711 {
1712         struct btrfs_fs_info *fs_info = root->fs_info;
1713         struct scrub_dev *sdev;
1714
1715         mutex_lock(&fs_info->scrub_lock);
1716         sdev = dev->scrub_device;
1717         if (!sdev) {
1718                 mutex_unlock(&fs_info->scrub_lock);
1719                 return -ENOTCONN;
1720         }
1721         atomic_inc(&sdev->cancel_req);
1722         while (dev->scrub_device) {
1723                 mutex_unlock(&fs_info->scrub_lock);
1724                 wait_event(fs_info->scrub_pause_wait,
1725                            dev->scrub_device == NULL);
1726                 mutex_lock(&fs_info->scrub_lock);
1727         }
1728         mutex_unlock(&fs_info->scrub_lock);
1729
1730         return 0;
1731 }
1732
1733 int btrfs_scrub_cancel_devid(struct btrfs_root *root, u64 devid)
1734 {
1735         struct btrfs_fs_info *fs_info = root->fs_info;
1736         struct btrfs_device *dev;
1737         int ret;
1738
1739         /*
1740          * we have to hold the device_list_mutex here so the device
1741          * does not go away in cancel_dev. FIXME: find a better solution
1742          */
1743         mutex_lock(&fs_info->fs_devices->device_list_mutex);
1744         dev = btrfs_find_device(root, devid, NULL, NULL);
1745         if (!dev) {
1746                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
1747                 return -ENODEV;
1748         }
1749         ret = btrfs_scrub_cancel_dev(root, dev);
1750         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
1751
1752         return ret;
1753 }
1754
1755 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
1756                          struct btrfs_scrub_progress *progress)
1757 {
1758         struct btrfs_device *dev;
1759         struct scrub_dev *sdev = NULL;
1760
1761         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
1762         dev = btrfs_find_device(root, devid, NULL, NULL);
1763         if (dev)
1764                 sdev = dev->scrub_device;
1765         if (sdev)
1766                 memcpy(progress, &sdev->stat, sizeof(*progress));
1767         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1768
1769         return dev ? (sdev ? 0 : -ENOTCONN) : -ENODEV;
1770 }