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
  29 /*
  30  * This is only the first step towards a full-features scrub. It reads all
  31  * extent and super block and verifies the checksums. In case a bad checksum
  32  * is found or the extent cannot be read, good data will be written back if
  33  * any can be found.
  34  *
  35  * Future enhancements:
  36  *  - In case an unrepairable extent is encountered, track which files are
  37  *    affected and report them
  38  *  - In case of a read error on files with nodatasum, map the file and read
  39  *    the extent to trigger a writeback of the good copy
  40  *  - track and record media errors, throw out bad devices
  41  *  - add a mode to also read unallocated space
  42  */
  43
  44 struct scrub_bio;
  45 struct scrub_page;
  46 struct scrub_dev;
  47 static void scrub_bio_end_io(struct bio *bio, int err);
  48 static void scrub_checksum(struct btrfs_work *work);
  49 static int scrub_checksum_data(struct scrub_dev *sdev,
  50                                struct scrub_page *spag, void *buffer);
  51 static int scrub_checksum_tree_block(struct scrub_dev *sdev,
  52                                      struct scrub_page *spag, u64 logical,
  53                                      void *buffer);
  54 static int scrub_checksum_super(struct scrub_bio *sbio, void *buffer);
  55 static int scrub_fixup_check(struct scrub_bio *sbio, int ix);
  56 static void scrub_fixup_end_io(struct bio *bio, int err);
  57 static int scrub_fixup_io(int rw, struct block_device *bdev, sector_t sector,
  58                           struct page *page);
  59 static void scrub_fixup(struct scrub_bio *sbio, int ix);
  60
  61 #define SCRUB_PAGES_PER_BIO     16      /* 64k per bio */
  62 #define SCRUB_BIOS_PER_DEV      16      /* 1 MB per device in flight */
  63
  64 struct scrub_page {
  65         u64                     flags;  /* extent flags */
  66         u64                     generation;
  67         int                     mirror_num;
  68         int                     have_csum;
  69         u8                      csum[BTRFS_CSUM_SIZE];
  70 };
  71
  72 struct scrub_bio {
  73         int                     index;
  74         struct scrub_dev        *sdev;
  75         struct bio              *bio;
  76         int                     err;
  77         u64                     logical;
  78         u64                     physical;
  79         struct scrub_page       spag[SCRUB_PAGES_PER_BIO];
  80         u64                     count;
  81         int                     next_free;
  82         struct btrfs_work       work;
  83 };
  84
  85 struct scrub_dev {
  86         struct scrub_bio        *bios[SCRUB_BIOS_PER_DEV];
  87         struct btrfs_device     *dev;
  88         int                     first_free;
  89         int                     curr;
  90         atomic_t                in_flight;
  91         atomic_t                fixup_cnt;
  92         spinlock_t              list_lock;
  93         wait_queue_head_t       list_wait;
  94         u16                     csum_size;
  95         struct list_head        csum_list;
  96         atomic_t                cancel_req;
  97         int                     readonly;
  98         /*
  99          * statistics
 100          */
 101         struct btrfs_scrub_progress stat;
 102         spinlock_t              stat_lock;
 103 };
 104
 105 struct scrub_fixup_nodatasum {
 106         struct scrub_dev        *sdev;
 107         u64                     logical;
 108         struct btrfs_root       *root;
 109         struct btrfs_work       work;
 110         int                     mirror_num;
 111 };
 112
 113 struct scrub_warning {
 114         struct btrfs_path       *path;
 115         u64                     extent_item_size;
 116         char                    *scratch_buf;
 117         char                    *msg_buf;
 118         const char              *errstr;
 119         sector_t                sector;
 120         u64                     logical;
 121         struct btrfs_device     *dev;
 122         int                     msg_bufsize;
 123         int                     scratch_bufsize;
 124 };
 125
 126 static void scrub_free_csums(struct scrub_dev *sdev)
 127 {
 128         while (!list_empty(&sdev->csum_list)) {
 129                 struct btrfs_ordered_sum *sum;
 130                 sum = list_first_entry(&sdev->csum_list,
 131                                        struct btrfs_ordered_sum, list);
 132                 list_del(&sum->list);
 133                 kfree(sum);
 134         }
 135 }
 136
 137 static void scrub_free_bio(struct bio *bio)
 138 {
 139         int i;
 140         struct page *last_page = NULL;
 141
 142         if (!bio)
 143                 return;
 144
 145         for (i = 0; i < bio->bi_vcnt; ++i) {
 146                 if (bio->bi_io_vec[i].bv_page == last_page)
 147                         continue;
 148                 last_page = bio->bi_io_vec[i].bv_page;
 149                 __free_page(last_page);
 150         }
 151         bio_put(bio);
 152 }
 153
 154 static noinline_for_stack void scrub_free_dev(struct scrub_dev *sdev)
 155 {
 156         int i;
 157
 158         if (!sdev)
 159                 return;
 160
 161         for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
 162                 struct scrub_bio *sbio = sdev->bios[i];
 163
 164                 if (!sbio)
 165                         break;
 166
 167                 scrub_free_bio(sbio->bio);
 168                 kfree(sbio);
 169         }
 170
 171         scrub_free_csums(sdev);
 172         kfree(sdev);
 173 }
 174
 175 static noinline_for_stack
 176 struct scrub_dev *scrub_setup_dev(struct btrfs_device *dev)
 177 {
 178         struct scrub_dev *sdev;
 179         int             i;
 180         struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
 181
 182         sdev = kzalloc(sizeof(*sdev), GFP_NOFS);
 183         if (!sdev)
 184                 goto nomem;
 185         sdev->dev = dev;
 186         for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
 187                 struct scrub_bio *sbio;
 188
 189                 sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
 190                 if (!sbio)
 191                         goto nomem;
 192                 sdev->bios[i] = sbio;
 193
 194                 sbio->index = i;
 195                 sbio->sdev = sdev;
 196                 sbio->count = 0;
 197                 sbio->work.func = scrub_checksum;
 198
 199                 if (i != SCRUB_BIOS_PER_DEV-1)
 200                         sdev->bios[i]->next_free = i + 1;
 201                 else
 202                         sdev->bios[i]->next_free = -1;
 203         }
 204         sdev->first_free = 0;
 205         sdev->curr = -1;
 206         atomic_set(&sdev->in_flight, 0);
 207         atomic_set(&sdev->fixup_cnt, 0);
 208         atomic_set(&sdev->cancel_req, 0);
 209         sdev->csum_size = btrfs_super_csum_size(fs_info->super_copy);
 210         INIT_LIST_HEAD(&sdev->csum_list);
 211
 212         spin_lock_init(&sdev->list_lock);
 213         spin_lock_init(&sdev->stat_lock);
 214         init_waitqueue_head(&sdev->list_wait);
 215         return sdev;
 216
 217 nomem:
 218         scrub_free_dev(sdev);
 219         return ERR_PTR(-ENOMEM);
 220 }
 221
 222 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root, void *ctx)
 223 {
 224         u64 isize;
 225         u32 nlink;
 226         int ret;
 227         int i;
 228         struct extent_buffer *eb;
 229         struct btrfs_inode_item *inode_item;
 230         struct scrub_warning *swarn = ctx;
 231         struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
 232         struct inode_fs_paths *ipath = NULL;
 233         struct btrfs_root *local_root;
 234         struct btrfs_key root_key;
 235
 236         root_key.objectid = root;
 237         root_key.type = BTRFS_ROOT_ITEM_KEY;
 238         root_key.offset = (u64)-1;
 239         local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
 240         if (IS_ERR(local_root)) {
 241                 ret = PTR_ERR(local_root);
 242                 goto err;
 243         }
 244
 245         ret = inode_item_info(inum, 0, local_root, swarn->path);
 246         if (ret) {
 247                 btrfs_release_path(swarn->path);
 248                 goto err;
 249         }
 250
 251         eb = swarn->path->nodes[0];
 252         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 253                                         struct btrfs_inode_item);
 254         isize = btrfs_inode_size(eb, inode_item);
 255         nlink = btrfs_inode_nlink(eb, inode_item);
 256         btrfs_release_path(swarn->path);
 257
 258         ipath = init_ipath(4096, local_root, swarn->path);
 259         if (IS_ERR(ipath)) {
 260                 ret = PTR_ERR(ipath);
 261                 ipath = NULL;
 262                 goto err;
 263         }
 264         ret = paths_from_inode(inum, ipath);
 265
 266         if (ret < 0)
 267                 goto err;
 268
 269         /*
 270          * we deliberately ignore the bit ipath might have been too small to
 271          * hold all of the paths here
 272          */
 273         for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 274                 printk(KERN_WARNING "btrfs: %s at logical %llu on dev "
 275                         "%s, sector %llu, root %llu, inode %llu, offset %llu, "
 276                         "length %llu, links %u (path: %s)\n", swarn->errstr,
 277                         swarn->logical, swarn->dev->name,
 278                         (unsigned long long)swarn->sector, root, inum, offset,
 279                         min(isize - offset, (u64)PAGE_SIZE), nlink,
 280                         (char *)(unsigned long)ipath->fspath->val[i]);
 281
 282         free_ipath(ipath);
 283         return 0;
 284
 285 err:
 286         printk(KERN_WARNING "btrfs: %s at logical %llu on dev "
 287                 "%s, sector %llu, root %llu, inode %llu, offset %llu: path "
 288                 "resolving failed with ret=%d\n", swarn->errstr,
 289                 swarn->logical, swarn->dev->name,
 290                 (unsigned long long)swarn->sector, root, inum, offset, ret);
 291
 292         free_ipath(ipath);
 293         return 0;
 294 }
 295
 296 static void scrub_print_warning(const char *errstr, struct scrub_bio *sbio,
 297                                 int ix)
 298 {
 299         struct btrfs_device *dev = sbio->sdev->dev;
 300         struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
 301         struct btrfs_path *path;
 302         struct btrfs_key found_key;
 303         struct extent_buffer *eb;
 304         struct btrfs_extent_item *ei;
 305         struct scrub_warning swarn;
 306         u32 item_size;
 307         int ret;
 308         u64 ref_root;
 309         u8 ref_level;
 310         unsigned long ptr = 0;
 311         const int bufsize = 4096;
 312         u64 extent_item_pos;
 313
 314         path = btrfs_alloc_path();
 315
 316         swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
 317         swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
 318         swarn.sector = (sbio->physical + ix * PAGE_SIZE) >> 9;
 319         swarn.logical = sbio->logical + ix * PAGE_SIZE;
 320         swarn.errstr = errstr;
 321         swarn.dev = dev;
 322         swarn.msg_bufsize = bufsize;
 323         swarn.scratch_bufsize = bufsize;
 324
 325         if (!path || !swarn.scratch_buf || !swarn.msg_buf)
 326                 goto out;
 327
 328         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key);
 329         if (ret < 0)
 330                 goto out;
 331
 332         extent_item_pos = swarn.logical - found_key.objectid;
 333         swarn.extent_item_size = found_key.offset;
 334
 335         eb = path->nodes[0];
 336         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 337         item_size = btrfs_item_size_nr(eb, path->slots[0]);
 338         btrfs_release_path(path);
 339
 340         if (ret & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 341                 do {
 342                         ret = tree_backref_for_extent(&ptr, eb, ei, item_size,
 343                                                         &ref_root, &ref_level);
 344                         printk(KERN_WARNING "%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         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 int scrub_submit(struct scrub_dev *sdev)
 951 {
 952         struct scrub_bio *sbio;
 953
 954         if (sdev->curr == -1)
 955                 return 0;
 956
 957         sbio = sdev->bios[sdev->curr];
 958         sbio->err = 0;
 959         sdev->curr = -1;
 960         atomic_inc(&sdev->in_flight);
 961
 962         submit_bio(READ, sbio->bio);
 963
 964         return 0;
 965 }
 966
 967 static int scrub_page(struct scrub_dev *sdev, u64 logical, u64 len,
 968                       u64 physical, u64 flags, u64 gen, int mirror_num,
 969                       u8 *csum, int force)
 970 {
 971         struct scrub_bio *sbio;
 972         struct page *page;
 973         int ret;
 974
 975 again:
 976         /*
 977          * grab a fresh bio or wait for one to become available
 978          */
 979         while (sdev->curr == -1) {
 980                 spin_lock(&sdev->list_lock);
 981                 sdev->curr = sdev->first_free;
 982                 if (sdev->curr != -1) {
 983                         sdev->first_free = sdev->bios[sdev->curr]->next_free;
 984                         sdev->bios[sdev->curr]->next_free = -1;
 985                         sdev->bios[sdev->curr]->count = 0;
 986                         spin_unlock(&sdev->list_lock);
 987                 } else {
 988                         spin_unlock(&sdev->list_lock);
 989                         wait_event(sdev->list_wait, sdev->first_free != -1);
 990                 }
 991         }
 992         sbio = sdev->bios[sdev->curr];
 993         if (sbio->count == 0) {
 994                 struct bio *bio;
 995
 996                 sbio->physical = physical;
 997                 sbio->logical = logical;
 998                 bio = bio_alloc(GFP_NOFS, SCRUB_PAGES_PER_BIO);
 999                 if (!bio)
1000                         return -ENOMEM;
1001
1002                 bio->bi_private = sbio;
1003                 bio->bi_end_io = scrub_bio_end_io;
1004                 bio->bi_bdev = sdev->dev->bdev;
1005                 bio->bi_sector = sbio->physical >> 9;
1006                 sbio->err = 0;
1007                 sbio->bio = bio;
1008         } else if (sbio->physical + sbio->count * PAGE_SIZE != physical ||
1009                    sbio->logical + sbio->count * PAGE_SIZE != logical) {
1010                 ret = scrub_submit(sdev);
1011                 if (ret)
1012                         return ret;
1013                 goto again;
1014         }
1015         sbio->spag[sbio->count].flags = flags;
1016         sbio->spag[sbio->count].generation = gen;
1017         sbio->spag[sbio->count].have_csum = 0;
1018         sbio->spag[sbio->count].mirror_num = mirror_num;
1019
1020         page = alloc_page(GFP_NOFS);
1021         if (!page)
1022                 return -ENOMEM;
1023
1024         ret = bio_add_page(sbio->bio, page, PAGE_SIZE, 0);
1025         if (!ret) {
1026                 __free_page(page);
1027                 ret = scrub_submit(sdev);
1028                 if (ret)
1029                         return ret;
1030                 goto again;
1031         }
1032
1033         if (csum) {
1034                 sbio->spag[sbio->count].have_csum = 1;
1035                 memcpy(sbio->spag[sbio->count].csum, csum, sdev->csum_size);
1036         }
1037         ++sbio->count;
1038         if (sbio->count == SCRUB_PAGES_PER_BIO || force) {
1039                 int ret;
1040
1041                 ret = scrub_submit(sdev);
1042                 if (ret)
1043                         return ret;
1044         }
1045
1046         return 0;
1047 }
1048
1049 static int scrub_find_csum(struct scrub_dev *sdev, u64 logical, u64 len,
1050                            u8 *csum)
1051 {
1052         struct btrfs_ordered_sum *sum = NULL;
1053         int ret = 0;
1054         unsigned long i;
1055         unsigned long num_sectors;
1056         u32 sectorsize = sdev->dev->dev_root->sectorsize;
1057
1058         while (!list_empty(&sdev->csum_list)) {
1059                 sum = list_first_entry(&sdev->csum_list,
1060                                        struct btrfs_ordered_sum, list);
1061                 if (sum->bytenr > logical)
1062                         return 0;
1063                 if (sum->bytenr + sum->len > logical)
1064                         break;
1065
1066                 ++sdev->stat.csum_discards;
1067                 list_del(&sum->list);
1068                 kfree(sum);
1069                 sum = NULL;
1070         }
1071         if (!sum)
1072                 return 0;
1073
1074         num_sectors = sum->len / sectorsize;
1075         for (i = 0; i < num_sectors; ++i) {
1076                 if (sum->sums[i].bytenr == logical) {
1077                         memcpy(csum, &sum->sums[i].sum, sdev->csum_size);
1078                         ret = 1;
1079                         break;
1080                 }
1081         }
1082         if (ret && i == num_sectors - 1) {
1083                 list_del(&sum->list);
1084                 kfree(sum);
1085         }
1086         return ret;
1087 }
1088
1089 /* scrub extent tries to collect up to 64 kB for each bio */
1090 static int scrub_extent(struct scrub_dev *sdev, u64 logical, u64 len,
1091                         u64 physical, u64 flags, u64 gen, int mirror_num)
1092 {
1093         int ret;
1094         u8 csum[BTRFS_CSUM_SIZE];
1095
1096         while (len) {
1097                 u64 l = min_t(u64, len, PAGE_SIZE);
1098                 int have_csum = 0;
1099
1100                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
1101                         /* push csums to sbio */
1102                         have_csum = scrub_find_csum(sdev, logical, l, csum);
1103                         if (have_csum == 0)
1104                                 ++sdev->stat.no_csum;
1105                 }
1106                 ret = scrub_page(sdev, logical, l, physical, flags, gen,
1107                                  mirror_num, have_csum ? csum : NULL, 0);
1108                 if (ret)
1109                         return ret;
1110                 len -= l;
1111                 logical += l;
1112                 physical += l;
1113         }
1114         return 0;
1115 }
1116
1117 static noinline_for_stack int scrub_stripe(struct scrub_dev *sdev,
1118         struct map_lookup *map, int num, u64 base, u64 length)
1119 {
1120         struct btrfs_path *path;
1121         struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
1122         struct btrfs_root *root = fs_info->extent_root;
1123         struct btrfs_root *csum_root = fs_info->csum_root;
1124         struct btrfs_extent_item *extent;
1125         struct blk_plug plug;
1126         u64 flags;
1127         int ret;
1128         int slot;
1129         int i;
1130         u64 nstripes;
1131         struct extent_buffer *l;
1132         struct btrfs_key key;
1133         u64 physical;
1134         u64 logical;
1135         u64 generation;
1136         int mirror_num;
1137         struct reada_control *reada1;
1138         struct reada_control *reada2;
1139         struct btrfs_key key_start;
1140         struct btrfs_key key_end;
1141
1142         u64 increment = map->stripe_len;
1143         u64 offset;
1144
1145         nstripes = length;
1146         offset = 0;
1147         do_div(nstripes, map->stripe_len);
1148         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
1149                 offset = map->stripe_len * num;
1150                 increment = map->stripe_len * map->num_stripes;
1151                 mirror_num = 1;
1152         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
1153                 int factor = map->num_stripes / map->sub_stripes;
1154                 offset = map->stripe_len * (num / map->sub_stripes);
1155                 increment = map->stripe_len * factor;
1156                 mirror_num = num % map->sub_stripes + 1;
1157         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
1158                 increment = map->stripe_len;
1159                 mirror_num = num % map->num_stripes + 1;
1160         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
1161                 increment = map->stripe_len;
1162                 mirror_num = num % map->num_stripes + 1;
1163         } else {
1164                 increment = map->stripe_len;
1165                 mirror_num = 1;
1166         }
1167
1168         path = btrfs_alloc_path();
1169         if (!path)
1170                 return -ENOMEM;
1171
1172         path->search_commit_root = 1;
1173         path->skip_locking = 1;
1174
1175         /*
1176          * trigger the readahead for extent tree csum tree and wait for
1177          * completion. During readahead, the scrub is officially paused
1178          * to not hold off transaction commits
1179          */
1180         logical = base + offset;
1181
1182         wait_event(sdev->list_wait,
1183                    atomic_read(&sdev->in_flight) == 0);
1184         atomic_inc(&fs_info->scrubs_paused);
1185         wake_up(&fs_info->scrub_pause_wait);
1186
1187         /* FIXME it might be better to start readahead at commit root */
1188         key_start.objectid = logical;
1189         key_start.type = BTRFS_EXTENT_ITEM_KEY;
1190         key_start.offset = (u64)0;
1191         key_end.objectid = base + offset + nstripes * increment;
1192         key_end.type = BTRFS_EXTENT_ITEM_KEY;
1193         key_end.offset = (u64)0;
1194         reada1 = btrfs_reada_add(root, &key_start, &key_end);
1195
1196         key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
1197         key_start.type = BTRFS_EXTENT_CSUM_KEY;
1198         key_start.offset = logical;
1199         key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
1200         key_end.type = BTRFS_EXTENT_CSUM_KEY;
1201         key_end.offset = base + offset + nstripes * increment;
1202         reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
1203
1204         if (!IS_ERR(reada1))
1205                 btrfs_reada_wait(reada1);
1206         if (!IS_ERR(reada2))
1207                 btrfs_reada_wait(reada2);
1208
1209         mutex_lock(&fs_info->scrub_lock);
1210         while (atomic_read(&fs_info->scrub_pause_req)) {
1211                 mutex_unlock(&fs_info->scrub_lock);
1212                 wait_event(fs_info->scrub_pause_wait,
1213                    atomic_read(&fs_info->scrub_pause_req) == 0);
1214                 mutex_lock(&fs_info->scrub_lock);
1215         }
1216         atomic_dec(&fs_info->scrubs_paused);
1217         mutex_unlock(&fs_info->scrub_lock);
1218         wake_up(&fs_info->scrub_pause_wait);
1219
1220         /*
1221          * collect all data csums for the stripe to avoid seeking during
1222          * the scrub. This might currently (crc32) end up to be about 1MB
1223          */
1224         blk_start_plug(&plug);
1225
1226         /*
1227          * now find all extents for each stripe and scrub them
1228          */
1229         logical = base + offset;
1230         physical = map->stripes[num].physical;
1231         ret = 0;
1232         for (i = 0; i < nstripes; ++i) {
1233                 /*
1234                  * canceled?
1235                  */
1236                 if (atomic_read(&fs_info->scrub_cancel_req) ||
1237                     atomic_read(&sdev->cancel_req)) {
1238                         ret = -ECANCELED;
1239                         goto out;
1240                 }
1241                 /*
1242                  * check to see if we have to pause
1243                  */
1244                 if (atomic_read(&fs_info->scrub_pause_req)) {
1245                         /* push queued extents */
1246                         scrub_submit(sdev);
1247                         wait_event(sdev->list_wait,
1248                                    atomic_read(&sdev->in_flight) == 0);
1249                         atomic_inc(&fs_info->scrubs_paused);
1250                         wake_up(&fs_info->scrub_pause_wait);
1251                         mutex_lock(&fs_info->scrub_lock);
1252                         while (atomic_read(&fs_info->scrub_pause_req)) {
1253                                 mutex_unlock(&fs_info->scrub_lock);
1254                                 wait_event(fs_info->scrub_pause_wait,
1255                                    atomic_read(&fs_info->scrub_pause_req) == 0);
1256                                 mutex_lock(&fs_info->scrub_lock);
1257                         }
1258                         atomic_dec(&fs_info->scrubs_paused);
1259                         mutex_unlock(&fs_info->scrub_lock);
1260                         wake_up(&fs_info->scrub_pause_wait);
1261                 }
1262
1263                 ret = btrfs_lookup_csums_range(csum_root, logical,
1264                                                logical + map->stripe_len - 1,
1265                                                &sdev->csum_list, 1);
1266                 if (ret)
1267                         goto out;
1268
1269                 key.objectid = logical;
1270                 key.type = BTRFS_EXTENT_ITEM_KEY;
1271                 key.offset = (u64)0;
1272
1273                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1274                 if (ret < 0)
1275                         goto out;
1276                 if (ret > 0) {
1277                         ret = btrfs_previous_item(root, path, 0,
1278                                                   BTRFS_EXTENT_ITEM_KEY);
1279                         if (ret < 0)
1280                                 goto out;
1281                         if (ret > 0) {
1282                                 /* there's no smaller item, so stick with the
1283                                  * larger one */
1284                                 btrfs_release_path(path);
1285                                 ret = btrfs_search_slot(NULL, root, &key,
1286                                                         path, 0, 0);
1287                                 if (ret < 0)
1288                                         goto out;
1289                         }
1290                 }
1291
1292                 while (1) {
1293                         l = path->nodes[0];
1294                         slot = path->slots[0];
1295                         if (slot >= btrfs_header_nritems(l)) {
1296                                 ret = btrfs_next_leaf(root, path);
1297                                 if (ret == 0)
1298                                         continue;
1299                                 if (ret < 0)
1300                                         goto out;
1301
1302                                 break;
1303                         }
1304                         btrfs_item_key_to_cpu(l, &key, slot);
1305
1306                         if (key.objectid + key.offset <= logical)
1307                                 goto next;
1308
1309                         if (key.objectid >= logical + map->stripe_len)
1310                                 break;
1311
1312                         if (btrfs_key_type(&key) != BTRFS_EXTENT_ITEM_KEY)
1313                                 goto next;
1314
1315                         extent = btrfs_item_ptr(l, slot,
1316                                                 struct btrfs_extent_item);
1317                         flags = btrfs_extent_flags(l, extent);
1318                         generation = btrfs_extent_generation(l, extent);
1319
1320                         if (key.objectid < logical &&
1321                             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
1322                                 printk(KERN_ERR
1323                                        "btrfs scrub: tree block %llu spanning "
1324                                        "stripes, ignored. logical=%llu\n",
1325                                        (unsigned long long)key.objectid,
1326                                        (unsigned long long)logical);
1327                                 goto next;
1328                         }
1329
1330                         /*
1331                          * trim extent to this stripe
1332                          */
1333                         if (key.objectid < logical) {
1334                                 key.offset -= logical - key.objectid;
1335                                 key.objectid = logical;
1336                         }
1337                         if (key.objectid + key.offset >
1338                             logical + map->stripe_len) {
1339                                 key.offset = logical + map->stripe_len -
1340                                              key.objectid;
1341                         }
1342
1343                         ret = scrub_extent(sdev, key.objectid, key.offset,
1344                                            key.objectid - logical + physical,
1345                                            flags, generation, mirror_num);
1346                         if (ret)
1347                                 goto out;
1348
1349 next:
1350                         path->slots[0]++;
1351                 }
1352                 btrfs_release_path(path);
1353                 logical += increment;
1354                 physical += map->stripe_len;
1355                 spin_lock(&sdev->stat_lock);
1356                 sdev->stat.last_physical = physical;
1357                 spin_unlock(&sdev->stat_lock);
1358         }
1359         /* push queued extents */
1360         scrub_submit(sdev);
1361
1362 out:
1363         blk_finish_plug(&plug);
1364         btrfs_free_path(path);
1365         return ret < 0 ? ret : 0;
1366 }
1367
1368 static noinline_for_stack int scrub_chunk(struct scrub_dev *sdev,
1369         u64 chunk_tree, u64 chunk_objectid, u64 chunk_offset, u64 length)
1370 {
1371         struct btrfs_mapping_tree *map_tree =
1372                 &sdev->dev->dev_root->fs_info->mapping_tree;
1373         struct map_lookup *map;
1374         struct extent_map *em;
1375         int i;
1376         int ret = -EINVAL;
1377
1378         read_lock(&map_tree->map_tree.lock);
1379         em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
1380         read_unlock(&map_tree->map_tree.lock);
1381
1382         if (!em)
1383                 return -EINVAL;
1384
1385         map = (struct map_lookup *)em->bdev;
1386         if (em->start != chunk_offset)
1387                 goto out;
1388
1389         if (em->len < length)
1390                 goto out;
1391
1392         for (i = 0; i < map->num_stripes; ++i) {
1393                 if (map->stripes[i].dev == sdev->dev) {
1394                         ret = scrub_stripe(sdev, map, i, chunk_offset, length);
1395                         if (ret)
1396                                 goto out;
1397                 }
1398         }
1399 out:
1400         free_extent_map(em);
1401
1402         return ret;
1403 }
1404
1405 static noinline_for_stack
1406 int scrub_enumerate_chunks(struct scrub_dev *sdev, u64 start, u64 end)
1407 {
1408         struct btrfs_dev_extent *dev_extent = NULL;
1409         struct btrfs_path *path;
1410         struct btrfs_root *root = sdev->dev->dev_root;
1411         struct btrfs_fs_info *fs_info = root->fs_info;
1412         u64 length;
1413         u64 chunk_tree;
1414         u64 chunk_objectid;
1415         u64 chunk_offset;
1416         int ret;
1417         int slot;
1418         struct extent_buffer *l;
1419         struct btrfs_key key;
1420         struct btrfs_key found_key;
1421         struct btrfs_block_group_cache *cache;
1422
1423         path = btrfs_alloc_path();
1424         if (!path)
1425                 return -ENOMEM;
1426
1427         path->reada = 2;
1428         path->search_commit_root = 1;
1429         path->skip_locking = 1;
1430
1431         key.objectid = sdev->dev->devid;
1432         key.offset = 0ull;
1433         key.type = BTRFS_DEV_EXTENT_KEY;
1434
1435
1436         while (1) {
1437                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1438                 if (ret < 0)
1439                         break;
1440                 if (ret > 0) {
1441                         if (path->slots[0] >=
1442                             btrfs_header_nritems(path->nodes[0])) {
1443                                 ret = btrfs_next_leaf(root, path);
1444                                 if (ret)
1445                                         break;
1446                         }
1447                 }
1448
1449                 l = path->nodes[0];
1450                 slot = path->slots[0];
1451
1452                 btrfs_item_key_to_cpu(l, &found_key, slot);
1453
1454                 if (found_key.objectid != sdev->dev->devid)
1455                         break;
1456
1457                 if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
1458                         break;
1459
1460                 if (found_key.offset >= end)
1461                         break;
1462
1463                 if (found_key.offset < key.offset)
1464                         break;
1465
1466                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
1467                 length = btrfs_dev_extent_length(l, dev_extent);
1468
1469                 if (found_key.offset + length <= start) {
1470                         key.offset = found_key.offset + length;
1471                         btrfs_release_path(path);
1472                         continue;
1473                 }
1474
1475                 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
1476                 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
1477                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
1478
1479                 /*
1480                  * get a reference on the corresponding block group to prevent
1481                  * the chunk from going away while we scrub it
1482                  */
1483                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
1484                 if (!cache) {
1485                         ret = -ENOENT;
1486                         break;
1487                 }
1488                 ret = scrub_chunk(sdev, chunk_tree, chunk_objectid,
1489                                   chunk_offset, length);
1490                 btrfs_put_block_group(cache);
1491                 if (ret)
1492                         break;
1493
1494                 key.offset = found_key.offset + length;
1495                 btrfs_release_path(path);
1496         }
1497
1498         btrfs_free_path(path);
1499
1500         /*
1501          * ret can still be 1 from search_slot or next_leaf,
1502          * that's not an error
1503          */
1504         return ret < 0 ? ret : 0;
1505 }
1506
1507 static noinline_for_stack int scrub_supers(struct scrub_dev *sdev)
1508 {
1509         int     i;
1510         u64     bytenr;
1511         u64     gen;
1512         int     ret;
1513         struct btrfs_device *device = sdev->dev;
1514         struct btrfs_root *root = device->dev_root;
1515
1516         gen = root->fs_info->last_trans_committed;
1517
1518         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1519                 bytenr = btrfs_sb_offset(i);
1520                 if (bytenr + BTRFS_SUPER_INFO_SIZE >= device->total_bytes)
1521                         break;
1522
1523                 ret = scrub_page(sdev, bytenr, PAGE_SIZE, bytenr,
1524                                  BTRFS_EXTENT_FLAG_SUPER, gen, i, NULL, 1);
1525                 if (ret)
1526                         return ret;
1527         }
1528         wait_event(sdev->list_wait, atomic_read(&sdev->in_flight) == 0);
1529
1530         return 0;
1531 }
1532
1533 /*
1534  * get a reference count on fs_info->scrub_workers. start worker if necessary
1535  */
1536 static noinline_for_stack int scrub_workers_get(struct btrfs_root *root)
1537 {
1538         struct btrfs_fs_info *fs_info = root->fs_info;
1539         int ret = 0;
1540
1541         mutex_lock(&fs_info->scrub_lock);
1542         if (fs_info->scrub_workers_refcnt == 0) {
1543                 btrfs_init_workers(&fs_info->scrub_workers, "scrub",
1544                            fs_info->thread_pool_size, &fs_info->generic_worker);
1545                 fs_info->scrub_workers.idle_thresh = 4;
1546                 ret = btrfs_start_workers(&fs_info->scrub_workers);
1547                 if (ret)
1548                         goto out;
1549         }
1550         ++fs_info->scrub_workers_refcnt;
1551 out:
1552         mutex_unlock(&fs_info->scrub_lock);
1553
1554         return ret;
1555 }
1556
1557 static noinline_for_stack void scrub_workers_put(struct btrfs_root *root)
1558 {
1559         struct btrfs_fs_info *fs_info = root->fs_info;
1560
1561         mutex_lock(&fs_info->scrub_lock);
1562         if (--fs_info->scrub_workers_refcnt == 0)
1563                 btrfs_stop_workers(&fs_info->scrub_workers);
1564         WARN_ON(fs_info->scrub_workers_refcnt < 0);
1565         mutex_unlock(&fs_info->scrub_lock);
1566 }
1567
1568
1569 int btrfs_scrub_dev(struct btrfs_root *root, u64 devid, u64 start, u64 end,
1570                     struct btrfs_scrub_progress *progress, int readonly)
1571 {
1572         struct scrub_dev *sdev;
1573         struct btrfs_fs_info *fs_info = root->fs_info;
1574         int ret;
1575         struct btrfs_device *dev;
1576
1577         if (btrfs_fs_closing(root->fs_info))
1578                 return -EINVAL;
1579
1580         /*
1581          * check some assumptions
1582          */
1583         if (root->sectorsize != PAGE_SIZE ||
1584             root->sectorsize != root->leafsize ||
1585             root->sectorsize != root->nodesize) {
1586                 printk(KERN_ERR "btrfs_scrub: size assumptions fail\n");
1587                 return -EINVAL;
1588         }
1589
1590         ret = scrub_workers_get(root);
1591         if (ret)
1592                 return ret;
1593
1594         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
1595         dev = btrfs_find_device(root, devid, NULL, NULL);
1596         if (!dev || dev->missing) {
1597                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1598                 scrub_workers_put(root);
1599                 return -ENODEV;
1600         }
1601         mutex_lock(&fs_info->scrub_lock);
1602
1603         if (!dev->in_fs_metadata) {
1604                 mutex_unlock(&fs_info->scrub_lock);
1605                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1606                 scrub_workers_put(root);
1607                 return -ENODEV;
1608         }
1609
1610         if (dev->scrub_device) {
1611                 mutex_unlock(&fs_info->scrub_lock);
1612                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1613                 scrub_workers_put(root);
1614                 return -EINPROGRESS;
1615         }
1616         sdev = scrub_setup_dev(dev);
1617         if (IS_ERR(sdev)) {
1618                 mutex_unlock(&fs_info->scrub_lock);
1619                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1620                 scrub_workers_put(root);
1621                 return PTR_ERR(sdev);
1622         }
1623         sdev->readonly = readonly;
1624         dev->scrub_device = sdev;
1625
1626         atomic_inc(&fs_info->scrubs_running);
1627         mutex_unlock(&fs_info->scrub_lock);
1628         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1629
1630         down_read(&fs_info->scrub_super_lock);
1631         ret = scrub_supers(sdev);
1632         up_read(&fs_info->scrub_super_lock);
1633
1634         if (!ret)
1635                 ret = scrub_enumerate_chunks(sdev, start, end);
1636
1637         wait_event(sdev->list_wait, atomic_read(&sdev->in_flight) == 0);
1638         atomic_dec(&fs_info->scrubs_running);
1639         wake_up(&fs_info->scrub_pause_wait);
1640
1641         wait_event(sdev->list_wait, atomic_read(&sdev->fixup_cnt) == 0);
1642
1643         if (progress)
1644                 memcpy(progress, &sdev->stat, sizeof(*progress));
1645
1646         mutex_lock(&fs_info->scrub_lock);
1647         dev->scrub_device = NULL;
1648         mutex_unlock(&fs_info->scrub_lock);
1649
1650         scrub_free_dev(sdev);
1651         scrub_workers_put(root);
1652
1653         return ret;
1654 }
1655
1656 int btrfs_scrub_pause(struct btrfs_root *root)
1657 {
1658         struct btrfs_fs_info *fs_info = root->fs_info;
1659
1660         mutex_lock(&fs_info->scrub_lock);
1661         atomic_inc(&fs_info->scrub_pause_req);
1662         while (atomic_read(&fs_info->scrubs_paused) !=
1663                atomic_read(&fs_info->scrubs_running)) {
1664                 mutex_unlock(&fs_info->scrub_lock);
1665                 wait_event(fs_info->scrub_pause_wait,
1666                            atomic_read(&fs_info->scrubs_paused) ==
1667                            atomic_read(&fs_info->scrubs_running));
1668                 mutex_lock(&fs_info->scrub_lock);
1669         }
1670         mutex_unlock(&fs_info->scrub_lock);
1671
1672         return 0;
1673 }
1674
1675 int btrfs_scrub_continue(struct btrfs_root *root)
1676 {
1677         struct btrfs_fs_info *fs_info = root->fs_info;
1678
1679         atomic_dec(&fs_info->scrub_pause_req);
1680         wake_up(&fs_info->scrub_pause_wait);
1681         return 0;
1682 }
1683
1684 int btrfs_scrub_pause_super(struct btrfs_root *root)
1685 {
1686         down_write(&root->fs_info->scrub_super_lock);
1687         return 0;
1688 }
1689
1690 int btrfs_scrub_continue_super(struct btrfs_root *root)
1691 {
1692         up_write(&root->fs_info->scrub_super_lock);
1693         return 0;
1694 }
1695
1696 int btrfs_scrub_cancel(struct btrfs_root *root)
1697 {
1698         struct btrfs_fs_info *fs_info = root->fs_info;
1699
1700         mutex_lock(&fs_info->scrub_lock);
1701         if (!atomic_read(&fs_info->scrubs_running)) {
1702                 mutex_unlock(&fs_info->scrub_lock);
1703                 return -ENOTCONN;
1704         }
1705
1706         atomic_inc(&fs_info->scrub_cancel_req);
1707         while (atomic_read(&fs_info->scrubs_running)) {
1708                 mutex_unlock(&fs_info->scrub_lock);
1709                 wait_event(fs_info->scrub_pause_wait,
1710                            atomic_read(&fs_info->scrubs_running) == 0);
1711                 mutex_lock(&fs_info->scrub_lock);
1712         }
1713         atomic_dec(&fs_info->scrub_cancel_req);
1714         mutex_unlock(&fs_info->scrub_lock);
1715
1716         return 0;
1717 }
1718
1719 int btrfs_scrub_cancel_dev(struct btrfs_root *root, struct btrfs_device *dev)
1720 {
1721         struct btrfs_fs_info *fs_info = root->fs_info;
1722         struct scrub_dev *sdev;
1723
1724         mutex_lock(&fs_info->scrub_lock);
1725         sdev = dev->scrub_device;
1726         if (!sdev) {
1727                 mutex_unlock(&fs_info->scrub_lock);
1728                 return -ENOTCONN;
1729         }
1730         atomic_inc(&sdev->cancel_req);
1731         while (dev->scrub_device) {
1732                 mutex_unlock(&fs_info->scrub_lock);
1733                 wait_event(fs_info->scrub_pause_wait,
1734                            dev->scrub_device == NULL);
1735                 mutex_lock(&fs_info->scrub_lock);
1736         }
1737         mutex_unlock(&fs_info->scrub_lock);
1738
1739         return 0;
1740 }
1741 int btrfs_scrub_cancel_devid(struct btrfs_root *root, u64 devid)
1742 {
1743         struct btrfs_fs_info *fs_info = root->fs_info;
1744         struct btrfs_device *dev;
1745         int ret;
1746
1747         /*
1748          * we have to hold the device_list_mutex here so the device
1749          * does not go away in cancel_dev. FIXME: find a better solution
1750          */
1751         mutex_lock(&fs_info->fs_devices->device_list_mutex);
1752         dev = btrfs_find_device(root, devid, NULL, NULL);
1753         if (!dev) {
1754                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
1755                 return -ENODEV;
1756         }
1757         ret = btrfs_scrub_cancel_dev(root, dev);
1758         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
1759
1760         return ret;
1761 }
1762
1763 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
1764                          struct btrfs_scrub_progress *progress)
1765 {
1766         struct btrfs_device *dev;
1767         struct scrub_dev *sdev = NULL;
1768
1769         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
1770         dev = btrfs_find_device(root, devid, NULL, NULL);
1771         if (dev)
1772                 sdev = dev->scrub_device;
1773         if (sdev)
1774                 memcpy(progress, &sdev->stat, sizeof(*progress));
1775         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1776
1777         return dev ? (sdev ? 0 : -ENOTCONN) : -ENODEV;
1778 }