fs/btrfs/scrub.c

   1 /*
   2  * Copyright (C) 2011 STRATO.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/blkdev.h>
  20 #include <linux/ratelimit.h>
  21 #include "ctree.h"
  22 #include "volumes.h"
  23 #include "disk-io.h"
  24 #include "ordered-data.h"
  25 #include "transaction.h"
  26 #include "backref.h"
  27 #include "extent_io.h"
  28 #include "check-integrity.h"
  29
  30 /*
  31  * This is only the first step towards a full-features scrub. It reads all
  32  * extent and super block and verifies the checksums. In case a bad checksum
  33  * is found or the extent cannot be read, good data will be written back if
  34  * any can be found.
  35  *
  36  * Future enhancements:
  37  *  - In case an unrepairable extent is encountered, track which files are
  38  *    affected and report them
  39  *  - track and record media errors, throw out bad devices
  40  *  - add a mode to also read unallocated space
  41  */
  42
  43 struct scrub_block;
  44 struct scrub_dev;
  45
  46 #define SCRUB_PAGES_PER_BIO     16      /* 64k per bio */
  47 #define SCRUB_BIOS_PER_DEV      16      /* 1 MB per device in flight */
  48 #define SCRUB_MAX_PAGES_PER_BLOCK       16      /* 64k per node/leaf/sector */
  49
  50 struct scrub_page {
  51         struct scrub_block      *sblock;
  52         struct page             *page;
  53         struct block_device     *bdev;
  54         u64                     flags;  /* extent flags */
  55         u64                     generation;
  56         u64                     logical;
  57         u64                     physical;
  58         struct {
  59                 unsigned int    mirror_num:8;
  60                 unsigned int    have_csum:1;
  61                 unsigned int    io_error:1;
  62         };
  63         u8                      csum[BTRFS_CSUM_SIZE];
  64 };
  65
  66 struct scrub_bio {
  67         int                     index;
  68         struct scrub_dev        *sdev;
  69         struct bio              *bio;
  70         int                     err;
  71         u64                     logical;
  72         u64                     physical;
  73         struct scrub_page       *pagev[SCRUB_PAGES_PER_BIO];
  74         int                     page_count;
  75         int                     next_free;
  76         struct btrfs_work       work;
  77 };
  78
  79 struct scrub_block {
  80         struct scrub_page       pagev[SCRUB_MAX_PAGES_PER_BLOCK];
  81         int                     page_count;
  82         atomic_t                outstanding_pages;
  83         atomic_t                ref_count; /* free mem on transition to zero */
  84         struct scrub_dev        *sdev;
  85         struct {
  86                 unsigned int    header_error:1;
  87                 unsigned int    checksum_error:1;
  88                 unsigned int    no_io_error_seen:1;
  89         };
  90 };
  91
  92 struct scrub_dev {
  93         struct scrub_bio        *bios[SCRUB_BIOS_PER_DEV];
  94         struct btrfs_device     *dev;
  95         int                     first_free;
  96         int                     curr;
  97         atomic_t                in_flight;
  98         atomic_t                fixup_cnt;
  99         spinlock_t              list_lock;
 100         wait_queue_head_t       list_wait;
 101         u16                     csum_size;
 102         struct list_head        csum_list;
 103         atomic_t                cancel_req;
 104         int                     readonly;
 105         int                     pages_per_bio; /* <= SCRUB_PAGES_PER_BIO */
 106         u32                     sectorsize;
 107         u32                     nodesize;
 108         u32                     leafsize;
 109         /*
 110          * statistics
 111          */
 112         struct btrfs_scrub_progress stat;
 113         spinlock_t              stat_lock;
 114 };
 115
 116 struct scrub_fixup_nodatasum {
 117         struct scrub_dev        *sdev;
 118         u64                     logical;
 119         struct btrfs_root       *root;
 120         struct btrfs_work       work;
 121         int                     mirror_num;
 122 };
 123
 124 struct scrub_warning {
 125         struct btrfs_path       *path;
 126         u64                     extent_item_size;
 127         char                    *scratch_buf;
 128         char                    *msg_buf;
 129         const char              *errstr;
 130         sector_t                sector;
 131         u64                     logical;
 132         struct btrfs_device     *dev;
 133         int                     msg_bufsize;
 134         int                     scratch_bufsize;
 135 };
 136
 137
 138 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
 139 static int scrub_setup_recheck_block(struct scrub_dev *sdev,
 140                                      struct btrfs_mapping_tree *map_tree,
 141                                      u64 length, u64 logical,
 142                                      struct scrub_block *sblock);
 143 static int scrub_recheck_block(struct btrfs_fs_info *fs_info,
 144                                struct scrub_block *sblock, int is_metadata,
 145                                int have_csum, u8 *csum, u64 generation,
 146                                u16 csum_size);
 147 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
 148                                          struct scrub_block *sblock,
 149                                          int is_metadata, int have_csum,
 150                                          const u8 *csum, u64 generation,
 151                                          u16 csum_size);
 152 static void scrub_complete_bio_end_io(struct bio *bio, int err);
 153 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
 154                                              struct scrub_block *sblock_good,
 155                                              int force_write);
 156 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
 157                                             struct scrub_block *sblock_good,
 158                                             int page_num, int force_write);
 159 static int scrub_checksum_data(struct scrub_block *sblock);
 160 static int scrub_checksum_tree_block(struct scrub_block *sblock);
 161 static int scrub_checksum_super(struct scrub_block *sblock);
 162 static void scrub_block_get(struct scrub_block *sblock);
 163 static void scrub_block_put(struct scrub_block *sblock);
 164 static int scrub_add_page_to_bio(struct scrub_dev *sdev,
 165                                  struct scrub_page *spage);
 166 static int scrub_pages(struct scrub_dev *sdev, u64 logical, u64 len,
 167                        u64 physical, u64 flags, u64 gen, int mirror_num,
 168                        u8 *csum, int force);
 169 static void scrub_bio_end_io(struct bio *bio, int err);
 170 static void scrub_bio_end_io_worker(struct btrfs_work *work);
 171 static void scrub_block_complete(struct scrub_block *sblock);
 172
 173
 174 static void scrub_free_csums(struct scrub_dev *sdev)
 175 {
 176         while (!list_empty(&sdev->csum_list)) {
 177                 struct btrfs_ordered_sum *sum;
 178                 sum = list_first_entry(&sdev->csum_list,
 179                                        struct btrfs_ordered_sum, list);
 180                 list_del(&sum->list);
 181                 kfree(sum);
 182         }
 183 }
 184
 185 static noinline_for_stack void scrub_free_dev(struct scrub_dev *sdev)
 186 {
 187         int i;
 188
 189         if (!sdev)
 190                 return;
 191
 192         /* this can happen when scrub is cancelled */
 193         if (sdev->curr != -1) {
 194                 struct scrub_bio *sbio = sdev->bios[sdev->curr];
 195
 196                 for (i = 0; i < sbio->page_count; i++) {
 197                         BUG_ON(!sbio->pagev[i]);
 198                         BUG_ON(!sbio->pagev[i]->page);
 199                         scrub_block_put(sbio->pagev[i]->sblock);
 200                 }
 201                 bio_put(sbio->bio);
 202         }
 203
 204         for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
 205                 struct scrub_bio *sbio = sdev->bios[i];
 206
 207                 if (!sbio)
 208                         break;
 209                 kfree(sbio);
 210         }
 211
 212         scrub_free_csums(sdev);
 213         kfree(sdev);
 214 }
 215
 216 static noinline_for_stack
 217 struct scrub_dev *scrub_setup_dev(struct btrfs_device *dev)
 218 {
 219         struct scrub_dev *sdev;
 220         int             i;
 221         struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
 222         int pages_per_bio;
 223
 224         pages_per_bio = min_t(int, SCRUB_PAGES_PER_BIO,
 225                               bio_get_nr_vecs(dev->bdev));
 226         sdev = kzalloc(sizeof(*sdev), GFP_NOFS);
 227         if (!sdev)
 228                 goto nomem;
 229         sdev->dev = dev;
 230         sdev->pages_per_bio = pages_per_bio;
 231         sdev->curr = -1;
 232         for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
 233                 struct scrub_bio *sbio;
 234
 235                 sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
 236                 if (!sbio)
 237                         goto nomem;
 238                 sdev->bios[i] = sbio;
 239
 240                 sbio->index = i;
 241                 sbio->sdev = sdev;
 242                 sbio->page_count = 0;
 243                 sbio->work.func = scrub_bio_end_io_worker;
 244
 245                 if (i != SCRUB_BIOS_PER_DEV-1)
 246                         sdev->bios[i]->next_free = i + 1;
 247                 else
 248                         sdev->bios[i]->next_free = -1;
 249         }
 250         sdev->first_free = 0;
 251         sdev->nodesize = dev->dev_root->nodesize;
 252         sdev->leafsize = dev->dev_root->leafsize;
 253         sdev->sectorsize = dev->dev_root->sectorsize;
 254         atomic_set(&sdev->in_flight, 0);
 255         atomic_set(&sdev->fixup_cnt, 0);
 256         atomic_set(&sdev->cancel_req, 0);
 257         sdev->csum_size = btrfs_super_csum_size(fs_info->super_copy);
 258         INIT_LIST_HEAD(&sdev->csum_list);
 259
 260         spin_lock_init(&sdev->list_lock);
 261         spin_lock_init(&sdev->stat_lock);
 262         init_waitqueue_head(&sdev->list_wait);
 263         return sdev;
 264
 265 nomem:
 266         scrub_free_dev(sdev);
 267         return ERR_PTR(-ENOMEM);
 268 }
 269
 270 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root, void *ctx)
 271 {
 272         u64 isize;
 273         u32 nlink;
 274         int ret;
 275         int i;
 276         struct extent_buffer *eb;
 277         struct btrfs_inode_item *inode_item;
 278         struct scrub_warning *swarn = ctx;
 279         struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
 280         struct inode_fs_paths *ipath = NULL;
 281         struct btrfs_root *local_root;
 282         struct btrfs_key root_key;
 283
 284         root_key.objectid = root;
 285         root_key.type = BTRFS_ROOT_ITEM_KEY;
 286         root_key.offset = (u64)-1;
 287         local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
 288         if (IS_ERR(local_root)) {
 289                 ret = PTR_ERR(local_root);
 290                 goto err;
 291         }
 292
 293         ret = inode_item_info(inum, 0, local_root, swarn->path);
 294         if (ret) {
 295                 btrfs_release_path(swarn->path);
 296                 goto err;
 297         }
 298
 299         eb = swarn->path->nodes[0];
 300         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 301                                         struct btrfs_inode_item);
 302         isize = btrfs_inode_size(eb, inode_item);
 303         nlink = btrfs_inode_nlink(eb, inode_item);
 304         btrfs_release_path(swarn->path);
 305
 306         ipath = init_ipath(4096, local_root, swarn->path);
 307         if (IS_ERR(ipath)) {
 308                 ret = PTR_ERR(ipath);
 309                 ipath = NULL;
 310                 goto err;
 311         }
 312         ret = paths_from_inode(inum, ipath);
 313
 314         if (ret < 0)
 315                 goto err;
 316
 317         /*
 318          * we deliberately ignore the bit ipath might have been too small to
 319          * hold all of the paths here
 320          */
 321         for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 322                 printk(KERN_WARNING "btrfs: %s at logical %llu on dev "
 323                         "%s, sector %llu, root %llu, inode %llu, offset %llu, "
 324                         "length %llu, links %u (path: %s)\n", swarn->errstr,
 325                         swarn->logical, swarn->dev->name,
 326                         (unsigned long long)swarn->sector, root, inum, offset,
 327                         min(isize - offset, (u64)PAGE_SIZE), nlink,
 328                         (char *)(unsigned long)ipath->fspath->val[i]);
 329
 330         free_ipath(ipath);
 331         return 0;
 332
 333 err:
 334         printk(KERN_WARNING "btrfs: %s at logical %llu on dev "
 335                 "%s, sector %llu, root %llu, inode %llu, offset %llu: path "
 336                 "resolving failed with ret=%d\n", swarn->errstr,
 337                 swarn->logical, swarn->dev->name,
 338                 (unsigned long long)swarn->sector, root, inum, offset, ret);
 339
 340         free_ipath(ipath);
 341         return 0;
 342 }
 343
 344 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
 345 {
 346         struct btrfs_device *dev = sblock->sdev->dev;
 347         struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
 348         struct btrfs_path *path;
 349         struct btrfs_key found_key;
 350         struct extent_buffer *eb;
 351         struct btrfs_extent_item *ei;
 352         struct scrub_warning swarn;
 353         u32 item_size;
 354         int ret;
 355         u64 ref_root;
 356         u8 ref_level;
 357         unsigned long ptr = 0;
 358         const int bufsize = 4096;
 359         u64 extent_item_pos;
 360
 361         path = btrfs_alloc_path();
 362
 363         swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
 364         swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
 365         BUG_ON(sblock->page_count < 1);
 366         swarn.sector = (sblock->pagev[0].physical) >> 9;
 367         swarn.logical = sblock->pagev[0].logical;
 368         swarn.errstr = errstr;
 369         swarn.dev = dev;
 370         swarn.msg_bufsize = bufsize;
 371         swarn.scratch_bufsize = bufsize;
 372
 373         if (!path || !swarn.scratch_buf || !swarn.msg_buf)
 374                 goto out;
 375
 376         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key);
 377         if (ret < 0)
 378                 goto out;
 379
 380         extent_item_pos = swarn.logical - found_key.objectid;
 381         swarn.extent_item_size = found_key.offset;
 382
 383         eb = path->nodes[0];
 384         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 385         item_size = btrfs_item_size_nr(eb, path->slots[0]);
 386         btrfs_release_path(path);
 387
 388         if (ret & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 389                 do {
 390                         ret = tree_backref_for_extent(&ptr, eb, ei, item_size,
 391                                                         &ref_root, &ref_level);
 392                         printk(KERN_WARNING
 393                                 "btrfs: %s at logical %llu on dev %s, "
 394                                 "sector %llu: metadata %s (level %d) in tree "
 395                                 "%llu\n", errstr, swarn.logical, dev->name,
 396                                 (unsigned long long)swarn.sector,
 397                                 ref_level ? "node" : "leaf",
 398                                 ret < 0 ? -1 : ref_level,
 399                                 ret < 0 ? -1 : ref_root);
 400                 } while (ret != 1);
 401         } else {
 402                 swarn.path = path;
 403                 iterate_extent_inodes(fs_info, found_key.objectid,
 404                                         extent_item_pos, 1,
 405                                         scrub_print_warning_inode, &swarn);
 406         }
 407
 408 out:
 409         btrfs_free_path(path);
 410         kfree(swarn.scratch_buf);
 411         kfree(swarn.msg_buf);
 412 }
 413
 414 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *ctx)
 415 {
 416         struct page *page = NULL;
 417         unsigned long index;
 418         struct scrub_fixup_nodatasum *fixup = ctx;
 419         int ret;
 420         int corrected = 0;
 421         struct btrfs_key key;
 422         struct inode *inode = NULL;
 423         u64 end = offset + PAGE_SIZE - 1;
 424         struct btrfs_root *local_root;
 425
 426         key.objectid = root;
 427         key.type = BTRFS_ROOT_ITEM_KEY;
 428         key.offset = (u64)-1;
 429         local_root = btrfs_read_fs_root_no_name(fixup->root->fs_info, &key);
 430         if (IS_ERR(local_root))
 431                 return PTR_ERR(local_root);
 432
 433         key.type = BTRFS_INODE_ITEM_KEY;
 434         key.objectid = inum;
 435         key.offset = 0;
 436         inode = btrfs_iget(fixup->root->fs_info->sb, &key, local_root, NULL);
 437         if (IS_ERR(inode))
 438                 return PTR_ERR(inode);
 439
 440         index = offset >> PAGE_CACHE_SHIFT;
 441
 442         page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
 443         if (!page) {
 444                 ret = -ENOMEM;
 445                 goto out;
 446         }
 447
 448         if (PageUptodate(page)) {
 449                 struct btrfs_mapping_tree *map_tree;
 450                 if (PageDirty(page)) {
 451                         /*
 452                          * we need to write the data to the defect sector. the
 453                          * data that was in that sector is not in memory,
 454                          * because the page was modified. we must not write the
 455                          * modified page to that sector.
 456                          *
 457                          * TODO: what could be done here: wait for the delalloc
 458                          *       runner to write out that page (might involve
 459                          *       COW) and see whether the sector is still
 460                          *       referenced afterwards.
 461                          *
 462                          * For the meantime, we'll treat this error
 463                          * incorrectable, although there is a chance that a
 464                          * later scrub will find the bad sector again and that
 465                          * there's no dirty page in memory, then.
 466                          */
 467                         ret = -EIO;
 468                         goto out;
 469                 }
 470                 map_tree = &BTRFS_I(inode)->root->fs_info->mapping_tree;
 471                 ret = repair_io_failure(map_tree, offset, PAGE_SIZE,
 472                                         fixup->logical, page,
 473                                         fixup->mirror_num);
 474                 unlock_page(page);
 475                 corrected = !ret;
 476         } else {
 477                 /*
 478                  * we need to get good data first. the general readpage path
 479                  * will call repair_io_failure for us, we just have to make
 480                  * sure we read the bad mirror.
 481                  */
 482                 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 483                                         EXTENT_DAMAGED, GFP_NOFS);
 484                 if (ret) {
 485                         /* set_extent_bits should give proper error */
 486                         WARN_ON(ret > 0);
 487                         if (ret > 0)
 488                                 ret = -EFAULT;
 489                         goto out;
 490                 }
 491
 492                 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
 493                                                 btrfs_get_extent,
 494                                                 fixup->mirror_num);
 495                 wait_on_page_locked(page);
 496
 497                 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
 498                                                 end, EXTENT_DAMAGED, 0, NULL);
 499                 if (!corrected)
 500                         clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 501                                                 EXTENT_DAMAGED, GFP_NOFS);
 502         }
 503
 504 out:
 505         if (page)
 506                 put_page(page);
 507         if (inode)
 508                 iput(inode);
 509
 510         if (ret < 0)
 511                 return ret;
 512
 513         if (ret == 0 && corrected) {
 514                 /*
 515                  * we only need to call readpage for one of the inodes belonging
 516                  * to this extent. so make iterate_extent_inodes stop
 517                  */
 518                 return 1;
 519         }
 520
 521         return -EIO;
 522 }
 523
 524 static void scrub_fixup_nodatasum(struct btrfs_work *work)
 525 {
 526         int ret;
 527         struct scrub_fixup_nodatasum *fixup;
 528         struct scrub_dev *sdev;
 529         struct btrfs_trans_handle *trans = NULL;
 530         struct btrfs_fs_info *fs_info;
 531         struct btrfs_path *path;
 532         int uncorrectable = 0;
 533
 534         fixup = container_of(work, struct scrub_fixup_nodatasum, work);
 535         sdev = fixup->sdev;
 536         fs_info = fixup->root->fs_info;
 537
 538         path = btrfs_alloc_path();
 539         if (!path) {
 540                 spin_lock(&sdev->stat_lock);
 541                 ++sdev->stat.malloc_errors;
 542                 spin_unlock(&sdev->stat_lock);
 543                 uncorrectable = 1;
 544                 goto out;
 545         }
 546
 547         trans = btrfs_join_transaction(fixup->root);
 548         if (IS_ERR(trans)) {
 549                 uncorrectable = 1;
 550                 goto out;
 551         }
 552
 553         /*
 554          * the idea is to trigger a regular read through the standard path. we
 555          * read a page from the (failed) logical address by specifying the
 556          * corresponding copynum of the failed sector. thus, that readpage is
 557          * expected to fail.
 558          * that is the point where on-the-fly error correction will kick in
 559          * (once it's finished) and rewrite the failed sector if a good copy
 560          * can be found.
 561          */
 562         ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
 563                                                 path, scrub_fixup_readpage,
 564                                                 fixup);
 565         if (ret < 0) {
 566                 uncorrectable = 1;
 567                 goto out;
 568         }
 569         WARN_ON(ret != 1);
 570
 571         spin_lock(&sdev->stat_lock);
 572         ++sdev->stat.corrected_errors;
 573         spin_unlock(&sdev->stat_lock);
 574
 575 out:
 576         if (trans && !IS_ERR(trans))
 577                 btrfs_end_transaction(trans, fixup->root);
 578         if (uncorrectable) {
 579                 spin_lock(&sdev->stat_lock);
 580                 ++sdev->stat.uncorrectable_errors;
 581                 spin_unlock(&sdev->stat_lock);
 582                 printk_ratelimited(KERN_ERR
 583                         "btrfs: unable to fixup (nodatasum) error at logical %llu on dev %s\n",
 584                         (unsigned long long)fixup->logical, sdev->dev->name);
 585         }
 586
 587         btrfs_free_path(path);
 588         kfree(fixup);
 589
 590         /* see caller why we're pretending to be paused in the scrub counters */
 591         mutex_lock(&fs_info->scrub_lock);
 592         atomic_dec(&fs_info->scrubs_running);
 593         atomic_dec(&fs_info->scrubs_paused);
 594         mutex_unlock(&fs_info->scrub_lock);
 595         atomic_dec(&sdev->fixup_cnt);
 596         wake_up(&fs_info->scrub_pause_wait);
 597         wake_up(&sdev->list_wait);
 598 }
 599
 600 /*
 601  * scrub_handle_errored_block gets called when either verification of the
 602  * pages failed or the bio failed to read, e.g. with EIO. In the latter
 603  * case, this function handles all pages in the bio, even though only one
 604  * may be bad.
 605  * The goal of this function is to repair the errored block by using the
 606  * contents of one of the mirrors.
 607  */
 608 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
 609 {
 610         struct scrub_dev *sdev = sblock_to_check->sdev;
 611         struct btrfs_fs_info *fs_info;
 612         u64 length;
 613         u64 logical;
 614         u64 generation;
 615         unsigned int failed_mirror_index;
 616         unsigned int is_metadata;
 617         unsigned int have_csum;
 618         u8 *csum;
 619         struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
 620         struct scrub_block *sblock_bad;
 621         int ret;
 622         int mirror_index;
 623         int page_num;
 624         int success;
 625         static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
 626                                       DEFAULT_RATELIMIT_BURST);
 627
 628         BUG_ON(sblock_to_check->page_count < 1);
 629         fs_info = sdev->dev->dev_root->fs_info;
 630         length = sblock_to_check->page_count * PAGE_SIZE;
 631         logical = sblock_to_check->pagev[0].logical;
 632         generation = sblock_to_check->pagev[0].generation;
 633         BUG_ON(sblock_to_check->pagev[0].mirror_num < 1);
 634         failed_mirror_index = sblock_to_check->pagev[0].mirror_num - 1;
 635         is_metadata = !(sblock_to_check->pagev[0].flags &
 636                         BTRFS_EXTENT_FLAG_DATA);
 637         have_csum = sblock_to_check->pagev[0].have_csum;
 638         csum = sblock_to_check->pagev[0].csum;
 639
 640         /*
 641          * read all mirrors one after the other. This includes to
 642          * re-read the extent or metadata block that failed (that was
 643          * the cause that this fixup code is called) another time,
 644          * page by page this time in order to know which pages
 645          * caused I/O errors and which ones are good (for all mirrors).
 646          * It is the goal to handle the situation when more than one
 647          * mirror contains I/O errors, but the errors do not
 648          * overlap, i.e. the data can be repaired by selecting the
 649          * pages from those mirrors without I/O error on the
 650          * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
 651          * would be that mirror #1 has an I/O error on the first page,
 652          * the second page is good, and mirror #2 has an I/O error on
 653          * the second page, but the first page is good.
 654          * Then the first page of the first mirror can be repaired by
 655          * taking the first page of the second mirror, and the
 656          * second page of the second mirror can be repaired by
 657          * copying the contents of the 2nd page of the 1st mirror.
 658          * One more note: if the pages of one mirror contain I/O
 659          * errors, the checksum cannot be verified. In order to get
 660          * the best data for repairing, the first attempt is to find
 661          * a mirror without I/O errors and with a validated checksum.
 662          * Only if this is not possible, the pages are picked from
 663          * mirrors with I/O errors without considering the checksum.
 664          * If the latter is the case, at the end, the checksum of the
 665          * repaired area is verified in order to correctly maintain
 666          * the statistics.
 667          */
 668
 669         sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
 670                                      sizeof(*sblocks_for_recheck),
 671                                      GFP_NOFS);
 672         if (!sblocks_for_recheck) {
 673                 spin_lock(&sdev->stat_lock);
 674                 sdev->stat.malloc_errors++;
 675                 sdev->stat.read_errors++;
 676                 sdev->stat.uncorrectable_errors++;
 677                 spin_unlock(&sdev->stat_lock);
 678                 goto out;
 679         }
 680
 681         /* setup the context, map the logical blocks and alloc the pages */
 682         ret = scrub_setup_recheck_block(sdev, &fs_info->mapping_tree, length,
 683                                         logical, sblocks_for_recheck);
 684         if (ret) {
 685                 spin_lock(&sdev->stat_lock);
 686                 sdev->stat.read_errors++;
 687                 sdev->stat.uncorrectable_errors++;
 688                 spin_unlock(&sdev->stat_lock);
 689                 goto out;
 690         }
 691         BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
 692         sblock_bad = sblocks_for_recheck + failed_mirror_index;
 693
 694         /* build and submit the bios for the failed mirror, check checksums */
 695         ret = scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
 696                                   csum, generation, sdev->csum_size);
 697         if (ret) {
 698                 spin_lock(&sdev->stat_lock);
 699                 sdev->stat.read_errors++;
 700                 sdev->stat.uncorrectable_errors++;
 701                 spin_unlock(&sdev->stat_lock);
 702                 goto out;
 703         }
 704
 705         if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
 706             sblock_bad->no_io_error_seen) {
 707                 /*
 708                  * the error disappeared after reading page by page, or
 709                  * the area was part of a huge bio and other parts of the
 710                  * bio caused I/O errors, or the block layer merged several
 711                  * read requests into one and the error is caused by a
 712                  * different bio (usually one of the two latter cases is
 713                  * the cause)
 714                  */
 715                 spin_lock(&sdev->stat_lock);
 716                 sdev->stat.unverified_errors++;
 717                 spin_unlock(&sdev->stat_lock);
 718
 719                 goto out;
 720         }
 721
 722         if (!sblock_bad->no_io_error_seen) {
 723                 spin_lock(&sdev->stat_lock);
 724                 sdev->stat.read_errors++;
 725                 spin_unlock(&sdev->stat_lock);
 726                 if (__ratelimit(&_rs))
 727                         scrub_print_warning("i/o error", sblock_to_check);
 728         } else if (sblock_bad->checksum_error) {
 729                 spin_lock(&sdev->stat_lock);
 730                 sdev->stat.csum_errors++;
 731                 spin_unlock(&sdev->stat_lock);
 732                 if (__ratelimit(&_rs))
 733                         scrub_print_warning("checksum error", sblock_to_check);
 734         } else if (sblock_bad->header_error) {
 735                 spin_lock(&sdev->stat_lock);
 736                 sdev->stat.verify_errors++;
 737                 spin_unlock(&sdev->stat_lock);
 738                 if (__ratelimit(&_rs))
 739                         scrub_print_warning("checksum/header error",
 740                                             sblock_to_check);
 741         }
 742
 743         if (sdev->readonly)
 744                 goto did_not_correct_error;
 745
 746         if (!is_metadata && !have_csum) {
 747                 struct scrub_fixup_nodatasum *fixup_nodatasum;
 748
 749                 /*
 750                  * !is_metadata and !have_csum, this means that the data
 751                  * might not be COW'ed, that it might be modified
 752                  * concurrently. The general strategy to work on the
 753                  * commit root does not help in the case when COW is not
 754                  * used.
 755                  */
 756                 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
 757                 if (!fixup_nodatasum)
 758                         goto did_not_correct_error;
 759                 fixup_nodatasum->sdev = sdev;
 760                 fixup_nodatasum->logical = logical;
 761                 fixup_nodatasum->root = fs_info->extent_root;
 762                 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
 763                 /*
 764                  * increment scrubs_running to prevent cancel requests from
 765                  * completing as long as a fixup worker is running. we must also
 766                  * increment scrubs_paused to prevent deadlocking on pause
 767                  * requests used for transactions commits (as the worker uses a
 768                  * transaction context). it is safe to regard the fixup worker
 769                  * as paused for all matters practical. effectively, we only
 770                  * avoid cancellation requests from completing.
 771                  */
 772                 mutex_lock(&fs_info->scrub_lock);
 773                 atomic_inc(&fs_info->scrubs_running);
 774                 atomic_inc(&fs_info->scrubs_paused);
 775                 mutex_unlock(&fs_info->scrub_lock);
 776                 atomic_inc(&sdev->fixup_cnt);
 777                 fixup_nodatasum->work.func = scrub_fixup_nodatasum;
 778                 btrfs_queue_worker(&fs_info->scrub_workers,
 779                                    &fixup_nodatasum->work);
 780                 goto out;
 781         }
 782
 783         /*
 784          * now build and submit the bios for the other mirrors, check
 785          * checksums
 786          */
 787         for (mirror_index = 0;
 788              mirror_index < BTRFS_MAX_MIRRORS &&
 789              sblocks_for_recheck[mirror_index].page_count > 0;
 790              mirror_index++) {
 791                 if (mirror_index == failed_mirror_index)
 792                         continue;
 793
 794                 /* build and submit the bios, check checksums */
 795                 ret = scrub_recheck_block(fs_info,
 796                                           sblocks_for_recheck + mirror_index,
 797                                           is_metadata, have_csum, csum,
 798                                           generation, sdev->csum_size);
 799                 if (ret)
 800                         goto did_not_correct_error;
 801         }
 802
 803         /*
 804          * first try to pick the mirror which is completely without I/O
 805          * errors and also does not have a checksum error.
 806          * If one is found, and if a checksum is present, the full block
 807          * that is known to contain an error is rewritten. Afterwards
 808          * the block is known to be corrected.
 809          * If a mirror is found which is completely correct, and no
 810          * checksum is present, only those pages are rewritten that had
 811          * an I/O error in the block to be repaired, since it cannot be
 812          * determined, which copy of the other pages is better (and it
 813          * could happen otherwise that a correct page would be
 814          * overwritten by a bad one).
 815          */
 816         for (mirror_index = 0;
 817              mirror_index < BTRFS_MAX_MIRRORS &&
 818              sblocks_for_recheck[mirror_index].page_count > 0;
 819              mirror_index++) {
 820                 struct scrub_block *sblock_other = sblocks_for_recheck +
 821                                                    mirror_index;
 822
 823                 if (!sblock_other->header_error &&
 824                     !sblock_other->checksum_error &&
 825                     sblock_other->no_io_error_seen) {
 826                         int force_write = is_metadata || have_csum;
 827
 828                         ret = scrub_repair_block_from_good_copy(sblock_bad,
 829                                                                 sblock_other,
 830                                                                 force_write);
 831                         if (0 == ret)
 832                                 goto corrected_error;
 833                 }
 834         }
 835
 836         /*
 837          * in case of I/O errors in the area that is supposed to be
 838          * repaired, continue by picking good copies of those pages.
 839          * Select the good pages from mirrors to rewrite bad pages from
 840          * the area to fix. Afterwards verify the checksum of the block
 841          * that is supposed to be repaired. This verification step is
 842          * only done for the purpose of statistic counting and for the
 843          * final scrub report, whether errors remain.
 844          * A perfect algorithm could make use of the checksum and try
 845          * all possible combinations of pages from the different mirrors
 846          * until the checksum verification succeeds. For example, when
 847          * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
 848          * of mirror #2 is readable but the final checksum test fails,
 849          * then the 2nd page of mirror #3 could be tried, whether now
 850          * the final checksum succeedes. But this would be a rare
 851          * exception and is therefore not implemented. At least it is
 852          * avoided that the good copy is overwritten.
 853          * A more useful improvement would be to pick the sectors
 854          * without I/O error based on sector sizes (512 bytes on legacy
 855          * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
 856          * mirror could be repaired by taking 512 byte of a different
 857          * mirror, even if other 512 byte sectors in the same PAGE_SIZE
 858          * area are unreadable.
 859          */
 860
 861         /* can only fix I/O errors from here on */
 862         if (sblock_bad->no_io_error_seen)
 863                 goto did_not_correct_error;
 864
 865         success = 1;
 866         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
 867                 struct scrub_page *page_bad = sblock_bad->pagev + page_num;
 868
 869                 if (!page_bad->io_error)
 870                         continue;
 871
 872                 for (mirror_index = 0;
 873                      mirror_index < BTRFS_MAX_MIRRORS &&
 874                      sblocks_for_recheck[mirror_index].page_count > 0;
 875                      mirror_index++) {
 876                         struct scrub_block *sblock_other = sblocks_for_recheck +
 877                                                            mirror_index;
 878                         struct scrub_page *page_other = sblock_other->pagev +
 879                                                         page_num;
 880
 881                         if (!page_other->io_error) {
 882                                 ret = scrub_repair_page_from_good_copy(
 883                                         sblock_bad, sblock_other, page_num, 0);
 884                                 if (0 == ret) {
 885                                         page_bad->io_error = 0;
 886                                         break; /* succeeded for this page */
 887                                 }
 888                         }
 889                 }
 890
 891                 if (page_bad->io_error) {
 892                         /* did not find a mirror to copy the page from */
 893                         success = 0;
 894                 }
 895         }
 896
 897         if (success) {
 898                 if (is_metadata || have_csum) {
 899                         /*
 900                          * need to verify the checksum now that all
 901                          * sectors on disk are repaired (the write
 902                          * request for data to be repaired is on its way).
 903                          * Just be lazy and use scrub_recheck_block()
 904                          * which re-reads the data before the checksum
 905                          * is verified, but most likely the data comes out
 906                          * of the page cache.
 907                          */
 908                         ret = scrub_recheck_block(fs_info, sblock_bad,
 909                                                   is_metadata, have_csum, csum,
 910                                                   generation, sdev->csum_size);
 911                         if (!ret && !sblock_bad->header_error &&
 912                             !sblock_bad->checksum_error &&
 913                             sblock_bad->no_io_error_seen)
 914                                 goto corrected_error;
 915                         else
 916                                 goto did_not_correct_error;
 917                 } else {
 918 corrected_error:
 919                         spin_lock(&sdev->stat_lock);
 920                         sdev->stat.corrected_errors++;
 921                         spin_unlock(&sdev->stat_lock);
 922                         printk_ratelimited(KERN_ERR
 923                                 "btrfs: fixed up error at logical %llu on dev %s\n",
 924                                 (unsigned long long)logical, sdev->dev->name);
 925                 }
 926         } else {
 927 did_not_correct_error:
 928                 spin_lock(&sdev->stat_lock);
 929                 sdev->stat.uncorrectable_errors++;
 930                 spin_unlock(&sdev->stat_lock);
 931                 printk_ratelimited(KERN_ERR
 932                         "btrfs: unable to fixup (regular) error at logical %llu on dev %s\n",
 933                         (unsigned long long)logical, sdev->dev->name);
 934         }
 935
 936 out:
 937         if (sblocks_for_recheck) {
 938                 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
 939                      mirror_index++) {
 940                         struct scrub_block *sblock = sblocks_for_recheck +
 941                                                      mirror_index;
 942                         int page_index;
 943
 944                         for (page_index = 0; page_index < SCRUB_PAGES_PER_BIO;
 945                              page_index++)
 946                                 if (sblock->pagev[page_index].page)
 947                                         __free_page(
 948                                                 sblock->pagev[page_index].page);
 949                 }
 950                 kfree(sblocks_for_recheck);
 951         }
 952
 953         return 0;
 954 }
 955
 956 static int scrub_setup_recheck_block(struct scrub_dev *sdev,
 957                                      struct btrfs_mapping_tree *map_tree,
 958                                      u64 length, u64 logical,
 959                                      struct scrub_block *sblocks_for_recheck)
 960 {
 961         int page_index;
 962         int mirror_index;
 963         int ret;
 964
 965         /*
 966          * note: the three members sdev, ref_count and outstanding_pages
 967          * are not used (and not set) in the blocks that are used for
 968          * the recheck procedure
 969          */
 970
 971         page_index = 0;
 972         while (length > 0) {
 973                 u64 sublen = min_t(u64, length, PAGE_SIZE);
 974                 u64 mapped_length = sublen;
 975                 struct btrfs_bio *bbio = NULL;
 976
 977                 /*
 978                  * with a length of PAGE_SIZE, each returned stripe
 979                  * represents one mirror
 980                  */
 981                 ret = btrfs_map_block(map_tree, WRITE, logical, &mapped_length,
 982                                       &bbio, 0);
 983                 if (ret || !bbio || mapped_length < sublen) {
 984                         kfree(bbio);
 985                         return -EIO;
 986                 }
 987
 988                 BUG_ON(page_index >= SCRUB_PAGES_PER_BIO);
 989                 for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
 990                      mirror_index++) {
 991                         struct scrub_block *sblock;
 992                         struct scrub_page *page;
 993
 994                         if (mirror_index >= BTRFS_MAX_MIRRORS)
 995                                 continue;
 996
 997                         sblock = sblocks_for_recheck + mirror_index;
 998                         page = sblock->pagev + page_index;
 999                         page->logical = logical;
1000                         page->physical = bbio->stripes[mirror_index].physical;
1001                         /* for missing devices, bdev is NULL */
1002                         page->bdev = bbio->stripes[mirror_index].dev->bdev;
1003                         page->mirror_num = mirror_index + 1;
1004                         page->page = alloc_page(GFP_NOFS);
1005                         if (!page->page) {
1006                                 spin_lock(&sdev->stat_lock);
1007                                 sdev->stat.malloc_errors++;
1008                                 spin_unlock(&sdev->stat_lock);
1009                                 return -ENOMEM;
1010                         }
1011                         sblock->page_count++;
1012                 }
1013                 kfree(bbio);
1014                 length -= sublen;
1015                 logical += sublen;
1016                 page_index++;
1017         }
1018
1019         return 0;
1020 }
1021
1022 /*
1023  * this function will check the on disk data for checksum errors, header
1024  * errors and read I/O errors. If any I/O errors happen, the exact pages
1025  * which are errored are marked as being bad. The goal is to enable scrub
1026  * to take those pages that are not errored from all the mirrors so that
1027  * the pages that are errored in the just handled mirror can be repaired.
1028  */
1029 static int scrub_recheck_block(struct btrfs_fs_info *fs_info,
1030                                struct scrub_block *sblock, int is_metadata,
1031                                int have_csum, u8 *csum, u64 generation,
1032                                u16 csum_size)
1033 {
1034         int page_num;
1035
1036         sblock->no_io_error_seen = 1;
1037         sblock->header_error = 0;
1038         sblock->checksum_error = 0;
1039
1040         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1041                 struct bio *bio;
1042                 int ret;
1043                 struct scrub_page *page = sblock->pagev + page_num;
1044                 DECLARE_COMPLETION_ONSTACK(complete);
1045
1046                 if (page->bdev == NULL) {
1047                         page->io_error = 1;
1048                         sblock->no_io_error_seen = 0;
1049                         continue;
1050                 }
1051
1052                 BUG_ON(!page->page);
1053                 bio = bio_alloc(GFP_NOFS, 1);
1054                 if (!bio)
1055                         return -EIO;
1056                 bio->bi_bdev = page->bdev;
1057                 bio->bi_sector = page->physical >> 9;
1058                 bio->bi_end_io = scrub_complete_bio_end_io;
1059                 bio->bi_private = &complete;
1060
1061                 ret = bio_add_page(bio, page->page, PAGE_SIZE, 0);
1062                 if (PAGE_SIZE != ret) {
1063                         bio_put(bio);
1064                         return -EIO;
1065                 }
1066                 btrfsic_submit_bio(READ, bio);
1067
1068                 /* this will also unplug the queue */
1069                 wait_for_completion(&complete);
1070
1071                 page->io_error = !test_bit(BIO_UPTODATE, &bio->bi_flags);
1072                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1073                         sblock->no_io_error_seen = 0;
1074                 bio_put(bio);
1075         }
1076
1077         if (sblock->no_io_error_seen)
1078                 scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1079                                              have_csum, csum, generation,
1080                                              csum_size);
1081
1082         return 0;
1083 }
1084
1085 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1086                                          struct scrub_block *sblock,
1087                                          int is_metadata, int have_csum,
1088                                          const u8 *csum, u64 generation,
1089                                          u16 csum_size)
1090 {
1091         int page_num;
1092         u8 calculated_csum[BTRFS_CSUM_SIZE];
1093         u32 crc = ~(u32)0;
1094         struct btrfs_root *root = fs_info->extent_root;
1095         void *mapped_buffer;
1096
1097         BUG_ON(!sblock->pagev[0].page);
1098         if (is_metadata) {
1099                 struct btrfs_header *h;
1100
1101                 mapped_buffer = kmap_atomic(sblock->pagev[0].page);
1102                 h = (struct btrfs_header *)mapped_buffer;
1103
1104                 if (sblock->pagev[0].logical != le64_to_cpu(h->bytenr) ||
1105                     generation != le64_to_cpu(h->generation) ||
1106                     memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
1107                     memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1108                            BTRFS_UUID_SIZE))
1109                         sblock->header_error = 1;
1110                 csum = h->csum;
1111         } else {
1112                 if (!have_csum)
1113                         return;
1114
1115                 mapped_buffer = kmap_atomic(sblock->pagev[0].page);
1116         }
1117
1118         for (page_num = 0;;) {
1119                 if (page_num == 0 && is_metadata)
1120                         crc = btrfs_csum_data(root,
1121                                 ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1122                                 crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1123                 else
1124                         crc = btrfs_csum_data(root, mapped_buffer, crc,
1125                                               PAGE_SIZE);
1126
1127                 kunmap_atomic(mapped_buffer);
1128                 page_num++;
1129                 if (page_num >= sblock->page_count)
1130                         break;
1131                 BUG_ON(!sblock->pagev[page_num].page);
1132
1133                 mapped_buffer = kmap_atomic(sblock->pagev[page_num].page);
1134         }
1135
1136         btrfs_csum_final(crc, calculated_csum);
1137         if (memcmp(calculated_csum, csum, csum_size))
1138                 sblock->checksum_error = 1;
1139 }
1140
1141 static void scrub_complete_bio_end_io(struct bio *bio, int err)
1142 {
1143         complete((struct completion *)bio->bi_private);
1144 }
1145
1146 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1147                                              struct scrub_block *sblock_good,
1148                                              int force_write)
1149 {
1150         int page_num;
1151         int ret = 0;
1152
1153         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1154                 int ret_sub;
1155
1156                 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1157                                                            sblock_good,
1158                                                            page_num,
1159                                                            force_write);
1160                 if (ret_sub)
1161                         ret = ret_sub;
1162         }
1163
1164         return ret;
1165 }
1166
1167 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1168                                             struct scrub_block *sblock_good,
1169                                             int page_num, int force_write)
1170 {
1171         struct scrub_page *page_bad = sblock_bad->pagev + page_num;
1172         struct scrub_page *page_good = sblock_good->pagev + page_num;
1173
1174         BUG_ON(sblock_bad->pagev[page_num].page == NULL);
1175         BUG_ON(sblock_good->pagev[page_num].page == NULL);
1176         if (force_write || sblock_bad->header_error ||
1177             sblock_bad->checksum_error || page_bad->io_error) {
1178                 struct bio *bio;
1179                 int ret;
1180                 DECLARE_COMPLETION_ONSTACK(complete);
1181
1182                 bio = bio_alloc(GFP_NOFS, 1);
1183                 if (!bio)
1184                         return -EIO;
1185                 bio->bi_bdev = page_bad->bdev;
1186                 bio->bi_sector = page_bad->physical >> 9;
1187                 bio->bi_end_io = scrub_complete_bio_end_io;
1188                 bio->bi_private = &complete;
1189
1190                 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1191                 if (PAGE_SIZE != ret) {
1192                         bio_put(bio);
1193                         return -EIO;
1194                 }
1195                 btrfsic_submit_bio(WRITE, bio);
1196
1197                 /* this will also unplug the queue */
1198                 wait_for_completion(&complete);
1199                 bio_put(bio);
1200         }
1201
1202         return 0;
1203 }
1204
1205 static void scrub_checksum(struct scrub_block *sblock)
1206 {
1207         u64 flags;
1208         int ret;
1209
1210         BUG_ON(sblock->page_count < 1);
1211         flags = sblock->pagev[0].flags;
1212         ret = 0;
1213         if (flags & BTRFS_EXTENT_FLAG_DATA)
1214                 ret = scrub_checksum_data(sblock);
1215         else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1216                 ret = scrub_checksum_tree_block(sblock);
1217         else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1218                 (void)scrub_checksum_super(sblock);
1219         else
1220                 WARN_ON(1);
1221         if (ret)
1222                 scrub_handle_errored_block(sblock);
1223 }
1224
1225 static int scrub_checksum_data(struct scrub_block *sblock)
1226 {
1227         struct scrub_dev *sdev = sblock->sdev;
1228         u8 csum[BTRFS_CSUM_SIZE];
1229         u8 *on_disk_csum;
1230         struct page *page;
1231         void *buffer;
1232         u32 crc = ~(u32)0;
1233         int fail = 0;
1234         struct btrfs_root *root = sdev->dev->dev_root;
1235         u64 len;
1236         int index;
1237
1238         BUG_ON(sblock->page_count < 1);
1239         if (!sblock->pagev[0].have_csum)
1240                 return 0;
1241
1242         on_disk_csum = sblock->pagev[0].csum;
1243         page = sblock->pagev[0].page;
1244         buffer = kmap_atomic(page);
1245
1246         len = sdev->sectorsize;
1247         index = 0;
1248         for (;;) {
1249                 u64 l = min_t(u64, len, PAGE_SIZE);
1250
1251                 crc = btrfs_csum_data(root, buffer, crc, l);
1252                 kunmap_atomic(buffer);
1253                 len -= l;
1254                 if (len == 0)
1255                         break;
1256                 index++;
1257                 BUG_ON(index >= sblock->page_count);
1258                 BUG_ON(!sblock->pagev[index].page);
1259                 page = sblock->pagev[index].page;
1260                 buffer = kmap_atomic(page);
1261         }
1262
1263         btrfs_csum_final(crc, csum);
1264         if (memcmp(csum, on_disk_csum, sdev->csum_size))
1265                 fail = 1;
1266
1267         return fail;
1268 }
1269
1270 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1271 {
1272         struct scrub_dev *sdev = sblock->sdev;
1273         struct btrfs_header *h;
1274         struct btrfs_root *root = sdev->dev->dev_root;
1275         struct btrfs_fs_info *fs_info = root->fs_info;
1276         u8 calculated_csum[BTRFS_CSUM_SIZE];
1277         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1278         struct page *page;
1279         void *mapped_buffer;
1280         u64 mapped_size;
1281         void *p;
1282         u32 crc = ~(u32)0;
1283         int fail = 0;
1284         int crc_fail = 0;
1285         u64 len;
1286         int index;
1287
1288         BUG_ON(sblock->page_count < 1);
1289         page = sblock->pagev[0].page;
1290         mapped_buffer = kmap_atomic(page);
1291         h = (struct btrfs_header *)mapped_buffer;
1292         memcpy(on_disk_csum, h->csum, sdev->csum_size);
1293
1294         /*
1295          * we don't use the getter functions here, as we
1296          * a) don't have an extent buffer and
1297          * b) the page is already kmapped
1298          */
1299
1300         if (sblock->pagev[0].logical != le64_to_cpu(h->bytenr))
1301                 ++fail;
1302
1303         if (sblock->pagev[0].generation != le64_to_cpu(h->generation))
1304                 ++fail;
1305
1306         if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1307                 ++fail;
1308
1309         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1310                    BTRFS_UUID_SIZE))
1311                 ++fail;
1312
1313         BUG_ON(sdev->nodesize != sdev->leafsize);
1314         len = sdev->nodesize - BTRFS_CSUM_SIZE;
1315         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1316         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1317         index = 0;
1318         for (;;) {
1319                 u64 l = min_t(u64, len, mapped_size);
1320
1321                 crc = btrfs_csum_data(root, p, crc, l);
1322                 kunmap_atomic(mapped_buffer);
1323                 len -= l;
1324                 if (len == 0)
1325                         break;
1326                 index++;
1327                 BUG_ON(index >= sblock->page_count);
1328                 BUG_ON(!sblock->pagev[index].page);
1329                 page = sblock->pagev[index].page;
1330                 mapped_buffer = kmap_atomic(page);
1331                 mapped_size = PAGE_SIZE;
1332                 p = mapped_buffer;
1333         }
1334
1335         btrfs_csum_final(crc, calculated_csum);
1336         if (memcmp(calculated_csum, on_disk_csum, sdev->csum_size))
1337                 ++crc_fail;
1338
1339         return fail || crc_fail;
1340 }
1341
1342 static int scrub_checksum_super(struct scrub_block *sblock)
1343 {
1344         struct btrfs_super_block *s;
1345         struct scrub_dev *sdev = sblock->sdev;
1346         struct btrfs_root *root = sdev->dev->dev_root;
1347         struct btrfs_fs_info *fs_info = root->fs_info;
1348         u8 calculated_csum[BTRFS_CSUM_SIZE];
1349         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1350         struct page *page;
1351         void *mapped_buffer;
1352         u64 mapped_size;
1353         void *p;
1354         u32 crc = ~(u32)0;
1355         int fail = 0;
1356         u64 len;
1357         int index;
1358
1359         BUG_ON(sblock->page_count < 1);
1360         page = sblock->pagev[0].page;
1361         mapped_buffer = kmap_atomic(page);
1362         s = (struct btrfs_super_block *)mapped_buffer;
1363         memcpy(on_disk_csum, s->csum, sdev->csum_size);
1364
1365         if (sblock->pagev[0].logical != le64_to_cpu(s->bytenr))
1366                 ++fail;
1367
1368         if (sblock->pagev[0].generation != le64_to_cpu(s->generation))
1369                 ++fail;
1370
1371         if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1372                 ++fail;
1373
1374         len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1375         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1376         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1377         index = 0;
1378         for (;;) {
1379                 u64 l = min_t(u64, len, mapped_size);
1380
1381                 crc = btrfs_csum_data(root, p, crc, l);
1382                 kunmap_atomic(mapped_buffer);
1383                 len -= l;
1384                 if (len == 0)
1385                         break;
1386                 index++;
1387                 BUG_ON(index >= sblock->page_count);
1388                 BUG_ON(!sblock->pagev[index].page);
1389                 page = sblock->pagev[index].page;
1390                 mapped_buffer = kmap_atomic(page);
1391                 mapped_size = PAGE_SIZE;
1392                 p = mapped_buffer;
1393         }
1394
1395         btrfs_csum_final(crc, calculated_csum);
1396         if (memcmp(calculated_csum, on_disk_csum, sdev->csum_size))
1397                 ++fail;
1398
1399         if (fail) {
1400                 /*
1401                  * if we find an error in a super block, we just report it.
1402                  * They will get written with the next transaction commit
1403                  * anyway
1404                  */
1405                 spin_lock(&sdev->stat_lock);
1406                 ++sdev->stat.super_errors;
1407                 spin_unlock(&sdev->stat_lock);
1408         }
1409
1410         return fail;
1411 }
1412
1413 static void scrub_block_get(struct scrub_block *sblock)
1414 {
1415         atomic_inc(&sblock->ref_count);
1416 }
1417
1418 static void scrub_block_put(struct scrub_block *sblock)
1419 {
1420         if (atomic_dec_and_test(&sblock->ref_count)) {
1421                 int i;
1422
1423                 for (i = 0; i < sblock->page_count; i++)
1424                         if (sblock->pagev[i].page)
1425                                 __free_page(sblock->pagev[i].page);
1426                 kfree(sblock);
1427         }
1428 }
1429
1430 static void scrub_submit(struct scrub_dev *sdev)
1431 {
1432         struct scrub_bio *sbio;
1433
1434         if (sdev->curr == -1)
1435                 return;
1436
1437         sbio = sdev->bios[sdev->curr];
1438         sdev->curr = -1;
1439         atomic_inc(&sdev->in_flight);
1440
1441         btrfsic_submit_bio(READ, sbio->bio);
1442 }
1443
1444 static int scrub_add_page_to_bio(struct scrub_dev *sdev,
1445                                  struct scrub_page *spage)
1446 {
1447         struct scrub_block *sblock = spage->sblock;
1448         struct scrub_bio *sbio;
1449         int ret;
1450
1451 again:
1452         /*
1453          * grab a fresh bio or wait for one to become available
1454          */
1455         while (sdev->curr == -1) {
1456                 spin_lock(&sdev->list_lock);
1457                 sdev->curr = sdev->first_free;
1458                 if (sdev->curr != -1) {
1459                         sdev->first_free = sdev->bios[sdev->curr]->next_free;
1460                         sdev->bios[sdev->curr]->next_free = -1;
1461                         sdev->bios[sdev->curr]->page_count = 0;
1462                         spin_unlock(&sdev->list_lock);
1463                 } else {
1464                         spin_unlock(&sdev->list_lock);
1465                         wait_event(sdev->list_wait, sdev->first_free != -1);
1466                 }
1467         }
1468         sbio = sdev->bios[sdev->curr];
1469         if (sbio->page_count == 0) {
1470                 struct bio *bio;
1471
1472                 sbio->physical = spage->physical;
1473                 sbio->logical = spage->logical;
1474                 bio = sbio->bio;
1475                 if (!bio) {
1476                         bio = bio_alloc(GFP_NOFS, sdev->pages_per_bio);
1477                         if (!bio)
1478                                 return -ENOMEM;
1479                         sbio->bio = bio;
1480                 }
1481
1482                 bio->bi_private = sbio;
1483                 bio->bi_end_io = scrub_bio_end_io;
1484                 bio->bi_bdev = sdev->dev->bdev;
1485                 bio->bi_sector = spage->physical >> 9;
1486                 sbio->err = 0;
1487         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1488                    spage->physical ||
1489                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1490                    spage->logical) {
1491                 scrub_submit(sdev);
1492                 goto again;
1493         }
1494
1495         sbio->pagev[sbio->page_count] = spage;
1496         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1497         if (ret != PAGE_SIZE) {
1498                 if (sbio->page_count < 1) {
1499                         bio_put(sbio->bio);
1500                         sbio->bio = NULL;
1501                         return -EIO;
1502                 }
1503                 scrub_submit(sdev);
1504                 goto again;
1505         }
1506
1507         scrub_block_get(sblock); /* one for the added page */
1508         atomic_inc(&sblock->outstanding_pages);
1509         sbio->page_count++;
1510         if (sbio->page_count == sdev->pages_per_bio)
1511                 scrub_submit(sdev);
1512
1513         return 0;
1514 }
1515
1516 static int scrub_pages(struct scrub_dev *sdev, u64 logical, u64 len,
1517                        u64 physical, u64 flags, u64 gen, int mirror_num,
1518                        u8 *csum, int force)
1519 {
1520         struct scrub_block *sblock;
1521         int index;
1522
1523         sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
1524         if (!sblock) {
1525                 spin_lock(&sdev->stat_lock);
1526                 sdev->stat.malloc_errors++;
1527                 spin_unlock(&sdev->stat_lock);
1528                 return -ENOMEM;
1529         }
1530
1531         /* one ref inside this function, plus one for each page later on */
1532         atomic_set(&sblock->ref_count, 1);
1533         sblock->sdev = sdev;
1534         sblock->no_io_error_seen = 1;
1535
1536         for (index = 0; len > 0; index++) {
1537                 struct scrub_page *spage = sblock->pagev + index;
1538                 u64 l = min_t(u64, len, PAGE_SIZE);
1539
1540                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
1541                 spage->page = alloc_page(GFP_NOFS);
1542                 if (!spage->page) {
1543                         spin_lock(&sdev->stat_lock);
1544                         sdev->stat.malloc_errors++;
1545                         spin_unlock(&sdev->stat_lock);
1546                         while (index > 0) {
1547                                 index--;
1548                                 __free_page(sblock->pagev[index].page);
1549                         }
1550                         kfree(sblock);
1551                         return -ENOMEM;
1552                 }
1553                 spage->sblock = sblock;
1554                 spage->bdev = sdev->dev->bdev;
1555                 spage->flags = flags;
1556                 spage->generation = gen;
1557                 spage->logical = logical;
1558                 spage->physical = physical;
1559                 spage->mirror_num = mirror_num;
1560                 if (csum) {
1561                         spage->have_csum = 1;
1562                         memcpy(spage->csum, csum, sdev->csum_size);
1563                 } else {
1564                         spage->have_csum = 0;
1565                 }
1566                 sblock->page_count++;
1567                 len -= l;
1568                 logical += l;
1569                 physical += l;
1570         }
1571
1572         BUG_ON(sblock->page_count == 0);
1573         for (index = 0; index < sblock->page_count; index++) {
1574                 struct scrub_page *spage = sblock->pagev + index;
1575                 int ret;
1576
1577                 ret = scrub_add_page_to_bio(sdev, spage);
1578                 if (ret) {
1579                         scrub_block_put(sblock);
1580                         return ret;
1581                 }
1582         }
1583
1584         if (force)
1585                 scrub_submit(sdev);
1586
1587         /* last one frees, either here or in bio completion for last page */
1588         scrub_block_put(sblock);
1589         return 0;
1590 }
1591
1592 static void scrub_bio_end_io(struct bio *bio, int err)
1593 {
1594         struct scrub_bio *sbio = bio->bi_private;
1595         struct scrub_dev *sdev = sbio->sdev;
1596         struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
1597
1598         sbio->err = err;
1599         sbio->bio = bio;
1600
1601         btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
1602 }
1603
1604 static void scrub_bio_end_io_worker(struct btrfs_work *work)
1605 {
1606         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1607         struct scrub_dev *sdev = sbio->sdev;
1608         int i;
1609
1610         BUG_ON(sbio->page_count > SCRUB_PAGES_PER_BIO);
1611         if (sbio->err) {
1612                 for (i = 0; i < sbio->page_count; i++) {
1613                         struct scrub_page *spage = sbio->pagev[i];
1614
1615                         spage->io_error = 1;
1616                         spage->sblock->no_io_error_seen = 0;
1617                 }
1618         }
1619
1620         /* now complete the scrub_block items that have all pages completed */
1621         for (i = 0; i < sbio->page_count; i++) {
1622                 struct scrub_page *spage = sbio->pagev[i];
1623                 struct scrub_block *sblock = spage->sblock;
1624
1625                 if (atomic_dec_and_test(&sblock->outstanding_pages))
1626                         scrub_block_complete(sblock);
1627                 scrub_block_put(sblock);
1628         }
1629
1630         if (sbio->err) {
1631                 /* what is this good for??? */
1632                 sbio->bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1633                 sbio->bio->bi_flags |= 1 << BIO_UPTODATE;
1634                 sbio->bio->bi_phys_segments = 0;
1635                 sbio->bio->bi_idx = 0;
1636
1637                 for (i = 0; i < sbio->page_count; i++) {
1638                         struct bio_vec *bi;
1639                         bi = &sbio->bio->bi_io_vec[i];
1640                         bi->bv_offset = 0;
1641                         bi->bv_len = PAGE_SIZE;
1642                 }
1643         }
1644
1645         bio_put(sbio->bio);
1646         sbio->bio = NULL;
1647         spin_lock(&sdev->list_lock);
1648         sbio->next_free = sdev->first_free;
1649         sdev->first_free = sbio->index;
1650         spin_unlock(&sdev->list_lock);
1651         atomic_dec(&sdev->in_flight);
1652         wake_up(&sdev->list_wait);
1653 }
1654
1655 static void scrub_block_complete(struct scrub_block *sblock)
1656 {
1657         if (!sblock->no_io_error_seen)
1658                 scrub_handle_errored_block(sblock);
1659         else
1660                 scrub_checksum(sblock);
1661 }
1662
1663 static int scrub_find_csum(struct scrub_dev *sdev, u64 logical, u64 len,
1664                            u8 *csum)
1665 {
1666         struct btrfs_ordered_sum *sum = NULL;
1667         int ret = 0;
1668         unsigned long i;
1669         unsigned long num_sectors;
1670
1671         while (!list_empty(&sdev->csum_list)) {
1672                 sum = list_first_entry(&sdev->csum_list,
1673                                        struct btrfs_ordered_sum, list);
1674                 if (sum->bytenr > logical)
1675                         return 0;
1676                 if (sum->bytenr + sum->len > logical)
1677                         break;
1678
1679                 ++sdev->stat.csum_discards;
1680                 list_del(&sum->list);
1681                 kfree(sum);
1682                 sum = NULL;
1683         }
1684         if (!sum)
1685                 return 0;
1686
1687         num_sectors = sum->len / sdev->sectorsize;
1688         for (i = 0; i < num_sectors; ++i) {
1689                 if (sum->sums[i].bytenr == logical) {
1690                         memcpy(csum, &sum->sums[i].sum, sdev->csum_size);
1691                         ret = 1;
1692                         break;
1693                 }
1694         }
1695         if (ret && i == num_sectors - 1) {
1696                 list_del(&sum->list);
1697                 kfree(sum);
1698         }
1699         return ret;
1700 }
1701
1702 /* scrub extent tries to collect up to 64 kB for each bio */
1703 static int scrub_extent(struct scrub_dev *sdev, u64 logical, u64 len,
1704                         u64 physical, u64 flags, u64 gen, int mirror_num)
1705 {
1706         int ret;
1707         u8 csum[BTRFS_CSUM_SIZE];
1708         u32 blocksize;
1709
1710         if (flags & BTRFS_EXTENT_FLAG_DATA) {
1711                 blocksize = sdev->sectorsize;
1712                 spin_lock(&sdev->stat_lock);
1713                 sdev->stat.data_extents_scrubbed++;
1714                 sdev->stat.data_bytes_scrubbed += len;
1715                 spin_unlock(&sdev->stat_lock);
1716         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1717                 BUG_ON(sdev->nodesize != sdev->leafsize);
1718                 blocksize = sdev->nodesize;
1719                 spin_lock(&sdev->stat_lock);
1720                 sdev->stat.tree_extents_scrubbed++;
1721                 sdev->stat.tree_bytes_scrubbed += len;
1722                 spin_unlock(&sdev->stat_lock);
1723         } else {
1724                 blocksize = sdev->sectorsize;
1725                 BUG_ON(1);
1726         }
1727
1728         while (len) {
1729                 u64 l = min_t(u64, len, blocksize);
1730                 int have_csum = 0;
1731
1732                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
1733                         /* push csums to sbio */
1734                         have_csum = scrub_find_csum(sdev, logical, l, csum);
1735                         if (have_csum == 0)
1736                                 ++sdev->stat.no_csum;
1737                 }
1738                 ret = scrub_pages(sdev, logical, l, physical, flags, gen,
1739                                   mirror_num, have_csum ? csum : NULL, 0);
1740                 if (ret)
1741                         return ret;
1742                 len -= l;
1743                 logical += l;
1744                 physical += l;
1745         }
1746         return 0;
1747 }
1748
1749 static noinline_for_stack int scrub_stripe(struct scrub_dev *sdev,
1750         struct map_lookup *map, int num, u64 base, u64 length)
1751 {
1752         struct btrfs_path *path;
1753         struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
1754         struct btrfs_root *root = fs_info->extent_root;
1755         struct btrfs_root *csum_root = fs_info->csum_root;
1756         struct btrfs_extent_item *extent;
1757         struct blk_plug plug;
1758         u64 flags;
1759         int ret;
1760         int slot;
1761         int i;
1762         u64 nstripes;
1763         struct extent_buffer *l;
1764         struct btrfs_key key;
1765         u64 physical;
1766         u64 logical;
1767         u64 generation;
1768         int mirror_num;
1769         struct reada_control *reada1;
1770         struct reada_control *reada2;
1771         struct btrfs_key key_start;
1772         struct btrfs_key key_end;
1773
1774         u64 increment = map->stripe_len;
1775         u64 offset;
1776
1777         nstripes = length;
1778         offset = 0;
1779         do_div(nstripes, map->stripe_len);
1780         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
1781                 offset = map->stripe_len * num;
1782                 increment = map->stripe_len * map->num_stripes;
1783                 mirror_num = 1;
1784         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
1785                 int factor = map->num_stripes / map->sub_stripes;
1786                 offset = map->stripe_len * (num / map->sub_stripes);
1787                 increment = map->stripe_len * factor;
1788                 mirror_num = num % map->sub_stripes + 1;
1789         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
1790                 increment = map->stripe_len;
1791                 mirror_num = num % map->num_stripes + 1;
1792         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
1793                 increment = map->stripe_len;
1794                 mirror_num = num % map->num_stripes + 1;
1795         } else {
1796                 increment = map->stripe_len;
1797                 mirror_num = 1;
1798         }
1799
1800         path = btrfs_alloc_path();
1801         if (!path)
1802                 return -ENOMEM;
1803
1804         /*
1805          * work on commit root. The related disk blocks are static as
1806          * long as COW is applied. This means, it is save to rewrite
1807          * them to repair disk errors without any race conditions
1808          */
1809         path->search_commit_root = 1;
1810         path->skip_locking = 1;
1811
1812         /*
1813          * trigger the readahead for extent tree csum tree and wait for
1814          * completion. During readahead, the scrub is officially paused
1815          * to not hold off transaction commits
1816          */
1817         logical = base + offset;
1818
1819         wait_event(sdev->list_wait,
1820                    atomic_read(&sdev->in_flight) == 0);
1821         atomic_inc(&fs_info->scrubs_paused);
1822         wake_up(&fs_info->scrub_pause_wait);
1823
1824         /* FIXME it might be better to start readahead at commit root */
1825         key_start.objectid = logical;
1826         key_start.type = BTRFS_EXTENT_ITEM_KEY;
1827         key_start.offset = (u64)0;
1828         key_end.objectid = base + offset + nstripes * increment;
1829         key_end.type = BTRFS_EXTENT_ITEM_KEY;
1830         key_end.offset = (u64)0;
1831         reada1 = btrfs_reada_add(root, &key_start, &key_end);
1832
1833         key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
1834         key_start.type = BTRFS_EXTENT_CSUM_KEY;
1835         key_start.offset = logical;
1836         key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
1837         key_end.type = BTRFS_EXTENT_CSUM_KEY;
1838         key_end.offset = base + offset + nstripes * increment;
1839         reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
1840
1841         if (!IS_ERR(reada1))
1842                 btrfs_reada_wait(reada1);
1843         if (!IS_ERR(reada2))
1844                 btrfs_reada_wait(reada2);
1845
1846         mutex_lock(&fs_info->scrub_lock);
1847         while (atomic_read(&fs_info->scrub_pause_req)) {
1848                 mutex_unlock(&fs_info->scrub_lock);
1849                 wait_event(fs_info->scrub_pause_wait,
1850                    atomic_read(&fs_info->scrub_pause_req) == 0);
1851                 mutex_lock(&fs_info->scrub_lock);
1852         }
1853         atomic_dec(&fs_info->scrubs_paused);
1854         mutex_unlock(&fs_info->scrub_lock);
1855         wake_up(&fs_info->scrub_pause_wait);
1856
1857         /*
1858          * collect all data csums for the stripe to avoid seeking during
1859          * the scrub. This might currently (crc32) end up to be about 1MB
1860          */
1861         blk_start_plug(&plug);
1862
1863         /*
1864          * now find all extents for each stripe and scrub them
1865          */
1866         logical = base + offset;
1867         physical = map->stripes[num].physical;
1868         ret = 0;
1869         for (i = 0; i < nstripes; ++i) {
1870                 /*
1871                  * canceled?
1872                  */
1873                 if (atomic_read(&fs_info->scrub_cancel_req) ||
1874                     atomic_read(&sdev->cancel_req)) {
1875                         ret = -ECANCELED;
1876                         goto out;
1877                 }
1878                 /*
1879                  * check to see if we have to pause
1880                  */
1881                 if (atomic_read(&fs_info->scrub_pause_req)) {
1882                         /* push queued extents */
1883                         scrub_submit(sdev);
1884                         wait_event(sdev->list_wait,
1885                                    atomic_read(&sdev->in_flight) == 0);
1886                         atomic_inc(&fs_info->scrubs_paused);
1887                         wake_up(&fs_info->scrub_pause_wait);
1888                         mutex_lock(&fs_info->scrub_lock);
1889                         while (atomic_read(&fs_info->scrub_pause_req)) {
1890                                 mutex_unlock(&fs_info->scrub_lock);
1891                                 wait_event(fs_info->scrub_pause_wait,
1892                                    atomic_read(&fs_info->scrub_pause_req) == 0);
1893                                 mutex_lock(&fs_info->scrub_lock);
1894                         }
1895                         atomic_dec(&fs_info->scrubs_paused);
1896                         mutex_unlock(&fs_info->scrub_lock);
1897                         wake_up(&fs_info->scrub_pause_wait);
1898                 }
1899
1900                 ret = btrfs_lookup_csums_range(csum_root, logical,
1901                                                logical + map->stripe_len - 1,
1902                                                &sdev->csum_list, 1);
1903                 if (ret)
1904                         goto out;
1905
1906                 key.objectid = logical;
1907                 key.type = BTRFS_EXTENT_ITEM_KEY;
1908                 key.offset = (u64)0;
1909
1910                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1911                 if (ret < 0)
1912                         goto out;
1913                 if (ret > 0) {
1914                         ret = btrfs_previous_item(root, path, 0,
1915                                                   BTRFS_EXTENT_ITEM_KEY);
1916                         if (ret < 0)
1917                                 goto out;
1918                         if (ret > 0) {
1919                                 /* there's no smaller item, so stick with the
1920                                  * larger one */
1921                                 btrfs_release_path(path);
1922                                 ret = btrfs_search_slot(NULL, root, &key,
1923                                                         path, 0, 0);
1924                                 if (ret < 0)
1925                                         goto out;
1926                         }
1927                 }
1928
1929                 while (1) {
1930                         l = path->nodes[0];
1931                         slot = path->slots[0];
1932                         if (slot >= btrfs_header_nritems(l)) {
1933                                 ret = btrfs_next_leaf(root, path);
1934                                 if (ret == 0)
1935                                         continue;
1936                                 if (ret < 0)
1937                                         goto out;
1938
1939                                 break;
1940                         }
1941                         btrfs_item_key_to_cpu(l, &key, slot);
1942
1943                         if (key.objectid + key.offset <= logical)
1944                                 goto next;
1945
1946                         if (key.objectid >= logical + map->stripe_len)
1947                                 break;
1948
1949                         if (btrfs_key_type(&key) != BTRFS_EXTENT_ITEM_KEY)
1950                                 goto next;
1951
1952                         extent = btrfs_item_ptr(l, slot,
1953                                                 struct btrfs_extent_item);
1954                         flags = btrfs_extent_flags(l, extent);
1955                         generation = btrfs_extent_generation(l, extent);
1956
1957                         if (key.objectid < logical &&
1958                             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
1959                                 printk(KERN_ERR
1960                                        "btrfs scrub: tree block %llu spanning "
1961                                        "stripes, ignored. logical=%llu\n",
1962                                        (unsigned long long)key.objectid,
1963                                        (unsigned long long)logical);
1964                                 goto next;
1965                         }
1966
1967                         /*
1968                          * trim extent to this stripe
1969                          */
1970                         if (key.objectid < logical) {
1971                                 key.offset -= logical - key.objectid;
1972                                 key.objectid = logical;
1973                         }
1974                         if (key.objectid + key.offset >
1975                             logical + map->stripe_len) {
1976                                 key.offset = logical + map->stripe_len -
1977                                              key.objectid;
1978                         }
1979
1980                         ret = scrub_extent(sdev, key.objectid, key.offset,
1981                                            key.objectid - logical + physical,
1982                                            flags, generation, mirror_num);
1983                         if (ret)
1984                                 goto out;
1985
1986 next:
1987                         path->slots[0]++;
1988                 }
1989                 btrfs_release_path(path);
1990                 logical += increment;
1991                 physical += map->stripe_len;
1992                 spin_lock(&sdev->stat_lock);
1993                 sdev->stat.last_physical = physical;
1994                 spin_unlock(&sdev->stat_lock);
1995         }
1996         /* push queued extents */
1997         scrub_submit(sdev);
1998
1999 out:
2000         blk_finish_plug(&plug);
2001         btrfs_free_path(path);
2002         return ret < 0 ? ret : 0;
2003 }
2004
2005 static noinline_for_stack int scrub_chunk(struct scrub_dev *sdev,
2006         u64 chunk_tree, u64 chunk_objectid, u64 chunk_offset, u64 length,
2007         u64 dev_offset)
2008 {
2009         struct btrfs_mapping_tree *map_tree =
2010                 &sdev->dev->dev_root->fs_info->mapping_tree;
2011         struct map_lookup *map;
2012         struct extent_map *em;
2013         int i;
2014         int ret = -EINVAL;
2015
2016         read_lock(&map_tree->map_tree.lock);
2017         em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2018         read_unlock(&map_tree->map_tree.lock);
2019
2020         if (!em)
2021                 return -EINVAL;
2022
2023         map = (struct map_lookup *)em->bdev;
2024         if (em->start != chunk_offset)
2025                 goto out;
2026
2027         if (em->len < length)
2028                 goto out;
2029
2030         for (i = 0; i < map->num_stripes; ++i) {
2031                 if (map->stripes[i].dev == sdev->dev &&
2032                     map->stripes[i].physical == dev_offset) {
2033                         ret = scrub_stripe(sdev, map, i, chunk_offset, length);
2034                         if (ret)
2035                                 goto out;
2036                 }
2037         }
2038 out:
2039         free_extent_map(em);
2040
2041         return ret;
2042 }
2043
2044 static noinline_for_stack
2045 int scrub_enumerate_chunks(struct scrub_dev *sdev, u64 start, u64 end)
2046 {
2047         struct btrfs_dev_extent *dev_extent = NULL;
2048         struct btrfs_path *path;
2049         struct btrfs_root *root = sdev->dev->dev_root;
2050         struct btrfs_fs_info *fs_info = root->fs_info;
2051         u64 length;
2052         u64 chunk_tree;
2053         u64 chunk_objectid;
2054         u64 chunk_offset;
2055         int ret;
2056         int slot;
2057         struct extent_buffer *l;
2058         struct btrfs_key key;
2059         struct btrfs_key found_key;
2060         struct btrfs_block_group_cache *cache;
2061
2062         path = btrfs_alloc_path();
2063         if (!path)
2064                 return -ENOMEM;
2065
2066         path->reada = 2;
2067         path->search_commit_root = 1;
2068         path->skip_locking = 1;
2069
2070         key.objectid = sdev->dev->devid;
2071         key.offset = 0ull;
2072         key.type = BTRFS_DEV_EXTENT_KEY;
2073
2074
2075         while (1) {
2076                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2077                 if (ret < 0)
2078                         break;
2079                 if (ret > 0) {
2080                         if (path->slots[0] >=
2081                             btrfs_header_nritems(path->nodes[0])) {
2082                                 ret = btrfs_next_leaf(root, path);
2083                                 if (ret)
2084                                         break;
2085                         }
2086                 }
2087
2088                 l = path->nodes[0];
2089                 slot = path->slots[0];
2090
2091                 btrfs_item_key_to_cpu(l, &found_key, slot);
2092
2093                 if (found_key.objectid != sdev->dev->devid)
2094                         break;
2095
2096                 if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
2097                         break;
2098
2099                 if (found_key.offset >= end)
2100                         break;
2101
2102                 if (found_key.offset < key.offset)
2103                         break;
2104
2105                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2106                 length = btrfs_dev_extent_length(l, dev_extent);
2107
2108                 if (found_key.offset + length <= start) {
2109                         key.offset = found_key.offset + length;
2110                         btrfs_release_path(path);
2111                         continue;
2112                 }
2113
2114                 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2115                 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2116                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2117
2118                 /*
2119                  * get a reference on the corresponding block group to prevent
2120                  * the chunk from going away while we scrub it
2121                  */
2122                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2123                 if (!cache) {
2124                         ret = -ENOENT;
2125                         break;
2126                 }
2127                 ret = scrub_chunk(sdev, chunk_tree, chunk_objectid,
2128                                   chunk_offset, length, found_key.offset);
2129                 btrfs_put_block_group(cache);
2130                 if (ret)
2131                         break;
2132
2133                 key.offset = found_key.offset + length;
2134                 btrfs_release_path(path);
2135         }
2136
2137         btrfs_free_path(path);
2138
2139         /*
2140          * ret can still be 1 from search_slot or next_leaf,
2141          * that's not an error
2142          */
2143         return ret < 0 ? ret : 0;
2144 }
2145
2146 static noinline_for_stack int scrub_supers(struct scrub_dev *sdev)
2147 {
2148         int     i;
2149         u64     bytenr;
2150         u64     gen;
2151         int     ret;
2152         struct btrfs_device *device = sdev->dev;
2153         struct btrfs_root *root = device->dev_root;
2154
2155         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
2156                 return -EIO;
2157
2158         gen = root->fs_info->last_trans_committed;
2159
2160         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2161                 bytenr = btrfs_sb_offset(i);
2162                 if (bytenr + BTRFS_SUPER_INFO_SIZE > device->total_bytes)
2163                         break;
2164
2165                 ret = scrub_pages(sdev, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2166                                      BTRFS_EXTENT_FLAG_SUPER, gen, i, NULL, 1);
2167                 if (ret)
2168                         return ret;
2169         }
2170         wait_event(sdev->list_wait, atomic_read(&sdev->in_flight) == 0);
2171
2172         return 0;
2173 }
2174
2175 /*
2176  * get a reference count on fs_info->scrub_workers. start worker if necessary
2177  */
2178 static noinline_for_stack int scrub_workers_get(struct btrfs_root *root)
2179 {
2180         struct btrfs_fs_info *fs_info = root->fs_info;
2181         int ret = 0;
2182
2183         mutex_lock(&fs_info->scrub_lock);
2184         if (fs_info->scrub_workers_refcnt == 0) {
2185                 btrfs_init_workers(&fs_info->scrub_workers, "scrub",
2186                            fs_info->thread_pool_size, &fs_info->generic_worker);
2187                 fs_info->scrub_workers.idle_thresh = 4;
2188                 ret = btrfs_start_workers(&fs_info->scrub_workers);
2189                 if (ret)
2190                         goto out;
2191         }
2192         ++fs_info->scrub_workers_refcnt;
2193 out:
2194         mutex_unlock(&fs_info->scrub_lock);
2195
2196         return ret;
2197 }
2198
2199 static noinline_for_stack void scrub_workers_put(struct btrfs_root *root)
2200 {
2201         struct btrfs_fs_info *fs_info = root->fs_info;
2202
2203         mutex_lock(&fs_info->scrub_lock);
2204         if (--fs_info->scrub_workers_refcnt == 0)
2205                 btrfs_stop_workers(&fs_info->scrub_workers);
2206         WARN_ON(fs_info->scrub_workers_refcnt < 0);
2207         mutex_unlock(&fs_info->scrub_lock);
2208 }
2209
2210
2211 int btrfs_scrub_dev(struct btrfs_root *root, u64 devid, u64 start, u64 end,
2212                     struct btrfs_scrub_progress *progress, int readonly)
2213 {
2214         struct scrub_dev *sdev;
2215         struct btrfs_fs_info *fs_info = root->fs_info;
2216         int ret;
2217         struct btrfs_device *dev;
2218
2219         if (btrfs_fs_closing(root->fs_info))
2220                 return -EINVAL;
2221
2222         /*
2223          * check some assumptions
2224          */
2225         if (root->nodesize != root->leafsize) {
2226                 printk(KERN_ERR
2227                        "btrfs_scrub: size assumption nodesize == leafsize (%d == %d) fails\n",
2228                        root->nodesize, root->leafsize);
2229                 return -EINVAL;
2230         }
2231
2232         if (root->nodesize > BTRFS_STRIPE_LEN) {
2233                 /*
2234                  * in this case scrub is unable to calculate the checksum
2235                  * the way scrub is implemented. Do not handle this
2236                  * situation at all because it won't ever happen.
2237                  */
2238                 printk(KERN_ERR
2239                        "btrfs_scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails\n",
2240                        root->nodesize, BTRFS_STRIPE_LEN);
2241                 return -EINVAL;
2242         }
2243
2244         if (root->sectorsize != PAGE_SIZE) {
2245                 /* not supported for data w/o checksums */
2246                 printk(KERN_ERR
2247                        "btrfs_scrub: size assumption sectorsize != PAGE_SIZE (%d != %lld) fails\n",
2248                        root->sectorsize, (unsigned long long)PAGE_SIZE);
2249                 return -EINVAL;
2250         }
2251
2252         ret = scrub_workers_get(root);
2253         if (ret)
2254                 return ret;
2255
2256         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
2257         dev = btrfs_find_device(root, devid, NULL, NULL);
2258         if (!dev || dev->missing) {
2259                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2260                 scrub_workers_put(root);
2261                 return -ENODEV;
2262         }
2263         mutex_lock(&fs_info->scrub_lock);
2264
2265         if (!dev->in_fs_metadata) {
2266                 mutex_unlock(&fs_info->scrub_lock);
2267                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2268                 scrub_workers_put(root);
2269                 return -ENODEV;
2270         }
2271
2272         if (dev->scrub_device) {
2273                 mutex_unlock(&fs_info->scrub_lock);
2274                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2275                 scrub_workers_put(root);
2276                 return -EINPROGRESS;
2277         }
2278         sdev = scrub_setup_dev(dev);
2279         if (IS_ERR(sdev)) {
2280                 mutex_unlock(&fs_info->scrub_lock);
2281                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2282                 scrub_workers_put(root);
2283                 return PTR_ERR(sdev);
2284         }
2285         sdev->readonly = readonly;
2286         dev->scrub_device = sdev;
2287
2288         atomic_inc(&fs_info->scrubs_running);
2289         mutex_unlock(&fs_info->scrub_lock);
2290         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2291
2292         down_read(&fs_info->scrub_super_lock);
2293         ret = scrub_supers(sdev);
2294         up_read(&fs_info->scrub_super_lock);
2295
2296         if (!ret)
2297                 ret = scrub_enumerate_chunks(sdev, start, end);
2298
2299         wait_event(sdev->list_wait, atomic_read(&sdev->in_flight) == 0);
2300         atomic_dec(&fs_info->scrubs_running);
2301         wake_up(&fs_info->scrub_pause_wait);
2302
2303         wait_event(sdev->list_wait, atomic_read(&sdev->fixup_cnt) == 0);
2304
2305         if (progress)
2306                 memcpy(progress, &sdev->stat, sizeof(*progress));
2307
2308         mutex_lock(&fs_info->scrub_lock);
2309         dev->scrub_device = NULL;
2310         mutex_unlock(&fs_info->scrub_lock);
2311
2312         scrub_free_dev(sdev);
2313         scrub_workers_put(root);
2314
2315         return ret;
2316 }
2317
2318 void btrfs_scrub_pause(struct btrfs_root *root)
2319 {
2320         struct btrfs_fs_info *fs_info = root->fs_info;
2321
2322         mutex_lock(&fs_info->scrub_lock);
2323         atomic_inc(&fs_info->scrub_pause_req);
2324         while (atomic_read(&fs_info->scrubs_paused) !=
2325                atomic_read(&fs_info->scrubs_running)) {
2326                 mutex_unlock(&fs_info->scrub_lock);
2327                 wait_event(fs_info->scrub_pause_wait,
2328                            atomic_read(&fs_info->scrubs_paused) ==
2329                            atomic_read(&fs_info->scrubs_running));
2330                 mutex_lock(&fs_info->scrub_lock);
2331         }
2332         mutex_unlock(&fs_info->scrub_lock);
2333 }
2334
2335 void btrfs_scrub_continue(struct btrfs_root *root)
2336 {
2337         struct btrfs_fs_info *fs_info = root->fs_info;
2338
2339         atomic_dec(&fs_info->scrub_pause_req);
2340         wake_up(&fs_info->scrub_pause_wait);
2341 }
2342
2343 void btrfs_scrub_pause_super(struct btrfs_root *root)
2344 {
2345         down_write(&root->fs_info->scrub_super_lock);
2346 }
2347
2348 void btrfs_scrub_continue_super(struct btrfs_root *root)
2349 {
2350         up_write(&root->fs_info->scrub_super_lock);
2351 }
2352
2353 int __btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
2354 {
2355
2356         mutex_lock(&fs_info->scrub_lock);
2357         if (!atomic_read(&fs_info->scrubs_running)) {
2358                 mutex_unlock(&fs_info->scrub_lock);
2359                 return -ENOTCONN;
2360         }
2361
2362         atomic_inc(&fs_info->scrub_cancel_req);
2363         while (atomic_read(&fs_info->scrubs_running)) {
2364                 mutex_unlock(&fs_info->scrub_lock);
2365                 wait_event(fs_info->scrub_pause_wait,
2366                            atomic_read(&fs_info->scrubs_running) == 0);
2367                 mutex_lock(&fs_info->scrub_lock);
2368         }
2369         atomic_dec(&fs_info->scrub_cancel_req);
2370         mutex_unlock(&fs_info->scrub_lock);
2371
2372         return 0;
2373 }
2374
2375 int btrfs_scrub_cancel(struct btrfs_root *root)
2376 {
2377         return __btrfs_scrub_cancel(root->fs_info);
2378 }
2379
2380 int btrfs_scrub_cancel_dev(struct btrfs_root *root, struct btrfs_device *dev)
2381 {
2382         struct btrfs_fs_info *fs_info = root->fs_info;
2383         struct scrub_dev *sdev;
2384
2385         mutex_lock(&fs_info->scrub_lock);
2386         sdev = dev->scrub_device;
2387         if (!sdev) {
2388                 mutex_unlock(&fs_info->scrub_lock);
2389                 return -ENOTCONN;
2390         }
2391         atomic_inc(&sdev->cancel_req);
2392         while (dev->scrub_device) {
2393                 mutex_unlock(&fs_info->scrub_lock);
2394                 wait_event(fs_info->scrub_pause_wait,
2395                            dev->scrub_device == NULL);
2396                 mutex_lock(&fs_info->scrub_lock);
2397         }
2398         mutex_unlock(&fs_info->scrub_lock);
2399
2400         return 0;
2401 }
2402
2403 int btrfs_scrub_cancel_devid(struct btrfs_root *root, u64 devid)
2404 {
2405         struct btrfs_fs_info *fs_info = root->fs_info;
2406         struct btrfs_device *dev;
2407         int ret;
2408
2409         /*
2410          * we have to hold the device_list_mutex here so the device
2411          * does not go away in cancel_dev. FIXME: find a better solution
2412          */
2413         mutex_lock(&fs_info->fs_devices->device_list_mutex);
2414         dev = btrfs_find_device(root, devid, NULL, NULL);
2415         if (!dev) {
2416                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2417                 return -ENODEV;
2418         }
2419         ret = btrfs_scrub_cancel_dev(root, dev);
2420         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2421
2422         return ret;
2423 }
2424
2425 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
2426                          struct btrfs_scrub_progress *progress)
2427 {
2428         struct btrfs_device *dev;
2429         struct scrub_dev *sdev = NULL;
2430
2431         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
2432         dev = btrfs_find_device(root, devid, NULL, NULL);
2433         if (dev)
2434                 sdev = dev->scrub_device;
2435         if (sdev)
2436                 memcpy(progress, &sdev->stat, sizeof(*progress));
2437         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2438
2439         return dev ? (sdev ? 0 : -ENOTCONN) : -ENODEV;
2440 }