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 #include "rcu-string.h"
  30
  31 /*
  32  * This is only the first step towards a full-features scrub. It reads all
  33  * extent and super block and verifies the checksums. In case a bad checksum
  34  * is found or the extent cannot be read, good data will be written back if
  35  * any can be found.
  36  *
  37  * Future enhancements:
  38  *  - In case an unrepairable extent is encountered, track which files are
  39  *    affected and report them
  40  *  - track and record media errors, throw out bad devices
  41  *  - add a mode to also read unallocated space
  42  */
  43
  44 struct scrub_block;
  45 struct scrub_dev;
  46
  47 #define SCRUB_PAGES_PER_BIO     16      /* 64k per bio */
  48 #define SCRUB_BIOS_PER_DEV      16      /* 1 MB per device in flight */
  49 #define SCRUB_MAX_PAGES_PER_BLOCK       16      /* 64k per node/leaf/sector */
  50
  51 struct scrub_page {
  52         struct scrub_block      *sblock;
  53         struct page             *page;
  54         struct btrfs_device     *dev;
  55         u64                     flags;  /* extent flags */
  56         u64                     generation;
  57         u64                     logical;
  58         u64                     physical;
  59         struct {
  60                 unsigned int    mirror_num:8;
  61                 unsigned int    have_csum:1;
  62                 unsigned int    io_error:1;
  63         };
  64         u8                      csum[BTRFS_CSUM_SIZE];
  65 };
  66
  67 struct scrub_bio {
  68         int                     index;
  69         struct scrub_dev        *sdev;
  70         struct bio              *bio;
  71         int                     err;
  72         u64                     logical;
  73         u64                     physical;
  74         struct scrub_page       *pagev[SCRUB_PAGES_PER_BIO];
  75         int                     page_count;
  76         int                     next_free;
  77         struct btrfs_work       work;
  78 };
  79
  80 struct scrub_block {
  81         struct scrub_page       pagev[SCRUB_MAX_PAGES_PER_BLOCK];
  82         int                     page_count;
  83         atomic_t                outstanding_pages;
  84         atomic_t                ref_count; /* free mem on transition to zero */
  85         struct scrub_dev        *sdev;
  86         struct {
  87                 unsigned int    header_error:1;
  88                 unsigned int    checksum_error:1;
  89                 unsigned int    no_io_error_seen:1;
  90                 unsigned int    generation_error:1; /* also sets header_error */
  91         };
  92 };
  93
  94 struct scrub_dev {
  95         struct scrub_bio        *bios[SCRUB_BIOS_PER_DEV];
  96         struct btrfs_device     *dev;
  97         int                     first_free;
  98         int                     curr;
  99         atomic_t                in_flight;
 100         atomic_t                fixup_cnt;
 101         spinlock_t              list_lock;
 102         wait_queue_head_t       list_wait;
 103         u16                     csum_size;
 104         struct list_head        csum_list;
 105         atomic_t                cancel_req;
 106         int                     readonly;
 107         int                     pages_per_bio; /* <= SCRUB_PAGES_PER_BIO */
 108         u32                     sectorsize;
 109         u32                     nodesize;
 110         u32                     leafsize;
 111         /*
 112          * statistics
 113          */
 114         struct btrfs_scrub_progress stat;
 115         spinlock_t              stat_lock;
 116 };
 117
 118 struct scrub_fixup_nodatasum {
 119         struct scrub_dev        *sdev;
 120         u64                     logical;
 121         struct btrfs_root       *root;
 122         struct btrfs_work       work;
 123         int                     mirror_num;
 124 };
 125
 126 struct scrub_warning {
 127         struct btrfs_path       *path;
 128         u64                     extent_item_size;
 129         char                    *scratch_buf;
 130         char                    *msg_buf;
 131         const char              *errstr;
 132         sector_t                sector;
 133         u64                     logical;
 134         struct btrfs_device     *dev;
 135         int                     msg_bufsize;
 136         int                     scratch_bufsize;
 137 };
 138
 139
 140 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
 141 static int scrub_setup_recheck_block(struct scrub_dev *sdev,
 142                                      struct btrfs_mapping_tree *map_tree,
 143                                      u64 length, u64 logical,
 144                                      struct scrub_block *sblock);
 145 static int scrub_recheck_block(struct btrfs_fs_info *fs_info,
 146                                struct scrub_block *sblock, int is_metadata,
 147                                int have_csum, u8 *csum, u64 generation,
 148                                u16 csum_size);
 149 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
 150                                          struct scrub_block *sblock,
 151                                          int is_metadata, int have_csum,
 152                                          const u8 *csum, u64 generation,
 153                                          u16 csum_size);
 154 static void scrub_complete_bio_end_io(struct bio *bio, int err);
 155 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
 156                                              struct scrub_block *sblock_good,
 157                                              int force_write);
 158 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
 159                                             struct scrub_block *sblock_good,
 160                                             int page_num, int force_write);
 161 static int scrub_checksum_data(struct scrub_block *sblock);
 162 static int scrub_checksum_tree_block(struct scrub_block *sblock);
 163 static int scrub_checksum_super(struct scrub_block *sblock);
 164 static void scrub_block_get(struct scrub_block *sblock);
 165 static void scrub_block_put(struct scrub_block *sblock);
 166 static int scrub_add_page_to_bio(struct scrub_dev *sdev,
 167                                  struct scrub_page *spage);
 168 static int scrub_pages(struct scrub_dev *sdev, u64 logical, u64 len,
 169                        u64 physical, u64 flags, u64 gen, int mirror_num,
 170                        u8 *csum, int force);
 171 static void scrub_bio_end_io(struct bio *bio, int err);
 172 static void scrub_bio_end_io_worker(struct btrfs_work *work);
 173 static void scrub_block_complete(struct scrub_block *sblock);
 174
 175
 176 static void scrub_free_csums(struct scrub_dev *sdev)
 177 {
 178         while (!list_empty(&sdev->csum_list)) {
 179                 struct btrfs_ordered_sum *sum;
 180                 sum = list_first_entry(&sdev->csum_list,
 181                                        struct btrfs_ordered_sum, list);
 182                 list_del(&sum->list);
 183                 kfree(sum);
 184         }
 185 }
 186
 187 static noinline_for_stack void scrub_free_dev(struct scrub_dev *sdev)
 188 {
 189         int i;
 190
 191         if (!sdev)
 192                 return;
 193
 194         /* this can happen when scrub is cancelled */
 195         if (sdev->curr != -1) {
 196                 struct scrub_bio *sbio = sdev->bios[sdev->curr];
 197
 198                 for (i = 0; i < sbio->page_count; i++) {
 199                         BUG_ON(!sbio->pagev[i]);
 200                         BUG_ON(!sbio->pagev[i]->page);
 201                         scrub_block_put(sbio->pagev[i]->sblock);
 202                 }
 203                 bio_put(sbio->bio);
 204         }
 205
 206         for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
 207                 struct scrub_bio *sbio = sdev->bios[i];
 208
 209                 if (!sbio)
 210                         break;
 211                 kfree(sbio);
 212         }
 213
 214         scrub_free_csums(sdev);
 215         kfree(sdev);
 216 }
 217
 218 static noinline_for_stack
 219 struct scrub_dev *scrub_setup_dev(struct btrfs_device *dev)
 220 {
 221         struct scrub_dev *sdev;
 222         int             i;
 223         struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
 224         int pages_per_bio;
 225
 226         pages_per_bio = min_t(int, SCRUB_PAGES_PER_BIO,
 227                               bio_get_nr_vecs(dev->bdev));
 228         sdev = kzalloc(sizeof(*sdev), GFP_NOFS);
 229         if (!sdev)
 230                 goto nomem;
 231         sdev->dev = dev;
 232         sdev->pages_per_bio = pages_per_bio;
 233         sdev->curr = -1;
 234         for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
 235                 struct scrub_bio *sbio;
 236
 237                 sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
 238                 if (!sbio)
 239                         goto nomem;
 240                 sdev->bios[i] = sbio;
 241
 242                 sbio->index = i;
 243                 sbio->sdev = sdev;
 244                 sbio->page_count = 0;
 245                 sbio->work.func = scrub_bio_end_io_worker;
 246
 247                 if (i != SCRUB_BIOS_PER_DEV-1)
 248                         sdev->bios[i]->next_free = i + 1;
 249                 else
 250                         sdev->bios[i]->next_free = -1;
 251         }
 252         sdev->first_free = 0;
 253         sdev->nodesize = dev->dev_root->nodesize;
 254         sdev->leafsize = dev->dev_root->leafsize;
 255         sdev->sectorsize = dev->dev_root->sectorsize;
 256         atomic_set(&sdev->in_flight, 0);
 257         atomic_set(&sdev->fixup_cnt, 0);
 258         atomic_set(&sdev->cancel_req, 0);
 259         sdev->csum_size = btrfs_super_csum_size(fs_info->super_copy);
 260         INIT_LIST_HEAD(&sdev->csum_list);
 261
 262         spin_lock_init(&sdev->list_lock);
 263         spin_lock_init(&sdev->stat_lock);
 264         init_waitqueue_head(&sdev->list_wait);
 265         return sdev;
 266
 267 nomem:
 268         scrub_free_dev(sdev);
 269         return ERR_PTR(-ENOMEM);
 270 }
 271
 272 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root, void *ctx)
 273 {
 274         u64 isize;
 275         u32 nlink;
 276         int ret;
 277         int i;
 278         struct extent_buffer *eb;
 279         struct btrfs_inode_item *inode_item;
 280         struct scrub_warning *swarn = ctx;
 281         struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
 282         struct inode_fs_paths *ipath = NULL;
 283         struct btrfs_root *local_root;
 284         struct btrfs_key root_key;
 285
 286         root_key.objectid = root;
 287         root_key.type = BTRFS_ROOT_ITEM_KEY;
 288         root_key.offset = (u64)-1;
 289         local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
 290         if (IS_ERR(local_root)) {
 291                 ret = PTR_ERR(local_root);
 292                 goto err;
 293         }
 294
 295         ret = inode_item_info(inum, 0, local_root, swarn->path);
 296         if (ret) {
 297                 btrfs_release_path(swarn->path);
 298                 goto err;
 299         }
 300
 301         eb = swarn->path->nodes[0];
 302         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 303                                         struct btrfs_inode_item);
 304         isize = btrfs_inode_size(eb, inode_item);
 305         nlink = btrfs_inode_nlink(eb, inode_item);
 306         btrfs_release_path(swarn->path);
 307
 308         ipath = init_ipath(4096, local_root, swarn->path);
 309         if (IS_ERR(ipath)) {
 310                 ret = PTR_ERR(ipath);
 311                 ipath = NULL;
 312                 goto err;
 313         }
 314         ret = paths_from_inode(inum, ipath);
 315
 316         if (ret < 0)
 317                 goto err;
 318
 319         /*
 320          * we deliberately ignore the bit ipath might have been too small to
 321          * hold all of the paths here
 322          */
 323         for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 324                 printk_in_rcu(KERN_WARNING "btrfs: %s at logical %llu on dev "
 325                         "%s, sector %llu, root %llu, inode %llu, offset %llu, "
 326                         "length %llu, links %u (path: %s)\n", swarn->errstr,
 327                         swarn->logical, rcu_str_deref(swarn->dev->name),
 328                         (unsigned long long)swarn->sector, root, inum, offset,
 329                         min(isize - offset, (u64)PAGE_SIZE), nlink,
 330                         (char *)(unsigned long)ipath->fspath->val[i]);
 331
 332         free_ipath(ipath);
 333         return 0;
 334
 335 err:
 336         printk_in_rcu(KERN_WARNING "btrfs: %s at logical %llu on dev "
 337                 "%s, sector %llu, root %llu, inode %llu, offset %llu: path "
 338                 "resolving failed with ret=%d\n", swarn->errstr,
 339                 swarn->logical, rcu_str_deref(swarn->dev->name),
 340                 (unsigned long long)swarn->sector, root, inum, offset, ret);
 341
 342         free_ipath(ipath);
 343         return 0;
 344 }
 345
 346 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
 347 {
 348         struct btrfs_device *dev = sblock->sdev->dev;
 349         struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
 350         struct btrfs_path *path;
 351         struct btrfs_key found_key;
 352         struct extent_buffer *eb;
 353         struct btrfs_extent_item *ei;
 354         struct scrub_warning swarn;
 355         u32 item_size;
 356         int ret;
 357         u64 ref_root;
 358         u8 ref_level;
 359         unsigned long ptr = 0;
 360         const int bufsize = 4096;
 361         u64 extent_item_pos;
 362
 363         path = btrfs_alloc_path();
 364
 365         swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
 366         swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
 367         BUG_ON(sblock->page_count < 1);
 368         swarn.sector = (sblock->pagev[0].physical) >> 9;
 369         swarn.logical = sblock->pagev[0].logical;
 370         swarn.errstr = errstr;
 371         swarn.dev = dev;
 372         swarn.msg_bufsize = bufsize;
 373         swarn.scratch_bufsize = bufsize;
 374
 375         if (!path || !swarn.scratch_buf || !swarn.msg_buf)
 376                 goto out;
 377
 378         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key);
 379         if (ret < 0)
 380                 goto out;
 381
 382         extent_item_pos = swarn.logical - found_key.objectid;
 383         swarn.extent_item_size = found_key.offset;
 384
 385         eb = path->nodes[0];
 386         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 387         item_size = btrfs_item_size_nr(eb, path->slots[0]);
 388         btrfs_release_path(path);
 389
 390         if (ret & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 391                 do {
 392                         ret = tree_backref_for_extent(&ptr, eb, ei, item_size,
 393                                                         &ref_root, &ref_level);
 394                         printk_in_rcu(KERN_WARNING
 395                                 "btrfs: %s at logical %llu on dev %s, "
 396                                 "sector %llu: metadata %s (level %d) in tree "
 397                                 "%llu\n", errstr, swarn.logical,
 398                                 rcu_str_deref(dev->name),
 399                                 (unsigned long long)swarn.sector,
 400                                 ref_level ? "node" : "leaf",
 401                                 ret < 0 ? -1 : ref_level,
 402                                 ret < 0 ? -1 : ref_root);
 403                 } while (ret != 1);
 404         } else {
 405                 swarn.path = path;
 406                 iterate_extent_inodes(fs_info, found_key.objectid,
 407                                         extent_item_pos, 1,
 408                                         scrub_print_warning_inode, &swarn);
 409         }
 410
 411 out:
 412         btrfs_free_path(path);
 413         kfree(swarn.scratch_buf);
 414         kfree(swarn.msg_buf);
 415 }
 416
 417 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *ctx)
 418 {
 419         struct page *page = NULL;
 420         unsigned long index;
 421         struct scrub_fixup_nodatasum *fixup = ctx;
 422         int ret;
 423         int corrected = 0;
 424         struct btrfs_key key;
 425         struct inode *inode = NULL;
 426         u64 end = offset + PAGE_SIZE - 1;
 427         struct btrfs_root *local_root;
 428
 429         key.objectid = root;
 430         key.type = BTRFS_ROOT_ITEM_KEY;
 431         key.offset = (u64)-1;
 432         local_root = btrfs_read_fs_root_no_name(fixup->root->fs_info, &key);
 433         if (IS_ERR(local_root))
 434                 return PTR_ERR(local_root);
 435
 436         key.type = BTRFS_INODE_ITEM_KEY;
 437         key.objectid = inum;
 438         key.offset = 0;
 439         inode = btrfs_iget(fixup->root->fs_info->sb, &key, local_root, NULL);
 440         if (IS_ERR(inode))
 441                 return PTR_ERR(inode);
 442
 443         index = offset >> PAGE_CACHE_SHIFT;
 444
 445         page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
 446         if (!page) {
 447                 ret = -ENOMEM;
 448                 goto out;
 449         }
 450
 451         if (PageUptodate(page)) {
 452                 struct btrfs_mapping_tree *map_tree;
 453                 if (PageDirty(page)) {
 454                         /*
 455                          * we need to write the data to the defect sector. the
 456                          * data that was in that sector is not in memory,
 457                          * because the page was modified. we must not write the
 458                          * modified page to that sector.
 459                          *
 460                          * TODO: what could be done here: wait for the delalloc
 461                          *       runner to write out that page (might involve
 462                          *       COW) and see whether the sector is still
 463                          *       referenced afterwards.
 464                          *
 465                          * For the meantime, we'll treat this error
 466                          * incorrectable, although there is a chance that a
 467                          * later scrub will find the bad sector again and that
 468                          * there's no dirty page in memory, then.
 469                          */
 470                         ret = -EIO;
 471                         goto out;
 472                 }
 473                 map_tree = &BTRFS_I(inode)->root->fs_info->mapping_tree;
 474                 ret = repair_io_failure(map_tree, offset, PAGE_SIZE,
 475                                         fixup->logical, page,
 476                                         fixup->mirror_num);
 477                 unlock_page(page);
 478                 corrected = !ret;
 479         } else {
 480                 /*
 481                  * we need to get good data first. the general readpage path
 482                  * will call repair_io_failure for us, we just have to make
 483                  * sure we read the bad mirror.
 484                  */
 485                 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 486                                         EXTENT_DAMAGED, GFP_NOFS);
 487                 if (ret) {
 488                         /* set_extent_bits should give proper error */
 489                         WARN_ON(ret > 0);
 490                         if (ret > 0)
 491                                 ret = -EFAULT;
 492                         goto out;
 493                 }
 494
 495                 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
 496                                                 btrfs_get_extent,
 497                                                 fixup->mirror_num);
 498                 wait_on_page_locked(page);
 499
 500                 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
 501                                                 end, EXTENT_DAMAGED, 0, NULL);
 502                 if (!corrected)
 503                         clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 504                                                 EXTENT_DAMAGED, GFP_NOFS);
 505         }
 506
 507 out:
 508         if (page)
 509                 put_page(page);
 510         if (inode)
 511                 iput(inode);
 512
 513         if (ret < 0)
 514                 return ret;
 515
 516         if (ret == 0 && corrected) {
 517                 /*
 518                  * we only need to call readpage for one of the inodes belonging
 519                  * to this extent. so make iterate_extent_inodes stop
 520                  */
 521                 return 1;
 522         }
 523
 524         return -EIO;
 525 }
 526
 527 static void scrub_fixup_nodatasum(struct btrfs_work *work)
 528 {
 529         int ret;
 530         struct scrub_fixup_nodatasum *fixup;
 531         struct scrub_dev *sdev;
 532         struct btrfs_trans_handle *trans = NULL;
 533         struct btrfs_fs_info *fs_info;
 534         struct btrfs_path *path;
 535         int uncorrectable = 0;
 536
 537         fixup = container_of(work, struct scrub_fixup_nodatasum, work);
 538         sdev = fixup->sdev;
 539         fs_info = fixup->root->fs_info;
 540
 541         path = btrfs_alloc_path();
 542         if (!path) {
 543                 spin_lock(&sdev->stat_lock);
 544                 ++sdev->stat.malloc_errors;
 545                 spin_unlock(&sdev->stat_lock);
 546                 uncorrectable = 1;
 547                 goto out;
 548         }
 549
 550         trans = btrfs_join_transaction(fixup->root);
 551         if (IS_ERR(trans)) {
 552                 uncorrectable = 1;
 553                 goto out;
 554         }
 555
 556         /*
 557          * the idea is to trigger a regular read through the standard path. we
 558          * read a page from the (failed) logical address by specifying the
 559          * corresponding copynum of the failed sector. thus, that readpage is
 560          * expected to fail.
 561          * that is the point where on-the-fly error correction will kick in
 562          * (once it's finished) and rewrite the failed sector if a good copy
 563          * can be found.
 564          */
 565         ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
 566                                                 path, scrub_fixup_readpage,
 567                                                 fixup);
 568         if (ret < 0) {
 569                 uncorrectable = 1;
 570                 goto out;
 571         }
 572         WARN_ON(ret != 1);
 573
 574         spin_lock(&sdev->stat_lock);
 575         ++sdev->stat.corrected_errors;
 576         spin_unlock(&sdev->stat_lock);
 577
 578 out:
 579         if (trans && !IS_ERR(trans))
 580                 btrfs_end_transaction(trans, fixup->root);
 581         if (uncorrectable) {
 582                 spin_lock(&sdev->stat_lock);
 583                 ++sdev->stat.uncorrectable_errors;
 584                 spin_unlock(&sdev->stat_lock);
 585
 586                 printk_ratelimited_in_rcu(KERN_ERR
 587                         "btrfs: unable to fixup (nodatasum) error at logical %llu on dev %s\n",
 588                         (unsigned long long)fixup->logical,
 589                         rcu_str_deref(sdev->dev->name));
 590         }
 591
 592         btrfs_free_path(path);
 593         kfree(fixup);
 594
 595         /* see caller why we're pretending to be paused in the scrub counters */
 596         mutex_lock(&fs_info->scrub_lock);
 597         atomic_dec(&fs_info->scrubs_running);
 598         atomic_dec(&fs_info->scrubs_paused);
 599         mutex_unlock(&fs_info->scrub_lock);
 600         atomic_dec(&sdev->fixup_cnt);
 601         wake_up(&fs_info->scrub_pause_wait);
 602         wake_up(&sdev->list_wait);
 603 }
 604
 605 /*
 606  * scrub_handle_errored_block gets called when either verification of the
 607  * pages failed or the bio failed to read, e.g. with EIO. In the latter
 608  * case, this function handles all pages in the bio, even though only one
 609  * may be bad.
 610  * The goal of this function is to repair the errored block by using the
 611  * contents of one of the mirrors.
 612  */
 613 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
 614 {
 615         struct scrub_dev *sdev = sblock_to_check->sdev;
 616         struct btrfs_fs_info *fs_info;
 617         u64 length;
 618         u64 logical;
 619         u64 generation;
 620         unsigned int failed_mirror_index;
 621         unsigned int is_metadata;
 622         unsigned int have_csum;
 623         u8 *csum;
 624         struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
 625         struct scrub_block *sblock_bad;
 626         int ret;
 627         int mirror_index;
 628         int page_num;
 629         int success;
 630         static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
 631                                       DEFAULT_RATELIMIT_BURST);
 632
 633         BUG_ON(sblock_to_check->page_count < 1);
 634         fs_info = sdev->dev->dev_root->fs_info;
 635         length = sblock_to_check->page_count * PAGE_SIZE;
 636         logical = sblock_to_check->pagev[0].logical;
 637         generation = sblock_to_check->pagev[0].generation;
 638         BUG_ON(sblock_to_check->pagev[0].mirror_num < 1);
 639         failed_mirror_index = sblock_to_check->pagev[0].mirror_num - 1;
 640         is_metadata = !(sblock_to_check->pagev[0].flags &
 641                         BTRFS_EXTENT_FLAG_DATA);
 642         have_csum = sblock_to_check->pagev[0].have_csum;
 643         csum = sblock_to_check->pagev[0].csum;
 644
 645         /*
 646          * read all mirrors one after the other. This includes to
 647          * re-read the extent or metadata block that failed (that was
 648          * the cause that this fixup code is called) another time,
 649          * page by page this time in order to know which pages
 650          * caused I/O errors and which ones are good (for all mirrors).
 651          * It is the goal to handle the situation when more than one
 652          * mirror contains I/O errors, but the errors do not
 653          * overlap, i.e. the data can be repaired by selecting the
 654          * pages from those mirrors without I/O error on the
 655          * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
 656          * would be that mirror #1 has an I/O error on the first page,
 657          * the second page is good, and mirror #2 has an I/O error on
 658          * the second page, but the first page is good.
 659          * Then the first page of the first mirror can be repaired by
 660          * taking the first page of the second mirror, and the
 661          * second page of the second mirror can be repaired by
 662          * copying the contents of the 2nd page of the 1st mirror.
 663          * One more note: if the pages of one mirror contain I/O
 664          * errors, the checksum cannot be verified. In order to get
 665          * the best data for repairing, the first attempt is to find
 666          * a mirror without I/O errors and with a validated checksum.
 667          * Only if this is not possible, the pages are picked from
 668          * mirrors with I/O errors without considering the checksum.
 669          * If the latter is the case, at the end, the checksum of the
 670          * repaired area is verified in order to correctly maintain
 671          * the statistics.
 672          */
 673
 674         sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
 675                                      sizeof(*sblocks_for_recheck),
 676                                      GFP_NOFS);
 677         if (!sblocks_for_recheck) {
 678                 spin_lock(&sdev->stat_lock);
 679                 sdev->stat.malloc_errors++;
 680                 sdev->stat.read_errors++;
 681                 sdev->stat.uncorrectable_errors++;
 682                 spin_unlock(&sdev->stat_lock);
 683                 btrfs_dev_stat_inc_and_print(sdev->dev,
 684                                              BTRFS_DEV_STAT_READ_ERRS);
 685                 goto out;
 686         }
 687
 688         /* setup the context, map the logical blocks and alloc the pages */
 689         ret = scrub_setup_recheck_block(sdev, &fs_info->mapping_tree, length,
 690                                         logical, sblocks_for_recheck);
 691         if (ret) {
 692                 spin_lock(&sdev->stat_lock);
 693                 sdev->stat.read_errors++;
 694                 sdev->stat.uncorrectable_errors++;
 695                 spin_unlock(&sdev->stat_lock);
 696                 btrfs_dev_stat_inc_and_print(sdev->dev,
 697                                              BTRFS_DEV_STAT_READ_ERRS);
 698                 goto out;
 699         }
 700         BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
 701         sblock_bad = sblocks_for_recheck + failed_mirror_index;
 702
 703         /* build and submit the bios for the failed mirror, check checksums */
 704         ret = scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
 705                                   csum, generation, sdev->csum_size);
 706         if (ret) {
 707                 spin_lock(&sdev->stat_lock);
 708                 sdev->stat.read_errors++;
 709                 sdev->stat.uncorrectable_errors++;
 710                 spin_unlock(&sdev->stat_lock);
 711                 btrfs_dev_stat_inc_and_print(sdev->dev,
 712                                              BTRFS_DEV_STAT_READ_ERRS);
 713                 goto out;
 714         }
 715
 716         if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
 717             sblock_bad->no_io_error_seen) {
 718                 /*
 719                  * the error disappeared after reading page by page, or
 720                  * the area was part of a huge bio and other parts of the
 721                  * bio caused I/O errors, or the block layer merged several
 722                  * read requests into one and the error is caused by a
 723                  * different bio (usually one of the two latter cases is
 724                  * the cause)
 725                  */
 726                 spin_lock(&sdev->stat_lock);
 727                 sdev->stat.unverified_errors++;
 728                 spin_unlock(&sdev->stat_lock);
 729
 730                 goto out;
 731         }
 732
 733         if (!sblock_bad->no_io_error_seen) {
 734                 spin_lock(&sdev->stat_lock);
 735                 sdev->stat.read_errors++;
 736                 spin_unlock(&sdev->stat_lock);
 737                 if (__ratelimit(&_rs))
 738                         scrub_print_warning("i/o error", sblock_to_check);
 739                 btrfs_dev_stat_inc_and_print(sdev->dev,
 740                                              BTRFS_DEV_STAT_READ_ERRS);
 741         } else if (sblock_bad->checksum_error) {
 742                 spin_lock(&sdev->stat_lock);
 743                 sdev->stat.csum_errors++;
 744                 spin_unlock(&sdev->stat_lock);
 745                 if (__ratelimit(&_rs))
 746                         scrub_print_warning("checksum error", sblock_to_check);
 747                 btrfs_dev_stat_inc_and_print(sdev->dev,
 748                                              BTRFS_DEV_STAT_CORRUPTION_ERRS);
 749         } else if (sblock_bad->header_error) {
 750                 spin_lock(&sdev->stat_lock);
 751                 sdev->stat.verify_errors++;
 752                 spin_unlock(&sdev->stat_lock);
 753                 if (__ratelimit(&_rs))
 754                         scrub_print_warning("checksum/header error",
 755                                             sblock_to_check);
 756                 if (sblock_bad->generation_error)
 757                         btrfs_dev_stat_inc_and_print(sdev->dev,
 758                                 BTRFS_DEV_STAT_GENERATION_ERRS);
 759                 else
 760                         btrfs_dev_stat_inc_and_print(sdev->dev,
 761                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
 762         }
 763
 764         if (sdev->readonly)
 765                 goto did_not_correct_error;
 766
 767         if (!is_metadata && !have_csum) {
 768                 struct scrub_fixup_nodatasum *fixup_nodatasum;
 769
 770                 /*
 771                  * !is_metadata and !have_csum, this means that the data
 772                  * might not be COW'ed, that it might be modified
 773                  * concurrently. The general strategy to work on the
 774                  * commit root does not help in the case when COW is not
 775                  * used.
 776                  */
 777                 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
 778                 if (!fixup_nodatasum)
 779                         goto did_not_correct_error;
 780                 fixup_nodatasum->sdev = sdev;
 781                 fixup_nodatasum->logical = logical;
 782                 fixup_nodatasum->root = fs_info->extent_root;
 783                 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
 784                 /*
 785                  * increment scrubs_running to prevent cancel requests from
 786                  * completing as long as a fixup worker is running. we must also
 787                  * increment scrubs_paused to prevent deadlocking on pause
 788                  * requests used for transactions commits (as the worker uses a
 789                  * transaction context). it is safe to regard the fixup worker
 790                  * as paused for all matters practical. effectively, we only
 791                  * avoid cancellation requests from completing.
 792                  */
 793                 mutex_lock(&fs_info->scrub_lock);
 794                 atomic_inc(&fs_info->scrubs_running);
 795                 atomic_inc(&fs_info->scrubs_paused);
 796                 mutex_unlock(&fs_info->scrub_lock);
 797                 atomic_inc(&sdev->fixup_cnt);
 798                 fixup_nodatasum->work.func = scrub_fixup_nodatasum;
 799                 btrfs_queue_worker(&fs_info->scrub_workers,
 800                                    &fixup_nodatasum->work);
 801                 goto out;
 802         }
 803
 804         /*
 805          * now build and submit the bios for the other mirrors, check
 806          * checksums
 807          */
 808         for (mirror_index = 0;
 809              mirror_index < BTRFS_MAX_MIRRORS &&
 810              sblocks_for_recheck[mirror_index].page_count > 0;
 811              mirror_index++) {
 812                 if (mirror_index == failed_mirror_index)
 813                         continue;
 814
 815                 /* build and submit the bios, check checksums */
 816                 ret = scrub_recheck_block(fs_info,
 817                                           sblocks_for_recheck + mirror_index,
 818                                           is_metadata, have_csum, csum,
 819                                           generation, sdev->csum_size);
 820                 if (ret)
 821                         goto did_not_correct_error;
 822         }
 823
 824         /*
 825          * first try to pick the mirror which is completely without I/O
 826          * errors and also does not have a checksum error.
 827          * If one is found, and if a checksum is present, the full block
 828          * that is known to contain an error is rewritten. Afterwards
 829          * the block is known to be corrected.
 830          * If a mirror is found which is completely correct, and no
 831          * checksum is present, only those pages are rewritten that had
 832          * an I/O error in the block to be repaired, since it cannot be
 833          * determined, which copy of the other pages is better (and it
 834          * could happen otherwise that a correct page would be
 835          * overwritten by a bad one).
 836          */
 837         for (mirror_index = 0;
 838              mirror_index < BTRFS_MAX_MIRRORS &&
 839              sblocks_for_recheck[mirror_index].page_count > 0;
 840              mirror_index++) {
 841                 struct scrub_block *sblock_other = sblocks_for_recheck +
 842                                                    mirror_index;
 843
 844                 if (!sblock_other->header_error &&
 845                     !sblock_other->checksum_error &&
 846                     sblock_other->no_io_error_seen) {
 847                         int force_write = is_metadata || have_csum;
 848
 849                         ret = scrub_repair_block_from_good_copy(sblock_bad,
 850                                                                 sblock_other,
 851                                                                 force_write);
 852                         if (0 == ret)
 853                                 goto corrected_error;
 854                 }
 855         }
 856
 857         /*
 858          * in case of I/O errors in the area that is supposed to be
 859          * repaired, continue by picking good copies of those pages.
 860          * Select the good pages from mirrors to rewrite bad pages from
 861          * the area to fix. Afterwards verify the checksum of the block
 862          * that is supposed to be repaired. This verification step is
 863          * only done for the purpose of statistic counting and for the
 864          * final scrub report, whether errors remain.
 865          * A perfect algorithm could make use of the checksum and try
 866          * all possible combinations of pages from the different mirrors
 867          * until the checksum verification succeeds. For example, when
 868          * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
 869          * of mirror #2 is readable but the final checksum test fails,
 870          * then the 2nd page of mirror #3 could be tried, whether now
 871          * the final checksum succeedes. But this would be a rare
 872          * exception and is therefore not implemented. At least it is
 873          * avoided that the good copy is overwritten.
 874          * A more useful improvement would be to pick the sectors
 875          * without I/O error based on sector sizes (512 bytes on legacy
 876          * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
 877          * mirror could be repaired by taking 512 byte of a different
 878          * mirror, even if other 512 byte sectors in the same PAGE_SIZE
 879          * area are unreadable.
 880          */
 881
 882         /* can only fix I/O errors from here on */
 883         if (sblock_bad->no_io_error_seen)
 884                 goto did_not_correct_error;
 885
 886         success = 1;
 887         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
 888                 struct scrub_page *page_bad = sblock_bad->pagev + page_num;
 889
 890                 if (!page_bad->io_error)
 891                         continue;
 892
 893                 for (mirror_index = 0;
 894                      mirror_index < BTRFS_MAX_MIRRORS &&
 895                      sblocks_for_recheck[mirror_index].page_count > 0;
 896                      mirror_index++) {
 897                         struct scrub_block *sblock_other = sblocks_for_recheck +
 898                                                            mirror_index;
 899                         struct scrub_page *page_other = sblock_other->pagev +
 900                                                         page_num;
 901
 902                         if (!page_other->io_error) {
 903                                 ret = scrub_repair_page_from_good_copy(
 904                                         sblock_bad, sblock_other, page_num, 0);
 905                                 if (0 == ret) {
 906                                         page_bad->io_error = 0;
 907                                         break; /* succeeded for this page */
 908                                 }
 909                         }
 910                 }
 911
 912                 if (page_bad->io_error) {
 913                         /* did not find a mirror to copy the page from */
 914                         success = 0;
 915                 }
 916         }
 917
 918         if (success) {
 919                 if (is_metadata || have_csum) {
 920                         /*
 921                          * need to verify the checksum now that all
 922                          * sectors on disk are repaired (the write
 923                          * request for data to be repaired is on its way).
 924                          * Just be lazy and use scrub_recheck_block()
 925                          * which re-reads the data before the checksum
 926                          * is verified, but most likely the data comes out
 927                          * of the page cache.
 928                          */
 929                         ret = scrub_recheck_block(fs_info, sblock_bad,
 930                                                   is_metadata, have_csum, csum,
 931                                                   generation, sdev->csum_size);
 932                         if (!ret && !sblock_bad->header_error &&
 933                             !sblock_bad->checksum_error &&
 934                             sblock_bad->no_io_error_seen)
 935                                 goto corrected_error;
 936                         else
 937                                 goto did_not_correct_error;
 938                 } else {
 939 corrected_error:
 940                         spin_lock(&sdev->stat_lock);
 941                         sdev->stat.corrected_errors++;
 942                         spin_unlock(&sdev->stat_lock);
 943                         printk_ratelimited_in_rcu(KERN_ERR
 944                                 "btrfs: fixed up error at logical %llu on dev %s\n",
 945                                 (unsigned long long)logical,
 946                                 rcu_str_deref(sdev->dev->name));
 947                 }
 948         } else {
 949 did_not_correct_error:
 950                 spin_lock(&sdev->stat_lock);
 951                 sdev->stat.uncorrectable_errors++;
 952                 spin_unlock(&sdev->stat_lock);
 953                 printk_ratelimited_in_rcu(KERN_ERR
 954                         "btrfs: unable to fixup (regular) error at logical %llu on dev %s\n",
 955                         (unsigned long long)logical,
 956                         rcu_str_deref(sdev->dev->name));
 957         }
 958
 959 out:
 960         if (sblocks_for_recheck) {
 961                 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
 962                      mirror_index++) {
 963                         struct scrub_block *sblock = sblocks_for_recheck +
 964                                                      mirror_index;
 965                         int page_index;
 966
 967                         for (page_index = 0; page_index < SCRUB_PAGES_PER_BIO;
 968                              page_index++)
 969                                 if (sblock->pagev[page_index].page)
 970                                         __free_page(
 971                                                 sblock->pagev[page_index].page);
 972                 }
 973                 kfree(sblocks_for_recheck);
 974         }
 975
 976         return 0;
 977 }
 978
 979 static int scrub_setup_recheck_block(struct scrub_dev *sdev,
 980                                      struct btrfs_mapping_tree *map_tree,
 981                                      u64 length, u64 logical,
 982                                      struct scrub_block *sblocks_for_recheck)
 983 {
 984         int page_index;
 985         int mirror_index;
 986         int ret;
 987
 988         /*
 989          * note: the three members sdev, ref_count and outstanding_pages
 990          * are not used (and not set) in the blocks that are used for
 991          * the recheck procedure
 992          */
 993
 994         page_index = 0;
 995         while (length > 0) {
 996                 u64 sublen = min_t(u64, length, PAGE_SIZE);
 997                 u64 mapped_length = sublen;
 998                 struct btrfs_bio *bbio = NULL;
 999
1000                 /*
1001                  * with a length of PAGE_SIZE, each returned stripe
1002                  * represents one mirror
1003                  */
1004                 ret = btrfs_map_block(map_tree, WRITE, logical, &mapped_length,
1005                                       &bbio, 0);
1006                 if (ret || !bbio || mapped_length < sublen) {
1007                         kfree(bbio);
1008                         return -EIO;
1009                 }
1010
1011                 BUG_ON(page_index >= SCRUB_PAGES_PER_BIO);
1012                 for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
1013                      mirror_index++) {
1014                         struct scrub_block *sblock;
1015                         struct scrub_page *page;
1016
1017                         if (mirror_index >= BTRFS_MAX_MIRRORS)
1018                                 continue;
1019
1020                         sblock = sblocks_for_recheck + mirror_index;
1021                         page = sblock->pagev + page_index;
1022                         page->logical = logical;
1023                         page->physical = bbio->stripes[mirror_index].physical;
1024                         /* for missing devices, dev->bdev is NULL */
1025                         page->dev = bbio->stripes[mirror_index].dev;
1026                         page->mirror_num = mirror_index + 1;
1027                         page->page = alloc_page(GFP_NOFS);
1028                         if (!page->page) {
1029                                 spin_lock(&sdev->stat_lock);
1030                                 sdev->stat.malloc_errors++;
1031                                 spin_unlock(&sdev->stat_lock);
1032                                 return -ENOMEM;
1033                         }
1034                         sblock->page_count++;
1035                 }
1036                 kfree(bbio);
1037                 length -= sublen;
1038                 logical += sublen;
1039                 page_index++;
1040         }
1041
1042         return 0;
1043 }
1044
1045 /*
1046  * this function will check the on disk data for checksum errors, header
1047  * errors and read I/O errors. If any I/O errors happen, the exact pages
1048  * which are errored are marked as being bad. The goal is to enable scrub
1049  * to take those pages that are not errored from all the mirrors so that
1050  * the pages that are errored in the just handled mirror can be repaired.
1051  */
1052 static int scrub_recheck_block(struct btrfs_fs_info *fs_info,
1053                                struct scrub_block *sblock, int is_metadata,
1054                                int have_csum, u8 *csum, u64 generation,
1055                                u16 csum_size)
1056 {
1057         int page_num;
1058
1059         sblock->no_io_error_seen = 1;
1060         sblock->header_error = 0;
1061         sblock->checksum_error = 0;
1062
1063         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1064                 struct bio *bio;
1065                 int ret;
1066                 struct scrub_page *page = sblock->pagev + page_num;
1067                 DECLARE_COMPLETION_ONSTACK(complete);
1068
1069                 if (page->dev->bdev == NULL) {
1070                         page->io_error = 1;
1071                         sblock->no_io_error_seen = 0;
1072                         continue;
1073                 }
1074
1075                 BUG_ON(!page->page);
1076                 bio = bio_alloc(GFP_NOFS, 1);
1077                 if (!bio)
1078                         return -EIO;
1079                 bio->bi_bdev = page->dev->bdev;
1080                 bio->bi_sector = page->physical >> 9;
1081                 bio->bi_end_io = scrub_complete_bio_end_io;
1082                 bio->bi_private = &complete;
1083
1084                 ret = bio_add_page(bio, page->page, PAGE_SIZE, 0);
1085                 if (PAGE_SIZE != ret) {
1086                         bio_put(bio);
1087                         return -EIO;
1088                 }
1089                 btrfsic_submit_bio(READ, bio);
1090
1091                 /* this will also unplug the queue */
1092                 wait_for_completion(&complete);
1093
1094                 page->io_error = !test_bit(BIO_UPTODATE, &bio->bi_flags);
1095                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1096                         sblock->no_io_error_seen = 0;
1097                 bio_put(bio);
1098         }
1099
1100         if (sblock->no_io_error_seen)
1101                 scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1102                                              have_csum, csum, generation,
1103                                              csum_size);
1104
1105         return 0;
1106 }
1107
1108 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1109                                          struct scrub_block *sblock,
1110                                          int is_metadata, int have_csum,
1111                                          const u8 *csum, u64 generation,
1112                                          u16 csum_size)
1113 {
1114         int page_num;
1115         u8 calculated_csum[BTRFS_CSUM_SIZE];
1116         u32 crc = ~(u32)0;
1117         struct btrfs_root *root = fs_info->extent_root;
1118         void *mapped_buffer;
1119
1120         BUG_ON(!sblock->pagev[0].page);
1121         if (is_metadata) {
1122                 struct btrfs_header *h;
1123
1124                 mapped_buffer = kmap_atomic(sblock->pagev[0].page);
1125                 h = (struct btrfs_header *)mapped_buffer;
1126
1127                 if (sblock->pagev[0].logical != le64_to_cpu(h->bytenr) ||
1128                     memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
1129                     memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1130                            BTRFS_UUID_SIZE)) {
1131                         sblock->header_error = 1;
1132                 } else if (generation != le64_to_cpu(h->generation)) {
1133                         sblock->header_error = 1;
1134                         sblock->generation_error = 1;
1135                 }
1136                 csum = h->csum;
1137         } else {
1138                 if (!have_csum)
1139                         return;
1140
1141                 mapped_buffer = kmap_atomic(sblock->pagev[0].page);
1142         }
1143
1144         for (page_num = 0;;) {
1145                 if (page_num == 0 && is_metadata)
1146                         crc = btrfs_csum_data(root,
1147                                 ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1148                                 crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1149                 else
1150                         crc = btrfs_csum_data(root, mapped_buffer, crc,
1151                                               PAGE_SIZE);
1152
1153                 kunmap_atomic(mapped_buffer);
1154                 page_num++;
1155                 if (page_num >= sblock->page_count)
1156                         break;
1157                 BUG_ON(!sblock->pagev[page_num].page);
1158
1159                 mapped_buffer = kmap_atomic(sblock->pagev[page_num].page);
1160         }
1161
1162         btrfs_csum_final(crc, calculated_csum);
1163         if (memcmp(calculated_csum, csum, csum_size))
1164                 sblock->checksum_error = 1;
1165 }
1166
1167 static void scrub_complete_bio_end_io(struct bio *bio, int err)
1168 {
1169         complete((struct completion *)bio->bi_private);
1170 }
1171
1172 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1173                                              struct scrub_block *sblock_good,
1174                                              int force_write)
1175 {
1176         int page_num;
1177         int ret = 0;
1178
1179         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1180                 int ret_sub;
1181
1182                 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1183                                                            sblock_good,
1184                                                            page_num,
1185                                                            force_write);
1186                 if (ret_sub)
1187                         ret = ret_sub;
1188         }
1189
1190         return ret;
1191 }
1192
1193 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1194                                             struct scrub_block *sblock_good,
1195                                             int page_num, int force_write)
1196 {
1197         struct scrub_page *page_bad = sblock_bad->pagev + page_num;
1198         struct scrub_page *page_good = sblock_good->pagev + page_num;
1199
1200         BUG_ON(sblock_bad->pagev[page_num].page == NULL);
1201         BUG_ON(sblock_good->pagev[page_num].page == NULL);
1202         if (force_write || sblock_bad->header_error ||
1203             sblock_bad->checksum_error || page_bad->io_error) {
1204                 struct bio *bio;
1205                 int ret;
1206                 DECLARE_COMPLETION_ONSTACK(complete);
1207
1208                 bio = bio_alloc(GFP_NOFS, 1);
1209                 if (!bio)
1210                         return -EIO;
1211                 bio->bi_bdev = page_bad->dev->bdev;
1212                 bio->bi_sector = page_bad->physical >> 9;
1213                 bio->bi_end_io = scrub_complete_bio_end_io;
1214                 bio->bi_private = &complete;
1215
1216                 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1217                 if (PAGE_SIZE != ret) {
1218                         bio_put(bio);
1219                         return -EIO;
1220                 }
1221                 btrfsic_submit_bio(WRITE, bio);
1222
1223                 /* this will also unplug the queue */
1224                 wait_for_completion(&complete);
1225                 if (!bio_flagged(bio, BIO_UPTODATE)) {
1226                         btrfs_dev_stat_inc_and_print(page_bad->dev,
1227                                 BTRFS_DEV_STAT_WRITE_ERRS);
1228                         bio_put(bio);
1229                         return -EIO;
1230                 }
1231                 bio_put(bio);
1232         }
1233
1234         return 0;
1235 }
1236
1237 static void scrub_checksum(struct scrub_block *sblock)
1238 {
1239         u64 flags;
1240         int ret;
1241
1242         BUG_ON(sblock->page_count < 1);
1243         flags = sblock->pagev[0].flags;
1244         ret = 0;
1245         if (flags & BTRFS_EXTENT_FLAG_DATA)
1246                 ret = scrub_checksum_data(sblock);
1247         else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1248                 ret = scrub_checksum_tree_block(sblock);
1249         else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1250                 (void)scrub_checksum_super(sblock);
1251         else
1252                 WARN_ON(1);
1253         if (ret)
1254                 scrub_handle_errored_block(sblock);
1255 }
1256
1257 static int scrub_checksum_data(struct scrub_block *sblock)
1258 {
1259         struct scrub_dev *sdev = sblock->sdev;
1260         u8 csum[BTRFS_CSUM_SIZE];
1261         u8 *on_disk_csum;
1262         struct page *page;
1263         void *buffer;
1264         u32 crc = ~(u32)0;
1265         int fail = 0;
1266         struct btrfs_root *root = sdev->dev->dev_root;
1267         u64 len;
1268         int index;
1269
1270         BUG_ON(sblock->page_count < 1);
1271         if (!sblock->pagev[0].have_csum)
1272                 return 0;
1273
1274         on_disk_csum = sblock->pagev[0].csum;
1275         page = sblock->pagev[0].page;
1276         buffer = kmap_atomic(page);
1277
1278         len = sdev->sectorsize;
1279         index = 0;
1280         for (;;) {
1281                 u64 l = min_t(u64, len, PAGE_SIZE);
1282
1283                 crc = btrfs_csum_data(root, buffer, crc, l);
1284                 kunmap_atomic(buffer);
1285                 len -= l;
1286                 if (len == 0)
1287                         break;
1288                 index++;
1289                 BUG_ON(index >= sblock->page_count);
1290                 BUG_ON(!sblock->pagev[index].page);
1291                 page = sblock->pagev[index].page;
1292                 buffer = kmap_atomic(page);
1293         }
1294
1295         btrfs_csum_final(crc, csum);
1296         if (memcmp(csum, on_disk_csum, sdev->csum_size))
1297                 fail = 1;
1298
1299         return fail;
1300 }
1301
1302 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1303 {
1304         struct scrub_dev *sdev = sblock->sdev;
1305         struct btrfs_header *h;
1306         struct btrfs_root *root = sdev->dev->dev_root;
1307         struct btrfs_fs_info *fs_info = root->fs_info;
1308         u8 calculated_csum[BTRFS_CSUM_SIZE];
1309         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1310         struct page *page;
1311         void *mapped_buffer;
1312         u64 mapped_size;
1313         void *p;
1314         u32 crc = ~(u32)0;
1315         int fail = 0;
1316         int crc_fail = 0;
1317         u64 len;
1318         int index;
1319
1320         BUG_ON(sblock->page_count < 1);
1321         page = sblock->pagev[0].page;
1322         mapped_buffer = kmap_atomic(page);
1323         h = (struct btrfs_header *)mapped_buffer;
1324         memcpy(on_disk_csum, h->csum, sdev->csum_size);
1325
1326         /*
1327          * we don't use the getter functions here, as we
1328          * a) don't have an extent buffer and
1329          * b) the page is already kmapped
1330          */
1331
1332         if (sblock->pagev[0].logical != le64_to_cpu(h->bytenr))
1333                 ++fail;
1334
1335         if (sblock->pagev[0].generation != le64_to_cpu(h->generation))
1336                 ++fail;
1337
1338         if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1339                 ++fail;
1340
1341         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1342                    BTRFS_UUID_SIZE))
1343                 ++fail;
1344
1345         BUG_ON(sdev->nodesize != sdev->leafsize);
1346         len = sdev->nodesize - BTRFS_CSUM_SIZE;
1347         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1348         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1349         index = 0;
1350         for (;;) {
1351                 u64 l = min_t(u64, len, mapped_size);
1352
1353                 crc = btrfs_csum_data(root, p, crc, l);
1354                 kunmap_atomic(mapped_buffer);
1355                 len -= l;
1356                 if (len == 0)
1357                         break;
1358                 index++;
1359                 BUG_ON(index >= sblock->page_count);
1360                 BUG_ON(!sblock->pagev[index].page);
1361                 page = sblock->pagev[index].page;
1362                 mapped_buffer = kmap_atomic(page);
1363                 mapped_size = PAGE_SIZE;
1364                 p = mapped_buffer;
1365         }
1366
1367         btrfs_csum_final(crc, calculated_csum);
1368         if (memcmp(calculated_csum, on_disk_csum, sdev->csum_size))
1369                 ++crc_fail;
1370
1371         return fail || crc_fail;
1372 }
1373
1374 static int scrub_checksum_super(struct scrub_block *sblock)
1375 {
1376         struct btrfs_super_block *s;
1377         struct scrub_dev *sdev = sblock->sdev;
1378         struct btrfs_root *root = sdev->dev->dev_root;
1379         struct btrfs_fs_info *fs_info = root->fs_info;
1380         u8 calculated_csum[BTRFS_CSUM_SIZE];
1381         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1382         struct page *page;
1383         void *mapped_buffer;
1384         u64 mapped_size;
1385         void *p;
1386         u32 crc = ~(u32)0;
1387         int fail_gen = 0;
1388         int fail_cor = 0;
1389         u64 len;
1390         int index;
1391
1392         BUG_ON(sblock->page_count < 1);
1393         page = sblock->pagev[0].page;
1394         mapped_buffer = kmap_atomic(page);
1395         s = (struct btrfs_super_block *)mapped_buffer;
1396         memcpy(on_disk_csum, s->csum, sdev->csum_size);
1397
1398         if (sblock->pagev[0].logical != le64_to_cpu(s->bytenr))
1399                 ++fail_cor;
1400
1401         if (sblock->pagev[0].generation != le64_to_cpu(s->generation))
1402                 ++fail_gen;
1403
1404         if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1405                 ++fail_cor;
1406
1407         len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1408         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1409         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1410         index = 0;
1411         for (;;) {
1412                 u64 l = min_t(u64, len, mapped_size);
1413
1414                 crc = btrfs_csum_data(root, p, crc, l);
1415                 kunmap_atomic(mapped_buffer);
1416                 len -= l;
1417                 if (len == 0)
1418                         break;
1419                 index++;
1420                 BUG_ON(index >= sblock->page_count);
1421                 BUG_ON(!sblock->pagev[index].page);
1422                 page = sblock->pagev[index].page;
1423                 mapped_buffer = kmap_atomic(page);
1424                 mapped_size = PAGE_SIZE;
1425                 p = mapped_buffer;
1426         }
1427
1428         btrfs_csum_final(crc, calculated_csum);
1429         if (memcmp(calculated_csum, on_disk_csum, sdev->csum_size))
1430                 ++fail_cor;
1431
1432         if (fail_cor + fail_gen) {
1433                 /*
1434                  * if we find an error in a super block, we just report it.
1435                  * They will get written with the next transaction commit
1436                  * anyway
1437                  */
1438                 spin_lock(&sdev->stat_lock);
1439                 ++sdev->stat.super_errors;
1440                 spin_unlock(&sdev->stat_lock);
1441                 if (fail_cor)
1442                         btrfs_dev_stat_inc_and_print(sdev->dev,
1443                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1444                 else
1445                         btrfs_dev_stat_inc_and_print(sdev->dev,
1446                                 BTRFS_DEV_STAT_GENERATION_ERRS);
1447         }
1448
1449         return fail_cor + fail_gen;
1450 }
1451
1452 static void scrub_block_get(struct scrub_block *sblock)
1453 {
1454         atomic_inc(&sblock->ref_count);
1455 }
1456
1457 static void scrub_block_put(struct scrub_block *sblock)
1458 {
1459         if (atomic_dec_and_test(&sblock->ref_count)) {
1460                 int i;
1461
1462                 for (i = 0; i < sblock->page_count; i++)
1463                         if (sblock->pagev[i].page)
1464                                 __free_page(sblock->pagev[i].page);
1465                 kfree(sblock);
1466         }
1467 }
1468
1469 static void scrub_submit(struct scrub_dev *sdev)
1470 {
1471         struct scrub_bio *sbio;
1472
1473         if (sdev->curr == -1)
1474                 return;
1475
1476         sbio = sdev->bios[sdev->curr];
1477         sdev->curr = -1;
1478         atomic_inc(&sdev->in_flight);
1479
1480         btrfsic_submit_bio(READ, sbio->bio);
1481 }
1482
1483 static int scrub_add_page_to_bio(struct scrub_dev *sdev,
1484                                  struct scrub_page *spage)
1485 {
1486         struct scrub_block *sblock = spage->sblock;
1487         struct scrub_bio *sbio;
1488         int ret;
1489
1490 again:
1491         /*
1492          * grab a fresh bio or wait for one to become available
1493          */
1494         while (sdev->curr == -1) {
1495                 spin_lock(&sdev->list_lock);
1496                 sdev->curr = sdev->first_free;
1497                 if (sdev->curr != -1) {
1498                         sdev->first_free = sdev->bios[sdev->curr]->next_free;
1499                         sdev->bios[sdev->curr]->next_free = -1;
1500                         sdev->bios[sdev->curr]->page_count = 0;
1501                         spin_unlock(&sdev->list_lock);
1502                 } else {
1503                         spin_unlock(&sdev->list_lock);
1504                         wait_event(sdev->list_wait, sdev->first_free != -1);
1505                 }
1506         }
1507         sbio = sdev->bios[sdev->curr];
1508         if (sbio->page_count == 0) {
1509                 struct bio *bio;
1510
1511                 sbio->physical = spage->physical;
1512                 sbio->logical = spage->logical;
1513                 bio = sbio->bio;
1514                 if (!bio) {
1515                         bio = bio_alloc(GFP_NOFS, sdev->pages_per_bio);
1516                         if (!bio)
1517                                 return -ENOMEM;
1518                         sbio->bio = bio;
1519                 }
1520
1521                 bio->bi_private = sbio;
1522                 bio->bi_end_io = scrub_bio_end_io;
1523                 bio->bi_bdev = sdev->dev->bdev;
1524                 bio->bi_sector = spage->physical >> 9;
1525                 sbio->err = 0;
1526         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1527                    spage->physical ||
1528                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1529                    spage->logical) {
1530                 scrub_submit(sdev);
1531                 goto again;
1532         }
1533
1534         sbio->pagev[sbio->page_count] = spage;
1535         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1536         if (ret != PAGE_SIZE) {
1537                 if (sbio->page_count < 1) {
1538                         bio_put(sbio->bio);
1539                         sbio->bio = NULL;
1540                         return -EIO;
1541                 }
1542                 scrub_submit(sdev);
1543                 goto again;
1544         }
1545
1546         scrub_block_get(sblock); /* one for the added page */
1547         atomic_inc(&sblock->outstanding_pages);
1548         sbio->page_count++;
1549         if (sbio->page_count == sdev->pages_per_bio)
1550                 scrub_submit(sdev);
1551
1552         return 0;
1553 }
1554
1555 static int scrub_pages(struct scrub_dev *sdev, u64 logical, u64 len,
1556                        u64 physical, u64 flags, u64 gen, int mirror_num,
1557                        u8 *csum, int force)
1558 {
1559         struct scrub_block *sblock;
1560         int index;
1561
1562         sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
1563         if (!sblock) {
1564                 spin_lock(&sdev->stat_lock);
1565                 sdev->stat.malloc_errors++;
1566                 spin_unlock(&sdev->stat_lock);
1567                 return -ENOMEM;
1568         }
1569
1570         /* one ref inside this function, plus one for each page later on */
1571         atomic_set(&sblock->ref_count, 1);
1572         sblock->sdev = sdev;
1573         sblock->no_io_error_seen = 1;
1574
1575         for (index = 0; len > 0; index++) {
1576                 struct scrub_page *spage = sblock->pagev + index;
1577                 u64 l = min_t(u64, len, PAGE_SIZE);
1578
1579                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
1580                 spage->page = alloc_page(GFP_NOFS);
1581                 if (!spage->page) {
1582                         spin_lock(&sdev->stat_lock);
1583                         sdev->stat.malloc_errors++;
1584                         spin_unlock(&sdev->stat_lock);
1585                         while (index > 0) {
1586                                 index--;
1587                                 __free_page(sblock->pagev[index].page);
1588                         }
1589                         kfree(sblock);
1590                         return -ENOMEM;
1591                 }
1592                 spage->sblock = sblock;
1593                 spage->dev = sdev->dev;
1594                 spage->flags = flags;
1595                 spage->generation = gen;
1596                 spage->logical = logical;
1597                 spage->physical = physical;
1598                 spage->mirror_num = mirror_num;
1599                 if (csum) {
1600                         spage->have_csum = 1;
1601                         memcpy(spage->csum, csum, sdev->csum_size);
1602                 } else {
1603                         spage->have_csum = 0;
1604                 }
1605                 sblock->page_count++;
1606                 len -= l;
1607                 logical += l;
1608                 physical += l;
1609         }
1610
1611         BUG_ON(sblock->page_count == 0);
1612         for (index = 0; index < sblock->page_count; index++) {
1613                 struct scrub_page *spage = sblock->pagev + index;
1614                 int ret;
1615
1616                 ret = scrub_add_page_to_bio(sdev, spage);
1617                 if (ret) {
1618                         scrub_block_put(sblock);
1619                         return ret;
1620                 }
1621         }
1622
1623         if (force)
1624                 scrub_submit(sdev);
1625
1626         /* last one frees, either here or in bio completion for last page */
1627         scrub_block_put(sblock);
1628         return 0;
1629 }
1630
1631 static void scrub_bio_end_io(struct bio *bio, int err)
1632 {
1633         struct scrub_bio *sbio = bio->bi_private;
1634         struct scrub_dev *sdev = sbio->sdev;
1635         struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
1636
1637         sbio->err = err;
1638         sbio->bio = bio;
1639
1640         btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
1641 }
1642
1643 static void scrub_bio_end_io_worker(struct btrfs_work *work)
1644 {
1645         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1646         struct scrub_dev *sdev = sbio->sdev;
1647         int i;
1648
1649         BUG_ON(sbio->page_count > SCRUB_PAGES_PER_BIO);
1650         if (sbio->err) {
1651                 for (i = 0; i < sbio->page_count; i++) {
1652                         struct scrub_page *spage = sbio->pagev[i];
1653
1654                         spage->io_error = 1;
1655                         spage->sblock->no_io_error_seen = 0;
1656                 }
1657         }
1658
1659         /* now complete the scrub_block items that have all pages completed */
1660         for (i = 0; i < sbio->page_count; i++) {
1661                 struct scrub_page *spage = sbio->pagev[i];
1662                 struct scrub_block *sblock = spage->sblock;
1663
1664                 if (atomic_dec_and_test(&sblock->outstanding_pages))
1665                         scrub_block_complete(sblock);
1666                 scrub_block_put(sblock);
1667         }
1668
1669         if (sbio->err) {
1670                 /* what is this good for??? */
1671                 sbio->bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1672                 sbio->bio->bi_flags |= 1 << BIO_UPTODATE;
1673                 sbio->bio->bi_phys_segments = 0;
1674                 sbio->bio->bi_idx = 0;
1675
1676                 for (i = 0; i < sbio->page_count; i++) {
1677                         struct bio_vec *bi;
1678                         bi = &sbio->bio->bi_io_vec[i];
1679                         bi->bv_offset = 0;
1680                         bi->bv_len = PAGE_SIZE;
1681                 }
1682         }
1683
1684         bio_put(sbio->bio);
1685         sbio->bio = NULL;
1686         spin_lock(&sdev->list_lock);
1687         sbio->next_free = sdev->first_free;
1688         sdev->first_free = sbio->index;
1689         spin_unlock(&sdev->list_lock);
1690         atomic_dec(&sdev->in_flight);
1691         wake_up(&sdev->list_wait);
1692 }
1693
1694 static void scrub_block_complete(struct scrub_block *sblock)
1695 {
1696         if (!sblock->no_io_error_seen)
1697                 scrub_handle_errored_block(sblock);
1698         else
1699                 scrub_checksum(sblock);
1700 }
1701
1702 static int scrub_find_csum(struct scrub_dev *sdev, u64 logical, u64 len,
1703                            u8 *csum)
1704 {
1705         struct btrfs_ordered_sum *sum = NULL;
1706         int ret = 0;
1707         unsigned long i;
1708         unsigned long num_sectors;
1709
1710         while (!list_empty(&sdev->csum_list)) {
1711                 sum = list_first_entry(&sdev->csum_list,
1712                                        struct btrfs_ordered_sum, list);
1713                 if (sum->bytenr > logical)
1714                         return 0;
1715                 if (sum->bytenr + sum->len > logical)
1716                         break;
1717
1718                 ++sdev->stat.csum_discards;
1719                 list_del(&sum->list);
1720                 kfree(sum);
1721                 sum = NULL;
1722         }
1723         if (!sum)
1724                 return 0;
1725
1726         num_sectors = sum->len / sdev->sectorsize;
1727         for (i = 0; i < num_sectors; ++i) {
1728                 if (sum->sums[i].bytenr == logical) {
1729                         memcpy(csum, &sum->sums[i].sum, sdev->csum_size);
1730                         ret = 1;
1731                         break;
1732                 }
1733         }
1734         if (ret && i == num_sectors - 1) {
1735                 list_del(&sum->list);
1736                 kfree(sum);
1737         }
1738         return ret;
1739 }
1740
1741 /* scrub extent tries to collect up to 64 kB for each bio */
1742 static int scrub_extent(struct scrub_dev *sdev, u64 logical, u64 len,
1743                         u64 physical, u64 flags, u64 gen, int mirror_num)
1744 {
1745         int ret;
1746         u8 csum[BTRFS_CSUM_SIZE];
1747         u32 blocksize;
1748
1749         if (flags & BTRFS_EXTENT_FLAG_DATA) {
1750                 blocksize = sdev->sectorsize;
1751                 spin_lock(&sdev->stat_lock);
1752                 sdev->stat.data_extents_scrubbed++;
1753                 sdev->stat.data_bytes_scrubbed += len;
1754                 spin_unlock(&sdev->stat_lock);
1755         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1756                 BUG_ON(sdev->nodesize != sdev->leafsize);
1757                 blocksize = sdev->nodesize;
1758                 spin_lock(&sdev->stat_lock);
1759                 sdev->stat.tree_extents_scrubbed++;
1760                 sdev->stat.tree_bytes_scrubbed += len;
1761                 spin_unlock(&sdev->stat_lock);
1762         } else {
1763                 blocksize = sdev->sectorsize;
1764                 BUG_ON(1);
1765         }
1766
1767         while (len) {
1768                 u64 l = min_t(u64, len, blocksize);
1769                 int have_csum = 0;
1770
1771                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
1772                         /* push csums to sbio */
1773                         have_csum = scrub_find_csum(sdev, logical, l, csum);
1774                         if (have_csum == 0)
1775                                 ++sdev->stat.no_csum;
1776                 }
1777                 ret = scrub_pages(sdev, logical, l, physical, flags, gen,
1778                                   mirror_num, have_csum ? csum : NULL, 0);
1779                 if (ret)
1780                         return ret;
1781                 len -= l;
1782                 logical += l;
1783                 physical += l;
1784         }
1785         return 0;
1786 }
1787
1788 static noinline_for_stack int scrub_stripe(struct scrub_dev *sdev,
1789         struct map_lookup *map, int num, u64 base, u64 length)
1790 {
1791         struct btrfs_path *path;
1792         struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
1793         struct btrfs_root *root = fs_info->extent_root;
1794         struct btrfs_root *csum_root = fs_info->csum_root;
1795         struct btrfs_extent_item *extent;
1796         struct blk_plug plug;
1797         u64 flags;
1798         int ret;
1799         int slot;
1800         int i;
1801         u64 nstripes;
1802         struct extent_buffer *l;
1803         struct btrfs_key key;
1804         u64 physical;
1805         u64 logical;
1806         u64 generation;
1807         int mirror_num;
1808         struct reada_control *reada1;
1809         struct reada_control *reada2;
1810         struct btrfs_key key_start;
1811         struct btrfs_key key_end;
1812
1813         u64 increment = map->stripe_len;
1814         u64 offset;
1815
1816         nstripes = length;
1817         offset = 0;
1818         do_div(nstripes, map->stripe_len);
1819         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
1820                 offset = map->stripe_len * num;
1821                 increment = map->stripe_len * map->num_stripes;
1822                 mirror_num = 1;
1823         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
1824                 int factor = map->num_stripes / map->sub_stripes;
1825                 offset = map->stripe_len * (num / map->sub_stripes);
1826                 increment = map->stripe_len * factor;
1827                 mirror_num = num % map->sub_stripes + 1;
1828         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
1829                 increment = map->stripe_len;
1830                 mirror_num = num % map->num_stripes + 1;
1831         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
1832                 increment = map->stripe_len;
1833                 mirror_num = num % map->num_stripes + 1;
1834         } else {
1835                 increment = map->stripe_len;
1836                 mirror_num = 1;
1837         }
1838
1839         path = btrfs_alloc_path();
1840         if (!path)
1841                 return -ENOMEM;
1842
1843         /*
1844          * work on commit root. The related disk blocks are static as
1845          * long as COW is applied. This means, it is save to rewrite
1846          * them to repair disk errors without any race conditions
1847          */
1848         path->search_commit_root = 1;
1849         path->skip_locking = 1;
1850
1851         /*
1852          * trigger the readahead for extent tree csum tree and wait for
1853          * completion. During readahead, the scrub is officially paused
1854          * to not hold off transaction commits
1855          */
1856         logical = base + offset;
1857
1858         wait_event(sdev->list_wait,
1859                    atomic_read(&sdev->in_flight) == 0);
1860         atomic_inc(&fs_info->scrubs_paused);
1861         wake_up(&fs_info->scrub_pause_wait);
1862
1863         /* FIXME it might be better to start readahead at commit root */
1864         key_start.objectid = logical;
1865         key_start.type = BTRFS_EXTENT_ITEM_KEY;
1866         key_start.offset = (u64)0;
1867         key_end.objectid = base + offset + nstripes * increment;
1868         key_end.type = BTRFS_EXTENT_ITEM_KEY;
1869         key_end.offset = (u64)0;
1870         reada1 = btrfs_reada_add(root, &key_start, &key_end);
1871
1872         key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
1873         key_start.type = BTRFS_EXTENT_CSUM_KEY;
1874         key_start.offset = logical;
1875         key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
1876         key_end.type = BTRFS_EXTENT_CSUM_KEY;
1877         key_end.offset = base + offset + nstripes * increment;
1878         reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
1879
1880         if (!IS_ERR(reada1))
1881                 btrfs_reada_wait(reada1);
1882         if (!IS_ERR(reada2))
1883                 btrfs_reada_wait(reada2);
1884
1885         mutex_lock(&fs_info->scrub_lock);
1886         while (atomic_read(&fs_info->scrub_pause_req)) {
1887                 mutex_unlock(&fs_info->scrub_lock);
1888                 wait_event(fs_info->scrub_pause_wait,
1889                    atomic_read(&fs_info->scrub_pause_req) == 0);
1890                 mutex_lock(&fs_info->scrub_lock);
1891         }
1892         atomic_dec(&fs_info->scrubs_paused);
1893         mutex_unlock(&fs_info->scrub_lock);
1894         wake_up(&fs_info->scrub_pause_wait);
1895
1896         /*
1897          * collect all data csums for the stripe to avoid seeking during
1898          * the scrub. This might currently (crc32) end up to be about 1MB
1899          */
1900         blk_start_plug(&plug);
1901
1902         /*
1903          * now find all extents for each stripe and scrub them
1904          */
1905         logical = base + offset;
1906         physical = map->stripes[num].physical;
1907         ret = 0;
1908         for (i = 0; i < nstripes; ++i) {
1909                 /*
1910                  * canceled?
1911                  */
1912                 if (atomic_read(&fs_info->scrub_cancel_req) ||
1913                     atomic_read(&sdev->cancel_req)) {
1914                         ret = -ECANCELED;
1915                         goto out;
1916                 }
1917                 /*
1918                  * check to see if we have to pause
1919                  */
1920                 if (atomic_read(&fs_info->scrub_pause_req)) {
1921                         /* push queued extents */
1922                         scrub_submit(sdev);
1923                         wait_event(sdev->list_wait,
1924                                    atomic_read(&sdev->in_flight) == 0);
1925                         atomic_inc(&fs_info->scrubs_paused);
1926                         wake_up(&fs_info->scrub_pause_wait);
1927                         mutex_lock(&fs_info->scrub_lock);
1928                         while (atomic_read(&fs_info->scrub_pause_req)) {
1929                                 mutex_unlock(&fs_info->scrub_lock);
1930                                 wait_event(fs_info->scrub_pause_wait,
1931                                    atomic_read(&fs_info->scrub_pause_req) == 0);
1932                                 mutex_lock(&fs_info->scrub_lock);
1933                         }
1934                         atomic_dec(&fs_info->scrubs_paused);
1935                         mutex_unlock(&fs_info->scrub_lock);
1936                         wake_up(&fs_info->scrub_pause_wait);
1937                 }
1938
1939                 ret = btrfs_lookup_csums_range(csum_root, logical,
1940                                                logical + map->stripe_len - 1,
1941                                                &sdev->csum_list, 1);
1942                 if (ret)
1943                         goto out;
1944
1945                 key.objectid = logical;
1946                 key.type = BTRFS_EXTENT_ITEM_KEY;
1947                 key.offset = (u64)0;
1948
1949                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1950                 if (ret < 0)
1951                         goto out;
1952                 if (ret > 0) {
1953                         ret = btrfs_previous_item(root, path, 0,
1954                                                   BTRFS_EXTENT_ITEM_KEY);
1955                         if (ret < 0)
1956                                 goto out;
1957                         if (ret > 0) {
1958                                 /* there's no smaller item, so stick with the
1959                                  * larger one */
1960                                 btrfs_release_path(path);
1961                                 ret = btrfs_search_slot(NULL, root, &key,
1962                                                         path, 0, 0);
1963                                 if (ret < 0)
1964                                         goto out;
1965                         }
1966                 }
1967
1968                 while (1) {
1969                         l = path->nodes[0];
1970                         slot = path->slots[0];
1971                         if (slot >= btrfs_header_nritems(l)) {
1972                                 ret = btrfs_next_leaf(root, path);
1973                                 if (ret == 0)
1974                                         continue;
1975                                 if (ret < 0)
1976                                         goto out;
1977
1978                                 break;
1979                         }
1980                         btrfs_item_key_to_cpu(l, &key, slot);
1981
1982                         if (key.objectid + key.offset <= logical)
1983                                 goto next;
1984
1985                         if (key.objectid >= logical + map->stripe_len)
1986                                 break;
1987
1988                         if (btrfs_key_type(&key) != BTRFS_EXTENT_ITEM_KEY)
1989                                 goto next;
1990
1991                         extent = btrfs_item_ptr(l, slot,
1992                                                 struct btrfs_extent_item);
1993                         flags = btrfs_extent_flags(l, extent);
1994                         generation = btrfs_extent_generation(l, extent);
1995
1996                         if (key.objectid < logical &&
1997                             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
1998                                 printk(KERN_ERR
1999                                        "btrfs scrub: tree block %llu spanning "
2000                                        "stripes, ignored. logical=%llu\n",
2001                                        (unsigned long long)key.objectid,
2002                                        (unsigned long long)logical);
2003                                 goto next;
2004                         }
2005
2006                         /*
2007                          * trim extent to this stripe
2008                          */
2009                         if (key.objectid < logical) {
2010                                 key.offset -= logical - key.objectid;
2011                                 key.objectid = logical;
2012                         }
2013                         if (key.objectid + key.offset >
2014                             logical + map->stripe_len) {
2015                                 key.offset = logical + map->stripe_len -
2016                                              key.objectid;
2017                         }
2018
2019                         ret = scrub_extent(sdev, key.objectid, key.offset,
2020                                            key.objectid - logical + physical,
2021                                            flags, generation, mirror_num);
2022                         if (ret)
2023                                 goto out;
2024
2025 next:
2026                         path->slots[0]++;
2027                 }
2028                 btrfs_release_path(path);
2029                 logical += increment;
2030                 physical += map->stripe_len;
2031                 spin_lock(&sdev->stat_lock);
2032                 sdev->stat.last_physical = physical;
2033                 spin_unlock(&sdev->stat_lock);
2034         }
2035         /* push queued extents */
2036         scrub_submit(sdev);
2037
2038 out:
2039         blk_finish_plug(&plug);
2040         btrfs_free_path(path);
2041         return ret < 0 ? ret : 0;
2042 }
2043
2044 static noinline_for_stack int scrub_chunk(struct scrub_dev *sdev,
2045         u64 chunk_tree, u64 chunk_objectid, u64 chunk_offset, u64 length,
2046         u64 dev_offset)
2047 {
2048         struct btrfs_mapping_tree *map_tree =
2049                 &sdev->dev->dev_root->fs_info->mapping_tree;
2050         struct map_lookup *map;
2051         struct extent_map *em;
2052         int i;
2053         int ret = -EINVAL;
2054
2055         read_lock(&map_tree->map_tree.lock);
2056         em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2057         read_unlock(&map_tree->map_tree.lock);
2058
2059         if (!em)
2060                 return -EINVAL;
2061
2062         map = (struct map_lookup *)em->bdev;
2063         if (em->start != chunk_offset)
2064                 goto out;
2065
2066         if (em->len < length)
2067                 goto out;
2068
2069         for (i = 0; i < map->num_stripes; ++i) {
2070                 if (map->stripes[i].dev == sdev->dev &&
2071                     map->stripes[i].physical == dev_offset) {
2072                         ret = scrub_stripe(sdev, map, i, chunk_offset, length);
2073                         if (ret)
2074                                 goto out;
2075                 }
2076         }
2077 out:
2078         free_extent_map(em);
2079
2080         return ret;
2081 }
2082
2083 static noinline_for_stack
2084 int scrub_enumerate_chunks(struct scrub_dev *sdev, u64 start, u64 end)
2085 {
2086         struct btrfs_dev_extent *dev_extent = NULL;
2087         struct btrfs_path *path;
2088         struct btrfs_root *root = sdev->dev->dev_root;
2089         struct btrfs_fs_info *fs_info = root->fs_info;
2090         u64 length;
2091         u64 chunk_tree;
2092         u64 chunk_objectid;
2093         u64 chunk_offset;
2094         int ret;
2095         int slot;
2096         struct extent_buffer *l;
2097         struct btrfs_key key;
2098         struct btrfs_key found_key;
2099         struct btrfs_block_group_cache *cache;
2100
2101         path = btrfs_alloc_path();
2102         if (!path)
2103                 return -ENOMEM;
2104
2105         path->reada = 2;
2106         path->search_commit_root = 1;
2107         path->skip_locking = 1;
2108
2109         key.objectid = sdev->dev->devid;
2110         key.offset = 0ull;
2111         key.type = BTRFS_DEV_EXTENT_KEY;
2112
2113
2114         while (1) {
2115                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2116                 if (ret < 0)
2117                         break;
2118                 if (ret > 0) {
2119                         if (path->slots[0] >=
2120                             btrfs_header_nritems(path->nodes[0])) {
2121                                 ret = btrfs_next_leaf(root, path);
2122                                 if (ret)
2123                                         break;
2124                         }
2125                 }
2126
2127                 l = path->nodes[0];
2128                 slot = path->slots[0];
2129
2130                 btrfs_item_key_to_cpu(l, &found_key, slot);
2131
2132                 if (found_key.objectid != sdev->dev->devid)
2133                         break;
2134
2135                 if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
2136                         break;
2137
2138                 if (found_key.offset >= end)
2139                         break;
2140
2141                 if (found_key.offset < key.offset)
2142                         break;
2143
2144                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2145                 length = btrfs_dev_extent_length(l, dev_extent);
2146
2147                 if (found_key.offset + length <= start) {
2148                         key.offset = found_key.offset + length;
2149                         btrfs_release_path(path);
2150                         continue;
2151                 }
2152
2153                 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2154                 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2155                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2156
2157                 /*
2158                  * get a reference on the corresponding block group to prevent
2159                  * the chunk from going away while we scrub it
2160                  */
2161                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2162                 if (!cache) {
2163                         ret = -ENOENT;
2164                         break;
2165                 }
2166                 ret = scrub_chunk(sdev, chunk_tree, chunk_objectid,
2167                                   chunk_offset, length, found_key.offset);
2168                 btrfs_put_block_group(cache);
2169                 if (ret)
2170                         break;
2171
2172                 key.offset = found_key.offset + length;
2173                 btrfs_release_path(path);
2174         }
2175
2176         btrfs_free_path(path);
2177
2178         /*
2179          * ret can still be 1 from search_slot or next_leaf,
2180          * that's not an error
2181          */
2182         return ret < 0 ? ret : 0;
2183 }
2184
2185 static noinline_for_stack int scrub_supers(struct scrub_dev *sdev)
2186 {
2187         int     i;
2188         u64     bytenr;
2189         u64     gen;
2190         int     ret;
2191         struct btrfs_device *device = sdev->dev;
2192         struct btrfs_root *root = device->dev_root;
2193
2194         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
2195                 return -EIO;
2196
2197         gen = root->fs_info->last_trans_committed;
2198
2199         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2200                 bytenr = btrfs_sb_offset(i);
2201                 if (bytenr + BTRFS_SUPER_INFO_SIZE > device->total_bytes)
2202                         break;
2203
2204                 ret = scrub_pages(sdev, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2205                                      BTRFS_EXTENT_FLAG_SUPER, gen, i, NULL, 1);
2206                 if (ret)
2207                         return ret;
2208         }
2209         wait_event(sdev->list_wait, atomic_read(&sdev->in_flight) == 0);
2210
2211         return 0;
2212 }
2213
2214 /*
2215  * get a reference count on fs_info->scrub_workers. start worker if necessary
2216  */
2217 static noinline_for_stack int scrub_workers_get(struct btrfs_root *root)
2218 {
2219         struct btrfs_fs_info *fs_info = root->fs_info;
2220         int ret = 0;
2221
2222         mutex_lock(&fs_info->scrub_lock);
2223         if (fs_info->scrub_workers_refcnt == 0) {
2224                 btrfs_init_workers(&fs_info->scrub_workers, "scrub",
2225                            fs_info->thread_pool_size, &fs_info->generic_worker);
2226                 fs_info->scrub_workers.idle_thresh = 4;
2227                 ret = btrfs_start_workers(&fs_info->scrub_workers);
2228                 if (ret)
2229                         goto out;
2230         }
2231         ++fs_info->scrub_workers_refcnt;
2232 out:
2233         mutex_unlock(&fs_info->scrub_lock);
2234
2235         return ret;
2236 }
2237
2238 static noinline_for_stack void scrub_workers_put(struct btrfs_root *root)
2239 {
2240         struct btrfs_fs_info *fs_info = root->fs_info;
2241
2242         mutex_lock(&fs_info->scrub_lock);
2243         if (--fs_info->scrub_workers_refcnt == 0)
2244                 btrfs_stop_workers(&fs_info->scrub_workers);
2245         WARN_ON(fs_info->scrub_workers_refcnt < 0);
2246         mutex_unlock(&fs_info->scrub_lock);
2247 }
2248
2249
2250 int btrfs_scrub_dev(struct btrfs_root *root, u64 devid, u64 start, u64 end,
2251                     struct btrfs_scrub_progress *progress, int readonly)
2252 {
2253         struct scrub_dev *sdev;
2254         struct btrfs_fs_info *fs_info = root->fs_info;
2255         int ret;
2256         struct btrfs_device *dev;
2257
2258         if (btrfs_fs_closing(root->fs_info))
2259                 return -EINVAL;
2260
2261         /*
2262          * check some assumptions
2263          */
2264         if (root->nodesize != root->leafsize) {
2265                 printk(KERN_ERR
2266                        "btrfs_scrub: size assumption nodesize == leafsize (%d == %d) fails\n",
2267                        root->nodesize, root->leafsize);
2268                 return -EINVAL;
2269         }
2270
2271         if (root->nodesize > BTRFS_STRIPE_LEN) {
2272                 /*
2273                  * in this case scrub is unable to calculate the checksum
2274                  * the way scrub is implemented. Do not handle this
2275                  * situation at all because it won't ever happen.
2276                  */
2277                 printk(KERN_ERR
2278                        "btrfs_scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails\n",
2279                        root->nodesize, BTRFS_STRIPE_LEN);
2280                 return -EINVAL;
2281         }
2282
2283         if (root->sectorsize != PAGE_SIZE) {
2284                 /* not supported for data w/o checksums */
2285                 printk(KERN_ERR
2286                        "btrfs_scrub: size assumption sectorsize != PAGE_SIZE (%d != %lld) fails\n",
2287                        root->sectorsize, (unsigned long long)PAGE_SIZE);
2288                 return -EINVAL;
2289         }
2290
2291         ret = scrub_workers_get(root);
2292         if (ret)
2293                 return ret;
2294
2295         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
2296         dev = btrfs_find_device(root, devid, NULL, NULL);
2297         if (!dev || dev->missing) {
2298                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2299                 scrub_workers_put(root);
2300                 return -ENODEV;
2301         }
2302         mutex_lock(&fs_info->scrub_lock);
2303
2304         if (!dev->in_fs_metadata) {
2305                 mutex_unlock(&fs_info->scrub_lock);
2306                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2307                 scrub_workers_put(root);
2308                 return -ENODEV;
2309         }
2310
2311         if (dev->scrub_device) {
2312                 mutex_unlock(&fs_info->scrub_lock);
2313                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2314                 scrub_workers_put(root);
2315                 return -EINPROGRESS;
2316         }
2317         sdev = scrub_setup_dev(dev);
2318         if (IS_ERR(sdev)) {
2319                 mutex_unlock(&fs_info->scrub_lock);
2320                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2321                 scrub_workers_put(root);
2322                 return PTR_ERR(sdev);
2323         }
2324         sdev->readonly = readonly;
2325         dev->scrub_device = sdev;
2326
2327         atomic_inc(&fs_info->scrubs_running);
2328         mutex_unlock(&fs_info->scrub_lock);
2329         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2330
2331         down_read(&fs_info->scrub_super_lock);
2332         ret = scrub_supers(sdev);
2333         up_read(&fs_info->scrub_super_lock);
2334
2335         if (!ret)
2336                 ret = scrub_enumerate_chunks(sdev, start, end);
2337
2338         wait_event(sdev->list_wait, atomic_read(&sdev->in_flight) == 0);
2339         atomic_dec(&fs_info->scrubs_running);
2340         wake_up(&fs_info->scrub_pause_wait);
2341
2342         wait_event(sdev->list_wait, atomic_read(&sdev->fixup_cnt) == 0);
2343
2344         if (progress)
2345                 memcpy(progress, &sdev->stat, sizeof(*progress));
2346
2347         mutex_lock(&fs_info->scrub_lock);
2348         dev->scrub_device = NULL;
2349         mutex_unlock(&fs_info->scrub_lock);
2350
2351         scrub_free_dev(sdev);
2352         scrub_workers_put(root);
2353
2354         return ret;
2355 }
2356
2357 void btrfs_scrub_pause(struct btrfs_root *root)
2358 {
2359         struct btrfs_fs_info *fs_info = root->fs_info;
2360
2361         mutex_lock(&fs_info->scrub_lock);
2362         atomic_inc(&fs_info->scrub_pause_req);
2363         while (atomic_read(&fs_info->scrubs_paused) !=
2364                atomic_read(&fs_info->scrubs_running)) {
2365                 mutex_unlock(&fs_info->scrub_lock);
2366                 wait_event(fs_info->scrub_pause_wait,
2367                            atomic_read(&fs_info->scrubs_paused) ==
2368                            atomic_read(&fs_info->scrubs_running));
2369                 mutex_lock(&fs_info->scrub_lock);
2370         }
2371         mutex_unlock(&fs_info->scrub_lock);
2372 }
2373
2374 void btrfs_scrub_continue(struct btrfs_root *root)
2375 {
2376         struct btrfs_fs_info *fs_info = root->fs_info;
2377
2378         atomic_dec(&fs_info->scrub_pause_req);
2379         wake_up(&fs_info->scrub_pause_wait);
2380 }
2381
2382 void btrfs_scrub_pause_super(struct btrfs_root *root)
2383 {
2384         down_write(&root->fs_info->scrub_super_lock);
2385 }
2386
2387 void btrfs_scrub_continue_super(struct btrfs_root *root)
2388 {
2389         up_write(&root->fs_info->scrub_super_lock);
2390 }
2391
2392 int __btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
2393 {
2394
2395         mutex_lock(&fs_info->scrub_lock);
2396         if (!atomic_read(&fs_info->scrubs_running)) {
2397                 mutex_unlock(&fs_info->scrub_lock);
2398                 return -ENOTCONN;
2399         }
2400
2401         atomic_inc(&fs_info->scrub_cancel_req);
2402         while (atomic_read(&fs_info->scrubs_running)) {
2403                 mutex_unlock(&fs_info->scrub_lock);
2404                 wait_event(fs_info->scrub_pause_wait,
2405                            atomic_read(&fs_info->scrubs_running) == 0);
2406                 mutex_lock(&fs_info->scrub_lock);
2407         }
2408         atomic_dec(&fs_info->scrub_cancel_req);
2409         mutex_unlock(&fs_info->scrub_lock);
2410
2411         return 0;
2412 }
2413
2414 int btrfs_scrub_cancel(struct btrfs_root *root)
2415 {
2416         return __btrfs_scrub_cancel(root->fs_info);
2417 }
2418
2419 int btrfs_scrub_cancel_dev(struct btrfs_root *root, struct btrfs_device *dev)
2420 {
2421         struct btrfs_fs_info *fs_info = root->fs_info;
2422         struct scrub_dev *sdev;
2423
2424         mutex_lock(&fs_info->scrub_lock);
2425         sdev = dev->scrub_device;
2426         if (!sdev) {
2427                 mutex_unlock(&fs_info->scrub_lock);
2428                 return -ENOTCONN;
2429         }
2430         atomic_inc(&sdev->cancel_req);
2431         while (dev->scrub_device) {
2432                 mutex_unlock(&fs_info->scrub_lock);
2433                 wait_event(fs_info->scrub_pause_wait,
2434                            dev->scrub_device == NULL);
2435                 mutex_lock(&fs_info->scrub_lock);
2436         }
2437         mutex_unlock(&fs_info->scrub_lock);
2438
2439         return 0;
2440 }
2441
2442 int btrfs_scrub_cancel_devid(struct btrfs_root *root, u64 devid)
2443 {
2444         struct btrfs_fs_info *fs_info = root->fs_info;
2445         struct btrfs_device *dev;
2446         int ret;
2447
2448         /*
2449          * we have to hold the device_list_mutex here so the device
2450          * does not go away in cancel_dev. FIXME: find a better solution
2451          */
2452         mutex_lock(&fs_info->fs_devices->device_list_mutex);
2453         dev = btrfs_find_device(root, devid, NULL, NULL);
2454         if (!dev) {
2455                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2456                 return -ENODEV;
2457         }
2458         ret = btrfs_scrub_cancel_dev(root, dev);
2459         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2460
2461         return ret;
2462 }
2463
2464 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
2465                          struct btrfs_scrub_progress *progress)
2466 {
2467         struct btrfs_device *dev;
2468         struct scrub_dev *sdev = NULL;
2469
2470         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
2471         dev = btrfs_find_device(root, devid, NULL, NULL);
2472         if (dev)
2473                 sdev = dev->scrub_device;
2474         if (sdev)
2475                 memcpy(progress, &sdev->stat, sizeof(*progress));
2476         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2477
2478         return dev ? (sdev ? 0 : -ENOTCONN) : -ENODEV;
2479 }