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