fs/btrfs/free-space-tree.c

   1 /*
   2  * Copyright (C) 2015 Facebook.  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/kernel.h>
  20 #include <linux/vmalloc.h>
  21 #include "ctree.h"
  22 #include "disk-io.h"
  23 #include "locking.h"
  24 #include "free-space-tree.h"
  25 #include "transaction.h"
  26 #include "sysfs.h"
  27
  28 static int __add_block_group_free_space(struct btrfs_trans_handle *trans,
  29                                         struct btrfs_fs_info *fs_info,
  30                                         struct btrfs_block_group_cache *block_group,
  31                                         struct btrfs_path *path);
  32
  33 void set_free_space_tree_thresholds(struct btrfs_block_group_cache *cache)
  34 {
  35         u32 bitmap_range;
  36         size_t bitmap_size;
  37         u64 num_bitmaps, total_bitmap_size;
  38
  39         /*
  40          * We convert to bitmaps when the disk space required for using extents
  41          * exceeds that required for using bitmaps.
  42          */
  43         bitmap_range = cache->sectorsize * BTRFS_FREE_SPACE_BITMAP_BITS;
  44         num_bitmaps = div_u64(cache->key.offset + bitmap_range - 1,
  45                               bitmap_range);
  46         bitmap_size = sizeof(struct btrfs_item) + BTRFS_FREE_SPACE_BITMAP_SIZE;
  47         total_bitmap_size = num_bitmaps * bitmap_size;
  48         cache->bitmap_high_thresh = div_u64(total_bitmap_size,
  49                                             sizeof(struct btrfs_item));
  50
  51         /*
  52          * We allow for a small buffer between the high threshold and low
  53          * threshold to avoid thrashing back and forth between the two formats.
  54          */
  55         if (cache->bitmap_high_thresh > 100)
  56                 cache->bitmap_low_thresh = cache->bitmap_high_thresh - 100;
  57         else
  58                 cache->bitmap_low_thresh = 0;
  59 }
  60
  61 static int add_new_free_space_info(struct btrfs_trans_handle *trans,
  62                                    struct btrfs_fs_info *fs_info,
  63                                    struct btrfs_block_group_cache *block_group,
  64                                    struct btrfs_path *path)
  65 {
  66         struct btrfs_root *root = fs_info->free_space_root;
  67         struct btrfs_free_space_info *info;
  68         struct btrfs_key key;
  69         struct extent_buffer *leaf;
  70         int ret;
  71
  72         key.objectid = block_group->key.objectid;
  73         key.type = BTRFS_FREE_SPACE_INFO_KEY;
  74         key.offset = block_group->key.offset;
  75
  76         ret = btrfs_insert_empty_item(trans, root, path, &key, sizeof(*info));
  77         if (ret)
  78                 goto out;
  79
  80         leaf = path->nodes[0];
  81         info = btrfs_item_ptr(leaf, path->slots[0],
  82                               struct btrfs_free_space_info);
  83         btrfs_set_free_space_extent_count(leaf, info, 0);
  84         btrfs_set_free_space_flags(leaf, info, 0);
  85         btrfs_mark_buffer_dirty(leaf);
  86
  87         ret = 0;
  88 out:
  89         btrfs_release_path(path);
  90         return ret;
  91 }
  92
  93 struct btrfs_free_space_info *
  94 search_free_space_info(struct btrfs_trans_handle *trans,
  95                        struct btrfs_fs_info *fs_info,
  96                        struct btrfs_block_group_cache *block_group,
  97                        struct btrfs_path *path, int cow)
  98 {
  99         struct btrfs_root *root = fs_info->free_space_root;
 100         struct btrfs_key key;
 101         int ret;
 102
 103         key.objectid = block_group->key.objectid;
 104         key.type = BTRFS_FREE_SPACE_INFO_KEY;
 105         key.offset = block_group->key.offset;
 106
 107         ret = btrfs_search_slot(trans, root, &key, path, 0, cow);
 108         if (ret < 0)
 109                 return ERR_PTR(ret);
 110         if (ret != 0) {
 111                 btrfs_warn(fs_info, "missing free space info for %llu\n",
 112                            block_group->key.objectid);
 113                 ASSERT(0);
 114                 return ERR_PTR(-ENOENT);
 115         }
 116
 117         return btrfs_item_ptr(path->nodes[0], path->slots[0],
 118                               struct btrfs_free_space_info);
 119 }
 120
 121 /*
 122  * btrfs_search_slot() but we're looking for the greatest key less than the
 123  * passed key.
 124  */
 125 static int btrfs_search_prev_slot(struct btrfs_trans_handle *trans,
 126                                   struct btrfs_root *root,
 127                                   struct btrfs_key *key, struct btrfs_path *p,
 128                                   int ins_len, int cow)
 129 {
 130         int ret;
 131
 132         ret = btrfs_search_slot(trans, root, key, p, ins_len, cow);
 133         if (ret < 0)
 134                 return ret;
 135
 136         if (ret == 0) {
 137                 ASSERT(0);
 138                 return -EIO;
 139         }
 140
 141         if (p->slots[0] == 0) {
 142                 ASSERT(0);
 143                 return -EIO;
 144         }
 145         p->slots[0]--;
 146
 147         return 0;
 148 }
 149
 150 static inline u32 free_space_bitmap_size(u64 size, u32 sectorsize)
 151 {
 152         return DIV_ROUND_UP((u32)div_u64(size, sectorsize), BITS_PER_BYTE);
 153 }
 154
 155 static unsigned long *alloc_bitmap(u32 bitmap_size)
 156 {
 157         return __vmalloc(bitmap_size, GFP_NOFS | __GFP_HIGHMEM | __GFP_ZERO,
 158                          PAGE_KERNEL);
 159 }
 160
 161 int convert_free_space_to_bitmaps(struct btrfs_trans_handle *trans,
 162                                   struct btrfs_fs_info *fs_info,
 163                                   struct btrfs_block_group_cache *block_group,
 164                                   struct btrfs_path *path)
 165 {
 166         struct btrfs_root *root = fs_info->free_space_root;
 167         struct btrfs_free_space_info *info;
 168         struct btrfs_key key, found_key;
 169         struct extent_buffer *leaf;
 170         unsigned long *bitmap;
 171         char *bitmap_cursor;
 172         u64 start, end;
 173         u64 bitmap_range, i;
 174         u32 bitmap_size, flags, expected_extent_count;
 175         u32 extent_count = 0;
 176         int done = 0, nr;
 177         int ret;
 178
 179         bitmap_size = free_space_bitmap_size(block_group->key.offset,
 180                                              block_group->sectorsize);
 181         bitmap = alloc_bitmap(bitmap_size);
 182         if (!bitmap) {
 183                 ret = -ENOMEM;
 184                 goto out;
 185         }
 186
 187         start = block_group->key.objectid;
 188         end = block_group->key.objectid + block_group->key.offset;
 189
 190         key.objectid = end - 1;
 191         key.type = (u8)-1;
 192         key.offset = (u64)-1;
 193
 194         while (!done) {
 195                 ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
 196                 if (ret)
 197                         goto out;
 198
 199                 leaf = path->nodes[0];
 200                 nr = 0;
 201                 path->slots[0]++;
 202                 while (path->slots[0] > 0) {
 203                         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0] - 1);
 204
 205                         if (found_key.type == BTRFS_FREE_SPACE_INFO_KEY) {
 206                                 ASSERT(found_key.objectid == block_group->key.objectid);
 207                                 ASSERT(found_key.offset == block_group->key.offset);
 208                                 done = 1;
 209                                 break;
 210                         } else if (found_key.type == BTRFS_FREE_SPACE_EXTENT_KEY) {
 211                                 u64 first, last;
 212
 213                                 ASSERT(found_key.objectid >= start);
 214                                 ASSERT(found_key.objectid < end);
 215                                 ASSERT(found_key.objectid + found_key.offset <= end);
 216
 217                                 first = div_u64(found_key.objectid - start,
 218                                                 block_group->sectorsize);
 219                                 last = div_u64(found_key.objectid + found_key.offset - start,
 220                                                block_group->sectorsize);
 221                                 bitmap_set(bitmap, first, last - first);
 222
 223                                 extent_count++;
 224                                 nr++;
 225                                 path->slots[0]--;
 226                         } else {
 227                                 ASSERT(0);
 228                         }
 229                 }
 230
 231                 ret = btrfs_del_items(trans, root, path, path->slots[0], nr);
 232                 if (ret)
 233                         goto out;
 234                 btrfs_release_path(path);
 235         }
 236
 237         info = search_free_space_info(trans, fs_info, block_group, path, 1);
 238         if (IS_ERR(info)) {
 239                 ret = PTR_ERR(info);
 240                 goto out;
 241         }
 242         leaf = path->nodes[0];
 243         flags = btrfs_free_space_flags(leaf, info);
 244         flags |= BTRFS_FREE_SPACE_USING_BITMAPS;
 245         btrfs_set_free_space_flags(leaf, info, flags);
 246         expected_extent_count = btrfs_free_space_extent_count(leaf, info);
 247         btrfs_mark_buffer_dirty(leaf);
 248         btrfs_release_path(path);
 249
 250         if (extent_count != expected_extent_count) {
 251                 btrfs_err(fs_info, "incorrect extent count for %llu; counted %u, expected %u",
 252                           block_group->key.objectid, extent_count,
 253                           expected_extent_count);
 254                 ASSERT(0);
 255                 ret = -EIO;
 256                 goto out;
 257         }
 258
 259         bitmap_cursor = (char *)bitmap;
 260         bitmap_range = block_group->sectorsize * BTRFS_FREE_SPACE_BITMAP_BITS;
 261         i = start;
 262         while (i < end) {
 263                 unsigned long ptr;
 264                 u64 extent_size;
 265                 u32 data_size;
 266
 267                 extent_size = min(end - i, bitmap_range);
 268                 data_size = free_space_bitmap_size(extent_size,
 269                                                    block_group->sectorsize);
 270
 271                 key.objectid = i;
 272                 key.type = BTRFS_FREE_SPACE_BITMAP_KEY;
 273                 key.offset = extent_size;
 274
 275                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 276                                               data_size);
 277                 if (ret)
 278                         goto out;
 279
 280                 leaf = path->nodes[0];
 281                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
 282                 write_extent_buffer(leaf, bitmap_cursor, ptr,
 283                                     data_size);
 284                 btrfs_mark_buffer_dirty(leaf);
 285                 btrfs_release_path(path);
 286
 287                 i += extent_size;
 288                 bitmap_cursor += data_size;
 289         }
 290
 291         ret = 0;
 292 out:
 293         vfree(bitmap);
 294         if (ret)
 295                 btrfs_abort_transaction(trans, root, ret);
 296         return ret;
 297 }
 298
 299 int convert_free_space_to_extents(struct btrfs_trans_handle *trans,
 300                                   struct btrfs_fs_info *fs_info,
 301                                   struct btrfs_block_group_cache *block_group,
 302                                   struct btrfs_path *path)
 303 {
 304         struct btrfs_root *root = fs_info->free_space_root;
 305         struct btrfs_free_space_info *info;
 306         struct btrfs_key key, found_key;
 307         struct extent_buffer *leaf;
 308         unsigned long *bitmap;
 309         u64 start, end;
 310         /* Initialize to silence GCC. */
 311         u64 extent_start = 0;
 312         u64 offset;
 313         u32 bitmap_size, flags, expected_extent_count;
 314         int prev_bit = 0, bit, bitnr;
 315         u32 extent_count = 0;
 316         int done = 0, nr;
 317         int ret;
 318
 319         bitmap_size = free_space_bitmap_size(block_group->key.offset,
 320                                              block_group->sectorsize);
 321         bitmap = alloc_bitmap(bitmap_size);
 322         if (!bitmap) {
 323                 ret = -ENOMEM;
 324                 goto out;
 325         }
 326
 327         start = block_group->key.objectid;
 328         end = block_group->key.objectid + block_group->key.offset;
 329
 330         key.objectid = end - 1;
 331         key.type = (u8)-1;
 332         key.offset = (u64)-1;
 333
 334         while (!done) {
 335                 ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
 336                 if (ret)
 337                         goto out;
 338
 339                 leaf = path->nodes[0];
 340                 nr = 0;
 341                 path->slots[0]++;
 342                 while (path->slots[0] > 0) {
 343                         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0] - 1);
 344
 345                         if (found_key.type == BTRFS_FREE_SPACE_INFO_KEY) {
 346                                 ASSERT(found_key.objectid == block_group->key.objectid);
 347                                 ASSERT(found_key.offset == block_group->key.offset);
 348                                 done = 1;
 349                                 break;
 350                         } else if (found_key.type == BTRFS_FREE_SPACE_BITMAP_KEY) {
 351                                 unsigned long ptr;
 352                                 char *bitmap_cursor;
 353                                 u32 bitmap_pos, data_size;
 354
 355                                 ASSERT(found_key.objectid >= start);
 356                                 ASSERT(found_key.objectid < end);
 357                                 ASSERT(found_key.objectid + found_key.offset <= end);
 358
 359                                 bitmap_pos = div_u64(found_key.objectid - start,
 360                                                      block_group->sectorsize *
 361                                                      BITS_PER_BYTE);
 362                                 bitmap_cursor = ((char *)bitmap) + bitmap_pos;
 363                                 data_size = free_space_bitmap_size(found_key.offset,
 364                                                                    block_group->sectorsize);
 365
 366                                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0] - 1);
 367                                 read_extent_buffer(leaf, bitmap_cursor, ptr,
 368                                                    data_size);
 369
 370                                 nr++;
 371                                 path->slots[0]--;
 372                         } else {
 373                                 ASSERT(0);
 374                         }
 375                 }
 376
 377                 ret = btrfs_del_items(trans, root, path, path->slots[0], nr);
 378                 if (ret)
 379                         goto out;
 380                 btrfs_release_path(path);
 381         }
 382
 383         info = search_free_space_info(trans, fs_info, block_group, path, 1);
 384         if (IS_ERR(info)) {
 385                 ret = PTR_ERR(info);
 386                 goto out;
 387         }
 388         leaf = path->nodes[0];
 389         flags = btrfs_free_space_flags(leaf, info);
 390         flags &= ~BTRFS_FREE_SPACE_USING_BITMAPS;
 391         btrfs_set_free_space_flags(leaf, info, flags);
 392         expected_extent_count = btrfs_free_space_extent_count(leaf, info);
 393         btrfs_mark_buffer_dirty(leaf);
 394         btrfs_release_path(path);
 395
 396         offset = start;
 397         bitnr = 0;
 398         while (offset < end) {
 399                 bit = !!test_bit(bitnr, bitmap);
 400                 if (prev_bit == 0 && bit == 1) {
 401                         extent_start = offset;
 402                 } else if (prev_bit == 1 && bit == 0) {
 403                         key.objectid = extent_start;
 404                         key.type = BTRFS_FREE_SPACE_EXTENT_KEY;
 405                         key.offset = offset - extent_start;
 406
 407                         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
 408                         if (ret)
 409                                 goto out;
 410                         btrfs_release_path(path);
 411
 412                         extent_count++;
 413                 }
 414                 prev_bit = bit;
 415                 offset += block_group->sectorsize;
 416                 bitnr++;
 417         }
 418         if (prev_bit == 1) {
 419                 key.objectid = extent_start;
 420                 key.type = BTRFS_FREE_SPACE_EXTENT_KEY;
 421                 key.offset = end - extent_start;
 422
 423                 ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
 424                 if (ret)
 425                         goto out;
 426                 btrfs_release_path(path);
 427
 428                 extent_count++;
 429         }
 430
 431         if (extent_count != expected_extent_count) {
 432                 btrfs_err(fs_info, "incorrect extent count for %llu; counted %u, expected %u",
 433                           block_group->key.objectid, extent_count,
 434                           expected_extent_count);
 435                 ASSERT(0);
 436                 ret = -EIO;
 437                 goto out;
 438         }
 439
 440         ret = 0;
 441 out:
 442         vfree(bitmap);
 443         if (ret)
 444                 btrfs_abort_transaction(trans, root, ret);
 445         return ret;
 446 }
 447
 448 static int update_free_space_extent_count(struct btrfs_trans_handle *trans,
 449                                           struct btrfs_fs_info *fs_info,
 450                                           struct btrfs_block_group_cache *block_group,
 451                                           struct btrfs_path *path,
 452                                           int new_extents)
 453 {
 454         struct btrfs_free_space_info *info;
 455         u32 flags;
 456         u32 extent_count;
 457         int ret = 0;
 458
 459         if (new_extents == 0)
 460                 return 0;
 461
 462         info = search_free_space_info(trans, fs_info, block_group, path, 1);
 463         if (IS_ERR(info)) {
 464                 ret = PTR_ERR(info);
 465                 goto out;
 466         }
 467         flags = btrfs_free_space_flags(path->nodes[0], info);
 468         extent_count = btrfs_free_space_extent_count(path->nodes[0], info);
 469
 470         extent_count += new_extents;
 471         btrfs_set_free_space_extent_count(path->nodes[0], info, extent_count);
 472         btrfs_mark_buffer_dirty(path->nodes[0]);
 473         btrfs_release_path(path);
 474
 475         if (!(flags & BTRFS_FREE_SPACE_USING_BITMAPS) &&
 476             extent_count > block_group->bitmap_high_thresh) {
 477                 ret = convert_free_space_to_bitmaps(trans, fs_info, block_group,
 478                                                     path);
 479         } else if ((flags & BTRFS_FREE_SPACE_USING_BITMAPS) &&
 480                    extent_count < block_group->bitmap_low_thresh) {
 481                 ret = convert_free_space_to_extents(trans, fs_info, block_group,
 482                                                     path);
 483         }
 484
 485 out:
 486         return ret;
 487 }
 488
 489 int free_space_test_bit(struct btrfs_block_group_cache *block_group,
 490                         struct btrfs_path *path, u64 offset)
 491 {
 492         struct extent_buffer *leaf;
 493         struct btrfs_key key;
 494         u64 found_start, found_end;
 495         unsigned long ptr, i;
 496
 497         leaf = path->nodes[0];
 498         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 499         ASSERT(key.type == BTRFS_FREE_SPACE_BITMAP_KEY);
 500
 501         found_start = key.objectid;
 502         found_end = key.objectid + key.offset;
 503         ASSERT(offset >= found_start && offset < found_end);
 504
 505         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
 506         i = div_u64(offset - found_start, block_group->sectorsize);
 507         return !!extent_buffer_test_bit(leaf, ptr, i);
 508 }
 509
 510 static void free_space_set_bits(struct btrfs_block_group_cache *block_group,
 511                                 struct btrfs_path *path, u64 *start, u64 *size,
 512                                 int bit)
 513 {
 514         struct extent_buffer *leaf;
 515         struct btrfs_key key;
 516         u64 end = *start + *size;
 517         u64 found_start, found_end;
 518         unsigned long ptr, first, last;
 519
 520         leaf = path->nodes[0];
 521         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 522         ASSERT(key.type == BTRFS_FREE_SPACE_BITMAP_KEY);
 523
 524         found_start = key.objectid;
 525         found_end = key.objectid + key.offset;
 526         ASSERT(*start >= found_start && *start < found_end);
 527         ASSERT(end > found_start);
 528
 529         if (end > found_end)
 530                 end = found_end;
 531
 532         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
 533         first = div_u64(*start - found_start, block_group->sectorsize);
 534         last = div_u64(end - found_start, block_group->sectorsize);
 535         if (bit)
 536                 extent_buffer_bitmap_set(leaf, ptr, first, last - first);
 537         else
 538                 extent_buffer_bitmap_clear(leaf, ptr, first, last - first);
 539         btrfs_mark_buffer_dirty(leaf);
 540
 541         *size -= end - *start;
 542         *start = end;
 543 }
 544
 545 /*
 546  * We can't use btrfs_next_item() in modify_free_space_bitmap() because
 547  * btrfs_next_leaf() doesn't get the path for writing. We can forgo the fancy
 548  * tree walking in btrfs_next_leaf() anyways because we know exactly what we're
 549  * looking for.
 550  */
 551 static int free_space_next_bitmap(struct btrfs_trans_handle *trans,
 552                                   struct btrfs_root *root, struct btrfs_path *p)
 553 {
 554         struct btrfs_key key;
 555
 556         if (p->slots[0] + 1 < btrfs_header_nritems(p->nodes[0])) {
 557                 p->slots[0]++;
 558                 return 0;
 559         }
 560
 561         btrfs_item_key_to_cpu(p->nodes[0], &key, p->slots[0]);
 562         btrfs_release_path(p);
 563
 564         key.objectid += key.offset;
 565         key.type = (u8)-1;
 566         key.offset = (u64)-1;
 567
 568         return btrfs_search_prev_slot(trans, root, &key, p, 0, 1);
 569 }
 570
 571 /*
 572  * If remove is 1, then we are removing free space, thus clearing bits in the
 573  * bitmap. If remove is 0, then we are adding free space, thus setting bits in
 574  * the bitmap.
 575  */
 576 static int modify_free_space_bitmap(struct btrfs_trans_handle *trans,
 577                                     struct btrfs_fs_info *fs_info,
 578                                     struct btrfs_block_group_cache *block_group,
 579                                     struct btrfs_path *path,
 580                                     u64 start, u64 size, int remove)
 581 {
 582         struct btrfs_root *root = fs_info->free_space_root;
 583         struct btrfs_key key;
 584         u64 end = start + size;
 585         u64 cur_start, cur_size;
 586         int prev_bit, next_bit;
 587         int new_extents;
 588         int ret;
 589
 590         /*
 591          * Read the bit for the block immediately before the extent of space if
 592          * that block is within the block group.
 593          */
 594         if (start > block_group->key.objectid) {
 595                 u64 prev_block = start - block_group->sectorsize;
 596
 597                 key.objectid = prev_block;
 598                 key.type = (u8)-1;
 599                 key.offset = (u64)-1;
 600
 601                 ret = btrfs_search_prev_slot(trans, root, &key, path, 0, 1);
 602                 if (ret)
 603                         goto out;
 604
 605                 prev_bit = free_space_test_bit(block_group, path, prev_block);
 606
 607                 /* The previous block may have been in the previous bitmap. */
 608                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 609                 if (start >= key.objectid + key.offset) {
 610                         ret = free_space_next_bitmap(trans, root, path);
 611                         if (ret)
 612                                 goto out;
 613                 }
 614         } else {
 615                 key.objectid = start;
 616                 key.type = (u8)-1;
 617                 key.offset = (u64)-1;
 618
 619                 ret = btrfs_search_prev_slot(trans, root, &key, path, 0, 1);
 620                 if (ret)
 621                         goto out;
 622
 623                 prev_bit = -1;
 624         }
 625
 626         /*
 627          * Iterate over all of the bitmaps overlapped by the extent of space,
 628          * clearing/setting bits as required.
 629          */
 630         cur_start = start;
 631         cur_size = size;
 632         while (1) {
 633                 free_space_set_bits(block_group, path, &cur_start, &cur_size,
 634                                     !remove);
 635                 if (cur_size == 0)
 636                         break;
 637                 ret = free_space_next_bitmap(trans, root, path);
 638                 if (ret)
 639                         goto out;
 640         }
 641
 642         /*
 643          * Read the bit for the block immediately after the extent of space if
 644          * that block is within the block group.
 645          */
 646         if (end < block_group->key.objectid + block_group->key.offset) {
 647                 /* The next block may be in the next bitmap. */
 648                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 649                 if (end >= key.objectid + key.offset) {
 650                         ret = free_space_next_bitmap(trans, root, path);
 651                         if (ret)
 652                                 goto out;
 653                 }
 654
 655                 next_bit = free_space_test_bit(block_group, path, end);
 656         } else {
 657                 next_bit = -1;
 658         }
 659
 660         if (remove) {
 661                 new_extents = -1;
 662                 if (prev_bit == 1) {
 663                         /* Leftover on the left. */
 664                         new_extents++;
 665                 }
 666                 if (next_bit == 1) {
 667                         /* Leftover on the right. */
 668                         new_extents++;
 669                 }
 670         } else {
 671                 new_extents = 1;
 672                 if (prev_bit == 1) {
 673                         /* Merging with neighbor on the left. */
 674                         new_extents--;
 675                 }
 676                 if (next_bit == 1) {
 677                         /* Merging with neighbor on the right. */
 678                         new_extents--;
 679                 }
 680         }
 681
 682         btrfs_release_path(path);
 683         ret = update_free_space_extent_count(trans, fs_info, block_group, path,
 684                                              new_extents);
 685
 686 out:
 687         return ret;
 688 }
 689
 690 static int remove_free_space_extent(struct btrfs_trans_handle *trans,
 691                                     struct btrfs_fs_info *fs_info,
 692                                     struct btrfs_block_group_cache *block_group,
 693                                     struct btrfs_path *path,
 694                                     u64 start, u64 size)
 695 {
 696         struct btrfs_root *root = fs_info->free_space_root;
 697         struct btrfs_key key;
 698         u64 found_start, found_end;
 699         u64 end = start + size;
 700         int new_extents = -1;
 701         int ret;
 702
 703         key.objectid = start;
 704         key.type = (u8)-1;
 705         key.offset = (u64)-1;
 706
 707         ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
 708         if (ret)
 709                 goto out;
 710
 711         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 712
 713         ASSERT(key.type == BTRFS_FREE_SPACE_EXTENT_KEY);
 714
 715         found_start = key.objectid;
 716         found_end = key.objectid + key.offset;
 717         ASSERT(start >= found_start && end <= found_end);
 718
 719         /*
 720          * Okay, now that we've found the free space extent which contains the
 721          * free space that we are removing, there are four cases:
 722          *
 723          * 1. We're using the whole extent: delete the key we found and
 724          * decrement the free space extent count.
 725          * 2. We are using part of the extent starting at the beginning: delete
 726          * the key we found and insert a new key representing the leftover at
 727          * the end. There is no net change in the number of extents.
 728          * 3. We are using part of the extent ending at the end: delete the key
 729          * we found and insert a new key representing the leftover at the
 730          * beginning. There is no net change in the number of extents.
 731          * 4. We are using part of the extent in the middle: delete the key we
 732          * found and insert two new keys representing the leftovers on each
 733          * side. Where we used to have one extent, we now have two, so increment
 734          * the extent count. We may need to convert the block group to bitmaps
 735          * as a result.
 736          */
 737
 738         /* Delete the existing key (cases 1-4). */
 739         ret = btrfs_del_item(trans, root, path);
 740         if (ret)
 741                 goto out;
 742
 743         /* Add a key for leftovers at the beginning (cases 3 and 4). */
 744         if (start > found_start) {
 745                 key.objectid = found_start;
 746                 key.type = BTRFS_FREE_SPACE_EXTENT_KEY;
 747                 key.offset = start - found_start;
 748
 749                 btrfs_release_path(path);
 750                 ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
 751                 if (ret)
 752                         goto out;
 753                 new_extents++;
 754         }
 755
 756         /* Add a key for leftovers at the end (cases 2 and 4). */
 757         if (end < found_end) {
 758                 key.objectid = end;
 759                 key.type = BTRFS_FREE_SPACE_EXTENT_KEY;
 760                 key.offset = found_end - end;
 761
 762                 btrfs_release_path(path);
 763                 ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
 764                 if (ret)
 765                         goto out;
 766                 new_extents++;
 767         }
 768
 769         btrfs_release_path(path);
 770         ret = update_free_space_extent_count(trans, fs_info, block_group, path,
 771                                              new_extents);
 772
 773 out:
 774         return ret;
 775 }
 776
 777 int __remove_from_free_space_tree(struct btrfs_trans_handle *trans,
 778                                   struct btrfs_fs_info *fs_info,
 779                                   struct btrfs_block_group_cache *block_group,
 780                                   struct btrfs_path *path, u64 start, u64 size)
 781 {
 782         struct btrfs_free_space_info *info;
 783         u32 flags;
 784         int ret;
 785
 786         if (block_group->needs_free_space) {
 787                 ret = __add_block_group_free_space(trans, fs_info, block_group,
 788                                                    path);
 789                 if (ret)
 790                         return ret;
 791         }
 792
 793         info = search_free_space_info(NULL, fs_info, block_group, path, 0);
 794         if (IS_ERR(info))
 795                 return PTR_ERR(info);
 796         flags = btrfs_free_space_flags(path->nodes[0], info);
 797         btrfs_release_path(path);
 798
 799         if (flags & BTRFS_FREE_SPACE_USING_BITMAPS) {
 800                 return modify_free_space_bitmap(trans, fs_info, block_group,
 801                                                 path, start, size, 1);
 802         } else {
 803                 return remove_free_space_extent(trans, fs_info, block_group,
 804                                                 path, start, size);
 805         }
 806 }
 807
 808 int remove_from_free_space_tree(struct btrfs_trans_handle *trans,
 809                                 struct btrfs_fs_info *fs_info,
 810                                 u64 start, u64 size)
 811 {
 812         struct btrfs_block_group_cache *block_group;
 813         struct btrfs_path *path;
 814         int ret;
 815
 816         if (!btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 817                 return 0;
 818
 819         path = btrfs_alloc_path();
 820         if (!path) {
 821                 ret = -ENOMEM;
 822                 goto out;
 823         }
 824
 825         block_group = btrfs_lookup_block_group(fs_info, start);
 826         if (!block_group) {
 827                 ASSERT(0);
 828                 ret = -ENOENT;
 829                 goto out;
 830         }
 831
 832         mutex_lock(&block_group->free_space_lock);
 833         ret = __remove_from_free_space_tree(trans, fs_info, block_group, path,
 834                                             start, size);
 835         mutex_unlock(&block_group->free_space_lock);
 836
 837         btrfs_put_block_group(block_group);
 838 out:
 839         btrfs_free_path(path);
 840         if (ret)
 841                 btrfs_abort_transaction(trans, fs_info->free_space_root, ret);
 842         return ret;
 843 }
 844
 845 static int add_free_space_extent(struct btrfs_trans_handle *trans,
 846                                  struct btrfs_fs_info *fs_info,
 847                                  struct btrfs_block_group_cache *block_group,
 848                                  struct btrfs_path *path,
 849                                  u64 start, u64 size)
 850 {
 851         struct btrfs_root *root = fs_info->free_space_root;
 852         struct btrfs_key key, new_key;
 853         u64 found_start, found_end;
 854         u64 end = start + size;
 855         int new_extents = 1;
 856         int ret;
 857
 858         /*
 859          * We are adding a new extent of free space, but we need to merge
 860          * extents. There are four cases here:
 861          *
 862          * 1. The new extent does not have any immediate neighbors to merge
 863          * with: add the new key and increment the free space extent count. We
 864          * may need to convert the block group to bitmaps as a result.
 865          * 2. The new extent has an immediate neighbor before it: remove the
 866          * previous key and insert a new key combining both of them. There is no
 867          * net change in the number of extents.
 868          * 3. The new extent has an immediate neighbor after it: remove the next
 869          * key and insert a new key combining both of them. There is no net
 870          * change in the number of extents.
 871          * 4. The new extent has immediate neighbors on both sides: remove both
 872          * of the keys and insert a new key combining all of them. Where we used
 873          * to have two extents, we now have one, so decrement the extent count.
 874          */
 875
 876         new_key.objectid = start;
 877         new_key.type = BTRFS_FREE_SPACE_EXTENT_KEY;
 878         new_key.offset = size;
 879
 880         /* Search for a neighbor on the left. */
 881         if (start == block_group->key.objectid)
 882                 goto right;
 883         key.objectid = start - 1;
 884         key.type = (u8)-1;
 885         key.offset = (u64)-1;
 886
 887         ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
 888         if (ret)
 889                 goto out;
 890
 891         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 892
 893         if (key.type != BTRFS_FREE_SPACE_EXTENT_KEY) {
 894                 ASSERT(key.type == BTRFS_FREE_SPACE_INFO_KEY);
 895                 btrfs_release_path(path);
 896                 goto right;
 897         }
 898
 899         found_start = key.objectid;
 900         found_end = key.objectid + key.offset;
 901         ASSERT(found_start >= block_group->key.objectid &&
 902                found_end > block_group->key.objectid);
 903         ASSERT(found_start < start && found_end <= start);
 904
 905         /*
 906          * Delete the neighbor on the left and absorb it into the new key (cases
 907          * 2 and 4).
 908          */
 909         if (found_end == start) {
 910                 ret = btrfs_del_item(trans, root, path);
 911                 if (ret)
 912                         goto out;
 913                 new_key.objectid = found_start;
 914                 new_key.offset += key.offset;
 915                 new_extents--;
 916         }
 917         btrfs_release_path(path);
 918
 919 right:
 920         /* Search for a neighbor on the right. */
 921         if (end == block_group->key.objectid + block_group->key.offset)
 922                 goto insert;
 923         key.objectid = end;
 924         key.type = (u8)-1;
 925         key.offset = (u64)-1;
 926
 927         ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
 928         if (ret)
 929                 goto out;
 930
 931         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 932
 933         if (key.type != BTRFS_FREE_SPACE_EXTENT_KEY) {
 934                 ASSERT(key.type == BTRFS_FREE_SPACE_INFO_KEY);
 935                 btrfs_release_path(path);
 936                 goto insert;
 937         }
 938
 939         found_start = key.objectid;
 940         found_end = key.objectid + key.offset;
 941         ASSERT(found_start >= block_group->key.objectid &&
 942                found_end > block_group->key.objectid);
 943         ASSERT((found_start < start && found_end <= start) ||
 944                (found_start >= end && found_end > end));
 945
 946         /*
 947          * Delete the neighbor on the right and absorb it into the new key
 948          * (cases 3 and 4).
 949          */
 950         if (found_start == end) {
 951                 ret = btrfs_del_item(trans, root, path);
 952                 if (ret)
 953                         goto out;
 954                 new_key.offset += key.offset;
 955                 new_extents--;
 956         }
 957         btrfs_release_path(path);
 958
 959 insert:
 960         /* Insert the new key (cases 1-4). */
 961         ret = btrfs_insert_empty_item(trans, root, path, &new_key, 0);
 962         if (ret)
 963                 goto out;
 964
 965         btrfs_release_path(path);
 966         ret = update_free_space_extent_count(trans, fs_info, block_group, path,
 967                                              new_extents);
 968
 969 out:
 970         return ret;
 971 }
 972
 973 int __add_to_free_space_tree(struct btrfs_trans_handle *trans,
 974                              struct btrfs_fs_info *fs_info,
 975                              struct btrfs_block_group_cache *block_group,
 976                              struct btrfs_path *path, u64 start, u64 size)
 977 {
 978         struct btrfs_free_space_info *info;
 979         u32 flags;
 980         int ret;
 981
 982         if (block_group->needs_free_space) {
 983                 ret = __add_block_group_free_space(trans, fs_info, block_group,
 984                                                    path);
 985                 if (ret)
 986                         return ret;
 987         }
 988
 989         info = search_free_space_info(NULL, fs_info, block_group, path, 0);
 990         if (IS_ERR(info))
 991                 return PTR_ERR(info);
 992         flags = btrfs_free_space_flags(path->nodes[0], info);
 993         btrfs_release_path(path);
 994
 995         if (flags & BTRFS_FREE_SPACE_USING_BITMAPS) {
 996                 return modify_free_space_bitmap(trans, fs_info, block_group,
 997                                                 path, start, size, 0);
 998         } else {
 999                 return add_free_space_extent(trans, fs_info, block_group, path,
1000                                              start, size);
1001         }
1002 }
1003
1004 int add_to_free_space_tree(struct btrfs_trans_handle *trans,
1005                            struct btrfs_fs_info *fs_info,
1006                            u64 start, u64 size)
1007 {
1008         struct btrfs_block_group_cache *block_group;
1009         struct btrfs_path *path;
1010         int ret;
1011
1012         if (!btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
1013                 return 0;
1014
1015         path = btrfs_alloc_path();
1016         if (!path) {
1017                 ret = -ENOMEM;
1018                 goto out;
1019         }
1020
1021         block_group = btrfs_lookup_block_group(fs_info, start);
1022         if (!block_group) {
1023                 ASSERT(0);
1024                 ret = -ENOENT;
1025                 goto out;
1026         }
1027
1028         mutex_lock(&block_group->free_space_lock);
1029         ret = __add_to_free_space_tree(trans, fs_info, block_group, path, start,
1030                                        size);
1031         mutex_unlock(&block_group->free_space_lock);
1032
1033         btrfs_put_block_group(block_group);
1034 out:
1035         btrfs_free_path(path);
1036         if (ret)
1037                 btrfs_abort_transaction(trans, fs_info->free_space_root, ret);
1038         return ret;
1039 }
1040
1041 /*
1042  * Populate the free space tree by walking the extent tree. Operations on the
1043  * extent tree that happen as a result of writes to the free space tree will go
1044  * through the normal add/remove hooks.
1045  */
1046 static int populate_free_space_tree(struct btrfs_trans_handle *trans,
1047                                     struct btrfs_fs_info *fs_info,
1048                                     struct btrfs_block_group_cache *block_group)
1049 {
1050         struct btrfs_root *extent_root = fs_info->extent_root;
1051         struct btrfs_path *path, *path2;
1052         struct btrfs_key key;
1053         u64 start, end;
1054         int ret;
1055
1056         path = btrfs_alloc_path();
1057         if (!path)
1058                 return -ENOMEM;
1059         path->reada = 1;
1060
1061         path2 = btrfs_alloc_path();
1062         if (!path2) {
1063                 btrfs_free_path(path);
1064                 return -ENOMEM;
1065         }
1066
1067         ret = add_new_free_space_info(trans, fs_info, block_group, path2);
1068         if (ret)
1069                 goto out;
1070
1071         mutex_lock(&block_group->free_space_lock);
1072
1073         /*
1074          * Iterate through all of the extent and metadata items in this block
1075          * group, adding the free space between them and the free space at the
1076          * end. Note that EXTENT_ITEM and METADATA_ITEM are less than
1077          * BLOCK_GROUP_ITEM, so an extent may precede the block group that it's
1078          * contained in.
1079          */
1080         key.objectid = block_group->key.objectid;
1081         key.type = BTRFS_EXTENT_ITEM_KEY;
1082         key.offset = 0;
1083
1084         ret = btrfs_search_slot_for_read(extent_root, &key, path, 1, 0);
1085         if (ret < 0)
1086                 goto out_locked;
1087         ASSERT(ret == 0);
1088
1089         start = block_group->key.objectid;
1090         end = block_group->key.objectid + block_group->key.offset;
1091         while (1) {
1092                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1093
1094                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
1095                     key.type == BTRFS_METADATA_ITEM_KEY) {
1096                         if (key.objectid >= end)
1097                                 break;
1098
1099                         if (start < key.objectid) {
1100                                 ret = __add_to_free_space_tree(trans, fs_info,
1101                                                                block_group,
1102                                                                path2, start,
1103                                                                key.objectid -
1104                                                                start);
1105                                 if (ret)
1106                                         goto out_locked;
1107                         }
1108                         start = key.objectid;
1109                         if (key.type == BTRFS_METADATA_ITEM_KEY)
1110                                 start += fs_info->tree_root->nodesize;
1111                         else
1112                                 start += key.offset;
1113                 } else if (key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
1114                         if (key.objectid != block_group->key.objectid)
1115                                 break;
1116                 }
1117
1118                 ret = btrfs_next_item(extent_root, path);
1119                 if (ret < 0)
1120                         goto out_locked;
1121                 if (ret)
1122                         break;
1123         }
1124         if (start < end) {
1125                 ret = __add_to_free_space_tree(trans, fs_info, block_group,
1126                                                path2, start, end - start);
1127                 if (ret)
1128                         goto out_locked;
1129         }
1130
1131         ret = 0;
1132 out_locked:
1133         mutex_unlock(&block_group->free_space_lock);
1134 out:
1135         btrfs_free_path(path2);
1136         btrfs_free_path(path);
1137         return ret;
1138 }
1139
1140 int btrfs_create_free_space_tree(struct btrfs_fs_info *fs_info)
1141 {
1142         struct btrfs_trans_handle *trans;
1143         struct btrfs_root *tree_root = fs_info->tree_root;
1144         struct btrfs_root *free_space_root;
1145         struct btrfs_block_group_cache *block_group;
1146         struct rb_node *node;
1147         int ret;
1148
1149         trans = btrfs_start_transaction(tree_root, 0);
1150         if (IS_ERR(trans))
1151                 return PTR_ERR(trans);
1152
1153         fs_info->creating_free_space_tree = 1;
1154         free_space_root = btrfs_create_tree(trans, fs_info,
1155                                             BTRFS_FREE_SPACE_TREE_OBJECTID);
1156         if (IS_ERR(free_space_root)) {
1157                 ret = PTR_ERR(free_space_root);
1158                 goto abort;
1159         }
1160         fs_info->free_space_root = free_space_root;
1161
1162         node = rb_first(&fs_info->block_group_cache_tree);
1163         while (node) {
1164                 block_group = rb_entry(node, struct btrfs_block_group_cache,
1165                                        cache_node);
1166                 ret = populate_free_space_tree(trans, fs_info, block_group);
1167                 if (ret)
1168                         goto abort;
1169                 node = rb_next(node);
1170         }
1171
1172         btrfs_set_fs_compat_ro(fs_info, FREE_SPACE_TREE);
1173         btrfs_sysfs_feature_update(fs_info,
1174                 BTRFS_FEATURE_COMPAT_RO_FREE_SPACE_TREE, FEAT_COMPAT_RO);
1175
1176         fs_info->creating_free_space_tree = 0;
1177
1178         ret = btrfs_commit_transaction(trans, tree_root);
1179         if (ret)
1180                 return ret;
1181
1182         return 0;
1183
1184 abort:
1185         fs_info->creating_free_space_tree = 0;
1186         btrfs_abort_transaction(trans, tree_root, ret);
1187         btrfs_end_transaction(trans, tree_root);
1188         return ret;
1189 }
1190
1191 static int clear_free_space_tree(struct btrfs_trans_handle *trans,
1192                                  struct btrfs_root *root)
1193 {
1194         struct btrfs_path *path;
1195         struct btrfs_key key;
1196         int nr;
1197         int ret;
1198
1199         path = btrfs_alloc_path();
1200         if (!path)
1201                 return -ENOMEM;
1202
1203         path->leave_spinning = 1;
1204
1205         key.objectid = 0;
1206         key.type = 0;
1207         key.offset = 0;
1208
1209         while (1) {
1210                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1211                 if (ret < 0)
1212                         goto out;
1213
1214                 nr = btrfs_header_nritems(path->nodes[0]);
1215                 if (!nr)
1216                         break;
1217
1218                 path->slots[0] = 0;
1219                 ret = btrfs_del_items(trans, root, path, 0, nr);
1220                 if (ret)
1221                         goto out;
1222
1223                 btrfs_release_path(path);
1224         }
1225
1226         ret = 0;
1227 out:
1228         btrfs_free_path(path);
1229         return ret;
1230 }
1231
1232 int btrfs_clear_free_space_tree(struct btrfs_fs_info *fs_info)
1233 {
1234         struct btrfs_trans_handle *trans;
1235         struct btrfs_root *tree_root = fs_info->tree_root;
1236         struct btrfs_root *free_space_root = fs_info->free_space_root;
1237         int ret;
1238
1239         trans = btrfs_start_transaction(tree_root, 0);
1240         if (IS_ERR(trans))
1241                 return PTR_ERR(trans);
1242
1243         btrfs_clear_fs_compat_ro(fs_info, FREE_SPACE_TREE);
1244         btrfs_sysfs_feature_update(fs_info,
1245                 BTRFS_FEATURE_COMPAT_RO_FREE_SPACE_TREE, FEAT_COMPAT_RO);
1246
1247         fs_info->free_space_root = NULL;
1248
1249         ret = clear_free_space_tree(trans, free_space_root);
1250         if (ret)
1251                 goto abort;
1252
1253         ret = btrfs_del_root(trans, tree_root, &free_space_root->root_key);
1254         if (ret)
1255                 goto abort;
1256
1257         list_del(&free_space_root->dirty_list);
1258
1259         btrfs_tree_lock(free_space_root->node);
1260         clean_tree_block(trans, tree_root->fs_info, free_space_root->node);
1261         btrfs_tree_unlock(free_space_root->node);
1262         btrfs_free_tree_block(trans, free_space_root, free_space_root->node,
1263                               0, 1);
1264
1265         free_extent_buffer(free_space_root->node);
1266         free_extent_buffer(free_space_root->commit_root);
1267         kfree(free_space_root);
1268
1269         ret = btrfs_commit_transaction(trans, tree_root);
1270         if (ret)
1271                 return ret;
1272
1273         return 0;
1274
1275 abort:
1276         btrfs_abort_transaction(trans, tree_root, ret);
1277         btrfs_end_transaction(trans, tree_root);
1278         return ret;
1279 }
1280
1281 static int __add_block_group_free_space(struct btrfs_trans_handle *trans,
1282                                         struct btrfs_fs_info *fs_info,
1283                                         struct btrfs_block_group_cache *block_group,
1284                                         struct btrfs_path *path)
1285 {
1286         u64 start, end;
1287         int ret;
1288
1289         start = block_group->key.objectid;
1290         end = block_group->key.objectid + block_group->key.offset;
1291
1292         block_group->needs_free_space = 0;
1293
1294         ret = add_new_free_space_info(trans, fs_info, block_group, path);
1295         if (ret)
1296                 return ret;
1297
1298         return __add_to_free_space_tree(trans, fs_info, block_group, path,
1299                                         block_group->key.objectid,
1300                                         block_group->key.offset);
1301 }
1302
1303 int add_block_group_free_space(struct btrfs_trans_handle *trans,
1304                                struct btrfs_fs_info *fs_info,
1305                                struct btrfs_block_group_cache *block_group)
1306 {
1307         struct btrfs_path *path = NULL;
1308         int ret = 0;
1309
1310         if (!btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
1311                 return 0;
1312
1313         mutex_lock(&block_group->free_space_lock);
1314         if (!block_group->needs_free_space)
1315                 goto out;
1316
1317         path = btrfs_alloc_path();
1318         if (!path) {
1319                 ret = -ENOMEM;
1320                 goto out;
1321         }
1322
1323         ret = __add_block_group_free_space(trans, fs_info, block_group, path);
1324
1325 out:
1326         btrfs_free_path(path);
1327         mutex_unlock(&block_group->free_space_lock);
1328         if (ret)
1329                 btrfs_abort_transaction(trans, fs_info->free_space_root, ret);
1330         return ret;
1331 }
1332
1333 int remove_block_group_free_space(struct btrfs_trans_handle *trans,
1334                                   struct btrfs_fs_info *fs_info,
1335                                   struct btrfs_block_group_cache *block_group)
1336 {
1337         struct btrfs_root *root = fs_info->free_space_root;
1338         struct btrfs_path *path;
1339         struct btrfs_key key, found_key;
1340         struct extent_buffer *leaf;
1341         u64 start, end;
1342         int done = 0, nr;
1343         int ret;
1344
1345         if (!btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
1346                 return 0;
1347
1348         if (block_group->needs_free_space) {
1349                 /* We never added this block group to the free space tree. */
1350                 return 0;
1351         }
1352
1353         path = btrfs_alloc_path();
1354         if (!path) {
1355                 ret = -ENOMEM;
1356                 goto out;
1357         }
1358
1359         start = block_group->key.objectid;
1360         end = block_group->key.objectid + block_group->key.offset;
1361
1362         key.objectid = end - 1;
1363         key.type = (u8)-1;
1364         key.offset = (u64)-1;
1365
1366         while (!done) {
1367                 ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
1368                 if (ret)
1369                         goto out;
1370
1371                 leaf = path->nodes[0];
1372                 nr = 0;
1373                 path->slots[0]++;
1374                 while (path->slots[0] > 0) {
1375                         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0] - 1);
1376
1377                         if (found_key.type == BTRFS_FREE_SPACE_INFO_KEY) {
1378                                 ASSERT(found_key.objectid == block_group->key.objectid);
1379                                 ASSERT(found_key.offset == block_group->key.offset);
1380                                 done = 1;
1381                                 nr++;
1382                                 path->slots[0]--;
1383                                 break;
1384                         } else if (found_key.type == BTRFS_FREE_SPACE_EXTENT_KEY ||
1385                                    found_key.type == BTRFS_FREE_SPACE_BITMAP_KEY) {
1386                                 ASSERT(found_key.objectid >= start);
1387                                 ASSERT(found_key.objectid < end);
1388                                 ASSERT(found_key.objectid + found_key.offset <= end);
1389                                 nr++;
1390                                 path->slots[0]--;
1391                         } else {
1392                                 ASSERT(0);
1393                         }
1394                 }
1395
1396                 ret = btrfs_del_items(trans, root, path, path->slots[0], nr);
1397                 if (ret)
1398                         goto out;
1399                 btrfs_release_path(path);
1400         }
1401
1402         ret = 0;
1403 out:
1404         btrfs_free_path(path);
1405         if (ret)
1406                 btrfs_abort_transaction(trans, root, ret);
1407         return ret;
1408 }
1409
1410 static int load_free_space_bitmaps(struct btrfs_caching_control *caching_ctl,
1411                                    struct btrfs_path *path,
1412                                    u32 expected_extent_count)
1413 {
1414         struct btrfs_block_group_cache *block_group;
1415         struct btrfs_fs_info *fs_info;
1416         struct btrfs_root *root;
1417         struct btrfs_key key;
1418         int prev_bit = 0, bit;
1419         /* Initialize to silence GCC. */
1420         u64 extent_start = 0;
1421         u64 end, offset;
1422         u64 total_found = 0;
1423         u32 extent_count = 0;
1424         int ret;
1425
1426         block_group = caching_ctl->block_group;
1427         fs_info = block_group->fs_info;
1428         root = fs_info->free_space_root;
1429
1430         end = block_group->key.objectid + block_group->key.offset;
1431
1432         while (1) {
1433                 ret = btrfs_next_item(root, path);
1434                 if (ret < 0)
1435                         goto out;
1436                 if (ret)
1437                         break;
1438
1439                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1440
1441                 if (key.type == BTRFS_FREE_SPACE_INFO_KEY)
1442                         break;
1443
1444                 ASSERT(key.type == BTRFS_FREE_SPACE_BITMAP_KEY);
1445                 ASSERT(key.objectid < end && key.objectid + key.offset <= end);
1446
1447                 caching_ctl->progress = key.objectid;
1448
1449                 offset = key.objectid;
1450                 while (offset < key.objectid + key.offset) {
1451                         bit = free_space_test_bit(block_group, path, offset);
1452                         if (prev_bit == 0 && bit == 1) {
1453                                 extent_start = offset;
1454                         } else if (prev_bit == 1 && bit == 0) {
1455                                 total_found += add_new_free_space(block_group,
1456                                                                   fs_info,
1457                                                                   extent_start,
1458                                                                   offset);
1459                                 if (total_found > CACHING_CTL_WAKE_UP) {
1460                                         total_found = 0;
1461                                         wake_up(&caching_ctl->wait);
1462                                 }
1463                                 extent_count++;
1464                         }
1465                         prev_bit = bit;
1466                         offset += block_group->sectorsize;
1467                 }
1468         }
1469         if (prev_bit == 1) {
1470                 total_found += add_new_free_space(block_group, fs_info,
1471                                                   extent_start, end);
1472                 extent_count++;
1473         }
1474
1475         if (extent_count != expected_extent_count) {
1476                 btrfs_err(fs_info, "incorrect extent count for %llu; counted %u, expected %u",
1477                           block_group->key.objectid, extent_count,
1478                           expected_extent_count);
1479                 ASSERT(0);
1480                 ret = -EIO;
1481                 goto out;
1482         }
1483
1484         caching_ctl->progress = (u64)-1;
1485
1486         ret = 0;
1487 out:
1488         return ret;
1489 }
1490
1491 static int load_free_space_extents(struct btrfs_caching_control *caching_ctl,
1492                                    struct btrfs_path *path,
1493                                    u32 expected_extent_count)
1494 {
1495         struct btrfs_block_group_cache *block_group;
1496         struct btrfs_fs_info *fs_info;
1497         struct btrfs_root *root;
1498         struct btrfs_key key;
1499         u64 end;
1500         u64 total_found = 0;
1501         u32 extent_count = 0;
1502         int ret;
1503
1504         block_group = caching_ctl->block_group;
1505         fs_info = block_group->fs_info;
1506         root = fs_info->free_space_root;
1507
1508         end = block_group->key.objectid + block_group->key.offset;
1509
1510         while (1) {
1511                 ret = btrfs_next_item(root, path);
1512                 if (ret < 0)
1513                         goto out;
1514                 if (ret)
1515                         break;
1516
1517                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1518
1519                 if (key.type == BTRFS_FREE_SPACE_INFO_KEY)
1520                         break;
1521
1522                 ASSERT(key.type == BTRFS_FREE_SPACE_EXTENT_KEY);
1523                 ASSERT(key.objectid < end && key.objectid + key.offset <= end);
1524
1525                 caching_ctl->progress = key.objectid;
1526
1527                 total_found += add_new_free_space(block_group, fs_info,
1528                                                   key.objectid,
1529                                                   key.objectid + key.offset);
1530                 if (total_found > CACHING_CTL_WAKE_UP) {
1531                         total_found = 0;
1532                         wake_up(&caching_ctl->wait);
1533                 }
1534                 extent_count++;
1535         }
1536
1537         if (extent_count != expected_extent_count) {
1538                 btrfs_err(fs_info, "incorrect extent count for %llu; counted %u, expected %u",
1539                           block_group->key.objectid, extent_count,
1540                           expected_extent_count);
1541                 ASSERT(0);
1542                 ret = -EIO;
1543                 goto out;
1544         }
1545
1546         caching_ctl->progress = (u64)-1;
1547
1548         ret = 0;
1549 out:
1550         return ret;
1551 }
1552
1553 int load_free_space_tree(struct btrfs_caching_control *caching_ctl)
1554 {
1555         struct btrfs_block_group_cache *block_group;
1556         struct btrfs_fs_info *fs_info;
1557         struct btrfs_free_space_info *info;
1558         struct btrfs_path *path;
1559         u32 extent_count, flags;
1560         int ret;
1561
1562         block_group = caching_ctl->block_group;
1563         fs_info = block_group->fs_info;
1564
1565         path = btrfs_alloc_path();
1566         if (!path)
1567                 return -ENOMEM;
1568
1569         /*
1570          * Just like caching_thread() doesn't want to deadlock on the extent
1571          * tree, we don't want to deadlock on the free space tree.
1572          */
1573         path->skip_locking = 1;
1574         path->search_commit_root = 1;
1575         path->reada = 1;
1576
1577         info = search_free_space_info(NULL, fs_info, block_group, path, 0);
1578         if (IS_ERR(info)) {
1579                 ret = PTR_ERR(info);
1580                 goto out;
1581         }
1582         extent_count = btrfs_free_space_extent_count(path->nodes[0], info);
1583         flags = btrfs_free_space_flags(path->nodes[0], info);
1584
1585         /*
1586          * We left path pointing to the free space info item, so now
1587          * load_free_space_foo can just iterate through the free space tree from
1588          * there.
1589          */
1590         if (flags & BTRFS_FREE_SPACE_USING_BITMAPS)
1591                 ret = load_free_space_bitmaps(caching_ctl, path, extent_count);
1592         else
1593                 ret = load_free_space_extents(caching_ctl, path, extent_count);
1594
1595 out:
1596         btrfs_free_path(path);
1597         return ret;
1598 }