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
  27 static int __add_block_group_free_space(struct btrfs_trans_handle *trans,
  28                                         struct btrfs_fs_info *fs_info,
  29                                         struct btrfs_block_group_cache *block_group,
  30                                         struct btrfs_path *path);
  31
  32 void set_free_space_tree_thresholds(struct btrfs_block_group_cache *cache)
  33 {
  34         u32 bitmap_range;
  35         size_t bitmap_size;
  36         u64 num_bitmaps, total_bitmap_size;
  37
  38         /*
  39          * We convert to bitmaps when the disk space required for using extents
  40          * exceeds that required for using bitmaps.
  41          */
  42         bitmap_range = cache->fs_info->sectorsize * BTRFS_FREE_SPACE_BITMAP_BITS;
  43         num_bitmaps = div_u64(cache->key.offset + bitmap_range - 1,
  44                               bitmap_range);
  45         bitmap_size = sizeof(struct btrfs_item) + BTRFS_FREE_SPACE_BITMAP_SIZE;
  46         total_bitmap_size = num_bitmaps * bitmap_size;
  47         cache->bitmap_high_thresh = div_u64(total_bitmap_size,
  48                                             sizeof(struct btrfs_item));
  49
  50         /*
  51          * We allow for a small buffer between the high threshold and low
  52          * threshold to avoid thrashing back and forth between the two formats.
  53          */
  54         if (cache->bitmap_high_thresh > 100)
  55                 cache->bitmap_low_thresh = cache->bitmap_high_thresh - 100;
  56         else
  57                 cache->bitmap_low_thresh = 0;
  58 }
  59
  60 static int add_new_free_space_info(struct btrfs_trans_handle *trans,
  61                                    struct btrfs_fs_info *fs_info,
  62                                    struct btrfs_block_group_cache *block_group,
  63                                    struct btrfs_path *path)
  64 {
  65         struct btrfs_root *root = fs_info->free_space_root;
  66         struct btrfs_free_space_info *info;
  67         struct btrfs_key key;
  68         struct extent_buffer *leaf;
  69         int ret;
  70
  71         key.objectid = block_group->key.objectid;
  72         key.type = BTRFS_FREE_SPACE_INFO_KEY;
  73         key.offset = block_group->key.offset;
  74
  75         ret = btrfs_insert_empty_item(trans, root, path, &key, sizeof(*info));
  76         if (ret)
  77                 goto out;
  78
  79         leaf = path->nodes[0];
  80         info = btrfs_item_ptr(leaf, path->slots[0],
  81                               struct btrfs_free_space_info);
  82         btrfs_set_free_space_extent_count(leaf, info, 0);
  83         btrfs_set_free_space_flags(leaf, info, 0);
  84         btrfs_mark_buffer_dirty(leaf);
  85
  86         ret = 0;
  87 out:
  88         btrfs_release_path(path);
  89         return ret;
  90 }
  91
  92 struct btrfs_free_space_info *
  93 search_free_space_info(struct btrfs_trans_handle *trans,
  94                        struct btrfs_fs_info *fs_info,
  95                        struct btrfs_block_group_cache *block_group,
  96                        struct btrfs_path *path, int cow)
  97 {
  98         struct btrfs_root *root = fs_info->free_space_root;
  99         struct btrfs_key key;
 100         int ret;
 101
 102         key.objectid = block_group->key.objectid;
 103         key.type = BTRFS_FREE_SPACE_INFO_KEY;
 104         key.offset = block_group->key.offset;
 105
 106         ret = btrfs_search_slot(trans, root, &key, path, 0, cow);
 107         if (ret < 0)
 108                 return ERR_PTR(ret);
 109         if (ret != 0) {
 110                 btrfs_warn(fs_info, "missing free space info for %llu",
 111                            block_group->key.objectid);
 112                 ASSERT(0);
 113                 return ERR_PTR(-ENOENT);
 114         }
 115
 116         return btrfs_item_ptr(path->nodes[0], path->slots[0],
 117                               struct btrfs_free_space_info);
 118 }
 119
 120 /*
 121  * btrfs_search_slot() but we're looking for the greatest key less than the
 122  * passed key.
 123  */
 124 static int btrfs_search_prev_slot(struct btrfs_trans_handle *trans,
 125                                   struct btrfs_root *root,
 126                                   struct btrfs_key *key, struct btrfs_path *p,
 127                                   int ins_len, int cow)
 128 {
 129         int ret;
 130
 131         ret = btrfs_search_slot(trans, root, key, p, ins_len, cow);
 132         if (ret < 0)
 133                 return ret;
 134
 135         if (ret == 0) {
 136                 ASSERT(0);
 137                 return -EIO;
 138         }
 139
 140         if (p->slots[0] == 0) {
 141                 ASSERT(0);
 142                 return -EIO;
 143         }
 144         p->slots[0]--;
 145
 146         return 0;
 147 }
 148
 149 static inline u32 free_space_bitmap_size(u64 size, u32 sectorsize)
 150 {
 151         return DIV_ROUND_UP((u32)div_u64(size, sectorsize), BITS_PER_BYTE);
 152 }
 153
 154 static u8 *alloc_bitmap(u32 bitmap_size)
 155 {
 156         void *mem;
 157
 158         /*
 159          * The allocation size varies, observed numbers were < 4K up to 16K.
 160          * Using vmalloc unconditionally would be too heavy, we'll try
 161          * contiguous allocations first.
 162          */
 163         if  (bitmap_size <= PAGE_SIZE)
 164                 return kzalloc(bitmap_size, GFP_NOFS);
 165
 166         mem = kzalloc(bitmap_size, GFP_NOFS | __GFP_NOWARN);
 167         if (mem)
 168                 return mem;
 169
 170         return __vmalloc(bitmap_size, GFP_NOFS | __GFP_HIGHMEM | __GFP_ZERO,
 171                          PAGE_KERNEL);
 172 }
 173
 174 int convert_free_space_to_bitmaps(struct btrfs_trans_handle *trans,
 175                                   struct btrfs_fs_info *fs_info,
 176                                   struct btrfs_block_group_cache *block_group,
 177                                   struct btrfs_path *path)
 178 {
 179         struct btrfs_root *root = fs_info->free_space_root;
 180         struct btrfs_free_space_info *info;
 181         struct btrfs_key key, found_key;
 182         struct extent_buffer *leaf;
 183         u8 *bitmap, *bitmap_cursor;
 184         u64 start, end;
 185         u64 bitmap_range, i;
 186         u32 bitmap_size, flags, expected_extent_count;
 187         u32 extent_count = 0;
 188         int done = 0, nr;
 189         int ret;
 190
 191         bitmap_size = free_space_bitmap_size(block_group->key.offset,
 192                                              fs_info->sectorsize);
 193         bitmap = alloc_bitmap(bitmap_size);
 194         if (!bitmap) {
 195                 ret = -ENOMEM;
 196                 goto out;
 197         }
 198
 199         start = block_group->key.objectid;
 200         end = block_group->key.objectid + block_group->key.offset;
 201
 202         key.objectid = end - 1;
 203         key.type = (u8)-1;
 204         key.offset = (u64)-1;
 205
 206         while (!done) {
 207                 ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
 208                 if (ret)
 209                         goto out;
 210
 211                 leaf = path->nodes[0];
 212                 nr = 0;
 213                 path->slots[0]++;
 214                 while (path->slots[0] > 0) {
 215                         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0] - 1);
 216
 217                         if (found_key.type == BTRFS_FREE_SPACE_INFO_KEY) {
 218                                 ASSERT(found_key.objectid == block_group->key.objectid);
 219                                 ASSERT(found_key.offset == block_group->key.offset);
 220                                 done = 1;
 221                                 break;
 222                         } else if (found_key.type == BTRFS_FREE_SPACE_EXTENT_KEY) {
 223                                 u64 first, last;
 224
 225                                 ASSERT(found_key.objectid >= start);
 226                                 ASSERT(found_key.objectid < end);
 227                                 ASSERT(found_key.objectid + found_key.offset <= end);
 228
 229                                 first = div_u64(found_key.objectid - start,
 230                                                 fs_info->sectorsize);
 231                                 last = div_u64(found_key.objectid + found_key.offset - start,
 232                                                fs_info->sectorsize);
 233                                 le_bitmap_set(bitmap, first, last - first);
 234
 235                                 extent_count++;
 236                                 nr++;
 237                                 path->slots[0]--;
 238                         } else {
 239                                 ASSERT(0);
 240                         }
 241                 }
 242
 243                 ret = btrfs_del_items(trans, root, path, path->slots[0], nr);
 244                 if (ret)
 245                         goto out;
 246                 btrfs_release_path(path);
 247         }
 248
 249         info = search_free_space_info(trans, fs_info, block_group, path, 1);
 250         if (IS_ERR(info)) {
 251                 ret = PTR_ERR(info);
 252                 goto out;
 253         }
 254         leaf = path->nodes[0];
 255         flags = btrfs_free_space_flags(leaf, info);
 256         flags |= BTRFS_FREE_SPACE_USING_BITMAPS;
 257         btrfs_set_free_space_flags(leaf, info, flags);
 258         expected_extent_count = btrfs_free_space_extent_count(leaf, info);
 259         btrfs_mark_buffer_dirty(leaf);
 260         btrfs_release_path(path);
 261
 262         if (extent_count != expected_extent_count) {
 263                 btrfs_err(fs_info,
 264                           "incorrect extent count for %llu; counted %u, expected %u",
 265                           block_group->key.objectid, extent_count,
 266                           expected_extent_count);
 267                 ASSERT(0);
 268                 ret = -EIO;
 269                 goto out;
 270         }
 271
 272         bitmap_cursor = bitmap;
 273         bitmap_range = fs_info->sectorsize * BTRFS_FREE_SPACE_BITMAP_BITS;
 274         i = start;
 275         while (i < end) {
 276                 unsigned long ptr;
 277                 u64 extent_size;
 278                 u32 data_size;
 279
 280                 extent_size = min(end - i, bitmap_range);
 281                 data_size = free_space_bitmap_size(extent_size,
 282                                                    fs_info->sectorsize);
 283
 284                 key.objectid = i;
 285                 key.type = BTRFS_FREE_SPACE_BITMAP_KEY;
 286                 key.offset = extent_size;
 287
 288                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 289                                               data_size);
 290                 if (ret)
 291                         goto out;
 292
 293                 leaf = path->nodes[0];
 294                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
 295                 write_extent_buffer(leaf, bitmap_cursor, ptr,
 296                                     data_size);
 297                 btrfs_mark_buffer_dirty(leaf);
 298                 btrfs_release_path(path);
 299
 300                 i += extent_size;
 301                 bitmap_cursor += data_size;
 302         }
 303
 304         ret = 0;
 305 out:
 306         kvfree(bitmap);
 307         if (ret)
 308                 btrfs_abort_transaction(trans, ret);
 309         return ret;
 310 }
 311
 312 int convert_free_space_to_extents(struct btrfs_trans_handle *trans,
 313                                   struct btrfs_fs_info *fs_info,
 314                                   struct btrfs_block_group_cache *block_group,
 315                                   struct btrfs_path *path)
 316 {
 317         struct btrfs_root *root = fs_info->free_space_root;
 318         struct btrfs_free_space_info *info;
 319         struct btrfs_key key, found_key;
 320         struct extent_buffer *leaf;
 321         u8 *bitmap;
 322         u64 start, end;
 323         /* Initialize to silence GCC. */
 324         u64 extent_start = 0;
 325         u64 offset;
 326         u32 bitmap_size, flags, expected_extent_count;
 327         int prev_bit = 0, bit, bitnr;
 328         u32 extent_count = 0;
 329         int done = 0, nr;
 330         int ret;
 331
 332         bitmap_size = free_space_bitmap_size(block_group->key.offset,
 333                                              fs_info->sectorsize);
 334         bitmap = alloc_bitmap(bitmap_size);
 335         if (!bitmap) {
 336                 ret = -ENOMEM;
 337                 goto out;
 338         }
 339
 340         start = block_group->key.objectid;
 341         end = block_group->key.objectid + block_group->key.offset;
 342
 343         key.objectid = end - 1;
 344         key.type = (u8)-1;
 345         key.offset = (u64)-1;
 346
 347         while (!done) {
 348                 ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
 349                 if (ret)
 350                         goto out;
 351
 352                 leaf = path->nodes[0];
 353                 nr = 0;
 354                 path->slots[0]++;
 355                 while (path->slots[0] > 0) {
 356                         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0] - 1);
 357
 358                         if (found_key.type == BTRFS_FREE_SPACE_INFO_KEY) {
 359                                 ASSERT(found_key.objectid == block_group->key.objectid);
 360                                 ASSERT(found_key.offset == block_group->key.offset);
 361                                 done = 1;
 362                                 break;
 363                         } else if (found_key.type == BTRFS_FREE_SPACE_BITMAP_KEY) {
 364                                 unsigned long ptr;
 365                                 u8 *bitmap_cursor;
 366                                 u32 bitmap_pos, data_size;
 367
 368                                 ASSERT(found_key.objectid >= start);
 369                                 ASSERT(found_key.objectid < end);
 370                                 ASSERT(found_key.objectid + found_key.offset <= end);
 371
 372                                 bitmap_pos = div_u64(found_key.objectid - start,
 373                                                      fs_info->sectorsize *
 374                                                      BITS_PER_BYTE);
 375                                 bitmap_cursor = bitmap + bitmap_pos;
 376                                 data_size = free_space_bitmap_size(found_key.offset,
 377                                                                    fs_info->sectorsize);
 378
 379                                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0] - 1);
 380                                 read_extent_buffer(leaf, bitmap_cursor, ptr,
 381                                                    data_size);
 382
 383                                 nr++;
 384                                 path->slots[0]--;
 385                         } else {
 386                                 ASSERT(0);
 387                         }
 388                 }
 389
 390                 ret = btrfs_del_items(trans, root, path, path->slots[0], nr);
 391                 if (ret)
 392                         goto out;
 393                 btrfs_release_path(path);
 394         }
 395
 396         info = search_free_space_info(trans, fs_info, block_group, path, 1);
 397         if (IS_ERR(info)) {
 398                 ret = PTR_ERR(info);
 399                 goto out;
 400         }
 401         leaf = path->nodes[0];
 402         flags = btrfs_free_space_flags(leaf, info);
 403         flags &= ~BTRFS_FREE_SPACE_USING_BITMAPS;
 404         btrfs_set_free_space_flags(leaf, info, flags);
 405         expected_extent_count = btrfs_free_space_extent_count(leaf, info);
 406         btrfs_mark_buffer_dirty(leaf);
 407         btrfs_release_path(path);
 408
 409         offset = start;
 410         bitnr = 0;
 411         while (offset < end) {
 412                 bit = !!le_test_bit(bitnr, bitmap);
 413                 if (prev_bit == 0 && bit == 1) {
 414                         extent_start = offset;
 415                 } else if (prev_bit == 1 && bit == 0) {
 416                         key.objectid = extent_start;
 417                         key.type = BTRFS_FREE_SPACE_EXTENT_KEY;
 418                         key.offset = offset - extent_start;
 419
 420                         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
 421                         if (ret)
 422                                 goto out;
 423                         btrfs_release_path(path);
 424
 425                         extent_count++;
 426                 }
 427                 prev_bit = bit;
 428                 offset += fs_info->sectorsize;
 429                 bitnr++;
 430         }
 431         if (prev_bit == 1) {
 432                 key.objectid = extent_start;
 433                 key.type = BTRFS_FREE_SPACE_EXTENT_KEY;
 434                 key.offset = end - extent_start;
 435
 436                 ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
 437                 if (ret)
 438                         goto out;
 439                 btrfs_release_path(path);
 440
 441                 extent_count++;
 442         }
 443
 444         if (extent_count != expected_extent_count) {
 445                 btrfs_err(fs_info,
 446                           "incorrect extent count for %llu; counted %u, expected %u",
 447                           block_group->key.objectid, extent_count,
 448                           expected_extent_count);
 449                 ASSERT(0);
 450                 ret = -EIO;
 451                 goto out;
 452         }
 453
 454         ret = 0;
 455 out:
 456         kvfree(bitmap);
 457         if (ret)
 458                 btrfs_abort_transaction(trans, ret);
 459         return ret;
 460 }
 461
 462 static int update_free_space_extent_count(struct btrfs_trans_handle *trans,
 463                                           struct btrfs_fs_info *fs_info,
 464                                           struct btrfs_block_group_cache *block_group,
 465                                           struct btrfs_path *path,
 466                                           int new_extents)
 467 {
 468         struct btrfs_free_space_info *info;
 469         u32 flags;
 470         u32 extent_count;
 471         int ret = 0;
 472
 473         if (new_extents == 0)
 474                 return 0;
 475
 476         info = search_free_space_info(trans, fs_info, block_group, path, 1);
 477         if (IS_ERR(info)) {
 478                 ret = PTR_ERR(info);
 479                 goto out;
 480         }
 481         flags = btrfs_free_space_flags(path->nodes[0], info);
 482         extent_count = btrfs_free_space_extent_count(path->nodes[0], info);
 483
 484         extent_count += new_extents;
 485         btrfs_set_free_space_extent_count(path->nodes[0], info, extent_count);
 486         btrfs_mark_buffer_dirty(path->nodes[0]);
 487         btrfs_release_path(path);
 488
 489         if (!(flags & BTRFS_FREE_SPACE_USING_BITMAPS) &&
 490             extent_count > block_group->bitmap_high_thresh) {
 491                 ret = convert_free_space_to_bitmaps(trans, fs_info, block_group,
 492                                                     path);
 493         } else if ((flags & BTRFS_FREE_SPACE_USING_BITMAPS) &&
 494                    extent_count < block_group->bitmap_low_thresh) {
 495                 ret = convert_free_space_to_extents(trans, fs_info, block_group,
 496                                                     path);
 497         }
 498
 499 out:
 500         return ret;
 501 }
 502
 503 int free_space_test_bit(struct btrfs_block_group_cache *block_group,
 504                         struct btrfs_path *path, u64 offset)
 505 {
 506         struct extent_buffer *leaf;
 507         struct btrfs_key key;
 508         u64 found_start, found_end;
 509         unsigned long ptr, i;
 510
 511         leaf = path->nodes[0];
 512         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 513         ASSERT(key.type == BTRFS_FREE_SPACE_BITMAP_KEY);
 514
 515         found_start = key.objectid;
 516         found_end = key.objectid + key.offset;
 517         ASSERT(offset >= found_start && offset < found_end);
 518
 519         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
 520         i = div_u64(offset - found_start,
 521                     block_group->fs_info->sectorsize);
 522         return !!extent_buffer_test_bit(leaf, ptr, i);
 523 }
 524
 525 static void free_space_set_bits(struct btrfs_block_group_cache *block_group,
 526                                 struct btrfs_path *path, u64 *start, u64 *size,
 527                                 int bit)
 528 {
 529         struct btrfs_fs_info *fs_info = block_group->fs_info;
 530         struct extent_buffer *leaf;
 531         struct btrfs_key key;
 532         u64 end = *start + *size;
 533         u64 found_start, found_end;
 534         unsigned long ptr, first, last;
 535
 536         leaf = path->nodes[0];
 537         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 538         ASSERT(key.type == BTRFS_FREE_SPACE_BITMAP_KEY);
 539
 540         found_start = key.objectid;
 541         found_end = key.objectid + key.offset;
 542         ASSERT(*start >= found_start && *start < found_end);
 543         ASSERT(end > found_start);
 544
 545         if (end > found_end)
 546                 end = found_end;
 547
 548         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
 549         first = div_u64(*start - found_start, fs_info->sectorsize);
 550         last = div_u64(end - found_start, fs_info->sectorsize);
 551         if (bit)
 552                 extent_buffer_bitmap_set(leaf, ptr, first, last - first);
 553         else
 554                 extent_buffer_bitmap_clear(leaf, ptr, first, last - first);
 555         btrfs_mark_buffer_dirty(leaf);
 556
 557         *size -= end - *start;
 558         *start = end;
 559 }
 560
 561 /*
 562  * We can't use btrfs_next_item() in modify_free_space_bitmap() because
 563  * btrfs_next_leaf() doesn't get the path for writing. We can forgo the fancy
 564  * tree walking in btrfs_next_leaf() anyways because we know exactly what we're
 565  * looking for.
 566  */
 567 static int free_space_next_bitmap(struct btrfs_trans_handle *trans,
 568                                   struct btrfs_root *root, struct btrfs_path *p)
 569 {
 570         struct btrfs_key key;
 571
 572         if (p->slots[0] + 1 < btrfs_header_nritems(p->nodes[0])) {
 573                 p->slots[0]++;
 574                 return 0;
 575         }
 576
 577         btrfs_item_key_to_cpu(p->nodes[0], &key, p->slots[0]);
 578         btrfs_release_path(p);
 579
 580         key.objectid += key.offset;
 581         key.type = (u8)-1;
 582         key.offset = (u64)-1;
 583
 584         return btrfs_search_prev_slot(trans, root, &key, p, 0, 1);
 585 }
 586
 587 /*
 588  * If remove is 1, then we are removing free space, thus clearing bits in the
 589  * bitmap. If remove is 0, then we are adding free space, thus setting bits in
 590  * the bitmap.
 591  */
 592 static int modify_free_space_bitmap(struct btrfs_trans_handle *trans,
 593                                     struct btrfs_fs_info *fs_info,
 594                                     struct btrfs_block_group_cache *block_group,
 595                                     struct btrfs_path *path,
 596                                     u64 start, u64 size, int remove)
 597 {
 598         struct btrfs_root *root = fs_info->free_space_root;
 599         struct btrfs_key key;
 600         u64 end = start + size;
 601         u64 cur_start, cur_size;
 602         int prev_bit, next_bit;
 603         int new_extents;
 604         int ret;
 605
 606         /*
 607          * Read the bit for the block immediately before the extent of space if
 608          * that block is within the block group.
 609          */
 610         if (start > block_group->key.objectid) {
 611                 u64 prev_block = start - block_group->fs_info->sectorsize;
 612
 613                 key.objectid = prev_block;
 614                 key.type = (u8)-1;
 615                 key.offset = (u64)-1;
 616
 617                 ret = btrfs_search_prev_slot(trans, root, &key, path, 0, 1);
 618                 if (ret)
 619                         goto out;
 620
 621                 prev_bit = free_space_test_bit(block_group, path, prev_block);
 622
 623                 /* The previous block may have been in the previous bitmap. */
 624                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 625                 if (start >= key.objectid + key.offset) {
 626                         ret = free_space_next_bitmap(trans, root, path);
 627                         if (ret)
 628                                 goto out;
 629                 }
 630         } else {
 631                 key.objectid = start;
 632                 key.type = (u8)-1;
 633                 key.offset = (u64)-1;
 634
 635                 ret = btrfs_search_prev_slot(trans, root, &key, path, 0, 1);
 636                 if (ret)
 637                         goto out;
 638
 639                 prev_bit = -1;
 640         }
 641
 642         /*
 643          * Iterate over all of the bitmaps overlapped by the extent of space,
 644          * clearing/setting bits as required.
 645          */
 646         cur_start = start;
 647         cur_size = size;
 648         while (1) {
 649                 free_space_set_bits(block_group, path, &cur_start, &cur_size,
 650                                     !remove);
 651                 if (cur_size == 0)
 652                         break;
 653                 ret = free_space_next_bitmap(trans, root, path);
 654                 if (ret)
 655                         goto out;
 656         }
 657
 658         /*
 659          * Read the bit for the block immediately after the extent of space if
 660          * that block is within the block group.
 661          */
 662         if (end < block_group->key.objectid + block_group->key.offset) {
 663                 /* The next block may be in the next bitmap. */
 664                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 665                 if (end >= key.objectid + key.offset) {
 666                         ret = free_space_next_bitmap(trans, root, path);
 667                         if (ret)
 668                                 goto out;
 669                 }
 670
 671                 next_bit = free_space_test_bit(block_group, path, end);
 672         } else {
 673                 next_bit = -1;
 674         }
 675
 676         if (remove) {
 677                 new_extents = -1;
 678                 if (prev_bit == 1) {
 679                         /* Leftover on the left. */
 680                         new_extents++;
 681                 }
 682                 if (next_bit == 1) {
 683                         /* Leftover on the right. */
 684                         new_extents++;
 685                 }
 686         } else {
 687                 new_extents = 1;
 688                 if (prev_bit == 1) {
 689                         /* Merging with neighbor on the left. */
 690                         new_extents--;
 691                 }
 692                 if (next_bit == 1) {
 693                         /* Merging with neighbor on the right. */
 694                         new_extents--;
 695                 }
 696         }
 697
 698         btrfs_release_path(path);
 699         ret = update_free_space_extent_count(trans, fs_info, block_group, path,
 700                                              new_extents);
 701
 702 out:
 703         return ret;
 704 }
 705
 706 static int remove_free_space_extent(struct btrfs_trans_handle *trans,
 707                                     struct btrfs_fs_info *fs_info,
 708                                     struct btrfs_block_group_cache *block_group,
 709                                     struct btrfs_path *path,
 710                                     u64 start, u64 size)
 711 {
 712         struct btrfs_root *root = fs_info->free_space_root;
 713         struct btrfs_key key;
 714         u64 found_start, found_end;
 715         u64 end = start + size;
 716         int new_extents = -1;
 717         int ret;
 718
 719         key.objectid = start;
 720         key.type = (u8)-1;
 721         key.offset = (u64)-1;
 722
 723         ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
 724         if (ret)
 725                 goto out;
 726
 727         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 728
 729         ASSERT(key.type == BTRFS_FREE_SPACE_EXTENT_KEY);
 730
 731         found_start = key.objectid;
 732         found_end = key.objectid + key.offset;
 733         ASSERT(start >= found_start && end <= found_end);
 734
 735         /*
 736          * Okay, now that we've found the free space extent which contains the
 737          * free space that we are removing, there are four cases:
 738          *
 739          * 1. We're using the whole extent: delete the key we found and
 740          * decrement the free space extent count.
 741          * 2. We are using part of the extent starting at the beginning: delete
 742          * the key we found and insert a new key representing the leftover at
 743          * the end. There is no net change in the number of extents.
 744          * 3. We are using part of the extent ending at the end: delete the key
 745          * we found and insert a new key representing the leftover at the
 746          * beginning. There is no net change in the number of extents.
 747          * 4. We are using part of the extent in the middle: delete the key we
 748          * found and insert two new keys representing the leftovers on each
 749          * side. Where we used to have one extent, we now have two, so increment
 750          * the extent count. We may need to convert the block group to bitmaps
 751          * as a result.
 752          */
 753
 754         /* Delete the existing key (cases 1-4). */
 755         ret = btrfs_del_item(trans, root, path);
 756         if (ret)
 757                 goto out;
 758
 759         /* Add a key for leftovers at the beginning (cases 3 and 4). */
 760         if (start > found_start) {
 761                 key.objectid = found_start;
 762                 key.type = BTRFS_FREE_SPACE_EXTENT_KEY;
 763                 key.offset = start - found_start;
 764
 765                 btrfs_release_path(path);
 766                 ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
 767                 if (ret)
 768                         goto out;
 769                 new_extents++;
 770         }
 771
 772         /* Add a key for leftovers at the end (cases 2 and 4). */
 773         if (end < found_end) {
 774                 key.objectid = end;
 775                 key.type = BTRFS_FREE_SPACE_EXTENT_KEY;
 776                 key.offset = found_end - end;
 777
 778                 btrfs_release_path(path);
 779                 ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
 780                 if (ret)
 781                         goto out;
 782                 new_extents++;
 783         }
 784
 785         btrfs_release_path(path);
 786         ret = update_free_space_extent_count(trans, fs_info, block_group, path,
 787                                              new_extents);
 788
 789 out:
 790         return ret;
 791 }
 792
 793 int __remove_from_free_space_tree(struct btrfs_trans_handle *trans,
 794                                   struct btrfs_fs_info *fs_info,
 795                                   struct btrfs_block_group_cache *block_group,
 796                                   struct btrfs_path *path, u64 start, u64 size)
 797 {
 798         struct btrfs_free_space_info *info;
 799         u32 flags;
 800         int ret;
 801
 802         if (block_group->needs_free_space) {
 803                 ret = __add_block_group_free_space(trans, fs_info, block_group,
 804                                                    path);
 805                 if (ret)
 806                         return ret;
 807         }
 808
 809         info = search_free_space_info(NULL, fs_info, block_group, path, 0);
 810         if (IS_ERR(info))
 811                 return PTR_ERR(info);
 812         flags = btrfs_free_space_flags(path->nodes[0], info);
 813         btrfs_release_path(path);
 814
 815         if (flags & BTRFS_FREE_SPACE_USING_BITMAPS) {
 816                 return modify_free_space_bitmap(trans, fs_info, block_group,
 817                                                 path, start, size, 1);
 818         } else {
 819                 return remove_free_space_extent(trans, fs_info, block_group,
 820                                                 path, start, size);
 821         }
 822 }
 823
 824 int remove_from_free_space_tree(struct btrfs_trans_handle *trans,
 825                                 struct btrfs_fs_info *fs_info,
 826                                 u64 start, u64 size)
 827 {
 828         struct btrfs_block_group_cache *block_group;
 829         struct btrfs_path *path;
 830         int ret;
 831
 832         if (!btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 833                 return 0;
 834
 835         path = btrfs_alloc_path();
 836         if (!path) {
 837                 ret = -ENOMEM;
 838                 goto out;
 839         }
 840
 841         block_group = btrfs_lookup_block_group(fs_info, start);
 842         if (!block_group) {
 843                 ASSERT(0);
 844                 ret = -ENOENT;
 845                 goto out;
 846         }
 847
 848         mutex_lock(&block_group->free_space_lock);
 849         ret = __remove_from_free_space_tree(trans, fs_info, block_group, path,
 850                                             start, size);
 851         mutex_unlock(&block_group->free_space_lock);
 852
 853         btrfs_put_block_group(block_group);
 854 out:
 855         btrfs_free_path(path);
 856         if (ret)
 857                 btrfs_abort_transaction(trans, ret);
 858         return ret;
 859 }
 860
 861 static int add_free_space_extent(struct btrfs_trans_handle *trans,
 862                                  struct btrfs_fs_info *fs_info,
 863                                  struct btrfs_block_group_cache *block_group,
 864                                  struct btrfs_path *path,
 865                                  u64 start, u64 size)
 866 {
 867         struct btrfs_root *root = fs_info->free_space_root;
 868         struct btrfs_key key, new_key;
 869         u64 found_start, found_end;
 870         u64 end = start + size;
 871         int new_extents = 1;
 872         int ret;
 873
 874         /*
 875          * We are adding a new extent of free space, but we need to merge
 876          * extents. There are four cases here:
 877          *
 878          * 1. The new extent does not have any immediate neighbors to merge
 879          * with: add the new key and increment the free space extent count. We
 880          * may need to convert the block group to bitmaps as a result.
 881          * 2. The new extent has an immediate neighbor before it: remove the
 882          * previous key and insert a new key combining both of them. There is no
 883          * net change in the number of extents.
 884          * 3. The new extent has an immediate neighbor after it: remove the next
 885          * key and insert a new key combining both of them. There is no net
 886          * change in the number of extents.
 887          * 4. The new extent has immediate neighbors on both sides: remove both
 888          * of the keys and insert a new key combining all of them. Where we used
 889          * to have two extents, we now have one, so decrement the extent count.
 890          */
 891
 892         new_key.objectid = start;
 893         new_key.type = BTRFS_FREE_SPACE_EXTENT_KEY;
 894         new_key.offset = size;
 895
 896         /* Search for a neighbor on the left. */
 897         if (start == block_group->key.objectid)
 898                 goto right;
 899         key.objectid = start - 1;
 900         key.type = (u8)-1;
 901         key.offset = (u64)-1;
 902
 903         ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
 904         if (ret)
 905                 goto out;
 906
 907         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 908
 909         if (key.type != BTRFS_FREE_SPACE_EXTENT_KEY) {
 910                 ASSERT(key.type == BTRFS_FREE_SPACE_INFO_KEY);
 911                 btrfs_release_path(path);
 912                 goto right;
 913         }
 914
 915         found_start = key.objectid;
 916         found_end = key.objectid + key.offset;
 917         ASSERT(found_start >= block_group->key.objectid &&
 918                found_end > block_group->key.objectid);
 919         ASSERT(found_start < start && found_end <= start);
 920
 921         /*
 922          * Delete the neighbor on the left and absorb it into the new key (cases
 923          * 2 and 4).
 924          */
 925         if (found_end == start) {
 926                 ret = btrfs_del_item(trans, root, path);
 927                 if (ret)
 928                         goto out;
 929                 new_key.objectid = found_start;
 930                 new_key.offset += key.offset;
 931                 new_extents--;
 932         }
 933         btrfs_release_path(path);
 934
 935 right:
 936         /* Search for a neighbor on the right. */
 937         if (end == block_group->key.objectid + block_group->key.offset)
 938                 goto insert;
 939         key.objectid = end;
 940         key.type = (u8)-1;
 941         key.offset = (u64)-1;
 942
 943         ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
 944         if (ret)
 945                 goto out;
 946
 947         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 948
 949         if (key.type != BTRFS_FREE_SPACE_EXTENT_KEY) {
 950                 ASSERT(key.type == BTRFS_FREE_SPACE_INFO_KEY);
 951                 btrfs_release_path(path);
 952                 goto insert;
 953         }
 954
 955         found_start = key.objectid;
 956         found_end = key.objectid + key.offset;
 957         ASSERT(found_start >= block_group->key.objectid &&
 958                found_end > block_group->key.objectid);
 959         ASSERT((found_start < start && found_end <= start) ||
 960                (found_start >= end && found_end > end));
 961
 962         /*
 963          * Delete the neighbor on the right and absorb it into the new key
 964          * (cases 3 and 4).
 965          */
 966         if (found_start == end) {
 967                 ret = btrfs_del_item(trans, root, path);
 968                 if (ret)
 969                         goto out;
 970                 new_key.offset += key.offset;
 971                 new_extents--;
 972         }
 973         btrfs_release_path(path);
 974
 975 insert:
 976         /* Insert the new key (cases 1-4). */
 977         ret = btrfs_insert_empty_item(trans, root, path, &new_key, 0);
 978         if (ret)
 979                 goto out;
 980
 981         btrfs_release_path(path);
 982         ret = update_free_space_extent_count(trans, fs_info, block_group, path,
 983                                              new_extents);
 984
 985 out:
 986         return ret;
 987 }
 988
 989 int __add_to_free_space_tree(struct btrfs_trans_handle *trans,
 990                              struct btrfs_fs_info *fs_info,
 991                              struct btrfs_block_group_cache *block_group,
 992                              struct btrfs_path *path, u64 start, u64 size)
 993 {
 994         struct btrfs_free_space_info *info;
 995         u32 flags;
 996         int ret;
 997
 998         if (block_group->needs_free_space) {
 999                 ret = __add_block_group_free_space(trans, fs_info, block_group,
1000                                                    path);
1001                 if (ret)
1002                         return ret;
1003         }
1004
1005         info = search_free_space_info(NULL, fs_info, block_group, path, 0);
1006         if (IS_ERR(info))
1007                 return PTR_ERR(info);
1008         flags = btrfs_free_space_flags(path->nodes[0], info);
1009         btrfs_release_path(path);
1010
1011         if (flags & BTRFS_FREE_SPACE_USING_BITMAPS) {
1012                 return modify_free_space_bitmap(trans, fs_info, block_group,
1013                                                 path, start, size, 0);
1014         } else {
1015                 return add_free_space_extent(trans, fs_info, block_group, path,
1016                                              start, size);
1017         }
1018 }
1019
1020 int add_to_free_space_tree(struct btrfs_trans_handle *trans,
1021                            struct btrfs_fs_info *fs_info,
1022                            u64 start, u64 size)
1023 {
1024         struct btrfs_block_group_cache *block_group;
1025         struct btrfs_path *path;
1026         int ret;
1027
1028         if (!btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
1029                 return 0;
1030
1031         path = btrfs_alloc_path();
1032         if (!path) {
1033                 ret = -ENOMEM;
1034                 goto out;
1035         }
1036
1037         block_group = btrfs_lookup_block_group(fs_info, start);
1038         if (!block_group) {
1039                 ASSERT(0);
1040                 ret = -ENOENT;
1041                 goto out;
1042         }
1043
1044         mutex_lock(&block_group->free_space_lock);
1045         ret = __add_to_free_space_tree(trans, fs_info, block_group, path, start,
1046                                        size);
1047         mutex_unlock(&block_group->free_space_lock);
1048
1049         btrfs_put_block_group(block_group);
1050 out:
1051         btrfs_free_path(path);
1052         if (ret)
1053                 btrfs_abort_transaction(trans, ret);
1054         return ret;
1055 }
1056
1057 /*
1058  * Populate the free space tree by walking the extent tree. Operations on the
1059  * extent tree that happen as a result of writes to the free space tree will go
1060  * through the normal add/remove hooks.
1061  */
1062 static int populate_free_space_tree(struct btrfs_trans_handle *trans,
1063                                     struct btrfs_fs_info *fs_info,
1064                                     struct btrfs_block_group_cache *block_group)
1065 {
1066         struct btrfs_root *extent_root = fs_info->extent_root;
1067         struct btrfs_path *path, *path2;
1068         struct btrfs_key key;
1069         u64 start, end;
1070         int ret;
1071
1072         path = btrfs_alloc_path();
1073         if (!path)
1074                 return -ENOMEM;
1075         path->reada = 1;
1076
1077         path2 = btrfs_alloc_path();
1078         if (!path2) {
1079                 btrfs_free_path(path);
1080                 return -ENOMEM;
1081         }
1082
1083         ret = add_new_free_space_info(trans, fs_info, block_group, path2);
1084         if (ret)
1085                 goto out;
1086
1087         mutex_lock(&block_group->free_space_lock);
1088
1089         /*
1090          * Iterate through all of the extent and metadata items in this block
1091          * group, adding the free space between them and the free space at the
1092          * end. Note that EXTENT_ITEM and METADATA_ITEM are less than
1093          * BLOCK_GROUP_ITEM, so an extent may precede the block group that it's
1094          * contained in.
1095          */
1096         key.objectid = block_group->key.objectid;
1097         key.type = BTRFS_EXTENT_ITEM_KEY;
1098         key.offset = 0;
1099
1100         ret = btrfs_search_slot_for_read(extent_root, &key, path, 1, 0);
1101         if (ret < 0)
1102                 goto out_locked;
1103         ASSERT(ret == 0);
1104
1105         start = block_group->key.objectid;
1106         end = block_group->key.objectid + block_group->key.offset;
1107         while (1) {
1108                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1109
1110                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
1111                     key.type == BTRFS_METADATA_ITEM_KEY) {
1112                         if (key.objectid >= end)
1113                                 break;
1114
1115                         if (start < key.objectid) {
1116                                 ret = __add_to_free_space_tree(trans, fs_info,
1117                                                                block_group,
1118                                                                path2, start,
1119                                                                key.objectid -
1120                                                                start);
1121                                 if (ret)
1122                                         goto out_locked;
1123                         }
1124                         start = key.objectid;
1125                         if (key.type == BTRFS_METADATA_ITEM_KEY)
1126                                 start += fs_info->nodesize;
1127                         else
1128                                 start += key.offset;
1129                 } else if (key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
1130                         if (key.objectid != block_group->key.objectid)
1131                                 break;
1132                 }
1133
1134                 ret = btrfs_next_item(extent_root, path);
1135                 if (ret < 0)
1136                         goto out_locked;
1137                 if (ret)
1138                         break;
1139         }
1140         if (start < end) {
1141                 ret = __add_to_free_space_tree(trans, fs_info, block_group,
1142                                                path2, start, end - start);
1143                 if (ret)
1144                         goto out_locked;
1145         }
1146
1147         ret = 0;
1148 out_locked:
1149         mutex_unlock(&block_group->free_space_lock);
1150 out:
1151         btrfs_free_path(path2);
1152         btrfs_free_path(path);
1153         return ret;
1154 }
1155
1156 int btrfs_create_free_space_tree(struct btrfs_fs_info *fs_info)
1157 {
1158         struct btrfs_trans_handle *trans;
1159         struct btrfs_root *tree_root = fs_info->tree_root;
1160         struct btrfs_root *free_space_root;
1161         struct btrfs_block_group_cache *block_group;
1162         struct rb_node *node;
1163         int ret;
1164
1165         trans = btrfs_start_transaction(tree_root, 0);
1166         if (IS_ERR(trans))
1167                 return PTR_ERR(trans);
1168
1169         set_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags);
1170         free_space_root = btrfs_create_tree(trans, fs_info,
1171                                             BTRFS_FREE_SPACE_TREE_OBJECTID);
1172         if (IS_ERR(free_space_root)) {
1173                 ret = PTR_ERR(free_space_root);
1174                 goto abort;
1175         }
1176         fs_info->free_space_root = free_space_root;
1177
1178         node = rb_first(&fs_info->block_group_cache_tree);
1179         while (node) {
1180                 block_group = rb_entry(node, struct btrfs_block_group_cache,
1181                                        cache_node);
1182                 ret = populate_free_space_tree(trans, fs_info, block_group);
1183                 if (ret)
1184                         goto abort;
1185                 node = rb_next(node);
1186         }
1187
1188         btrfs_set_fs_compat_ro(fs_info, FREE_SPACE_TREE);
1189         btrfs_set_fs_compat_ro(fs_info, FREE_SPACE_TREE_VALID);
1190         clear_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags);
1191
1192         ret = btrfs_commit_transaction(trans);
1193         if (ret)
1194                 return ret;
1195
1196         return 0;
1197
1198 abort:
1199         clear_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags);
1200         btrfs_abort_transaction(trans, ret);
1201         btrfs_end_transaction(trans);
1202         return ret;
1203 }
1204
1205 static int clear_free_space_tree(struct btrfs_trans_handle *trans,
1206                                  struct btrfs_root *root)
1207 {
1208         struct btrfs_path *path;
1209         struct btrfs_key key;
1210         int nr;
1211         int ret;
1212
1213         path = btrfs_alloc_path();
1214         if (!path)
1215                 return -ENOMEM;
1216
1217         path->leave_spinning = 1;
1218
1219         key.objectid = 0;
1220         key.type = 0;
1221         key.offset = 0;
1222
1223         while (1) {
1224                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1225                 if (ret < 0)
1226                         goto out;
1227
1228                 nr = btrfs_header_nritems(path->nodes[0]);
1229                 if (!nr)
1230                         break;
1231
1232                 path->slots[0] = 0;
1233                 ret = btrfs_del_items(trans, root, path, 0, nr);
1234                 if (ret)
1235                         goto out;
1236
1237                 btrfs_release_path(path);
1238         }
1239
1240         ret = 0;
1241 out:
1242         btrfs_free_path(path);
1243         return ret;
1244 }
1245
1246 int btrfs_clear_free_space_tree(struct btrfs_fs_info *fs_info)
1247 {
1248         struct btrfs_trans_handle *trans;
1249         struct btrfs_root *tree_root = fs_info->tree_root;
1250         struct btrfs_root *free_space_root = fs_info->free_space_root;
1251         int ret;
1252
1253         trans = btrfs_start_transaction(tree_root, 0);
1254         if (IS_ERR(trans))
1255                 return PTR_ERR(trans);
1256
1257         btrfs_clear_fs_compat_ro(fs_info, FREE_SPACE_TREE);
1258         btrfs_clear_fs_compat_ro(fs_info, FREE_SPACE_TREE_VALID);
1259         fs_info->free_space_root = NULL;
1260
1261         ret = clear_free_space_tree(trans, free_space_root);
1262         if (ret)
1263                 goto abort;
1264
1265         ret = btrfs_del_root(trans, tree_root, &free_space_root->root_key);
1266         if (ret)
1267                 goto abort;
1268
1269         list_del(&free_space_root->dirty_list);
1270
1271         btrfs_tree_lock(free_space_root->node);
1272         clean_tree_block(fs_info, free_space_root->node);
1273         btrfs_tree_unlock(free_space_root->node);
1274         btrfs_free_tree_block(trans, free_space_root, free_space_root->node,
1275                               0, 1);
1276
1277         free_extent_buffer(free_space_root->node);
1278         free_extent_buffer(free_space_root->commit_root);
1279         kfree(free_space_root);
1280
1281         ret = btrfs_commit_transaction(trans);
1282         if (ret)
1283                 return ret;
1284
1285         return 0;
1286
1287 abort:
1288         btrfs_abort_transaction(trans, ret);
1289         btrfs_end_transaction(trans);
1290         return ret;
1291 }
1292
1293 static int __add_block_group_free_space(struct btrfs_trans_handle *trans,
1294                                         struct btrfs_fs_info *fs_info,
1295                                         struct btrfs_block_group_cache *block_group,
1296                                         struct btrfs_path *path)
1297 {
1298         u64 start, end;
1299         int ret;
1300
1301         start = block_group->key.objectid;
1302         end = block_group->key.objectid + block_group->key.offset;
1303
1304         block_group->needs_free_space = 0;
1305
1306         ret = add_new_free_space_info(trans, fs_info, block_group, path);
1307         if (ret)
1308                 return ret;
1309
1310         return __add_to_free_space_tree(trans, fs_info, block_group, path,
1311                                         block_group->key.objectid,
1312                                         block_group->key.offset);
1313 }
1314
1315 int add_block_group_free_space(struct btrfs_trans_handle *trans,
1316                                struct btrfs_fs_info *fs_info,
1317                                struct btrfs_block_group_cache *block_group)
1318 {
1319         struct btrfs_path *path = NULL;
1320         int ret = 0;
1321
1322         if (!btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
1323                 return 0;
1324
1325         mutex_lock(&block_group->free_space_lock);
1326         if (!block_group->needs_free_space)
1327                 goto out;
1328
1329         path = btrfs_alloc_path();
1330         if (!path) {
1331                 ret = -ENOMEM;
1332                 goto out;
1333         }
1334
1335         ret = __add_block_group_free_space(trans, fs_info, block_group, path);
1336
1337 out:
1338         btrfs_free_path(path);
1339         mutex_unlock(&block_group->free_space_lock);
1340         if (ret)
1341                 btrfs_abort_transaction(trans, ret);
1342         return ret;
1343 }
1344
1345 int remove_block_group_free_space(struct btrfs_trans_handle *trans,
1346                                   struct btrfs_fs_info *fs_info,
1347                                   struct btrfs_block_group_cache *block_group)
1348 {
1349         struct btrfs_root *root = fs_info->free_space_root;
1350         struct btrfs_path *path;
1351         struct btrfs_key key, found_key;
1352         struct extent_buffer *leaf;
1353         u64 start, end;
1354         int done = 0, nr;
1355         int ret;
1356
1357         if (!btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
1358                 return 0;
1359
1360         if (block_group->needs_free_space) {
1361                 /* We never added this block group to the free space tree. */
1362                 return 0;
1363         }
1364
1365         path = btrfs_alloc_path();
1366         if (!path) {
1367                 ret = -ENOMEM;
1368                 goto out;
1369         }
1370
1371         start = block_group->key.objectid;
1372         end = block_group->key.objectid + block_group->key.offset;
1373
1374         key.objectid = end - 1;
1375         key.type = (u8)-1;
1376         key.offset = (u64)-1;
1377
1378         while (!done) {
1379                 ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
1380                 if (ret)
1381                         goto out;
1382
1383                 leaf = path->nodes[0];
1384                 nr = 0;
1385                 path->slots[0]++;
1386                 while (path->slots[0] > 0) {
1387                         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0] - 1);
1388
1389                         if (found_key.type == BTRFS_FREE_SPACE_INFO_KEY) {
1390                                 ASSERT(found_key.objectid == block_group->key.objectid);
1391                                 ASSERT(found_key.offset == block_group->key.offset);
1392                                 done = 1;
1393                                 nr++;
1394                                 path->slots[0]--;
1395                                 break;
1396                         } else if (found_key.type == BTRFS_FREE_SPACE_EXTENT_KEY ||
1397                                    found_key.type == BTRFS_FREE_SPACE_BITMAP_KEY) {
1398                                 ASSERT(found_key.objectid >= start);
1399                                 ASSERT(found_key.objectid < end);
1400                                 ASSERT(found_key.objectid + found_key.offset <= end);
1401                                 nr++;
1402                                 path->slots[0]--;
1403                         } else {
1404                                 ASSERT(0);
1405                         }
1406                 }
1407
1408                 ret = btrfs_del_items(trans, root, path, path->slots[0], nr);
1409                 if (ret)
1410                         goto out;
1411                 btrfs_release_path(path);
1412         }
1413
1414         ret = 0;
1415 out:
1416         btrfs_free_path(path);
1417         if (ret)
1418                 btrfs_abort_transaction(trans, ret);
1419         return ret;
1420 }
1421
1422 static int load_free_space_bitmaps(struct btrfs_caching_control *caching_ctl,
1423                                    struct btrfs_path *path,
1424                                    u32 expected_extent_count)
1425 {
1426         struct btrfs_block_group_cache *block_group;
1427         struct btrfs_fs_info *fs_info;
1428         struct btrfs_root *root;
1429         struct btrfs_key key;
1430         int prev_bit = 0, bit;
1431         /* Initialize to silence GCC. */
1432         u64 extent_start = 0;
1433         u64 end, offset;
1434         u64 total_found = 0;
1435         u32 extent_count = 0;
1436         int ret;
1437
1438         block_group = caching_ctl->block_group;
1439         fs_info = block_group->fs_info;
1440         root = fs_info->free_space_root;
1441
1442         end = block_group->key.objectid + block_group->key.offset;
1443
1444         while (1) {
1445                 ret = btrfs_next_item(root, path);
1446                 if (ret < 0)
1447                         goto out;
1448                 if (ret)
1449                         break;
1450
1451                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1452
1453                 if (key.type == BTRFS_FREE_SPACE_INFO_KEY)
1454                         break;
1455
1456                 ASSERT(key.type == BTRFS_FREE_SPACE_BITMAP_KEY);
1457                 ASSERT(key.objectid < end && key.objectid + key.offset <= end);
1458
1459                 caching_ctl->progress = key.objectid;
1460
1461                 offset = key.objectid;
1462                 while (offset < key.objectid + key.offset) {
1463                         bit = free_space_test_bit(block_group, path, offset);
1464                         if (prev_bit == 0 && bit == 1) {
1465                                 extent_start = offset;
1466                         } else if (prev_bit == 1 && bit == 0) {
1467                                 total_found += add_new_free_space(block_group,
1468                                                                   fs_info,
1469                                                                   extent_start,
1470                                                                   offset);
1471                                 if (total_found > CACHING_CTL_WAKE_UP) {
1472                                         total_found = 0;
1473                                         wake_up(&caching_ctl->wait);
1474                                 }
1475                                 extent_count++;
1476                         }
1477                         prev_bit = bit;
1478                         offset += fs_info->sectorsize;
1479                 }
1480         }
1481         if (prev_bit == 1) {
1482                 total_found += add_new_free_space(block_group, fs_info,
1483                                                   extent_start, end);
1484                 extent_count++;
1485         }
1486
1487         if (extent_count != expected_extent_count) {
1488                 btrfs_err(fs_info,
1489                           "incorrect extent count for %llu; counted %u, expected %u",
1490                           block_group->key.objectid, extent_count,
1491                           expected_extent_count);
1492                 ASSERT(0);
1493                 ret = -EIO;
1494                 goto out;
1495         }
1496
1497         caching_ctl->progress = (u64)-1;
1498
1499         ret = 0;
1500 out:
1501         return ret;
1502 }
1503
1504 static int load_free_space_extents(struct btrfs_caching_control *caching_ctl,
1505                                    struct btrfs_path *path,
1506                                    u32 expected_extent_count)
1507 {
1508         struct btrfs_block_group_cache *block_group;
1509         struct btrfs_fs_info *fs_info;
1510         struct btrfs_root *root;
1511         struct btrfs_key key;
1512         u64 end;
1513         u64 total_found = 0;
1514         u32 extent_count = 0;
1515         int ret;
1516
1517         block_group = caching_ctl->block_group;
1518         fs_info = block_group->fs_info;
1519         root = fs_info->free_space_root;
1520
1521         end = block_group->key.objectid + block_group->key.offset;
1522
1523         while (1) {
1524                 ret = btrfs_next_item(root, path);
1525                 if (ret < 0)
1526                         goto out;
1527                 if (ret)
1528                         break;
1529
1530                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1531
1532                 if (key.type == BTRFS_FREE_SPACE_INFO_KEY)
1533                         break;
1534
1535                 ASSERT(key.type == BTRFS_FREE_SPACE_EXTENT_KEY);
1536                 ASSERT(key.objectid < end && key.objectid + key.offset <= end);
1537
1538                 caching_ctl->progress = key.objectid;
1539
1540                 total_found += add_new_free_space(block_group, fs_info,
1541                                                   key.objectid,
1542                                                   key.objectid + key.offset);
1543                 if (total_found > CACHING_CTL_WAKE_UP) {
1544                         total_found = 0;
1545                         wake_up(&caching_ctl->wait);
1546                 }
1547                 extent_count++;
1548         }
1549
1550         if (extent_count != expected_extent_count) {
1551                 btrfs_err(fs_info,
1552                           "incorrect extent count for %llu; counted %u, expected %u",
1553                           block_group->key.objectid, extent_count,
1554                           expected_extent_count);
1555                 ASSERT(0);
1556                 ret = -EIO;
1557                 goto out;
1558         }
1559
1560         caching_ctl->progress = (u64)-1;
1561
1562         ret = 0;
1563 out:
1564         return ret;
1565 }
1566
1567 int load_free_space_tree(struct btrfs_caching_control *caching_ctl)
1568 {
1569         struct btrfs_block_group_cache *block_group;
1570         struct btrfs_fs_info *fs_info;
1571         struct btrfs_free_space_info *info;
1572         struct btrfs_path *path;
1573         u32 extent_count, flags;
1574         int ret;
1575
1576         block_group = caching_ctl->block_group;
1577         fs_info = block_group->fs_info;
1578
1579         path = btrfs_alloc_path();
1580         if (!path)
1581                 return -ENOMEM;
1582
1583         /*
1584          * Just like caching_thread() doesn't want to deadlock on the extent
1585          * tree, we don't want to deadlock on the free space tree.
1586          */
1587         path->skip_locking = 1;
1588         path->search_commit_root = 1;
1589         path->reada = 1;
1590
1591         info = search_free_space_info(NULL, fs_info, block_group, path, 0);
1592         if (IS_ERR(info)) {
1593                 ret = PTR_ERR(info);
1594                 goto out;
1595         }
1596         extent_count = btrfs_free_space_extent_count(path->nodes[0], info);
1597         flags = btrfs_free_space_flags(path->nodes[0], info);
1598
1599         /*
1600          * We left path pointing to the free space info item, so now
1601          * load_free_space_foo can just iterate through the free space tree from
1602          * there.
1603          */
1604         if (flags & BTRFS_FREE_SPACE_USING_BITMAPS)
1605                 ret = load_free_space_bitmaps(caching_ctl, path, extent_count);
1606         else
1607                 ret = load_free_space_extents(caching_ctl, path, extent_count);
1608
1609 out:
1610         btrfs_free_path(path);
1611         return ret;
1612 }