fs/btrfs/extent-tree.c

   1 #include <linux/module.h>
   2 #include "ctree.h"
   3 #include "disk-io.h"
   4 #include "print-tree.h"
   5 #include "transaction.h"
   6
   7 static int find_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
   8                             *orig_root, u64 num_blocks, u64 search_start, u64
   9                             search_end, struct btrfs_key *ins);
  10 static int finish_current_insert(struct btrfs_trans_handle *trans, struct
  11                                  btrfs_root *extent_root);
  12 static int del_pending_extents(struct btrfs_trans_handle *trans, struct
  13                                btrfs_root *extent_root);
  14
  15 static int find_search_start(struct btrfs_root *root, int data)
  16 {
  17         struct btrfs_block_group_cache *cache[8];
  18         struct btrfs_fs_info *info = root->fs_info;
  19         u64 used;
  20         u64 last;
  21         int i;
  22         int ret;
  23
  24         cache[0] = info->block_group_cache;
  25         if (!cache[0])
  26                 goto find_new;
  27         used = btrfs_block_group_used(&cache[0]->item);
  28         if (used < (cache[0]->key.offset * 3 / 2))
  29                 return 0;
  30 find_new:
  31         last = 0;
  32         while(1) {
  33                 ret = radix_tree_gang_lookup_tag(&info->block_group_radix,
  34                                                  (void **)cache,
  35                                                  last, ARRAY_SIZE(cache),
  36                                                  BTRFS_BLOCK_GROUP_DIRTY);
  37                 if (!ret)
  38                         break;
  39                 for (i = 0; i < ret; i++) {
  40                         used = btrfs_block_group_used(&cache[i]->item);
  41                         if (used < (cache[i]->key.offset * 3 / 2)) {
  42                                 info->block_group_cache = cache[i];
  43                                 cache[i]->last_alloc = cache[i]->first_free;
  44                                 return 0;
  45                         }
  46                         last = cache[i]->key.objectid +
  47                                 cache[i]->key.offset - 1;
  48                 }
  49         }
  50         last = 0;
  51         while(1) {
  52                 ret = radix_tree_gang_lookup(&info->block_group_radix,
  53                                                  (void **)cache,
  54                                                  last, ARRAY_SIZE(cache));
  55                 if (!ret)
  56                         break;
  57                 for (i = 0; i < ret; i++) {
  58                         used = btrfs_block_group_used(&cache[i]->item);
  59                         if (used < (cache[i]->key.offset * 3 / 2)) {
  60                                 info->block_group_cache = cache[i];
  61                                 cache[i]->last_alloc = cache[i]->first_free;
  62                                 return 0;
  63                         }
  64                         last = cache[i]->key.objectid +
  65                                 cache[i]->key.offset - 1;
  66                 }
  67         }
  68         info->block_group_cache = NULL;
  69         return 0;
  70 }
  71
  72 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
  73                                 struct btrfs_root *root,
  74                                 u64 blocknr, u64 num_blocks)
  75 {
  76         struct btrfs_path *path;
  77         int ret;
  78         struct btrfs_key key;
  79         struct btrfs_leaf *l;
  80         struct btrfs_extent_item *item;
  81         struct btrfs_key ins;
  82         u32 refs;
  83
  84         find_free_extent(trans, root->fs_info->extent_root, 0, 0, (u64)-1,
  85                          &ins);
  86         path = btrfs_alloc_path();
  87         BUG_ON(!path);
  88         btrfs_init_path(path);
  89         key.objectid = blocknr;
  90         key.flags = 0;
  91         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
  92         key.offset = num_blocks;
  93         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key, path,
  94                                 0, 1);
  95         if (ret != 0) {
  96 printk("can't find block %Lu %Lu\n", blocknr, num_blocks);
  97                 BUG();
  98         }
  99         BUG_ON(ret != 0);
 100         l = btrfs_buffer_leaf(path->nodes[0]);
 101         item = btrfs_item_ptr(l, path->slots[0], struct btrfs_extent_item);
 102         refs = btrfs_extent_refs(item);
 103         btrfs_set_extent_refs(item, refs + 1);
 104         btrfs_mark_buffer_dirty(path->nodes[0]);
 105
 106         btrfs_release_path(root->fs_info->extent_root, path);
 107         btrfs_free_path(path);
 108         finish_current_insert(trans, root->fs_info->extent_root);
 109         del_pending_extents(trans, root->fs_info->extent_root);
 110         return 0;
 111 }
 112
 113 static int lookup_extent_ref(struct btrfs_trans_handle *trans,
 114                              struct btrfs_root *root, u64 blocknr,
 115                              u64 num_blocks, u32 *refs)
 116 {
 117         struct btrfs_path *path;
 118         int ret;
 119         struct btrfs_key key;
 120         struct btrfs_leaf *l;
 121         struct btrfs_extent_item *item;
 122
 123         path = btrfs_alloc_path();
 124         btrfs_init_path(path);
 125         key.objectid = blocknr;
 126         key.offset = num_blocks;
 127         key.flags = 0;
 128         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 129         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key, path,
 130                                 0, 0);
 131         if (ret != 0)
 132                 BUG();
 133         l = btrfs_buffer_leaf(path->nodes[0]);
 134         item = btrfs_item_ptr(l, path->slots[0], struct btrfs_extent_item);
 135         *refs = btrfs_extent_refs(item);
 136         btrfs_release_path(root->fs_info->extent_root, path);
 137         btrfs_free_path(path);
 138         return 0;
 139 }
 140
 141 int btrfs_inc_root_ref(struct btrfs_trans_handle *trans,
 142                        struct btrfs_root *root)
 143 {
 144         return btrfs_inc_extent_ref(trans, root, bh_blocknr(root->node), 1);
 145 }
 146
 147 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
 148                   struct buffer_head *buf)
 149 {
 150         u64 blocknr;
 151         struct btrfs_node *buf_node;
 152         struct btrfs_leaf *buf_leaf;
 153         struct btrfs_disk_key *key;
 154         struct btrfs_file_extent_item *fi;
 155         int i;
 156         int leaf;
 157         int ret;
 158
 159         if (!root->ref_cows)
 160                 return 0;
 161         buf_node = btrfs_buffer_node(buf);
 162         leaf = btrfs_is_leaf(buf_node);
 163         buf_leaf = btrfs_buffer_leaf(buf);
 164         for (i = 0; i < btrfs_header_nritems(&buf_node->header); i++) {
 165                 if (leaf) {
 166                         key = &buf_leaf->items[i].key;
 167                         if (btrfs_disk_key_type(key) != BTRFS_EXTENT_DATA_KEY)
 168                                 continue;
 169                         fi = btrfs_item_ptr(buf_leaf, i,
 170                                             struct btrfs_file_extent_item);
 171                         if (btrfs_file_extent_type(fi) ==
 172                             BTRFS_FILE_EXTENT_INLINE)
 173                                 continue;
 174                         ret = btrfs_inc_extent_ref(trans, root,
 175                                     btrfs_file_extent_disk_blocknr(fi),
 176                                     btrfs_file_extent_disk_num_blocks(fi));
 177                         BUG_ON(ret);
 178                 } else {
 179                         blocknr = btrfs_node_blockptr(buf_node, i);
 180                         ret = btrfs_inc_extent_ref(trans, root, blocknr, 1);
 181                         BUG_ON(ret);
 182                 }
 183         }
 184         return 0;
 185 }
 186
 187 static int write_one_cache_group(struct btrfs_trans_handle *trans,
 188                                  struct btrfs_root *root,
 189                                  struct btrfs_path *path,
 190                                  struct btrfs_block_group_cache *cache)
 191 {
 192         int ret;
 193         int pending_ret;
 194         struct btrfs_root *extent_root = root->fs_info->extent_root;
 195         struct btrfs_block_group_item *bi;
 196         struct btrfs_key ins;
 197
 198         find_free_extent(trans, extent_root, 0, 0, (u64)-1, &ins);
 199         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
 200         BUG_ON(ret);
 201         bi = btrfs_item_ptr(btrfs_buffer_leaf(path->nodes[0]), path->slots[0],
 202                             struct btrfs_block_group_item);
 203         memcpy(bi, &cache->item, sizeof(*bi));
 204         mark_buffer_dirty(path->nodes[0]);
 205         btrfs_release_path(extent_root, path);
 206
 207         finish_current_insert(trans, extent_root);
 208         pending_ret = del_pending_extents(trans, extent_root);
 209         if (ret)
 210                 return ret;
 211         if (pending_ret)
 212                 return pending_ret;
 213         return 0;
 214
 215 }
 216
 217 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
 218                                     struct btrfs_root *root)
 219 {
 220         struct btrfs_block_group_cache *cache[8];
 221         int ret;
 222         int err = 0;
 223         int werr = 0;
 224         struct radix_tree_root *radix = &root->fs_info->block_group_radix;
 225         int i;
 226         struct btrfs_path *path;
 227
 228         path = btrfs_alloc_path();
 229         if (!path)
 230                 return -ENOMEM;
 231
 232         while(1) {
 233                 ret = radix_tree_gang_lookup_tag(radix, (void **)cache,
 234                                                  0, ARRAY_SIZE(cache),
 235                                                  BTRFS_BLOCK_GROUP_DIRTY);
 236                 if (!ret)
 237                         break;
 238                 for (i = 0; i < ret; i++) {
 239                         radix_tree_tag_clear(radix, cache[i]->key.objectid +
 240                                              cache[i]->key.offset - 1,
 241                                              BTRFS_BLOCK_GROUP_DIRTY);
 242                         err = write_one_cache_group(trans, root,
 243                                                     path, cache[i]);
 244                         if (err)
 245                                 werr = err;
 246                 }
 247         }
 248         btrfs_free_path(path);
 249         return werr;
 250 }
 251
 252 static int update_block_group(struct btrfs_trans_handle *trans,
 253                               struct btrfs_root *root,
 254                               u64 blocknr, u64 num, int alloc)
 255 {
 256         struct btrfs_block_group_cache *cache;
 257         struct btrfs_fs_info *info = root->fs_info;
 258         u64 total = num;
 259         u64 old_val;
 260         u64 block_in_group;
 261         int ret;
 262         while(total) {
 263                 ret = radix_tree_gang_lookup(&info->block_group_radix,
 264                                              (void **)&cache, blocknr, 1);
 265                 if (!ret) {
 266                         printk(KERN_CRIT "blocknr %Lu lookup failed\n",
 267                                blocknr);
 268                         return -1;
 269                 }
 270                 block_in_group = blocknr - cache->key.objectid;
 271                 WARN_ON(block_in_group > cache->key.offset);
 272                 radix_tree_tag_set(&info->block_group_radix,
 273                                    cache->key.objectid + cache->key.offset - 1,
 274                                    BTRFS_BLOCK_GROUP_DIRTY);
 275
 276                 old_val = btrfs_block_group_used(&cache->item);
 277                 num = min(total, cache->key.offset - block_in_group);
 278                 total -= num;
 279                 blocknr += num;
 280                 if (alloc) {
 281                         old_val += num;
 282                         if (blocknr > cache->last_alloc)
 283                                 cache->last_alloc = blocknr;
 284                 } else {
 285                         old_val -= num;
 286                         if (blocknr < cache->first_free)
 287                                 cache->first_free = blocknr;
 288                 }
 289                 btrfs_set_block_group_used(&cache->item, old_val);
 290         }
 291         return 0;
 292 }
 293
 294 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans, struct
 295                                btrfs_root *root)
 296 {
 297         unsigned long gang[8];
 298         u64 first = 0;
 299         int ret;
 300         int i;
 301         struct radix_tree_root *pinned_radix = &root->fs_info->pinned_radix;
 302
 303         while(1) {
 304                 ret = find_first_radix_bit(pinned_radix, gang,
 305                                            ARRAY_SIZE(gang));
 306                 if (!ret)
 307                         break;
 308                 if (!first)
 309                         first = gang[0];
 310                 for (i = 0; i < ret; i++) {
 311                         clear_radix_bit(pinned_radix, gang[i]);
 312                 }
 313         }
 314         root->fs_info->block_group_cache = NULL;
 315         return 0;
 316 }
 317
 318 static int finish_current_insert(struct btrfs_trans_handle *trans, struct
 319                                  btrfs_root *extent_root)
 320 {
 321         struct btrfs_key ins;
 322         struct btrfs_extent_item extent_item;
 323         int i;
 324         int ret;
 325         u64 super_blocks_used;
 326         struct btrfs_fs_info *info = extent_root->fs_info;
 327
 328         btrfs_set_extent_refs(&extent_item, 1);
 329         ins.offset = 1;
 330         ins.flags = 0;
 331         btrfs_set_key_type(&ins, BTRFS_EXTENT_ITEM_KEY);
 332         btrfs_set_extent_owner(&extent_item, extent_root->root_key.objectid);
 333
 334         for (i = 0; i < extent_root->fs_info->extent_tree_insert_nr; i++) {
 335                 ins.objectid = extent_root->fs_info->extent_tree_insert[i];
 336                 super_blocks_used = btrfs_super_blocks_used(info->disk_super);
 337                 btrfs_set_super_blocks_used(info->disk_super,
 338                                             super_blocks_used + 1);
 339                 ret = btrfs_insert_item(trans, extent_root, &ins, &extent_item,
 340                                         sizeof(extent_item));
 341                 BUG_ON(ret);
 342         }
 343         extent_root->fs_info->extent_tree_insert_nr = 0;
 344         extent_root->fs_info->extent_tree_prealloc_nr = 0;
 345         return 0;
 346 }
 347
 348 static int pin_down_block(struct btrfs_root *root, u64 blocknr, int pending)
 349 {
 350         int err;
 351         struct btrfs_header *header;
 352         struct buffer_head *bh;
 353
 354         if (!pending) {
 355                 bh = btrfs_find_tree_block(root, blocknr);
 356                 if (bh) {
 357                         if (buffer_uptodate(bh)) {
 358                                 u64 transid =
 359                                     root->fs_info->running_transaction->transid;
 360                                 header = btrfs_buffer_header(bh);
 361                                 if (btrfs_header_generation(header) ==
 362                                     transid) {
 363                                         btrfs_block_release(root, bh);
 364                                         return 0;
 365                                 }
 366                         }
 367                         btrfs_block_release(root, bh);
 368                 }
 369                 err = set_radix_bit(&root->fs_info->pinned_radix, blocknr);
 370         } else {
 371                 err = set_radix_bit(&root->fs_info->pending_del_radix, blocknr);
 372         }
 373         BUG_ON(err);
 374         return 0;
 375 }
 376
 377 /*
 378  * remove an extent from the root, returns 0 on success
 379  */
 380 static int __free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
 381                          *root, u64 blocknr, u64 num_blocks, int pin)
 382 {
 383         struct btrfs_path *path;
 384         struct btrfs_key key;
 385         struct btrfs_fs_info *info = root->fs_info;
 386         struct btrfs_root *extent_root = info->extent_root;
 387         int ret;
 388         struct btrfs_extent_item *ei;
 389         struct btrfs_key ins;
 390         u32 refs;
 391
 392         key.objectid = blocknr;
 393         key.flags = 0;
 394         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 395         key.offset = num_blocks;
 396
 397         find_free_extent(trans, root, 0, 0, (u64)-1, &ins);
 398         path = btrfs_alloc_path();
 399         BUG_ON(!path);
 400         btrfs_init_path(path);
 401
 402         ret = btrfs_search_slot(trans, extent_root, &key, path, -1, 1);
 403         if (ret) {
 404                 printk("failed to find %Lu\n", key.objectid);
 405                 btrfs_print_tree(extent_root, extent_root->node);
 406                 printk("failed to find %Lu\n", key.objectid);
 407                 BUG();
 408         }
 409         ei = btrfs_item_ptr(btrfs_buffer_leaf(path->nodes[0]), path->slots[0],
 410                             struct btrfs_extent_item);
 411         BUG_ON(ei->refs == 0);
 412         refs = btrfs_extent_refs(ei) - 1;
 413         btrfs_set_extent_refs(ei, refs);
 414         btrfs_mark_buffer_dirty(path->nodes[0]);
 415         if (refs == 0) {
 416                 u64 super_blocks_used;
 417
 418                 if (pin) {
 419                         ret = pin_down_block(root, blocknr, 0);
 420                         BUG_ON(ret);
 421                 }
 422
 423                 super_blocks_used = btrfs_super_blocks_used(info->disk_super);
 424                 btrfs_set_super_blocks_used(info->disk_super,
 425                                             super_blocks_used - num_blocks);
 426                 ret = btrfs_del_item(trans, extent_root, path);
 427                 if (ret)
 428                         BUG();
 429                 ret = update_block_group(trans, root, blocknr, num_blocks, 0);
 430                 BUG_ON(ret);
 431         }
 432         btrfs_release_path(extent_root, path);
 433         btrfs_free_path(path);
 434         finish_current_insert(trans, extent_root);
 435         return ret;
 436 }
 437
 438 /*
 439  * find all the blocks marked as pending in the radix tree and remove
 440  * them from the extent map
 441  */
 442 static int del_pending_extents(struct btrfs_trans_handle *trans, struct
 443                                btrfs_root *extent_root)
 444 {
 445         int ret;
 446         int wret;
 447         int err = 0;
 448         unsigned long gang[4];
 449         int i;
 450         struct radix_tree_root *pending_radix;
 451         struct radix_tree_root *pinned_radix;
 452
 453         pending_radix = &extent_root->fs_info->pending_del_radix;
 454         pinned_radix = &extent_root->fs_info->pinned_radix;
 455
 456         while(1) {
 457                 ret = find_first_radix_bit(pending_radix, gang,
 458                                            ARRAY_SIZE(gang));
 459                 if (!ret)
 460                         break;
 461                 for (i = 0; i < ret; i++) {
 462                         wret = set_radix_bit(pinned_radix, gang[i]);
 463                         BUG_ON(wret);
 464                         wret = clear_radix_bit(pending_radix, gang[i]);
 465                         BUG_ON(wret);
 466                         wret = __free_extent(trans, extent_root,
 467                                              gang[i], 1, 0);
 468                         if (wret)
 469                                 err = wret;
 470                 }
 471         }
 472         return err;
 473 }
 474
 475 /*
 476  * remove an extent from the root, returns 0 on success
 477  */
 478 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
 479                       *root, u64 blocknr, u64 num_blocks, int pin)
 480 {
 481         struct btrfs_root *extent_root = root->fs_info->extent_root;
 482         int pending_ret;
 483         int ret;
 484
 485         if (root == extent_root) {
 486                 pin_down_block(root, blocknr, 1);
 487                 return 0;
 488         }
 489         ret = __free_extent(trans, root, blocknr, num_blocks, pin);
 490         pending_ret = del_pending_extents(trans, root->fs_info->extent_root);
 491         return ret ? ret : pending_ret;
 492 }
 493
 494 /*
 495  * walks the btree of allocated extents and find a hole of a given size.
 496  * The key ins is changed to record the hole:
 497  * ins->objectid == block start
 498  * ins->flags = BTRFS_EXTENT_ITEM_KEY
 499  * ins->offset == number of blocks
 500  * Any available blocks before search_start are skipped.
 501  */
 502 static int find_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
 503                             *orig_root, u64 num_blocks, u64 search_start, u64
 504                             search_end, struct btrfs_key *ins)
 505 {
 506         struct btrfs_path *path;
 507         struct btrfs_key key;
 508         int ret;
 509         u64 hole_size = 0;
 510         int slot = 0;
 511         u64 last_block = 0;
 512         u64 test_block;
 513         int start_found;
 514         struct btrfs_leaf *l;
 515         struct btrfs_root * root = orig_root->fs_info->extent_root;
 516         struct btrfs_fs_info *info = root->fs_info;
 517         int total_needed = num_blocks;
 518         int total_found = 0;
 519         int fill_prealloc = 0;
 520         int level;
 521
 522         path = btrfs_alloc_path();
 523         ins->flags = 0;
 524         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
 525
 526         level = btrfs_header_level(btrfs_buffer_header(root->node));
 527         if (num_blocks == 0) {
 528                 fill_prealloc = 1;
 529                 num_blocks = 1;
 530                 total_needed = min(level + 2, BTRFS_MAX_LEVEL) * 3;
 531         }
 532         find_search_start(root, 0);
 533         if (info->block_group_cache &&
 534             info->block_group_cache->last_alloc > search_start)
 535                 search_start = info->block_group_cache->last_alloc;
 536
 537 check_failed:
 538         btrfs_init_path(path);
 539         ins->objectid = search_start;
 540         ins->offset = 0;
 541         start_found = 0;
 542         ret = btrfs_search_slot(trans, root, ins, path, 0, 0);
 543         if (ret < 0)
 544                 goto error;
 545
 546         if (path->slots[0] > 0)
 547                 path->slots[0]--;
 548
 549         while (1) {
 550                 l = btrfs_buffer_leaf(path->nodes[0]);
 551                 slot = path->slots[0];
 552                 if (slot >= btrfs_header_nritems(&l->header)) {
 553                         if (fill_prealloc) {
 554                                 info->extent_tree_prealloc_nr = 0;
 555                                 total_found = 0;
 556                         }
 557                         ret = btrfs_next_leaf(root, path);
 558                         if (ret == 0)
 559                                 continue;
 560                         if (ret < 0)
 561                                 goto error;
 562                         if (!start_found) {
 563                                 ins->objectid = search_start;
 564                                 ins->offset = (u64)-1 - search_start;
 565                                 start_found = 1;
 566                                 goto check_pending;
 567                         }
 568                         ins->objectid = last_block > search_start ?
 569                                         last_block : search_start;
 570                         ins->offset = (u64)-1 - ins->objectid;
 571                         goto check_pending;
 572                 }
 573                 btrfs_disk_key_to_cpu(&key, &l->items[slot].key);
 574                 if (btrfs_key_type(&key) != BTRFS_EXTENT_ITEM_KEY)
 575                         goto next;
 576                 if (key.objectid >= search_start) {
 577                         if (start_found) {
 578                                 if (last_block < search_start)
 579                                         last_block = search_start;
 580                                 hole_size = key.objectid - last_block;
 581                                 if (hole_size > num_blocks) {
 582                                         ins->objectid = last_block;
 583                                         ins->offset = hole_size;
 584                                         goto check_pending;
 585                                 }
 586                         }
 587                 }
 588                 start_found = 1;
 589                 last_block = key.objectid + key.offset;
 590 next:
 591                 path->slots[0]++;
 592         }
 593         // FIXME -ENOSPC
 594 check_pending:
 595         /* we have to make sure we didn't find an extent that has already
 596          * been allocated by the map tree or the original allocation
 597          */
 598         btrfs_release_path(root, path);
 599         BUG_ON(ins->objectid < search_start);
 600         for (test_block = ins->objectid;
 601              test_block < ins->objectid + num_blocks; test_block++) {
 602                 if (test_radix_bit(&info->pinned_radix, test_block)) {
 603                         search_start = test_block + 1;
 604                         goto check_failed;
 605                 }
 606         }
 607         if (!fill_prealloc && info->extent_tree_insert_nr) {
 608                 u64 last =
 609                   info->extent_tree_insert[info->extent_tree_insert_nr - 1];
 610                 if (ins->objectid + num_blocks >
 611                     info->extent_tree_insert[0] &&
 612                     ins->objectid <= last) {
 613                         search_start = last + 1;
 614                         WARN_ON(1);
 615                         goto check_failed;
 616                 }
 617         }
 618         if (!fill_prealloc && info->extent_tree_prealloc_nr) {
 619                 u64 first =
 620                   info->extent_tree_prealloc[info->extent_tree_prealloc_nr - 1];
 621                 if (ins->objectid + num_blocks > first &&
 622                     ins->objectid <= info->extent_tree_prealloc[0]) {
 623                         search_start = info->extent_tree_prealloc[0] + 1;
 624                         WARN_ON(1);
 625                         goto check_failed;
 626                 }
 627         }
 628         if (fill_prealloc) {
 629                 int nr;
 630                 test_block = ins->objectid;
 631                 while(test_block < ins->objectid + ins->offset &&
 632                       total_found < total_needed) {
 633                         nr = total_needed - total_found - 1;
 634                         BUG_ON(nr < 0);
 635                         info->extent_tree_prealloc[nr] = test_block;
 636                         total_found++;
 637                         test_block++;
 638                 }
 639                 if (total_found < total_needed) {
 640                         search_start = test_block;
 641                         goto check_failed;
 642                 }
 643                 info->extent_tree_prealloc_nr = total_found;
 644         }
 645         ret = radix_tree_gang_lookup(&info->block_group_radix,
 646                                      (void **)&info->block_group_cache,
 647                                      ins->objectid, 1);
 648         if (ret) {
 649                 info->block_group_cache->last_alloc = ins->objectid;
 650         }
 651         ins->offset = num_blocks;
 652         btrfs_free_path(path);
 653         return 0;
 654 error:
 655         btrfs_release_path(root, path);
 656         btrfs_free_path(path);
 657         return ret;
 658 }
 659 /*
 660  * finds a free extent and does all the dirty work required for allocation
 661  * returns the key for the extent through ins, and a tree buffer for
 662  * the first block of the extent through buf.
 663  *
 664  * returns 0 if everything worked, non-zero otherwise.
 665  */
 666 int btrfs_alloc_extent(struct btrfs_trans_handle *trans,
 667                        struct btrfs_root *root, u64 owner,
 668                        u64 num_blocks, u64 search_start,
 669                        u64 search_end, struct btrfs_key *ins)
 670 {
 671         int ret;
 672         int pending_ret;
 673         u64 super_blocks_used;
 674         struct btrfs_fs_info *info = root->fs_info;
 675         struct btrfs_root *extent_root = info->extent_root;
 676         struct btrfs_extent_item extent_item;
 677         struct btrfs_key prealloc_key;
 678
 679         btrfs_set_extent_refs(&extent_item, 1);
 680         btrfs_set_extent_owner(&extent_item, owner);
 681
 682         if (root == extent_root) {
 683                 int nr;
 684                 BUG_ON(info->extent_tree_prealloc_nr == 0);
 685                 BUG_ON(num_blocks != 1);
 686                 ins->offset = 1;
 687                 info->extent_tree_prealloc_nr--;
 688                 nr = info->extent_tree_prealloc_nr;
 689                 ins->objectid = info->extent_tree_prealloc[nr];
 690                 info->extent_tree_insert[info->extent_tree_insert_nr++] =
 691                         ins->objectid;
 692                 ret = update_block_group(trans, root,
 693                                          ins->objectid, ins->offset, 1);
 694                 BUG_ON(ret);
 695                 return 0;
 696         }
 697         /* do the real allocation */
 698         ret = find_free_extent(trans, root, num_blocks, search_start,
 699                                search_end, ins);
 700         if (ret)
 701                 return ret;
 702
 703         /* then do prealloc for the extent tree */
 704         ret = find_free_extent(trans, root, 0, ins->objectid + ins->offset,
 705                                search_end, &prealloc_key);
 706         if (ret)
 707                 return ret;
 708
 709         super_blocks_used = btrfs_super_blocks_used(info->disk_super);
 710         btrfs_set_super_blocks_used(info->disk_super, super_blocks_used +
 711                                     num_blocks);
 712         ret = btrfs_insert_item(trans, extent_root, ins, &extent_item,
 713                                 sizeof(extent_item));
 714
 715         finish_current_insert(trans, extent_root);
 716         pending_ret = del_pending_extents(trans, extent_root);
 717         if (ret)
 718                 return ret;
 719         if (pending_ret)
 720                 return pending_ret;
 721         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
 722         return 0;
 723 }
 724
 725 /*
 726  * helper function to allocate a block for a given tree
 727  * returns the tree buffer or NULL.
 728  */
 729 struct buffer_head *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
 730                                            struct btrfs_root *root)
 731 {
 732         struct btrfs_key ins;
 733         int ret;
 734         struct buffer_head *buf;
 735
 736         ret = btrfs_alloc_extent(trans, root, root->root_key.objectid,
 737                                  1, 0, (unsigned long)-1, &ins);
 738         if (ret) {
 739                 BUG();
 740                 return NULL;
 741         }
 742         BUG_ON(ret);
 743         buf = btrfs_find_create_tree_block(root, ins.objectid);
 744         set_buffer_uptodate(buf);
 745         return buf;
 746 }
 747
 748 static int drop_leaf_ref(struct btrfs_trans_handle *trans,
 749                          struct btrfs_root *root, struct buffer_head *cur)
 750 {
 751         struct btrfs_disk_key *key;
 752         struct btrfs_leaf *leaf;
 753         struct btrfs_file_extent_item *fi;
 754         int i;
 755         int nritems;
 756         int ret;
 757
 758         BUG_ON(!btrfs_is_leaf(btrfs_buffer_node(cur)));
 759         leaf = btrfs_buffer_leaf(cur);
 760         nritems = btrfs_header_nritems(&leaf->header);
 761         for (i = 0; i < nritems; i++) {
 762                 key = &leaf->items[i].key;
 763                 if (btrfs_disk_key_type(key) != BTRFS_EXTENT_DATA_KEY)
 764                         continue;
 765                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
 766                 if (btrfs_file_extent_type(fi) == BTRFS_FILE_EXTENT_INLINE)
 767                         continue;
 768                 /*
 769                  * FIXME make sure to insert a trans record that
 770                  * repeats the snapshot del on crash
 771                  */
 772                 ret = btrfs_free_extent(trans, root,
 773                                         btrfs_file_extent_disk_blocknr(fi),
 774                                         btrfs_file_extent_disk_num_blocks(fi),
 775                                         0);
 776                 BUG_ON(ret);
 777         }
 778         return 0;
 779 }
 780
 781 /*
 782  * helper function for drop_snapshot, this walks down the tree dropping ref
 783  * counts as it goes.
 784  */
 785 static int walk_down_tree(struct btrfs_trans_handle *trans, struct btrfs_root
 786                           *root, struct btrfs_path *path, int *level)
 787 {
 788         struct buffer_head *next;
 789         struct buffer_head *cur;
 790         u64 blocknr;
 791         int ret;
 792         u32 refs;
 793
 794         WARN_ON(*level < 0);
 795         WARN_ON(*level >= BTRFS_MAX_LEVEL);
 796         ret = lookup_extent_ref(trans, root, bh_blocknr(path->nodes[*level]),
 797                                1, &refs);
 798         BUG_ON(ret);
 799         if (refs > 1)
 800                 goto out;
 801         /*
 802          * walk down to the last node level and free all the leaves
 803          */
 804         while(*level >= 0) {
 805                 WARN_ON(*level < 0);
 806                 WARN_ON(*level >= BTRFS_MAX_LEVEL);
 807                 cur = path->nodes[*level];
 808                 if (btrfs_header_level(btrfs_buffer_header(cur)) != *level)
 809                         WARN_ON(1);
 810                 if (path->slots[*level] >=
 811                     btrfs_header_nritems(btrfs_buffer_header(cur)))
 812                         break;
 813                 if (*level == 0) {
 814                         ret = drop_leaf_ref(trans, root, cur);
 815                         BUG_ON(ret);
 816                         break;
 817                 }
 818                 blocknr = btrfs_node_blockptr(btrfs_buffer_node(cur),
 819                                               path->slots[*level]);
 820                 ret = lookup_extent_ref(trans, root, blocknr, 1, &refs);
 821                 BUG_ON(ret);
 822                 if (refs != 1) {
 823                         path->slots[*level]++;
 824                         ret = btrfs_free_extent(trans, root, blocknr, 1, 1);
 825                         BUG_ON(ret);
 826                         continue;
 827                 }
 828                 next = read_tree_block(root, blocknr);
 829                 WARN_ON(*level <= 0);
 830                 if (path->nodes[*level-1])
 831                         btrfs_block_release(root, path->nodes[*level-1]);
 832                 path->nodes[*level-1] = next;
 833                 *level = btrfs_header_level(btrfs_buffer_header(next));
 834                 path->slots[*level] = 0;
 835         }
 836 out:
 837         WARN_ON(*level < 0);
 838         WARN_ON(*level >= BTRFS_MAX_LEVEL);
 839         ret = btrfs_free_extent(trans, root,
 840                                 bh_blocknr(path->nodes[*level]), 1, 1);
 841         btrfs_block_release(root, path->nodes[*level]);
 842         path->nodes[*level] = NULL;
 843         *level += 1;
 844         BUG_ON(ret);
 845         return 0;
 846 }
 847
 848 /*
 849  * helper for dropping snapshots.  This walks back up the tree in the path
 850  * to find the first node higher up where we haven't yet gone through
 851  * all the slots
 852  */
 853 static int walk_up_tree(struct btrfs_trans_handle *trans, struct btrfs_root
 854                         *root, struct btrfs_path *path, int *level)
 855 {
 856         int i;
 857         int slot;
 858         int ret;
 859         for(i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
 860                 slot = path->slots[i];
 861                 if (slot < btrfs_header_nritems(
 862                     btrfs_buffer_header(path->nodes[i])) - 1) {
 863                         path->slots[i]++;
 864                         *level = i;
 865                         return 0;
 866                 } else {
 867                         ret = btrfs_free_extent(trans, root,
 868                                                 bh_blocknr(path->nodes[*level]),
 869                                                 1, 1);
 870                         BUG_ON(ret);
 871                         btrfs_block_release(root, path->nodes[*level]);
 872                         path->nodes[*level] = NULL;
 873                         *level = i + 1;
 874                 }
 875         }
 876         return 1;
 877 }
 878
 879 /*
 880  * drop the reference count on the tree rooted at 'snap'.  This traverses
 881  * the tree freeing any blocks that have a ref count of zero after being
 882  * decremented.
 883  */
 884 int btrfs_drop_snapshot(struct btrfs_trans_handle *trans, struct btrfs_root
 885                         *root, struct buffer_head *snap)
 886 {
 887         int ret = 0;
 888         int wret;
 889         int level;
 890         struct btrfs_path *path;
 891         int i;
 892         int orig_level;
 893
 894         path = btrfs_alloc_path();
 895         BUG_ON(!path);
 896         btrfs_init_path(path);
 897
 898         level = btrfs_header_level(btrfs_buffer_header(snap));
 899         orig_level = level;
 900         path->nodes[level] = snap;
 901         path->slots[level] = 0;
 902         while(1) {
 903                 wret = walk_down_tree(trans, root, path, &level);
 904                 if (wret > 0)
 905                         break;
 906                 if (wret < 0)
 907                         ret = wret;
 908
 909                 wret = walk_up_tree(trans, root, path, &level);
 910                 if (wret > 0)
 911                         break;
 912                 if (wret < 0)
 913                         ret = wret;
 914         }
 915         for (i = 0; i <= orig_level; i++) {
 916                 if (path->nodes[i]) {
 917                         btrfs_block_release(root, path->nodes[i]);
 918                 }
 919         }
 920         btrfs_free_path(path);
 921         return ret;
 922 }
 923
 924 int btrfs_free_block_groups(struct btrfs_fs_info *info)
 925 {
 926         int ret;
 927         struct btrfs_block_group_cache *cache[8];
 928         int i;
 929
 930         while(1) {
 931                 ret = radix_tree_gang_lookup(&info->block_group_radix,
 932                                              (void **)cache, 0,
 933                                              ARRAY_SIZE(cache));
 934                 if (!ret)
 935                         break;
 936                 for (i = 0; i < ret; i++) {
 937                         radix_tree_delete(&info->block_group_radix,
 938                                           cache[i]->key.objectid +
 939                                           cache[i]->key.offset - 1);
 940                         kfree(cache[i]);
 941                 }
 942         }
 943         return 0;
 944 }
 945
 946 int btrfs_read_block_groups(struct btrfs_root *root)
 947 {
 948         struct btrfs_path *path;
 949         int ret;
 950         int err = 0;
 951         struct btrfs_block_group_item *bi;
 952         struct btrfs_block_group_cache *cache;
 953         struct btrfs_key key;
 954         struct btrfs_key found_key;
 955         struct btrfs_leaf *leaf;
 956         u64 group_size_blocks = BTRFS_BLOCK_GROUP_SIZE / root->blocksize;
 957
 958         root = root->fs_info->extent_root;
 959         key.objectid = 0;
 960         key.offset = group_size_blocks;
 961         key.flags = 0;
 962         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
 963
 964         path = btrfs_alloc_path();
 965         if (!path)
 966                 return -ENOMEM;
 967
 968         while(1) {
 969                 ret = btrfs_search_slot(NULL, root->fs_info->extent_root,
 970                                         &key, path, 0, 0);
 971                 if (ret != 0) {
 972                         err = ret;
 973                         break;
 974                 }
 975                 leaf = btrfs_buffer_leaf(path->nodes[0]);
 976                 btrfs_disk_key_to_cpu(&found_key,
 977                                       &leaf->items[path->slots[0]].key);
 978                 cache = kmalloc(sizeof(*cache), GFP_NOFS);
 979                 if (!cache) {
 980                         err = -1;
 981                         break;
 982                 }
 983                 bi = btrfs_item_ptr(leaf, path->slots[0],
 984                                     struct btrfs_block_group_item);
 985                 memcpy(&cache->item, bi, sizeof(*bi));
 986                 memcpy(&cache->key, &found_key, sizeof(found_key));
 987                 cache->last_alloc = 0;
 988                 cache->first_free = 0;
 989                 key.objectid = found_key.objectid + found_key.offset;
 990                 btrfs_release_path(root, path);
 991                 ret = radix_tree_insert(&root->fs_info->block_group_radix,
 992                                         found_key.objectid +
 993                                         found_key.offset - 1,
 994                                         (void *)cache);
 995                 BUG_ON(ret);
 996                 if (key.objectid >=
 997                     btrfs_super_total_blocks(root->fs_info->disk_super))
 998                         break;
 999         }
1000
1001         btrfs_free_path(path);
1002         return 0;
1003 }