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         if (root->fs_info->block_group_cache) {
 315                 root->fs_info->block_group_cache->last_alloc =
 316                         root->fs_info->block_group_cache->first_free;
 317         }
 318         return 0;
 319 }
 320
 321 static int finish_current_insert(struct btrfs_trans_handle *trans, struct
 322                                  btrfs_root *extent_root)
 323 {
 324         struct btrfs_key ins;
 325         struct btrfs_extent_item extent_item;
 326         int i;
 327         int ret;
 328         u64 super_blocks_used;
 329         struct btrfs_fs_info *info = extent_root->fs_info;
 330
 331         btrfs_set_extent_refs(&extent_item, 1);
 332         ins.offset = 1;
 333         ins.flags = 0;
 334         btrfs_set_key_type(&ins, BTRFS_EXTENT_ITEM_KEY);
 335         btrfs_set_extent_owner(&extent_item, extent_root->root_key.objectid);
 336
 337         for (i = 0; i < extent_root->fs_info->extent_tree_insert_nr; i++) {
 338                 ins.objectid = extent_root->fs_info->extent_tree_insert[i];
 339                 super_blocks_used = btrfs_super_blocks_used(info->disk_super);
 340                 btrfs_set_super_blocks_used(info->disk_super,
 341                                             super_blocks_used + 1);
 342                 ret = btrfs_insert_item(trans, extent_root, &ins, &extent_item,
 343                                         sizeof(extent_item));
 344                 BUG_ON(ret);
 345         }
 346         extent_root->fs_info->extent_tree_insert_nr = 0;
 347         extent_root->fs_info->extent_tree_prealloc_nr = 0;
 348         return 0;
 349 }
 350
 351 static int pin_down_block(struct btrfs_root *root, u64 blocknr, int pending)
 352 {
 353         int err;
 354         struct btrfs_header *header;
 355         struct buffer_head *bh;
 356
 357         if (!pending) {
 358                 bh = btrfs_find_tree_block(root, blocknr);
 359                 if (bh) {
 360                         if (buffer_uptodate(bh)) {
 361                                 u64 transid =
 362                                     root->fs_info->running_transaction->transid;
 363                                 header = btrfs_buffer_header(bh);
 364                                 if (btrfs_header_generation(header) ==
 365                                     transid) {
 366                                         btrfs_block_release(root, bh);
 367                                         return 0;
 368                                 }
 369                         }
 370                         btrfs_block_release(root, bh);
 371                 }
 372                 err = set_radix_bit(&root->fs_info->pinned_radix, blocknr);
 373         } else {
 374                 err = set_radix_bit(&root->fs_info->pending_del_radix, blocknr);
 375         }
 376         BUG_ON(err);
 377         return 0;
 378 }
 379
 380 /*
 381  * remove an extent from the root, returns 0 on success
 382  */
 383 static int __free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
 384                          *root, u64 blocknr, u64 num_blocks, int pin)
 385 {
 386         struct btrfs_path *path;
 387         struct btrfs_key key;
 388         struct btrfs_fs_info *info = root->fs_info;
 389         struct btrfs_root *extent_root = info->extent_root;
 390         int ret;
 391         struct btrfs_extent_item *ei;
 392         struct btrfs_key ins;
 393         u32 refs;
 394
 395         key.objectid = blocknr;
 396         key.flags = 0;
 397         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 398         key.offset = num_blocks;
 399
 400         find_free_extent(trans, root, 0, 0, (u64)-1, &ins);
 401         path = btrfs_alloc_path();
 402         BUG_ON(!path);
 403         btrfs_init_path(path);
 404
 405         ret = btrfs_search_slot(trans, extent_root, &key, path, -1, 1);
 406         if (ret) {
 407                 printk("failed to find %Lu\n", key.objectid);
 408                 btrfs_print_tree(extent_root, extent_root->node);
 409                 printk("failed to find %Lu\n", key.objectid);
 410                 BUG();
 411         }
 412         ei = btrfs_item_ptr(btrfs_buffer_leaf(path->nodes[0]), path->slots[0],
 413                             struct btrfs_extent_item);
 414         BUG_ON(ei->refs == 0);
 415         refs = btrfs_extent_refs(ei) - 1;
 416         btrfs_set_extent_refs(ei, refs);
 417         btrfs_mark_buffer_dirty(path->nodes[0]);
 418         if (refs == 0) {
 419                 u64 super_blocks_used;
 420
 421                 if (pin) {
 422                         ret = pin_down_block(root, blocknr, 0);
 423                         BUG_ON(ret);
 424                 }
 425
 426                 super_blocks_used = btrfs_super_blocks_used(info->disk_super);
 427                 btrfs_set_super_blocks_used(info->disk_super,
 428                                             super_blocks_used - num_blocks);
 429                 ret = btrfs_del_item(trans, extent_root, path);
 430                 if (ret)
 431                         BUG();
 432                 ret = update_block_group(trans, root, blocknr, num_blocks, 0);
 433                 BUG_ON(ret);
 434         }
 435         btrfs_release_path(extent_root, path);
 436         btrfs_free_path(path);
 437         finish_current_insert(trans, extent_root);
 438         return ret;
 439 }
 440
 441 /*
 442  * find all the blocks marked as pending in the radix tree and remove
 443  * them from the extent map
 444  */
 445 static int del_pending_extents(struct btrfs_trans_handle *trans, struct
 446                                btrfs_root *extent_root)
 447 {
 448         int ret;
 449         int wret;
 450         int err = 0;
 451         unsigned long gang[4];
 452         int i;
 453         struct radix_tree_root *pending_radix;
 454         struct radix_tree_root *pinned_radix;
 455
 456         pending_radix = &extent_root->fs_info->pending_del_radix;
 457         pinned_radix = &extent_root->fs_info->pinned_radix;
 458
 459         while(1) {
 460                 ret = find_first_radix_bit(pending_radix, gang,
 461                                            ARRAY_SIZE(gang));
 462                 if (!ret)
 463                         break;
 464                 for (i = 0; i < ret; i++) {
 465                         wret = set_radix_bit(pinned_radix, gang[i]);
 466                         BUG_ON(wret);
 467                         wret = clear_radix_bit(pending_radix, gang[i]);
 468                         BUG_ON(wret);
 469                         wret = __free_extent(trans, extent_root,
 470                                              gang[i], 1, 0);
 471                         if (wret)
 472                                 err = wret;
 473                 }
 474         }
 475         return err;
 476 }
 477
 478 /*
 479  * remove an extent from the root, returns 0 on success
 480  */
 481 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
 482                       *root, u64 blocknr, u64 num_blocks, int pin)
 483 {
 484         struct btrfs_root *extent_root = root->fs_info->extent_root;
 485         int pending_ret;
 486         int ret;
 487
 488         if (root == extent_root) {
 489                 pin_down_block(root, blocknr, 1);
 490                 return 0;
 491         }
 492         ret = __free_extent(trans, root, blocknr, num_blocks, pin);
 493         pending_ret = del_pending_extents(trans, root->fs_info->extent_root);
 494         return ret ? ret : pending_ret;
 495 }
 496
 497 /*
 498  * walks the btree of allocated extents and find a hole of a given size.
 499  * The key ins is changed to record the hole:
 500  * ins->objectid == block start
 501  * ins->flags = BTRFS_EXTENT_ITEM_KEY
 502  * ins->offset == number of blocks
 503  * Any available blocks before search_start are skipped.
 504  */
 505 static int find_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
 506                             *orig_root, u64 num_blocks, u64 search_start, u64
 507                             search_end, struct btrfs_key *ins)
 508 {
 509         struct btrfs_path *path;
 510         struct btrfs_key key;
 511         int ret;
 512         u64 hole_size = 0;
 513         int slot = 0;
 514         u64 last_block = 0;
 515         u64 test_block;
 516         int start_found;
 517         struct btrfs_leaf *l;
 518         struct btrfs_root * root = orig_root->fs_info->extent_root;
 519         struct btrfs_fs_info *info = root->fs_info;
 520         int total_needed = num_blocks;
 521         int total_found = 0;
 522         int fill_prealloc = 0;
 523         int level;
 524
 525         path = btrfs_alloc_path();
 526         ins->flags = 0;
 527         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
 528
 529         level = btrfs_header_level(btrfs_buffer_header(root->node));
 530         if (num_blocks == 0) {
 531                 fill_prealloc = 1;
 532                 num_blocks = 1;
 533                 total_needed = min(level + 2, BTRFS_MAX_LEVEL) * 3;
 534         }
 535         find_search_start(root, 0);
 536         if (info->block_group_cache &&
 537             info->block_group_cache->last_alloc > search_start)
 538                 search_start = info->block_group_cache->last_alloc;
 539
 540 check_failed:
 541         btrfs_init_path(path);
 542         ins->objectid = search_start;
 543         ins->offset = 0;
 544         start_found = 0;
 545         ret = btrfs_search_slot(trans, root, ins, path, 0, 0);
 546         if (ret < 0)
 547                 goto error;
 548
 549         if (path->slots[0] > 0)
 550                 path->slots[0]--;
 551
 552         while (1) {
 553                 l = btrfs_buffer_leaf(path->nodes[0]);
 554                 slot = path->slots[0];
 555                 if (slot >= btrfs_header_nritems(&l->header)) {
 556                         if (fill_prealloc) {
 557                                 info->extent_tree_prealloc_nr = 0;
 558                                 total_found = 0;
 559                         }
 560                         ret = btrfs_next_leaf(root, path);
 561                         if (ret == 0)
 562                                 continue;
 563                         if (ret < 0)
 564                                 goto error;
 565                         if (!start_found) {
 566                                 ins->objectid = search_start;
 567                                 ins->offset = (u64)-1 - search_start;
 568                                 start_found = 1;
 569                                 goto check_pending;
 570                         }
 571                         ins->objectid = last_block > search_start ?
 572                                         last_block : search_start;
 573                         ins->offset = (u64)-1 - ins->objectid;
 574                         goto check_pending;
 575                 }
 576                 btrfs_disk_key_to_cpu(&key, &l->items[slot].key);
 577                 if (btrfs_key_type(&key) != BTRFS_EXTENT_ITEM_KEY)
 578                         goto next;
 579                 if (key.objectid >= search_start) {
 580                         if (start_found) {
 581                                 if (last_block < search_start)
 582                                         last_block = search_start;
 583                                 hole_size = key.objectid - last_block;
 584                                 if (hole_size >= num_blocks) {
 585                                         ins->objectid = last_block;
 586                                         ins->offset = hole_size;
 587                                         goto check_pending;
 588                                 }
 589                         }
 590                 }
 591                 start_found = 1;
 592                 last_block = key.objectid + key.offset;
 593 next:
 594                 path->slots[0]++;
 595         }
 596         // FIXME -ENOSPC
 597 check_pending:
 598         /* we have to make sure we didn't find an extent that has already
 599          * been allocated by the map tree or the original allocation
 600          */
 601         btrfs_release_path(root, path);
 602         BUG_ON(ins->objectid < search_start);
 603         for (test_block = ins->objectid;
 604              test_block < ins->objectid + num_blocks; test_block++) {
 605                 if (test_radix_bit(&info->pinned_radix, test_block)) {
 606                         search_start = test_block + 1;
 607                         goto check_failed;
 608                 }
 609         }
 610         if (!fill_prealloc && info->extent_tree_insert_nr) {
 611                 u64 last =
 612                   info->extent_tree_insert[info->extent_tree_insert_nr - 1];
 613                 if (ins->objectid + num_blocks >
 614                     info->extent_tree_insert[0] &&
 615                     ins->objectid <= last) {
 616                         search_start = last + 1;
 617                         WARN_ON(1);
 618                         goto check_failed;
 619                 }
 620         }
 621         if (!fill_prealloc && info->extent_tree_prealloc_nr) {
 622                 u64 first =
 623                   info->extent_tree_prealloc[info->extent_tree_prealloc_nr - 1];
 624                 if (ins->objectid + num_blocks > first &&
 625                     ins->objectid <= info->extent_tree_prealloc[0]) {
 626                         search_start = info->extent_tree_prealloc[0] + 1;
 627                         WARN_ON(1);
 628                         goto check_failed;
 629                 }
 630         }
 631         if (fill_prealloc) {
 632                 int nr;
 633                 test_block = ins->objectid;
 634                 while(test_block < ins->objectid + ins->offset &&
 635                       total_found < total_needed) {
 636                         nr = total_needed - total_found - 1;
 637                         BUG_ON(nr < 0);
 638                         info->extent_tree_prealloc[nr] = test_block;
 639                         total_found++;
 640                         test_block++;
 641                 }
 642                 if (total_found < total_needed) {
 643                         search_start = test_block;
 644                         goto check_failed;
 645                 }
 646                 info->extent_tree_prealloc_nr = total_found;
 647         }
 648         ret = radix_tree_gang_lookup(&info->block_group_radix,
 649                                      (void **)&info->block_group_cache,
 650                                      ins->objectid, 1);
 651         if (ret) {
 652                 info->block_group_cache->last_alloc = ins->objectid;
 653         }
 654         ins->offset = num_blocks;
 655         btrfs_free_path(path);
 656         return 0;
 657 error:
 658         btrfs_release_path(root, path);
 659         btrfs_free_path(path);
 660         return ret;
 661 }
 662 /*
 663  * finds a free extent and does all the dirty work required for allocation
 664  * returns the key for the extent through ins, and a tree buffer for
 665  * the first block of the extent through buf.
 666  *
 667  * returns 0 if everything worked, non-zero otherwise.
 668  */
 669 int btrfs_alloc_extent(struct btrfs_trans_handle *trans,
 670                        struct btrfs_root *root, u64 owner,
 671                        u64 num_blocks, u64 search_start,
 672                        u64 search_end, struct btrfs_key *ins)
 673 {
 674         int ret;
 675         int pending_ret;
 676         u64 super_blocks_used;
 677         struct btrfs_fs_info *info = root->fs_info;
 678         struct btrfs_root *extent_root = info->extent_root;
 679         struct btrfs_extent_item extent_item;
 680         struct btrfs_key prealloc_key;
 681
 682         btrfs_set_extent_refs(&extent_item, 1);
 683         btrfs_set_extent_owner(&extent_item, owner);
 684
 685         if (root == extent_root) {
 686                 int nr;
 687                 BUG_ON(info->extent_tree_prealloc_nr == 0);
 688                 BUG_ON(num_blocks != 1);
 689                 ins->offset = 1;
 690                 info->extent_tree_prealloc_nr--;
 691                 nr = info->extent_tree_prealloc_nr;
 692                 ins->objectid = info->extent_tree_prealloc[nr];
 693                 info->extent_tree_insert[info->extent_tree_insert_nr++] =
 694                         ins->objectid;
 695                 ret = update_block_group(trans, root,
 696                                          ins->objectid, ins->offset, 1);
 697                 BUG_ON(ret);
 698                 return 0;
 699         }
 700         /* do the real allocation */
 701         ret = find_free_extent(trans, root, num_blocks, search_start,
 702                                search_end, ins);
 703         if (ret)
 704                 return ret;
 705
 706         /* then do prealloc for the extent tree */
 707         ret = find_free_extent(trans, root, 0, ins->objectid + ins->offset,
 708                                search_end, &prealloc_key);
 709         if (ret)
 710                 return ret;
 711
 712         super_blocks_used = btrfs_super_blocks_used(info->disk_super);
 713         btrfs_set_super_blocks_used(info->disk_super, super_blocks_used +
 714                                     num_blocks);
 715         ret = btrfs_insert_item(trans, extent_root, ins, &extent_item,
 716                                 sizeof(extent_item));
 717
 718         finish_current_insert(trans, extent_root);
 719         pending_ret = del_pending_extents(trans, extent_root);
 720         if (ret)
 721                 return ret;
 722         if (pending_ret)
 723                 return pending_ret;
 724         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
 725         return 0;
 726 }
 727
 728 /*
 729  * helper function to allocate a block for a given tree
 730  * returns the tree buffer or NULL.
 731  */
 732 struct buffer_head *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
 733                                            struct btrfs_root *root)
 734 {
 735         struct btrfs_key ins;
 736         int ret;
 737         struct buffer_head *buf;
 738
 739         ret = btrfs_alloc_extent(trans, root, root->root_key.objectid,
 740                                  1, 0, (unsigned long)-1, &ins);
 741         if (ret) {
 742                 BUG();
 743                 return NULL;
 744         }
 745         BUG_ON(ret);
 746         buf = btrfs_find_create_tree_block(root, ins.objectid);
 747         set_buffer_uptodate(buf);
 748         return buf;
 749 }
 750
 751 static int drop_leaf_ref(struct btrfs_trans_handle *trans,
 752                          struct btrfs_root *root, struct buffer_head *cur)
 753 {
 754         struct btrfs_disk_key *key;
 755         struct btrfs_leaf *leaf;
 756         struct btrfs_file_extent_item *fi;
 757         int i;
 758         int nritems;
 759         int ret;
 760
 761         BUG_ON(!btrfs_is_leaf(btrfs_buffer_node(cur)));
 762         leaf = btrfs_buffer_leaf(cur);
 763         nritems = btrfs_header_nritems(&leaf->header);
 764         for (i = 0; i < nritems; i++) {
 765                 key = &leaf->items[i].key;
 766                 if (btrfs_disk_key_type(key) != BTRFS_EXTENT_DATA_KEY)
 767                         continue;
 768                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
 769                 if (btrfs_file_extent_type(fi) == BTRFS_FILE_EXTENT_INLINE)
 770                         continue;
 771                 /*
 772                  * FIXME make sure to insert a trans record that
 773                  * repeats the snapshot del on crash
 774                  */
 775                 ret = btrfs_free_extent(trans, root,
 776                                         btrfs_file_extent_disk_blocknr(fi),
 777                                         btrfs_file_extent_disk_num_blocks(fi),
 778                                         0);
 779                 BUG_ON(ret);
 780         }
 781         return 0;
 782 }
 783
 784 /*
 785  * helper function for drop_snapshot, this walks down the tree dropping ref
 786  * counts as it goes.
 787  */
 788 static int walk_down_tree(struct btrfs_trans_handle *trans, struct btrfs_root
 789                           *root, struct btrfs_path *path, int *level)
 790 {
 791         struct buffer_head *next;
 792         struct buffer_head *cur;
 793         u64 blocknr;
 794         int ret;
 795         u32 refs;
 796
 797         WARN_ON(*level < 0);
 798         WARN_ON(*level >= BTRFS_MAX_LEVEL);
 799         ret = lookup_extent_ref(trans, root, bh_blocknr(path->nodes[*level]),
 800                                1, &refs);
 801         BUG_ON(ret);
 802         if (refs > 1)
 803                 goto out;
 804         /*
 805          * walk down to the last node level and free all the leaves
 806          */
 807         while(*level >= 0) {
 808                 WARN_ON(*level < 0);
 809                 WARN_ON(*level >= BTRFS_MAX_LEVEL);
 810                 cur = path->nodes[*level];
 811                 if (btrfs_header_level(btrfs_buffer_header(cur)) != *level)
 812                         WARN_ON(1);
 813                 if (path->slots[*level] >=
 814                     btrfs_header_nritems(btrfs_buffer_header(cur)))
 815                         break;
 816                 if (*level == 0) {
 817                         ret = drop_leaf_ref(trans, root, cur);
 818                         BUG_ON(ret);
 819                         break;
 820                 }
 821                 blocknr = btrfs_node_blockptr(btrfs_buffer_node(cur),
 822                                               path->slots[*level]);
 823                 ret = lookup_extent_ref(trans, root, blocknr, 1, &refs);
 824                 BUG_ON(ret);
 825                 if (refs != 1) {
 826                         path->slots[*level]++;
 827                         ret = btrfs_free_extent(trans, root, blocknr, 1, 1);
 828                         BUG_ON(ret);
 829                         continue;
 830                 }
 831                 next = read_tree_block(root, blocknr);
 832                 WARN_ON(*level <= 0);
 833                 if (path->nodes[*level-1])
 834                         btrfs_block_release(root, path->nodes[*level-1]);
 835                 path->nodes[*level-1] = next;
 836                 *level = btrfs_header_level(btrfs_buffer_header(next));
 837                 path->slots[*level] = 0;
 838         }
 839 out:
 840         WARN_ON(*level < 0);
 841         WARN_ON(*level >= BTRFS_MAX_LEVEL);
 842         ret = btrfs_free_extent(trans, root,
 843                                 bh_blocknr(path->nodes[*level]), 1, 1);
 844         btrfs_block_release(root, path->nodes[*level]);
 845         path->nodes[*level] = NULL;
 846         *level += 1;
 847         BUG_ON(ret);
 848         return 0;
 849 }
 850
 851 /*
 852  * helper for dropping snapshots.  This walks back up the tree in the path
 853  * to find the first node higher up where we haven't yet gone through
 854  * all the slots
 855  */
 856 static int walk_up_tree(struct btrfs_trans_handle *trans, struct btrfs_root
 857                         *root, struct btrfs_path *path, int *level)
 858 {
 859         int i;
 860         int slot;
 861         int ret;
 862         for(i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
 863                 slot = path->slots[i];
 864                 if (slot < btrfs_header_nritems(
 865                     btrfs_buffer_header(path->nodes[i])) - 1) {
 866                         path->slots[i]++;
 867                         *level = i;
 868                         return 0;
 869                 } else {
 870                         ret = btrfs_free_extent(trans, root,
 871                                                 bh_blocknr(path->nodes[*level]),
 872                                                 1, 1);
 873                         BUG_ON(ret);
 874                         btrfs_block_release(root, path->nodes[*level]);
 875                         path->nodes[*level] = NULL;
 876                         *level = i + 1;
 877                 }
 878         }
 879         return 1;
 880 }
 881
 882 /*
 883  * drop the reference count on the tree rooted at 'snap'.  This traverses
 884  * the tree freeing any blocks that have a ref count of zero after being
 885  * decremented.
 886  */
 887 int btrfs_drop_snapshot(struct btrfs_trans_handle *trans, struct btrfs_root
 888                         *root, struct buffer_head *snap)
 889 {
 890         int ret = 0;
 891         int wret;
 892         int level;
 893         struct btrfs_path *path;
 894         int i;
 895         int orig_level;
 896
 897         path = btrfs_alloc_path();
 898         BUG_ON(!path);
 899         btrfs_init_path(path);
 900
 901         level = btrfs_header_level(btrfs_buffer_header(snap));
 902         orig_level = level;
 903         path->nodes[level] = snap;
 904         path->slots[level] = 0;
 905         while(1) {
 906                 wret = walk_down_tree(trans, root, path, &level);
 907                 if (wret > 0)
 908                         break;
 909                 if (wret < 0)
 910                         ret = wret;
 911
 912                 wret = walk_up_tree(trans, root, path, &level);
 913                 if (wret > 0)
 914                         break;
 915                 if (wret < 0)
 916                         ret = wret;
 917         }
 918         for (i = 0; i <= orig_level; i++) {
 919                 if (path->nodes[i]) {
 920                         btrfs_block_release(root, path->nodes[i]);
 921                 }
 922         }
 923         btrfs_free_path(path);
 924         return ret;
 925 }
 926
 927 int btrfs_free_block_groups(struct btrfs_fs_info *info)
 928 {
 929         int ret;
 930         struct btrfs_block_group_cache *cache[8];
 931         int i;
 932
 933         while(1) {
 934                 ret = radix_tree_gang_lookup(&info->block_group_radix,
 935                                              (void **)cache, 0,
 936                                              ARRAY_SIZE(cache));
 937                 if (!ret)
 938                         break;
 939                 for (i = 0; i < ret; i++) {
 940                         radix_tree_delete(&info->block_group_radix,
 941                                           cache[i]->key.objectid +
 942                                           cache[i]->key.offset - 1);
 943                         kfree(cache[i]);
 944                 }
 945         }
 946         return 0;
 947 }
 948
 949 int btrfs_read_block_groups(struct btrfs_root *root)
 950 {
 951         struct btrfs_path *path;
 952         int ret;
 953         int err = 0;
 954         struct btrfs_block_group_item *bi;
 955         struct btrfs_block_group_cache *cache;
 956         struct btrfs_key key;
 957         struct btrfs_key found_key;
 958         struct btrfs_leaf *leaf;
 959         u64 group_size_blocks = BTRFS_BLOCK_GROUP_SIZE / root->blocksize;
 960
 961         root = root->fs_info->extent_root;
 962         key.objectid = 0;
 963         key.offset = group_size_blocks;
 964         key.flags = 0;
 965         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
 966
 967         path = btrfs_alloc_path();
 968         if (!path)
 969                 return -ENOMEM;
 970
 971         while(1) {
 972                 ret = btrfs_search_slot(NULL, root->fs_info->extent_root,
 973                                         &key, path, 0, 0);
 974                 if (ret != 0) {
 975                         err = ret;
 976                         break;
 977                 }
 978                 leaf = btrfs_buffer_leaf(path->nodes[0]);
 979                 btrfs_disk_key_to_cpu(&found_key,
 980                                       &leaf->items[path->slots[0]].key);
 981                 cache = kmalloc(sizeof(*cache), GFP_NOFS);
 982                 if (!cache) {
 983                         err = -1;
 984                         break;
 985                 }
 986                 bi = btrfs_item_ptr(leaf, path->slots[0],
 987                                     struct btrfs_block_group_item);
 988                 memcpy(&cache->item, bi, sizeof(*bi));
 989                 memcpy(&cache->key, &found_key, sizeof(found_key));
 990                 cache->last_alloc = 0;
 991                 cache->first_free = 0;
 992                 key.objectid = found_key.objectid + found_key.offset;
 993                 btrfs_release_path(root, path);
 994                 ret = radix_tree_insert(&root->fs_info->block_group_radix,
 995                                         found_key.objectid +
 996                                         found_key.offset - 1,
 997                                         (void *)cache);
 998                 BUG_ON(ret);
 999                 if (key.objectid >=
1000                     btrfs_super_total_blocks(root->fs_info->disk_super))
1001                         break;
1002         }
1003
1004         btrfs_free_path(path);
1005         return 0;
1006 }