ctree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  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 #include "ctree.h"
  19 #include "disk-io.h"
  20 #include "transaction.h"
  21 #include "print-tree.h"
  22 #include "repair.h"
  23
  24 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  25                       *root, struct btrfs_path *path, int level);
  26 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  27                       *root, struct btrfs_key *ins_key,
  28                       struct btrfs_path *path, int data_size, int extend);
  29 static int push_node_left(struct btrfs_trans_handle *trans,
  30                           struct btrfs_root *root, struct extent_buffer *dst,
  31                           struct extent_buffer *src, int empty);
  32 static int balance_node_right(struct btrfs_trans_handle *trans,
  33                               struct btrfs_root *root,
  34                               struct extent_buffer *dst_buf,
  35                               struct extent_buffer *src_buf);
  36
  37 inline void btrfs_init_path(struct btrfs_path *p)
  38 {
  39         memset(p, 0, sizeof(*p));
  40 }
  41
  42 struct btrfs_path *btrfs_alloc_path(void)
  43 {
  44         struct btrfs_path *path;
  45         path = kzalloc(sizeof(struct btrfs_path), GFP_NOFS);
  46         return path;
  47 }
  48
  49 void btrfs_free_path(struct btrfs_path *p)
  50 {
  51         btrfs_release_path(p);
  52         kfree(p);
  53 }
  54
  55 void btrfs_release_path(struct btrfs_path *p)
  56 {
  57         int i;
  58         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  59                 if (!p->nodes[i])
  60                         continue;
  61                 free_extent_buffer(p->nodes[i]);
  62         }
  63         memset(p, 0, sizeof(*p));
  64 }
  65
  66 void add_root_to_dirty_list(struct btrfs_root *root)
  67 {
  68         if (root->track_dirty && list_empty(&root->dirty_list)) {
  69                 list_add(&root->dirty_list,
  70                          &root->fs_info->dirty_cowonly_roots);
  71         }
  72 }
  73
  74 int btrfs_copy_root(struct btrfs_trans_handle *trans,
  75                       struct btrfs_root *root,
  76                       struct extent_buffer *buf,
  77                       struct extent_buffer **cow_ret, u64 new_root_objectid)
  78 {
  79         struct extent_buffer *cow;
  80         int ret = 0;
  81         int level;
  82         struct btrfs_root *new_root;
  83         struct btrfs_disk_key disk_key;
  84
  85         new_root = kmalloc(sizeof(*new_root), GFP_NOFS);
  86         if (!new_root)
  87                 return -ENOMEM;
  88
  89         memcpy(new_root, root, sizeof(*new_root));
  90         new_root->root_key.objectid = new_root_objectid;
  91
  92         WARN_ON(root->ref_cows && trans->transid !=
  93                 root->fs_info->running_transaction->transid);
  94         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
  95
  96         level = btrfs_header_level(buf);
  97         if (level == 0)
  98                 btrfs_item_key(buf, &disk_key, 0);
  99         else
 100                 btrfs_node_key(buf, &disk_key, 0);
 101         cow = btrfs_alloc_free_block(trans, new_root, buf->len,
 102                                      new_root_objectid, &disk_key,
 103                                      level, buf->start, 0);
 104         if (IS_ERR(cow)) {
 105                 kfree(new_root);
 106                 return PTR_ERR(cow);
 107         }
 108
 109         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 110         btrfs_set_header_bytenr(cow, cow->start);
 111         btrfs_set_header_generation(cow, trans->transid);
 112         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 113         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 114                                      BTRFS_HEADER_FLAG_RELOC);
 115         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 116                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 117         else
 118                 btrfs_set_header_owner(cow, new_root_objectid);
 119
 120         write_extent_buffer(cow, root->fs_info->fsid,
 121                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
 122
 123         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 124         ret = btrfs_inc_ref(trans, new_root, cow, 0);
 125         kfree(new_root);
 126
 127         if (ret)
 128                 return ret;
 129
 130         btrfs_mark_buffer_dirty(cow);
 131         *cow_ret = cow;
 132         return 0;
 133 }
 134
 135 /*
 136  * check if the tree block can be shared by multiple trees
 137  */
 138 static int btrfs_block_can_be_shared(struct btrfs_root *root,
 139                                      struct extent_buffer *buf)
 140 {
 141         /*
 142          * Tree blocks not in refernece counted trees and tree roots
 143          * are never shared. If a block was allocated after the last
 144          * snapshot and the block was not allocated by tree relocation,
 145          * we know the block is not shared.
 146          */
 147         if (root->ref_cows &&
 148             buf != root->node && buf != root->commit_root &&
 149             (btrfs_header_generation(buf) <=
 150              btrfs_root_last_snapshot(&root->root_item) ||
 151              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 152                 return 1;
 153 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 154         if (root->ref_cows &&
 155             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 156                 return 1;
 157 #endif
 158         return 0;
 159 }
 160
 161 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
 162                                        struct btrfs_root *root,
 163                                        struct extent_buffer *buf,
 164                                        struct extent_buffer *cow)
 165 {
 166         u64 refs;
 167         u64 owner;
 168         u64 flags;
 169         u64 new_flags = 0;
 170         int ret;
 171
 172         /*
 173          * Backrefs update rules:
 174          *
 175          * Always use full backrefs for extent pointers in tree block
 176          * allocated by tree relocation.
 177          *
 178          * If a shared tree block is no longer referenced by its owner
 179          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
 180          * use full backrefs for extent pointers in tree block.
 181          *
 182          * If a tree block is been relocating
 183          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
 184          * use full backrefs for extent pointers in tree block.
 185          * The reason for this is some operations (such as drop tree)
 186          * are only allowed for blocks use full backrefs.
 187          */
 188
 189         if (btrfs_block_can_be_shared(root, buf)) {
 190                 ret = btrfs_lookup_extent_info(trans, root, buf->start,
 191                                                btrfs_header_level(buf), 1,
 192                                                &refs, &flags);
 193                 BUG_ON(ret);
 194                 BUG_ON(refs == 0);
 195         } else {
 196                 refs = 1;
 197                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 198                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 199                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 200                 else
 201                         flags = 0;
 202         }
 203
 204         owner = btrfs_header_owner(buf);
 205         BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) &&
 206                owner == BTRFS_TREE_RELOC_OBJECTID);
 207
 208         if (refs > 1) {
 209                 if ((owner == root->root_key.objectid ||
 210                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
 211                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
 212                         ret = btrfs_inc_ref(trans, root, buf, 1);
 213                         BUG_ON(ret);
 214
 215                         if (root->root_key.objectid ==
 216                             BTRFS_TREE_RELOC_OBJECTID) {
 217                                 ret = btrfs_dec_ref(trans, root, buf, 0);
 218                                 BUG_ON(ret);
 219                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 220                                 BUG_ON(ret);
 221                         }
 222                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
 223                 } else {
 224
 225                         if (root->root_key.objectid ==
 226                             BTRFS_TREE_RELOC_OBJECTID)
 227                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 228                         else
 229                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 230                         BUG_ON(ret);
 231                 }
 232                 if (new_flags != 0) {
 233                         ret = btrfs_set_block_flags(trans, root, buf->start,
 234                                                     btrfs_header_level(buf),
 235                                                     new_flags);
 236                         BUG_ON(ret);
 237                 }
 238         } else {
 239                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
 240                         if (root->root_key.objectid ==
 241                             BTRFS_TREE_RELOC_OBJECTID)
 242                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 243                         else
 244                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 245                         BUG_ON(ret);
 246                         ret = btrfs_dec_ref(trans, root, buf, 1);
 247                         BUG_ON(ret);
 248                 }
 249                 clean_tree_block(trans, root, buf);
 250         }
 251         return 0;
 252 }
 253
 254 int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 255                              struct btrfs_root *root,
 256                              struct extent_buffer *buf,
 257                              struct extent_buffer *parent, int parent_slot,
 258                              struct extent_buffer **cow_ret,
 259                              u64 search_start, u64 empty_size)
 260 {
 261         struct extent_buffer *cow;
 262         struct btrfs_disk_key disk_key;
 263         int level;
 264
 265         WARN_ON(root->ref_cows && trans->transid !=
 266                 root->fs_info->running_transaction->transid);
 267         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 268
 269         level = btrfs_header_level(buf);
 270
 271         if (level == 0)
 272                 btrfs_item_key(buf, &disk_key, 0);
 273         else
 274                 btrfs_node_key(buf, &disk_key, 0);
 275
 276         cow = btrfs_alloc_free_block(trans, root, buf->len,
 277                                      root->root_key.objectid, &disk_key,
 278                                      level, search_start, empty_size);
 279         if (IS_ERR(cow))
 280                 return PTR_ERR(cow);
 281
 282         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 283         btrfs_set_header_bytenr(cow, cow->start);
 284         btrfs_set_header_generation(cow, trans->transid);
 285         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 286         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 287                                      BTRFS_HEADER_FLAG_RELOC);
 288         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 289                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 290         else
 291                 btrfs_set_header_owner(cow, root->root_key.objectid);
 292
 293         write_extent_buffer(cow, root->fs_info->fsid,
 294                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
 295
 296         WARN_ON(!(buf->flags & EXTENT_BAD_TRANSID) &&
 297                 btrfs_header_generation(buf) > trans->transid);
 298
 299         update_ref_for_cow(trans, root, buf, cow);
 300
 301         if (buf == root->node) {
 302                 root->node = cow;
 303                 extent_buffer_get(cow);
 304
 305                 btrfs_free_extent(trans, root, buf->start, buf->len,
 306                                   0, root->root_key.objectid, level, 0);
 307                 free_extent_buffer(buf);
 308                 add_root_to_dirty_list(root);
 309         } else {
 310                 btrfs_set_node_blockptr(parent, parent_slot,
 311                                         cow->start);
 312                 WARN_ON(trans->transid == 0);
 313                 btrfs_set_node_ptr_generation(parent, parent_slot,
 314                                               trans->transid);
 315                 btrfs_mark_buffer_dirty(parent);
 316                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 317
 318                 btrfs_free_extent(trans, root, buf->start, buf->len,
 319                                   0, root->root_key.objectid, level, 1);
 320         }
 321         if (!list_empty(&buf->recow)) {
 322                 list_del_init(&buf->recow);
 323                 free_extent_buffer(buf);
 324         }
 325         free_extent_buffer(buf);
 326         btrfs_mark_buffer_dirty(cow);
 327         *cow_ret = cow;
 328         return 0;
 329 }
 330
 331 static inline int should_cow_block(struct btrfs_trans_handle *trans,
 332                                    struct btrfs_root *root,
 333                                    struct extent_buffer *buf)
 334 {
 335         if (btrfs_header_generation(buf) == trans->transid &&
 336             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
 337             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
 338               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 339                 return 0;
 340         return 1;
 341 }
 342
 343 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 344                     struct btrfs_root *root, struct extent_buffer *buf,
 345                     struct extent_buffer *parent, int parent_slot,
 346                     struct extent_buffer **cow_ret)
 347 {
 348         u64 search_start;
 349         int ret;
 350         /*
 351         if (trans->transaction != root->fs_info->running_transaction) {
 352                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 353                        root->fs_info->running_transaction->transid);
 354                 WARN_ON(1);
 355         }
 356         */
 357         if (trans->transid != root->fs_info->generation) {
 358                 printk(KERN_CRIT "trans %llu running %llu\n",
 359                         (unsigned long long)trans->transid,
 360                         (unsigned long long)root->fs_info->generation);
 361                 WARN_ON(1);
 362         }
 363         if (!should_cow_block(trans, root, buf)) {
 364                 *cow_ret = buf;
 365                 return 0;
 366         }
 367
 368         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
 369         ret = __btrfs_cow_block(trans, root, buf, parent,
 370                                  parent_slot, cow_ret, search_start, 0);
 371         return ret;
 372 }
 373
 374 /*
 375  * compare two keys in a memcmp fashion
 376  */
 377 static int btrfs_comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 378 {
 379         struct btrfs_key k1;
 380
 381         btrfs_disk_key_to_cpu(&k1, disk);
 382
 383         if (k1.objectid > k2->objectid)
 384                 return 1;
 385         if (k1.objectid < k2->objectid)
 386                 return -1;
 387         if (k1.type > k2->type)
 388                 return 1;
 389         if (k1.type < k2->type)
 390                 return -1;
 391         if (k1.offset > k2->offset)
 392                 return 1;
 393         if (k1.offset < k2->offset)
 394                 return -1;
 395         return 0;
 396 }
 397
 398 /*
 399  * The leaf data grows from end-to-front in the node.
 400  * this returns the address of the start of the last item,
 401  * which is the stop of the leaf data stack
 402  */
 403 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 404                                          struct extent_buffer *leaf)
 405 {
 406         u32 nr = btrfs_header_nritems(leaf);
 407         if (nr == 0)
 408                 return BTRFS_LEAF_DATA_SIZE(root);
 409         return btrfs_item_offset_nr(leaf, nr - 1);
 410 }
 411
 412 int btrfs_check_node(struct btrfs_root *root,
 413                       struct btrfs_disk_key *parent_key,
 414                       struct extent_buffer *buf)
 415 {
 416         int i;
 417         struct btrfs_key cpukey;
 418         struct btrfs_disk_key key;
 419         u32 nritems = btrfs_header_nritems(buf);
 420
 421         if (nritems == 0 || nritems > BTRFS_NODEPTRS_PER_BLOCK(root))
 422                 goto fail;
 423
 424         if (parent_key && parent_key->type) {
 425                 btrfs_node_key(buf, &key, 0);
 426                 if (memcmp(parent_key, &key, sizeof(key)))
 427                         goto fail;
 428         }
 429         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 430                 btrfs_node_key(buf, &key, i);
 431                 btrfs_node_key_to_cpu(buf, &cpukey, i + 1);
 432                 if (btrfs_comp_keys(&key, &cpukey) >= 0)
 433                         goto fail;
 434         }
 435         return 0;
 436 fail:
 437         if (btrfs_header_owner(buf) == BTRFS_EXTENT_TREE_OBJECTID) {
 438                 if (parent_key)
 439                         btrfs_disk_key_to_cpu(&cpukey, parent_key);
 440                 else
 441                         btrfs_node_key_to_cpu(buf, &cpukey, 0);
 442                 btrfs_add_corrupt_extent_record(root->fs_info, &cpukey,
 443                                                 buf->start, buf->len,
 444                                                 btrfs_header_level(buf));
 445         }
 446         return -EIO;
 447 }
 448
 449 int btrfs_check_leaf(struct btrfs_root *root,
 450                       struct btrfs_disk_key *parent_key,
 451                       struct extent_buffer *buf)
 452 {
 453         int i;
 454         struct btrfs_key cpukey;
 455         struct btrfs_disk_key key;
 456         u32 nritems = btrfs_header_nritems(buf);
 457
 458         if (nritems * sizeof(struct btrfs_item) > buf->len)  {
 459                 fprintf(stderr, "invalid number of items %llu\n",
 460                         (unsigned long long)buf->start);
 461                 goto fail;
 462         }
 463
 464         if (btrfs_header_level(buf) != 0) {
 465                 fprintf(stderr, "leaf is not a leaf %llu\n",
 466                        (unsigned long long)btrfs_header_bytenr(buf));
 467                 goto fail;
 468         }
 469         if (btrfs_leaf_free_space(root, buf) < 0) {
 470                 fprintf(stderr, "leaf free space incorrect %llu %d\n",
 471                         (unsigned long long)btrfs_header_bytenr(buf),
 472                         btrfs_leaf_free_space(root, buf));
 473                 goto fail;
 474         }
 475
 476         if (nritems == 0)
 477                 return 0;
 478
 479         btrfs_item_key(buf, &key, 0);
 480         if (parent_key && parent_key->type &&
 481             memcmp(parent_key, &key, sizeof(key))) {
 482                 fprintf(stderr, "leaf parent key incorrect %llu\n",
 483                        (unsigned long long)btrfs_header_bytenr(buf));
 484                 goto fail;
 485         }
 486         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 487                 btrfs_item_key(buf, &key, i);
 488                 btrfs_item_key_to_cpu(buf, &cpukey, i + 1);
 489                 if (btrfs_comp_keys(&key, &cpukey) >= 0) {
 490                         fprintf(stderr, "bad key ordering %d %d\n", i, i+1);
 491                         goto fail;
 492                 }
 493                 if (btrfs_item_offset_nr(buf, i) !=
 494                         btrfs_item_end_nr(buf, i + 1)) {
 495                         fprintf(stderr, "incorrect offsets %u %u\n",
 496                                 btrfs_item_offset_nr(buf, i),
 497                                 btrfs_item_end_nr(buf, i + 1));
 498                         goto fail;
 499                 }
 500                 if (i == 0 && btrfs_item_end_nr(buf, i) !=
 501                     BTRFS_LEAF_DATA_SIZE(root)) {
 502                         fprintf(stderr, "bad item end %u wanted %u\n",
 503                                 btrfs_item_end_nr(buf, i),
 504                                 (unsigned)BTRFS_LEAF_DATA_SIZE(root));
 505                         goto fail;
 506                 }
 507         }
 508         return 0;
 509 fail:
 510         if (btrfs_header_owner(buf) == BTRFS_EXTENT_TREE_OBJECTID) {
 511                 if (parent_key)
 512                         btrfs_disk_key_to_cpu(&cpukey, parent_key);
 513                 else
 514                         btrfs_item_key_to_cpu(buf, &cpukey, 0);
 515
 516                 btrfs_add_corrupt_extent_record(root->fs_info, &cpukey,
 517                                                 buf->start, buf->len, 0);
 518         }
 519         return -EIO;
 520 }
 521
 522 static int noinline check_block(struct btrfs_root *root,
 523                                 struct btrfs_path *path, int level)
 524 {
 525         struct btrfs_disk_key key;
 526         struct btrfs_disk_key *key_ptr = NULL;
 527         struct extent_buffer *parent;
 528
 529         if (path->nodes[level + 1]) {
 530                 parent = path->nodes[level + 1];
 531                 btrfs_node_key(parent, &key, path->slots[level + 1]);
 532                 key_ptr = &key;
 533         }
 534         if (level == 0)
 535                 return btrfs_check_leaf(root, key_ptr, path->nodes[0]);
 536         return btrfs_check_node(root, key_ptr, path->nodes[level]);
 537 }
 538
 539 /*
 540  * search for key in the extent_buffer.  The items start at offset p,
 541  * and they are item_size apart.  There are 'max' items in p.
 542  *
 543  * the slot in the array is returned via slot, and it points to
 544  * the place where you would insert key if it is not found in
 545  * the array.
 546  *
 547  * slot may point to max if the key is bigger than all of the keys
 548  */
 549 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 550                               int item_size, struct btrfs_key *key,
 551                               int max, int *slot)
 552 {
 553         int low = 0;
 554         int high = max;
 555         int mid;
 556         int ret;
 557         unsigned long offset;
 558         struct btrfs_disk_key *tmp;
 559
 560         while(low < high) {
 561                 mid = (low + high) / 2;
 562                 offset = p + mid * item_size;
 563
 564                 tmp = (struct btrfs_disk_key *)(eb->data + offset);
 565                 ret = btrfs_comp_keys(tmp, key);
 566
 567                 if (ret < 0)
 568                         low = mid + 1;
 569                 else if (ret > 0)
 570                         high = mid;
 571                 else {
 572                         *slot = mid;
 573                         return 0;
 574                 }
 575         }
 576         *slot = low;
 577         return 1;
 578 }
 579
 580 /*
 581  * simple bin_search frontend that does the right thing for
 582  * leaves vs nodes
 583  */
 584 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 585                       int level, int *slot)
 586 {
 587         if (level == 0)
 588                 return generic_bin_search(eb,
 589                                           offsetof(struct btrfs_leaf, items),
 590                                           sizeof(struct btrfs_item),
 591                                           key, btrfs_header_nritems(eb),
 592                                           slot);
 593         else
 594                 return generic_bin_search(eb,
 595                                           offsetof(struct btrfs_node, ptrs),
 596                                           sizeof(struct btrfs_key_ptr),
 597                                           key, btrfs_header_nritems(eb),
 598                                           slot);
 599 }
 600
 601 struct extent_buffer *read_node_slot(struct btrfs_root *root,
 602                                    struct extent_buffer *parent, int slot)
 603 {
 604         int level = btrfs_header_level(parent);
 605         if (slot < 0)
 606                 return NULL;
 607         if (slot >= btrfs_header_nritems(parent))
 608                 return NULL;
 609
 610         if (level == 0)
 611                 return NULL;
 612
 613         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 614                        btrfs_level_size(root, level - 1),
 615                        btrfs_node_ptr_generation(parent, slot));
 616 }
 617
 618 static int balance_level(struct btrfs_trans_handle *trans,
 619                          struct btrfs_root *root,
 620                          struct btrfs_path *path, int level)
 621 {
 622         struct extent_buffer *right = NULL;
 623         struct extent_buffer *mid;
 624         struct extent_buffer *left = NULL;
 625         struct extent_buffer *parent = NULL;
 626         int ret = 0;
 627         int wret;
 628         int pslot;
 629         int orig_slot = path->slots[level];
 630         u64 orig_ptr;
 631
 632         if (level == 0)
 633                 return 0;
 634
 635         mid = path->nodes[level];
 636         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 637
 638         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 639
 640         if (level < BTRFS_MAX_LEVEL - 1) {
 641                 parent = path->nodes[level + 1];
 642                 pslot = path->slots[level + 1];
 643         }
 644
 645         /*
 646          * deal with the case where there is only one pointer in the root
 647          * by promoting the node below to a root
 648          */
 649         if (!parent) {
 650                 struct extent_buffer *child;
 651
 652                 if (btrfs_header_nritems(mid) != 1)
 653                         return 0;
 654
 655                 /* promote the child to a root */
 656                 child = read_node_slot(root, mid, 0);
 657                 BUG_ON(!child);
 658                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
 659                 BUG_ON(ret);
 660
 661                 root->node = child;
 662                 add_root_to_dirty_list(root);
 663                 path->nodes[level] = NULL;
 664                 clean_tree_block(trans, root, mid);
 665                 wait_on_tree_block_writeback(root, mid);
 666                 /* once for the path */
 667                 free_extent_buffer(mid);
 668
 669                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 670                                         0, root->root_key.objectid,
 671                                         level, 1);
 672                 /* once for the root ptr */
 673                 free_extent_buffer(mid);
 674                 return ret;
 675         }
 676         if (btrfs_header_nritems(mid) >
 677             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 678                 return 0;
 679
 680         left = read_node_slot(root, parent, pslot - 1);
 681         if (left) {
 682                 wret = btrfs_cow_block(trans, root, left,
 683                                        parent, pslot - 1, &left);
 684                 if (wret) {
 685                         ret = wret;
 686                         goto enospc;
 687                 }
 688         }
 689         right = read_node_slot(root, parent, pslot + 1);
 690         if (right) {
 691                 wret = btrfs_cow_block(trans, root, right,
 692                                        parent, pslot + 1, &right);
 693                 if (wret) {
 694                         ret = wret;
 695                         goto enospc;
 696                 }
 697         }
 698
 699         /* first, try to make some room in the middle buffer */
 700         if (left) {
 701                 orig_slot += btrfs_header_nritems(left);
 702                 wret = push_node_left(trans, root, left, mid, 1);
 703                 if (wret < 0)
 704                         ret = wret;
 705         }
 706
 707         /*
 708          * then try to empty the right most buffer into the middle
 709          */
 710         if (right) {
 711                 wret = push_node_left(trans, root, mid, right, 1);
 712                 if (wret < 0 && wret != -ENOSPC)
 713                         ret = wret;
 714                 if (btrfs_header_nritems(right) == 0) {
 715                         u64 bytenr = right->start;
 716                         u32 blocksize = right->len;
 717
 718                         clean_tree_block(trans, root, right);
 719                         wait_on_tree_block_writeback(root, right);
 720                         free_extent_buffer(right);
 721                         right = NULL;
 722                         wret = btrfs_del_ptr(trans, root, path,
 723                                              level + 1, pslot + 1);
 724                         if (wret)
 725                                 ret = wret;
 726                         wret = btrfs_free_extent(trans, root, bytenr,
 727                                                  blocksize, 0,
 728                                                  root->root_key.objectid,
 729                                                  level, 0);
 730                         if (wret)
 731                                 ret = wret;
 732                 } else {
 733                         struct btrfs_disk_key right_key;
 734                         btrfs_node_key(right, &right_key, 0);
 735                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 736                         btrfs_mark_buffer_dirty(parent);
 737                 }
 738         }
 739         if (btrfs_header_nritems(mid) == 1) {
 740                 /*
 741                  * we're not allowed to leave a node with one item in the
 742                  * tree during a delete.  A deletion from lower in the tree
 743                  * could try to delete the only pointer in this node.
 744                  * So, pull some keys from the left.
 745                  * There has to be a left pointer at this point because
 746                  * otherwise we would have pulled some pointers from the
 747                  * right
 748                  */
 749                 BUG_ON(!left);
 750                 wret = balance_node_right(trans, root, mid, left);
 751                 if (wret < 0) {
 752                         ret = wret;
 753                         goto enospc;
 754                 }
 755                 if (wret == 1) {
 756                         wret = push_node_left(trans, root, left, mid, 1);
 757                         if (wret < 0)
 758                                 ret = wret;
 759                 }
 760                 BUG_ON(wret == 1);
 761         }
 762         if (btrfs_header_nritems(mid) == 0) {
 763                 /* we've managed to empty the middle node, drop it */
 764                 u64 bytenr = mid->start;
 765                 u32 blocksize = mid->len;
 766                 clean_tree_block(trans, root, mid);
 767                 wait_on_tree_block_writeback(root, mid);
 768                 free_extent_buffer(mid);
 769                 mid = NULL;
 770                 wret = btrfs_del_ptr(trans, root, path, level + 1, pslot);
 771                 if (wret)
 772                         ret = wret;
 773                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 774                                          0, root->root_key.objectid,
 775                                          level, 0);
 776                 if (wret)
 777                         ret = wret;
 778         } else {
 779                 /* update the parent key to reflect our changes */
 780                 struct btrfs_disk_key mid_key;
 781                 btrfs_node_key(mid, &mid_key, 0);
 782                 btrfs_set_node_key(parent, &mid_key, pslot);
 783                 btrfs_mark_buffer_dirty(parent);
 784         }
 785
 786         /* update the path */
 787         if (left) {
 788                 if (btrfs_header_nritems(left) > orig_slot) {
 789                         extent_buffer_get(left);
 790                         path->nodes[level] = left;
 791                         path->slots[level + 1] -= 1;
 792                         path->slots[level] = orig_slot;
 793                         if (mid)
 794                                 free_extent_buffer(mid);
 795                 } else {
 796                         orig_slot -= btrfs_header_nritems(left);
 797                         path->slots[level] = orig_slot;
 798                 }
 799         }
 800         /* double check we haven't messed things up */
 801         check_block(root, path, level);
 802         if (orig_ptr !=
 803             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 804                 BUG();
 805 enospc:
 806         if (right)
 807                 free_extent_buffer(right);
 808         if (left)
 809                 free_extent_buffer(left);
 810         return ret;
 811 }
 812
 813 /* returns zero if the push worked, non-zero otherwise */
 814 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
 815                                           struct btrfs_root *root,
 816                                           struct btrfs_path *path, int level)
 817 {
 818         struct extent_buffer *right = NULL;
 819         struct extent_buffer *mid;
 820         struct extent_buffer *left = NULL;
 821         struct extent_buffer *parent = NULL;
 822         int ret = 0;
 823         int wret;
 824         int pslot;
 825         int orig_slot = path->slots[level];
 826
 827         if (level == 0)
 828                 return 1;
 829
 830         mid = path->nodes[level];
 831         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 832
 833         if (level < BTRFS_MAX_LEVEL - 1) {
 834                 parent = path->nodes[level + 1];
 835                 pslot = path->slots[level + 1];
 836         }
 837
 838         if (!parent)
 839                 return 1;
 840
 841         left = read_node_slot(root, parent, pslot - 1);
 842
 843         /* first, try to make some room in the middle buffer */
 844         if (left) {
 845                 u32 left_nr;
 846                 left_nr = btrfs_header_nritems(left);
 847                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 848                         wret = 1;
 849                 } else {
 850                         ret = btrfs_cow_block(trans, root, left, parent,
 851                                               pslot - 1, &left);
 852                         if (ret)
 853                                 wret = 1;
 854                         else {
 855                                 wret = push_node_left(trans, root,
 856                                                       left, mid, 0);
 857                         }
 858                 }
 859                 if (wret < 0)
 860                         ret = wret;
 861                 if (wret == 0) {
 862                         struct btrfs_disk_key disk_key;
 863                         orig_slot += left_nr;
 864                         btrfs_node_key(mid, &disk_key, 0);
 865                         btrfs_set_node_key(parent, &disk_key, pslot);
 866                         btrfs_mark_buffer_dirty(parent);
 867                         if (btrfs_header_nritems(left) > orig_slot) {
 868                                 path->nodes[level] = left;
 869                                 path->slots[level + 1] -= 1;
 870                                 path->slots[level] = orig_slot;
 871                                 free_extent_buffer(mid);
 872                         } else {
 873                                 orig_slot -=
 874                                         btrfs_header_nritems(left);
 875                                 path->slots[level] = orig_slot;
 876                                 free_extent_buffer(left);
 877                         }
 878                         return 0;
 879                 }
 880                 free_extent_buffer(left);
 881         }
 882         right= read_node_slot(root, parent, pslot + 1);
 883
 884         /*
 885          * then try to empty the right most buffer into the middle
 886          */
 887         if (right) {
 888                 u32 right_nr;
 889                 right_nr = btrfs_header_nritems(right);
 890                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 891                         wret = 1;
 892                 } else {
 893                         ret = btrfs_cow_block(trans, root, right,
 894                                               parent, pslot + 1,
 895                                               &right);
 896                         if (ret)
 897                                 wret = 1;
 898                         else {
 899                                 wret = balance_node_right(trans, root,
 900                                                           right, mid);
 901                         }
 902                 }
 903                 if (wret < 0)
 904                         ret = wret;
 905                 if (wret == 0) {
 906                         struct btrfs_disk_key disk_key;
 907
 908                         btrfs_node_key(right, &disk_key, 0);
 909                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
 910                         btrfs_mark_buffer_dirty(parent);
 911
 912                         if (btrfs_header_nritems(mid) <= orig_slot) {
 913                                 path->nodes[level] = right;
 914                                 path->slots[level + 1] += 1;
 915                                 path->slots[level] = orig_slot -
 916                                         btrfs_header_nritems(mid);
 917                                 free_extent_buffer(mid);
 918                         } else {
 919                                 free_extent_buffer(right);
 920                         }
 921                         return 0;
 922                 }
 923                 free_extent_buffer(right);
 924         }
 925         return 1;
 926 }
 927
 928 /*
 929  * readahead one full node of leaves
 930  */
 931 void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
 932                              int level, int slot, u64 objectid)
 933 {
 934         struct extent_buffer *node;
 935         struct btrfs_disk_key disk_key;
 936         u32 nritems;
 937         u64 search;
 938         u64 lowest_read;
 939         u64 highest_read;
 940         u64 nread = 0;
 941         int direction = path->reada;
 942         struct extent_buffer *eb;
 943         u32 nr;
 944         u32 blocksize;
 945         u32 nscan = 0;
 946
 947         if (level != 1)
 948                 return;
 949
 950         if (!path->nodes[level])
 951                 return;
 952
 953         node = path->nodes[level];
 954         search = btrfs_node_blockptr(node, slot);
 955         blocksize = btrfs_level_size(root, level - 1);
 956         eb = btrfs_find_tree_block(root, search, blocksize);
 957         if (eb) {
 958                 free_extent_buffer(eb);
 959                 return;
 960         }
 961
 962         highest_read = search;
 963         lowest_read = search;
 964
 965         nritems = btrfs_header_nritems(node);
 966         nr = slot;
 967         while(1) {
 968                 if (direction < 0) {
 969                         if (nr == 0)
 970                                 break;
 971                         nr--;
 972                 } else if (direction > 0) {
 973                         nr++;
 974                         if (nr >= nritems)
 975                                 break;
 976                 }
 977                 if (path->reada < 0 && objectid) {
 978                         btrfs_node_key(node, &disk_key, nr);
 979                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
 980                                 break;
 981                 }
 982                 search = btrfs_node_blockptr(node, nr);
 983                 if ((search >= lowest_read && search <= highest_read) ||
 984                     (search < lowest_read && lowest_read - search <= 32768) ||
 985                     (search > highest_read && search - highest_read <= 32768)) {
 986                         readahead_tree_block(root, search, blocksize,
 987                                      btrfs_node_ptr_generation(node, nr));
 988                         nread += blocksize;
 989                 }
 990                 nscan++;
 991                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
 992                         break;
 993                 if(nread > (1024 * 1024) || nscan > 128)
 994                         break;
 995
 996                 if (search < lowest_read)
 997                         lowest_read = search;
 998                 if (search > highest_read)
 999                         highest_read = search;
1000         }
1001 }
1002
1003 /*
1004  * look for key in the tree.  path is filled in with nodes along the way
1005  * if key is found, we return zero and you can find the item in the leaf
1006  * level of the path (level 0)
1007  *
1008  * If the key isn't found, the path points to the slot where it should
1009  * be inserted, and 1 is returned.  If there are other errors during the
1010  * search a negative error number is returned.
1011  *
1012  * if ins_len > 0, nodes and leaves will be split as we walk down the
1013  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1014  * possible)
1015  */
1016 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1017                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1018                       ins_len, int cow)
1019 {
1020         struct extent_buffer *b;
1021         int slot;
1022         int ret;
1023         int level;
1024         int should_reada = p->reada;
1025         u8 lowest_level = 0;
1026
1027         lowest_level = p->lowest_level;
1028         WARN_ON(lowest_level && ins_len > 0);
1029         WARN_ON(p->nodes[0] != NULL);
1030         /*
1031         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
1032         */
1033 again:
1034         b = root->node;
1035         extent_buffer_get(b);
1036         while (b) {
1037                 level = btrfs_header_level(b);
1038                 if (cow) {
1039                         int wret;
1040                         wret = btrfs_cow_block(trans, root, b,
1041                                                p->nodes[level + 1],
1042                                                p->slots[level + 1],
1043                                                &b);
1044                         if (wret) {
1045                                 free_extent_buffer(b);
1046                                 return wret;
1047                         }
1048                 }
1049                 BUG_ON(!cow && ins_len);
1050                 if (level != btrfs_header_level(b))
1051                         WARN_ON(1);
1052                 level = btrfs_header_level(b);
1053                 p->nodes[level] = b;
1054                 ret = check_block(root, p, level);
1055                 if (ret)
1056                         return -1;
1057                 ret = bin_search(b, key, level, &slot);
1058                 if (level != 0) {
1059                         if (ret && slot > 0)
1060                                 slot -= 1;
1061                         p->slots[level] = slot;
1062                         if ((p->search_for_split || ins_len > 0) &&
1063                             btrfs_header_nritems(b) >=
1064                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1065                                 int sret = split_node(trans, root, p, level);
1066                                 BUG_ON(sret > 0);
1067                                 if (sret)
1068                                         return sret;
1069                                 b = p->nodes[level];
1070                                 slot = p->slots[level];
1071                         } else if (ins_len < 0) {
1072                                 int sret = balance_level(trans, root, p,
1073                                                          level);
1074                                 if (sret)
1075                                         return sret;
1076                                 b = p->nodes[level];
1077                                 if (!b) {
1078                                         btrfs_release_path(p);
1079                                         goto again;
1080                                 }
1081                                 slot = p->slots[level];
1082                                 BUG_ON(btrfs_header_nritems(b) == 1);
1083                         }
1084                         /* this is only true while dropping a snapshot */
1085                         if (level == lowest_level)
1086                                 break;
1087
1088                         if (should_reada)
1089                                 reada_for_search(root, p, level, slot,
1090                                                  key->objectid);
1091
1092                         b = read_node_slot(root, b, slot);
1093                         if (!extent_buffer_uptodate(b))
1094                                 return -EIO;
1095                 } else {
1096                         p->slots[level] = slot;
1097                         if (ins_len > 0 &&
1098                             ins_len > btrfs_leaf_free_space(root, b)) {
1099                                 int sret = split_leaf(trans, root, key,
1100                                                       p, ins_len, ret == 0);
1101                                 BUG_ON(sret > 0);
1102                                 if (sret)
1103                                         return sret;
1104                         }
1105                         return ret;
1106                 }
1107         }
1108         return 1;
1109 }
1110
1111 /*
1112  * adjust the pointers going up the tree, starting at level
1113  * making sure the right key of each node is points to 'key'.
1114  * This is used after shifting pointers to the left, so it stops
1115  * fixing up pointers when a given leaf/node is not in slot 0 of the
1116  * higher levels
1117  *
1118  * If this fails to write a tree block, it returns -1, but continues
1119  * fixing up the blocks in ram so the tree is consistent.
1120  */
1121 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1122                           struct btrfs_root *root, struct btrfs_path *path,
1123                           struct btrfs_disk_key *key, int level)
1124 {
1125         int i;
1126         int ret = 0;
1127         struct extent_buffer *t;
1128
1129         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1130                 int tslot = path->slots[i];
1131                 if (!path->nodes[i])
1132                         break;
1133                 t = path->nodes[i];
1134                 btrfs_set_node_key(t, key, tslot);
1135                 btrfs_mark_buffer_dirty(path->nodes[i]);
1136                 if (tslot != 0)
1137                         break;
1138         }
1139         return ret;
1140 }
1141
1142 /*
1143  * update item key.
1144  *
1145  * This function isn't completely safe. It's the caller's responsibility
1146  * that the new key won't break the order
1147  */
1148 int btrfs_set_item_key_safe(struct btrfs_trans_handle *trans,
1149                             struct btrfs_root *root, struct btrfs_path *path,
1150                             struct btrfs_key *new_key)
1151 {
1152         struct btrfs_disk_key disk_key;
1153         struct extent_buffer *eb;
1154         int slot;
1155
1156         eb = path->nodes[0];
1157         slot = path->slots[0];
1158         if (slot > 0) {
1159                 btrfs_item_key(eb, &disk_key, slot - 1);
1160                 if (btrfs_comp_keys(&disk_key, new_key) >= 0)
1161                         return -1;
1162         }
1163         if (slot < btrfs_header_nritems(eb) - 1) {
1164                 btrfs_item_key(eb, &disk_key, slot + 1);
1165                 if (btrfs_comp_keys(&disk_key, new_key) <= 0)
1166                         return -1;
1167         }
1168
1169         btrfs_cpu_key_to_disk(&disk_key, new_key);
1170         btrfs_set_item_key(eb, &disk_key, slot);
1171         btrfs_mark_buffer_dirty(eb);
1172         if (slot == 0)
1173                 fixup_low_keys(trans, root, path, &disk_key, 1);
1174         return 0;
1175 }
1176
1177 /*
1178  * try to push data from one node into the next node left in the
1179  * tree.
1180  *
1181  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1182  * error, and > 0 if there was no room in the left hand block.
1183  */
1184 static int push_node_left(struct btrfs_trans_handle *trans,
1185                           struct btrfs_root *root, struct extent_buffer *dst,
1186                           struct extent_buffer *src, int empty)
1187 {
1188         int push_items = 0;
1189         int src_nritems;
1190         int dst_nritems;
1191         int ret = 0;
1192
1193         src_nritems = btrfs_header_nritems(src);
1194         dst_nritems = btrfs_header_nritems(dst);
1195         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1196         WARN_ON(btrfs_header_generation(src) != trans->transid);
1197         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1198
1199         if (!empty && src_nritems <= 8)
1200                 return 1;
1201
1202         if (push_items <= 0) {
1203                 return 1;
1204         }
1205
1206         if (empty) {
1207                 push_items = min(src_nritems, push_items);
1208                 if (push_items < src_nritems) {
1209                         /* leave at least 8 pointers in the node if
1210                          * we aren't going to empty it
1211                          */
1212                         if (src_nritems - push_items < 8) {
1213                                 if (push_items <= 8)
1214                                         return 1;
1215                                 push_items -= 8;
1216                         }
1217                 }
1218         } else
1219                 push_items = min(src_nritems - 8, push_items);
1220
1221         copy_extent_buffer(dst, src,
1222                            btrfs_node_key_ptr_offset(dst_nritems),
1223                            btrfs_node_key_ptr_offset(0),
1224                            push_items * sizeof(struct btrfs_key_ptr));
1225
1226         if (push_items < src_nritems) {
1227                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1228                                       btrfs_node_key_ptr_offset(push_items),
1229                                       (src_nritems - push_items) *
1230                                       sizeof(struct btrfs_key_ptr));
1231         }
1232         btrfs_set_header_nritems(src, src_nritems - push_items);
1233         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1234         btrfs_mark_buffer_dirty(src);
1235         btrfs_mark_buffer_dirty(dst);
1236
1237         return ret;
1238 }
1239
1240 /*
1241  * try to push data from one node into the next node right in the
1242  * tree.
1243  *
1244  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1245  * error, and > 0 if there was no room in the right hand block.
1246  *
1247  * this will  only push up to 1/2 the contents of the left node over
1248  */
1249 static int balance_node_right(struct btrfs_trans_handle *trans,
1250                               struct btrfs_root *root,
1251                               struct extent_buffer *dst,
1252                               struct extent_buffer *src)
1253 {
1254         int push_items = 0;
1255         int max_push;
1256         int src_nritems;
1257         int dst_nritems;
1258         int ret = 0;
1259
1260         WARN_ON(btrfs_header_generation(src) != trans->transid);
1261         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1262
1263         src_nritems = btrfs_header_nritems(src);
1264         dst_nritems = btrfs_header_nritems(dst);
1265         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1266         if (push_items <= 0) {
1267                 return 1;
1268         }
1269
1270         if (src_nritems < 4) {
1271                 return 1;
1272         }
1273
1274         max_push = src_nritems / 2 + 1;
1275         /* don't try to empty the node */
1276         if (max_push >= src_nritems) {
1277                 return 1;
1278         }
1279
1280         if (max_push < push_items)
1281                 push_items = max_push;
1282
1283         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1284                                       btrfs_node_key_ptr_offset(0),
1285                                       (dst_nritems) *
1286                                       sizeof(struct btrfs_key_ptr));
1287
1288         copy_extent_buffer(dst, src,
1289                            btrfs_node_key_ptr_offset(0),
1290                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1291                            push_items * sizeof(struct btrfs_key_ptr));
1292
1293         btrfs_set_header_nritems(src, src_nritems - push_items);
1294         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1295
1296         btrfs_mark_buffer_dirty(src);
1297         btrfs_mark_buffer_dirty(dst);
1298
1299         return ret;
1300 }
1301
1302 /*
1303  * helper function to insert a new root level in the tree.
1304  * A new node is allocated, and a single item is inserted to
1305  * point to the existing root
1306  *
1307  * returns zero on success or < 0 on failure.
1308  */
1309 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1310                            struct btrfs_root *root,
1311                            struct btrfs_path *path, int level)
1312 {
1313         u64 lower_gen;
1314         struct extent_buffer *lower;
1315         struct extent_buffer *c;
1316         struct extent_buffer *old;
1317         struct btrfs_disk_key lower_key;
1318
1319         BUG_ON(path->nodes[level]);
1320         BUG_ON(path->nodes[level-1] != root->node);
1321
1322         lower = path->nodes[level-1];
1323         if (level == 1)
1324                 btrfs_item_key(lower, &lower_key, 0);
1325         else
1326                 btrfs_node_key(lower, &lower_key, 0);
1327
1328         c = btrfs_alloc_free_block(trans, root, root->nodesize,
1329                                    root->root_key.objectid, &lower_key,
1330                                    level, root->node->start, 0);
1331
1332         if (IS_ERR(c))
1333                 return PTR_ERR(c);
1334
1335         memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
1336         btrfs_set_header_nritems(c, 1);
1337         btrfs_set_header_level(c, level);
1338         btrfs_set_header_bytenr(c, c->start);
1339         btrfs_set_header_generation(c, trans->transid);
1340         btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
1341         btrfs_set_header_owner(c, root->root_key.objectid);
1342
1343         write_extent_buffer(c, root->fs_info->fsid,
1344                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
1345
1346         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1347                             btrfs_header_chunk_tree_uuid(c),
1348                             BTRFS_UUID_SIZE);
1349
1350         btrfs_set_node_key(c, &lower_key, 0);
1351         btrfs_set_node_blockptr(c, 0, lower->start);
1352         lower_gen = btrfs_header_generation(lower);
1353         WARN_ON(lower_gen != trans->transid);
1354
1355         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1356
1357         btrfs_mark_buffer_dirty(c);
1358
1359         old = root->node;
1360         root->node = c;
1361
1362         /* the super has an extra ref to root->node */
1363         free_extent_buffer(old);
1364
1365         add_root_to_dirty_list(root);
1366         extent_buffer_get(c);
1367         path->nodes[level] = c;
1368         path->slots[level] = 0;
1369         return 0;
1370 }
1371
1372 /*
1373  * worker function to insert a single pointer in a node.
1374  * the node should have enough room for the pointer already
1375  *
1376  * slot and level indicate where you want the key to go, and
1377  * blocknr is the block the key points to.
1378  *
1379  * returns zero on success and < 0 on any error
1380  */
1381 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1382                       *root, struct btrfs_path *path, struct btrfs_disk_key
1383                       *key, u64 bytenr, int slot, int level)
1384 {
1385         struct extent_buffer *lower;
1386         int nritems;
1387
1388         BUG_ON(!path->nodes[level]);
1389         lower = path->nodes[level];
1390         nritems = btrfs_header_nritems(lower);
1391         if (slot > nritems)
1392                 BUG();
1393         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1394                 BUG();
1395         if (slot != nritems) {
1396                 memmove_extent_buffer(lower,
1397                               btrfs_node_key_ptr_offset(slot + 1),
1398                               btrfs_node_key_ptr_offset(slot),
1399                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1400         }
1401         btrfs_set_node_key(lower, key, slot);
1402         btrfs_set_node_blockptr(lower, slot, bytenr);
1403         WARN_ON(trans->transid == 0);
1404         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1405         btrfs_set_header_nritems(lower, nritems + 1);
1406         btrfs_mark_buffer_dirty(lower);
1407         return 0;
1408 }
1409
1410 /*
1411  * split the node at the specified level in path in two.
1412  * The path is corrected to point to the appropriate node after the split
1413  *
1414  * Before splitting this tries to make some room in the node by pushing
1415  * left and right, if either one works, it returns right away.
1416  *
1417  * returns 0 on success and < 0 on failure
1418  */
1419 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1420                       *root, struct btrfs_path *path, int level)
1421 {
1422         struct extent_buffer *c;
1423         struct extent_buffer *split;
1424         struct btrfs_disk_key disk_key;
1425         int mid;
1426         int ret;
1427         int wret;
1428         u32 c_nritems;
1429
1430         c = path->nodes[level];
1431         WARN_ON(btrfs_header_generation(c) != trans->transid);
1432         if (c == root->node) {
1433                 /* trying to split the root, lets make a new one */
1434                 ret = insert_new_root(trans, root, path, level + 1);
1435                 if (ret)
1436                         return ret;
1437         } else {
1438                 ret = push_nodes_for_insert(trans, root, path, level);
1439                 c = path->nodes[level];
1440                 if (!ret && btrfs_header_nritems(c) <
1441                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
1442                         return 0;
1443                 if (ret < 0)
1444                         return ret;
1445         }
1446
1447         c_nritems = btrfs_header_nritems(c);
1448         mid = (c_nritems + 1) / 2;
1449         btrfs_node_key(c, &disk_key, mid);
1450
1451         split = btrfs_alloc_free_block(trans, root, root->nodesize,
1452                                         root->root_key.objectid,
1453                                         &disk_key, level, c->start, 0);
1454         if (IS_ERR(split))
1455                 return PTR_ERR(split);
1456
1457         memset_extent_buffer(split, 0, 0, sizeof(struct btrfs_header));
1458         btrfs_set_header_level(split, btrfs_header_level(c));
1459         btrfs_set_header_bytenr(split, split->start);
1460         btrfs_set_header_generation(split, trans->transid);
1461         btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
1462         btrfs_set_header_owner(split, root->root_key.objectid);
1463         write_extent_buffer(split, root->fs_info->fsid,
1464                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
1465         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1466                             btrfs_header_chunk_tree_uuid(split),
1467                             BTRFS_UUID_SIZE);
1468
1469
1470         copy_extent_buffer(split, c,
1471                            btrfs_node_key_ptr_offset(0),
1472                            btrfs_node_key_ptr_offset(mid),
1473                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1474         btrfs_set_header_nritems(split, c_nritems - mid);
1475         btrfs_set_header_nritems(c, mid);
1476         ret = 0;
1477
1478         btrfs_mark_buffer_dirty(c);
1479         btrfs_mark_buffer_dirty(split);
1480
1481         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1482                           path->slots[level + 1] + 1,
1483                           level + 1);
1484         if (wret)
1485                 ret = wret;
1486
1487         if (path->slots[level] >= mid) {
1488                 path->slots[level] -= mid;
1489                 free_extent_buffer(c);
1490                 path->nodes[level] = split;
1491                 path->slots[level + 1] += 1;
1492         } else {
1493                 free_extent_buffer(split);
1494         }
1495         return ret;
1496 }
1497
1498 /*
1499  * how many bytes are required to store the items in a leaf.  start
1500  * and nr indicate which items in the leaf to check.  This totals up the
1501  * space used both by the item structs and the item data
1502  */
1503 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1504 {
1505         int data_len;
1506         int nritems = btrfs_header_nritems(l);
1507         int end = min(nritems, start + nr) - 1;
1508
1509         if (!nr)
1510                 return 0;
1511         data_len = btrfs_item_end_nr(l, start);
1512         data_len = data_len - btrfs_item_offset_nr(l, end);
1513         data_len += sizeof(struct btrfs_item) * nr;
1514         WARN_ON(data_len < 0);
1515         return data_len;
1516 }
1517
1518 /*
1519  * The space between the end of the leaf items and
1520  * the start of the leaf data.  IOW, how much room
1521  * the leaf has left for both items and data
1522  */
1523 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1524 {
1525         int nritems = btrfs_header_nritems(leaf);
1526         int ret;
1527         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1528         if (ret < 0) {
1529                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1530                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1531                        leaf_space_used(leaf, 0, nritems), nritems);
1532         }
1533         return ret;
1534 }
1535
1536 /*
1537  * push some data in the path leaf to the right, trying to free up at
1538  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1539  *
1540  * returns 1 if the push failed because the other node didn't have enough
1541  * room, 0 if everything worked out and < 0 if there were major errors.
1542  */
1543 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1544                            *root, struct btrfs_path *path, int data_size,
1545                            int empty)
1546 {
1547         struct extent_buffer *left = path->nodes[0];
1548         struct extent_buffer *right;
1549         struct extent_buffer *upper;
1550         struct btrfs_disk_key disk_key;
1551         int slot;
1552         u32 i;
1553         int free_space;
1554         int push_space = 0;
1555         int push_items = 0;
1556         struct btrfs_item *item;
1557         u32 left_nritems;
1558         u32 nr;
1559         u32 right_nritems;
1560         u32 data_end;
1561         u32 this_item_size;
1562         int ret;
1563
1564         slot = path->slots[1];
1565         if (!path->nodes[1]) {
1566                 return 1;
1567         }
1568         upper = path->nodes[1];
1569         if (slot >= btrfs_header_nritems(upper) - 1)
1570                 return 1;
1571
1572         right = read_node_slot(root, upper, slot + 1);
1573         free_space = btrfs_leaf_free_space(root, right);
1574         if (free_space < data_size) {
1575                 free_extent_buffer(right);
1576                 return 1;
1577         }
1578
1579         /* cow and double check */
1580         ret = btrfs_cow_block(trans, root, right, upper,
1581                               slot + 1, &right);
1582         if (ret) {
1583                 free_extent_buffer(right);
1584                 return 1;
1585         }
1586         free_space = btrfs_leaf_free_space(root, right);
1587         if (free_space < data_size) {
1588                 free_extent_buffer(right);
1589                 return 1;
1590         }
1591
1592         left_nritems = btrfs_header_nritems(left);
1593         if (left_nritems == 0) {
1594                 free_extent_buffer(right);
1595                 return 1;
1596         }
1597
1598         if (empty)
1599                 nr = 0;
1600         else
1601                 nr = 1;
1602
1603         i = left_nritems - 1;
1604         while (i >= nr) {
1605                 item = btrfs_item_nr(i);
1606
1607                 if (path->slots[0] == i)
1608                         push_space += data_size + sizeof(*item);
1609
1610                 this_item_size = btrfs_item_size(left, item);
1611                 if (this_item_size + sizeof(*item) + push_space > free_space)
1612                         break;
1613                 push_items++;
1614                 push_space += this_item_size + sizeof(*item);
1615                 if (i == 0)
1616                         break;
1617                 i--;
1618         }
1619
1620         if (push_items == 0) {
1621                 free_extent_buffer(right);
1622                 return 1;
1623         }
1624
1625         if (!empty && push_items == left_nritems)
1626                 WARN_ON(1);
1627
1628         /* push left to right */
1629         right_nritems = btrfs_header_nritems(right);
1630
1631         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1632         push_space -= leaf_data_end(root, left);
1633
1634         /* make room in the right data area */
1635         data_end = leaf_data_end(root, right);
1636         memmove_extent_buffer(right,
1637                               btrfs_leaf_data(right) + data_end - push_space,
1638                               btrfs_leaf_data(right) + data_end,
1639                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1640
1641         /* copy from the left data area */
1642         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1643                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1644                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1645                      push_space);
1646
1647         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1648                               btrfs_item_nr_offset(0),
1649                               right_nritems * sizeof(struct btrfs_item));
1650
1651         /* copy the items from left to right */
1652         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1653                    btrfs_item_nr_offset(left_nritems - push_items),
1654                    push_items * sizeof(struct btrfs_item));
1655
1656         /* update the item pointers */
1657         right_nritems += push_items;
1658         btrfs_set_header_nritems(right, right_nritems);
1659         push_space = BTRFS_LEAF_DATA_SIZE(root);
1660         for (i = 0; i < right_nritems; i++) {
1661                 item = btrfs_item_nr(i);
1662                 push_space -= btrfs_item_size(right, item);
1663                 btrfs_set_item_offset(right, item, push_space);
1664         }
1665
1666         left_nritems -= push_items;
1667         btrfs_set_header_nritems(left, left_nritems);
1668
1669         if (left_nritems)
1670                 btrfs_mark_buffer_dirty(left);
1671         btrfs_mark_buffer_dirty(right);
1672
1673         btrfs_item_key(right, &disk_key, 0);
1674         btrfs_set_node_key(upper, &disk_key, slot + 1);
1675         btrfs_mark_buffer_dirty(upper);
1676
1677         /* then fixup the leaf pointer in the path */
1678         if (path->slots[0] >= left_nritems) {
1679                 path->slots[0] -= left_nritems;
1680                 free_extent_buffer(path->nodes[0]);
1681                 path->nodes[0] = right;
1682                 path->slots[1] += 1;
1683         } else {
1684                 free_extent_buffer(right);
1685         }
1686         return 0;
1687 }
1688 /*
1689  * push some data in the path leaf to the left, trying to free up at
1690  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1691  */
1692 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1693                           *root, struct btrfs_path *path, int data_size,
1694                           int empty)
1695 {
1696         struct btrfs_disk_key disk_key;
1697         struct extent_buffer *right = path->nodes[0];
1698         struct extent_buffer *left;
1699         int slot;
1700         int i;
1701         int free_space;
1702         int push_space = 0;
1703         int push_items = 0;
1704         struct btrfs_item *item;
1705         u32 old_left_nritems;
1706         u32 right_nritems;
1707         u32 nr;
1708         int ret = 0;
1709         int wret;
1710         u32 this_item_size;
1711         u32 old_left_item_size;
1712
1713         slot = path->slots[1];
1714         if (slot == 0)
1715                 return 1;
1716         if (!path->nodes[1])
1717                 return 1;
1718
1719         right_nritems = btrfs_header_nritems(right);
1720         if (right_nritems == 0) {
1721                 return 1;
1722         }
1723
1724         left = read_node_slot(root, path->nodes[1], slot - 1);
1725         free_space = btrfs_leaf_free_space(root, left);
1726         if (free_space < data_size) {
1727                 free_extent_buffer(left);
1728                 return 1;
1729         }
1730
1731         /* cow and double check */
1732         ret = btrfs_cow_block(trans, root, left,
1733                               path->nodes[1], slot - 1, &left);
1734         if (ret) {
1735                 /* we hit -ENOSPC, but it isn't fatal here */
1736                 free_extent_buffer(left);
1737                 return 1;
1738         }
1739
1740         free_space = btrfs_leaf_free_space(root, left);
1741         if (free_space < data_size) {
1742                 free_extent_buffer(left);
1743                 return 1;
1744         }
1745
1746         if (empty)
1747                 nr = right_nritems;
1748         else
1749                 nr = right_nritems - 1;
1750
1751         for (i = 0; i < nr; i++) {
1752                 item = btrfs_item_nr(i);
1753
1754                 if (path->slots[0] == i)
1755                         push_space += data_size + sizeof(*item);
1756
1757                 this_item_size = btrfs_item_size(right, item);
1758                 if (this_item_size + sizeof(*item) + push_space > free_space)
1759                         break;
1760
1761                 push_items++;
1762                 push_space += this_item_size + sizeof(*item);
1763         }
1764
1765         if (push_items == 0) {
1766                 free_extent_buffer(left);
1767                 return 1;
1768         }
1769         if (!empty && push_items == btrfs_header_nritems(right))
1770                 WARN_ON(1);
1771
1772         /* push data from right to left */
1773         copy_extent_buffer(left, right,
1774                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1775                            btrfs_item_nr_offset(0),
1776                            push_items * sizeof(struct btrfs_item));
1777
1778         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1779                      btrfs_item_offset_nr(right, push_items -1);
1780
1781         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1782                      leaf_data_end(root, left) - push_space,
1783                      btrfs_leaf_data(right) +
1784                      btrfs_item_offset_nr(right, push_items - 1),
1785                      push_space);
1786         old_left_nritems = btrfs_header_nritems(left);
1787         BUG_ON(old_left_nritems == 0);
1788
1789         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1790         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1791                 u32 ioff;
1792
1793                 item = btrfs_item_nr(i);
1794                 ioff = btrfs_item_offset(left, item);
1795                 btrfs_set_item_offset(left, item,
1796                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1797         }
1798         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1799
1800         /* fixup right node */
1801         if (push_items > right_nritems) {
1802                 printk("push items %d nr %u\n", push_items, right_nritems);
1803                 WARN_ON(1);
1804         }
1805
1806         if (push_items < right_nritems) {
1807                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1808                                                   leaf_data_end(root, right);
1809                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
1810                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1811                                       btrfs_leaf_data(right) +
1812                                       leaf_data_end(root, right), push_space);
1813
1814                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1815                               btrfs_item_nr_offset(push_items),
1816                              (btrfs_header_nritems(right) - push_items) *
1817                              sizeof(struct btrfs_item));
1818         }
1819         right_nritems -= push_items;
1820         btrfs_set_header_nritems(right, right_nritems);
1821         push_space = BTRFS_LEAF_DATA_SIZE(root);
1822         for (i = 0; i < right_nritems; i++) {
1823                 item = btrfs_item_nr(i);
1824                 push_space = push_space - btrfs_item_size(right, item);
1825                 btrfs_set_item_offset(right, item, push_space);
1826         }
1827
1828         btrfs_mark_buffer_dirty(left);
1829         if (right_nritems)
1830                 btrfs_mark_buffer_dirty(right);
1831
1832         btrfs_item_key(right, &disk_key, 0);
1833         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1834         if (wret)
1835                 ret = wret;
1836
1837         /* then fixup the leaf pointer in the path */
1838         if (path->slots[0] < push_items) {
1839                 path->slots[0] += old_left_nritems;
1840                 free_extent_buffer(path->nodes[0]);
1841                 path->nodes[0] = left;
1842                 path->slots[1] -= 1;
1843         } else {
1844                 free_extent_buffer(left);
1845                 path->slots[0] -= push_items;
1846         }
1847         BUG_ON(path->slots[0] < 0);
1848         return ret;
1849 }
1850
1851 /*
1852  * split the path's leaf in two, making sure there is at least data_size
1853  * available for the resulting leaf level of the path.
1854  *
1855  * returns 0 if all went well and < 0 on failure.
1856  */
1857 static noinline int copy_for_split(struct btrfs_trans_handle *trans,
1858                                struct btrfs_root *root,
1859                                struct btrfs_path *path,
1860                                struct extent_buffer *l,
1861                                struct extent_buffer *right,
1862                                int slot, int mid, int nritems)
1863 {
1864         int data_copy_size;
1865         int rt_data_off;
1866         int i;
1867         int ret = 0;
1868         int wret;
1869         struct btrfs_disk_key disk_key;
1870
1871         nritems = nritems - mid;
1872         btrfs_set_header_nritems(right, nritems);
1873         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
1874
1875         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
1876                            btrfs_item_nr_offset(mid),
1877                            nritems * sizeof(struct btrfs_item));
1878
1879         copy_extent_buffer(right, l,
1880                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
1881                      data_copy_size, btrfs_leaf_data(l) +
1882                      leaf_data_end(root, l), data_copy_size);
1883
1884         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
1885                       btrfs_item_end_nr(l, mid);
1886
1887         for (i = 0; i < nritems; i++) {
1888                 struct btrfs_item *item = btrfs_item_nr(i);
1889                 u32 ioff = btrfs_item_offset(right, item);
1890                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
1891         }
1892
1893         btrfs_set_header_nritems(l, mid);
1894         ret = 0;
1895         btrfs_item_key(right, &disk_key, 0);
1896         wret = insert_ptr(trans, root, path, &disk_key, right->start,
1897                           path->slots[1] + 1, 1);
1898         if (wret)
1899                 ret = wret;
1900
1901         btrfs_mark_buffer_dirty(right);
1902         btrfs_mark_buffer_dirty(l);
1903         BUG_ON(path->slots[0] != slot);
1904
1905         if (mid <= slot) {
1906                 free_extent_buffer(path->nodes[0]);
1907                 path->nodes[0] = right;
1908                 path->slots[0] -= mid;
1909                 path->slots[1] += 1;
1910         } else {
1911                 free_extent_buffer(right);
1912         }
1913
1914         BUG_ON(path->slots[0] < 0);
1915
1916         return ret;
1917 }
1918
1919 /*
1920  * split the path's leaf in two, making sure there is at least data_size
1921  * available for the resulting leaf level of the path.
1922  *
1923  * returns 0 if all went well and < 0 on failure.
1924  */
1925 static noinline int split_leaf(struct btrfs_trans_handle *trans,
1926                                struct btrfs_root *root,
1927                                struct btrfs_key *ins_key,
1928                                struct btrfs_path *path, int data_size,
1929                                int extend)
1930 {
1931         struct btrfs_disk_key disk_key;
1932         struct extent_buffer *l;
1933         u32 nritems;
1934         int mid;
1935         int slot;
1936         struct extent_buffer *right;
1937         int ret = 0;
1938         int wret;
1939         int split;
1940         int num_doubles = 0;
1941
1942         /* first try to make some room by pushing left and right */
1943         if (data_size && ins_key->type != BTRFS_DIR_ITEM_KEY) {
1944                 wret = push_leaf_right(trans, root, path, data_size, 0);
1945                 if (wret < 0)
1946                         return wret;
1947                 if (wret) {
1948                         wret = push_leaf_left(trans, root, path, data_size, 0);
1949                         if (wret < 0)
1950                                 return wret;
1951                 }
1952                 l = path->nodes[0];
1953
1954                 /* did the pushes work? */
1955                 if (btrfs_leaf_free_space(root, l) >= data_size)
1956                         return 0;
1957         }
1958
1959         if (!path->nodes[1]) {
1960                 ret = insert_new_root(trans, root, path, 1);
1961                 if (ret)
1962                         return ret;
1963         }
1964 again:
1965         split = 1;
1966         l = path->nodes[0];
1967         slot = path->slots[0];
1968         nritems = btrfs_header_nritems(l);
1969         mid = (nritems + 1) / 2;
1970
1971         if (mid <= slot) {
1972                 if (nritems == 1 ||
1973                     leaf_space_used(l, mid, nritems - mid) + data_size >
1974                         BTRFS_LEAF_DATA_SIZE(root)) {
1975                         if (slot >= nritems) {
1976                                 split = 0;
1977                         } else {
1978                                 mid = slot;
1979                                 if (mid != nritems &&
1980                                     leaf_space_used(l, mid, nritems - mid) +
1981                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
1982                                         split = 2;
1983                                 }
1984                         }
1985                 }
1986         } else {
1987                 if (leaf_space_used(l, 0, mid) + data_size >
1988                         BTRFS_LEAF_DATA_SIZE(root)) {
1989                         if (!extend && data_size && slot == 0) {
1990                                 split = 0;
1991                         } else if ((extend || !data_size) && slot == 0) {
1992                                 mid = 1;
1993                         } else {
1994                                 mid = slot;
1995                                 if (mid != nritems &&
1996                                     leaf_space_used(l, mid, nritems - mid) +
1997                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
1998                                         split = 2 ;
1999                                 }
2000                         }
2001                 }
2002         }
2003
2004         if (split == 0)
2005                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2006         else
2007                 btrfs_item_key(l, &disk_key, mid);
2008
2009         right = btrfs_alloc_free_block(trans, root, root->leafsize,
2010                                         root->root_key.objectid,
2011                                         &disk_key, 0, l->start, 0);
2012         if (IS_ERR(right)) {
2013                 BUG_ON(1);
2014                 return PTR_ERR(right);
2015         }
2016
2017         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2018         btrfs_set_header_bytenr(right, right->start);
2019         btrfs_set_header_generation(right, trans->transid);
2020         btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
2021         btrfs_set_header_owner(right, root->root_key.objectid);
2022         btrfs_set_header_level(right, 0);
2023         write_extent_buffer(right, root->fs_info->fsid,
2024                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
2025
2026         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2027                             btrfs_header_chunk_tree_uuid(right),
2028                             BTRFS_UUID_SIZE);
2029
2030         if (split == 0) {
2031                 if (mid <= slot) {
2032                         btrfs_set_header_nritems(right, 0);
2033                         wret = insert_ptr(trans, root, path,
2034                                           &disk_key, right->start,
2035                                           path->slots[1] + 1, 1);
2036                         if (wret)
2037                                 ret = wret;
2038
2039                         free_extent_buffer(path->nodes[0]);
2040                         path->nodes[0] = right;
2041                         path->slots[0] = 0;
2042                         path->slots[1] += 1;
2043                 } else {
2044                         btrfs_set_header_nritems(right, 0);
2045                         wret = insert_ptr(trans, root, path,
2046                                           &disk_key,
2047                                           right->start,
2048                                           path->slots[1], 1);
2049                         if (wret)
2050                                 ret = wret;
2051                         free_extent_buffer(path->nodes[0]);
2052                         path->nodes[0] = right;
2053                         path->slots[0] = 0;
2054                         if (path->slots[1] == 0) {
2055                                 wret = fixup_low_keys(trans, root,
2056                                                 path, &disk_key, 1);
2057                                 if (wret)
2058                                         ret = wret;
2059                         }
2060                 }
2061                 btrfs_mark_buffer_dirty(right);
2062                 return ret;
2063         }
2064
2065         ret = copy_for_split(trans, root, path, l, right, slot, mid, nritems);
2066         BUG_ON(ret);
2067
2068         if (split == 2) {
2069                 BUG_ON(num_doubles != 0);
2070                 num_doubles++;
2071                 goto again;
2072         }
2073
2074         return ret;
2075 }
2076
2077 /*
2078  * This function splits a single item into two items,
2079  * giving 'new_key' to the new item and splitting the
2080  * old one at split_offset (from the start of the item).
2081  *
2082  * The path may be released by this operation.  After
2083  * the split, the path is pointing to the old item.  The
2084  * new item is going to be in the same node as the old one.
2085  *
2086  * Note, the item being split must be smaller enough to live alone on
2087  * a tree block with room for one extra struct btrfs_item
2088  *
2089  * This allows us to split the item in place, keeping a lock on the
2090  * leaf the entire time.
2091  */
2092 int btrfs_split_item(struct btrfs_trans_handle *trans,
2093                      struct btrfs_root *root,
2094                      struct btrfs_path *path,
2095                      struct btrfs_key *new_key,
2096                      unsigned long split_offset)
2097 {
2098         u32 item_size;
2099         struct extent_buffer *leaf;
2100         struct btrfs_key orig_key;
2101         struct btrfs_item *item;
2102         struct btrfs_item *new_item;
2103         int ret = 0;
2104         int slot;
2105         u32 nritems;
2106         u32 orig_offset;
2107         struct btrfs_disk_key disk_key;
2108         char *buf;
2109
2110         leaf = path->nodes[0];
2111         btrfs_item_key_to_cpu(leaf, &orig_key, path->slots[0]);
2112         if (btrfs_leaf_free_space(root, leaf) >= sizeof(struct btrfs_item))
2113                 goto split;
2114
2115         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2116         btrfs_release_path(path);
2117
2118         path->search_for_split = 1;
2119
2120         ret = btrfs_search_slot(trans, root, &orig_key, path, 0, 1);
2121         path->search_for_split = 0;
2122
2123         /* if our item isn't there or got smaller, return now */
2124         if (ret != 0 || item_size != btrfs_item_size_nr(path->nodes[0],
2125                                                         path->slots[0])) {
2126                 return -EAGAIN;
2127         }
2128
2129         ret = split_leaf(trans, root, &orig_key, path, 0, 0);
2130         BUG_ON(ret);
2131
2132         BUG_ON(btrfs_leaf_free_space(root, leaf) < sizeof(struct btrfs_item));
2133         leaf = path->nodes[0];
2134
2135 split:
2136         item = btrfs_item_nr(path->slots[0]);
2137         orig_offset = btrfs_item_offset(leaf, item);
2138         item_size = btrfs_item_size(leaf, item);
2139
2140
2141         buf = kmalloc(item_size, GFP_NOFS);
2142         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
2143                             path->slots[0]), item_size);
2144         slot = path->slots[0] + 1;
2145         leaf = path->nodes[0];
2146
2147         nritems = btrfs_header_nritems(leaf);
2148
2149         if (slot != nritems) {
2150                 /* shift the items */
2151                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
2152                               btrfs_item_nr_offset(slot),
2153                               (nritems - slot) * sizeof(struct btrfs_item));
2154
2155         }
2156
2157         btrfs_cpu_key_to_disk(&disk_key, new_key);
2158         btrfs_set_item_key(leaf, &disk_key, slot);
2159
2160         new_item = btrfs_item_nr(slot);
2161
2162         btrfs_set_item_offset(leaf, new_item, orig_offset);
2163         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
2164
2165         btrfs_set_item_offset(leaf, item,
2166                               orig_offset + item_size - split_offset);
2167         btrfs_set_item_size(leaf, item, split_offset);
2168
2169         btrfs_set_header_nritems(leaf, nritems + 1);
2170
2171         /* write the data for the start of the original item */
2172         write_extent_buffer(leaf, buf,
2173                             btrfs_item_ptr_offset(leaf, path->slots[0]),
2174                             split_offset);
2175
2176         /* write the data for the new item */
2177         write_extent_buffer(leaf, buf + split_offset,
2178                             btrfs_item_ptr_offset(leaf, slot),
2179                             item_size - split_offset);
2180         btrfs_mark_buffer_dirty(leaf);
2181
2182         ret = 0;
2183         if (btrfs_leaf_free_space(root, leaf) < 0) {
2184                 btrfs_print_leaf(root, leaf);
2185                 BUG();
2186         }
2187         kfree(buf);
2188         return ret;
2189 }
2190
2191 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2192                         struct btrfs_root *root,
2193                         struct btrfs_path *path,
2194                         u32 new_size, int from_end)
2195 {
2196         int ret = 0;
2197         int slot;
2198         struct extent_buffer *leaf;
2199         struct btrfs_item *item;
2200         u32 nritems;
2201         unsigned int data_end;
2202         unsigned int old_data_start;
2203         unsigned int old_size;
2204         unsigned int size_diff;
2205         int i;
2206
2207         leaf = path->nodes[0];
2208         slot = path->slots[0];
2209
2210         old_size = btrfs_item_size_nr(leaf, slot);
2211         if (old_size == new_size)
2212                 return 0;
2213
2214         nritems = btrfs_header_nritems(leaf);
2215         data_end = leaf_data_end(root, leaf);
2216
2217         old_data_start = btrfs_item_offset_nr(leaf, slot);
2218
2219         size_diff = old_size - new_size;
2220
2221         BUG_ON(slot < 0);
2222         BUG_ON(slot >= nritems);
2223
2224         /*
2225          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2226          */
2227         /* first correct the data pointers */
2228         for (i = slot; i < nritems; i++) {
2229                 u32 ioff;
2230                 item = btrfs_item_nr(i);
2231                 ioff = btrfs_item_offset(leaf, item);
2232                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2233         }
2234
2235         /* shift the data */
2236         if (from_end) {
2237                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2238                               data_end + size_diff, btrfs_leaf_data(leaf) +
2239                               data_end, old_data_start + new_size - data_end);
2240         } else {
2241                 struct btrfs_disk_key disk_key;
2242                 u64 offset;
2243
2244                 btrfs_item_key(leaf, &disk_key, slot);
2245
2246                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2247                         unsigned long ptr;
2248                         struct btrfs_file_extent_item *fi;
2249
2250                         fi = btrfs_item_ptr(leaf, slot,
2251                                             struct btrfs_file_extent_item);
2252                         fi = (struct btrfs_file_extent_item *)(
2253                              (unsigned long)fi - size_diff);
2254
2255                         if (btrfs_file_extent_type(leaf, fi) ==
2256                             BTRFS_FILE_EXTENT_INLINE) {
2257                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2258                                 memmove_extent_buffer(leaf, ptr,
2259                                         (unsigned long)fi,
2260                                         offsetof(struct btrfs_file_extent_item,
2261                                                  disk_bytenr));
2262                         }
2263                 }
2264
2265                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2266                               data_end + size_diff, btrfs_leaf_data(leaf) +
2267                               data_end, old_data_start - data_end);
2268
2269                 offset = btrfs_disk_key_offset(&disk_key);
2270                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2271                 btrfs_set_item_key(leaf, &disk_key, slot);
2272                 if (slot == 0)
2273                         fixup_low_keys(trans, root, path, &disk_key, 1);
2274         }
2275
2276         item = btrfs_item_nr(slot);
2277         btrfs_set_item_size(leaf, item, new_size);
2278         btrfs_mark_buffer_dirty(leaf);
2279
2280         ret = 0;
2281         if (btrfs_leaf_free_space(root, leaf) < 0) {
2282                 btrfs_print_leaf(root, leaf);
2283                 BUG();
2284         }
2285         return ret;
2286 }
2287
2288 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2289                       struct btrfs_root *root, struct btrfs_path *path,
2290                       u32 data_size)
2291 {
2292         int ret = 0;
2293         int slot;
2294         struct extent_buffer *leaf;
2295         struct btrfs_item *item;
2296         u32 nritems;
2297         unsigned int data_end;
2298         unsigned int old_data;
2299         unsigned int old_size;
2300         int i;
2301
2302         leaf = path->nodes[0];
2303
2304         nritems = btrfs_header_nritems(leaf);
2305         data_end = leaf_data_end(root, leaf);
2306
2307         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2308                 btrfs_print_leaf(root, leaf);
2309                 BUG();
2310         }
2311         slot = path->slots[0];
2312         old_data = btrfs_item_end_nr(leaf, slot);
2313
2314         BUG_ON(slot < 0);
2315         if (slot >= nritems) {
2316                 btrfs_print_leaf(root, leaf);
2317                 printk("slot %d too large, nritems %d\n", slot, nritems);
2318                 BUG_ON(1);
2319         }
2320
2321         /*
2322          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2323          */
2324         /* first correct the data pointers */
2325         for (i = slot; i < nritems; i++) {
2326                 u32 ioff;
2327                 item = btrfs_item_nr(i);
2328                 ioff = btrfs_item_offset(leaf, item);
2329                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2330         }
2331
2332         /* shift the data */
2333         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2334                       data_end - data_size, btrfs_leaf_data(leaf) +
2335                       data_end, old_data - data_end);
2336
2337         data_end = old_data;
2338         old_size = btrfs_item_size_nr(leaf, slot);
2339         item = btrfs_item_nr(slot);
2340         btrfs_set_item_size(leaf, item, old_size + data_size);
2341         btrfs_mark_buffer_dirty(leaf);
2342
2343         ret = 0;
2344         if (btrfs_leaf_free_space(root, leaf) < 0) {
2345                 btrfs_print_leaf(root, leaf);
2346                 BUG();
2347         }
2348         return ret;
2349 }
2350
2351 /*
2352  * Given a key and some data, insert an item into the tree.
2353  * This does all the path init required, making room in the tree if needed.
2354  */
2355 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2356                             struct btrfs_root *root,
2357                             struct btrfs_path *path,
2358                             struct btrfs_key *cpu_key, u32 *data_size,
2359                             int nr)
2360 {
2361         struct extent_buffer *leaf;
2362         struct btrfs_item *item;
2363         int ret = 0;
2364         int slot;
2365         int i;
2366         u32 nritems;
2367         u32 total_size = 0;
2368         u32 total_data = 0;
2369         unsigned int data_end;
2370         struct btrfs_disk_key disk_key;
2371
2372         for (i = 0; i < nr; i++) {
2373                 total_data += data_size[i];
2374         }
2375
2376         /* create a root if there isn't one */
2377         if (!root->node)
2378                 BUG();
2379
2380         total_size = total_data + nr * sizeof(struct btrfs_item);
2381         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2382         if (ret == 0) {
2383                 return -EEXIST;
2384         }
2385         if (ret < 0)
2386                 goto out;
2387
2388         leaf = path->nodes[0];
2389
2390         nritems = btrfs_header_nritems(leaf);
2391         data_end = leaf_data_end(root, leaf);
2392
2393         if (btrfs_leaf_free_space(root, leaf) < total_size) {
2394                 btrfs_print_leaf(root, leaf);
2395                 printk("not enough freespace need %u have %d\n",
2396                        total_size, btrfs_leaf_free_space(root, leaf));
2397                 BUG();
2398         }
2399
2400         slot = path->slots[0];
2401         BUG_ON(slot < 0);
2402
2403         if (slot != nritems) {
2404                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2405
2406                 if (old_data < data_end) {
2407                         btrfs_print_leaf(root, leaf);
2408                         printk("slot %d old_data %d data_end %d\n",
2409                                slot, old_data, data_end);
2410                         BUG_ON(1);
2411                 }
2412                 /*
2413                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2414                  */
2415                 /* first correct the data pointers */
2416                 for (i = slot; i < nritems; i++) {
2417                         u32 ioff;
2418
2419                         item = btrfs_item_nr(i);
2420                         ioff = btrfs_item_offset(leaf, item);
2421                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2422                 }
2423
2424                 /* shift the items */
2425                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2426                               btrfs_item_nr_offset(slot),
2427                               (nritems - slot) * sizeof(struct btrfs_item));
2428
2429                 /* shift the data */
2430                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2431                               data_end - total_data, btrfs_leaf_data(leaf) +
2432                               data_end, old_data - data_end);
2433                 data_end = old_data;
2434         }
2435
2436         /* setup the item for the new data */
2437         for (i = 0; i < nr; i++) {
2438                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2439                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2440                 item = btrfs_item_nr(slot + i);
2441                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2442                 data_end -= data_size[i];
2443                 btrfs_set_item_size(leaf, item, data_size[i]);
2444         }
2445         btrfs_set_header_nritems(leaf, nritems + nr);
2446         btrfs_mark_buffer_dirty(leaf);
2447
2448         ret = 0;
2449         if (slot == 0) {
2450                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2451                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2452         }
2453
2454         if (btrfs_leaf_free_space(root, leaf) < 0) {
2455                 btrfs_print_leaf(root, leaf);
2456                 BUG();
2457         }
2458
2459 out:
2460         return ret;
2461 }
2462
2463 /*
2464  * Given a key and some data, insert an item into the tree.
2465  * This does all the path init required, making room in the tree if needed.
2466  */
2467 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2468                       *root, struct btrfs_key *cpu_key, void *data, u32
2469                       data_size)
2470 {
2471         int ret = 0;
2472         struct btrfs_path *path;
2473         struct extent_buffer *leaf;
2474         unsigned long ptr;
2475
2476         path = btrfs_alloc_path();
2477         BUG_ON(!path);
2478         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2479         if (!ret) {
2480                 leaf = path->nodes[0];
2481                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2482                 write_extent_buffer(leaf, data, ptr, data_size);
2483                 btrfs_mark_buffer_dirty(leaf);
2484         }
2485         btrfs_free_path(path);
2486         return ret;
2487 }
2488
2489 /*
2490  * delete the pointer from a given node.
2491  *
2492  * If the delete empties a node, the node is removed from the tree,
2493  * continuing all the way the root if required.  The root is converted into
2494  * a leaf if all the nodes are emptied.
2495  */
2496 int btrfs_del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2497                    struct btrfs_path *path, int level, int slot)
2498 {
2499         struct extent_buffer *parent = path->nodes[level];
2500         u32 nritems;
2501         int ret = 0;
2502         int wret;
2503
2504         nritems = btrfs_header_nritems(parent);
2505         if (slot != nritems -1) {
2506                 memmove_extent_buffer(parent,
2507                               btrfs_node_key_ptr_offset(slot),
2508                               btrfs_node_key_ptr_offset(slot + 1),
2509                               sizeof(struct btrfs_key_ptr) *
2510                               (nritems - slot - 1));
2511         }
2512         nritems--;
2513         btrfs_set_header_nritems(parent, nritems);
2514         if (nritems == 0 && parent == root->node) {
2515                 BUG_ON(btrfs_header_level(root->node) != 1);
2516                 /* just turn the root into a leaf and break */
2517                 btrfs_set_header_level(root->node, 0);
2518         } else if (slot == 0) {
2519                 struct btrfs_disk_key disk_key;
2520
2521                 btrfs_node_key(parent, &disk_key, 0);
2522                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2523                 if (wret)
2524                         ret = wret;
2525         }
2526         btrfs_mark_buffer_dirty(parent);
2527         return ret;
2528 }
2529
2530 /*
2531  * a helper function to delete the leaf pointed to by path->slots[1] and
2532  * path->nodes[1].
2533  *
2534  * This deletes the pointer in path->nodes[1] and frees the leaf
2535  * block extent.  zero is returned if it all worked out, < 0 otherwise.
2536  *
2537  * The path must have already been setup for deleting the leaf, including
2538  * all the proper balancing.  path->nodes[1] must be locked.
2539  */
2540 static noinline int btrfs_del_leaf(struct btrfs_trans_handle *trans,
2541                                    struct btrfs_root *root,
2542                                    struct btrfs_path *path,
2543                                    struct extent_buffer *leaf)
2544 {
2545         int ret;
2546
2547         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
2548         ret = btrfs_del_ptr(trans, root, path, 1, path->slots[1]);
2549         if (ret)
2550                 return ret;
2551
2552         ret = btrfs_free_extent(trans, root, leaf->start, leaf->len,
2553                                 0, root->root_key.objectid, 0, 0);
2554         return ret;
2555 }
2556
2557 /*
2558  * delete the item at the leaf level in path.  If that empties
2559  * the leaf, remove it from the tree
2560  */
2561 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2562                     struct btrfs_path *path, int slot, int nr)
2563 {
2564         struct extent_buffer *leaf;
2565         struct btrfs_item *item;
2566         int last_off;
2567         int dsize = 0;
2568         int ret = 0;
2569         int wret;
2570         int i;
2571         u32 nritems;
2572
2573         leaf = path->nodes[0];
2574         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2575
2576         for (i = 0; i < nr; i++)
2577                 dsize += btrfs_item_size_nr(leaf, slot + i);
2578
2579         nritems = btrfs_header_nritems(leaf);
2580
2581         if (slot + nr != nritems) {
2582                 int data_end = leaf_data_end(root, leaf);
2583
2584                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2585                               data_end + dsize,
2586                               btrfs_leaf_data(leaf) + data_end,
2587                               last_off - data_end);
2588
2589                 for (i = slot + nr; i < nritems; i++) {
2590                         u32 ioff;
2591
2592                         item = btrfs_item_nr(i);
2593                         ioff = btrfs_item_offset(leaf, item);
2594                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2595                 }
2596
2597                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2598                               btrfs_item_nr_offset(slot + nr),
2599                               sizeof(struct btrfs_item) *
2600                               (nritems - slot - nr));
2601         }
2602         btrfs_set_header_nritems(leaf, nritems - nr);
2603         nritems -= nr;
2604
2605         /* delete the leaf if we've emptied it */
2606         if (nritems == 0) {
2607                 if (leaf == root->node) {
2608                         btrfs_set_header_level(leaf, 0);
2609                 } else {
2610                         clean_tree_block(trans, root, leaf);
2611                         wait_on_tree_block_writeback(root, leaf);
2612
2613                         wret = btrfs_del_leaf(trans, root, path, leaf);
2614                         BUG_ON(ret);
2615                         if (wret)
2616                                 ret = wret;
2617                 }
2618         } else {
2619                 int used = leaf_space_used(leaf, 0, nritems);
2620                 if (slot == 0) {
2621                         struct btrfs_disk_key disk_key;
2622
2623                         btrfs_item_key(leaf, &disk_key, 0);
2624                         wret = fixup_low_keys(trans, root, path,
2625                                               &disk_key, 1);
2626                         if (wret)
2627                                 ret = wret;
2628                 }
2629
2630                 /* delete the leaf if it is mostly empty */
2631                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2632                         /* push_leaf_left fixes the path.
2633                          * make sure the path still points to our leaf
2634                          * for possible call to del_ptr below
2635                          */
2636                         slot = path->slots[1];
2637                         extent_buffer_get(leaf);
2638
2639                         wret = push_leaf_left(trans, root, path, 1, 1);
2640                         if (wret < 0 && wret != -ENOSPC)
2641                                 ret = wret;
2642
2643                         if (path->nodes[0] == leaf &&
2644                             btrfs_header_nritems(leaf)) {
2645                                 wret = push_leaf_right(trans, root, path, 1, 1);
2646                                 if (wret < 0 && wret != -ENOSPC)
2647                                         ret = wret;
2648                         }
2649
2650                         if (btrfs_header_nritems(leaf) == 0) {
2651                                 clean_tree_block(trans, root, leaf);
2652                                 wait_on_tree_block_writeback(root, leaf);
2653
2654                                 path->slots[1] = slot;
2655                                 ret = btrfs_del_leaf(trans, root, path, leaf);
2656                                 BUG_ON(ret);
2657                                 free_extent_buffer(leaf);
2658
2659                         } else {
2660                                 btrfs_mark_buffer_dirty(leaf);
2661                                 free_extent_buffer(leaf);
2662                         }
2663                 } else {
2664                         btrfs_mark_buffer_dirty(leaf);
2665                 }
2666         }
2667         return ret;
2668 }
2669
2670 /*
2671  * walk up the tree as far as required to find the previous leaf.
2672  * returns 0 if it found something or 1 if there are no lesser leaves.
2673  * returns < 0 on io errors.
2674  */
2675 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2676 {
2677         int slot;
2678         int level = 1;
2679         struct extent_buffer *c;
2680         struct extent_buffer *next = NULL;
2681
2682         while(level < BTRFS_MAX_LEVEL) {
2683                 if (!path->nodes[level])
2684                         return 1;
2685
2686                 slot = path->slots[level];
2687                 c = path->nodes[level];
2688                 if (slot == 0) {
2689                         level++;
2690                         if (level == BTRFS_MAX_LEVEL)
2691                                 return 1;
2692                         continue;
2693                 }
2694                 slot--;
2695
2696                 next = read_node_slot(root, c, slot);
2697                 break;
2698         }
2699         path->slots[level] = slot;
2700         while(1) {
2701                 level--;
2702                 c = path->nodes[level];
2703                 free_extent_buffer(c);
2704                 slot = btrfs_header_nritems(next);
2705                 if (slot != 0)
2706                         slot--;
2707                 path->nodes[level] = next;
2708                 path->slots[level] = slot;
2709                 if (!level)
2710                         break;
2711                 next = read_node_slot(root, next, slot);
2712         }
2713         return 0;
2714 }
2715
2716 /*
2717  * walk up the tree as far as required to find the next leaf.
2718  * returns 0 if it found something or 1 if there are no greater leaves.
2719  * returns < 0 on io errors.
2720  */
2721 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2722 {
2723         int slot;
2724         int level = 1;
2725         struct extent_buffer *c;
2726         struct extent_buffer *next = NULL;
2727
2728         while(level < BTRFS_MAX_LEVEL) {
2729                 if (!path->nodes[level])
2730                         return 1;
2731
2732                 slot = path->slots[level] + 1;
2733                 c = path->nodes[level];
2734                 if (slot >= btrfs_header_nritems(c)) {
2735                         level++;
2736                         if (level == BTRFS_MAX_LEVEL)
2737                                 return 1;
2738                         continue;
2739                 }
2740
2741                 if (path->reada)
2742                         reada_for_search(root, path, level, slot, 0);
2743
2744                 next = read_node_slot(root, c, slot);
2745                 if (!next)
2746                         return -EIO;
2747                 break;
2748         }
2749         path->slots[level] = slot;
2750         while(1) {
2751                 level--;
2752                 c = path->nodes[level];
2753                 free_extent_buffer(c);
2754                 path->nodes[level] = next;
2755                 path->slots[level] = 0;
2756                 if (!level)
2757                         break;
2758                 if (path->reada)
2759                         reada_for_search(root, path, level, 0, 0);
2760                 next = read_node_slot(root, next, 0);
2761                 if (!next)
2762                         return -EIO;
2763         }
2764         return 0;
2765 }
2766
2767 int btrfs_previous_item(struct btrfs_root *root,
2768                         struct btrfs_path *path, u64 min_objectid,
2769                         int type)
2770 {
2771         struct btrfs_key found_key;
2772         struct extent_buffer *leaf;
2773         int ret;
2774
2775         while(1) {
2776                 if (path->slots[0] == 0) {
2777                         ret = btrfs_prev_leaf(root, path);
2778                         if (ret != 0)
2779                                 return ret;
2780                 } else {
2781                         path->slots[0]--;
2782                 }
2783                 leaf = path->nodes[0];
2784                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2785                 if (found_key.type == type)
2786                         return 0;
2787         }
2788         return 1;
2789 }
2790