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 = kmalloc(sizeof(struct btrfs_path), GFP_NOFS);
  46         if (path) {
  47                 btrfs_init_path(path);
  48                 path->reada = 0;
  49         }
  50         return path;
  51 }
  52
  53 void btrfs_free_path(struct btrfs_path *p)
  54 {
  55         btrfs_release_path(NULL, p);
  56         kfree(p);
  57 }
  58
  59 void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
  60 {
  61         int i;
  62         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  63                 if (!p->nodes[i])
  64                         continue;
  65                 free_extent_buffer(p->nodes[i]);
  66         }
  67         memset(p, 0, sizeof(*p));
  68 }
  69
  70 void add_root_to_dirty_list(struct btrfs_root *root)
  71 {
  72         if (root->track_dirty && list_empty(&root->dirty_list)) {
  73                 list_add(&root->dirty_list,
  74                          &root->fs_info->dirty_cowonly_roots);
  75         }
  76 }
  77
  78 int btrfs_copy_root(struct btrfs_trans_handle *trans,
  79                       struct btrfs_root *root,
  80                       struct extent_buffer *buf,
  81                       struct extent_buffer **cow_ret, u64 new_root_objectid)
  82 {
  83         struct extent_buffer *cow;
  84         int ret = 0;
  85         int level;
  86         struct btrfs_root *new_root;
  87         struct btrfs_disk_key disk_key;
  88
  89         new_root = kmalloc(sizeof(*new_root), GFP_NOFS);
  90         if (!new_root)
  91                 return -ENOMEM;
  92
  93         memcpy(new_root, root, sizeof(*new_root));
  94         new_root->root_key.objectid = new_root_objectid;
  95
  96         WARN_ON(root->ref_cows && trans->transid !=
  97                 root->fs_info->running_transaction->transid);
  98         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
  99
 100         level = btrfs_header_level(buf);
 101         if (level == 0)
 102                 btrfs_item_key(buf, &disk_key, 0);
 103         else
 104                 btrfs_node_key(buf, &disk_key, 0);
 105         cow = btrfs_alloc_free_block(trans, new_root, buf->len,
 106                                      new_root_objectid, &disk_key,
 107                                      level, buf->start, 0);
 108         if (IS_ERR(cow)) {
 109                 kfree(new_root);
 110                 return PTR_ERR(cow);
 111         }
 112
 113         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 114         btrfs_set_header_bytenr(cow, cow->start);
 115         btrfs_set_header_generation(cow, trans->transid);
 116         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 117         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 118                                      BTRFS_HEADER_FLAG_RELOC);
 119         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 120                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 121         else
 122                 btrfs_set_header_owner(cow, new_root_objectid);
 123
 124         write_extent_buffer(cow, root->fs_info->fsid,
 125                             (unsigned long)btrfs_header_fsid(cow),
 126                             BTRFS_FSID_SIZE);
 127
 128         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 129         ret = btrfs_inc_ref(trans, new_root, cow, 0);
 130         kfree(new_root);
 131
 132         if (ret)
 133                 return ret;
 134
 135         btrfs_mark_buffer_dirty(cow);
 136         *cow_ret = cow;
 137         return 0;
 138 }
 139
 140 /*
 141  * check if the tree block can be shared by multiple trees
 142  */
 143 int btrfs_block_can_be_shared(struct btrfs_root *root,
 144                               struct extent_buffer *buf)
 145 {
 146         /*
 147          * Tree blocks not in refernece counted trees and tree roots
 148          * are never shared. If a block was allocated after the last
 149          * snapshot and the block was not allocated by tree relocation,
 150          * we know the block is not shared.
 151          */
 152         if (root->ref_cows &&
 153             buf != root->node && buf != root->commit_root &&
 154             (btrfs_header_generation(buf) <=
 155              btrfs_root_last_snapshot(&root->root_item) ||
 156              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 157                 return 1;
 158 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 159         if (root->ref_cows &&
 160             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 161                 return 1;
 162 #endif
 163         return 0;
 164 }
 165
 166 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
 167                                        struct btrfs_root *root,
 168                                        struct extent_buffer *buf,
 169                                        struct extent_buffer *cow)
 170 {
 171         u64 refs;
 172         u64 owner;
 173         u64 flags;
 174         u64 new_flags = 0;
 175         int ret;
 176
 177         /*
 178          * Backrefs update rules:
 179          *
 180          * Always use full backrefs for extent pointers in tree block
 181          * allocated by tree relocation.
 182          *
 183          * If a shared tree block is no longer referenced by its owner
 184          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
 185          * use full backrefs for extent pointers in tree block.
 186          *
 187          * If a tree block is been relocating
 188          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
 189          * use full backrefs for extent pointers in tree block.
 190          * The reason for this is some operations (such as drop tree)
 191          * are only allowed for blocks use full backrefs.
 192          */
 193
 194         if (btrfs_block_can_be_shared(root, buf)) {
 195                 ret = btrfs_lookup_extent_info(trans, root, buf->start,
 196                                                btrfs_header_level(buf), 1,
 197                                                &refs, &flags);
 198                 BUG_ON(ret);
 199                 BUG_ON(refs == 0);
 200         } else {
 201                 refs = 1;
 202                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 203                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 204                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 205                 else
 206                         flags = 0;
 207         }
 208
 209         owner = btrfs_header_owner(buf);
 210         BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) &&
 211                owner == BTRFS_TREE_RELOC_OBJECTID);
 212
 213         if (refs > 1) {
 214                 if ((owner == root->root_key.objectid ||
 215                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
 216                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
 217                         ret = btrfs_inc_ref(trans, root, buf, 1);
 218                         BUG_ON(ret);
 219
 220                         if (root->root_key.objectid ==
 221                             BTRFS_TREE_RELOC_OBJECTID) {
 222                                 ret = btrfs_dec_ref(trans, root, buf, 0);
 223                                 BUG_ON(ret);
 224                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 225                                 BUG_ON(ret);
 226                         }
 227                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
 228                 } else {
 229
 230                         if (root->root_key.objectid ==
 231                             BTRFS_TREE_RELOC_OBJECTID)
 232                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 233                         else
 234                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 235                         BUG_ON(ret);
 236                 }
 237                 if (new_flags != 0) {
 238                         ret = btrfs_set_block_flags(trans, root, buf->start,
 239                                                     btrfs_header_level(buf),
 240                                                     new_flags);
 241                         BUG_ON(ret);
 242                 }
 243         } else {
 244                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
 245                         if (root->root_key.objectid ==
 246                             BTRFS_TREE_RELOC_OBJECTID)
 247                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 248                         else
 249                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 250                         BUG_ON(ret);
 251                         ret = btrfs_dec_ref(trans, root, buf, 1);
 252                         BUG_ON(ret);
 253                 }
 254                 clean_tree_block(trans, root, buf);
 255         }
 256         return 0;
 257 }
 258
 259 int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 260                              struct btrfs_root *root,
 261                              struct extent_buffer *buf,
 262                              struct extent_buffer *parent, int parent_slot,
 263                              struct extent_buffer **cow_ret,
 264                              u64 search_start, u64 empty_size)
 265 {
 266         struct extent_buffer *cow;
 267         struct btrfs_disk_key disk_key;
 268         int level;
 269
 270         WARN_ON(root->ref_cows && trans->transid !=
 271                 root->fs_info->running_transaction->transid);
 272         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 273
 274         level = btrfs_header_level(buf);
 275
 276         if (level == 0)
 277                 btrfs_item_key(buf, &disk_key, 0);
 278         else
 279                 btrfs_node_key(buf, &disk_key, 0);
 280
 281         cow = btrfs_alloc_free_block(trans, root, buf->len,
 282                                      root->root_key.objectid, &disk_key,
 283                                      level, search_start, empty_size);
 284         if (IS_ERR(cow))
 285                 return PTR_ERR(cow);
 286
 287         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 288         btrfs_set_header_bytenr(cow, cow->start);
 289         btrfs_set_header_generation(cow, trans->transid);
 290         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 291         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 292                                      BTRFS_HEADER_FLAG_RELOC);
 293         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 294                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 295         else
 296                 btrfs_set_header_owner(cow, root->root_key.objectid);
 297
 298         write_extent_buffer(cow, root->fs_info->fsid,
 299                             (unsigned long)btrfs_header_fsid(cow),
 300                             BTRFS_FSID_SIZE);
 301
 302         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 303
 304         update_ref_for_cow(trans, root, buf, cow);
 305
 306         if (buf == root->node) {
 307                 root->node = cow;
 308                 extent_buffer_get(cow);
 309
 310                 btrfs_free_extent(trans, root, buf->start, buf->len,
 311                                   0, root->root_key.objectid, level, 0);
 312                 free_extent_buffer(buf);
 313                 add_root_to_dirty_list(root);
 314         } else {
 315                 btrfs_set_node_blockptr(parent, parent_slot,
 316                                         cow->start);
 317                 WARN_ON(trans->transid == 0);
 318                 btrfs_set_node_ptr_generation(parent, parent_slot,
 319                                               trans->transid);
 320                 btrfs_mark_buffer_dirty(parent);
 321                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 322
 323                 btrfs_free_extent(trans, root, buf->start, buf->len,
 324                                   0, root->root_key.objectid, level, 1);
 325         }
 326         free_extent_buffer(buf);
 327         btrfs_mark_buffer_dirty(cow);
 328         *cow_ret = cow;
 329         return 0;
 330 }
 331
 332 static inline int should_cow_block(struct btrfs_trans_handle *trans,
 333                                    struct btrfs_root *root,
 334                                    struct extent_buffer *buf)
 335 {
 336         if (btrfs_header_generation(buf) == trans->transid &&
 337             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
 338             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
 339               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 340                 return 0;
 341         return 1;
 342 }
 343
 344 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 345                     struct btrfs_root *root, struct extent_buffer *buf,
 346                     struct extent_buffer *parent, int parent_slot,
 347                     struct extent_buffer **cow_ret)
 348 {
 349         u64 search_start;
 350         int ret;
 351         /*
 352         if (trans->transaction != root->fs_info->running_transaction) {
 353                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 354                        root->fs_info->running_transaction->transid);
 355                 WARN_ON(1);
 356         }
 357         */
 358         if (trans->transid != root->fs_info->generation) {
 359                 printk(KERN_CRIT "trans %llu running %llu\n",
 360                         (unsigned long long)trans->transid,
 361                         (unsigned long long)root->fs_info->generation);
 362                 WARN_ON(1);
 363         }
 364         if (!should_cow_block(trans, root, buf)) {
 365                 *cow_ret = buf;
 366                 return 0;
 367         }
 368
 369         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
 370         ret = __btrfs_cow_block(trans, root, buf, parent,
 371                                  parent_slot, cow_ret, search_start, 0);
 372         return ret;
 373 }
 374
 375 /*
 376  * compare two keys in a memcmp fashion
 377  */
 378 int btrfs_comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 379 {
 380         struct btrfs_key k1;
 381
 382         btrfs_disk_key_to_cpu(&k1, disk);
 383
 384         if (k1.objectid > k2->objectid)
 385                 return 1;
 386         if (k1.objectid < k2->objectid)
 387                 return -1;
 388         if (k1.type > k2->type)
 389                 return 1;
 390         if (k1.type < k2->type)
 391                 return -1;
 392         if (k1.offset > k2->offset)
 393                 return 1;
 394         if (k1.offset < k2->offset)
 395                 return -1;
 396         return 0;
 397 }
 398
 399 /*
 400  * The leaf data grows from end-to-front in the node.
 401  * this returns the address of the start of the last item,
 402  * which is the stop of the leaf data stack
 403  */
 404 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 405                                          struct extent_buffer *leaf)
 406 {
 407         u32 nr = btrfs_header_nritems(leaf);
 408         if (nr == 0)
 409                 return BTRFS_LEAF_DATA_SIZE(root);
 410         return btrfs_item_offset_nr(leaf, nr - 1);
 411 }
 412
 413 int btrfs_check_node(struct btrfs_root *root,
 414                       struct btrfs_disk_key *parent_key,
 415                       struct extent_buffer *buf)
 416 {
 417         int i;
 418         struct btrfs_key cpukey;
 419         struct btrfs_disk_key key;
 420         u32 nritems = btrfs_header_nritems(buf);
 421
 422         if (nritems == 0 || nritems > BTRFS_NODEPTRS_PER_BLOCK(root))
 423                 goto fail;
 424
 425         if (parent_key && parent_key->type) {
 426                 btrfs_node_key(buf, &key, 0);
 427                 if (memcmp(parent_key, &key, sizeof(key)))
 428                         goto fail;
 429         }
 430         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 431                 btrfs_node_key(buf, &key, i);
 432                 btrfs_node_key_to_cpu(buf, &cpukey, i + 1);
 433                 if (btrfs_comp_keys(&key, &cpukey) >= 0)
 434                         goto fail;
 435         }
 436         return 0;
 437 fail:
 438         if (btrfs_header_owner(buf) == BTRFS_EXTENT_TREE_OBJECTID) {
 439                 if (parent_key)
 440                         btrfs_disk_key_to_cpu(&cpukey, parent_key);
 441                 else
 442                         btrfs_node_key_to_cpu(buf, &cpukey, 0);
 443                 btrfs_add_corrupt_extent_record(root->fs_info, &cpukey,
 444                                                 buf->start, buf->len,
 445                                                 btrfs_header_level(buf));
 446         }
 447         return -EIO;
 448 }
 449
 450 int btrfs_check_leaf(struct btrfs_root *root,
 451                       struct btrfs_disk_key *parent_key,
 452                       struct extent_buffer *buf)
 453 {
 454         int i;
 455         struct btrfs_key cpukey;
 456         struct btrfs_disk_key key;
 457         u32 nritems = btrfs_header_nritems(buf);
 458
 459         if (nritems * sizeof(struct btrfs_item) > buf->len)  {
 460                 fprintf(stderr, "invalid number of items %llu\n",
 461                         (unsigned long long)buf->start);
 462                 goto fail;
 463         }
 464
 465         if (btrfs_header_level(buf) != 0) {
 466                 fprintf(stderr, "leaf is not a leaf %llu\n",
 467                        (unsigned long long)btrfs_header_bytenr(buf));
 468                 goto fail;
 469         }
 470         if (btrfs_leaf_free_space(root, buf) < 0) {
 471                 fprintf(stderr, "leaf free space incorrect %llu %d\n",
 472                         (unsigned long long)btrfs_header_bytenr(buf),
 473                         btrfs_leaf_free_space(root, buf));
 474                 goto fail;
 475         }
 476
 477         if (nritems == 0)
 478                 return 0;
 479
 480         btrfs_item_key(buf, &key, 0);
 481         if (parent_key && parent_key->type &&
 482             memcmp(parent_key, &key, sizeof(key))) {
 483                 fprintf(stderr, "leaf parent key incorrect %llu\n",
 484                        (unsigned long long)btrfs_header_bytenr(buf));
 485                 goto fail;
 486         }
 487         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 488                 btrfs_item_key(buf, &key, i);
 489                 btrfs_item_key_to_cpu(buf, &cpukey, i + 1);
 490                 if (btrfs_comp_keys(&key, &cpukey) >= 0) {
 491                         fprintf(stderr, "bad key ordering %d %d\n", i, i+1);
 492                         goto fail;
 493                 }
 494                 if (btrfs_item_offset_nr(buf, i) !=
 495                         btrfs_item_end_nr(buf, i + 1)) {
 496                         fprintf(stderr, "incorrect offsets %u %u\n",
 497                                 btrfs_item_offset_nr(buf, i),
 498                                 btrfs_item_end_nr(buf, i + 1));
 499                         goto fail;
 500                 }
 501                 if (i == 0 && btrfs_item_end_nr(buf, i) !=
 502                     BTRFS_LEAF_DATA_SIZE(root)) {
 503                         fprintf(stderr, "bad item end %u wanted %u\n",
 504                                 btrfs_item_end_nr(buf, i),
 505                                 (unsigned)BTRFS_LEAF_DATA_SIZE(root));
 506                         goto fail;
 507                 }
 508         }
 509         return 0;
 510 fail:
 511         if (btrfs_header_owner(buf) == BTRFS_EXTENT_TREE_OBJECTID) {
 512                 if (parent_key)
 513                         btrfs_disk_key_to_cpu(&cpukey, parent_key);
 514                 else
 515                         btrfs_item_key_to_cpu(buf, &cpukey, 0);
 516
 517                 btrfs_add_corrupt_extent_record(root->fs_info, &cpukey,
 518                                                 buf->start, buf->len, 0);
 519         }
 520         return -EIO;
 521 }
 522
 523 static int noinline check_block(struct btrfs_root *root,
 524                                 struct btrfs_path *path, int level)
 525 {
 526         struct btrfs_disk_key key;
 527         struct btrfs_disk_key *key_ptr = NULL;
 528         struct extent_buffer *parent;
 529
 530         if (path->nodes[level + 1]) {
 531                 parent = path->nodes[level + 1];
 532                 btrfs_node_key(parent, &key, path->slots[level + 1]);
 533                 key_ptr = &key;
 534         }
 535         if (level == 0)
 536                 return btrfs_check_leaf(root, key_ptr, path->nodes[0]);
 537         return btrfs_check_node(root, key_ptr, path->nodes[level]);
 538 }
 539
 540 /*
 541  * search for key in the extent_buffer.  The items start at offset p,
 542  * and they are item_size apart.  There are 'max' items in p.
 543  *
 544  * the slot in the array is returned via slot, and it points to
 545  * the place where you would insert key if it is not found in
 546  * the array.
 547  *
 548  * slot may point to max if the key is bigger than all of the keys
 549  */
 550 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 551                               int item_size, struct btrfs_key *key,
 552                               int max, int *slot)
 553 {
 554         int low = 0;
 555         int high = max;
 556         int mid;
 557         int ret;
 558         unsigned long offset;
 559         struct btrfs_disk_key *tmp;
 560
 561         while(low < high) {
 562                 mid = (low + high) / 2;
 563                 offset = p + mid * item_size;
 564
 565                 tmp = (struct btrfs_disk_key *)(eb->data + offset);
 566                 ret = btrfs_comp_keys(tmp, key);
 567
 568                 if (ret < 0)
 569                         low = mid + 1;
 570                 else if (ret > 0)
 571                         high = mid;
 572                 else {
 573                         *slot = mid;
 574                         return 0;
 575                 }
 576         }
 577         *slot = low;
 578         return 1;
 579 }
 580
 581 /*
 582  * simple bin_search frontend that does the right thing for
 583  * leaves vs nodes
 584  */
 585 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 586                       int level, int *slot)
 587 {
 588         if (level == 0) {
 589                 return generic_bin_search(eb,
 590                                           offsetof(struct btrfs_leaf, items),
 591                                           sizeof(struct btrfs_item),
 592                                           key, btrfs_header_nritems(eb),
 593                                           slot);
 594         } else {
 595                 return generic_bin_search(eb,
 596                                           offsetof(struct btrfs_node, ptrs),
 597                                           sizeof(struct btrfs_key_ptr),
 598                                           key, btrfs_header_nritems(eb),
 599                                           slot);
 600         }
 601         return -1;
 602 }
 603
 604 struct extent_buffer *read_node_slot(struct btrfs_root *root,
 605                                    struct extent_buffer *parent, int slot)
 606 {
 607         int level = btrfs_header_level(parent);
 608         if (slot < 0)
 609                 return NULL;
 610         if (slot >= btrfs_header_nritems(parent))
 611                 return NULL;
 612
 613         if (level == 0)
 614                 return NULL;
 615
 616         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 617                        btrfs_level_size(root, level - 1),
 618                        btrfs_node_ptr_generation(parent, slot));
 619 }
 620
 621 static int balance_level(struct btrfs_trans_handle *trans,
 622                          struct btrfs_root *root,
 623                          struct btrfs_path *path, int level)
 624 {
 625         struct extent_buffer *right = NULL;
 626         struct extent_buffer *mid;
 627         struct extent_buffer *left = NULL;
 628         struct extent_buffer *parent = NULL;
 629         int ret = 0;
 630         int wret;
 631         int pslot;
 632         int orig_slot = path->slots[level];
 633         u64 orig_ptr;
 634
 635         if (level == 0)
 636                 return 0;
 637
 638         mid = path->nodes[level];
 639         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 640
 641         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 642
 643         if (level < BTRFS_MAX_LEVEL - 1) {
 644                 parent = path->nodes[level + 1];
 645                 pslot = path->slots[level + 1];
 646         }
 647
 648         /*
 649          * deal with the case where there is only one pointer in the root
 650          * by promoting the node below to a root
 651          */
 652         if (!parent) {
 653                 struct extent_buffer *child;
 654
 655                 if (btrfs_header_nritems(mid) != 1)
 656                         return 0;
 657
 658                 /* promote the child to a root */
 659                 child = read_node_slot(root, mid, 0);
 660                 BUG_ON(!child);
 661                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
 662                 BUG_ON(ret);
 663
 664                 root->node = child;
 665                 add_root_to_dirty_list(root);
 666                 path->nodes[level] = NULL;
 667                 clean_tree_block(trans, root, mid);
 668                 wait_on_tree_block_writeback(root, mid);
 669                 /* once for the path */
 670                 free_extent_buffer(mid);
 671
 672                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 673                                         0, root->root_key.objectid,
 674                                         level, 1);
 675                 /* once for the root ptr */
 676                 free_extent_buffer(mid);
 677                 return ret;
 678         }
 679         if (btrfs_header_nritems(mid) >
 680             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 681                 return 0;
 682
 683         left = read_node_slot(root, parent, pslot - 1);
 684         if (left) {
 685                 wret = btrfs_cow_block(trans, root, left,
 686                                        parent, pslot - 1, &left);
 687                 if (wret) {
 688                         ret = wret;
 689                         goto enospc;
 690                 }
 691         }
 692         right = read_node_slot(root, parent, pslot + 1);
 693         if (right) {
 694                 wret = btrfs_cow_block(trans, root, right,
 695                                        parent, pslot + 1, &right);
 696                 if (wret) {
 697                         ret = wret;
 698                         goto enospc;
 699                 }
 700         }
 701
 702         /* first, try to make some room in the middle buffer */
 703         if (left) {
 704                 orig_slot += btrfs_header_nritems(left);
 705                 wret = push_node_left(trans, root, left, mid, 1);
 706                 if (wret < 0)
 707                         ret = wret;
 708         }
 709
 710         /*
 711          * then try to empty the right most buffer into the middle
 712          */
 713         if (right) {
 714                 wret = push_node_left(trans, root, mid, right, 1);
 715                 if (wret < 0 && wret != -ENOSPC)
 716                         ret = wret;
 717                 if (btrfs_header_nritems(right) == 0) {
 718                         u64 bytenr = right->start;
 719                         u32 blocksize = right->len;
 720
 721                         clean_tree_block(trans, root, right);
 722                         wait_on_tree_block_writeback(root, right);
 723                         free_extent_buffer(right);
 724                         right = NULL;
 725                         wret = btrfs_del_ptr(trans, root, path,
 726                                              level + 1, pslot + 1);
 727                         if (wret)
 728                                 ret = wret;
 729                         wret = btrfs_free_extent(trans, root, bytenr,
 730                                                  blocksize, 0,
 731                                                  root->root_key.objectid,
 732                                                  level, 0);
 733                         if (wret)
 734                                 ret = wret;
 735                 } else {
 736                         struct btrfs_disk_key right_key;
 737                         btrfs_node_key(right, &right_key, 0);
 738                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 739                         btrfs_mark_buffer_dirty(parent);
 740                 }
 741         }
 742         if (btrfs_header_nritems(mid) == 1) {
 743                 /*
 744                  * we're not allowed to leave a node with one item in the
 745                  * tree during a delete.  A deletion from lower in the tree
 746                  * could try to delete the only pointer in this node.
 747                  * So, pull some keys from the left.
 748                  * There has to be a left pointer at this point because
 749                  * otherwise we would have pulled some pointers from the
 750                  * right
 751                  */
 752                 BUG_ON(!left);
 753                 wret = balance_node_right(trans, root, mid, left);
 754                 if (wret < 0) {
 755                         ret = wret;
 756                         goto enospc;
 757                 }
 758                 if (wret == 1) {
 759                         wret = push_node_left(trans, root, left, mid, 1);
 760                         if (wret < 0)
 761                                 ret = wret;
 762                 }
 763                 BUG_ON(wret == 1);
 764         }
 765         if (btrfs_header_nritems(mid) == 0) {
 766                 /* we've managed to empty the middle node, drop it */
 767                 u64 bytenr = mid->start;
 768                 u32 blocksize = mid->len;
 769                 clean_tree_block(trans, root, mid);
 770                 wait_on_tree_block_writeback(root, mid);
 771                 free_extent_buffer(mid);
 772                 mid = NULL;
 773                 wret = btrfs_del_ptr(trans, root, path, level + 1, pslot);
 774                 if (wret)
 775                         ret = wret;
 776                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 777                                          0, root->root_key.objectid,
 778                                          level, 0);
 779                 if (wret)
 780                         ret = wret;
 781         } else {
 782                 /* update the parent key to reflect our changes */
 783                 struct btrfs_disk_key mid_key;
 784                 btrfs_node_key(mid, &mid_key, 0);
 785                 btrfs_set_node_key(parent, &mid_key, pslot);
 786                 btrfs_mark_buffer_dirty(parent);
 787         }
 788
 789         /* update the path */
 790         if (left) {
 791                 if (btrfs_header_nritems(left) > orig_slot) {
 792                         extent_buffer_get(left);
 793                         path->nodes[level] = left;
 794                         path->slots[level + 1] -= 1;
 795                         path->slots[level] = orig_slot;
 796                         if (mid)
 797                                 free_extent_buffer(mid);
 798                 } else {
 799                         orig_slot -= btrfs_header_nritems(left);
 800                         path->slots[level] = orig_slot;
 801                 }
 802         }
 803         /* double check we haven't messed things up */
 804         check_block(root, path, level);
 805         if (orig_ptr !=
 806             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 807                 BUG();
 808 enospc:
 809         if (right)
 810                 free_extent_buffer(right);
 811         if (left)
 812                 free_extent_buffer(left);
 813         return ret;
 814 }
 815
 816 /* returns zero if the push worked, non-zero otherwise */
 817 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
 818                                           struct btrfs_root *root,
 819                                           struct btrfs_path *path, int level)
 820 {
 821         struct extent_buffer *right = NULL;
 822         struct extent_buffer *mid;
 823         struct extent_buffer *left = NULL;
 824         struct extent_buffer *parent = NULL;
 825         int ret = 0;
 826         int wret;
 827         int pslot;
 828         int orig_slot = path->slots[level];
 829
 830         if (level == 0)
 831                 return 1;
 832
 833         mid = path->nodes[level];
 834         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 835
 836         if (level < BTRFS_MAX_LEVEL - 1) {
 837                 parent = path->nodes[level + 1];
 838                 pslot = path->slots[level + 1];
 839         }
 840
 841         if (!parent)
 842                 return 1;
 843
 844         left = read_node_slot(root, parent, pslot - 1);
 845
 846         /* first, try to make some room in the middle buffer */
 847         if (left) {
 848                 u32 left_nr;
 849                 left_nr = btrfs_header_nritems(left);
 850                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 851                         wret = 1;
 852                 } else {
 853                         ret = btrfs_cow_block(trans, root, left, parent,
 854                                               pslot - 1, &left);
 855                         if (ret)
 856                                 wret = 1;
 857                         else {
 858                                 wret = push_node_left(trans, root,
 859                                                       left, mid, 0);
 860                         }
 861                 }
 862                 if (wret < 0)
 863                         ret = wret;
 864                 if (wret == 0) {
 865                         struct btrfs_disk_key disk_key;
 866                         orig_slot += left_nr;
 867                         btrfs_node_key(mid, &disk_key, 0);
 868                         btrfs_set_node_key(parent, &disk_key, pslot);
 869                         btrfs_mark_buffer_dirty(parent);
 870                         if (btrfs_header_nritems(left) > orig_slot) {
 871                                 path->nodes[level] = left;
 872                                 path->slots[level + 1] -= 1;
 873                                 path->slots[level] = orig_slot;
 874                                 free_extent_buffer(mid);
 875                         } else {
 876                                 orig_slot -=
 877                                         btrfs_header_nritems(left);
 878                                 path->slots[level] = orig_slot;
 879                                 free_extent_buffer(left);
 880                         }
 881                         return 0;
 882                 }
 883                 free_extent_buffer(left);
 884         }
 885         right= read_node_slot(root, parent, pslot + 1);
 886
 887         /*
 888          * then try to empty the right most buffer into the middle
 889          */
 890         if (right) {
 891                 u32 right_nr;
 892                 right_nr = btrfs_header_nritems(right);
 893                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 894                         wret = 1;
 895                 } else {
 896                         ret = btrfs_cow_block(trans, root, right,
 897                                               parent, pslot + 1,
 898                                               &right);
 899                         if (ret)
 900                                 wret = 1;
 901                         else {
 902                                 wret = balance_node_right(trans, root,
 903                                                           right, mid);
 904                         }
 905                 }
 906                 if (wret < 0)
 907                         ret = wret;
 908                 if (wret == 0) {
 909                         struct btrfs_disk_key disk_key;
 910
 911                         btrfs_node_key(right, &disk_key, 0);
 912                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
 913                         btrfs_mark_buffer_dirty(parent);
 914
 915                         if (btrfs_header_nritems(mid) <= orig_slot) {
 916                                 path->nodes[level] = right;
 917                                 path->slots[level + 1] += 1;
 918                                 path->slots[level] = orig_slot -
 919                                         btrfs_header_nritems(mid);
 920                                 free_extent_buffer(mid);
 921                         } else {
 922                                 free_extent_buffer(right);
 923                         }
 924                         return 0;
 925                 }
 926                 free_extent_buffer(right);
 927         }
 928         return 1;
 929 }
 930
 931 /*
 932  * readahead one full node of leaves
 933  */
 934 void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
 935                              int level, int slot, u64 objectid)
 936 {
 937         struct extent_buffer *node;
 938         struct btrfs_disk_key disk_key;
 939         u32 nritems;
 940         u64 search;
 941         u64 lowest_read;
 942         u64 highest_read;
 943         u64 nread = 0;
 944         int direction = path->reada;
 945         struct extent_buffer *eb;
 946         u32 nr;
 947         u32 blocksize;
 948         u32 nscan = 0;
 949
 950         if (level != 1)
 951                 return;
 952
 953         if (!path->nodes[level])
 954                 return;
 955
 956         node = path->nodes[level];
 957         search = btrfs_node_blockptr(node, slot);
 958         blocksize = btrfs_level_size(root, level - 1);
 959         eb = btrfs_find_tree_block(root, search, blocksize);
 960         if (eb) {
 961                 free_extent_buffer(eb);
 962                 return;
 963         }
 964
 965         highest_read = search;
 966         lowest_read = search;
 967
 968         nritems = btrfs_header_nritems(node);
 969         nr = slot;
 970         while(1) {
 971                 if (direction < 0) {
 972                         if (nr == 0)
 973                                 break;
 974                         nr--;
 975                 } else if (direction > 0) {
 976                         nr++;
 977                         if (nr >= nritems)
 978                                 break;
 979                 }
 980                 if (path->reada < 0 && objectid) {
 981                         btrfs_node_key(node, &disk_key, nr);
 982                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
 983                                 break;
 984                 }
 985                 search = btrfs_node_blockptr(node, nr);
 986                 if ((search >= lowest_read && search <= highest_read) ||
 987                     (search < lowest_read && lowest_read - search <= 32768) ||
 988                     (search > highest_read && search - highest_read <= 32768)) {
 989                         readahead_tree_block(root, search, blocksize,
 990                                      btrfs_node_ptr_generation(node, nr));
 991                         nread += blocksize;
 992                 }
 993                 nscan++;
 994                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
 995                         break;
 996                 if(nread > (1024 * 1024) || nscan > 128)
 997                         break;
 998
 999                 if (search < lowest_read)
1000                         lowest_read = search;
1001                 if (search > highest_read)
1002                         highest_read = search;
1003         }
1004 }
1005
1006 /*
1007  * look for key in the tree.  path is filled in with nodes along the way
1008  * if key is found, we return zero and you can find the item in the leaf
1009  * level of the path (level 0)
1010  *
1011  * If the key isn't found, the path points to the slot where it should
1012  * be inserted, and 1 is returned.  If there are other errors during the
1013  * search a negative error number is returned.
1014  *
1015  * if ins_len > 0, nodes and leaves will be split as we walk down the
1016  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1017  * possible)
1018  */
1019 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1020                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1021                       ins_len, int cow)
1022 {
1023         struct extent_buffer *b;
1024         int slot;
1025         int ret;
1026         int level;
1027         int should_reada = p->reada;
1028         u8 lowest_level = 0;
1029
1030         lowest_level = p->lowest_level;
1031         WARN_ON(lowest_level && ins_len > 0);
1032         WARN_ON(p->nodes[0] != NULL);
1033         /*
1034         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
1035         */
1036 again:
1037         b = root->node;
1038         extent_buffer_get(b);
1039         while (b) {
1040                 level = btrfs_header_level(b);
1041                 if (cow) {
1042                         int wret;
1043                         wret = btrfs_cow_block(trans, root, b,
1044                                                p->nodes[level + 1],
1045                                                p->slots[level + 1],
1046                                                &b);
1047                         if (wret) {
1048                                 free_extent_buffer(b);
1049                                 return wret;
1050                         }
1051                 }
1052                 BUG_ON(!cow && ins_len);
1053                 if (level != btrfs_header_level(b))
1054                         WARN_ON(1);
1055                 level = btrfs_header_level(b);
1056                 p->nodes[level] = b;
1057                 ret = check_block(root, p, level);
1058                 if (ret)
1059                         return -1;
1060                 ret = bin_search(b, key, level, &slot);
1061                 if (level != 0) {
1062                         if (ret && slot > 0)
1063                                 slot -= 1;
1064                         p->slots[level] = slot;
1065                         if ((p->search_for_split || ins_len > 0) &&
1066                             btrfs_header_nritems(b) >=
1067                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1068                                 int sret = split_node(trans, root, p, level);
1069                                 BUG_ON(sret > 0);
1070                                 if (sret)
1071                                         return sret;
1072                                 b = p->nodes[level];
1073                                 slot = p->slots[level];
1074                         } else if (ins_len < 0) {
1075                                 int sret = balance_level(trans, root, p,
1076                                                          level);
1077                                 if (sret)
1078                                         return sret;
1079                                 b = p->nodes[level];
1080                                 if (!b) {
1081                                         btrfs_release_path(NULL, p);
1082                                         goto again;
1083                                 }
1084                                 slot = p->slots[level];
1085                                 BUG_ON(btrfs_header_nritems(b) == 1);
1086                         }
1087                         /* this is only true while dropping a snapshot */
1088                         if (level == lowest_level)
1089                                 break;
1090
1091                         if (should_reada)
1092                                 reada_for_search(root, p, level, slot,
1093                                                  key->objectid);
1094
1095                         b = read_node_slot(root, b, slot);
1096                         if (!extent_buffer_uptodate(b))
1097                                 return -EIO;
1098                 } else {
1099                         p->slots[level] = slot;
1100                         if (ins_len > 0 &&
1101                             ins_len > btrfs_leaf_free_space(root, b)) {
1102                                 int sret = split_leaf(trans, root, key,
1103                                                       p, ins_len, ret == 0);
1104                                 BUG_ON(sret > 0);
1105                                 if (sret)
1106                                         return sret;
1107                         }
1108                         return ret;
1109                 }
1110         }
1111         return 1;
1112 }
1113
1114 /*
1115  * adjust the pointers going up the tree, starting at level
1116  * making sure the right key of each node is points to 'key'.
1117  * This is used after shifting pointers to the left, so it stops
1118  * fixing up pointers when a given leaf/node is not in slot 0 of the
1119  * higher levels
1120  *
1121  * If this fails to write a tree block, it returns -1, but continues
1122  * fixing up the blocks in ram so the tree is consistent.
1123  */
1124 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1125                           struct btrfs_root *root, struct btrfs_path *path,
1126                           struct btrfs_disk_key *key, int level)
1127 {
1128         int i;
1129         int ret = 0;
1130         struct extent_buffer *t;
1131
1132         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1133                 int tslot = path->slots[i];
1134                 if (!path->nodes[i])
1135                         break;
1136                 t = path->nodes[i];
1137                 btrfs_set_node_key(t, key, tslot);
1138                 btrfs_mark_buffer_dirty(path->nodes[i]);
1139                 if (tslot != 0)
1140                         break;
1141         }
1142         return ret;
1143 }
1144
1145 /*
1146  * update item key.
1147  *
1148  * This function isn't completely safe. It's the caller's responsibility
1149  * that the new key won't break the order
1150  */
1151 int btrfs_set_item_key_safe(struct btrfs_trans_handle *trans,
1152                             struct btrfs_root *root, struct btrfs_path *path,
1153                             struct btrfs_key *new_key)
1154 {
1155         struct btrfs_disk_key disk_key;
1156         struct extent_buffer *eb;
1157         int slot;
1158
1159         eb = path->nodes[0];
1160         slot = path->slots[0];
1161         if (slot > 0) {
1162                 btrfs_item_key(eb, &disk_key, slot - 1);
1163                 if (btrfs_comp_keys(&disk_key, new_key) >= 0)
1164                         return -1;
1165         }
1166         if (slot < btrfs_header_nritems(eb) - 1) {
1167                 btrfs_item_key(eb, &disk_key, slot + 1);
1168                 if (btrfs_comp_keys(&disk_key, new_key) <= 0)
1169                         return -1;
1170         }
1171
1172         btrfs_cpu_key_to_disk(&disk_key, new_key);
1173         btrfs_set_item_key(eb, &disk_key, slot);
1174         btrfs_mark_buffer_dirty(eb);
1175         if (slot == 0)
1176                 fixup_low_keys(trans, root, path, &disk_key, 1);
1177         return 0;
1178 }
1179
1180 /*
1181  * try to push data from one node into the next node left in the
1182  * tree.
1183  *
1184  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1185  * error, and > 0 if there was no room in the left hand block.
1186  */
1187 static int push_node_left(struct btrfs_trans_handle *trans,
1188                           struct btrfs_root *root, struct extent_buffer *dst,
1189                           struct extent_buffer *src, int empty)
1190 {
1191         int push_items = 0;
1192         int src_nritems;
1193         int dst_nritems;
1194         int ret = 0;
1195
1196         src_nritems = btrfs_header_nritems(src);
1197         dst_nritems = btrfs_header_nritems(dst);
1198         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1199         WARN_ON(btrfs_header_generation(src) != trans->transid);
1200         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1201
1202         if (!empty && src_nritems <= 8)
1203                 return 1;
1204
1205         if (push_items <= 0) {
1206                 return 1;
1207         }
1208
1209         if (empty) {
1210                 push_items = min(src_nritems, push_items);
1211                 if (push_items < src_nritems) {
1212                         /* leave at least 8 pointers in the node if
1213                          * we aren't going to empty it
1214                          */
1215                         if (src_nritems - push_items < 8) {
1216                                 if (push_items <= 8)
1217                                         return 1;
1218                                 push_items -= 8;
1219                         }
1220                 }
1221         } else
1222                 push_items = min(src_nritems - 8, push_items);
1223
1224         copy_extent_buffer(dst, src,
1225                            btrfs_node_key_ptr_offset(dst_nritems),
1226                            btrfs_node_key_ptr_offset(0),
1227                            push_items * sizeof(struct btrfs_key_ptr));
1228
1229         if (push_items < src_nritems) {
1230                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1231                                       btrfs_node_key_ptr_offset(push_items),
1232                                       (src_nritems - push_items) *
1233                                       sizeof(struct btrfs_key_ptr));
1234         }
1235         btrfs_set_header_nritems(src, src_nritems - push_items);
1236         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1237         btrfs_mark_buffer_dirty(src);
1238         btrfs_mark_buffer_dirty(dst);
1239
1240         return ret;
1241 }
1242
1243 /*
1244  * try to push data from one node into the next node right in the
1245  * tree.
1246  *
1247  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1248  * error, and > 0 if there was no room in the right hand block.
1249  *
1250  * this will  only push up to 1/2 the contents of the left node over
1251  */
1252 static int balance_node_right(struct btrfs_trans_handle *trans,
1253                               struct btrfs_root *root,
1254                               struct extent_buffer *dst,
1255                               struct extent_buffer *src)
1256 {
1257         int push_items = 0;
1258         int max_push;
1259         int src_nritems;
1260         int dst_nritems;
1261         int ret = 0;
1262
1263         WARN_ON(btrfs_header_generation(src) != trans->transid);
1264         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1265
1266         src_nritems = btrfs_header_nritems(src);
1267         dst_nritems = btrfs_header_nritems(dst);
1268         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1269         if (push_items <= 0) {
1270                 return 1;
1271         }
1272
1273         if (src_nritems < 4) {
1274                 return 1;
1275         }
1276
1277         max_push = src_nritems / 2 + 1;
1278         /* don't try to empty the node */
1279         if (max_push >= src_nritems) {
1280                 return 1;
1281         }
1282
1283         if (max_push < push_items)
1284                 push_items = max_push;
1285
1286         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1287                                       btrfs_node_key_ptr_offset(0),
1288                                       (dst_nritems) *
1289                                       sizeof(struct btrfs_key_ptr));
1290
1291         copy_extent_buffer(dst, src,
1292                            btrfs_node_key_ptr_offset(0),
1293                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1294                            push_items * sizeof(struct btrfs_key_ptr));
1295
1296         btrfs_set_header_nritems(src, src_nritems - push_items);
1297         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1298
1299         btrfs_mark_buffer_dirty(src);
1300         btrfs_mark_buffer_dirty(dst);
1301
1302         return ret;
1303 }
1304
1305 /*
1306  * helper function to insert a new root level in the tree.
1307  * A new node is allocated, and a single item is inserted to
1308  * point to the existing root
1309  *
1310  * returns zero on success or < 0 on failure.
1311  */
1312 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1313                            struct btrfs_root *root,
1314                            struct btrfs_path *path, int level)
1315 {
1316         u64 lower_gen;
1317         struct extent_buffer *lower;
1318         struct extent_buffer *c;
1319         struct extent_buffer *old;
1320         struct btrfs_disk_key lower_key;
1321
1322         BUG_ON(path->nodes[level]);
1323         BUG_ON(path->nodes[level-1] != root->node);
1324
1325         lower = path->nodes[level-1];
1326         if (level == 1)
1327                 btrfs_item_key(lower, &lower_key, 0);
1328         else
1329                 btrfs_node_key(lower, &lower_key, 0);
1330
1331         c = btrfs_alloc_free_block(trans, root, root->nodesize,
1332                                    root->root_key.objectid, &lower_key,
1333                                    level, root->node->start, 0);
1334
1335         if (IS_ERR(c))
1336                 return PTR_ERR(c);
1337
1338         memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
1339         btrfs_set_header_nritems(c, 1);
1340         btrfs_set_header_level(c, level);
1341         btrfs_set_header_bytenr(c, c->start);
1342         btrfs_set_header_generation(c, trans->transid);
1343         btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
1344         btrfs_set_header_owner(c, root->root_key.objectid);
1345
1346         write_extent_buffer(c, root->fs_info->fsid,
1347                             (unsigned long)btrfs_header_fsid(c),
1348                             BTRFS_FSID_SIZE);
1349
1350         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1351                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
1352                             BTRFS_UUID_SIZE);
1353
1354         btrfs_set_node_key(c, &lower_key, 0);
1355         btrfs_set_node_blockptr(c, 0, lower->start);
1356         lower_gen = btrfs_header_generation(lower);
1357         WARN_ON(lower_gen != trans->transid);
1358
1359         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1360
1361         btrfs_mark_buffer_dirty(c);
1362
1363         old = root->node;
1364         root->node = c;
1365
1366         /* the super has an extra ref to root->node */
1367         free_extent_buffer(old);
1368
1369         add_root_to_dirty_list(root);
1370         extent_buffer_get(c);
1371         path->nodes[level] = c;
1372         path->slots[level] = 0;
1373         return 0;
1374 }
1375
1376 /*
1377  * worker function to insert a single pointer in a node.
1378  * the node should have enough room for the pointer already
1379  *
1380  * slot and level indicate where you want the key to go, and
1381  * blocknr is the block the key points to.
1382  *
1383  * returns zero on success and < 0 on any error
1384  */
1385 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1386                       *root, struct btrfs_path *path, struct btrfs_disk_key
1387                       *key, u64 bytenr, int slot, int level)
1388 {
1389         struct extent_buffer *lower;
1390         int nritems;
1391
1392         BUG_ON(!path->nodes[level]);
1393         lower = path->nodes[level];
1394         nritems = btrfs_header_nritems(lower);
1395         if (slot > nritems)
1396                 BUG();
1397         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1398                 BUG();
1399         if (slot != nritems) {
1400                 memmove_extent_buffer(lower,
1401                               btrfs_node_key_ptr_offset(slot + 1),
1402                               btrfs_node_key_ptr_offset(slot),
1403                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1404         }
1405         btrfs_set_node_key(lower, key, slot);
1406         btrfs_set_node_blockptr(lower, slot, bytenr);
1407         WARN_ON(trans->transid == 0);
1408         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1409         btrfs_set_header_nritems(lower, nritems + 1);
1410         btrfs_mark_buffer_dirty(lower);
1411         return 0;
1412 }
1413
1414 /*
1415  * split the node at the specified level in path in two.
1416  * The path is corrected to point to the appropriate node after the split
1417  *
1418  * Before splitting this tries to make some room in the node by pushing
1419  * left and right, if either one works, it returns right away.
1420  *
1421  * returns 0 on success and < 0 on failure
1422  */
1423 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1424                       *root, struct btrfs_path *path, int level)
1425 {
1426         struct extent_buffer *c;
1427         struct extent_buffer *split;
1428         struct btrfs_disk_key disk_key;
1429         int mid;
1430         int ret;
1431         int wret;
1432         u32 c_nritems;
1433
1434         c = path->nodes[level];
1435         WARN_ON(btrfs_header_generation(c) != trans->transid);
1436         if (c == root->node) {
1437                 /* trying to split the root, lets make a new one */
1438                 ret = insert_new_root(trans, root, path, level + 1);
1439                 if (ret)
1440                         return ret;
1441         } else {
1442                 ret = push_nodes_for_insert(trans, root, path, level);
1443                 c = path->nodes[level];
1444                 if (!ret && btrfs_header_nritems(c) <
1445                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
1446                         return 0;
1447                 if (ret < 0)
1448                         return ret;
1449         }
1450
1451         c_nritems = btrfs_header_nritems(c);
1452         mid = (c_nritems + 1) / 2;
1453         btrfs_node_key(c, &disk_key, mid);
1454
1455         split = btrfs_alloc_free_block(trans, root, root->nodesize,
1456                                         root->root_key.objectid,
1457                                         &disk_key, level, c->start, 0);
1458         if (IS_ERR(split))
1459                 return PTR_ERR(split);
1460
1461         memset_extent_buffer(split, 0, 0, sizeof(struct btrfs_header));
1462         btrfs_set_header_level(split, btrfs_header_level(c));
1463         btrfs_set_header_bytenr(split, split->start);
1464         btrfs_set_header_generation(split, trans->transid);
1465         btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
1466         btrfs_set_header_owner(split, root->root_key.objectid);
1467         write_extent_buffer(split, root->fs_info->fsid,
1468                             (unsigned long)btrfs_header_fsid(split),
1469                             BTRFS_FSID_SIZE);
1470         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1471                             (unsigned long)btrfs_header_chunk_tree_uuid(split),
1472                             BTRFS_UUID_SIZE);
1473
1474
1475         copy_extent_buffer(split, c,
1476                            btrfs_node_key_ptr_offset(0),
1477                            btrfs_node_key_ptr_offset(mid),
1478                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1479         btrfs_set_header_nritems(split, c_nritems - mid);
1480         btrfs_set_header_nritems(c, mid);
1481         ret = 0;
1482
1483         btrfs_mark_buffer_dirty(c);
1484         btrfs_mark_buffer_dirty(split);
1485
1486         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1487                           path->slots[level + 1] + 1,
1488                           level + 1);
1489         if (wret)
1490                 ret = wret;
1491
1492         if (path->slots[level] >= mid) {
1493                 path->slots[level] -= mid;
1494                 free_extent_buffer(c);
1495                 path->nodes[level] = split;
1496                 path->slots[level + 1] += 1;
1497         } else {
1498                 free_extent_buffer(split);
1499         }
1500         return ret;
1501 }
1502
1503 /*
1504  * how many bytes are required to store the items in a leaf.  start
1505  * and nr indicate which items in the leaf to check.  This totals up the
1506  * space used both by the item structs and the item data
1507  */
1508 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1509 {
1510         int data_len;
1511         int nritems = btrfs_header_nritems(l);
1512         int end = min(nritems, start + nr) - 1;
1513
1514         if (!nr)
1515                 return 0;
1516         data_len = btrfs_item_end_nr(l, start);
1517         data_len = data_len - btrfs_item_offset_nr(l, end);
1518         data_len += sizeof(struct btrfs_item) * nr;
1519         WARN_ON(data_len < 0);
1520         return data_len;
1521 }
1522
1523 /*
1524  * The space between the end of the leaf items and
1525  * the start of the leaf data.  IOW, how much room
1526  * the leaf has left for both items and data
1527  */
1528 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1529 {
1530         int nritems = btrfs_header_nritems(leaf);
1531         int ret;
1532         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1533         if (ret < 0) {
1534                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1535                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1536                        leaf_space_used(leaf, 0, nritems), nritems);
1537         }
1538         return ret;
1539 }
1540
1541 /*
1542  * push some data in the path leaf to the right, trying to free up at
1543  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1544  *
1545  * returns 1 if the push failed because the other node didn't have enough
1546  * room, 0 if everything worked out and < 0 if there were major errors.
1547  */
1548 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1549                            *root, struct btrfs_path *path, int data_size,
1550                            int empty)
1551 {
1552         struct extent_buffer *left = path->nodes[0];
1553         struct extent_buffer *right;
1554         struct extent_buffer *upper;
1555         struct btrfs_disk_key disk_key;
1556         int slot;
1557         u32 i;
1558         int free_space;
1559         int push_space = 0;
1560         int push_items = 0;
1561         struct btrfs_item *item;
1562         u32 left_nritems;
1563         u32 nr;
1564         u32 right_nritems;
1565         u32 data_end;
1566         u32 this_item_size;
1567         int ret;
1568
1569         slot = path->slots[1];
1570         if (!path->nodes[1]) {
1571                 return 1;
1572         }
1573         upper = path->nodes[1];
1574         if (slot >= btrfs_header_nritems(upper) - 1)
1575                 return 1;
1576
1577         right = read_node_slot(root, upper, slot + 1);
1578         free_space = btrfs_leaf_free_space(root, right);
1579         if (free_space < data_size) {
1580                 free_extent_buffer(right);
1581                 return 1;
1582         }
1583
1584         /* cow and double check */
1585         ret = btrfs_cow_block(trans, root, right, upper,
1586                               slot + 1, &right);
1587         if (ret) {
1588                 free_extent_buffer(right);
1589                 return 1;
1590         }
1591         free_space = btrfs_leaf_free_space(root, right);
1592         if (free_space < data_size) {
1593                 free_extent_buffer(right);
1594                 return 1;
1595         }
1596
1597         left_nritems = btrfs_header_nritems(left);
1598         if (left_nritems == 0) {
1599                 free_extent_buffer(right);
1600                 return 1;
1601         }
1602
1603         if (empty)
1604                 nr = 0;
1605         else
1606                 nr = 1;
1607
1608         i = left_nritems - 1;
1609         while (i >= nr) {
1610                 item = btrfs_item_nr(left, i);
1611
1612                 if (path->slots[0] == i)
1613                         push_space += data_size + sizeof(*item);
1614
1615                 this_item_size = btrfs_item_size(left, item);
1616                 if (this_item_size + sizeof(*item) + push_space > free_space)
1617                         break;
1618                 push_items++;
1619                 push_space += this_item_size + sizeof(*item);
1620                 if (i == 0)
1621                         break;
1622                 i--;
1623         }
1624
1625         if (push_items == 0) {
1626                 free_extent_buffer(right);
1627                 return 1;
1628         }
1629
1630         if (!empty && push_items == left_nritems)
1631                 WARN_ON(1);
1632
1633         /* push left to right */
1634         right_nritems = btrfs_header_nritems(right);
1635
1636         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1637         push_space -= leaf_data_end(root, left);
1638
1639         /* make room in the right data area */
1640         data_end = leaf_data_end(root, right);
1641         memmove_extent_buffer(right,
1642                               btrfs_leaf_data(right) + data_end - push_space,
1643                               btrfs_leaf_data(right) + data_end,
1644                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1645
1646         /* copy from the left data area */
1647         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1648                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1649                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1650                      push_space);
1651
1652         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1653                               btrfs_item_nr_offset(0),
1654                               right_nritems * sizeof(struct btrfs_item));
1655
1656         /* copy the items from left to right */
1657         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1658                    btrfs_item_nr_offset(left_nritems - push_items),
1659                    push_items * sizeof(struct btrfs_item));
1660
1661         /* update the item pointers */
1662         right_nritems += push_items;
1663         btrfs_set_header_nritems(right, right_nritems);
1664         push_space = BTRFS_LEAF_DATA_SIZE(root);
1665         for (i = 0; i < right_nritems; i++) {
1666                 item = btrfs_item_nr(right, i);
1667                 push_space -= btrfs_item_size(right, item);
1668                 btrfs_set_item_offset(right, item, push_space);
1669         }
1670
1671         left_nritems -= push_items;
1672         btrfs_set_header_nritems(left, left_nritems);
1673
1674         if (left_nritems)
1675                 btrfs_mark_buffer_dirty(left);
1676         btrfs_mark_buffer_dirty(right);
1677
1678         btrfs_item_key(right, &disk_key, 0);
1679         btrfs_set_node_key(upper, &disk_key, slot + 1);
1680         btrfs_mark_buffer_dirty(upper);
1681
1682         /* then fixup the leaf pointer in the path */
1683         if (path->slots[0] >= left_nritems) {
1684                 path->slots[0] -= left_nritems;
1685                 free_extent_buffer(path->nodes[0]);
1686                 path->nodes[0] = right;
1687                 path->slots[1] += 1;
1688         } else {
1689                 free_extent_buffer(right);
1690         }
1691         return 0;
1692 }
1693 /*
1694  * push some data in the path leaf to the left, trying to free up at
1695  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1696  */
1697 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1698                           *root, struct btrfs_path *path, int data_size,
1699                           int empty)
1700 {
1701         struct btrfs_disk_key disk_key;
1702         struct extent_buffer *right = path->nodes[0];
1703         struct extent_buffer *left;
1704         int slot;
1705         int i;
1706         int free_space;
1707         int push_space = 0;
1708         int push_items = 0;
1709         struct btrfs_item *item;
1710         u32 old_left_nritems;
1711         u32 right_nritems;
1712         u32 nr;
1713         int ret = 0;
1714         int wret;
1715         u32 this_item_size;
1716         u32 old_left_item_size;
1717
1718         slot = path->slots[1];
1719         if (slot == 0)
1720                 return 1;
1721         if (!path->nodes[1])
1722                 return 1;
1723
1724         right_nritems = btrfs_header_nritems(right);
1725         if (right_nritems == 0) {
1726                 return 1;
1727         }
1728
1729         left = read_node_slot(root, path->nodes[1], slot - 1);
1730         free_space = btrfs_leaf_free_space(root, left);
1731         if (free_space < data_size) {
1732                 free_extent_buffer(left);
1733                 return 1;
1734         }
1735
1736         /* cow and double check */
1737         ret = btrfs_cow_block(trans, root, left,
1738                               path->nodes[1], slot - 1, &left);
1739         if (ret) {
1740                 /* we hit -ENOSPC, but it isn't fatal here */
1741                 free_extent_buffer(left);
1742                 return 1;
1743         }
1744
1745         free_space = btrfs_leaf_free_space(root, left);
1746         if (free_space < data_size) {
1747                 free_extent_buffer(left);
1748                 return 1;
1749         }
1750
1751         if (empty)
1752                 nr = right_nritems;
1753         else
1754                 nr = right_nritems - 1;
1755
1756         for (i = 0; i < nr; i++) {
1757                 item = btrfs_item_nr(right, i);
1758
1759                 if (path->slots[0] == i)
1760                         push_space += data_size + sizeof(*item);
1761
1762                 this_item_size = btrfs_item_size(right, item);
1763                 if (this_item_size + sizeof(*item) + push_space > free_space)
1764                         break;
1765
1766                 push_items++;
1767                 push_space += this_item_size + sizeof(*item);
1768         }
1769
1770         if (push_items == 0) {
1771                 free_extent_buffer(left);
1772                 return 1;
1773         }
1774         if (!empty && push_items == btrfs_header_nritems(right))
1775                 WARN_ON(1);
1776
1777         /* push data from right to left */
1778         copy_extent_buffer(left, right,
1779                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1780                            btrfs_item_nr_offset(0),
1781                            push_items * sizeof(struct btrfs_item));
1782
1783         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1784                      btrfs_item_offset_nr(right, push_items -1);
1785
1786         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1787                      leaf_data_end(root, left) - push_space,
1788                      btrfs_leaf_data(right) +
1789                      btrfs_item_offset_nr(right, push_items - 1),
1790                      push_space);
1791         old_left_nritems = btrfs_header_nritems(left);
1792         BUG_ON(old_left_nritems == 0);
1793
1794         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1795         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1796                 u32 ioff;
1797
1798                 item = btrfs_item_nr(left, i);
1799                 ioff = btrfs_item_offset(left, item);
1800                 btrfs_set_item_offset(left, item,
1801                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1802         }
1803         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1804
1805         /* fixup right node */
1806         if (push_items > right_nritems) {
1807                 printk("push items %d nr %u\n", push_items, right_nritems);
1808                 WARN_ON(1);
1809         }
1810
1811         if (push_items < right_nritems) {
1812                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1813                                                   leaf_data_end(root, right);
1814                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
1815                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1816                                       btrfs_leaf_data(right) +
1817                                       leaf_data_end(root, right), push_space);
1818
1819                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1820                               btrfs_item_nr_offset(push_items),
1821                              (btrfs_header_nritems(right) - push_items) *
1822                              sizeof(struct btrfs_item));
1823         }
1824         right_nritems -= push_items;
1825         btrfs_set_header_nritems(right, right_nritems);
1826         push_space = BTRFS_LEAF_DATA_SIZE(root);
1827         for (i = 0; i < right_nritems; i++) {
1828                 item = btrfs_item_nr(right, i);
1829                 push_space = push_space - btrfs_item_size(right, item);
1830                 btrfs_set_item_offset(right, item, push_space);
1831         }
1832
1833         btrfs_mark_buffer_dirty(left);
1834         if (right_nritems)
1835                 btrfs_mark_buffer_dirty(right);
1836
1837         btrfs_item_key(right, &disk_key, 0);
1838         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1839         if (wret)
1840                 ret = wret;
1841
1842         /* then fixup the leaf pointer in the path */
1843         if (path->slots[0] < push_items) {
1844                 path->slots[0] += old_left_nritems;
1845                 free_extent_buffer(path->nodes[0]);
1846                 path->nodes[0] = left;
1847                 path->slots[1] -= 1;
1848         } else {
1849                 free_extent_buffer(left);
1850                 path->slots[0] -= push_items;
1851         }
1852         BUG_ON(path->slots[0] < 0);
1853         return ret;
1854 }
1855
1856 /*
1857  * split the path's leaf in two, making sure there is at least data_size
1858  * available for the resulting leaf level of the path.
1859  *
1860  * returns 0 if all went well and < 0 on failure.
1861  */
1862 static noinline int copy_for_split(struct btrfs_trans_handle *trans,
1863                                struct btrfs_root *root,
1864                                struct btrfs_path *path,
1865                                struct extent_buffer *l,
1866                                struct extent_buffer *right,
1867                                int slot, int mid, int nritems)
1868 {
1869         int data_copy_size;
1870         int rt_data_off;
1871         int i;
1872         int ret = 0;
1873         int wret;
1874         struct btrfs_disk_key disk_key;
1875
1876         nritems = nritems - mid;
1877         btrfs_set_header_nritems(right, nritems);
1878         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
1879
1880         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
1881                            btrfs_item_nr_offset(mid),
1882                            nritems * sizeof(struct btrfs_item));
1883
1884         copy_extent_buffer(right, l,
1885                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
1886                      data_copy_size, btrfs_leaf_data(l) +
1887                      leaf_data_end(root, l), data_copy_size);
1888
1889         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
1890                       btrfs_item_end_nr(l, mid);
1891
1892         for (i = 0; i < nritems; i++) {
1893                 struct btrfs_item *item = btrfs_item_nr(right, i);
1894                 u32 ioff = btrfs_item_offset(right, item);
1895                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
1896         }
1897
1898         btrfs_set_header_nritems(l, mid);
1899         ret = 0;
1900         btrfs_item_key(right, &disk_key, 0);
1901         wret = insert_ptr(trans, root, path, &disk_key, right->start,
1902                           path->slots[1] + 1, 1);
1903         if (wret)
1904                 ret = wret;
1905
1906         btrfs_mark_buffer_dirty(right);
1907         btrfs_mark_buffer_dirty(l);
1908         BUG_ON(path->slots[0] != slot);
1909
1910         if (mid <= slot) {
1911                 free_extent_buffer(path->nodes[0]);
1912                 path->nodes[0] = right;
1913                 path->slots[0] -= mid;
1914                 path->slots[1] += 1;
1915         } else {
1916                 free_extent_buffer(right);
1917         }
1918
1919         BUG_ON(path->slots[0] < 0);
1920
1921         return ret;
1922 }
1923
1924 /*
1925  * split the path's leaf in two, making sure there is at least data_size
1926  * available for the resulting leaf level of the path.
1927  *
1928  * returns 0 if all went well and < 0 on failure.
1929  */
1930 static noinline int split_leaf(struct btrfs_trans_handle *trans,
1931                                struct btrfs_root *root,
1932                                struct btrfs_key *ins_key,
1933                                struct btrfs_path *path, int data_size,
1934                                int extend)
1935 {
1936         struct btrfs_disk_key disk_key;
1937         struct extent_buffer *l;
1938         u32 nritems;
1939         int mid;
1940         int slot;
1941         struct extent_buffer *right;
1942         int ret = 0;
1943         int wret;
1944         int split;
1945         int num_doubles = 0;
1946
1947         /* first try to make some room by pushing left and right */
1948         if (data_size && ins_key->type != BTRFS_DIR_ITEM_KEY) {
1949                 wret = push_leaf_right(trans, root, path, data_size, 0);
1950                 if (wret < 0)
1951                         return wret;
1952                 if (wret) {
1953                         wret = push_leaf_left(trans, root, path, data_size, 0);
1954                         if (wret < 0)
1955                                 return wret;
1956                 }
1957                 l = path->nodes[0];
1958
1959                 /* did the pushes work? */
1960                 if (btrfs_leaf_free_space(root, l) >= data_size)
1961                         return 0;
1962         }
1963
1964         if (!path->nodes[1]) {
1965                 ret = insert_new_root(trans, root, path, 1);
1966                 if (ret)
1967                         return ret;
1968         }
1969 again:
1970         split = 1;
1971         l = path->nodes[0];
1972         slot = path->slots[0];
1973         nritems = btrfs_header_nritems(l);
1974         mid = (nritems + 1) / 2;
1975
1976         if (mid <= slot) {
1977                 if (nritems == 1 ||
1978                     leaf_space_used(l, mid, nritems - mid) + data_size >
1979                         BTRFS_LEAF_DATA_SIZE(root)) {
1980                         if (slot >= nritems) {
1981                                 split = 0;
1982                         } else {
1983                                 mid = slot;
1984                                 if (mid != nritems &&
1985                                     leaf_space_used(l, mid, nritems - mid) +
1986                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
1987                                         split = 2;
1988                                 }
1989                         }
1990                 }
1991         } else {
1992                 if (leaf_space_used(l, 0, mid) + data_size >
1993                         BTRFS_LEAF_DATA_SIZE(root)) {
1994                         if (!extend && data_size && slot == 0) {
1995                                 split = 0;
1996                         } else if ((extend || !data_size) && slot == 0) {
1997                                 mid = 1;
1998                         } else {
1999                                 mid = slot;
2000                                 if (mid != nritems &&
2001                                     leaf_space_used(l, mid, nritems - mid) +
2002                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2003                                         split = 2 ;
2004                                 }
2005                         }
2006                 }
2007         }
2008
2009         if (split == 0)
2010                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2011         else
2012                 btrfs_item_key(l, &disk_key, mid);
2013
2014         right = btrfs_alloc_free_block(trans, root, root->leafsize,
2015                                         root->root_key.objectid,
2016                                         &disk_key, 0, l->start, 0);
2017         if (IS_ERR(right)) {
2018                 BUG_ON(1);
2019                 return PTR_ERR(right);
2020         }
2021
2022         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2023         btrfs_set_header_bytenr(right, right->start);
2024         btrfs_set_header_generation(right, trans->transid);
2025         btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
2026         btrfs_set_header_owner(right, root->root_key.objectid);
2027         btrfs_set_header_level(right, 0);
2028         write_extent_buffer(right, root->fs_info->fsid,
2029                             (unsigned long)btrfs_header_fsid(right),
2030                             BTRFS_FSID_SIZE);
2031
2032         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2033                             (unsigned long)btrfs_header_chunk_tree_uuid(right),
2034                             BTRFS_UUID_SIZE);
2035
2036         if (split == 0) {
2037                 if (mid <= slot) {
2038                         btrfs_set_header_nritems(right, 0);
2039                         wret = insert_ptr(trans, root, path,
2040                                           &disk_key, right->start,
2041                                           path->slots[1] + 1, 1);
2042                         if (wret)
2043                                 ret = wret;
2044
2045                         free_extent_buffer(path->nodes[0]);
2046                         path->nodes[0] = right;
2047                         path->slots[0] = 0;
2048                         path->slots[1] += 1;
2049                 } else {
2050                         btrfs_set_header_nritems(right, 0);
2051                         wret = insert_ptr(trans, root, path,
2052                                           &disk_key,
2053                                           right->start,
2054                                           path->slots[1], 1);
2055                         if (wret)
2056                                 ret = wret;
2057                         free_extent_buffer(path->nodes[0]);
2058                         path->nodes[0] = right;
2059                         path->slots[0] = 0;
2060                         if (path->slots[1] == 0) {
2061                                 wret = fixup_low_keys(trans, root,
2062                                                 path, &disk_key, 1);
2063                                 if (wret)
2064                                         ret = wret;
2065                         }
2066                 }
2067                 btrfs_mark_buffer_dirty(right);
2068                 return ret;
2069         }
2070
2071         ret = copy_for_split(trans, root, path, l, right, slot, mid, nritems);
2072         BUG_ON(ret);
2073
2074         if (split == 2) {
2075                 BUG_ON(num_doubles != 0);
2076                 num_doubles++;
2077                 goto again;
2078         }
2079
2080         return ret;
2081 }
2082
2083 /*
2084  * This function splits a single item into two items,
2085  * giving 'new_key' to the new item and splitting the
2086  * old one at split_offset (from the start of the item).
2087  *
2088  * The path may be released by this operation.  After
2089  * the split, the path is pointing to the old item.  The
2090  * new item is going to be in the same node as the old one.
2091  *
2092  * Note, the item being split must be smaller enough to live alone on
2093  * a tree block with room for one extra struct btrfs_item
2094  *
2095  * This allows us to split the item in place, keeping a lock on the
2096  * leaf the entire time.
2097  */
2098 int btrfs_split_item(struct btrfs_trans_handle *trans,
2099                      struct btrfs_root *root,
2100                      struct btrfs_path *path,
2101                      struct btrfs_key *new_key,
2102                      unsigned long split_offset)
2103 {
2104         u32 item_size;
2105         struct extent_buffer *leaf;
2106         struct btrfs_key orig_key;
2107         struct btrfs_item *item;
2108         struct btrfs_item *new_item;
2109         int ret = 0;
2110         int slot;
2111         u32 nritems;
2112         u32 orig_offset;
2113         struct btrfs_disk_key disk_key;
2114         char *buf;
2115
2116         leaf = path->nodes[0];
2117         btrfs_item_key_to_cpu(leaf, &orig_key, path->slots[0]);
2118         if (btrfs_leaf_free_space(root, leaf) >= sizeof(struct btrfs_item))
2119                 goto split;
2120
2121         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2122         btrfs_release_path(root, path);
2123
2124         path->search_for_split = 1;
2125
2126         ret = btrfs_search_slot(trans, root, &orig_key, path, 0, 1);
2127         path->search_for_split = 0;
2128
2129         /* if our item isn't there or got smaller, return now */
2130         if (ret != 0 || item_size != btrfs_item_size_nr(path->nodes[0],
2131                                                         path->slots[0])) {
2132                 return -EAGAIN;
2133         }
2134
2135         ret = split_leaf(trans, root, &orig_key, path, 0, 0);
2136         BUG_ON(ret);
2137
2138         BUG_ON(btrfs_leaf_free_space(root, leaf) < sizeof(struct btrfs_item));
2139         leaf = path->nodes[0];
2140
2141 split:
2142         item = btrfs_item_nr(leaf, path->slots[0]);
2143         orig_offset = btrfs_item_offset(leaf, item);
2144         item_size = btrfs_item_size(leaf, item);
2145
2146
2147         buf = kmalloc(item_size, GFP_NOFS);
2148         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
2149                             path->slots[0]), item_size);
2150         slot = path->slots[0] + 1;
2151         leaf = path->nodes[0];
2152
2153         nritems = btrfs_header_nritems(leaf);
2154
2155         if (slot != nritems) {
2156                 /* shift the items */
2157                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
2158                               btrfs_item_nr_offset(slot),
2159                               (nritems - slot) * sizeof(struct btrfs_item));
2160
2161         }
2162
2163         btrfs_cpu_key_to_disk(&disk_key, new_key);
2164         btrfs_set_item_key(leaf, &disk_key, slot);
2165
2166         new_item = btrfs_item_nr(leaf, slot);
2167
2168         btrfs_set_item_offset(leaf, new_item, orig_offset);
2169         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
2170
2171         btrfs_set_item_offset(leaf, item,
2172                               orig_offset + item_size - split_offset);
2173         btrfs_set_item_size(leaf, item, split_offset);
2174
2175         btrfs_set_header_nritems(leaf, nritems + 1);
2176
2177         /* write the data for the start of the original item */
2178         write_extent_buffer(leaf, buf,
2179                             btrfs_item_ptr_offset(leaf, path->slots[0]),
2180                             split_offset);
2181
2182         /* write the data for the new item */
2183         write_extent_buffer(leaf, buf + split_offset,
2184                             btrfs_item_ptr_offset(leaf, slot),
2185                             item_size - split_offset);
2186         btrfs_mark_buffer_dirty(leaf);
2187
2188         ret = 0;
2189         if (btrfs_leaf_free_space(root, leaf) < 0) {
2190                 btrfs_print_leaf(root, leaf);
2191                 BUG();
2192         }
2193         kfree(buf);
2194         return ret;
2195 }
2196
2197 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2198                         struct btrfs_root *root,
2199                         struct btrfs_path *path,
2200                         u32 new_size, int from_end)
2201 {
2202         int ret = 0;
2203         int slot;
2204         struct extent_buffer *leaf;
2205         struct btrfs_item *item;
2206         u32 nritems;
2207         unsigned int data_end;
2208         unsigned int old_data_start;
2209         unsigned int old_size;
2210         unsigned int size_diff;
2211         int i;
2212
2213         leaf = path->nodes[0];
2214         slot = path->slots[0];
2215
2216         old_size = btrfs_item_size_nr(leaf, slot);
2217         if (old_size == new_size)
2218                 return 0;
2219
2220         nritems = btrfs_header_nritems(leaf);
2221         data_end = leaf_data_end(root, leaf);
2222
2223         old_data_start = btrfs_item_offset_nr(leaf, slot);
2224
2225         size_diff = old_size - new_size;
2226
2227         BUG_ON(slot < 0);
2228         BUG_ON(slot >= nritems);
2229
2230         /*
2231          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2232          */
2233         /* first correct the data pointers */
2234         for (i = slot; i < nritems; i++) {
2235                 u32 ioff;
2236                 item = btrfs_item_nr(leaf, i);
2237                 ioff = btrfs_item_offset(leaf, item);
2238                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2239         }
2240
2241         /* shift the data */
2242         if (from_end) {
2243                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2244                               data_end + size_diff, btrfs_leaf_data(leaf) +
2245                               data_end, old_data_start + new_size - data_end);
2246         } else {
2247                 struct btrfs_disk_key disk_key;
2248                 u64 offset;
2249
2250                 btrfs_item_key(leaf, &disk_key, slot);
2251
2252                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2253                         unsigned long ptr;
2254                         struct btrfs_file_extent_item *fi;
2255
2256                         fi = btrfs_item_ptr(leaf, slot,
2257                                             struct btrfs_file_extent_item);
2258                         fi = (struct btrfs_file_extent_item *)(
2259                              (unsigned long)fi - size_diff);
2260
2261                         if (btrfs_file_extent_type(leaf, fi) ==
2262                             BTRFS_FILE_EXTENT_INLINE) {
2263                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2264                                 memmove_extent_buffer(leaf, ptr,
2265                                         (unsigned long)fi,
2266                                         offsetof(struct btrfs_file_extent_item,
2267                                                  disk_bytenr));
2268                         }
2269                 }
2270
2271                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2272                               data_end + size_diff, btrfs_leaf_data(leaf) +
2273                               data_end, old_data_start - data_end);
2274
2275                 offset = btrfs_disk_key_offset(&disk_key);
2276                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2277                 btrfs_set_item_key(leaf, &disk_key, slot);
2278                 if (slot == 0)
2279                         fixup_low_keys(trans, root, path, &disk_key, 1);
2280         }
2281
2282         item = btrfs_item_nr(leaf, slot);
2283         btrfs_set_item_size(leaf, item, new_size);
2284         btrfs_mark_buffer_dirty(leaf);
2285
2286         ret = 0;
2287         if (btrfs_leaf_free_space(root, leaf) < 0) {
2288                 btrfs_print_leaf(root, leaf);
2289                 BUG();
2290         }
2291         return ret;
2292 }
2293
2294 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2295                       struct btrfs_root *root, struct btrfs_path *path,
2296                       u32 data_size)
2297 {
2298         int ret = 0;
2299         int slot;
2300         struct extent_buffer *leaf;
2301         struct btrfs_item *item;
2302         u32 nritems;
2303         unsigned int data_end;
2304         unsigned int old_data;
2305         unsigned int old_size;
2306         int i;
2307
2308         leaf = path->nodes[0];
2309
2310         nritems = btrfs_header_nritems(leaf);
2311         data_end = leaf_data_end(root, leaf);
2312
2313         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2314                 btrfs_print_leaf(root, leaf);
2315                 BUG();
2316         }
2317         slot = path->slots[0];
2318         old_data = btrfs_item_end_nr(leaf, slot);
2319
2320         BUG_ON(slot < 0);
2321         if (slot >= nritems) {
2322                 btrfs_print_leaf(root, leaf);
2323                 printk("slot %d too large, nritems %d\n", slot, nritems);
2324                 BUG_ON(1);
2325         }
2326
2327         /*
2328          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2329          */
2330         /* first correct the data pointers */
2331         for (i = slot; i < nritems; i++) {
2332                 u32 ioff;
2333                 item = btrfs_item_nr(leaf, i);
2334                 ioff = btrfs_item_offset(leaf, item);
2335                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2336         }
2337
2338         /* shift the data */
2339         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2340                       data_end - data_size, btrfs_leaf_data(leaf) +
2341                       data_end, old_data - data_end);
2342
2343         data_end = old_data;
2344         old_size = btrfs_item_size_nr(leaf, slot);
2345         item = btrfs_item_nr(leaf, slot);
2346         btrfs_set_item_size(leaf, item, old_size + data_size);
2347         btrfs_mark_buffer_dirty(leaf);
2348
2349         ret = 0;
2350         if (btrfs_leaf_free_space(root, leaf) < 0) {
2351                 btrfs_print_leaf(root, leaf);
2352                 BUG();
2353         }
2354         return ret;
2355 }
2356
2357 /*
2358  * Given a key and some data, insert an item into the tree.
2359  * This does all the path init required, making room in the tree if needed.
2360  */
2361 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2362                             struct btrfs_root *root,
2363                             struct btrfs_path *path,
2364                             struct btrfs_key *cpu_key, u32 *data_size,
2365                             int nr)
2366 {
2367         struct extent_buffer *leaf;
2368         struct btrfs_item *item;
2369         int ret = 0;
2370         int slot;
2371         int i;
2372         u32 nritems;
2373         u32 total_size = 0;
2374         u32 total_data = 0;
2375         unsigned int data_end;
2376         struct btrfs_disk_key disk_key;
2377
2378         for (i = 0; i < nr; i++) {
2379                 total_data += data_size[i];
2380         }
2381
2382         /* create a root if there isn't one */
2383         if (!root->node)
2384                 BUG();
2385
2386         total_size = total_data + nr * sizeof(struct btrfs_item);
2387         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2388         if (ret == 0) {
2389                 return -EEXIST;
2390         }
2391         if (ret < 0)
2392                 goto out;
2393
2394         leaf = path->nodes[0];
2395
2396         nritems = btrfs_header_nritems(leaf);
2397         data_end = leaf_data_end(root, leaf);
2398
2399         if (btrfs_leaf_free_space(root, leaf) < total_size) {
2400                 btrfs_print_leaf(root, leaf);
2401                 printk("not enough freespace need %u have %d\n",
2402                        total_size, btrfs_leaf_free_space(root, leaf));
2403                 BUG();
2404         }
2405
2406         slot = path->slots[0];
2407         BUG_ON(slot < 0);
2408
2409         if (slot != nritems) {
2410                 int i;
2411                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2412
2413                 if (old_data < data_end) {
2414                         btrfs_print_leaf(root, leaf);
2415                         printk("slot %d old_data %d data_end %d\n",
2416                                slot, old_data, data_end);
2417                         BUG_ON(1);
2418                 }
2419                 /*
2420                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2421                  */
2422                 /* first correct the data pointers */
2423                 for (i = slot; i < nritems; i++) {
2424                         u32 ioff;
2425
2426                         item = btrfs_item_nr(leaf, i);
2427                         ioff = btrfs_item_offset(leaf, item);
2428                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2429                 }
2430
2431                 /* shift the items */
2432                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2433                               btrfs_item_nr_offset(slot),
2434                               (nritems - slot) * sizeof(struct btrfs_item));
2435
2436                 /* shift the data */
2437                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2438                               data_end - total_data, btrfs_leaf_data(leaf) +
2439                               data_end, old_data - data_end);
2440                 data_end = old_data;
2441         }
2442
2443         /* setup the item for the new data */
2444         for (i = 0; i < nr; i++) {
2445                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2446                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2447                 item = btrfs_item_nr(leaf, slot + i);
2448                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2449                 data_end -= data_size[i];
2450                 btrfs_set_item_size(leaf, item, data_size[i]);
2451         }
2452         btrfs_set_header_nritems(leaf, nritems + nr);
2453         btrfs_mark_buffer_dirty(leaf);
2454
2455         ret = 0;
2456         if (slot == 0) {
2457                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2458                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2459         }
2460
2461         if (btrfs_leaf_free_space(root, leaf) < 0) {
2462                 btrfs_print_leaf(root, leaf);
2463                 BUG();
2464         }
2465
2466 out:
2467         return ret;
2468 }
2469
2470 /*
2471  * Given a key and some data, insert an item into the tree.
2472  * This does all the path init required, making room in the tree if needed.
2473  */
2474 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2475                       *root, struct btrfs_key *cpu_key, void *data, u32
2476                       data_size)
2477 {
2478         int ret = 0;
2479         struct btrfs_path *path;
2480         struct extent_buffer *leaf;
2481         unsigned long ptr;
2482
2483         path = btrfs_alloc_path();
2484         BUG_ON(!path);
2485         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2486         if (!ret) {
2487                 leaf = path->nodes[0];
2488                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2489                 write_extent_buffer(leaf, data, ptr, data_size);
2490                 btrfs_mark_buffer_dirty(leaf);
2491         }
2492         btrfs_free_path(path);
2493         return ret;
2494 }
2495
2496 /*
2497  * delete the pointer from a given node.
2498  *
2499  * If the delete empties a node, the node is removed from the tree,
2500  * continuing all the way the root if required.  The root is converted into
2501  * a leaf if all the nodes are emptied.
2502  */
2503 int btrfs_del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2504                    struct btrfs_path *path, int level, int slot)
2505 {
2506         struct extent_buffer *parent = path->nodes[level];
2507         u32 nritems;
2508         int ret = 0;
2509         int wret;
2510
2511         nritems = btrfs_header_nritems(parent);
2512         if (slot != nritems -1) {
2513                 memmove_extent_buffer(parent,
2514                               btrfs_node_key_ptr_offset(slot),
2515                               btrfs_node_key_ptr_offset(slot + 1),
2516                               sizeof(struct btrfs_key_ptr) *
2517                               (nritems - slot - 1));
2518         }
2519         nritems--;
2520         btrfs_set_header_nritems(parent, nritems);
2521         if (nritems == 0 && parent == root->node) {
2522                 BUG_ON(btrfs_header_level(root->node) != 1);
2523                 /* just turn the root into a leaf and break */
2524                 btrfs_set_header_level(root->node, 0);
2525         } else if (slot == 0) {
2526                 struct btrfs_disk_key disk_key;
2527
2528                 btrfs_node_key(parent, &disk_key, 0);
2529                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2530                 if (wret)
2531                         ret = wret;
2532         }
2533         btrfs_mark_buffer_dirty(parent);
2534         return ret;
2535 }
2536
2537 /*
2538  * a helper function to delete the leaf pointed to by path->slots[1] and
2539  * path->nodes[1].
2540  *
2541  * This deletes the pointer in path->nodes[1] and frees the leaf
2542  * block extent.  zero is returned if it all worked out, < 0 otherwise.
2543  *
2544  * The path must have already been setup for deleting the leaf, including
2545  * all the proper balancing.  path->nodes[1] must be locked.
2546  */
2547 static noinline int btrfs_del_leaf(struct btrfs_trans_handle *trans,
2548                                    struct btrfs_root *root,
2549                                    struct btrfs_path *path,
2550                                    struct extent_buffer *leaf)
2551 {
2552         int ret;
2553
2554         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
2555         ret = btrfs_del_ptr(trans, root, path, 1, path->slots[1]);
2556         if (ret)
2557                 return ret;
2558
2559         ret = btrfs_free_extent(trans, root, leaf->start, leaf->len,
2560                                 0, root->root_key.objectid, 0, 0);
2561         return ret;
2562 }
2563
2564 /*
2565  * delete the item at the leaf level in path.  If that empties
2566  * the leaf, remove it from the tree
2567  */
2568 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2569                     struct btrfs_path *path, int slot, int nr)
2570 {
2571         struct extent_buffer *leaf;
2572         struct btrfs_item *item;
2573         int last_off;
2574         int dsize = 0;
2575         int ret = 0;
2576         int wret;
2577         int i;
2578         u32 nritems;
2579
2580         leaf = path->nodes[0];
2581         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2582
2583         for (i = 0; i < nr; i++)
2584                 dsize += btrfs_item_size_nr(leaf, slot + i);
2585
2586         nritems = btrfs_header_nritems(leaf);
2587
2588         if (slot + nr != nritems) {
2589                 int i;
2590                 int data_end = leaf_data_end(root, leaf);
2591
2592                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2593                               data_end + dsize,
2594                               btrfs_leaf_data(leaf) + data_end,
2595                               last_off - data_end);
2596
2597                 for (i = slot + nr; i < nritems; i++) {
2598                         u32 ioff;
2599
2600                         item = btrfs_item_nr(leaf, i);
2601                         ioff = btrfs_item_offset(leaf, item);
2602                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2603                 }
2604
2605                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2606                               btrfs_item_nr_offset(slot + nr),
2607                               sizeof(struct btrfs_item) *
2608                               (nritems - slot - nr));
2609         }
2610         btrfs_set_header_nritems(leaf, nritems - nr);
2611         nritems -= nr;
2612
2613         /* delete the leaf if we've emptied it */
2614         if (nritems == 0) {
2615                 if (leaf == root->node) {
2616                         btrfs_set_header_level(leaf, 0);
2617                 } else {
2618                         clean_tree_block(trans, root, leaf);
2619                         wait_on_tree_block_writeback(root, leaf);
2620
2621                         wret = btrfs_del_leaf(trans, root, path, leaf);
2622                         BUG_ON(ret);
2623                         if (wret)
2624                                 ret = wret;
2625                 }
2626         } else {
2627                 int used = leaf_space_used(leaf, 0, nritems);
2628                 if (slot == 0) {
2629                         struct btrfs_disk_key disk_key;
2630
2631                         btrfs_item_key(leaf, &disk_key, 0);
2632                         wret = fixup_low_keys(trans, root, path,
2633                                               &disk_key, 1);
2634                         if (wret)
2635                                 ret = wret;
2636                 }
2637
2638                 /* delete the leaf if it is mostly empty */
2639                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2640                         /* push_leaf_left fixes the path.
2641                          * make sure the path still points to our leaf
2642                          * for possible call to del_ptr below
2643                          */
2644                         slot = path->slots[1];
2645                         extent_buffer_get(leaf);
2646
2647                         wret = push_leaf_left(trans, root, path, 1, 1);
2648                         if (wret < 0 && wret != -ENOSPC)
2649                                 ret = wret;
2650
2651                         if (path->nodes[0] == leaf &&
2652                             btrfs_header_nritems(leaf)) {
2653                                 wret = push_leaf_right(trans, root, path, 1, 1);
2654                                 if (wret < 0 && wret != -ENOSPC)
2655                                         ret = wret;
2656                         }
2657
2658                         if (btrfs_header_nritems(leaf) == 0) {
2659                                 clean_tree_block(trans, root, leaf);
2660                                 wait_on_tree_block_writeback(root, leaf);
2661
2662                                 path->slots[1] = slot;
2663                                 ret = btrfs_del_leaf(trans, root, path, leaf);
2664                                 BUG_ON(ret);
2665                                 free_extent_buffer(leaf);
2666
2667                         } else {
2668                                 btrfs_mark_buffer_dirty(leaf);
2669                                 free_extent_buffer(leaf);
2670                         }
2671                 } else {
2672                         btrfs_mark_buffer_dirty(leaf);
2673                 }
2674         }
2675         return ret;
2676 }
2677
2678 /*
2679  * walk up the tree as far as required to find the previous leaf.
2680  * returns 0 if it found something or 1 if there are no lesser leaves.
2681  * returns < 0 on io errors.
2682  */
2683 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2684 {
2685         int slot;
2686         int level = 1;
2687         struct extent_buffer *c;
2688         struct extent_buffer *next = NULL;
2689
2690         while(level < BTRFS_MAX_LEVEL) {
2691                 if (!path->nodes[level])
2692                         return 1;
2693
2694                 slot = path->slots[level];
2695                 c = path->nodes[level];
2696                 if (slot == 0) {
2697                         level++;
2698                         if (level == BTRFS_MAX_LEVEL)
2699                                 return 1;
2700                         continue;
2701                 }
2702                 slot--;
2703
2704                 next = read_node_slot(root, c, slot);
2705                 break;
2706         }
2707         path->slots[level] = slot;
2708         while(1) {
2709                 level--;
2710                 c = path->nodes[level];
2711                 free_extent_buffer(c);
2712                 slot = btrfs_header_nritems(next);
2713                 if (slot != 0)
2714                         slot--;
2715                 path->nodes[level] = next;
2716                 path->slots[level] = slot;
2717                 if (!level)
2718                         break;
2719                 next = read_node_slot(root, next, slot);
2720         }
2721         return 0;
2722 }
2723
2724 /*
2725  * walk up the tree as far as required to find the next leaf.
2726  * returns 0 if it found something or 1 if there are no greater leaves.
2727  * returns < 0 on io errors.
2728  */
2729 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2730 {
2731         int slot;
2732         int level = 1;
2733         struct extent_buffer *c;
2734         struct extent_buffer *next = NULL;
2735
2736         while(level < BTRFS_MAX_LEVEL) {
2737                 if (!path->nodes[level])
2738                         return 1;
2739
2740                 slot = path->slots[level] + 1;
2741                 c = path->nodes[level];
2742                 if (slot >= btrfs_header_nritems(c)) {
2743                         level++;
2744                         if (level == BTRFS_MAX_LEVEL)
2745                                 return 1;
2746                         continue;
2747                 }
2748
2749                 if (path->reada)
2750                         reada_for_search(root, path, level, slot, 0);
2751
2752                 next = read_node_slot(root, c, slot);
2753                 if (!next)
2754                         return -EIO;
2755                 break;
2756         }
2757         path->slots[level] = slot;
2758         while(1) {
2759                 level--;
2760                 c = path->nodes[level];
2761                 free_extent_buffer(c);
2762                 path->nodes[level] = next;
2763                 path->slots[level] = 0;
2764                 if (!level)
2765                         break;
2766                 if (path->reada)
2767                         reada_for_search(root, path, level, 0, 0);
2768                 next = read_node_slot(root, next, 0);
2769                 if (!next)
2770                         return -EIO;
2771         }
2772         return 0;
2773 }
2774
2775 int btrfs_previous_item(struct btrfs_root *root,
2776                         struct btrfs_path *path, u64 min_objectid,
2777                         int type)
2778 {
2779         struct btrfs_key found_key;
2780         struct extent_buffer *leaf;
2781         int ret;
2782
2783         while(1) {
2784                 if (path->slots[0] == 0) {
2785                         ret = btrfs_prev_leaf(root, path);
2786                         if (ret != 0)
2787                                 return ret;
2788                 } else {
2789                         path->slots[0]--;
2790                 }
2791                 leaf = path->nodes[0];
2792                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2793                 if (found_key.type == type)
2794                         return 0;
2795         }
2796         return 1;
2797 }
2798