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
  23 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  24                       *root, struct btrfs_path *path, int level);
  25 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  26                       *root, struct btrfs_key *ins_key,
  27                       struct btrfs_path *path, int data_size, int extend);
  28 static int push_node_left(struct btrfs_trans_handle *trans,
  29                           struct btrfs_root *root, struct extent_buffer *dst,
  30                           struct extent_buffer *src, int empty);
  31 static int balance_node_right(struct btrfs_trans_handle *trans,
  32                               struct btrfs_root *root,
  33                               struct extent_buffer *dst_buf,
  34                               struct extent_buffer *src_buf);
  35 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  36                    struct btrfs_path *path, int level, int slot);
  37
  38 inline void btrfs_init_path(struct btrfs_path *p)
  39 {
  40         memset(p, 0, sizeof(*p));
  41 }
  42
  43 struct btrfs_path *btrfs_alloc_path(void)
  44 {
  45         struct btrfs_path *path;
  46         path = kmalloc(sizeof(struct btrfs_path), GFP_NOFS);
  47         if (path) {
  48                 btrfs_init_path(path);
  49                 path->reada = 0;
  50         }
  51         return path;
  52 }
  53
  54 void btrfs_free_path(struct btrfs_path *p)
  55 {
  56         btrfs_release_path(NULL, p);
  57         kfree(p);
  58 }
  59
  60 void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
  61 {
  62         int i;
  63         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  64                 if (!p->nodes[i])
  65                         continue;
  66                 free_extent_buffer(p->nodes[i]);
  67         }
  68         memset(p, 0, sizeof(*p));
  69 }
  70
  71 static void add_root_to_dirty_list(struct btrfs_root *root)
  72 {
  73         if (root->track_dirty && list_empty(&root->dirty_list)) {
  74                 list_add(&root->dirty_list,
  75                          &root->fs_info->dirty_cowonly_roots);
  76         }
  77 }
  78
  79 int btrfs_copy_root(struct btrfs_trans_handle *trans,
  80                       struct btrfs_root *root,
  81                       struct extent_buffer *buf,
  82                       struct extent_buffer **cow_ret, u64 new_root_objectid)
  83 {
  84         struct extent_buffer *cow;
  85         int ret = 0;
  86         int level;
  87         struct btrfs_root *new_root;
  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         cow = btrfs_alloc_free_block(trans, new_root, buf->len, 0,
 102                                      new_root_objectid, trans->transid,
 103                                      level, buf->start, 0);
 104         if (IS_ERR(cow)) {
 105                 kfree(new_root);
 106                 return PTR_ERR(cow);
 107         }
 108
 109         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 110         btrfs_set_header_bytenr(cow, cow->start);
 111         btrfs_set_header_generation(cow, trans->transid);
 112         btrfs_set_header_owner(cow, new_root_objectid);
 113         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN);
 114
 115         write_extent_buffer(cow, root->fs_info->fsid,
 116                             (unsigned long)btrfs_header_fsid(cow),
 117                             BTRFS_FSID_SIZE);
 118
 119         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 120         ret = btrfs_inc_ref(trans, new_root, buf, cow, NULL);
 121         kfree(new_root);
 122
 123         if (ret)
 124                 return ret;
 125
 126         btrfs_mark_buffer_dirty(cow);
 127         *cow_ret = cow;
 128         return 0;
 129 }
 130
 131 int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 132                              struct btrfs_root *root,
 133                              struct extent_buffer *buf,
 134                              struct extent_buffer *parent, int parent_slot,
 135                              struct extent_buffer **cow_ret,
 136                              u64 search_start, u64 empty_size)
 137 {
 138         u64 parent_start;
 139         struct extent_buffer *cow;
 140         u32 nritems;
 141         int ret = 0;
 142         int different_trans = 0;
 143         int level;
 144
 145         WARN_ON(root->ref_cows && trans->transid !=
 146                 root->fs_info->running_transaction->transid);
 147         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 148
 149         if (parent)
 150                 parent_start = parent->start;
 151         else
 152                 parent_start = 0;
 153
 154         level = btrfs_header_level(buf);
 155         nritems = btrfs_header_nritems(buf);
 156         cow = btrfs_alloc_free_block(trans, root, buf->len, parent_start,
 157                                      root->root_key.objectid, trans->transid,
 158                                      level, search_start, empty_size);
 159         if (IS_ERR(cow))
 160                 return PTR_ERR(cow);
 161
 162         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 163         btrfs_set_header_bytenr(cow, cow->start);
 164         btrfs_set_header_generation(cow, trans->transid);
 165         btrfs_set_header_owner(cow, root->root_key.objectid);
 166         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN);
 167
 168         write_extent_buffer(cow, root->fs_info->fsid,
 169                             (unsigned long)btrfs_header_fsid(cow),
 170                             BTRFS_FSID_SIZE);
 171
 172         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 173         if (btrfs_header_generation(buf) != trans->transid) {
 174                 different_trans = 1;
 175                 ret = btrfs_inc_ref(trans, root, buf, cow, NULL);
 176                 if (ret)
 177                         return ret;
 178         } else {
 179                 ret = btrfs_update_ref(trans, root, buf, cow, 0, nritems);
 180                 if (ret)
 181                         return ret;
 182                 clean_tree_block(trans, root, buf);
 183         }
 184
 185         if (buf == root->node) {
 186                 root->node = cow;
 187                 extent_buffer_get(cow);
 188                 if (buf != root->commit_root) {
 189                         btrfs_free_extent(trans, root, buf->start,
 190                                           buf->len, buf->start,
 191                                           root->root_key.objectid,
 192                                           btrfs_header_generation(buf),
 193                                           level, 1);
 194                 }
 195                 free_extent_buffer(buf);
 196                 add_root_to_dirty_list(root);
 197         } else {
 198                 btrfs_set_node_blockptr(parent, parent_slot,
 199                                         cow->start);
 200                 WARN_ON(trans->transid == 0);
 201                 btrfs_set_node_ptr_generation(parent, parent_slot,
 202                                               trans->transid);
 203                 btrfs_mark_buffer_dirty(parent);
 204                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 205                 btrfs_free_extent(trans, root, buf->start, buf->len,
 206                                   parent_start, btrfs_header_owner(parent),
 207                                   btrfs_header_generation(parent), level, 1);
 208         }
 209         free_extent_buffer(buf);
 210         btrfs_mark_buffer_dirty(cow);
 211         *cow_ret = cow;
 212         return 0;
 213 }
 214
 215 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 216                     struct btrfs_root *root, struct extent_buffer *buf,
 217                     struct extent_buffer *parent, int parent_slot,
 218                     struct extent_buffer **cow_ret)
 219 {
 220         u64 search_start;
 221         int ret;
 222         /*
 223         if (trans->transaction != root->fs_info->running_transaction) {
 224                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 225                        root->fs_info->running_transaction->transid);
 226                 WARN_ON(1);
 227         }
 228         */
 229         if (trans->transid != root->fs_info->generation) {
 230                 printk(KERN_CRIT "trans %llu running %llu\n",
 231                         (unsigned long long)trans->transid,
 232                         (unsigned long long)root->fs_info->generation);
 233                 WARN_ON(1);
 234         }
 235         if (btrfs_header_generation(buf) == trans->transid &&
 236             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
 237                 *cow_ret = buf;
 238                 return 0;
 239         }
 240
 241         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
 242         ret = __btrfs_cow_block(trans, root, buf, parent,
 243                                  parent_slot, cow_ret, search_start, 0);
 244         return ret;
 245 }
 246
 247 /*
 248 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 249 {
 250         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 251                 return 1;
 252         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 253                 return 1;
 254         return 0;
 255 }
 256 */
 257
 258 /*
 259  * compare two keys in a memcmp fashion
 260  */
 261 int btrfs_comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 262 {
 263         struct btrfs_key k1;
 264
 265         btrfs_disk_key_to_cpu(&k1, disk);
 266
 267         if (k1.objectid > k2->objectid)
 268                 return 1;
 269         if (k1.objectid < k2->objectid)
 270                 return -1;
 271         if (k1.type > k2->type)
 272                 return 1;
 273         if (k1.type < k2->type)
 274                 return -1;
 275         if (k1.offset > k2->offset)
 276                 return 1;
 277         if (k1.offset < k2->offset)
 278                 return -1;
 279         return 0;
 280 }
 281
 282
 283 #if 0
 284 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 285                        struct btrfs_root *root, struct extent_buffer *parent,
 286                        int start_slot, int cache_only, u64 *last_ret,
 287                        struct btrfs_key *progress)
 288 {
 289         struct extent_buffer *cur;
 290         struct extent_buffer *tmp;
 291         u64 blocknr;
 292         u64 gen;
 293         u64 search_start = *last_ret;
 294         u64 last_block = 0;
 295         u64 other;
 296         u32 parent_nritems;
 297         int end_slot;
 298         int i;
 299         int err = 0;
 300         int parent_level;
 301         int uptodate;
 302         u32 blocksize;
 303         int progress_passed = 0;
 304         struct btrfs_disk_key disk_key;
 305
 306         parent_level = btrfs_header_level(parent);
 307         if (cache_only && parent_level != 1)
 308                 return 0;
 309
 310         if (trans->transaction != root->fs_info->running_transaction) {
 311                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 312                        root->fs_info->running_transaction->transid);
 313                 WARN_ON(1);
 314         }
 315         if (trans->transid != root->fs_info->generation) {
 316                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 317                        root->fs_info->generation);
 318                 WARN_ON(1);
 319         }
 320
 321         parent_nritems = btrfs_header_nritems(parent);
 322         blocksize = btrfs_level_size(root, parent_level - 1);
 323         end_slot = parent_nritems;
 324
 325         if (parent_nritems == 1)
 326                 return 0;
 327
 328         for (i = start_slot; i < end_slot; i++) {
 329                 int close = 1;
 330
 331                 if (!parent->map_token) {
 332                         map_extent_buffer(parent,
 333                                         btrfs_node_key_ptr_offset(i),
 334                                         sizeof(struct btrfs_key_ptr),
 335                                         &parent->map_token, &parent->kaddr,
 336                                         &parent->map_start, &parent->map_len,
 337                                         KM_USER1);
 338                 }
 339                 btrfs_node_key(parent, &disk_key, i);
 340                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
 341                         continue;
 342
 343                 progress_passed = 1;
 344                 blocknr = btrfs_node_blockptr(parent, i);
 345                 gen = btrfs_node_ptr_generation(parent, i);
 346                 if (last_block == 0)
 347                         last_block = blocknr;
 348
 349                 if (i > 0) {
 350                         other = btrfs_node_blockptr(parent, i - 1);
 351                         close = close_blocks(blocknr, other, blocksize);
 352                 }
 353                 if (close && i < end_slot - 2) {
 354                         other = btrfs_node_blockptr(parent, i + 1);
 355                         close = close_blocks(blocknr, other, blocksize);
 356                 }
 357                 if (close) {
 358                         last_block = blocknr;
 359                         continue;
 360                 }
 361                 if (parent->map_token) {
 362                         unmap_extent_buffer(parent, parent->map_token,
 363                                             KM_USER1);
 364                         parent->map_token = NULL;
 365                 }
 366
 367                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
 368                 if (cur)
 369                         uptodate = btrfs_buffer_uptodate(cur, gen);
 370                 else
 371                         uptodate = 0;
 372                 if (!cur || !uptodate) {
 373                         if (cache_only) {
 374                                 free_extent_buffer(cur);
 375                                 continue;
 376                         }
 377                         if (!cur) {
 378                                 cur = read_tree_block(root, blocknr,
 379                                                          blocksize, gen);
 380                         } else if (!uptodate) {
 381                                 btrfs_read_buffer(cur, gen);
 382                         }
 383                 }
 384                 if (search_start == 0)
 385                         search_start = last_block;
 386
 387                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 388                                         &tmp, search_start,
 389                                         min(16 * blocksize,
 390                                             (end_slot - i) * blocksize));
 391                 if (err) {
 392                         free_extent_buffer(cur);
 393                         break;
 394                 }
 395                 search_start = tmp->start;
 396                 last_block = tmp->start;
 397                 *last_ret = search_start;
 398                 if (parent_level == 1)
 399                         btrfs_clear_buffer_defrag(tmp);
 400                 free_extent_buffer(tmp);
 401         }
 402         if (parent->map_token) {
 403                 unmap_extent_buffer(parent, parent->map_token,
 404                                     KM_USER1);
 405                 parent->map_token = NULL;
 406         }
 407         return err;
 408 }
 409 #endif
 410
 411 /*
 412  * The leaf data grows from end-to-front in the node.
 413  * this returns the address of the start of the last item,
 414  * which is the stop of the leaf data stack
 415  */
 416 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 417                                          struct extent_buffer *leaf)
 418 {
 419         u32 nr = btrfs_header_nritems(leaf);
 420         if (nr == 0)
 421                 return BTRFS_LEAF_DATA_SIZE(root);
 422         return btrfs_item_offset_nr(leaf, nr - 1);
 423 }
 424
 425 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
 426                       int level)
 427 {
 428         struct extent_buffer *parent = NULL;
 429         struct extent_buffer *node = path->nodes[level];
 430         struct btrfs_disk_key parent_key;
 431         struct btrfs_disk_key node_key;
 432         int parent_slot;
 433         int slot;
 434         struct btrfs_key cpukey;
 435         u32 nritems = btrfs_header_nritems(node);
 436
 437         if (path->nodes[level + 1])
 438                 parent = path->nodes[level + 1];
 439
 440         slot = path->slots[level];
 441         BUG_ON(nritems == 0);
 442         if (parent) {
 443                 parent_slot = path->slots[level + 1];
 444                 btrfs_node_key(parent, &parent_key, parent_slot);
 445                 btrfs_node_key(node, &node_key, 0);
 446                 BUG_ON(memcmp(&parent_key, &node_key,
 447                               sizeof(struct btrfs_disk_key)));
 448                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 449                        btrfs_header_bytenr(node));
 450         }
 451         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 452         if (slot != 0) {
 453                 btrfs_node_key_to_cpu(node, &cpukey, slot - 1);
 454                 btrfs_node_key(node, &node_key, slot);
 455                 BUG_ON(btrfs_comp_keys(&node_key, &cpukey) <= 0);
 456         }
 457         if (slot < nritems - 1) {
 458                 btrfs_node_key_to_cpu(node, &cpukey, slot + 1);
 459                 btrfs_node_key(node, &node_key, slot);
 460                 BUG_ON(btrfs_comp_keys(&node_key, &cpukey) >= 0);
 461         }
 462         return 0;
 463 }
 464
 465 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 466                       int level)
 467 {
 468         struct extent_buffer *leaf = path->nodes[level];
 469         struct extent_buffer *parent = NULL;
 470         int parent_slot;
 471         struct btrfs_key cpukey;
 472         struct btrfs_disk_key parent_key;
 473         struct btrfs_disk_key leaf_key;
 474         int slot = path->slots[0];
 475
 476         u32 nritems = btrfs_header_nritems(leaf);
 477
 478         if (path->nodes[level + 1])
 479                 parent = path->nodes[level + 1];
 480
 481         if (nritems == 0)
 482                 return 0;
 483
 484         if (parent) {
 485                 parent_slot = path->slots[level + 1];
 486                 btrfs_node_key(parent, &parent_key, parent_slot);
 487                 btrfs_item_key(leaf, &leaf_key, 0);
 488
 489                 BUG_ON(memcmp(&parent_key, &leaf_key,
 490                        sizeof(struct btrfs_disk_key)));
 491                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 492                        btrfs_header_bytenr(leaf));
 493         }
 494 #if 0
 495         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 496                 btrfs_item_key_to_cpu(leaf, &cpukey, i + 1);
 497                 btrfs_item_key(leaf, &leaf_key, i);
 498                 if (comp_keys(&leaf_key, &cpukey) >= 0) {
 499                         btrfs_print_leaf(root, leaf);
 500                         printk("slot %d offset bad key\n", i);
 501                         BUG_ON(1);
 502                 }
 503                 if (btrfs_item_offset_nr(leaf, i) !=
 504                         btrfs_item_end_nr(leaf, i + 1)) {
 505                         btrfs_print_leaf(root, leaf);
 506                         printk("slot %d offset bad\n", i);
 507                         BUG_ON(1);
 508                 }
 509                 if (i == 0) {
 510                         if (btrfs_item_offset_nr(leaf, i) +
 511                                btrfs_item_size_nr(leaf, i) !=
 512                                BTRFS_LEAF_DATA_SIZE(root)) {
 513                                 btrfs_print_leaf(root, leaf);
 514                                 printk("slot %d first offset bad\n", i);
 515                                 BUG_ON(1);
 516                         }
 517                 }
 518         }
 519         if (nritems > 0) {
 520                 if (btrfs_item_size_nr(leaf, nritems - 1) > 4096) {
 521                                 btrfs_print_leaf(root, leaf);
 522                                 printk("slot %d bad size \n", nritems - 1);
 523                                 BUG_ON(1);
 524                 }
 525         }
 526 #endif
 527         if (slot != 0 && slot < nritems - 1) {
 528                 btrfs_item_key(leaf, &leaf_key, slot);
 529                 btrfs_item_key_to_cpu(leaf, &cpukey, slot - 1);
 530                 if (btrfs_comp_keys(&leaf_key, &cpukey) <= 0) {
 531                         btrfs_print_leaf(root, leaf);
 532                         printk("slot %d offset bad key\n", slot);
 533                         BUG_ON(1);
 534                 }
 535                 if (btrfs_item_offset_nr(leaf, slot - 1) !=
 536                        btrfs_item_end_nr(leaf, slot)) {
 537                         btrfs_print_leaf(root, leaf);
 538                         printk("slot %d offset bad\n", slot);
 539                         BUG_ON(1);
 540                 }
 541         }
 542         if (slot < nritems - 1) {
 543                 btrfs_item_key(leaf, &leaf_key, slot);
 544                 btrfs_item_key_to_cpu(leaf, &cpukey, slot + 1);
 545                 BUG_ON(btrfs_comp_keys(&leaf_key, &cpukey) >= 0);
 546                 if (btrfs_item_offset_nr(leaf, slot) !=
 547                         btrfs_item_end_nr(leaf, slot + 1)) {
 548                         btrfs_print_leaf(root, leaf);
 549                         printk("slot %d offset bad\n", slot);
 550                         BUG_ON(1);
 551                 }
 552         }
 553         BUG_ON(btrfs_item_offset_nr(leaf, 0) +
 554                btrfs_item_size_nr(leaf, 0) != BTRFS_LEAF_DATA_SIZE(root));
 555         return 0;
 556 }
 557
 558 static int noinline check_block(struct btrfs_root *root,
 559                                 struct btrfs_path *path, int level)
 560 {
 561         return 0;
 562 #if 0
 563         struct extent_buffer *buf = path->nodes[level];
 564
 565         if (memcmp_extent_buffer(buf, root->fs_info->fsid,
 566                                  (unsigned long)btrfs_header_fsid(buf),
 567                                  BTRFS_FSID_SIZE)) {
 568                 printk("warning bad block %Lu\n", buf->start);
 569                 return 1;
 570         }
 571 #endif
 572         if (level == 0)
 573                 return check_leaf(root, path, level);
 574         return check_node(root, path, level);
 575 }
 576
 577 /*
 578  * search for key in the extent_buffer.  The items start at offset p,
 579  * and they are item_size apart.  There are 'max' items in p.
 580  *
 581  * the slot in the array is returned via slot, and it points to
 582  * the place where you would insert key if it is not found in
 583  * the array.
 584  *
 585  * slot may point to max if the key is bigger than all of the keys
 586  */
 587 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 588                               int item_size, struct btrfs_key *key,
 589                               int max, int *slot)
 590 {
 591         int low = 0;
 592         int high = max;
 593         int mid;
 594         int ret;
 595         unsigned long offset;
 596         struct btrfs_disk_key *tmp;
 597
 598         while(low < high) {
 599                 mid = (low + high) / 2;
 600                 offset = p + mid * item_size;
 601
 602                 tmp = (struct btrfs_disk_key *)(eb->data + offset);
 603                 ret = btrfs_comp_keys(tmp, key);
 604
 605                 if (ret < 0)
 606                         low = mid + 1;
 607                 else if (ret > 0)
 608                         high = mid;
 609                 else {
 610                         *slot = mid;
 611                         return 0;
 612                 }
 613         }
 614         *slot = low;
 615         return 1;
 616 }
 617
 618 /*
 619  * simple bin_search frontend that does the right thing for
 620  * leaves vs nodes
 621  */
 622 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 623                       int level, int *slot)
 624 {
 625         if (level == 0) {
 626                 return generic_bin_search(eb,
 627                                           offsetof(struct btrfs_leaf, items),
 628                                           sizeof(struct btrfs_item),
 629                                           key, btrfs_header_nritems(eb),
 630                                           slot);
 631         } else {
 632                 return generic_bin_search(eb,
 633                                           offsetof(struct btrfs_node, ptrs),
 634                                           sizeof(struct btrfs_key_ptr),
 635                                           key, btrfs_header_nritems(eb),
 636                                           slot);
 637         }
 638         return -1;
 639 }
 640
 641 static struct extent_buffer *read_node_slot(struct btrfs_root *root,
 642                                    struct extent_buffer *parent, int slot)
 643 {
 644         int level = btrfs_header_level(parent);
 645         if (slot < 0)
 646                 return NULL;
 647         if (slot >= btrfs_header_nritems(parent))
 648                 return NULL;
 649
 650         BUG_ON(level == 0);
 651
 652         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 653                        btrfs_level_size(root, level - 1),
 654                        btrfs_node_ptr_generation(parent, slot));
 655 }
 656
 657 static int balance_level(struct btrfs_trans_handle *trans,
 658                          struct btrfs_root *root,
 659                          struct btrfs_path *path, int level)
 660 {
 661         struct extent_buffer *right = NULL;
 662         struct extent_buffer *mid;
 663         struct extent_buffer *left = NULL;
 664         struct extent_buffer *parent = NULL;
 665         int ret = 0;
 666         int wret;
 667         int pslot;
 668         int orig_slot = path->slots[level];
 669         int err_on_enospc = 0;
 670         u64 orig_ptr;
 671
 672         if (level == 0)
 673                 return 0;
 674
 675         mid = path->nodes[level];
 676         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 677
 678         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 679
 680         if (level < BTRFS_MAX_LEVEL - 1)
 681                 parent = path->nodes[level + 1];
 682         pslot = path->slots[level + 1];
 683
 684         /*
 685          * deal with the case where there is only one pointer in the root
 686          * by promoting the node below to a root
 687          */
 688         if (!parent) {
 689                 struct extent_buffer *child;
 690
 691                 if (btrfs_header_nritems(mid) != 1)
 692                         return 0;
 693
 694                 /* promote the child to a root */
 695                 child = read_node_slot(root, mid, 0);
 696                 BUG_ON(!child);
 697                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
 698                 BUG_ON(ret);
 699
 700                 root->node = child;
 701
 702                 ret = btrfs_update_extent_ref(trans, root, child->start,
 703                                               mid->start, child->start,
 704                                               root->root_key.objectid,
 705                                               trans->transid, level - 1);
 706                 BUG_ON(ret);
 707
 708                 add_root_to_dirty_list(root);
 709                 path->nodes[level] = NULL;
 710                 clean_tree_block(trans, root, mid);
 711                 wait_on_tree_block_writeback(root, mid);
 712                 /* once for the path */
 713                 free_extent_buffer(mid);
 714                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 715                                         mid->start, root->root_key.objectid,
 716                                         btrfs_header_generation(mid),
 717                                         level, 1);
 718                 /* once for the root ptr */
 719                 free_extent_buffer(mid);
 720                 return ret;
 721         }
 722         if (btrfs_header_nritems(mid) >
 723             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 724                 return 0;
 725
 726         if (btrfs_header_nritems(mid) < 2)
 727                 err_on_enospc = 1;
 728
 729         left = read_node_slot(root, parent, pslot - 1);
 730         if (left) {
 731                 wret = btrfs_cow_block(trans, root, left,
 732                                        parent, pslot - 1, &left);
 733                 if (wret) {
 734                         ret = wret;
 735                         goto enospc;
 736                 }
 737         }
 738         right = read_node_slot(root, parent, pslot + 1);
 739         if (right) {
 740                 wret = btrfs_cow_block(trans, root, right,
 741                                        parent, pslot + 1, &right);
 742                 if (wret) {
 743                         ret = wret;
 744                         goto enospc;
 745                 }
 746         }
 747
 748         /* first, try to make some room in the middle buffer */
 749         if (left) {
 750                 orig_slot += btrfs_header_nritems(left);
 751                 wret = push_node_left(trans, root, left, mid, 1);
 752                 if (wret < 0)
 753                         ret = wret;
 754                 if (btrfs_header_nritems(mid) < 2)
 755                         err_on_enospc = 1;
 756         }
 757
 758         /*
 759          * then try to empty the right most buffer into the middle
 760          */
 761         if (right) {
 762                 wret = push_node_left(trans, root, mid, right, 1);
 763                 if (wret < 0 && wret != -ENOSPC)
 764                         ret = wret;
 765                 if (btrfs_header_nritems(right) == 0) {
 766                         u64 bytenr = right->start;
 767                         u64 generation = btrfs_header_generation(parent);
 768                         u32 blocksize = right->len;
 769
 770                         clean_tree_block(trans, root, right);
 771                         wait_on_tree_block_writeback(root, right);
 772                         free_extent_buffer(right);
 773                         right = NULL;
 774                         wret = del_ptr(trans, root, path, level + 1, pslot +
 775                                        1);
 776                         if (wret)
 777                                 ret = wret;
 778                         wret = btrfs_free_extent(trans, root, bytenr,
 779                                                  blocksize, parent->start,
 780                                                  btrfs_header_owner(parent),
 781                                                  generation, level, 1);
 782                         if (wret)
 783                                 ret = wret;
 784                 } else {
 785                         struct btrfs_disk_key right_key;
 786                         btrfs_node_key(right, &right_key, 0);
 787                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 788                         btrfs_mark_buffer_dirty(parent);
 789                 }
 790         }
 791         if (btrfs_header_nritems(mid) == 1) {
 792                 /*
 793                  * we're not allowed to leave a node with one item in the
 794                  * tree during a delete.  A deletion from lower in the tree
 795                  * could try to delete the only pointer in this node.
 796                  * So, pull some keys from the left.
 797                  * There has to be a left pointer at this point because
 798                  * otherwise we would have pulled some pointers from the
 799                  * right
 800                  */
 801                 BUG_ON(!left);
 802                 wret = balance_node_right(trans, root, mid, left);
 803                 if (wret < 0) {
 804                         ret = wret;
 805                         goto enospc;
 806                 }
 807                 if (wret == 1) {
 808                         wret = push_node_left(trans, root, left, mid, 1);
 809                         if (wret < 0)
 810                                 ret = wret;
 811                 }
 812                 BUG_ON(wret == 1);
 813         }
 814         if (btrfs_header_nritems(mid) == 0) {
 815                 /* we've managed to empty the middle node, drop it */
 816                 u64 root_gen = btrfs_header_generation(parent);
 817                 u64 bytenr = mid->start;
 818                 u32 blocksize = mid->len;
 819                 clean_tree_block(trans, root, mid);
 820                 wait_on_tree_block_writeback(root, mid);
 821                 free_extent_buffer(mid);
 822                 mid = NULL;
 823                 wret = del_ptr(trans, root, path, level + 1, pslot);
 824                 if (wret)
 825                         ret = wret;
 826                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 827                                          parent->start,
 828                                          btrfs_header_owner(parent),
 829                                          root_gen, level, 1);
 830                 if (wret)
 831                         ret = wret;
 832         } else {
 833                 /* update the parent key to reflect our changes */
 834                 struct btrfs_disk_key mid_key;
 835                 btrfs_node_key(mid, &mid_key, 0);
 836                 btrfs_set_node_key(parent, &mid_key, pslot);
 837                 btrfs_mark_buffer_dirty(parent);
 838         }
 839
 840         /* update the path */
 841         if (left) {
 842                 if (btrfs_header_nritems(left) > orig_slot) {
 843                         extent_buffer_get(left);
 844                         path->nodes[level] = left;
 845                         path->slots[level + 1] -= 1;
 846                         path->slots[level] = orig_slot;
 847                         if (mid)
 848                                 free_extent_buffer(mid);
 849                 } else {
 850                         orig_slot -= btrfs_header_nritems(left);
 851                         path->slots[level] = orig_slot;
 852                 }
 853         }
 854         /* double check we haven't messed things up */
 855         check_block(root, path, level);
 856         if (orig_ptr !=
 857             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 858                 BUG();
 859 enospc:
 860         if (right)
 861                 free_extent_buffer(right);
 862         if (left)
 863                 free_extent_buffer(left);
 864         return ret;
 865 }
 866
 867 /* returns zero if the push worked, non-zero otherwise */
 868 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
 869                                           struct btrfs_root *root,
 870                                           struct btrfs_path *path, int level)
 871 {
 872         struct extent_buffer *right = NULL;
 873         struct extent_buffer *mid;
 874         struct extent_buffer *left = NULL;
 875         struct extent_buffer *parent = NULL;
 876         int ret = 0;
 877         int wret;
 878         int pslot;
 879         int orig_slot = path->slots[level];
 880         u64 orig_ptr;
 881
 882         if (level == 0)
 883                 return 1;
 884
 885         mid = path->nodes[level];
 886         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 887         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 888
 889         if (level < BTRFS_MAX_LEVEL - 1)
 890                 parent = path->nodes[level + 1];
 891         pslot = path->slots[level + 1];
 892
 893         if (!parent)
 894                 return 1;
 895
 896         left = read_node_slot(root, parent, pslot - 1);
 897
 898         /* first, try to make some room in the middle buffer */
 899         if (left) {
 900                 u32 left_nr;
 901                 left_nr = btrfs_header_nritems(left);
 902                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 903                         wret = 1;
 904                 } else {
 905                         ret = btrfs_cow_block(trans, root, left, parent,
 906                                               pslot - 1, &left);
 907                         if (ret)
 908                                 wret = 1;
 909                         else {
 910                                 wret = push_node_left(trans, root,
 911                                                       left, mid, 0);
 912                         }
 913                 }
 914                 if (wret < 0)
 915                         ret = wret;
 916                 if (wret == 0) {
 917                         struct btrfs_disk_key disk_key;
 918                         orig_slot += left_nr;
 919                         btrfs_node_key(mid, &disk_key, 0);
 920                         btrfs_set_node_key(parent, &disk_key, pslot);
 921                         btrfs_mark_buffer_dirty(parent);
 922                         if (btrfs_header_nritems(left) > orig_slot) {
 923                                 path->nodes[level] = left;
 924                                 path->slots[level + 1] -= 1;
 925                                 path->slots[level] = orig_slot;
 926                                 free_extent_buffer(mid);
 927                         } else {
 928                                 orig_slot -=
 929                                         btrfs_header_nritems(left);
 930                                 path->slots[level] = orig_slot;
 931                                 free_extent_buffer(left);
 932                         }
 933                         return 0;
 934                 }
 935                 free_extent_buffer(left);
 936         }
 937         right= read_node_slot(root, parent, pslot + 1);
 938
 939         /*
 940          * then try to empty the right most buffer into the middle
 941          */
 942         if (right) {
 943                 u32 right_nr;
 944                 right_nr = btrfs_header_nritems(right);
 945                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 946                         wret = 1;
 947                 } else {
 948                         ret = btrfs_cow_block(trans, root, right,
 949                                               parent, pslot + 1,
 950                                               &right);
 951                         if (ret)
 952                                 wret = 1;
 953                         else {
 954                                 wret = balance_node_right(trans, root,
 955                                                           right, mid);
 956                         }
 957                 }
 958                 if (wret < 0)
 959                         ret = wret;
 960                 if (wret == 0) {
 961                         struct btrfs_disk_key disk_key;
 962
 963                         btrfs_node_key(right, &disk_key, 0);
 964                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
 965                         btrfs_mark_buffer_dirty(parent);
 966
 967                         if (btrfs_header_nritems(mid) <= orig_slot) {
 968                                 path->nodes[level] = right;
 969                                 path->slots[level + 1] += 1;
 970                                 path->slots[level] = orig_slot -
 971                                         btrfs_header_nritems(mid);
 972                                 free_extent_buffer(mid);
 973                         } else {
 974                                 free_extent_buffer(right);
 975                         }
 976                         return 0;
 977                 }
 978                 free_extent_buffer(right);
 979         }
 980         return 1;
 981 }
 982
 983 /*
 984  * readahead one full node of leaves
 985  */
 986 static void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
 987                              int level, int slot, u64 objectid)
 988 {
 989         struct extent_buffer *node;
 990         struct btrfs_disk_key disk_key;
 991         u32 nritems;
 992         u64 search;
 993         u64 lowest_read;
 994         u64 highest_read;
 995         u64 nread = 0;
 996         int direction = path->reada;
 997         struct extent_buffer *eb;
 998         u32 nr;
 999         u32 blocksize;
1000         u32 nscan = 0;
1001
1002         if (level != 1)
1003                 return;
1004
1005         if (!path->nodes[level])
1006                 return;
1007
1008         node = path->nodes[level];
1009         search = btrfs_node_blockptr(node, slot);
1010         blocksize = btrfs_level_size(root, level - 1);
1011         eb = btrfs_find_tree_block(root, search, blocksize);
1012         if (eb) {
1013                 free_extent_buffer(eb);
1014                 return;
1015         }
1016
1017         highest_read = search;
1018         lowest_read = search;
1019
1020         nritems = btrfs_header_nritems(node);
1021         nr = slot;
1022         while(1) {
1023                 if (direction < 0) {
1024                         if (nr == 0)
1025                                 break;
1026                         nr--;
1027                 } else if (direction > 0) {
1028                         nr++;
1029                         if (nr >= nritems)
1030                                 break;
1031                 }
1032                 if (path->reada < 0 && objectid) {
1033                         btrfs_node_key(node, &disk_key, nr);
1034                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1035                                 break;
1036                 }
1037                 search = btrfs_node_blockptr(node, nr);
1038                 if ((search >= lowest_read && search <= highest_read) ||
1039                     (search < lowest_read && lowest_read - search <= 32768) ||
1040                     (search > highest_read && search - highest_read <= 32768)) {
1041                         readahead_tree_block(root, search, blocksize,
1042                                      btrfs_node_ptr_generation(node, nr));
1043                         nread += blocksize;
1044                 }
1045                 nscan++;
1046                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
1047                         break;
1048                 if(nread > (1024 * 1024) || nscan > 128)
1049                         break;
1050
1051                 if (search < lowest_read)
1052                         lowest_read = search;
1053                 if (search > highest_read)
1054                         highest_read = search;
1055         }
1056 }
1057
1058 /*
1059  * look for key in the tree.  path is filled in with nodes along the way
1060  * if key is found, we return zero and you can find the item in the leaf
1061  * level of the path (level 0)
1062  *
1063  * If the key isn't found, the path points to the slot where it should
1064  * be inserted, and 1 is returned.  If there are other errors during the
1065  * search a negative error number is returned.
1066  *
1067  * if ins_len > 0, nodes and leaves will be split as we walk down the
1068  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1069  * possible)
1070  */
1071 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1072                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1073                       ins_len, int cow)
1074 {
1075         struct extent_buffer *b;
1076         int slot;
1077         int ret;
1078         int level;
1079         int should_reada = p->reada;
1080         u8 lowest_level = 0;
1081
1082         lowest_level = p->lowest_level;
1083         WARN_ON(lowest_level && ins_len);
1084         WARN_ON(p->nodes[0] != NULL);
1085         /*
1086         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
1087         */
1088 again:
1089         b = root->node;
1090         extent_buffer_get(b);
1091         while (b) {
1092                 level = btrfs_header_level(b);
1093                 if (cow) {
1094                         int wret;
1095                         wret = btrfs_cow_block(trans, root, b,
1096                                                p->nodes[level + 1],
1097                                                p->slots[level + 1],
1098                                                &b);
1099                         if (wret) {
1100                                 free_extent_buffer(b);
1101                                 return wret;
1102                         }
1103                 }
1104                 BUG_ON(!cow && ins_len);
1105                 if (level != btrfs_header_level(b))
1106                         WARN_ON(1);
1107                 level = btrfs_header_level(b);
1108                 p->nodes[level] = b;
1109                 ret = check_block(root, p, level);
1110                 if (ret)
1111                         return -1;
1112                 ret = bin_search(b, key, level, &slot);
1113                 if (level != 0) {
1114                         if (ret && slot > 0)
1115                                 slot -= 1;
1116                         p->slots[level] = slot;
1117                         if ((p->search_for_split || ins_len > 0) &&
1118                             btrfs_header_nritems(b) >=
1119                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1120                                 int sret = split_node(trans, root, p, level);
1121                                 BUG_ON(sret > 0);
1122                                 if (sret)
1123                                         return sret;
1124                                 b = p->nodes[level];
1125                                 slot = p->slots[level];
1126                         } else if (ins_len < 0) {
1127                                 int sret = balance_level(trans, root, p,
1128                                                          level);
1129                                 if (sret)
1130                                         return sret;
1131                                 b = p->nodes[level];
1132                                 if (!b) {
1133                                         btrfs_release_path(NULL, p);
1134                                         goto again;
1135                                 }
1136                                 slot = p->slots[level];
1137                                 BUG_ON(btrfs_header_nritems(b) == 1);
1138                         }
1139                         /* this is only true while dropping a snapshot */
1140                         if (level == lowest_level)
1141                                 break;
1142
1143                         if (should_reada)
1144                                 reada_for_search(root, p, level, slot,
1145                                                  key->objectid);
1146
1147                         b = read_node_slot(root, b, slot);
1148                 } else {
1149                         p->slots[level] = slot;
1150                         if (ins_len > 0 && btrfs_leaf_free_space(root, b) <
1151                             sizeof(struct btrfs_item) + ins_len) {
1152                                 int sret = split_leaf(trans, root, key,
1153                                                       p, ins_len, ret == 0);
1154                                 BUG_ON(sret > 0);
1155                                 if (sret)
1156                                         return sret;
1157                         }
1158                         return ret;
1159                 }
1160         }
1161         return 1;
1162 }
1163
1164 /*
1165  * adjust the pointers going up the tree, starting at level
1166  * making sure the right key of each node is points to 'key'.
1167  * This is used after shifting pointers to the left, so it stops
1168  * fixing up pointers when a given leaf/node is not in slot 0 of the
1169  * higher levels
1170  *
1171  * If this fails to write a tree block, it returns -1, but continues
1172  * fixing up the blocks in ram so the tree is consistent.
1173  */
1174 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1175                           struct btrfs_root *root, struct btrfs_path *path,
1176                           struct btrfs_disk_key *key, int level)
1177 {
1178         int i;
1179         int ret = 0;
1180         struct extent_buffer *t;
1181
1182         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1183                 int tslot = path->slots[i];
1184                 if (!path->nodes[i])
1185                         break;
1186                 t = path->nodes[i];
1187                 btrfs_set_node_key(t, key, tslot);
1188                 btrfs_mark_buffer_dirty(path->nodes[i]);
1189                 if (tslot != 0)
1190                         break;
1191         }
1192         return ret;
1193 }
1194
1195 /*
1196  * update item key.
1197  *
1198  * This function isn't completely safe. It's the caller's responsibility
1199  * that the new key won't break the order
1200  */
1201 int btrfs_set_item_key_safe(struct btrfs_trans_handle *trans,
1202                             struct btrfs_root *root, struct btrfs_path *path,
1203                             struct btrfs_key *new_key)
1204 {
1205         struct btrfs_disk_key disk_key;
1206         struct extent_buffer *eb;
1207         int slot;
1208
1209         eb = path->nodes[0];
1210         slot = path->slots[0];
1211         if (slot > 0) {
1212                 btrfs_item_key(eb, &disk_key, slot - 1);
1213                 if (btrfs_comp_keys(&disk_key, new_key) >= 0)
1214                         return -1;
1215         }
1216         if (slot < btrfs_header_nritems(eb) - 1) {
1217                 btrfs_item_key(eb, &disk_key, slot + 1);
1218                 if (btrfs_comp_keys(&disk_key, new_key) <= 0)
1219                         return -1;
1220         }
1221
1222         btrfs_cpu_key_to_disk(&disk_key, new_key);
1223         btrfs_set_item_key(eb, &disk_key, slot);
1224         btrfs_mark_buffer_dirty(eb);
1225         if (slot == 0)
1226                 fixup_low_keys(trans, root, path, &disk_key, 1);
1227         return 0;
1228 }
1229
1230 /*
1231  * try to push data from one node into the next node left in the
1232  * tree.
1233  *
1234  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1235  * error, and > 0 if there was no room in the left hand block.
1236  */
1237 static int push_node_left(struct btrfs_trans_handle *trans,
1238                           struct btrfs_root *root, struct extent_buffer *dst,
1239                           struct extent_buffer *src, int empty)
1240 {
1241         int push_items = 0;
1242         int src_nritems;
1243         int dst_nritems;
1244         int ret = 0;
1245
1246         src_nritems = btrfs_header_nritems(src);
1247         dst_nritems = btrfs_header_nritems(dst);
1248         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1249         WARN_ON(btrfs_header_generation(src) != trans->transid);
1250         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1251
1252         if (!empty && src_nritems <= 8)
1253                 return 1;
1254
1255         if (push_items <= 0) {
1256                 return 1;
1257         }
1258
1259         if (empty) {
1260                 push_items = min(src_nritems, push_items);
1261                 if (push_items < src_nritems) {
1262                         /* leave at least 8 pointers in the node if
1263                          * we aren't going to empty it
1264                          */
1265                         if (src_nritems - push_items < 8) {
1266                                 if (push_items <= 8)
1267                                         return 1;
1268                                 push_items -= 8;
1269                         }
1270                 }
1271         } else
1272                 push_items = min(src_nritems - 8, push_items);
1273
1274         copy_extent_buffer(dst, src,
1275                            btrfs_node_key_ptr_offset(dst_nritems),
1276                            btrfs_node_key_ptr_offset(0),
1277                            push_items * sizeof(struct btrfs_key_ptr));
1278
1279         if (push_items < src_nritems) {
1280                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1281                                       btrfs_node_key_ptr_offset(push_items),
1282                                       (src_nritems - push_items) *
1283                                       sizeof(struct btrfs_key_ptr));
1284         }
1285         btrfs_set_header_nritems(src, src_nritems - push_items);
1286         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1287         btrfs_mark_buffer_dirty(src);
1288         btrfs_mark_buffer_dirty(dst);
1289
1290         ret = btrfs_update_ref(trans, root, src, dst, dst_nritems, push_items);
1291         BUG_ON(ret);
1292         return ret;
1293 }
1294
1295 /*
1296  * try to push data from one node into the next node right in the
1297  * tree.
1298  *
1299  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1300  * error, and > 0 if there was no room in the right hand block.
1301  *
1302  * this will  only push up to 1/2 the contents of the left node over
1303  */
1304 static int balance_node_right(struct btrfs_trans_handle *trans,
1305                               struct btrfs_root *root,
1306                               struct extent_buffer *dst,
1307                               struct extent_buffer *src)
1308 {
1309         int push_items = 0;
1310         int max_push;
1311         int src_nritems;
1312         int dst_nritems;
1313         int ret = 0;
1314
1315         WARN_ON(btrfs_header_generation(src) != trans->transid);
1316         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1317
1318         src_nritems = btrfs_header_nritems(src);
1319         dst_nritems = btrfs_header_nritems(dst);
1320         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1321         if (push_items <= 0) {
1322                 return 1;
1323         }
1324
1325         if (src_nritems < 4) {
1326                 return 1;
1327         }
1328
1329         max_push = src_nritems / 2 + 1;
1330         /* don't try to empty the node */
1331         if (max_push >= src_nritems) {
1332                 return 1;
1333         }
1334
1335         if (max_push < push_items)
1336                 push_items = max_push;
1337
1338         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1339                                       btrfs_node_key_ptr_offset(0),
1340                                       (dst_nritems) *
1341                                       sizeof(struct btrfs_key_ptr));
1342
1343         copy_extent_buffer(dst, src,
1344                            btrfs_node_key_ptr_offset(0),
1345                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1346                            push_items * sizeof(struct btrfs_key_ptr));
1347
1348         btrfs_set_header_nritems(src, src_nritems - push_items);
1349         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1350
1351         btrfs_mark_buffer_dirty(src);
1352         btrfs_mark_buffer_dirty(dst);
1353
1354         ret = btrfs_update_ref(trans, root, src, dst, 0, push_items);
1355         BUG_ON(ret);
1356         return ret;
1357 }
1358
1359 /*
1360  * helper function to insert a new root level in the tree.
1361  * A new node is allocated, and a single item is inserted to
1362  * point to the existing root
1363  *
1364  * returns zero on success or < 0 on failure.
1365  */
1366 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1367                            struct btrfs_root *root,
1368                            struct btrfs_path *path, int level)
1369 {
1370         u64 lower_gen;
1371         struct extent_buffer *lower;
1372         struct extent_buffer *c;
1373         struct extent_buffer *old;
1374         struct btrfs_disk_key lower_key;
1375         int ret;
1376
1377         BUG_ON(path->nodes[level]);
1378         BUG_ON(path->nodes[level-1] != root->node);
1379
1380         lower = path->nodes[level-1];
1381         if (level == 1)
1382                 btrfs_item_key(lower, &lower_key, 0);
1383         else
1384                 btrfs_node_key(lower, &lower_key, 0);
1385
1386         c = btrfs_alloc_free_block(trans, root, root->nodesize, 0,
1387                                    root->root_key.objectid,
1388                                    trans->transid, level,
1389                                    root->node->start, 0);
1390         if (IS_ERR(c))
1391                 return PTR_ERR(c);
1392
1393         memset_extent_buffer(c, 0, 0, root->nodesize);
1394         btrfs_set_header_nritems(c, 1);
1395         btrfs_set_header_level(c, level);
1396         btrfs_set_header_bytenr(c, c->start);
1397         btrfs_set_header_generation(c, trans->transid);
1398         btrfs_set_header_owner(c, root->root_key.objectid);
1399
1400         write_extent_buffer(c, root->fs_info->fsid,
1401                             (unsigned long)btrfs_header_fsid(c),
1402                             BTRFS_FSID_SIZE);
1403
1404         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1405                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
1406                             BTRFS_UUID_SIZE);
1407
1408         btrfs_set_node_key(c, &lower_key, 0);
1409         btrfs_set_node_blockptr(c, 0, lower->start);
1410         lower_gen = btrfs_header_generation(lower);
1411         WARN_ON(lower_gen != trans->transid);
1412
1413         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1414
1415         btrfs_mark_buffer_dirty(c);
1416
1417         old = root->node;
1418         root->node = c;
1419
1420         ret = btrfs_update_extent_ref(trans, root, lower->start,
1421                                       lower->start, c->start,
1422                                       root->root_key.objectid,
1423                                       trans->transid, level - 1);
1424         BUG_ON(ret);
1425
1426         /* the super has an extra ref to root->node */
1427         free_extent_buffer(old);
1428
1429         add_root_to_dirty_list(root);
1430         extent_buffer_get(c);
1431         path->nodes[level] = c;
1432         path->slots[level] = 0;
1433         return 0;
1434 }
1435
1436 /*
1437  * worker function to insert a single pointer in a node.
1438  * the node should have enough room for the pointer already
1439  *
1440  * slot and level indicate where you want the key to go, and
1441  * blocknr is the block the key points to.
1442  *
1443  * returns zero on success and < 0 on any error
1444  */
1445 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1446                       *root, struct btrfs_path *path, struct btrfs_disk_key
1447                       *key, u64 bytenr, int slot, int level)
1448 {
1449         struct extent_buffer *lower;
1450         int nritems;
1451
1452         BUG_ON(!path->nodes[level]);
1453         lower = path->nodes[level];
1454         nritems = btrfs_header_nritems(lower);
1455         if (slot > nritems)
1456                 BUG();
1457         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1458                 BUG();
1459         if (slot != nritems) {
1460                 memmove_extent_buffer(lower,
1461                               btrfs_node_key_ptr_offset(slot + 1),
1462                               btrfs_node_key_ptr_offset(slot),
1463                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1464         }
1465         btrfs_set_node_key(lower, key, slot);
1466         btrfs_set_node_blockptr(lower, slot, bytenr);
1467         WARN_ON(trans->transid == 0);
1468         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1469         btrfs_set_header_nritems(lower, nritems + 1);
1470         btrfs_mark_buffer_dirty(lower);
1471         return 0;
1472 }
1473
1474 /*
1475  * split the node at the specified level in path in two.
1476  * The path is corrected to point to the appropriate node after the split
1477  *
1478  * Before splitting this tries to make some room in the node by pushing
1479  * left and right, if either one works, it returns right away.
1480  *
1481  * returns 0 on success and < 0 on failure
1482  */
1483 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1484                       *root, struct btrfs_path *path, int level)
1485 {
1486         struct extent_buffer *c;
1487         struct extent_buffer *split;
1488         struct btrfs_disk_key disk_key;
1489         int mid;
1490         int ret;
1491         int wret;
1492         u32 c_nritems;
1493
1494         c = path->nodes[level];
1495         WARN_ON(btrfs_header_generation(c) != trans->transid);
1496         if (c == root->node) {
1497                 /* trying to split the root, lets make a new one */
1498                 ret = insert_new_root(trans, root, path, level + 1);
1499                 if (ret)
1500                         return ret;
1501         } else {
1502                 ret = push_nodes_for_insert(trans, root, path, level);
1503                 c = path->nodes[level];
1504                 if (!ret && btrfs_header_nritems(c) <
1505                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
1506                         return 0;
1507                 if (ret < 0)
1508                         return ret;
1509         }
1510
1511         c_nritems = btrfs_header_nritems(c);
1512
1513         btrfs_node_key(c, &disk_key, 0);
1514         split = btrfs_alloc_free_block(trans, root, root->nodesize,
1515                                          path->nodes[level + 1]->start,
1516                                          root->root_key.objectid,
1517                                          trans->transid, level, c->start, 0);
1518         if (IS_ERR(split))
1519                 return PTR_ERR(split);
1520
1521         btrfs_set_header_flags(split, btrfs_header_flags(c));
1522         btrfs_set_header_level(split, btrfs_header_level(c));
1523         btrfs_set_header_bytenr(split, split->start);
1524         btrfs_set_header_generation(split, trans->transid);
1525         btrfs_set_header_owner(split, root->root_key.objectid);
1526         btrfs_set_header_flags(split, 0);
1527         write_extent_buffer(split, root->fs_info->fsid,
1528                             (unsigned long)btrfs_header_fsid(split),
1529                             BTRFS_FSID_SIZE);
1530         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1531                             (unsigned long)btrfs_header_chunk_tree_uuid(split),
1532                             BTRFS_UUID_SIZE);
1533
1534         mid = (c_nritems + 1) / 2;
1535
1536         copy_extent_buffer(split, c,
1537                            btrfs_node_key_ptr_offset(0),
1538                            btrfs_node_key_ptr_offset(mid),
1539                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1540         btrfs_set_header_nritems(split, c_nritems - mid);
1541         btrfs_set_header_nritems(c, mid);
1542         ret = 0;
1543
1544         btrfs_mark_buffer_dirty(c);
1545         btrfs_mark_buffer_dirty(split);
1546
1547         btrfs_node_key(split, &disk_key, 0);
1548         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1549                           path->slots[level + 1] + 1,
1550                           level + 1);
1551         if (wret)
1552                 ret = wret;
1553
1554         ret = btrfs_update_ref(trans, root, c, split, 0, c_nritems - mid);
1555         BUG_ON(ret);
1556
1557         if (path->slots[level] >= mid) {
1558                 path->slots[level] -= mid;
1559                 free_extent_buffer(c);
1560                 path->nodes[level] = split;
1561                 path->slots[level + 1] += 1;
1562         } else {
1563                 free_extent_buffer(split);
1564         }
1565         return ret;
1566 }
1567
1568 /*
1569  * how many bytes are required to store the items in a leaf.  start
1570  * and nr indicate which items in the leaf to check.  This totals up the
1571  * space used both by the item structs and the item data
1572  */
1573 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1574 {
1575         int data_len;
1576         int nritems = btrfs_header_nritems(l);
1577         int end = min(nritems, start + nr) - 1;
1578
1579         if (!nr)
1580                 return 0;
1581         data_len = btrfs_item_end_nr(l, start);
1582         data_len = data_len - btrfs_item_offset_nr(l, end);
1583         data_len += sizeof(struct btrfs_item) * nr;
1584         WARN_ON(data_len < 0);
1585         return data_len;
1586 }
1587
1588 /*
1589  * The space between the end of the leaf items and
1590  * the start of the leaf data.  IOW, how much room
1591  * the leaf has left for both items and data
1592  */
1593 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1594 {
1595         int nritems = btrfs_header_nritems(leaf);
1596         int ret;
1597         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1598         if (ret < 0) {
1599                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1600                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1601                        leaf_space_used(leaf, 0, nritems), nritems);
1602         }
1603         return ret;
1604 }
1605
1606 /*
1607  * push some data in the path leaf to the right, trying to free up at
1608  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1609  *
1610  * returns 1 if the push failed because the other node didn't have enough
1611  * room, 0 if everything worked out and < 0 if there were major errors.
1612  */
1613 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1614                            *root, struct btrfs_path *path, int data_size,
1615                            int empty)
1616 {
1617         struct extent_buffer *left = path->nodes[0];
1618         struct extent_buffer *right;
1619         struct extent_buffer *upper;
1620         struct btrfs_disk_key disk_key;
1621         int slot;
1622         u32 i;
1623         int free_space;
1624         int push_space = 0;
1625         int push_items = 0;
1626         struct btrfs_item *item;
1627         u32 left_nritems;
1628         u32 nr;
1629         u32 right_nritems;
1630         u32 data_end;
1631         u32 this_item_size;
1632         int ret;
1633
1634         slot = path->slots[1];
1635         if (!path->nodes[1]) {
1636                 return 1;
1637         }
1638         upper = path->nodes[1];
1639         if (slot >= btrfs_header_nritems(upper) - 1)
1640                 return 1;
1641
1642         right = read_node_slot(root, upper, slot + 1);
1643         free_space = btrfs_leaf_free_space(root, right);
1644         if (free_space < data_size + sizeof(struct btrfs_item)) {
1645                 free_extent_buffer(right);
1646                 return 1;
1647         }
1648
1649         /* cow and double check */
1650         ret = btrfs_cow_block(trans, root, right, upper,
1651                               slot + 1, &right);
1652         if (ret) {
1653                 free_extent_buffer(right);
1654                 return 1;
1655         }
1656         free_space = btrfs_leaf_free_space(root, right);
1657         if (free_space < data_size + sizeof(struct btrfs_item)) {
1658                 free_extent_buffer(right);
1659                 return 1;
1660         }
1661
1662         left_nritems = btrfs_header_nritems(left);
1663         if (left_nritems == 0) {
1664                 free_extent_buffer(right);
1665                 return 1;
1666         }
1667
1668         if (empty)
1669                 nr = 0;
1670         else
1671                 nr = 1;
1672
1673         i = left_nritems - 1;
1674         while (i >= nr) {
1675                 item = btrfs_item_nr(left, i);
1676
1677                 if (path->slots[0] == i)
1678                         push_space += data_size + sizeof(*item);
1679
1680                 this_item_size = btrfs_item_size(left, item);
1681                 if (this_item_size + sizeof(*item) + push_space > free_space)
1682                         break;
1683                 push_items++;
1684                 push_space += this_item_size + sizeof(*item);
1685                 if (i == 0)
1686                         break;
1687                 i--;
1688         }
1689
1690         if (push_items == 0) {
1691                 free_extent_buffer(right);
1692                 return 1;
1693         }
1694
1695         if (!empty && push_items == left_nritems)
1696                 WARN_ON(1);
1697
1698         /* push left to right */
1699         right_nritems = btrfs_header_nritems(right);
1700
1701         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1702         push_space -= leaf_data_end(root, left);
1703
1704         /* make room in the right data area */
1705         data_end = leaf_data_end(root, right);
1706         memmove_extent_buffer(right,
1707                               btrfs_leaf_data(right) + data_end - push_space,
1708                               btrfs_leaf_data(right) + data_end,
1709                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1710
1711         /* copy from the left data area */
1712         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1713                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1714                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1715                      push_space);
1716
1717         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1718                               btrfs_item_nr_offset(0),
1719                               right_nritems * sizeof(struct btrfs_item));
1720
1721         /* copy the items from left to right */
1722         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1723                    btrfs_item_nr_offset(left_nritems - push_items),
1724                    push_items * sizeof(struct btrfs_item));
1725
1726         /* update the item pointers */
1727         right_nritems += push_items;
1728         btrfs_set_header_nritems(right, right_nritems);
1729         push_space = BTRFS_LEAF_DATA_SIZE(root);
1730         for (i = 0; i < right_nritems; i++) {
1731                 item = btrfs_item_nr(right, i);
1732                 push_space -= btrfs_item_size(right, item);
1733                 btrfs_set_item_offset(right, item, push_space);
1734         }
1735
1736         left_nritems -= push_items;
1737         btrfs_set_header_nritems(left, left_nritems);
1738
1739         if (left_nritems)
1740                 btrfs_mark_buffer_dirty(left);
1741         btrfs_mark_buffer_dirty(right);
1742
1743         btrfs_item_key(right, &disk_key, 0);
1744         btrfs_set_node_key(upper, &disk_key, slot + 1);
1745         btrfs_mark_buffer_dirty(upper);
1746
1747         ret = btrfs_update_ref(trans, root, left, right, 0, push_items);
1748         BUG_ON(ret);
1749
1750         /* then fixup the leaf pointer in the path */
1751         if (path->slots[0] >= left_nritems) {
1752                 path->slots[0] -= left_nritems;
1753                 free_extent_buffer(path->nodes[0]);
1754                 path->nodes[0] = right;
1755                 path->slots[1] += 1;
1756         } else {
1757                 free_extent_buffer(right);
1758         }
1759         return 0;
1760 }
1761 /*
1762  * push some data in the path leaf to the left, trying to free up at
1763  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1764  */
1765 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1766                           *root, struct btrfs_path *path, int data_size,
1767                           int empty)
1768 {
1769         struct btrfs_disk_key disk_key;
1770         struct extent_buffer *right = path->nodes[0];
1771         struct extent_buffer *left;
1772         int slot;
1773         int i;
1774         int free_space;
1775         int push_space = 0;
1776         int push_items = 0;
1777         struct btrfs_item *item;
1778         u32 old_left_nritems;
1779         u32 right_nritems;
1780         u32 nr;
1781         int ret = 0;
1782         int wret;
1783         u32 this_item_size;
1784         u32 old_left_item_size;
1785
1786         slot = path->slots[1];
1787         if (slot == 0)
1788                 return 1;
1789         if (!path->nodes[1])
1790                 return 1;
1791
1792         right_nritems = btrfs_header_nritems(right);
1793         if (right_nritems == 0) {
1794                 return 1;
1795         }
1796
1797         left = read_node_slot(root, path->nodes[1], slot - 1);
1798         free_space = btrfs_leaf_free_space(root, left);
1799         if (free_space < data_size + sizeof(struct btrfs_item)) {
1800                 free_extent_buffer(left);
1801                 return 1;
1802         }
1803
1804         /* cow and double check */
1805         ret = btrfs_cow_block(trans, root, left,
1806                               path->nodes[1], slot - 1, &left);
1807         if (ret) {
1808                 /* we hit -ENOSPC, but it isn't fatal here */
1809                 free_extent_buffer(left);
1810                 return 1;
1811         }
1812
1813         free_space = btrfs_leaf_free_space(root, left);
1814         if (free_space < data_size + sizeof(struct btrfs_item)) {
1815                 free_extent_buffer(left);
1816                 return 1;
1817         }
1818
1819         if (empty)
1820                 nr = right_nritems;
1821         else
1822                 nr = right_nritems - 1;
1823
1824         for (i = 0; i < nr; i++) {
1825                 item = btrfs_item_nr(right, i);
1826
1827                 if (path->slots[0] == i)
1828                         push_space += data_size + sizeof(*item);
1829
1830                 this_item_size = btrfs_item_size(right, item);
1831                 if (this_item_size + sizeof(*item) + push_space > free_space)
1832                         break;
1833
1834                 push_items++;
1835                 push_space += this_item_size + sizeof(*item);
1836         }
1837
1838         if (push_items == 0) {
1839                 free_extent_buffer(left);
1840                 return 1;
1841         }
1842         if (!empty && push_items == btrfs_header_nritems(right))
1843                 WARN_ON(1);
1844
1845         /* push data from right to left */
1846         copy_extent_buffer(left, right,
1847                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1848                            btrfs_item_nr_offset(0),
1849                            push_items * sizeof(struct btrfs_item));
1850
1851         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1852                      btrfs_item_offset_nr(right, push_items -1);
1853
1854         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1855                      leaf_data_end(root, left) - push_space,
1856                      btrfs_leaf_data(right) +
1857                      btrfs_item_offset_nr(right, push_items - 1),
1858                      push_space);
1859         old_left_nritems = btrfs_header_nritems(left);
1860         BUG_ON(old_left_nritems < 0);
1861
1862         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1863         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1864                 u32 ioff;
1865
1866                 item = btrfs_item_nr(left, i);
1867                 ioff = btrfs_item_offset(left, item);
1868                 btrfs_set_item_offset(left, item,
1869                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1870         }
1871         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1872
1873         /* fixup right node */
1874         if (push_items > right_nritems) {
1875                 printk("push items %d nr %u\n", push_items, right_nritems);
1876                 WARN_ON(1);
1877         }
1878
1879         if (push_items < right_nritems) {
1880                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1881                                                   leaf_data_end(root, right);
1882                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
1883                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1884                                       btrfs_leaf_data(right) +
1885                                       leaf_data_end(root, right), push_space);
1886
1887                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1888                               btrfs_item_nr_offset(push_items),
1889                              (btrfs_header_nritems(right) - push_items) *
1890                              sizeof(struct btrfs_item));
1891         }
1892         right_nritems -= push_items;
1893         btrfs_set_header_nritems(right, right_nritems);
1894         push_space = BTRFS_LEAF_DATA_SIZE(root);
1895         for (i = 0; i < right_nritems; i++) {
1896                 item = btrfs_item_nr(right, i);
1897                 push_space = push_space - btrfs_item_size(right, item);
1898                 btrfs_set_item_offset(right, item, push_space);
1899         }
1900
1901         btrfs_mark_buffer_dirty(left);
1902         if (right_nritems)
1903                 btrfs_mark_buffer_dirty(right);
1904
1905         btrfs_item_key(right, &disk_key, 0);
1906         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1907         if (wret)
1908                 ret = wret;
1909
1910         ret = btrfs_update_ref(trans, root, right, left,
1911                                old_left_nritems, push_items);
1912         BUG_ON(ret);
1913
1914         /* then fixup the leaf pointer in the path */
1915         if (path->slots[0] < push_items) {
1916                 path->slots[0] += old_left_nritems;
1917                 free_extent_buffer(path->nodes[0]);
1918                 path->nodes[0] = left;
1919                 path->slots[1] -= 1;
1920         } else {
1921                 free_extent_buffer(left);
1922                 path->slots[0] -= push_items;
1923         }
1924         BUG_ON(path->slots[0] < 0);
1925         return ret;
1926 }
1927
1928 /*
1929  * split the path's leaf in two, making sure there is at least data_size
1930  * available for the resulting leaf level of the path.
1931  *
1932  * returns 0 if all went well and < 0 on failure.
1933  */
1934 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
1935                       *root, struct btrfs_key *ins_key,
1936                       struct btrfs_path *path, int data_size, int extend)
1937 {
1938         struct extent_buffer *l;
1939         u32 nritems;
1940         int mid;
1941         int slot;
1942         struct extent_buffer *right;
1943         int space_needed = data_size + sizeof(struct btrfs_item);
1944         int data_copy_size;
1945         int rt_data_off;
1946         int i;
1947         int ret = 0;
1948         int wret;
1949         int double_split;
1950         int num_doubles = 0;
1951         struct btrfs_disk_key disk_key;
1952
1953         if (extend && data_size)
1954                 space_needed = data_size;
1955
1956         /* first try to make some room by pushing left and right */
1957         if (data_size && ins_key->type != BTRFS_DIR_ITEM_KEY) {
1958                 wret = push_leaf_right(trans, root, path, data_size, 0);
1959                 if (wret < 0) {
1960                         return wret;
1961                 }
1962                 if (wret) {
1963                         wret = push_leaf_left(trans, root, path, data_size, 0);
1964                         if (wret < 0)
1965                                 return wret;
1966                 }
1967                 l = path->nodes[0];
1968
1969                 /* did the pushes work? */
1970                 if (btrfs_leaf_free_space(root, l) >= space_needed)
1971                         return 0;
1972         }
1973
1974         if (!path->nodes[1]) {
1975                 ret = insert_new_root(trans, root, path, 1);
1976                 if (ret)
1977                         return ret;
1978         }
1979 again:
1980         double_split = 0;
1981         l = path->nodes[0];
1982         slot = path->slots[0];
1983         nritems = btrfs_header_nritems(l);
1984         mid = (nritems + 1)/ 2;
1985
1986         right = btrfs_alloc_free_block(trans, root, root->leafsize,
1987                                         path->nodes[1]->start,
1988                                         root->root_key.objectid,
1989                                         trans->transid, 0, l->start, 0);
1990         if (IS_ERR(right)) {
1991                 BUG_ON(1);
1992                 return PTR_ERR(right);
1993         }
1994
1995         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
1996         btrfs_set_header_bytenr(right, right->start);
1997         btrfs_set_header_generation(right, trans->transid);
1998         btrfs_set_header_owner(right, root->root_key.objectid);
1999         btrfs_set_header_level(right, 0);
2000         write_extent_buffer(right, root->fs_info->fsid,
2001                             (unsigned long)btrfs_header_fsid(right),
2002                             BTRFS_FSID_SIZE);
2003
2004         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2005                             (unsigned long)btrfs_header_chunk_tree_uuid(right),
2006                             BTRFS_UUID_SIZE);
2007         if (mid <= slot) {
2008                 if (nritems == 1 ||
2009                     leaf_space_used(l, mid, nritems - mid) + space_needed >
2010                         BTRFS_LEAF_DATA_SIZE(root)) {
2011                         if (slot >= nritems) {
2012                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2013                                 btrfs_set_header_nritems(right, 0);
2014                                 wret = insert_ptr(trans, root, path,
2015                                                   &disk_key, right->start,
2016                                                   path->slots[1] + 1, 1);
2017                                 if (wret)
2018                                         ret = wret;
2019                                 free_extent_buffer(path->nodes[0]);
2020                                 path->nodes[0] = right;
2021                                 path->slots[0] = 0;
2022                                 path->slots[1] += 1;
2023                                 return ret;
2024                         }
2025                         mid = slot;
2026                         if (mid != nritems &&
2027                             leaf_space_used(l, mid, nritems - mid) +
2028                             space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2029                                 double_split = 1;
2030                         }
2031                 }
2032         } else {
2033                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
2034                         BTRFS_LEAF_DATA_SIZE(root)) {
2035                         if (!extend && data_size && slot == 0) {
2036                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2037                                 btrfs_set_header_nritems(right, 0);
2038                                 wret = insert_ptr(trans, root, path,
2039                                                   &disk_key,
2040                                                   right->start,
2041                                                   path->slots[1], 1);
2042                                 if (wret)
2043                                         ret = wret;
2044                                 free_extent_buffer(path->nodes[0]);
2045                                 path->nodes[0] = right;
2046                                 path->slots[0] = 0;
2047                                 if (path->slots[1] == 0) {
2048                                         wret = fixup_low_keys(trans, root,
2049                                                    path, &disk_key, 1);
2050                                         if (wret)
2051                                                 ret = wret;
2052                                 }
2053                                 return ret;
2054                         } else if ((extend || !data_size) && slot == 0) {
2055                                 mid = 1;
2056                         } else {
2057                                 mid = slot;
2058                                 if (mid != nritems &&
2059                                     leaf_space_used(l, mid, nritems - mid) +
2060                                     space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2061                                         double_split = 1;
2062                                 }
2063                         }
2064                 }
2065         }
2066         nritems = nritems - mid;
2067         btrfs_set_header_nritems(right, nritems);
2068         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2069
2070         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2071                            btrfs_item_nr_offset(mid),
2072                            nritems * sizeof(struct btrfs_item));
2073
2074         copy_extent_buffer(right, l,
2075                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2076                      data_copy_size, btrfs_leaf_data(l) +
2077                      leaf_data_end(root, l), data_copy_size);
2078
2079         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2080                       btrfs_item_end_nr(l, mid);
2081
2082         for (i = 0; i < nritems; i++) {
2083                 struct btrfs_item *item = btrfs_item_nr(right, i);
2084                 u32 ioff = btrfs_item_offset(right, item);
2085                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2086         }
2087
2088         btrfs_set_header_nritems(l, mid);
2089         ret = 0;
2090         btrfs_item_key(right, &disk_key, 0);
2091         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2092                           path->slots[1] + 1, 1);
2093         if (wret)
2094                 ret = wret;
2095
2096         btrfs_mark_buffer_dirty(right);
2097         btrfs_mark_buffer_dirty(l);
2098         BUG_ON(path->slots[0] != slot);
2099
2100         ret = btrfs_update_ref(trans, root, l, right, 0, nritems);
2101         BUG_ON(ret);
2102
2103         if (mid <= slot) {
2104                 free_extent_buffer(path->nodes[0]);
2105                 path->nodes[0] = right;
2106                 path->slots[0] -= mid;
2107                 path->slots[1] += 1;
2108         } else
2109                 free_extent_buffer(right);
2110
2111         BUG_ON(path->slots[0] < 0);
2112
2113         if (double_split) {
2114                 BUG_ON(num_doubles != 0);
2115                 num_doubles++;
2116                 goto again;
2117         }
2118         return ret;
2119 }
2120
2121 /*
2122  * This function splits a single item into two items,
2123  * giving 'new_key' to the new item and splitting the
2124  * old one at split_offset (from the start of the item).
2125  *
2126  * The path may be released by this operation.  After
2127  * the split, the path is pointing to the old item.  The
2128  * new item is going to be in the same node as the old one.
2129  *
2130  * Note, the item being split must be smaller enough to live alone on
2131  * a tree block with room for one extra struct btrfs_item
2132  *
2133  * This allows us to split the item in place, keeping a lock on the
2134  * leaf the entire time.
2135  */
2136 int btrfs_split_item(struct btrfs_trans_handle *trans,
2137                      struct btrfs_root *root,
2138                      struct btrfs_path *path,
2139                      struct btrfs_key *new_key,
2140                      unsigned long split_offset)
2141 {
2142         u32 item_size;
2143         struct extent_buffer *leaf;
2144         struct btrfs_key orig_key;
2145         struct btrfs_item *item;
2146         struct btrfs_item *new_item;
2147         int ret = 0;
2148         int slot;
2149         u32 nritems;
2150         u32 orig_offset;
2151         struct btrfs_disk_key disk_key;
2152         char *buf;
2153
2154         leaf = path->nodes[0];
2155         btrfs_item_key_to_cpu(leaf, &orig_key, path->slots[0]);
2156         if (btrfs_leaf_free_space(root, leaf) >= sizeof(struct btrfs_item))
2157                 goto split;
2158
2159         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2160         btrfs_release_path(root, path);
2161
2162         path->search_for_split = 1;
2163
2164         ret = btrfs_search_slot(trans, root, &orig_key, path, 0, 1);
2165         path->search_for_split = 0;
2166
2167         /* if our item isn't there or got smaller, return now */
2168         if (ret != 0 || item_size != btrfs_item_size_nr(path->nodes[0],
2169                                                         path->slots[0])) {
2170                 return -EAGAIN;
2171         }
2172
2173         ret = split_leaf(trans, root, &orig_key, path, 0, 0);
2174         BUG_ON(ret);
2175
2176         BUG_ON(btrfs_leaf_free_space(root, leaf) < sizeof(struct btrfs_item));
2177         leaf = path->nodes[0];
2178
2179 split:
2180         item = btrfs_item_nr(leaf, path->slots[0]);
2181         orig_offset = btrfs_item_offset(leaf, item);
2182         item_size = btrfs_item_size(leaf, item);
2183
2184
2185         buf = kmalloc(item_size, GFP_NOFS);
2186         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
2187                             path->slots[0]), item_size);
2188         slot = path->slots[0] + 1;
2189         leaf = path->nodes[0];
2190
2191         nritems = btrfs_header_nritems(leaf);
2192
2193         if (slot != nritems) {
2194                 /* shift the items */
2195                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
2196                               btrfs_item_nr_offset(slot),
2197                               (nritems - slot) * sizeof(struct btrfs_item));
2198
2199         }
2200
2201         btrfs_cpu_key_to_disk(&disk_key, new_key);
2202         btrfs_set_item_key(leaf, &disk_key, slot);
2203
2204         new_item = btrfs_item_nr(leaf, slot);
2205
2206         btrfs_set_item_offset(leaf, new_item, orig_offset);
2207         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
2208
2209         btrfs_set_item_offset(leaf, item,
2210                               orig_offset + item_size - split_offset);
2211         btrfs_set_item_size(leaf, item, split_offset);
2212
2213         btrfs_set_header_nritems(leaf, nritems + 1);
2214
2215         /* write the data for the start of the original item */
2216         write_extent_buffer(leaf, buf,
2217                             btrfs_item_ptr_offset(leaf, path->slots[0]),
2218                             split_offset);
2219
2220         /* write the data for the new item */
2221         write_extent_buffer(leaf, buf + split_offset,
2222                             btrfs_item_ptr_offset(leaf, slot),
2223                             item_size - split_offset);
2224         btrfs_mark_buffer_dirty(leaf);
2225
2226         ret = 0;
2227         if (btrfs_leaf_free_space(root, leaf) < 0) {
2228                 btrfs_print_leaf(root, leaf);
2229                 BUG();
2230         }
2231         kfree(buf);
2232         return ret;
2233 }
2234
2235 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2236                         struct btrfs_root *root,
2237                         struct btrfs_path *path,
2238                         u32 new_size, int from_end)
2239 {
2240         int ret = 0;
2241         int slot;
2242         int slot_orig;
2243         struct extent_buffer *leaf;
2244         struct btrfs_item *item;
2245         u32 nritems;
2246         unsigned int data_end;
2247         unsigned int old_data_start;
2248         unsigned int old_size;
2249         unsigned int size_diff;
2250         int i;
2251
2252         slot_orig = path->slots[0];
2253         leaf = path->nodes[0];
2254         slot = path->slots[0];
2255
2256         old_size = btrfs_item_size_nr(leaf, slot);
2257         if (old_size == new_size)
2258                 return 0;
2259
2260         nritems = btrfs_header_nritems(leaf);
2261         data_end = leaf_data_end(root, leaf);
2262
2263         old_data_start = btrfs_item_offset_nr(leaf, slot);
2264
2265         size_diff = old_size - new_size;
2266
2267         BUG_ON(slot < 0);
2268         BUG_ON(slot >= nritems);
2269
2270         /*
2271          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2272          */
2273         /* first correct the data pointers */
2274         for (i = slot; i < nritems; i++) {
2275                 u32 ioff;
2276                 item = btrfs_item_nr(leaf, i);
2277                 ioff = btrfs_item_offset(leaf, item);
2278                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2279         }
2280
2281         /* shift the data */
2282         if (from_end) {
2283                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2284                               data_end + size_diff, btrfs_leaf_data(leaf) +
2285                               data_end, old_data_start + new_size - data_end);
2286         } else {
2287                 struct btrfs_disk_key disk_key;
2288                 u64 offset;
2289
2290                 btrfs_item_key(leaf, &disk_key, slot);
2291
2292                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2293                         unsigned long ptr;
2294                         struct btrfs_file_extent_item *fi;
2295
2296                         fi = btrfs_item_ptr(leaf, slot,
2297                                             struct btrfs_file_extent_item);
2298                         fi = (struct btrfs_file_extent_item *)(
2299                              (unsigned long)fi - size_diff);
2300
2301                         if (btrfs_file_extent_type(leaf, fi) ==
2302                             BTRFS_FILE_EXTENT_INLINE) {
2303                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2304                                 memmove_extent_buffer(leaf, ptr,
2305                                         (unsigned long)fi,
2306                                         offsetof(struct btrfs_file_extent_item,
2307                                                  disk_bytenr));
2308                         }
2309                 }
2310
2311                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2312                               data_end + size_diff, btrfs_leaf_data(leaf) +
2313                               data_end, old_data_start - data_end);
2314
2315                 offset = btrfs_disk_key_offset(&disk_key);
2316                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2317                 btrfs_set_item_key(leaf, &disk_key, slot);
2318                 if (slot == 0)
2319                         fixup_low_keys(trans, root, path, &disk_key, 1);
2320         }
2321
2322         item = btrfs_item_nr(leaf, slot);
2323         btrfs_set_item_size(leaf, item, new_size);
2324         btrfs_mark_buffer_dirty(leaf);
2325
2326         ret = 0;
2327         if (btrfs_leaf_free_space(root, leaf) < 0) {
2328                 btrfs_print_leaf(root, leaf);
2329                 BUG();
2330         }
2331         return ret;
2332 }
2333
2334 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2335                       struct btrfs_root *root, struct btrfs_path *path,
2336                       u32 data_size)
2337 {
2338         int ret = 0;
2339         int slot;
2340         int slot_orig;
2341         struct extent_buffer *leaf;
2342         struct btrfs_item *item;
2343         u32 nritems;
2344         unsigned int data_end;
2345         unsigned int old_data;
2346         unsigned int old_size;
2347         int i;
2348
2349         slot_orig = path->slots[0];
2350         leaf = path->nodes[0];
2351
2352         nritems = btrfs_header_nritems(leaf);
2353         data_end = leaf_data_end(root, leaf);
2354
2355         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2356                 btrfs_print_leaf(root, leaf);
2357                 BUG();
2358         }
2359         slot = path->slots[0];
2360         old_data = btrfs_item_end_nr(leaf, slot);
2361
2362         BUG_ON(slot < 0);
2363         if (slot >= nritems) {
2364                 btrfs_print_leaf(root, leaf);
2365                 printk("slot %d too large, nritems %d\n", slot, nritems);
2366                 BUG_ON(1);
2367         }
2368
2369         /*
2370          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2371          */
2372         /* first correct the data pointers */
2373         for (i = slot; i < nritems; i++) {
2374                 u32 ioff;
2375                 item = btrfs_item_nr(leaf, i);
2376                 ioff = btrfs_item_offset(leaf, item);
2377                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2378         }
2379
2380         /* shift the data */
2381         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2382                       data_end - data_size, btrfs_leaf_data(leaf) +
2383                       data_end, old_data - data_end);
2384
2385         data_end = old_data;
2386         old_size = btrfs_item_size_nr(leaf, slot);
2387         item = btrfs_item_nr(leaf, slot);
2388         btrfs_set_item_size(leaf, item, old_size + data_size);
2389         btrfs_mark_buffer_dirty(leaf);
2390
2391         ret = 0;
2392         if (btrfs_leaf_free_space(root, leaf) < 0) {
2393                 btrfs_print_leaf(root, leaf);
2394                 BUG();
2395         }
2396         return ret;
2397 }
2398
2399 /*
2400  * Given a key and some data, insert an item into the tree.
2401  * This does all the path init required, making room in the tree if needed.
2402  */
2403 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2404                             struct btrfs_root *root,
2405                             struct btrfs_path *path,
2406                             struct btrfs_key *cpu_key, u32 *data_size,
2407                             int nr)
2408 {
2409         struct extent_buffer *leaf;
2410         struct btrfs_item *item;
2411         int ret = 0;
2412         int slot;
2413         int slot_orig;
2414         int i;
2415         u32 nritems;
2416         u32 total_size = 0;
2417         u32 total_data = 0;
2418         unsigned int data_end;
2419         struct btrfs_disk_key disk_key;
2420
2421         for (i = 0; i < nr; i++) {
2422                 total_data += data_size[i];
2423         }
2424
2425         /* create a root if there isn't one */
2426         if (!root->node)
2427                 BUG();
2428
2429         total_size = total_data + (nr - 1) * sizeof(struct btrfs_item);
2430         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2431         if (ret == 0) {
2432                 return -EEXIST;
2433         }
2434         if (ret < 0)
2435                 goto out;
2436
2437         slot_orig = path->slots[0];
2438         leaf = path->nodes[0];
2439
2440         nritems = btrfs_header_nritems(leaf);
2441         data_end = leaf_data_end(root, leaf);
2442
2443         if (btrfs_leaf_free_space(root, leaf) <
2444             sizeof(struct btrfs_item) + total_size) {
2445                 btrfs_print_leaf(root, leaf);
2446                 printk("not enough freespace need %u have %d\n",
2447                        total_size, btrfs_leaf_free_space(root, leaf));
2448                 BUG();
2449         }
2450
2451         slot = path->slots[0];
2452         BUG_ON(slot < 0);
2453
2454         if (slot != nritems) {
2455                 int i;
2456                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2457
2458                 if (old_data < data_end) {
2459                         btrfs_print_leaf(root, leaf);
2460                         printk("slot %d old_data %d data_end %d\n",
2461                                slot, old_data, data_end);
2462                         BUG_ON(1);
2463                 }
2464                 /*
2465                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2466                  */
2467                 /* first correct the data pointers */
2468                 for (i = slot; i < nritems; i++) {
2469                         u32 ioff;
2470
2471                         item = btrfs_item_nr(leaf, i);
2472                         ioff = btrfs_item_offset(leaf, item);
2473                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2474                 }
2475
2476                 /* shift the items */
2477                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2478                               btrfs_item_nr_offset(slot),
2479                               (nritems - slot) * sizeof(struct btrfs_item));
2480
2481                 /* shift the data */
2482                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2483                               data_end - total_data, btrfs_leaf_data(leaf) +
2484                               data_end, old_data - data_end);
2485                 data_end = old_data;
2486         }
2487
2488         /* setup the item for the new data */
2489         for (i = 0; i < nr; i++) {
2490                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2491                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2492                 item = btrfs_item_nr(leaf, slot + i);
2493                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2494                 data_end -= data_size[i];
2495                 btrfs_set_item_size(leaf, item, data_size[i]);
2496         }
2497         btrfs_set_header_nritems(leaf, nritems + nr);
2498         btrfs_mark_buffer_dirty(leaf);
2499
2500         ret = 0;
2501         if (slot == 0) {
2502                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2503                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2504         }
2505
2506         if (btrfs_leaf_free_space(root, leaf) < 0) {
2507                 btrfs_print_leaf(root, leaf);
2508                 BUG();
2509         }
2510
2511 out:
2512         return ret;
2513 }
2514
2515 /*
2516  * Given a key and some data, insert an item into the tree.
2517  * This does all the path init required, making room in the tree if needed.
2518  */
2519 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2520                       *root, struct btrfs_key *cpu_key, void *data, u32
2521                       data_size)
2522 {
2523         int ret = 0;
2524         struct btrfs_path *path;
2525         struct extent_buffer *leaf;
2526         unsigned long ptr;
2527
2528         path = btrfs_alloc_path();
2529         BUG_ON(!path);
2530         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2531         if (!ret) {
2532                 leaf = path->nodes[0];
2533                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2534                 write_extent_buffer(leaf, data, ptr, data_size);
2535                 btrfs_mark_buffer_dirty(leaf);
2536         }
2537         btrfs_free_path(path);
2538         return ret;
2539 }
2540
2541 /*
2542  * delete the pointer from a given node.
2543  *
2544  * If the delete empties a node, the node is removed from the tree,
2545  * continuing all the way the root if required.  The root is converted into
2546  * a leaf if all the nodes are emptied.
2547  */
2548 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2549                    struct btrfs_path *path, int level, int slot)
2550 {
2551         struct extent_buffer *parent = path->nodes[level];
2552         u32 nritems;
2553         int ret = 0;
2554         int wret;
2555
2556         nritems = btrfs_header_nritems(parent);
2557         if (slot != nritems -1) {
2558                 memmove_extent_buffer(parent,
2559                               btrfs_node_key_ptr_offset(slot),
2560                               btrfs_node_key_ptr_offset(slot + 1),
2561                               sizeof(struct btrfs_key_ptr) *
2562                               (nritems - slot - 1));
2563         }
2564         nritems--;
2565         btrfs_set_header_nritems(parent, nritems);
2566         if (nritems == 0 && parent == root->node) {
2567                 BUG_ON(btrfs_header_level(root->node) != 1);
2568                 /* just turn the root into a leaf and break */
2569                 btrfs_set_header_level(root->node, 0);
2570         } else if (slot == 0) {
2571                 struct btrfs_disk_key disk_key;
2572
2573                 btrfs_node_key(parent, &disk_key, 0);
2574                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2575                 if (wret)
2576                         ret = wret;
2577         }
2578         btrfs_mark_buffer_dirty(parent);
2579         return ret;
2580 }
2581
2582 /*
2583  * delete the item at the leaf level in path.  If that empties
2584  * the leaf, remove it from the tree
2585  */
2586 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2587                     struct btrfs_path *path, int slot, int nr)
2588 {
2589         struct extent_buffer *leaf;
2590         struct btrfs_item *item;
2591         int last_off;
2592         int dsize = 0;
2593         int ret = 0;
2594         int wret;
2595         int i;
2596         u32 nritems;
2597
2598         leaf = path->nodes[0];
2599         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2600
2601         for (i = 0; i < nr; i++)
2602                 dsize += btrfs_item_size_nr(leaf, slot + i);
2603
2604         nritems = btrfs_header_nritems(leaf);
2605
2606         if (slot + nr != nritems) {
2607                 int i;
2608                 int data_end = leaf_data_end(root, leaf);
2609
2610                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2611                               data_end + dsize,
2612                               btrfs_leaf_data(leaf) + data_end,
2613                               last_off - data_end);
2614
2615                 for (i = slot + nr; i < nritems; i++) {
2616                         u32 ioff;
2617
2618                         item = btrfs_item_nr(leaf, i);
2619                         ioff = btrfs_item_offset(leaf, item);
2620                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2621                 }
2622
2623                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2624                               btrfs_item_nr_offset(slot + nr),
2625                               sizeof(struct btrfs_item) *
2626                               (nritems - slot - nr));
2627         }
2628         btrfs_set_header_nritems(leaf, nritems - nr);
2629         nritems -= nr;
2630
2631         /* delete the leaf if we've emptied it */
2632         if (nritems == 0) {
2633                 if (leaf == root->node) {
2634                         btrfs_set_header_level(leaf, 0);
2635                 } else {
2636                         u64 root_gen = btrfs_header_generation(path->nodes[1]);
2637                         clean_tree_block(trans, root, leaf);
2638                         wait_on_tree_block_writeback(root, leaf);
2639                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
2640                         if (wret)
2641                                 ret = wret;
2642                         wret = btrfs_free_extent(trans, root,
2643                                          leaf->start, leaf->len,
2644                                          path->nodes[1]->start,
2645                                          btrfs_header_owner(path->nodes[1]),
2646                                          root_gen, 0, 1);
2647                         if (wret)
2648                                 ret = wret;
2649                 }
2650         } else {
2651                 int used = leaf_space_used(leaf, 0, nritems);
2652                 if (slot == 0) {
2653                         struct btrfs_disk_key disk_key;
2654
2655                         btrfs_item_key(leaf, &disk_key, 0);
2656                         wret = fixup_low_keys(trans, root, path,
2657                                               &disk_key, 1);
2658                         if (wret)
2659                                 ret = wret;
2660                 }
2661
2662                 /* delete the leaf if it is mostly empty */
2663                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2664                         /* push_leaf_left fixes the path.
2665                          * make sure the path still points to our leaf
2666                          * for possible call to del_ptr below
2667                          */
2668                         slot = path->slots[1];
2669                         extent_buffer_get(leaf);
2670
2671                         wret = push_leaf_left(trans, root, path, 1, 1);
2672                         if (wret < 0 && wret != -ENOSPC)
2673                                 ret = wret;
2674
2675                         if (path->nodes[0] == leaf &&
2676                             btrfs_header_nritems(leaf)) {
2677                                 wret = push_leaf_right(trans, root, path, 1, 1);
2678                                 if (wret < 0 && wret != -ENOSPC)
2679                                         ret = wret;
2680                         }
2681
2682                         if (btrfs_header_nritems(leaf) == 0) {
2683                                 u64 root_gen;
2684                                 u64 bytenr = leaf->start;
2685                                 u32 blocksize = leaf->len;
2686
2687                                 root_gen = btrfs_header_generation(
2688                                                            path->nodes[1]);
2689
2690                                 clean_tree_block(trans, root, leaf);
2691                                 wait_on_tree_block_writeback(root, leaf);
2692
2693                                 wret = del_ptr(trans, root, path, 1, slot);
2694                                 if (wret)
2695                                         ret = wret;
2696
2697                                 free_extent_buffer(leaf);
2698                                 wret = btrfs_free_extent(trans, root, bytenr,
2699                                              blocksize, path->nodes[1]->start,
2700                                              btrfs_header_owner(path->nodes[1]),
2701                                              root_gen, 0, 1);
2702                                 if (wret)
2703                                         ret = wret;
2704                         } else {
2705                                 btrfs_mark_buffer_dirty(leaf);
2706                                 free_extent_buffer(leaf);
2707                         }
2708                 } else {
2709                         btrfs_mark_buffer_dirty(leaf);
2710                 }
2711         }
2712         return ret;
2713 }
2714
2715 /*
2716  * walk up the tree as far as required to find the previous leaf.
2717  * returns 0 if it found something or 1 if there are no lesser leaves.
2718  * returns < 0 on io errors.
2719  */
2720 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2721 {
2722         int slot;
2723         int level = 1;
2724         struct extent_buffer *c;
2725         struct extent_buffer *next = NULL;
2726
2727         while(level < BTRFS_MAX_LEVEL) {
2728                 if (!path->nodes[level])
2729                         return 1;
2730
2731                 slot = path->slots[level];
2732                 c = path->nodes[level];
2733                 if (slot == 0) {
2734                         level++;
2735                         if (level == BTRFS_MAX_LEVEL)
2736                                 return 1;
2737                         continue;
2738                 }
2739                 slot--;
2740
2741                 if (next)
2742                         free_extent_buffer(next);
2743
2744                 next = read_node_slot(root, c, slot);
2745                 break;
2746         }
2747         path->slots[level] = slot;
2748         while(1) {
2749                 level--;
2750                 c = path->nodes[level];
2751                 free_extent_buffer(c);
2752                 slot = btrfs_header_nritems(next);
2753                 if (slot != 0)
2754                         slot--;
2755                 path->nodes[level] = next;
2756                 path->slots[level] = slot;
2757                 if (!level)
2758                         break;
2759                 next = read_node_slot(root, next, slot);
2760         }
2761         return 0;
2762 }
2763
2764 /*
2765  * walk up the tree as far as required to find the next leaf.
2766  * returns 0 if it found something or 1 if there are no greater leaves.
2767  * returns < 0 on io errors.
2768  */
2769 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2770 {
2771         int slot;
2772         int level = 1;
2773         struct extent_buffer *c;
2774         struct extent_buffer *next = NULL;
2775
2776         while(level < BTRFS_MAX_LEVEL) {
2777                 if (!path->nodes[level])
2778                         return 1;
2779
2780                 slot = path->slots[level] + 1;
2781                 c = path->nodes[level];
2782                 if (slot >= btrfs_header_nritems(c)) {
2783                         level++;
2784                         if (level == BTRFS_MAX_LEVEL)
2785                                 return 1;
2786                         continue;
2787                 }
2788
2789                 if (next)
2790                         free_extent_buffer(next);
2791
2792                 if (path->reada)
2793                         reada_for_search(root, path, level, slot, 0);
2794
2795                 next = read_node_slot(root, c, slot);
2796                 break;
2797         }
2798         path->slots[level] = slot;
2799         while(1) {
2800                 level--;
2801                 c = path->nodes[level];
2802                 free_extent_buffer(c);
2803                 path->nodes[level] = next;
2804                 path->slots[level] = 0;
2805                 if (!level)
2806                         break;
2807                 if (path->reada)
2808                         reada_for_search(root, path, level, 0, 0);
2809                 next = read_node_slot(root, next, 0);
2810         }
2811         return 0;
2812 }
2813
2814 int btrfs_previous_item(struct btrfs_root *root,
2815                         struct btrfs_path *path, u64 min_objectid,
2816                         int type)
2817 {
2818         struct btrfs_key found_key;
2819         struct extent_buffer *leaf;
2820         int ret;
2821
2822         while(1) {
2823                 if (path->slots[0] == 0) {
2824                         ret = btrfs_prev_leaf(root, path);
2825                         if (ret != 0)
2826                                 return ret;
2827                 } else {
2828                         path->slots[0]--;
2829                 }
2830                 leaf = path->nodes[0];
2831                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2832                 if (found_key.type == type)
2833                         return 0;
2834         }
2835         return 1;
2836 }
2837