fs/btrfs/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
  19 #include <linux/sched.h>
  20 #include "ctree.h"
  21 #include "disk-io.h"
  22 #include "transaction.h"
  23 #include "print-tree.h"
  24
  25 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  26                       *root, struct btrfs_path *path, int level);
  27 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  28                       *root, struct btrfs_key *ins_key,
  29                       struct btrfs_path *path, int data_size, int extend);
  30 static int push_node_left(struct btrfs_trans_handle *trans,
  31                           struct btrfs_root *root, struct extent_buffer *dst,
  32                           struct extent_buffer *src);
  33 static int balance_node_right(struct btrfs_trans_handle *trans,
  34                               struct btrfs_root *root,
  35                               struct extent_buffer *dst_buf,
  36                               struct extent_buffer *src_buf);
  37 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  38                    struct btrfs_path *path, int level, int slot);
  39
  40 inline void btrfs_init_path(struct btrfs_path *p)
  41 {
  42         memset(p, 0, sizeof(*p));
  43 }
  44
  45 struct btrfs_path *btrfs_alloc_path(void)
  46 {
  47         struct btrfs_path *path;
  48         path = kmem_cache_alloc(btrfs_path_cachep, GFP_NOFS);
  49         if (path) {
  50                 btrfs_init_path(path);
  51                 path->reada = 1;
  52         }
  53         return path;
  54 }
  55
  56 void btrfs_free_path(struct btrfs_path *p)
  57 {
  58         btrfs_release_path(NULL, p);
  59         kmem_cache_free(btrfs_path_cachep, p);
  60 }
  61
  62 void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
  63 {
  64         int i;
  65         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  66                 if (!p->nodes[i])
  67                         break;
  68                 free_extent_buffer(p->nodes[i]);
  69         }
  70         memset(p, 0, sizeof(*p));
  71 }
  72
  73 static void add_root_to_dirty_list(struct btrfs_root *root)
  74 {
  75         if (root->track_dirty && list_empty(&root->dirty_list)) {
  76                 list_add(&root->dirty_list,
  77                          &root->fs_info->dirty_cowonly_roots);
  78         }
  79 }
  80
  81 int btrfs_copy_root(struct btrfs_trans_handle *trans,
  82                       struct btrfs_root *root,
  83                       struct extent_buffer *buf,
  84                       struct extent_buffer **cow_ret, u64 new_root_objectid)
  85 {
  86         struct extent_buffer *cow;
  87         u32 nritems;
  88         int ret = 0;
  89         int level;
  90         struct btrfs_key first_key;
  91         struct btrfs_root *new_root;
  92
  93         new_root = kmalloc(sizeof(*new_root), GFP_NOFS);
  94         if (!new_root)
  95                 return -ENOMEM;
  96
  97         memcpy(new_root, root, sizeof(*new_root));
  98         new_root->root_key.objectid = new_root_objectid;
  99
 100         WARN_ON(root->ref_cows && trans->transid !=
 101                 root->fs_info->running_transaction->transid);
 102         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 103
 104         level = btrfs_header_level(buf);
 105         nritems = btrfs_header_nritems(buf);
 106         if (nritems) {
 107                 if (level == 0)
 108                         btrfs_item_key_to_cpu(buf, &first_key, 0);
 109                 else
 110                         btrfs_node_key_to_cpu(buf, &first_key, 0);
 111         } else {
 112                 first_key.objectid = 0;
 113         }
 114         cow = __btrfs_alloc_free_block(trans, new_root, buf->len,
 115                                        new_root_objectid,
 116                                        trans->transid, first_key.objectid,
 117                                        level, buf->start, 0);
 118         if (IS_ERR(cow)) {
 119                 kfree(new_root);
 120                 return PTR_ERR(cow);
 121         }
 122
 123         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 124         btrfs_set_header_bytenr(cow, cow->start);
 125         btrfs_set_header_generation(cow, trans->transid);
 126         btrfs_set_header_owner(cow, new_root_objectid);
 127         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN);
 128
 129         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 130         ret = btrfs_inc_ref(trans, new_root, buf);
 131         kfree(new_root);
 132
 133         if (ret)
 134                 return ret;
 135
 136         btrfs_mark_buffer_dirty(cow);
 137         *cow_ret = cow;
 138         return 0;
 139 }
 140
 141 int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 142                              struct btrfs_root *root,
 143                              struct extent_buffer *buf,
 144                              struct extent_buffer *parent, int parent_slot,
 145                              struct extent_buffer **cow_ret,
 146                              u64 search_start, u64 empty_size)
 147 {
 148         u64 root_gen;
 149         struct extent_buffer *cow;
 150         u32 nritems;
 151         int ret = 0;
 152         int different_trans = 0;
 153         int level;
 154         struct btrfs_key first_key;
 155
 156         if (root->ref_cows) {
 157                 root_gen = trans->transid;
 158         } else {
 159                 root_gen = 0;
 160         }
 161         WARN_ON(root->ref_cows && trans->transid !=
 162                 root->fs_info->running_transaction->transid);
 163         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 164
 165         level = btrfs_header_level(buf);
 166         nritems = btrfs_header_nritems(buf);
 167         if (nritems) {
 168                 if (level == 0)
 169                         btrfs_item_key_to_cpu(buf, &first_key, 0);
 170                 else
 171                         btrfs_node_key_to_cpu(buf, &first_key, 0);
 172         } else {
 173                 first_key.objectid = 0;
 174         }
 175         cow = __btrfs_alloc_free_block(trans, root, buf->len,
 176                                      root->root_key.objectid,
 177                                      root_gen, first_key.objectid, level,
 178                                      search_start, empty_size);
 179         if (IS_ERR(cow))
 180                 return PTR_ERR(cow);
 181
 182         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 183         btrfs_set_header_bytenr(cow, cow->start);
 184         btrfs_set_header_generation(cow, trans->transid);
 185         btrfs_set_header_owner(cow, root->root_key.objectid);
 186         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN);
 187
 188         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 189         if (btrfs_header_generation(buf) != trans->transid) {
 190                 different_trans = 1;
 191                 ret = btrfs_inc_ref(trans, root, buf);
 192                 if (ret)
 193                         return ret;
 194         } else {
 195                 clean_tree_block(trans, root, buf);
 196         }
 197
 198         if (buf == root->node) {
 199                 root_gen = btrfs_header_generation(buf);
 200                 root->node = cow;
 201                 extent_buffer_get(cow);
 202                 if (buf != root->commit_root) {
 203                         btrfs_free_extent(trans, root, buf->start,
 204                                           buf->len, root->root_key.objectid,
 205                                           root_gen, 0, 0, 1);
 206                 }
 207                 free_extent_buffer(buf);
 208                 add_root_to_dirty_list(root);
 209         } else {
 210                 root_gen = btrfs_header_generation(parent);
 211                 btrfs_set_node_blockptr(parent, parent_slot,
 212                                         cow->start);
 213                 WARN_ON(trans->transid == 0);
 214                 btrfs_set_node_ptr_generation(parent, parent_slot,
 215                                               trans->transid);
 216                 btrfs_mark_buffer_dirty(parent);
 217                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 218                 btrfs_free_extent(trans, root, buf->start, buf->len,
 219                                   btrfs_header_owner(parent), root_gen,
 220                                   0, 0, 1);
 221         }
 222         free_extent_buffer(buf);
 223         btrfs_mark_buffer_dirty(cow);
 224         *cow_ret = cow;
 225         return 0;
 226 }
 227
 228 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 229                     struct btrfs_root *root, struct extent_buffer *buf,
 230                     struct extent_buffer *parent, int parent_slot,
 231                     struct extent_buffer **cow_ret)
 232 {
 233         u64 search_start;
 234         u64 header_trans;
 235         int ret;
 236
 237         if (trans->transaction != root->fs_info->running_transaction) {
 238                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 239                        root->fs_info->running_transaction->transid);
 240                 WARN_ON(1);
 241         }
 242         if (trans->transid != root->fs_info->generation) {
 243                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 244                        root->fs_info->generation);
 245                 WARN_ON(1);
 246         }
 247
 248         header_trans = btrfs_header_generation(buf);
 249         spin_lock(&root->fs_info->hash_lock);
 250         if (header_trans == trans->transid &&
 251             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
 252                 *cow_ret = buf;
 253                 spin_unlock(&root->fs_info->hash_lock);
 254                 return 0;
 255         }
 256         spin_unlock(&root->fs_info->hash_lock);
 257         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
 258         ret = __btrfs_cow_block(trans, root, buf, parent,
 259                                  parent_slot, cow_ret, search_start, 0);
 260         return ret;
 261 }
 262
 263 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 264 {
 265         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 266                 return 1;
 267         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 268                 return 1;
 269         return 0;
 270 }
 271
 272 /*
 273  * compare two keys in a memcmp fashion
 274  */
 275 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 276 {
 277         struct btrfs_key k1;
 278
 279         btrfs_disk_key_to_cpu(&k1, disk);
 280
 281         if (k1.objectid > k2->objectid)
 282                 return 1;
 283         if (k1.objectid < k2->objectid)
 284                 return -1;
 285         if (k1.type > k2->type)
 286                 return 1;
 287         if (k1.type < k2->type)
 288                 return -1;
 289         if (k1.offset > k2->offset)
 290                 return 1;
 291         if (k1.offset < k2->offset)
 292                 return -1;
 293         return 0;
 294 }
 295
 296
 297 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 298                        struct btrfs_root *root, struct extent_buffer *parent,
 299                        int start_slot, int cache_only, u64 *last_ret,
 300                        struct btrfs_key *progress)
 301 {
 302         struct extent_buffer *cur;
 303         struct extent_buffer *tmp;
 304         u64 blocknr;
 305         u64 search_start = *last_ret;
 306         u64 last_block = 0;
 307         u64 other;
 308         u32 parent_nritems;
 309         int end_slot;
 310         int i;
 311         int err = 0;
 312         int parent_level;
 313         int uptodate;
 314         u32 blocksize;
 315         int progress_passed = 0;
 316         struct btrfs_disk_key disk_key;
 317
 318         parent_level = btrfs_header_level(parent);
 319         if (cache_only && parent_level != 1)
 320                 return 0;
 321
 322         if (trans->transaction != root->fs_info->running_transaction) {
 323                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 324                        root->fs_info->running_transaction->transid);
 325                 WARN_ON(1);
 326         }
 327         if (trans->transid != root->fs_info->generation) {
 328                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 329                        root->fs_info->generation);
 330                 WARN_ON(1);
 331         }
 332
 333         parent_nritems = btrfs_header_nritems(parent);
 334         blocksize = btrfs_level_size(root, parent_level - 1);
 335         end_slot = parent_nritems;
 336
 337         if (parent_nritems == 1)
 338                 return 0;
 339
 340         for (i = start_slot; i < end_slot; i++) {
 341                 int close = 1;
 342
 343                 if (!parent->map_token) {
 344                         map_extent_buffer(parent,
 345                                         btrfs_node_key_ptr_offset(i),
 346                                         sizeof(struct btrfs_key_ptr),
 347                                         &parent->map_token, &parent->kaddr,
 348                                         &parent->map_start, &parent->map_len,
 349                                         KM_USER1);
 350                 }
 351                 btrfs_node_key(parent, &disk_key, i);
 352                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
 353                         continue;
 354
 355                 progress_passed = 1;
 356                 blocknr = btrfs_node_blockptr(parent, i);
 357                 if (last_block == 0)
 358                         last_block = blocknr;
 359
 360                 if (i > 0) {
 361                         other = btrfs_node_blockptr(parent, i - 1);
 362                         close = close_blocks(blocknr, other, blocksize);
 363                 }
 364                 if (close && i < end_slot - 2) {
 365                         other = btrfs_node_blockptr(parent, i + 1);
 366                         close = close_blocks(blocknr, other, blocksize);
 367                 }
 368                 if (close) {
 369                         last_block = blocknr;
 370                         continue;
 371                 }
 372                 if (parent->map_token) {
 373                         unmap_extent_buffer(parent, parent->map_token,
 374                                             KM_USER1);
 375                         parent->map_token = NULL;
 376                 }
 377
 378                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
 379                 if (cur)
 380                         uptodate = btrfs_buffer_uptodate(cur);
 381                 else
 382                         uptodate = 0;
 383                 if (!cur || !uptodate) {
 384                         if (cache_only) {
 385                                 free_extent_buffer(cur);
 386                                 continue;
 387                         }
 388                         if (!cur) {
 389                                 cur = read_tree_block(root, blocknr,
 390                                                          blocksize);
 391                         } else if (!uptodate) {
 392                                 btrfs_read_buffer(cur);
 393                         }
 394                 }
 395                 if (search_start == 0)
 396                         search_start = last_block;
 397
 398                 btrfs_verify_block_csum(root, cur);
 399                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 400                                         &tmp, search_start,
 401                                         min(16 * blocksize,
 402                                             (end_slot - i) * blocksize));
 403                 if (err) {
 404                         free_extent_buffer(cur);
 405                         break;
 406                 }
 407                 search_start = tmp->start;
 408                 last_block = tmp->start;
 409                 *last_ret = search_start;
 410                 if (parent_level == 1)
 411                         btrfs_clear_buffer_defrag(tmp);
 412                 free_extent_buffer(tmp);
 413         }
 414         if (parent->map_token) {
 415                 unmap_extent_buffer(parent, parent->map_token,
 416                                     KM_USER1);
 417                 parent->map_token = NULL;
 418         }
 419         return err;
 420 }
 421
 422 /*
 423  * The leaf data grows from end-to-front in the node.
 424  * this returns the address of the start of the last item,
 425  * which is the stop of the leaf data stack
 426  */
 427 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 428                                          struct extent_buffer *leaf)
 429 {
 430         u32 nr = btrfs_header_nritems(leaf);
 431         if (nr == 0)
 432                 return BTRFS_LEAF_DATA_SIZE(root);
 433         return btrfs_item_offset_nr(leaf, nr - 1);
 434 }
 435
 436 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
 437                       int level)
 438 {
 439         struct extent_buffer *parent = NULL;
 440         struct extent_buffer *node = path->nodes[level];
 441         struct btrfs_disk_key parent_key;
 442         struct btrfs_disk_key node_key;
 443         int parent_slot;
 444         int slot;
 445         struct btrfs_key cpukey;
 446         u32 nritems = btrfs_header_nritems(node);
 447
 448         if (path->nodes[level + 1])
 449                 parent = path->nodes[level + 1];
 450
 451         slot = path->slots[level];
 452         BUG_ON(nritems == 0);
 453         if (parent) {
 454                 parent_slot = path->slots[level + 1];
 455                 btrfs_node_key(parent, &parent_key, parent_slot);
 456                 btrfs_node_key(node, &node_key, 0);
 457                 BUG_ON(memcmp(&parent_key, &node_key,
 458                               sizeof(struct btrfs_disk_key)));
 459                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 460                        btrfs_header_bytenr(node));
 461         }
 462         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 463         if (slot != 0) {
 464                 btrfs_node_key_to_cpu(node, &cpukey, slot - 1);
 465                 btrfs_node_key(node, &node_key, slot);
 466                 BUG_ON(comp_keys(&node_key, &cpukey) <= 0);
 467         }
 468         if (slot < nritems - 1) {
 469                 btrfs_node_key_to_cpu(node, &cpukey, slot + 1);
 470                 btrfs_node_key(node, &node_key, slot);
 471                 BUG_ON(comp_keys(&node_key, &cpukey) >= 0);
 472         }
 473         return 0;
 474 }
 475
 476 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 477                       int level)
 478 {
 479         struct extent_buffer *leaf = path->nodes[level];
 480         struct extent_buffer *parent = NULL;
 481         int parent_slot;
 482         struct btrfs_key cpukey;
 483         struct btrfs_disk_key parent_key;
 484         struct btrfs_disk_key leaf_key;
 485         int slot = path->slots[0];
 486
 487         u32 nritems = btrfs_header_nritems(leaf);
 488
 489         if (path->nodes[level + 1])
 490                 parent = path->nodes[level + 1];
 491
 492         if (nritems == 0)
 493                 return 0;
 494
 495         if (parent) {
 496                 parent_slot = path->slots[level + 1];
 497                 btrfs_node_key(parent, &parent_key, parent_slot);
 498                 btrfs_item_key(leaf, &leaf_key, 0);
 499
 500                 BUG_ON(memcmp(&parent_key, &leaf_key,
 501                        sizeof(struct btrfs_disk_key)));
 502                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 503                        btrfs_header_bytenr(leaf));
 504         }
 505 #if 0
 506         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 507                 btrfs_item_key_to_cpu(leaf, &cpukey, i + 1);
 508                 btrfs_item_key(leaf, &leaf_key, i);
 509                 if (comp_keys(&leaf_key, &cpukey) >= 0) {
 510                         btrfs_print_leaf(root, leaf);
 511                         printk("slot %d offset bad key\n", i);
 512                         BUG_ON(1);
 513                 }
 514                 if (btrfs_item_offset_nr(leaf, i) !=
 515                         btrfs_item_end_nr(leaf, i + 1)) {
 516                         btrfs_print_leaf(root, leaf);
 517                         printk("slot %d offset bad\n", i);
 518                         BUG_ON(1);
 519                 }
 520                 if (i == 0) {
 521                         if (btrfs_item_offset_nr(leaf, i) +
 522                                btrfs_item_size_nr(leaf, i) !=
 523                                BTRFS_LEAF_DATA_SIZE(root)) {
 524                                 btrfs_print_leaf(root, leaf);
 525                                 printk("slot %d first offset bad\n", i);
 526                                 BUG_ON(1);
 527                         }
 528                 }
 529         }
 530         if (nritems > 0) {
 531                 if (btrfs_item_size_nr(leaf, nritems - 1) > 4096) {
 532                                 btrfs_print_leaf(root, leaf);
 533                                 printk("slot %d bad size \n", nritems - 1);
 534                                 BUG_ON(1);
 535                 }
 536         }
 537 #endif
 538         if (slot != 0 && slot < nritems - 1) {
 539                 btrfs_item_key(leaf, &leaf_key, slot);
 540                 btrfs_item_key_to_cpu(leaf, &cpukey, slot - 1);
 541                 if (comp_keys(&leaf_key, &cpukey) <= 0) {
 542                         btrfs_print_leaf(root, leaf);
 543                         printk("slot %d offset bad key\n", slot);
 544                         BUG_ON(1);
 545                 }
 546                 if (btrfs_item_offset_nr(leaf, slot - 1) !=
 547                        btrfs_item_end_nr(leaf, slot)) {
 548                         btrfs_print_leaf(root, leaf);
 549                         printk("slot %d offset bad\n", slot);
 550                         BUG_ON(1);
 551                 }
 552         }
 553         if (slot < nritems - 1) {
 554                 btrfs_item_key(leaf, &leaf_key, slot);
 555                 btrfs_item_key_to_cpu(leaf, &cpukey, slot + 1);
 556                 BUG_ON(comp_keys(&leaf_key, &cpukey) >= 0);
 557                 if (btrfs_item_offset_nr(leaf, slot) !=
 558                         btrfs_item_end_nr(leaf, slot + 1)) {
 559                         btrfs_print_leaf(root, leaf);
 560                         printk("slot %d offset bad\n", slot);
 561                         BUG_ON(1);
 562                 }
 563         }
 564         BUG_ON(btrfs_item_offset_nr(leaf, 0) +
 565                btrfs_item_size_nr(leaf, 0) != BTRFS_LEAF_DATA_SIZE(root));
 566         return 0;
 567 }
 568
 569 static int noinline check_block(struct btrfs_root *root,
 570                                 struct btrfs_path *path, int level)
 571 {
 572         u64 found_start;
 573         return 0;
 574         if (btrfs_header_level(path->nodes[level]) != level)
 575             printk("warning: bad level %Lu wanted %d found %d\n",
 576                    path->nodes[level]->start, level,
 577                    btrfs_header_level(path->nodes[level]));
 578         found_start = btrfs_header_bytenr(path->nodes[level]);
 579         if (found_start != path->nodes[level]->start) {
 580             printk("warning: bad bytentr %Lu found %Lu\n",
 581                    path->nodes[level]->start, found_start);
 582         }
 583 #if 0
 584         struct extent_buffer *buf = path->nodes[level];
 585
 586         if (memcmp_extent_buffer(buf, root->fs_info->fsid,
 587                                  (unsigned long)btrfs_header_fsid(buf),
 588                                  BTRFS_FSID_SIZE)) {
 589                 printk("warning bad block %Lu\n", buf->start);
 590                 return 1;
 591         }
 592 #endif
 593         if (level == 0)
 594                 return check_leaf(root, path, level);
 595         return check_node(root, path, level);
 596 }
 597
 598 /*
 599  * search for key in the extent_buffer.  The items start at offset p,
 600  * and they are item_size apart.  There are 'max' items in p.
 601  *
 602  * the slot in the array is returned via slot, and it points to
 603  * the place where you would insert key if it is not found in
 604  * the array.
 605  *
 606  * slot may point to max if the key is bigger than all of the keys
 607  */
 608 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 609                               int item_size, struct btrfs_key *key,
 610                               int max, int *slot)
 611 {
 612         int low = 0;
 613         int high = max;
 614         int mid;
 615         int ret;
 616         struct btrfs_disk_key *tmp = NULL;
 617         struct btrfs_disk_key unaligned;
 618         unsigned long offset;
 619         char *map_token = NULL;
 620         char *kaddr = NULL;
 621         unsigned long map_start = 0;
 622         unsigned long map_len = 0;
 623         int err;
 624
 625         while(low < high) {
 626                 mid = (low + high) / 2;
 627                 offset = p + mid * item_size;
 628
 629                 if (!map_token || offset < map_start ||
 630                     (offset + sizeof(struct btrfs_disk_key)) >
 631                     map_start + map_len) {
 632                         if (map_token) {
 633                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 634                                 map_token = NULL;
 635                         }
 636                         err = map_extent_buffer(eb, offset,
 637                                                 sizeof(struct btrfs_disk_key),
 638                                                 &map_token, &kaddr,
 639                                                 &map_start, &map_len, KM_USER0);
 640
 641                         if (!err) {
 642                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
 643                                                         map_start);
 644                         } else {
 645                                 read_extent_buffer(eb, &unaligned,
 646                                                    offset, sizeof(unaligned));
 647                                 tmp = &unaligned;
 648                         }
 649
 650                 } else {
 651                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
 652                                                         map_start);
 653                 }
 654                 ret = comp_keys(tmp, key);
 655
 656                 if (ret < 0)
 657                         low = mid + 1;
 658                 else if (ret > 0)
 659                         high = mid;
 660                 else {
 661                         *slot = mid;
 662                         if (map_token)
 663                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 664                         return 0;
 665                 }
 666         }
 667         *slot = low;
 668         if (map_token)
 669                 unmap_extent_buffer(eb, map_token, KM_USER0);
 670         return 1;
 671 }
 672
 673 /*
 674  * simple bin_search frontend that does the right thing for
 675  * leaves vs nodes
 676  */
 677 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 678                       int level, int *slot)
 679 {
 680         if (level == 0) {
 681                 return generic_bin_search(eb,
 682                                           offsetof(struct btrfs_leaf, items),
 683                                           sizeof(struct btrfs_item),
 684                                           key, btrfs_header_nritems(eb),
 685                                           slot);
 686         } else {
 687                 return generic_bin_search(eb,
 688                                           offsetof(struct btrfs_node, ptrs),
 689                                           sizeof(struct btrfs_key_ptr),
 690                                           key, btrfs_header_nritems(eb),
 691                                           slot);
 692         }
 693         return -1;
 694 }
 695
 696 static struct extent_buffer *read_node_slot(struct btrfs_root *root,
 697                                    struct extent_buffer *parent, int slot)
 698 {
 699         if (slot < 0)
 700                 return NULL;
 701         if (slot >= btrfs_header_nritems(parent))
 702                 return NULL;
 703         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 704                        btrfs_level_size(root, btrfs_header_level(parent) - 1));
 705 }
 706
 707 static int balance_level(struct btrfs_trans_handle *trans,
 708                          struct btrfs_root *root,
 709                          struct btrfs_path *path, int level)
 710 {
 711         struct extent_buffer *right = NULL;
 712         struct extent_buffer *mid;
 713         struct extent_buffer *left = NULL;
 714         struct extent_buffer *parent = NULL;
 715         int ret = 0;
 716         int wret;
 717         int pslot;
 718         int orig_slot = path->slots[level];
 719         int err_on_enospc = 0;
 720         u64 orig_ptr;
 721
 722         if (level == 0)
 723                 return 0;
 724
 725         mid = path->nodes[level];
 726         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 727
 728         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 729
 730         if (level < BTRFS_MAX_LEVEL - 1)
 731                 parent = path->nodes[level + 1];
 732         pslot = path->slots[level + 1];
 733
 734         /*
 735          * deal with the case where there is only one pointer in the root
 736          * by promoting the node below to a root
 737          */
 738         if (!parent) {
 739                 struct extent_buffer *child;
 740
 741                 if (btrfs_header_nritems(mid) != 1)
 742                         return 0;
 743
 744                 /* promote the child to a root */
 745                 child = read_node_slot(root, mid, 0);
 746                 BUG_ON(!child);
 747                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
 748                 BUG_ON(ret);
 749
 750                 root->node = child;
 751                 add_root_to_dirty_list(root);
 752                 path->nodes[level] = NULL;
 753                 clean_tree_block(trans, root, mid);
 754                 wait_on_tree_block_writeback(root, mid);
 755                 /* once for the path */
 756                 free_extent_buffer(mid);
 757                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 758                                         root->root_key.objectid,
 759                                         btrfs_header_generation(mid), 0, 0, 1);
 760                 /* once for the root ptr */
 761                 free_extent_buffer(mid);
 762                 return ret;
 763         }
 764         if (btrfs_header_nritems(mid) >
 765             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 766                 return 0;
 767
 768         if (btrfs_header_nritems(mid) < 2)
 769                 err_on_enospc = 1;
 770
 771         left = read_node_slot(root, parent, pslot - 1);
 772         if (left) {
 773                 wret = btrfs_cow_block(trans, root, left,
 774                                        parent, pslot - 1, &left);
 775                 if (wret) {
 776                         ret = wret;
 777                         goto enospc;
 778                 }
 779         }
 780         right = read_node_slot(root, parent, pslot + 1);
 781         if (right) {
 782                 wret = btrfs_cow_block(trans, root, right,
 783                                        parent, pslot + 1, &right);
 784                 if (wret) {
 785                         ret = wret;
 786                         goto enospc;
 787                 }
 788         }
 789
 790         /* first, try to make some room in the middle buffer */
 791         if (left) {
 792                 orig_slot += btrfs_header_nritems(left);
 793                 wret = push_node_left(trans, root, left, mid);
 794                 if (wret < 0)
 795                         ret = wret;
 796                 if (btrfs_header_nritems(mid) < 2)
 797                         err_on_enospc = 1;
 798         }
 799
 800         /*
 801          * then try to empty the right most buffer into the middle
 802          */
 803         if (right) {
 804                 wret = push_node_left(trans, root, mid, right);
 805                 if (wret < 0 && wret != -ENOSPC)
 806                         ret = wret;
 807                 if (btrfs_header_nritems(right) == 0) {
 808                         u64 bytenr = right->start;
 809                         u64 generation = btrfs_header_generation(parent);
 810                         u32 blocksize = right->len;
 811
 812                         clean_tree_block(trans, root, right);
 813                         wait_on_tree_block_writeback(root, right);
 814                         free_extent_buffer(right);
 815                         right = NULL;
 816                         wret = del_ptr(trans, root, path, level + 1, pslot +
 817                                        1);
 818                         if (wret)
 819                                 ret = wret;
 820                         wret = btrfs_free_extent(trans, root, bytenr,
 821                                                  blocksize,
 822                                                  btrfs_header_owner(parent),
 823                                                  generation, 0, 0, 1);
 824                         if (wret)
 825                                 ret = wret;
 826                 } else {
 827                         struct btrfs_disk_key right_key;
 828                         btrfs_node_key(right, &right_key, 0);
 829                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 830                         btrfs_mark_buffer_dirty(parent);
 831                 }
 832         }
 833         if (btrfs_header_nritems(mid) == 1) {
 834                 /*
 835                  * we're not allowed to leave a node with one item in the
 836                  * tree during a delete.  A deletion from lower in the tree
 837                  * could try to delete the only pointer in this node.
 838                  * So, pull some keys from the left.
 839                  * There has to be a left pointer at this point because
 840                  * otherwise we would have pulled some pointers from the
 841                  * right
 842                  */
 843                 BUG_ON(!left);
 844                 wret = balance_node_right(trans, root, mid, left);
 845                 if (wret < 0) {
 846                         ret = wret;
 847                         goto enospc;
 848                 }
 849                 BUG_ON(wret == 1);
 850         }
 851         if (btrfs_header_nritems(mid) == 0) {
 852                 /* we've managed to empty the middle node, drop it */
 853                 u64 root_gen = btrfs_header_generation(parent);
 854                 u64 bytenr = mid->start;
 855                 u32 blocksize = mid->len;
 856                 clean_tree_block(trans, root, mid);
 857                 wait_on_tree_block_writeback(root, mid);
 858                 free_extent_buffer(mid);
 859                 mid = NULL;
 860                 wret = del_ptr(trans, root, path, level + 1, pslot);
 861                 if (wret)
 862                         ret = wret;
 863                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 864                                          btrfs_header_owner(parent),
 865                                          root_gen, 0, 0, 1);
 866                 if (wret)
 867                         ret = wret;
 868         } else {
 869                 /* update the parent key to reflect our changes */
 870                 struct btrfs_disk_key mid_key;
 871                 btrfs_node_key(mid, &mid_key, 0);
 872                 btrfs_set_node_key(parent, &mid_key, pslot);
 873                 btrfs_mark_buffer_dirty(parent);
 874         }
 875
 876         /* update the path */
 877         if (left) {
 878                 if (btrfs_header_nritems(left) > orig_slot) {
 879                         extent_buffer_get(left);
 880                         path->nodes[level] = left;
 881                         path->slots[level + 1] -= 1;
 882                         path->slots[level] = orig_slot;
 883                         if (mid)
 884                                 free_extent_buffer(mid);
 885                 } else {
 886                         orig_slot -= btrfs_header_nritems(left);
 887                         path->slots[level] = orig_slot;
 888                 }
 889         }
 890         /* double check we haven't messed things up */
 891         check_block(root, path, level);
 892         if (orig_ptr !=
 893             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 894                 BUG();
 895 enospc:
 896         if (right)
 897                 free_extent_buffer(right);
 898         if (left)
 899                 free_extent_buffer(left);
 900         return ret;
 901 }
 902
 903 /* returns zero if the push worked, non-zero otherwise */
 904 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
 905                                           struct btrfs_root *root,
 906                                           struct btrfs_path *path, int level)
 907 {
 908         struct extent_buffer *right = NULL;
 909         struct extent_buffer *mid;
 910         struct extent_buffer *left = NULL;
 911         struct extent_buffer *parent = NULL;
 912         int ret = 0;
 913         int wret;
 914         int pslot;
 915         int orig_slot = path->slots[level];
 916         u64 orig_ptr;
 917
 918         if (level == 0)
 919                 return 1;
 920
 921         mid = path->nodes[level];
 922         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 923         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 924
 925         if (level < BTRFS_MAX_LEVEL - 1)
 926                 parent = path->nodes[level + 1];
 927         pslot = path->slots[level + 1];
 928
 929         if (!parent)
 930                 return 1;
 931
 932         left = read_node_slot(root, parent, pslot - 1);
 933
 934         /* first, try to make some room in the middle buffer */
 935         if (left) {
 936                 u32 left_nr;
 937                 left_nr = btrfs_header_nritems(left);
 938                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 939                         wret = 1;
 940                 } else {
 941                         ret = btrfs_cow_block(trans, root, left, parent,
 942                                               pslot - 1, &left);
 943                         if (ret)
 944                                 wret = 1;
 945                         else {
 946                                 wret = push_node_left(trans, root,
 947                                                       left, mid);
 948                         }
 949                 }
 950                 if (wret < 0)
 951                         ret = wret;
 952                 if (wret == 0) {
 953                         struct btrfs_disk_key disk_key;
 954                         orig_slot += left_nr;
 955                         btrfs_node_key(mid, &disk_key, 0);
 956                         btrfs_set_node_key(parent, &disk_key, pslot);
 957                         btrfs_mark_buffer_dirty(parent);
 958                         if (btrfs_header_nritems(left) > orig_slot) {
 959                                 path->nodes[level] = left;
 960                                 path->slots[level + 1] -= 1;
 961                                 path->slots[level] = orig_slot;
 962                                 free_extent_buffer(mid);
 963                         } else {
 964                                 orig_slot -=
 965                                         btrfs_header_nritems(left);
 966                                 path->slots[level] = orig_slot;
 967                                 free_extent_buffer(left);
 968                         }
 969                         return 0;
 970                 }
 971                 free_extent_buffer(left);
 972         }
 973         right= read_node_slot(root, parent, pslot + 1);
 974
 975         /*
 976          * then try to empty the right most buffer into the middle
 977          */
 978         if (right) {
 979                 u32 right_nr;
 980                 right_nr = btrfs_header_nritems(right);
 981                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 982                         wret = 1;
 983                 } else {
 984                         ret = btrfs_cow_block(trans, root, right,
 985                                               parent, pslot + 1,
 986                                               &right);
 987                         if (ret)
 988                                 wret = 1;
 989                         else {
 990                                 wret = balance_node_right(trans, root,
 991                                                           right, mid);
 992                         }
 993                 }
 994                 if (wret < 0)
 995                         ret = wret;
 996                 if (wret == 0) {
 997                         struct btrfs_disk_key disk_key;
 998
 999                         btrfs_node_key(right, &disk_key, 0);
1000                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
1001                         btrfs_mark_buffer_dirty(parent);
1002
1003                         if (btrfs_header_nritems(mid) <= orig_slot) {
1004                                 path->nodes[level] = right;
1005                                 path->slots[level + 1] += 1;
1006                                 path->slots[level] = orig_slot -
1007                                         btrfs_header_nritems(mid);
1008                                 free_extent_buffer(mid);
1009                         } else {
1010                                 free_extent_buffer(right);
1011                         }
1012                         return 0;
1013                 }
1014                 free_extent_buffer(right);
1015         }
1016         return 1;
1017 }
1018
1019 /*
1020  * readahead one full node of leaves
1021  */
1022 static void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
1023                              int level, int slot, u64 objectid)
1024 {
1025         struct extent_buffer *node;
1026         struct btrfs_disk_key disk_key;
1027         u32 nritems;
1028         u64 search;
1029         u64 lowest_read;
1030         u64 highest_read;
1031         u64 nread = 0;
1032         int direction = path->reada;
1033         struct extent_buffer *eb;
1034         u32 nr;
1035         u32 blocksize;
1036         u32 nscan = 0;
1037
1038         if (level != 1)
1039                 return;
1040
1041         if (!path->nodes[level])
1042                 return;
1043
1044         node = path->nodes[level];
1045         search = btrfs_node_blockptr(node, slot);
1046         blocksize = btrfs_level_size(root, level - 1);
1047         eb = btrfs_find_tree_block(root, search, blocksize);
1048         if (eb) {
1049                 free_extent_buffer(eb);
1050                 return;
1051         }
1052
1053         highest_read = search;
1054         lowest_read = search;
1055
1056         nritems = btrfs_header_nritems(node);
1057         nr = slot;
1058         while(1) {
1059                 if (direction < 0) {
1060                         if (nr == 0)
1061                                 break;
1062                         nr--;
1063                 } else if (direction > 0) {
1064                         nr++;
1065                         if (nr >= nritems)
1066                                 break;
1067                 }
1068                 if (path->reada < 0 && objectid) {
1069                         btrfs_node_key(node, &disk_key, nr);
1070                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1071                                 break;
1072                 }
1073                 search = btrfs_node_blockptr(node, nr);
1074                 if ((search >= lowest_read && search <= highest_read) ||
1075                     (search < lowest_read && lowest_read - search <= 32768) ||
1076                     (search > highest_read && search - highest_read <= 32768)) {
1077                         readahead_tree_block(root, search, blocksize);
1078                         nread += blocksize;
1079                 }
1080                 nscan++;
1081                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
1082                         break;
1083                 if(nread > (1024 * 1024) || nscan > 128)
1084                         break;
1085
1086                 if (search < lowest_read)
1087                         lowest_read = search;
1088                 if (search > highest_read)
1089                         highest_read = search;
1090         }
1091 }
1092 /*
1093  * look for key in the tree.  path is filled in with nodes along the way
1094  * if key is found, we return zero and you can find the item in the leaf
1095  * level of the path (level 0)
1096  *
1097  * If the key isn't found, the path points to the slot where it should
1098  * be inserted, and 1 is returned.  If there are other errors during the
1099  * search a negative error number is returned.
1100  *
1101  * if ins_len > 0, nodes and leaves will be split as we walk down the
1102  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1103  * possible)
1104  */
1105 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1106                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1107                       ins_len, int cow)
1108 {
1109         struct extent_buffer *b;
1110         u64 bytenr;
1111         u64 ptr_gen;
1112         int slot;
1113         int ret;
1114         int level;
1115         int should_reada = p->reada;
1116         u8 lowest_level = 0;
1117
1118         lowest_level = p->lowest_level;
1119         WARN_ON(lowest_level && ins_len);
1120         WARN_ON(p->nodes[0] != NULL);
1121         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
1122 again:
1123         b = root->node;
1124         extent_buffer_get(b);
1125         while (b) {
1126                 level = btrfs_header_level(b);
1127                 if (cow) {
1128                         int wret;
1129                         wret = btrfs_cow_block(trans, root, b,
1130                                                p->nodes[level + 1],
1131                                                p->slots[level + 1],
1132                                                &b);
1133                         if (wret) {
1134                                 free_extent_buffer(b);
1135                                 return wret;
1136                         }
1137                 }
1138                 BUG_ON(!cow && ins_len);
1139                 if (level != btrfs_header_level(b))
1140                         WARN_ON(1);
1141                 level = btrfs_header_level(b);
1142                 p->nodes[level] = b;
1143                 ret = check_block(root, p, level);
1144                 if (ret)
1145                         return -1;
1146                 ret = bin_search(b, key, level, &slot);
1147                 if (level != 0) {
1148                         if (ret && slot > 0)
1149                                 slot -= 1;
1150                         p->slots[level] = slot;
1151                         if (ins_len > 0 && btrfs_header_nritems(b) >=
1152                             BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1153                                 int sret = split_node(trans, root, p, level);
1154                                 BUG_ON(sret > 0);
1155                                 if (sret)
1156                                         return sret;
1157                                 b = p->nodes[level];
1158                                 slot = p->slots[level];
1159                         } else if (ins_len < 0) {
1160                                 int sret = balance_level(trans, root, p,
1161                                                          level);
1162                                 if (sret)
1163                                         return sret;
1164                                 b = p->nodes[level];
1165                                 if (!b) {
1166                                         btrfs_release_path(NULL, p);
1167                                         goto again;
1168                                 }
1169                                 slot = p->slots[level];
1170                                 BUG_ON(btrfs_header_nritems(b) == 1);
1171                         }
1172                         /* this is only true while dropping a snapshot */
1173                         if (level == lowest_level)
1174                                 break;
1175                         bytenr = btrfs_node_blockptr(b, slot);
1176                         ptr_gen = btrfs_node_ptr_generation(b, slot);
1177                         if (should_reada)
1178                                 reada_for_search(root, p, level, slot,
1179                                                  key->objectid);
1180                         b = read_tree_block(root, bytenr,
1181                                             btrfs_level_size(root, level - 1));
1182                         if (ptr_gen != btrfs_header_generation(b)) {
1183                                 printk("block %llu bad gen wanted %llu "
1184                                        "found %llu\n",
1185                                 (unsigned long long)b->start,
1186                                 (unsigned long long)ptr_gen,
1187                                 (unsigned long long)btrfs_header_generation(b));
1188                         }
1189                 } else {
1190                         p->slots[level] = slot;
1191                         if (ins_len > 0 && btrfs_leaf_free_space(root, b) <
1192                             sizeof(struct btrfs_item) + ins_len) {
1193                                 int sret = split_leaf(trans, root, key,
1194                                                       p, ins_len, ret == 0);
1195                                 BUG_ON(sret > 0);
1196                                 if (sret)
1197                                         return sret;
1198                         }
1199                         return ret;
1200                 }
1201         }
1202         return 1;
1203 }
1204
1205 /*
1206  * adjust the pointers going up the tree, starting at level
1207  * making sure the right key of each node is points to 'key'.
1208  * This is used after shifting pointers to the left, so it stops
1209  * fixing up pointers when a given leaf/node is not in slot 0 of the
1210  * higher levels
1211  *
1212  * If this fails to write a tree block, it returns -1, but continues
1213  * fixing up the blocks in ram so the tree is consistent.
1214  */
1215 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1216                           struct btrfs_root *root, struct btrfs_path *path,
1217                           struct btrfs_disk_key *key, int level)
1218 {
1219         int i;
1220         int ret = 0;
1221         struct extent_buffer *t;
1222
1223         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1224                 int tslot = path->slots[i];
1225                 if (!path->nodes[i])
1226                         break;
1227                 t = path->nodes[i];
1228                 btrfs_set_node_key(t, key, tslot);
1229                 btrfs_mark_buffer_dirty(path->nodes[i]);
1230                 if (tslot != 0)
1231                         break;
1232         }
1233         return ret;
1234 }
1235
1236 /*
1237  * try to push data from one node into the next node left in the
1238  * tree.
1239  *
1240  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1241  * error, and > 0 if there was no room in the left hand block.
1242  */
1243 static int push_node_left(struct btrfs_trans_handle *trans,
1244                           struct btrfs_root *root, struct extent_buffer *dst,
1245                           struct extent_buffer *src)
1246 {
1247         int push_items = 0;
1248         int src_nritems;
1249         int dst_nritems;
1250         int ret = 0;
1251
1252         src_nritems = btrfs_header_nritems(src);
1253         dst_nritems = btrfs_header_nritems(dst);
1254         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1255         WARN_ON(btrfs_header_generation(src) != trans->transid);
1256         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1257
1258         if (push_items <= 0) {
1259                 return 1;
1260         }
1261
1262         if (src_nritems < push_items)
1263                 push_items = src_nritems;
1264
1265         copy_extent_buffer(dst, src,
1266                            btrfs_node_key_ptr_offset(dst_nritems),
1267                            btrfs_node_key_ptr_offset(0),
1268                            push_items * sizeof(struct btrfs_key_ptr));
1269
1270         if (push_items < src_nritems) {
1271                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1272                                       btrfs_node_key_ptr_offset(push_items),
1273                                       (src_nritems - push_items) *
1274                                       sizeof(struct btrfs_key_ptr));
1275         }
1276         btrfs_set_header_nritems(src, src_nritems - push_items);
1277         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1278         btrfs_mark_buffer_dirty(src);
1279         btrfs_mark_buffer_dirty(dst);
1280         return ret;
1281 }
1282
1283 /*
1284  * try to push data from one node into the next node right in the
1285  * tree.
1286  *
1287  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1288  * error, and > 0 if there was no room in the right hand block.
1289  *
1290  * this will  only push up to 1/2 the contents of the left node over
1291  */
1292 static int balance_node_right(struct btrfs_trans_handle *trans,
1293                               struct btrfs_root *root,
1294                               struct extent_buffer *dst,
1295                               struct extent_buffer *src)
1296 {
1297         int push_items = 0;
1298         int max_push;
1299         int src_nritems;
1300         int dst_nritems;
1301         int ret = 0;
1302
1303         WARN_ON(btrfs_header_generation(src) != trans->transid);
1304         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1305
1306         src_nritems = btrfs_header_nritems(src);
1307         dst_nritems = btrfs_header_nritems(dst);
1308         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1309         if (push_items <= 0)
1310                 return 1;
1311
1312         max_push = src_nritems / 2 + 1;
1313         /* don't try to empty the node */
1314         if (max_push >= src_nritems)
1315                 return 1;
1316
1317         if (max_push < push_items)
1318                 push_items = max_push;
1319
1320         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1321                                       btrfs_node_key_ptr_offset(0),
1322                                       (dst_nritems) *
1323                                       sizeof(struct btrfs_key_ptr));
1324
1325         copy_extent_buffer(dst, src,
1326                            btrfs_node_key_ptr_offset(0),
1327                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1328                            push_items * sizeof(struct btrfs_key_ptr));
1329
1330         btrfs_set_header_nritems(src, src_nritems - push_items);
1331         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1332
1333         btrfs_mark_buffer_dirty(src);
1334         btrfs_mark_buffer_dirty(dst);
1335         return ret;
1336 }
1337
1338 /*
1339  * helper function to insert a new root level in the tree.
1340  * A new node is allocated, and a single item is inserted to
1341  * point to the existing root
1342  *
1343  * returns zero on success or < 0 on failure.
1344  */
1345 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1346                            struct btrfs_root *root,
1347                            struct btrfs_path *path, int level)
1348 {
1349         u64 root_gen;
1350         u64 lower_gen;
1351         struct extent_buffer *lower;
1352         struct extent_buffer *c;
1353         struct btrfs_disk_key lower_key;
1354
1355         BUG_ON(path->nodes[level]);
1356         BUG_ON(path->nodes[level-1] != root->node);
1357
1358         if (root->ref_cows)
1359                 root_gen = trans->transid;
1360         else
1361                 root_gen = 0;
1362
1363         lower = path->nodes[level-1];
1364         if (level == 1)
1365                 btrfs_item_key(lower, &lower_key, 0);
1366         else
1367                 btrfs_node_key(lower, &lower_key, 0);
1368
1369         c = __btrfs_alloc_free_block(trans, root, root->nodesize,
1370                                    root->root_key.objectid,
1371                                    root_gen, lower_key.objectid, level,
1372                                    root->node->start, 0);
1373         if (IS_ERR(c))
1374                 return PTR_ERR(c);
1375         memset_extent_buffer(c, 0, 0, root->nodesize);
1376         btrfs_set_header_nritems(c, 1);
1377         btrfs_set_header_level(c, level);
1378         btrfs_set_header_bytenr(c, c->start);
1379         btrfs_set_header_generation(c, trans->transid);
1380         btrfs_set_header_owner(c, root->root_key.objectid);
1381
1382         write_extent_buffer(c, root->fs_info->fsid,
1383                             (unsigned long)btrfs_header_fsid(c),
1384                             BTRFS_FSID_SIZE);
1385
1386         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1387                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
1388                             BTRFS_UUID_SIZE);
1389
1390         btrfs_set_node_key(c, &lower_key, 0);
1391         btrfs_set_node_blockptr(c, 0, lower->start);
1392         lower_gen = btrfs_header_generation(lower);
1393         WARN_ON(lower_gen == 0);
1394
1395         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1396
1397         btrfs_mark_buffer_dirty(c);
1398
1399         /* the super has an extra ref to root->node */
1400         free_extent_buffer(root->node);
1401         root->node = c;
1402         add_root_to_dirty_list(root);
1403         extent_buffer_get(c);
1404         path->nodes[level] = c;
1405         path->slots[level] = 0;
1406
1407         if (root->ref_cows && lower_gen != trans->transid) {
1408                 struct btrfs_path *back_path = btrfs_alloc_path();
1409                 int ret;
1410                 ret = btrfs_insert_extent_backref(trans,
1411                                                   root->fs_info->extent_root,
1412                                                   path, lower->start,
1413                                                   root->root_key.objectid,
1414                                                   trans->transid, 0, 0);
1415                 BUG_ON(ret);
1416                 btrfs_free_path(back_path);
1417         }
1418         return 0;
1419 }
1420
1421 /*
1422  * worker function to insert a single pointer in a node.
1423  * the node should have enough room for the pointer already
1424  *
1425  * slot and level indicate where you want the key to go, and
1426  * blocknr is the block the key points to.
1427  *
1428  * returns zero on success and < 0 on any error
1429  */
1430 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1431                       *root, struct btrfs_path *path, struct btrfs_disk_key
1432                       *key, u64 bytenr, int slot, int level)
1433 {
1434         struct extent_buffer *lower;
1435         int nritems;
1436
1437         BUG_ON(!path->nodes[level]);
1438         lower = path->nodes[level];
1439         nritems = btrfs_header_nritems(lower);
1440         if (slot > nritems)
1441                 BUG();
1442         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1443                 BUG();
1444         if (slot != nritems) {
1445                 memmove_extent_buffer(lower,
1446                               btrfs_node_key_ptr_offset(slot + 1),
1447                               btrfs_node_key_ptr_offset(slot),
1448                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1449         }
1450         btrfs_set_node_key(lower, key, slot);
1451         btrfs_set_node_blockptr(lower, slot, bytenr);
1452         WARN_ON(trans->transid == 0);
1453         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1454         btrfs_set_header_nritems(lower, nritems + 1);
1455         btrfs_mark_buffer_dirty(lower);
1456         return 0;
1457 }
1458
1459 /*
1460  * split the node at the specified level in path in two.
1461  * The path is corrected to point to the appropriate node after the split
1462  *
1463  * Before splitting this tries to make some room in the node by pushing
1464  * left and right, if either one works, it returns right away.
1465  *
1466  * returns 0 on success and < 0 on failure
1467  */
1468 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1469                       *root, struct btrfs_path *path, int level)
1470 {
1471         u64 root_gen;
1472         struct extent_buffer *c;
1473         struct extent_buffer *split;
1474         struct btrfs_disk_key disk_key;
1475         int mid;
1476         int ret;
1477         int wret;
1478         u32 c_nritems;
1479
1480         c = path->nodes[level];
1481         WARN_ON(btrfs_header_generation(c) != trans->transid);
1482         if (c == root->node) {
1483                 /* trying to split the root, lets make a new one */
1484                 ret = insert_new_root(trans, root, path, level + 1);
1485                 if (ret)
1486                         return ret;
1487         } else {
1488                 ret = push_nodes_for_insert(trans, root, path, level);
1489                 c = path->nodes[level];
1490                 if (!ret && btrfs_header_nritems(c) <
1491                     BTRFS_NODEPTRS_PER_BLOCK(root) - 1)
1492                         return 0;
1493                 if (ret < 0)
1494                         return ret;
1495         }
1496
1497         c_nritems = btrfs_header_nritems(c);
1498         if (root->ref_cows)
1499                 root_gen = trans->transid;
1500         else
1501                 root_gen = 0;
1502
1503         btrfs_node_key(c, &disk_key, 0);
1504         split = __btrfs_alloc_free_block(trans, root, root->nodesize,
1505                                          root->root_key.objectid,
1506                                          root_gen,
1507                                          btrfs_disk_key_objectid(&disk_key),
1508                                          level, c->start, 0);
1509         if (IS_ERR(split))
1510                 return PTR_ERR(split);
1511
1512         btrfs_set_header_flags(split, btrfs_header_flags(c));
1513         btrfs_set_header_level(split, btrfs_header_level(c));
1514         btrfs_set_header_bytenr(split, split->start);
1515         btrfs_set_header_generation(split, trans->transid);
1516         btrfs_set_header_owner(split, root->root_key.objectid);
1517         btrfs_set_header_flags(split, 0);
1518         write_extent_buffer(split, root->fs_info->fsid,
1519                             (unsigned long)btrfs_header_fsid(split),
1520                             BTRFS_FSID_SIZE);
1521         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1522                             (unsigned long)btrfs_header_chunk_tree_uuid(split),
1523                             BTRFS_UUID_SIZE);
1524
1525         mid = (c_nritems + 1) / 2;
1526
1527         copy_extent_buffer(split, c,
1528                            btrfs_node_key_ptr_offset(0),
1529                            btrfs_node_key_ptr_offset(mid),
1530                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1531         btrfs_set_header_nritems(split, c_nritems - mid);
1532         btrfs_set_header_nritems(c, mid);
1533         ret = 0;
1534
1535         btrfs_mark_buffer_dirty(c);
1536         btrfs_mark_buffer_dirty(split);
1537
1538         btrfs_node_key(split, &disk_key, 0);
1539         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1540                           path->slots[level + 1] + 1,
1541                           level + 1);
1542         if (wret)
1543                 ret = wret;
1544
1545         if (path->slots[level] >= mid) {
1546                 path->slots[level] -= mid;
1547                 free_extent_buffer(c);
1548                 path->nodes[level] = split;
1549                 path->slots[level + 1] += 1;
1550         } else {
1551                 free_extent_buffer(split);
1552         }
1553         return ret;
1554 }
1555
1556 /*
1557  * how many bytes are required to store the items in a leaf.  start
1558  * and nr indicate which items in the leaf to check.  This totals up the
1559  * space used both by the item structs and the item data
1560  */
1561 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1562 {
1563         int data_len;
1564         int nritems = btrfs_header_nritems(l);
1565         int end = min(nritems, start + nr) - 1;
1566
1567         if (!nr)
1568                 return 0;
1569         data_len = btrfs_item_end_nr(l, start);
1570         data_len = data_len - btrfs_item_offset_nr(l, end);
1571         data_len += sizeof(struct btrfs_item) * nr;
1572         WARN_ON(data_len < 0);
1573         return data_len;
1574 }
1575
1576 /*
1577  * The space between the end of the leaf items and
1578  * the start of the leaf data.  IOW, how much room
1579  * the leaf has left for both items and data
1580  */
1581 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1582 {
1583         int nritems = btrfs_header_nritems(leaf);
1584         int ret;
1585         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1586         if (ret < 0) {
1587                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1588                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1589                        leaf_space_used(leaf, 0, nritems), nritems);
1590         }
1591         return ret;
1592 }
1593
1594 /*
1595  * push some data in the path leaf to the right, trying to free up at
1596  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1597  *
1598  * returns 1 if the push failed because the other node didn't have enough
1599  * room, 0 if everything worked out and < 0 if there were major errors.
1600  */
1601 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1602                            *root, struct btrfs_path *path, int data_size,
1603                            int empty)
1604 {
1605         struct extent_buffer *left = path->nodes[0];
1606         struct extent_buffer *right;
1607         struct extent_buffer *upper;
1608         struct btrfs_disk_key disk_key;
1609         int slot;
1610         u32 i;
1611         int free_space;
1612         int push_space = 0;
1613         int push_items = 0;
1614         struct btrfs_item *item;
1615         u32 left_nritems;
1616         u32 nr;
1617         u32 right_nritems;
1618         u32 data_end;
1619         u32 this_item_size;
1620         int ret;
1621
1622         slot = path->slots[1];
1623         if (!path->nodes[1]) {
1624                 return 1;
1625         }
1626         upper = path->nodes[1];
1627         if (slot >= btrfs_header_nritems(upper) - 1)
1628                 return 1;
1629
1630         right = read_tree_block(root, btrfs_node_blockptr(upper, slot + 1),
1631                                 root->leafsize);
1632         free_space = btrfs_leaf_free_space(root, right);
1633         if (free_space < data_size + sizeof(struct btrfs_item)) {
1634                 free_extent_buffer(right);
1635                 return 1;
1636         }
1637
1638         /* cow and double check */
1639         ret = btrfs_cow_block(trans, root, right, upper,
1640                               slot + 1, &right);
1641         if (ret) {
1642                 free_extent_buffer(right);
1643                 return 1;
1644         }
1645         free_space = btrfs_leaf_free_space(root, right);
1646         if (free_space < data_size + sizeof(struct btrfs_item)) {
1647                 free_extent_buffer(right);
1648                 return 1;
1649         }
1650
1651         left_nritems = btrfs_header_nritems(left);
1652         if (left_nritems == 0) {
1653                 free_extent_buffer(right);
1654                 return 1;
1655         }
1656
1657         if (empty)
1658                 nr = 0;
1659         else
1660                 nr = 1;
1661
1662         i = left_nritems - 1;
1663         while (i >= nr) {
1664                 item = btrfs_item_nr(left, i);
1665
1666                 if (path->slots[0] == i)
1667                         push_space += data_size + sizeof(*item);
1668
1669                 if (!left->map_token) {
1670                         map_extent_buffer(left, (unsigned long)item,
1671                                         sizeof(struct btrfs_item),
1672                                         &left->map_token, &left->kaddr,
1673                                         &left->map_start, &left->map_len,
1674                                         KM_USER1);
1675                 }
1676
1677                 this_item_size = btrfs_item_size(left, item);
1678                 if (this_item_size + sizeof(*item) + push_space > free_space)
1679                         break;
1680                 push_items++;
1681                 push_space += this_item_size + sizeof(*item);
1682                 if (i == 0)
1683                         break;
1684                 i--;
1685         }
1686         if (left->map_token) {
1687                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1688                 left->map_token = NULL;
1689         }
1690
1691         if (push_items == 0) {
1692                 free_extent_buffer(right);
1693                 return 1;
1694         }
1695
1696         if (!empty && push_items == left_nritems)
1697                 WARN_ON(1);
1698
1699         /* push left to right */
1700         right_nritems = btrfs_header_nritems(right);
1701
1702         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1703         push_space -= leaf_data_end(root, left);
1704
1705         /* make room in the right data area */
1706         data_end = leaf_data_end(root, right);
1707         memmove_extent_buffer(right,
1708                               btrfs_leaf_data(right) + data_end - push_space,
1709                               btrfs_leaf_data(right) + data_end,
1710                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1711
1712         /* copy from the left data area */
1713         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1714                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1715                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1716                      push_space);
1717
1718         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1719                               btrfs_item_nr_offset(0),
1720                               right_nritems * sizeof(struct btrfs_item));
1721
1722         /* copy the items from left to right */
1723         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1724                    btrfs_item_nr_offset(left_nritems - push_items),
1725                    push_items * sizeof(struct btrfs_item));
1726
1727         /* update the item pointers */
1728         right_nritems += push_items;
1729         btrfs_set_header_nritems(right, right_nritems);
1730         push_space = BTRFS_LEAF_DATA_SIZE(root);
1731         for (i = 0; i < right_nritems; i++) {
1732                 item = btrfs_item_nr(right, i);
1733                 if (!right->map_token) {
1734                         map_extent_buffer(right, (unsigned long)item,
1735                                         sizeof(struct btrfs_item),
1736                                         &right->map_token, &right->kaddr,
1737                                         &right->map_start, &right->map_len,
1738                                         KM_USER1);
1739                 }
1740                 push_space -= btrfs_item_size(right, item);
1741                 btrfs_set_item_offset(right, item, push_space);
1742         }
1743
1744         if (right->map_token) {
1745                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1746                 right->map_token = NULL;
1747         }
1748         left_nritems -= push_items;
1749         btrfs_set_header_nritems(left, left_nritems);
1750
1751         if (left_nritems)
1752                 btrfs_mark_buffer_dirty(left);
1753         btrfs_mark_buffer_dirty(right);
1754
1755         btrfs_item_key(right, &disk_key, 0);
1756         btrfs_set_node_key(upper, &disk_key, slot + 1);
1757         btrfs_mark_buffer_dirty(upper);
1758
1759         /* then fixup the leaf pointer in the path */
1760         if (path->slots[0] >= left_nritems) {
1761                 path->slots[0] -= left_nritems;
1762                 free_extent_buffer(path->nodes[0]);
1763                 path->nodes[0] = right;
1764                 path->slots[1] += 1;
1765         } else {
1766                 free_extent_buffer(right);
1767         }
1768         return 0;
1769 }
1770 /*
1771  * push some data in the path leaf to the left, trying to free up at
1772  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1773  */
1774 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1775                           *root, struct btrfs_path *path, int data_size,
1776                           int empty)
1777 {
1778         struct btrfs_disk_key disk_key;
1779         struct extent_buffer *right = path->nodes[0];
1780         struct extent_buffer *left;
1781         int slot;
1782         int i;
1783         int free_space;
1784         int push_space = 0;
1785         int push_items = 0;
1786         struct btrfs_item *item;
1787         u32 old_left_nritems;
1788         u32 right_nritems;
1789         u32 nr;
1790         int ret = 0;
1791         int wret;
1792         u32 this_item_size;
1793         u32 old_left_item_size;
1794
1795         slot = path->slots[1];
1796         if (slot == 0)
1797                 return 1;
1798         if (!path->nodes[1])
1799                 return 1;
1800
1801         right_nritems = btrfs_header_nritems(right);
1802         if (right_nritems == 0) {
1803                 return 1;
1804         }
1805
1806         left = read_tree_block(root, btrfs_node_blockptr(path->nodes[1],
1807                                slot - 1), root->leafsize);
1808         free_space = btrfs_leaf_free_space(root, left);
1809         if (free_space < data_size + sizeof(struct btrfs_item)) {
1810                 free_extent_buffer(left);
1811                 return 1;
1812         }
1813
1814         /* cow and double check */
1815         ret = btrfs_cow_block(trans, root, left,
1816                               path->nodes[1], slot - 1, &left);
1817         if (ret) {
1818                 /* we hit -ENOSPC, but it isn't fatal here */
1819                 free_extent_buffer(left);
1820                 return 1;
1821         }
1822
1823         free_space = btrfs_leaf_free_space(root, left);
1824         if (free_space < data_size + sizeof(struct btrfs_item)) {
1825                 free_extent_buffer(left);
1826                 return 1;
1827         }
1828
1829         if (empty)
1830                 nr = right_nritems;
1831         else
1832                 nr = right_nritems - 1;
1833
1834         for (i = 0; i < nr; i++) {
1835                 item = btrfs_item_nr(right, i);
1836                 if (!right->map_token) {
1837                         map_extent_buffer(right, (unsigned long)item,
1838                                         sizeof(struct btrfs_item),
1839                                         &right->map_token, &right->kaddr,
1840                                         &right->map_start, &right->map_len,
1841                                         KM_USER1);
1842                 }
1843
1844                 if (path->slots[0] == i)
1845                         push_space += data_size + sizeof(*item);
1846
1847                 this_item_size = btrfs_item_size(right, item);
1848                 if (this_item_size + sizeof(*item) + push_space > free_space)
1849                         break;
1850
1851                 push_items++;
1852                 push_space += this_item_size + sizeof(*item);
1853         }
1854
1855         if (right->map_token) {
1856                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1857                 right->map_token = NULL;
1858         }
1859
1860         if (push_items == 0) {
1861                 free_extent_buffer(left);
1862                 return 1;
1863         }
1864         if (!empty && push_items == btrfs_header_nritems(right))
1865                 WARN_ON(1);
1866
1867         /* push data from right to left */
1868         copy_extent_buffer(left, right,
1869                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1870                            btrfs_item_nr_offset(0),
1871                            push_items * sizeof(struct btrfs_item));
1872
1873         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1874                      btrfs_item_offset_nr(right, push_items -1);
1875
1876         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1877                      leaf_data_end(root, left) - push_space,
1878                      btrfs_leaf_data(right) +
1879                      btrfs_item_offset_nr(right, push_items - 1),
1880                      push_space);
1881         old_left_nritems = btrfs_header_nritems(left);
1882         BUG_ON(old_left_nritems < 0);
1883
1884         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1885         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1886                 u32 ioff;
1887
1888                 item = btrfs_item_nr(left, i);
1889                 if (!left->map_token) {
1890                         map_extent_buffer(left, (unsigned long)item,
1891                                         sizeof(struct btrfs_item),
1892                                         &left->map_token, &left->kaddr,
1893                                         &left->map_start, &left->map_len,
1894                                         KM_USER1);
1895                 }
1896
1897                 ioff = btrfs_item_offset(left, item);
1898                 btrfs_set_item_offset(left, item,
1899                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1900         }
1901         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1902         if (left->map_token) {
1903                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1904                 left->map_token = NULL;
1905         }
1906
1907         /* fixup right node */
1908         if (push_items > right_nritems) {
1909                 printk("push items %d nr %u\n", push_items, right_nritems);
1910                 WARN_ON(1);
1911         }
1912
1913         if (push_items < right_nritems) {
1914                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1915                                                   leaf_data_end(root, right);
1916                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
1917                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1918                                       btrfs_leaf_data(right) +
1919                                       leaf_data_end(root, right), push_space);
1920
1921                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1922                               btrfs_item_nr_offset(push_items),
1923                              (btrfs_header_nritems(right) - push_items) *
1924                              sizeof(struct btrfs_item));
1925         }
1926         right_nritems -= push_items;
1927         btrfs_set_header_nritems(right, right_nritems);
1928         push_space = BTRFS_LEAF_DATA_SIZE(root);
1929         for (i = 0; i < right_nritems; i++) {
1930                 item = btrfs_item_nr(right, i);
1931
1932                 if (!right->map_token) {
1933                         map_extent_buffer(right, (unsigned long)item,
1934                                         sizeof(struct btrfs_item),
1935                                         &right->map_token, &right->kaddr,
1936                                         &right->map_start, &right->map_len,
1937                                         KM_USER1);
1938                 }
1939
1940                 push_space = push_space - btrfs_item_size(right, item);
1941                 btrfs_set_item_offset(right, item, push_space);
1942         }
1943         if (right->map_token) {
1944                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1945                 right->map_token = NULL;
1946         }
1947
1948         btrfs_mark_buffer_dirty(left);
1949         if (right_nritems)
1950                 btrfs_mark_buffer_dirty(right);
1951
1952         btrfs_item_key(right, &disk_key, 0);
1953         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1954         if (wret)
1955                 ret = wret;
1956
1957         /* then fixup the leaf pointer in the path */
1958         if (path->slots[0] < push_items) {
1959                 path->slots[0] += old_left_nritems;
1960                 free_extent_buffer(path->nodes[0]);
1961                 path->nodes[0] = left;
1962                 path->slots[1] -= 1;
1963         } else {
1964                 free_extent_buffer(left);
1965                 path->slots[0] -= push_items;
1966         }
1967         BUG_ON(path->slots[0] < 0);
1968         return ret;
1969 }
1970
1971 /*
1972  * split the path's leaf in two, making sure there is at least data_size
1973  * available for the resulting leaf level of the path.
1974  *
1975  * returns 0 if all went well and < 0 on failure.
1976  */
1977 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
1978                       *root, struct btrfs_key *ins_key,
1979                       struct btrfs_path *path, int data_size, int extend)
1980 {
1981         u64 root_gen;
1982         struct extent_buffer *l;
1983         u32 nritems;
1984         int mid;
1985         int slot;
1986         struct extent_buffer *right;
1987         int space_needed = data_size + sizeof(struct btrfs_item);
1988         int data_copy_size;
1989         int rt_data_off;
1990         int i;
1991         int ret = 0;
1992         int wret;
1993         int double_split;
1994         int num_doubles = 0;
1995         struct btrfs_disk_key disk_key;
1996
1997         if (extend)
1998                 space_needed = data_size;
1999
2000         if (root->ref_cows)
2001                 root_gen = trans->transid;
2002         else
2003                 root_gen = 0;
2004
2005         /* first try to make some room by pushing left and right */
2006         if (ins_key->type != BTRFS_DIR_ITEM_KEY) {
2007                 wret = push_leaf_right(trans, root, path, data_size, 0);
2008                 if (wret < 0) {
2009                         return wret;
2010                 }
2011                 if (wret) {
2012                         wret = push_leaf_left(trans, root, path, data_size, 0);
2013                         if (wret < 0)
2014                                 return wret;
2015                 }
2016                 l = path->nodes[0];
2017
2018                 /* did the pushes work? */
2019                 if (btrfs_leaf_free_space(root, l) >= space_needed)
2020                         return 0;
2021         }
2022
2023         if (!path->nodes[1]) {
2024                 ret = insert_new_root(trans, root, path, 1);
2025                 if (ret)
2026                         return ret;
2027         }
2028 again:
2029         double_split = 0;
2030         l = path->nodes[0];
2031         slot = path->slots[0];
2032         nritems = btrfs_header_nritems(l);
2033         mid = (nritems + 1)/ 2;
2034
2035         btrfs_item_key(l, &disk_key, 0);
2036
2037         right = __btrfs_alloc_free_block(trans, root, root->leafsize,
2038                                          root->root_key.objectid,
2039                                          root_gen, disk_key.objectid, 0,
2040                                          l->start, 0);
2041         if (IS_ERR(right)) {
2042                 BUG_ON(1);
2043                 return PTR_ERR(right);
2044         }
2045
2046         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2047         btrfs_set_header_bytenr(right, right->start);
2048         btrfs_set_header_generation(right, trans->transid);
2049         btrfs_set_header_owner(right, root->root_key.objectid);
2050         btrfs_set_header_level(right, 0);
2051         write_extent_buffer(right, root->fs_info->fsid,
2052                             (unsigned long)btrfs_header_fsid(right),
2053                             BTRFS_FSID_SIZE);
2054
2055         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2056                             (unsigned long)btrfs_header_chunk_tree_uuid(right),
2057                             BTRFS_UUID_SIZE);
2058         if (mid <= slot) {
2059                 if (nritems == 1 ||
2060                     leaf_space_used(l, mid, nritems - mid) + space_needed >
2061                         BTRFS_LEAF_DATA_SIZE(root)) {
2062                         if (slot >= nritems) {
2063                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2064                                 btrfs_set_header_nritems(right, 0);
2065                                 wret = insert_ptr(trans, root, path,
2066                                                   &disk_key, right->start,
2067                                                   path->slots[1] + 1, 1);
2068                                 if (wret)
2069                                         ret = wret;
2070                                 free_extent_buffer(path->nodes[0]);
2071                                 path->nodes[0] = right;
2072                                 path->slots[0] = 0;
2073                                 path->slots[1] += 1;
2074                                 btrfs_mark_buffer_dirty(right);
2075                                 return ret;
2076                         }
2077                         mid = slot;
2078                         if (mid != nritems &&
2079                             leaf_space_used(l, mid, nritems - mid) +
2080                             space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2081                                 double_split = 1;
2082                         }
2083                 }
2084         } else {
2085                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
2086                         BTRFS_LEAF_DATA_SIZE(root)) {
2087                         if (!extend && slot == 0) {
2088                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2089                                 btrfs_set_header_nritems(right, 0);
2090                                 wret = insert_ptr(trans, root, path,
2091                                                   &disk_key,
2092                                                   right->start,
2093                                                   path->slots[1], 1);
2094                                 if (wret)
2095                                         ret = wret;
2096                                 free_extent_buffer(path->nodes[0]);
2097                                 path->nodes[0] = right;
2098                                 path->slots[0] = 0;
2099                                 if (path->slots[1] == 0) {
2100                                         wret = fixup_low_keys(trans, root,
2101                                                    path, &disk_key, 1);
2102                                         if (wret)
2103                                                 ret = wret;
2104                                 }
2105                                 btrfs_mark_buffer_dirty(right);
2106                                 return ret;
2107                         } else if (extend && slot == 0) {
2108                                 mid = 1;
2109                         } else {
2110                                 mid = slot;
2111                                 if (mid != nritems &&
2112                                     leaf_space_used(l, mid, nritems - mid) +
2113                                     space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2114                                         double_split = 1;
2115                                 }
2116                         }
2117                 }
2118         }
2119         nritems = nritems - mid;
2120         btrfs_set_header_nritems(right, nritems);
2121         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2122
2123         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2124                            btrfs_item_nr_offset(mid),
2125                            nritems * sizeof(struct btrfs_item));
2126
2127         copy_extent_buffer(right, l,
2128                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2129                      data_copy_size, btrfs_leaf_data(l) +
2130                      leaf_data_end(root, l), data_copy_size);
2131
2132         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2133                       btrfs_item_end_nr(l, mid);
2134
2135         for (i = 0; i < nritems; i++) {
2136                 struct btrfs_item *item = btrfs_item_nr(right, i);
2137                 u32 ioff;
2138
2139                 if (!right->map_token) {
2140                         map_extent_buffer(right, (unsigned long)item,
2141                                         sizeof(struct btrfs_item),
2142                                         &right->map_token, &right->kaddr,
2143                                         &right->map_start, &right->map_len,
2144                                         KM_USER1);
2145                 }
2146
2147                 ioff = btrfs_item_offset(right, item);
2148                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2149         }
2150
2151         if (right->map_token) {
2152                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2153                 right->map_token = NULL;
2154         }
2155
2156         btrfs_set_header_nritems(l, mid);
2157         ret = 0;
2158         btrfs_item_key(right, &disk_key, 0);
2159         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2160                           path->slots[1] + 1, 1);
2161         if (wret)
2162                 ret = wret;
2163
2164         btrfs_mark_buffer_dirty(right);
2165         btrfs_mark_buffer_dirty(l);
2166         BUG_ON(path->slots[0] != slot);
2167
2168         if (mid <= slot) {
2169                 free_extent_buffer(path->nodes[0]);
2170                 path->nodes[0] = right;
2171                 path->slots[0] -= mid;
2172                 path->slots[1] += 1;
2173         } else
2174                 free_extent_buffer(right);
2175
2176         BUG_ON(path->slots[0] < 0);
2177
2178         if (double_split) {
2179                 BUG_ON(num_doubles != 0);
2180                 num_doubles++;
2181                 goto again;
2182         }
2183         return ret;
2184 }
2185
2186 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2187                         struct btrfs_root *root,
2188                         struct btrfs_path *path,
2189                         u32 new_size, int from_end)
2190 {
2191         int ret = 0;
2192         int slot;
2193         int slot_orig;
2194         struct extent_buffer *leaf;
2195         struct btrfs_item *item;
2196         u32 nritems;
2197         unsigned int data_end;
2198         unsigned int old_data_start;
2199         unsigned int old_size;
2200         unsigned int size_diff;
2201         int i;
2202
2203         slot_orig = path->slots[0];
2204         leaf = path->nodes[0];
2205         slot = path->slots[0];
2206
2207         old_size = btrfs_item_size_nr(leaf, slot);
2208         if (old_size == new_size)
2209                 return 0;
2210
2211         nritems = btrfs_header_nritems(leaf);
2212         data_end = leaf_data_end(root, leaf);
2213
2214         old_data_start = btrfs_item_offset_nr(leaf, slot);
2215
2216         size_diff = old_size - new_size;
2217
2218         BUG_ON(slot < 0);
2219         BUG_ON(slot >= nritems);
2220
2221         /*
2222          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2223          */
2224         /* first correct the data pointers */
2225         for (i = slot; i < nritems; i++) {
2226                 u32 ioff;
2227                 item = btrfs_item_nr(leaf, i);
2228
2229                 if (!leaf->map_token) {
2230                         map_extent_buffer(leaf, (unsigned long)item,
2231                                         sizeof(struct btrfs_item),
2232                                         &leaf->map_token, &leaf->kaddr,
2233                                         &leaf->map_start, &leaf->map_len,
2234                                         KM_USER1);
2235                 }
2236
2237                 ioff = btrfs_item_offset(leaf, item);
2238                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2239         }
2240
2241         if (leaf->map_token) {
2242                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2243                 leaf->map_token = NULL;
2244         }
2245
2246         /* shift the data */
2247         if (from_end) {
2248                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2249                               data_end + size_diff, btrfs_leaf_data(leaf) +
2250                               data_end, old_data_start + new_size - data_end);
2251         } else {
2252                 struct btrfs_disk_key disk_key;
2253                 u64 offset;
2254
2255                 btrfs_item_key(leaf, &disk_key, slot);
2256
2257                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2258                         unsigned long ptr;
2259                         struct btrfs_file_extent_item *fi;
2260
2261                         fi = btrfs_item_ptr(leaf, slot,
2262                                             struct btrfs_file_extent_item);
2263                         fi = (struct btrfs_file_extent_item *)(
2264                              (unsigned long)fi - size_diff);
2265
2266                         if (btrfs_file_extent_type(leaf, fi) ==
2267                             BTRFS_FILE_EXTENT_INLINE) {
2268                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2269                                 memmove_extent_buffer(leaf, ptr,
2270                                         (unsigned long)fi,
2271                                         offsetof(struct btrfs_file_extent_item,
2272                                                  disk_bytenr));
2273                         }
2274                 }
2275
2276                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2277                               data_end + size_diff, btrfs_leaf_data(leaf) +
2278                               data_end, old_data_start - data_end);
2279
2280                 offset = btrfs_disk_key_offset(&disk_key);
2281                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2282                 btrfs_set_item_key(leaf, &disk_key, slot);
2283                 if (slot == 0)
2284                         fixup_low_keys(trans, root, path, &disk_key, 1);
2285         }
2286
2287         item = btrfs_item_nr(leaf, slot);
2288         btrfs_set_item_size(leaf, item, new_size);
2289         btrfs_mark_buffer_dirty(leaf);
2290
2291         ret = 0;
2292         if (btrfs_leaf_free_space(root, leaf) < 0) {
2293                 btrfs_print_leaf(root, leaf);
2294                 BUG();
2295         }
2296         return ret;
2297 }
2298
2299 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2300                       struct btrfs_root *root, struct btrfs_path *path,
2301                       u32 data_size)
2302 {
2303         int ret = 0;
2304         int slot;
2305         int slot_orig;
2306         struct extent_buffer *leaf;
2307         struct btrfs_item *item;
2308         u32 nritems;
2309         unsigned int data_end;
2310         unsigned int old_data;
2311         unsigned int old_size;
2312         int i;
2313
2314         slot_orig = path->slots[0];
2315         leaf = path->nodes[0];
2316
2317         nritems = btrfs_header_nritems(leaf);
2318         data_end = leaf_data_end(root, leaf);
2319
2320         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2321                 btrfs_print_leaf(root, leaf);
2322                 BUG();
2323         }
2324         slot = path->slots[0];
2325         old_data = btrfs_item_end_nr(leaf, slot);
2326
2327         BUG_ON(slot < 0);
2328         if (slot >= nritems) {
2329                 btrfs_print_leaf(root, leaf);
2330                 printk("slot %d too large, nritems %d\n", slot, nritems);
2331                 BUG_ON(1);
2332         }
2333
2334         /*
2335          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2336          */
2337         /* first correct the data pointers */
2338         for (i = slot; i < nritems; i++) {
2339                 u32 ioff;
2340                 item = btrfs_item_nr(leaf, i);
2341
2342                 if (!leaf->map_token) {
2343                         map_extent_buffer(leaf, (unsigned long)item,
2344                                         sizeof(struct btrfs_item),
2345                                         &leaf->map_token, &leaf->kaddr,
2346                                         &leaf->map_start, &leaf->map_len,
2347                                         KM_USER1);
2348                 }
2349                 ioff = btrfs_item_offset(leaf, item);
2350                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2351         }
2352
2353         if (leaf->map_token) {
2354                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2355                 leaf->map_token = NULL;
2356         }
2357
2358         /* shift the data */
2359         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2360                       data_end - data_size, btrfs_leaf_data(leaf) +
2361                       data_end, old_data - data_end);
2362
2363         data_end = old_data;
2364         old_size = btrfs_item_size_nr(leaf, slot);
2365         item = btrfs_item_nr(leaf, slot);
2366         btrfs_set_item_size(leaf, item, old_size + data_size);
2367         btrfs_mark_buffer_dirty(leaf);
2368
2369         ret = 0;
2370         if (btrfs_leaf_free_space(root, leaf) < 0) {
2371                 btrfs_print_leaf(root, leaf);
2372                 BUG();
2373         }
2374         return ret;
2375 }
2376
2377 /*
2378  * Given a key and some data, insert an item into the tree.
2379  * This does all the path init required, making room in the tree if needed.
2380  */
2381 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2382                             struct btrfs_root *root,
2383                             struct btrfs_path *path,
2384                             struct btrfs_key *cpu_key, u32 *data_size,
2385                             int nr)
2386 {
2387         struct extent_buffer *leaf;
2388         struct btrfs_item *item;
2389         int ret = 0;
2390         int slot;
2391         int slot_orig;
2392         int i;
2393         u32 nritems;
2394         u32 total_size = 0;
2395         u32 total_data = 0;
2396         unsigned int data_end;
2397         struct btrfs_disk_key disk_key;
2398
2399         for (i = 0; i < nr; i++) {
2400                 total_data += data_size[i];
2401         }
2402
2403         /* create a root if there isn't one */
2404         if (!root->node)
2405                 BUG();
2406
2407         total_size = total_data + (nr - 1) * sizeof(struct btrfs_item);
2408         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2409         if (ret == 0) {
2410                 return -EEXIST;
2411         }
2412         if (ret < 0)
2413                 goto out;
2414
2415         slot_orig = path->slots[0];
2416         leaf = path->nodes[0];
2417
2418         nritems = btrfs_header_nritems(leaf);
2419         data_end = leaf_data_end(root, leaf);
2420
2421         if (btrfs_leaf_free_space(root, leaf) <
2422             sizeof(struct btrfs_item) + total_size) {
2423                 btrfs_print_leaf(root, leaf);
2424                 printk("not enough freespace need %u have %d\n",
2425                        total_size, btrfs_leaf_free_space(root, leaf));
2426                 BUG();
2427         }
2428
2429         slot = path->slots[0];
2430         BUG_ON(slot < 0);
2431
2432         if (slot != nritems) {
2433                 int i;
2434                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2435
2436                 if (old_data < data_end) {
2437                         btrfs_print_leaf(root, leaf);
2438                         printk("slot %d old_data %d data_end %d\n",
2439                                slot, old_data, data_end);
2440                         BUG_ON(1);
2441                 }
2442                 /*
2443                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2444                  */
2445                 /* first correct the data pointers */
2446                 WARN_ON(leaf->map_token);
2447                 for (i = slot; i < nritems; i++) {
2448                         u32 ioff;
2449
2450                         item = btrfs_item_nr(leaf, i);
2451                         if (!leaf->map_token) {
2452                                 map_extent_buffer(leaf, (unsigned long)item,
2453                                         sizeof(struct btrfs_item),
2454                                         &leaf->map_token, &leaf->kaddr,
2455                                         &leaf->map_start, &leaf->map_len,
2456                                         KM_USER1);
2457                         }
2458
2459                         ioff = btrfs_item_offset(leaf, item);
2460                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2461                 }
2462                 if (leaf->map_token) {
2463                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2464                         leaf->map_token = NULL;
2465                 }
2466
2467                 /* shift the items */
2468                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2469                               btrfs_item_nr_offset(slot),
2470                               (nritems - slot) * sizeof(struct btrfs_item));
2471
2472                 /* shift the data */
2473                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2474                               data_end - total_data, btrfs_leaf_data(leaf) +
2475                               data_end, old_data - data_end);
2476                 data_end = old_data;
2477         }
2478
2479         /* setup the item for the new data */
2480         for (i = 0; i < nr; i++) {
2481                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2482                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2483                 item = btrfs_item_nr(leaf, slot + i);
2484                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2485                 data_end -= data_size[i];
2486                 btrfs_set_item_size(leaf, item, data_size[i]);
2487         }
2488         btrfs_set_header_nritems(leaf, nritems + nr);
2489         btrfs_mark_buffer_dirty(leaf);
2490
2491         ret = 0;
2492         if (slot == 0) {
2493                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2494                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2495         }
2496
2497         if (btrfs_leaf_free_space(root, leaf) < 0) {
2498                 btrfs_print_leaf(root, leaf);
2499                 BUG();
2500         }
2501
2502 out:
2503         return ret;
2504 }
2505
2506 /*
2507  * Given a key and some data, insert an item into the tree.
2508  * This does all the path init required, making room in the tree if needed.
2509  */
2510 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2511                       *root, struct btrfs_key *cpu_key, void *data, u32
2512                       data_size)
2513 {
2514         int ret = 0;
2515         struct btrfs_path *path;
2516         struct extent_buffer *leaf;
2517         unsigned long ptr;
2518
2519         path = btrfs_alloc_path();
2520         BUG_ON(!path);
2521         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2522         if (!ret) {
2523                 leaf = path->nodes[0];
2524                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2525                 write_extent_buffer(leaf, data, ptr, data_size);
2526                 btrfs_mark_buffer_dirty(leaf);
2527         }
2528         btrfs_free_path(path);
2529         return ret;
2530 }
2531
2532 /*
2533  * delete the pointer from a given node.
2534  *
2535  * If the delete empties a node, the node is removed from the tree,
2536  * continuing all the way the root if required.  The root is converted into
2537  * a leaf if all the nodes are emptied.
2538  */
2539 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2540                    struct btrfs_path *path, int level, int slot)
2541 {
2542         struct extent_buffer *parent = path->nodes[level];
2543         u32 nritems;
2544         int ret = 0;
2545         int wret;
2546
2547         nritems = btrfs_header_nritems(parent);
2548         if (slot != nritems -1) {
2549                 memmove_extent_buffer(parent,
2550                               btrfs_node_key_ptr_offset(slot),
2551                               btrfs_node_key_ptr_offset(slot + 1),
2552                               sizeof(struct btrfs_key_ptr) *
2553                               (nritems - slot - 1));
2554         }
2555         nritems--;
2556         btrfs_set_header_nritems(parent, nritems);
2557         if (nritems == 0 && parent == root->node) {
2558                 BUG_ON(btrfs_header_level(root->node) != 1);
2559                 /* just turn the root into a leaf and break */
2560                 btrfs_set_header_level(root->node, 0);
2561         } else if (slot == 0) {
2562                 struct btrfs_disk_key disk_key;
2563
2564                 btrfs_node_key(parent, &disk_key, 0);
2565                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2566                 if (wret)
2567                         ret = wret;
2568         }
2569         btrfs_mark_buffer_dirty(parent);
2570         return ret;
2571 }
2572
2573 /*
2574  * delete the item at the leaf level in path.  If that empties
2575  * the leaf, remove it from the tree
2576  */
2577 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2578                     struct btrfs_path *path, int slot, int nr)
2579 {
2580         struct extent_buffer *leaf;
2581         struct btrfs_item *item;
2582         int last_off;
2583         int dsize = 0;
2584         int ret = 0;
2585         int wret;
2586         int i;
2587         u32 nritems;
2588
2589         leaf = path->nodes[0];
2590         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2591
2592         for (i = 0; i < nr; i++)
2593                 dsize += btrfs_item_size_nr(leaf, slot + i);
2594
2595         nritems = btrfs_header_nritems(leaf);
2596
2597         if (slot + nr != nritems) {
2598                 int i;
2599                 int data_end = leaf_data_end(root, leaf);
2600
2601                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2602                               data_end + dsize,
2603                               btrfs_leaf_data(leaf) + data_end,
2604                               last_off - data_end);
2605
2606                 for (i = slot + nr; i < nritems; i++) {
2607                         u32 ioff;
2608
2609                         item = btrfs_item_nr(leaf, i);
2610                         if (!leaf->map_token) {
2611                                 map_extent_buffer(leaf, (unsigned long)item,
2612                                         sizeof(struct btrfs_item),
2613                                         &leaf->map_token, &leaf->kaddr,
2614                                         &leaf->map_start, &leaf->map_len,
2615                                         KM_USER1);
2616                         }
2617                         ioff = btrfs_item_offset(leaf, item);
2618                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2619                 }
2620
2621                 if (leaf->map_token) {
2622                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2623                         leaf->map_token = NULL;
2624                 }
2625
2626                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2627                               btrfs_item_nr_offset(slot + nr),
2628                               sizeof(struct btrfs_item) *
2629                               (nritems - slot - nr));
2630         }
2631         btrfs_set_header_nritems(leaf, nritems - nr);
2632         nritems -= nr;
2633
2634         /* delete the leaf if we've emptied it */
2635         if (nritems == 0) {
2636                 if (leaf == root->node) {
2637                         btrfs_set_header_level(leaf, 0);
2638                 } else {
2639                         u64 root_gen = btrfs_header_generation(path->nodes[1]);
2640                         clean_tree_block(trans, root, leaf);
2641                         wait_on_tree_block_writeback(root, leaf);
2642                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
2643                         if (wret)
2644                                 ret = wret;
2645                         wret = btrfs_free_extent(trans, root,
2646                                          leaf->start, leaf->len,
2647                                          btrfs_header_owner(path->nodes[1]),
2648                                          root_gen, 0, 0, 1);
2649                         if (wret)
2650                                 ret = wret;
2651                 }
2652         } else {
2653                 int used = leaf_space_used(leaf, 0, nritems);
2654                 if (slot == 0) {
2655                         struct btrfs_disk_key disk_key;
2656
2657                         btrfs_item_key(leaf, &disk_key, 0);
2658                         wret = fixup_low_keys(trans, root, path,
2659                                               &disk_key, 1);
2660                         if (wret)
2661                                 ret = wret;
2662                 }
2663
2664                 /* delete the leaf if it is mostly empty */
2665                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2666                         /* push_leaf_left fixes the path.
2667                          * make sure the path still points to our leaf
2668                          * for possible call to del_ptr below
2669                          */
2670                         slot = path->slots[1];
2671                         extent_buffer_get(leaf);
2672
2673                         wret = push_leaf_left(trans, root, path, 1, 1);
2674                         if (wret < 0 && wret != -ENOSPC)
2675                                 ret = wret;
2676
2677                         if (path->nodes[0] == leaf &&
2678                             btrfs_header_nritems(leaf)) {
2679                                 wret = push_leaf_right(trans, root, path, 1, 1);
2680                                 if (wret < 0 && wret != -ENOSPC)
2681                                         ret = wret;
2682                         }
2683
2684                         if (btrfs_header_nritems(leaf) == 0) {
2685                                 u64 root_gen;
2686                                 u64 bytenr = leaf->start;
2687                                 u32 blocksize = leaf->len;
2688
2689                                 root_gen = btrfs_header_generation(
2690                                                            path->nodes[1]);
2691
2692                                 clean_tree_block(trans, root, leaf);
2693                                 wait_on_tree_block_writeback(root, leaf);
2694
2695                                 wret = del_ptr(trans, root, path, 1, slot);
2696                                 if (wret)
2697                                         ret = wret;
2698
2699                                 free_extent_buffer(leaf);
2700                                 wret = btrfs_free_extent(trans, root, bytenr,
2701                                              blocksize,
2702                                              btrfs_header_owner(path->nodes[1]),
2703                                              root_gen, 0, 0, 1);
2704                                 if (wret)
2705                                         ret = wret;
2706                         } else {
2707                                 btrfs_mark_buffer_dirty(leaf);
2708                                 free_extent_buffer(leaf);
2709                         }
2710                 } else {
2711                         btrfs_mark_buffer_dirty(leaf);
2712                 }
2713         }
2714         return ret;
2715 }
2716
2717 /*
2718  * walk up the tree as far as required to find the previous leaf.
2719  * returns 0 if it found something or 1 if there are no lesser leaves.
2720  * returns < 0 on io errors.
2721  */
2722 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2723 {
2724         u64 bytenr;
2725         int slot;
2726         int level = 1;
2727         struct extent_buffer *c;
2728         struct extent_buffer *next = NULL;
2729
2730         while(level < BTRFS_MAX_LEVEL) {
2731                 if (!path->nodes[level])
2732                         return 1;
2733
2734                 slot = path->slots[level];
2735                 c = path->nodes[level];
2736                 if (slot == 0) {
2737                         level++;
2738                         if (level == BTRFS_MAX_LEVEL)
2739                                 return 1;
2740                         continue;
2741                 }
2742                 slot--;
2743
2744                 bytenr = btrfs_node_blockptr(c, slot);
2745                 if (next)
2746                         free_extent_buffer(next);
2747
2748                 next = read_tree_block(root, bytenr,
2749                                        btrfs_level_size(root, level - 1));
2750                 break;
2751         }
2752         path->slots[level] = slot;
2753         while(1) {
2754                 level--;
2755                 c = path->nodes[level];
2756                 free_extent_buffer(c);
2757                 slot = btrfs_header_nritems(next);
2758                 if (slot != 0)
2759                         slot--;
2760                 path->nodes[level] = next;
2761                 path->slots[level] = slot;
2762                 if (!level)
2763                         break;
2764                 next = read_tree_block(root, btrfs_node_blockptr(next, slot),
2765                                        btrfs_level_size(root, level - 1));
2766         }
2767         return 0;
2768 }
2769
2770 /*
2771  * walk up the tree as far as required to find the next leaf.
2772  * returns 0 if it found something or 1 if there are no greater leaves.
2773  * returns < 0 on io errors.
2774  */
2775 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2776 {
2777         int slot;
2778         int level = 1;
2779         u64 bytenr;
2780         struct extent_buffer *c;
2781         struct extent_buffer *next = NULL;
2782
2783         while(level < BTRFS_MAX_LEVEL) {
2784                 if (!path->nodes[level])
2785                         return 1;
2786
2787                 slot = path->slots[level] + 1;
2788                 c = path->nodes[level];
2789                 if (slot >= btrfs_header_nritems(c)) {
2790                         level++;
2791                         if (level == BTRFS_MAX_LEVEL)
2792                                 return 1;
2793                         continue;
2794                 }
2795
2796                 bytenr = btrfs_node_blockptr(c, slot);
2797                 if (next)
2798                         free_extent_buffer(next);
2799
2800                 if (path->reada)
2801                         reada_for_search(root, path, level, slot, 0);
2802
2803                 next = read_tree_block(root, bytenr,
2804                                        btrfs_level_size(root, level -1));
2805                 break;
2806         }
2807         path->slots[level] = slot;
2808         while(1) {
2809                 level--;
2810                 c = path->nodes[level];
2811                 free_extent_buffer(c);
2812                 path->nodes[level] = next;
2813                 path->slots[level] = 0;
2814                 if (!level)
2815                         break;
2816                 if (path->reada)
2817                         reada_for_search(root, path, level, 0, 0);
2818                 next = read_tree_block(root, btrfs_node_blockptr(next, 0),
2819                                        btrfs_level_size(root, level - 1));
2820         }
2821         return 0;
2822 }
2823
2824 int btrfs_previous_item(struct btrfs_root *root,
2825                         struct btrfs_path *path, u64 min_objectid,
2826                         int type)
2827 {
2828         struct btrfs_key found_key;
2829         struct extent_buffer *leaf;
2830         int ret;
2831
2832         while(1) {
2833                 if (path->slots[0] == 0) {
2834                         ret = btrfs_prev_leaf(root, path);
2835                         if (ret != 0)
2836                                 return ret;
2837                 } else {
2838                         path->slots[0]--;
2839                 }
2840                 leaf = path->nodes[0];
2841                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2842                 if (found_key.type == type)
2843                         return 0;
2844         }
2845         return 1;
2846 }
2847