SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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         btrfs_set_node_key(c, &lower_key, 0);
1386         btrfs_set_node_blockptr(c, 0, lower->start);
1387         lower_gen = btrfs_header_generation(lower);
1388         WARN_ON(lower_gen == 0);
1389
1390         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1391
1392         btrfs_mark_buffer_dirty(c);
1393
1394         /* the super has an extra ref to root->node */
1395         free_extent_buffer(root->node);
1396         root->node = c;
1397         add_root_to_dirty_list(root);
1398         extent_buffer_get(c);
1399         path->nodes[level] = c;
1400         path->slots[level] = 0;
1401
1402         if (root->ref_cows && lower_gen != trans->transid) {
1403                 struct btrfs_path *back_path = btrfs_alloc_path();
1404                 int ret;
1405                 ret = btrfs_insert_extent_backref(trans,
1406                                                   root->fs_info->extent_root,
1407                                                   path, lower->start,
1408                                                   root->root_key.objectid,
1409                                                   trans->transid, 0, 0);
1410                 BUG_ON(ret);
1411                 btrfs_free_path(back_path);
1412         }
1413         return 0;
1414 }
1415
1416 /*
1417  * worker function to insert a single pointer in a node.
1418  * the node should have enough room for the pointer already
1419  *
1420  * slot and level indicate where you want the key to go, and
1421  * blocknr is the block the key points to.
1422  *
1423  * returns zero on success and < 0 on any error
1424  */
1425 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1426                       *root, struct btrfs_path *path, struct btrfs_disk_key
1427                       *key, u64 bytenr, int slot, int level)
1428 {
1429         struct extent_buffer *lower;
1430         int nritems;
1431
1432         BUG_ON(!path->nodes[level]);
1433         lower = path->nodes[level];
1434         nritems = btrfs_header_nritems(lower);
1435         if (slot > nritems)
1436                 BUG();
1437         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1438                 BUG();
1439         if (slot != nritems) {
1440                 memmove_extent_buffer(lower,
1441                               btrfs_node_key_ptr_offset(slot + 1),
1442                               btrfs_node_key_ptr_offset(slot),
1443                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1444         }
1445         btrfs_set_node_key(lower, key, slot);
1446         btrfs_set_node_blockptr(lower, slot, bytenr);
1447         WARN_ON(trans->transid == 0);
1448         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1449         btrfs_set_header_nritems(lower, nritems + 1);
1450         btrfs_mark_buffer_dirty(lower);
1451         return 0;
1452 }
1453
1454 /*
1455  * split the node at the specified level in path in two.
1456  * The path is corrected to point to the appropriate node after the split
1457  *
1458  * Before splitting this tries to make some room in the node by pushing
1459  * left and right, if either one works, it returns right away.
1460  *
1461  * returns 0 on success and < 0 on failure
1462  */
1463 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1464                       *root, struct btrfs_path *path, int level)
1465 {
1466         u64 root_gen;
1467         struct extent_buffer *c;
1468         struct extent_buffer *split;
1469         struct btrfs_disk_key disk_key;
1470         int mid;
1471         int ret;
1472         int wret;
1473         u32 c_nritems;
1474
1475         c = path->nodes[level];
1476         WARN_ON(btrfs_header_generation(c) != trans->transid);
1477         if (c == root->node) {
1478                 /* trying to split the root, lets make a new one */
1479                 ret = insert_new_root(trans, root, path, level + 1);
1480                 if (ret)
1481                         return ret;
1482         } else {
1483                 ret = push_nodes_for_insert(trans, root, path, level);
1484                 c = path->nodes[level];
1485                 if (!ret && btrfs_header_nritems(c) <
1486                     BTRFS_NODEPTRS_PER_BLOCK(root) - 1)
1487                         return 0;
1488                 if (ret < 0)
1489                         return ret;
1490         }
1491
1492         c_nritems = btrfs_header_nritems(c);
1493         if (root->ref_cows)
1494                 root_gen = trans->transid;
1495         else
1496                 root_gen = 0;
1497
1498         btrfs_node_key(c, &disk_key, 0);
1499         split = __btrfs_alloc_free_block(trans, root, root->nodesize,
1500                                          root->root_key.objectid,
1501                                          root_gen,
1502                                          btrfs_disk_key_objectid(&disk_key),
1503                                          level, c->start, 0);
1504         if (IS_ERR(split))
1505                 return PTR_ERR(split);
1506
1507         btrfs_set_header_flags(split, btrfs_header_flags(c));
1508         btrfs_set_header_level(split, btrfs_header_level(c));
1509         btrfs_set_header_bytenr(split, split->start);
1510         btrfs_set_header_generation(split, trans->transid);
1511         btrfs_set_header_owner(split, root->root_key.objectid);
1512         btrfs_set_header_flags(split, 0);
1513         write_extent_buffer(split, root->fs_info->fsid,
1514                             (unsigned long)btrfs_header_fsid(split),
1515                             BTRFS_FSID_SIZE);
1516
1517         mid = (c_nritems + 1) / 2;
1518
1519         copy_extent_buffer(split, c,
1520                            btrfs_node_key_ptr_offset(0),
1521                            btrfs_node_key_ptr_offset(mid),
1522                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1523         btrfs_set_header_nritems(split, c_nritems - mid);
1524         btrfs_set_header_nritems(c, mid);
1525         ret = 0;
1526
1527         btrfs_mark_buffer_dirty(c);
1528         btrfs_mark_buffer_dirty(split);
1529
1530         btrfs_node_key(split, &disk_key, 0);
1531         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1532                           path->slots[level + 1] + 1,
1533                           level + 1);
1534         if (wret)
1535                 ret = wret;
1536
1537         if (path->slots[level] >= mid) {
1538                 path->slots[level] -= mid;
1539                 free_extent_buffer(c);
1540                 path->nodes[level] = split;
1541                 path->slots[level + 1] += 1;
1542         } else {
1543                 free_extent_buffer(split);
1544         }
1545         return ret;
1546 }
1547
1548 /*
1549  * how many bytes are required to store the items in a leaf.  start
1550  * and nr indicate which items in the leaf to check.  This totals up the
1551  * space used both by the item structs and the item data
1552  */
1553 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1554 {
1555         int data_len;
1556         int nritems = btrfs_header_nritems(l);
1557         int end = min(nritems, start + nr) - 1;
1558
1559         if (!nr)
1560                 return 0;
1561         data_len = btrfs_item_end_nr(l, start);
1562         data_len = data_len - btrfs_item_offset_nr(l, end);
1563         data_len += sizeof(struct btrfs_item) * nr;
1564         WARN_ON(data_len < 0);
1565         return data_len;
1566 }
1567
1568 /*
1569  * The space between the end of the leaf items and
1570  * the start of the leaf data.  IOW, how much room
1571  * the leaf has left for both items and data
1572  */
1573 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1574 {
1575         int nritems = btrfs_header_nritems(leaf);
1576         int ret;
1577         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1578         if (ret < 0) {
1579                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1580                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1581                        leaf_space_used(leaf, 0, nritems), nritems);
1582         }
1583         return ret;
1584 }
1585
1586 /*
1587  * push some data in the path leaf to the right, trying to free up at
1588  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1589  *
1590  * returns 1 if the push failed because the other node didn't have enough
1591  * room, 0 if everything worked out and < 0 if there were major errors.
1592  */
1593 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1594                            *root, struct btrfs_path *path, int data_size,
1595                            int empty)
1596 {
1597         struct extent_buffer *left = path->nodes[0];
1598         struct extent_buffer *right;
1599         struct extent_buffer *upper;
1600         struct btrfs_disk_key disk_key;
1601         int slot;
1602         u32 i;
1603         int free_space;
1604         int push_space = 0;
1605         int push_items = 0;
1606         struct btrfs_item *item;
1607         u32 left_nritems;
1608         u32 nr;
1609         u32 right_nritems;
1610         u32 data_end;
1611         u32 this_item_size;
1612         int ret;
1613
1614         slot = path->slots[1];
1615         if (!path->nodes[1]) {
1616                 return 1;
1617         }
1618         upper = path->nodes[1];
1619         if (slot >= btrfs_header_nritems(upper) - 1)
1620                 return 1;
1621
1622         right = read_tree_block(root, btrfs_node_blockptr(upper, slot + 1),
1623                                 root->leafsize);
1624         free_space = btrfs_leaf_free_space(root, right);
1625         if (free_space < data_size + sizeof(struct btrfs_item)) {
1626                 free_extent_buffer(right);
1627                 return 1;
1628         }
1629
1630         /* cow and double check */
1631         ret = btrfs_cow_block(trans, root, right, upper,
1632                               slot + 1, &right);
1633         if (ret) {
1634                 free_extent_buffer(right);
1635                 return 1;
1636         }
1637         free_space = btrfs_leaf_free_space(root, right);
1638         if (free_space < data_size + sizeof(struct btrfs_item)) {
1639                 free_extent_buffer(right);
1640                 return 1;
1641         }
1642
1643         left_nritems = btrfs_header_nritems(left);
1644         if (left_nritems == 0) {
1645                 free_extent_buffer(right);
1646                 return 1;
1647         }
1648
1649         if (empty)
1650                 nr = 0;
1651         else
1652                 nr = 1;
1653
1654         i = left_nritems - 1;
1655         while (i >= nr) {
1656                 item = btrfs_item_nr(left, i);
1657
1658                 if (path->slots[0] == i)
1659                         push_space += data_size + sizeof(*item);
1660
1661                 if (!left->map_token) {
1662                         map_extent_buffer(left, (unsigned long)item,
1663                                         sizeof(struct btrfs_item),
1664                                         &left->map_token, &left->kaddr,
1665                                         &left->map_start, &left->map_len,
1666                                         KM_USER1);
1667                 }
1668
1669                 this_item_size = btrfs_item_size(left, item);
1670                 if (this_item_size + sizeof(*item) + push_space > free_space)
1671                         break;
1672                 push_items++;
1673                 push_space += this_item_size + sizeof(*item);
1674                 if (i == 0)
1675                         break;
1676                 i--;
1677         }
1678         if (left->map_token) {
1679                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1680                 left->map_token = NULL;
1681         }
1682
1683         if (push_items == 0) {
1684                 free_extent_buffer(right);
1685                 return 1;
1686         }
1687
1688         if (!empty && push_items == left_nritems)
1689                 WARN_ON(1);
1690
1691         /* push left to right */
1692         right_nritems = btrfs_header_nritems(right);
1693
1694         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1695         push_space -= leaf_data_end(root, left);
1696
1697         /* make room in the right data area */
1698         data_end = leaf_data_end(root, right);
1699         memmove_extent_buffer(right,
1700                               btrfs_leaf_data(right) + data_end - push_space,
1701                               btrfs_leaf_data(right) + data_end,
1702                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1703
1704         /* copy from the left data area */
1705         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1706                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1707                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1708                      push_space);
1709
1710         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1711                               btrfs_item_nr_offset(0),
1712                               right_nritems * sizeof(struct btrfs_item));
1713
1714         /* copy the items from left to right */
1715         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1716                    btrfs_item_nr_offset(left_nritems - push_items),
1717                    push_items * sizeof(struct btrfs_item));
1718
1719         /* update the item pointers */
1720         right_nritems += push_items;
1721         btrfs_set_header_nritems(right, right_nritems);
1722         push_space = BTRFS_LEAF_DATA_SIZE(root);
1723         for (i = 0; i < right_nritems; i++) {
1724                 item = btrfs_item_nr(right, i);
1725                 if (!right->map_token) {
1726                         map_extent_buffer(right, (unsigned long)item,
1727                                         sizeof(struct btrfs_item),
1728                                         &right->map_token, &right->kaddr,
1729                                         &right->map_start, &right->map_len,
1730                                         KM_USER1);
1731                 }
1732                 push_space -= btrfs_item_size(right, item);
1733                 btrfs_set_item_offset(right, item, push_space);
1734         }
1735
1736         if (right->map_token) {
1737                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1738                 right->map_token = NULL;
1739         }
1740         left_nritems -= push_items;
1741         btrfs_set_header_nritems(left, left_nritems);
1742
1743         if (left_nritems)
1744                 btrfs_mark_buffer_dirty(left);
1745         btrfs_mark_buffer_dirty(right);
1746
1747         btrfs_item_key(right, &disk_key, 0);
1748         btrfs_set_node_key(upper, &disk_key, slot + 1);
1749         btrfs_mark_buffer_dirty(upper);
1750
1751         /* then fixup the leaf pointer in the path */
1752         if (path->slots[0] >= left_nritems) {
1753                 path->slots[0] -= left_nritems;
1754                 free_extent_buffer(path->nodes[0]);
1755                 path->nodes[0] = right;
1756                 path->slots[1] += 1;
1757         } else {
1758                 free_extent_buffer(right);
1759         }
1760         return 0;
1761 }
1762 /*
1763  * push some data in the path leaf to the left, trying to free up at
1764  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1765  */
1766 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1767                           *root, struct btrfs_path *path, int data_size,
1768                           int empty)
1769 {
1770         struct btrfs_disk_key disk_key;
1771         struct extent_buffer *right = path->nodes[0];
1772         struct extent_buffer *left;
1773         int slot;
1774         int i;
1775         int free_space;
1776         int push_space = 0;
1777         int push_items = 0;
1778         struct btrfs_item *item;
1779         u32 old_left_nritems;
1780         u32 right_nritems;
1781         u32 nr;
1782         int ret = 0;
1783         int wret;
1784         u32 this_item_size;
1785         u32 old_left_item_size;
1786
1787         slot = path->slots[1];
1788         if (slot == 0)
1789                 return 1;
1790         if (!path->nodes[1])
1791                 return 1;
1792
1793         right_nritems = btrfs_header_nritems(right);
1794         if (right_nritems == 0) {
1795                 return 1;
1796         }
1797
1798         left = read_tree_block(root, btrfs_node_blockptr(path->nodes[1],
1799                                slot - 1), root->leafsize);
1800         free_space = btrfs_leaf_free_space(root, left);
1801         if (free_space < data_size + sizeof(struct btrfs_item)) {
1802                 free_extent_buffer(left);
1803                 return 1;
1804         }
1805
1806         /* cow and double check */
1807         ret = btrfs_cow_block(trans, root, left,
1808                               path->nodes[1], slot - 1, &left);
1809         if (ret) {
1810                 /* we hit -ENOSPC, but it isn't fatal here */
1811                 free_extent_buffer(left);
1812                 return 1;
1813         }
1814
1815         free_space = btrfs_leaf_free_space(root, left);
1816         if (free_space < data_size + sizeof(struct btrfs_item)) {
1817                 free_extent_buffer(left);
1818                 return 1;
1819         }
1820
1821         if (empty)
1822                 nr = right_nritems;
1823         else
1824                 nr = right_nritems - 1;
1825
1826         for (i = 0; i < nr; i++) {
1827                 item = btrfs_item_nr(right, i);
1828                 if (!right->map_token) {
1829                         map_extent_buffer(right, (unsigned long)item,
1830                                         sizeof(struct btrfs_item),
1831                                         &right->map_token, &right->kaddr,
1832                                         &right->map_start, &right->map_len,
1833                                         KM_USER1);
1834                 }
1835
1836                 if (path->slots[0] == i)
1837                         push_space += data_size + sizeof(*item);
1838
1839                 this_item_size = btrfs_item_size(right, item);
1840                 if (this_item_size + sizeof(*item) + push_space > free_space)
1841                         break;
1842
1843                 push_items++;
1844                 push_space += this_item_size + sizeof(*item);
1845         }
1846
1847         if (right->map_token) {
1848                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1849                 right->map_token = NULL;
1850         }
1851
1852         if (push_items == 0) {
1853                 free_extent_buffer(left);
1854                 return 1;
1855         }
1856         if (!empty && push_items == btrfs_header_nritems(right))
1857                 WARN_ON(1);
1858
1859         /* push data from right to left */
1860         copy_extent_buffer(left, right,
1861                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1862                            btrfs_item_nr_offset(0),
1863                            push_items * sizeof(struct btrfs_item));
1864
1865         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1866                      btrfs_item_offset_nr(right, push_items -1);
1867
1868         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1869                      leaf_data_end(root, left) - push_space,
1870                      btrfs_leaf_data(right) +
1871                      btrfs_item_offset_nr(right, push_items - 1),
1872                      push_space);
1873         old_left_nritems = btrfs_header_nritems(left);
1874         BUG_ON(old_left_nritems < 0);
1875
1876         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1877         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1878                 u32 ioff;
1879
1880                 item = btrfs_item_nr(left, i);
1881                 if (!left->map_token) {
1882                         map_extent_buffer(left, (unsigned long)item,
1883                                         sizeof(struct btrfs_item),
1884                                         &left->map_token, &left->kaddr,
1885                                         &left->map_start, &left->map_len,
1886                                         KM_USER1);
1887                 }
1888
1889                 ioff = btrfs_item_offset(left, item);
1890                 btrfs_set_item_offset(left, item,
1891                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1892         }
1893         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1894         if (left->map_token) {
1895                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1896                 left->map_token = NULL;
1897         }
1898
1899         /* fixup right node */
1900         if (push_items > right_nritems) {
1901                 printk("push items %d nr %u\n", push_items, right_nritems);
1902                 WARN_ON(1);
1903         }
1904
1905         if (push_items < right_nritems) {
1906                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1907                                                   leaf_data_end(root, right);
1908                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
1909                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1910                                       btrfs_leaf_data(right) +
1911                                       leaf_data_end(root, right), push_space);
1912
1913                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1914                               btrfs_item_nr_offset(push_items),
1915                              (btrfs_header_nritems(right) - push_items) *
1916                              sizeof(struct btrfs_item));
1917         }
1918         right_nritems -= push_items;
1919         btrfs_set_header_nritems(right, right_nritems);
1920         push_space = BTRFS_LEAF_DATA_SIZE(root);
1921         for (i = 0; i < right_nritems; i++) {
1922                 item = btrfs_item_nr(right, i);
1923
1924                 if (!right->map_token) {
1925                         map_extent_buffer(right, (unsigned long)item,
1926                                         sizeof(struct btrfs_item),
1927                                         &right->map_token, &right->kaddr,
1928                                         &right->map_start, &right->map_len,
1929                                         KM_USER1);
1930                 }
1931
1932                 push_space = push_space - btrfs_item_size(right, item);
1933                 btrfs_set_item_offset(right, item, push_space);
1934         }
1935         if (right->map_token) {
1936                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1937                 right->map_token = NULL;
1938         }
1939
1940         btrfs_mark_buffer_dirty(left);
1941         if (right_nritems)
1942                 btrfs_mark_buffer_dirty(right);
1943
1944         btrfs_item_key(right, &disk_key, 0);
1945         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1946         if (wret)
1947                 ret = wret;
1948
1949         /* then fixup the leaf pointer in the path */
1950         if (path->slots[0] < push_items) {
1951                 path->slots[0] += old_left_nritems;
1952                 free_extent_buffer(path->nodes[0]);
1953                 path->nodes[0] = left;
1954                 path->slots[1] -= 1;
1955         } else {
1956                 free_extent_buffer(left);
1957                 path->slots[0] -= push_items;
1958         }
1959         BUG_ON(path->slots[0] < 0);
1960         return ret;
1961 }
1962
1963 /*
1964  * split the path's leaf in two, making sure there is at least data_size
1965  * available for the resulting leaf level of the path.
1966  *
1967  * returns 0 if all went well and < 0 on failure.
1968  */
1969 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
1970                       *root, struct btrfs_key *ins_key,
1971                       struct btrfs_path *path, int data_size, int extend)
1972 {
1973         u64 root_gen;
1974         struct extent_buffer *l;
1975         u32 nritems;
1976         int mid;
1977         int slot;
1978         struct extent_buffer *right;
1979         int space_needed = data_size + sizeof(struct btrfs_item);
1980         int data_copy_size;
1981         int rt_data_off;
1982         int i;
1983         int ret = 0;
1984         int wret;
1985         int double_split;
1986         int num_doubles = 0;
1987         struct btrfs_disk_key disk_key;
1988
1989         if (extend)
1990                 space_needed = data_size;
1991
1992         if (root->ref_cows)
1993                 root_gen = trans->transid;
1994         else
1995                 root_gen = 0;
1996
1997         /* first try to make some room by pushing left and right */
1998         if (ins_key->type != BTRFS_DIR_ITEM_KEY) {
1999                 wret = push_leaf_right(trans, root, path, data_size, 0);
2000                 if (wret < 0) {
2001                         return wret;
2002                 }
2003                 if (wret) {
2004                         wret = push_leaf_left(trans, root, path, data_size, 0);
2005                         if (wret < 0)
2006                                 return wret;
2007                 }
2008                 l = path->nodes[0];
2009
2010                 /* did the pushes work? */
2011                 if (btrfs_leaf_free_space(root, l) >= space_needed)
2012                         return 0;
2013         }
2014
2015         if (!path->nodes[1]) {
2016                 ret = insert_new_root(trans, root, path, 1);
2017                 if (ret)
2018                         return ret;
2019         }
2020 again:
2021         double_split = 0;
2022         l = path->nodes[0];
2023         slot = path->slots[0];
2024         nritems = btrfs_header_nritems(l);
2025         mid = (nritems + 1)/ 2;
2026
2027         btrfs_item_key(l, &disk_key, 0);
2028
2029         right = __btrfs_alloc_free_block(trans, root, root->leafsize,
2030                                          root->root_key.objectid,
2031                                          root_gen, disk_key.objectid, 0,
2032                                          l->start, 0);
2033         if (IS_ERR(right)) {
2034                 BUG_ON(1);
2035                 return PTR_ERR(right);
2036         }
2037
2038         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2039         btrfs_set_header_bytenr(right, right->start);
2040         btrfs_set_header_generation(right, trans->transid);
2041         btrfs_set_header_owner(right, root->root_key.objectid);
2042         btrfs_set_header_level(right, 0);
2043         write_extent_buffer(right, root->fs_info->fsid,
2044                             (unsigned long)btrfs_header_fsid(right),
2045                             BTRFS_FSID_SIZE);
2046         if (mid <= slot) {
2047                 if (nritems == 1 ||
2048                     leaf_space_used(l, mid, nritems - mid) + space_needed >
2049                         BTRFS_LEAF_DATA_SIZE(root)) {
2050                         if (slot >= nritems) {
2051                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2052                                 btrfs_set_header_nritems(right, 0);
2053                                 wret = insert_ptr(trans, root, path,
2054                                                   &disk_key, right->start,
2055                                                   path->slots[1] + 1, 1);
2056                                 if (wret)
2057                                         ret = wret;
2058                                 free_extent_buffer(path->nodes[0]);
2059                                 path->nodes[0] = right;
2060                                 path->slots[0] = 0;
2061                                 path->slots[1] += 1;
2062                                 btrfs_mark_buffer_dirty(right);
2063                                 return ret;
2064                         }
2065                         mid = slot;
2066                         if (mid != nritems &&
2067                             leaf_space_used(l, mid, nritems - mid) +
2068                             space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2069                                 double_split = 1;
2070                         }
2071                 }
2072         } else {
2073                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
2074                         BTRFS_LEAF_DATA_SIZE(root)) {
2075                         if (!extend && slot == 0) {
2076                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2077                                 btrfs_set_header_nritems(right, 0);
2078                                 wret = insert_ptr(trans, root, path,
2079                                                   &disk_key,
2080                                                   right->start,
2081                                                   path->slots[1], 1);
2082                                 if (wret)
2083                                         ret = wret;
2084                                 free_extent_buffer(path->nodes[0]);
2085                                 path->nodes[0] = right;
2086                                 path->slots[0] = 0;
2087                                 if (path->slots[1] == 0) {
2088                                         wret = fixup_low_keys(trans, root,
2089                                                    path, &disk_key, 1);
2090                                         if (wret)
2091                                                 ret = wret;
2092                                 }
2093                                 btrfs_mark_buffer_dirty(right);
2094                                 return ret;
2095                         } else if (extend && slot == 0) {
2096                                 mid = 1;
2097                         } else {
2098                                 mid = slot;
2099                                 if (mid != nritems &&
2100                                     leaf_space_used(l, mid, nritems - mid) +
2101                                     space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2102                                         double_split = 1;
2103                                 }
2104                         }
2105                 }
2106         }
2107         nritems = nritems - mid;
2108         btrfs_set_header_nritems(right, nritems);
2109         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2110
2111         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2112                            btrfs_item_nr_offset(mid),
2113                            nritems * sizeof(struct btrfs_item));
2114
2115         copy_extent_buffer(right, l,
2116                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2117                      data_copy_size, btrfs_leaf_data(l) +
2118                      leaf_data_end(root, l), data_copy_size);
2119
2120         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2121                       btrfs_item_end_nr(l, mid);
2122
2123         for (i = 0; i < nritems; i++) {
2124                 struct btrfs_item *item = btrfs_item_nr(right, i);
2125                 u32 ioff;
2126
2127                 if (!right->map_token) {
2128                         map_extent_buffer(right, (unsigned long)item,
2129                                         sizeof(struct btrfs_item),
2130                                         &right->map_token, &right->kaddr,
2131                                         &right->map_start, &right->map_len,
2132                                         KM_USER1);
2133                 }
2134
2135                 ioff = btrfs_item_offset(right, item);
2136                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2137         }
2138
2139         if (right->map_token) {
2140                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2141                 right->map_token = NULL;
2142         }
2143
2144         btrfs_set_header_nritems(l, mid);
2145         ret = 0;
2146         btrfs_item_key(right, &disk_key, 0);
2147         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2148                           path->slots[1] + 1, 1);
2149         if (wret)
2150                 ret = wret;
2151
2152         btrfs_mark_buffer_dirty(right);
2153         btrfs_mark_buffer_dirty(l);
2154         BUG_ON(path->slots[0] != slot);
2155
2156         if (mid <= slot) {
2157                 free_extent_buffer(path->nodes[0]);
2158                 path->nodes[0] = right;
2159                 path->slots[0] -= mid;
2160                 path->slots[1] += 1;
2161         } else
2162                 free_extent_buffer(right);
2163
2164         BUG_ON(path->slots[0] < 0);
2165
2166         if (double_split) {
2167                 BUG_ON(num_doubles != 0);
2168                 num_doubles++;
2169                 goto again;
2170         }
2171         return ret;
2172 }
2173
2174 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2175                         struct btrfs_root *root,
2176                         struct btrfs_path *path,
2177                         u32 new_size, int from_end)
2178 {
2179         int ret = 0;
2180         int slot;
2181         int slot_orig;
2182         struct extent_buffer *leaf;
2183         struct btrfs_item *item;
2184         u32 nritems;
2185         unsigned int data_end;
2186         unsigned int old_data_start;
2187         unsigned int old_size;
2188         unsigned int size_diff;
2189         int i;
2190
2191         slot_orig = path->slots[0];
2192         leaf = path->nodes[0];
2193         slot = path->slots[0];
2194
2195         old_size = btrfs_item_size_nr(leaf, slot);
2196         if (old_size == new_size)
2197                 return 0;
2198
2199         nritems = btrfs_header_nritems(leaf);
2200         data_end = leaf_data_end(root, leaf);
2201
2202         old_data_start = btrfs_item_offset_nr(leaf, slot);
2203
2204         size_diff = old_size - new_size;
2205
2206         BUG_ON(slot < 0);
2207         BUG_ON(slot >= nritems);
2208
2209         /*
2210          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2211          */
2212         /* first correct the data pointers */
2213         for (i = slot; i < nritems; i++) {
2214                 u32 ioff;
2215                 item = btrfs_item_nr(leaf, i);
2216
2217                 if (!leaf->map_token) {
2218                         map_extent_buffer(leaf, (unsigned long)item,
2219                                         sizeof(struct btrfs_item),
2220                                         &leaf->map_token, &leaf->kaddr,
2221                                         &leaf->map_start, &leaf->map_len,
2222                                         KM_USER1);
2223                 }
2224
2225                 ioff = btrfs_item_offset(leaf, item);
2226                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2227         }
2228
2229         if (leaf->map_token) {
2230                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2231                 leaf->map_token = NULL;
2232         }
2233
2234         /* shift the data */
2235         if (from_end) {
2236                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2237                               data_end + size_diff, btrfs_leaf_data(leaf) +
2238                               data_end, old_data_start + new_size - data_end);
2239         } else {
2240                 struct btrfs_disk_key disk_key;
2241                 u64 offset;
2242
2243                 btrfs_item_key(leaf, &disk_key, slot);
2244
2245                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2246                         unsigned long ptr;
2247                         struct btrfs_file_extent_item *fi;
2248
2249                         fi = btrfs_item_ptr(leaf, slot,
2250                                             struct btrfs_file_extent_item);
2251                         fi = (struct btrfs_file_extent_item *)(
2252                              (unsigned long)fi - size_diff);
2253
2254                         if (btrfs_file_extent_type(leaf, fi) ==
2255                             BTRFS_FILE_EXTENT_INLINE) {
2256                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2257                                 memmove_extent_buffer(leaf, ptr,
2258                                         (unsigned long)fi,
2259                                         offsetof(struct btrfs_file_extent_item,
2260                                                  disk_bytenr));
2261                         }
2262                 }
2263
2264                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2265                               data_end + size_diff, btrfs_leaf_data(leaf) +
2266                               data_end, old_data_start - data_end);
2267
2268                 offset = btrfs_disk_key_offset(&disk_key);
2269                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2270                 btrfs_set_item_key(leaf, &disk_key, slot);
2271                 if (slot == 0)
2272                         fixup_low_keys(trans, root, path, &disk_key, 1);
2273         }
2274
2275         item = btrfs_item_nr(leaf, slot);
2276         btrfs_set_item_size(leaf, item, new_size);
2277         btrfs_mark_buffer_dirty(leaf);
2278
2279         ret = 0;
2280         if (btrfs_leaf_free_space(root, leaf) < 0) {
2281                 btrfs_print_leaf(root, leaf);
2282                 BUG();
2283         }
2284         return ret;
2285 }
2286
2287 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2288                       struct btrfs_root *root, struct btrfs_path *path,
2289                       u32 data_size)
2290 {
2291         int ret = 0;
2292         int slot;
2293         int slot_orig;
2294         struct extent_buffer *leaf;
2295         struct btrfs_item *item;
2296         u32 nritems;
2297         unsigned int data_end;
2298         unsigned int old_data;
2299         unsigned int old_size;
2300         int i;
2301
2302         slot_orig = path->slots[0];
2303         leaf = path->nodes[0];
2304
2305         nritems = btrfs_header_nritems(leaf);
2306         data_end = leaf_data_end(root, leaf);
2307
2308         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2309                 btrfs_print_leaf(root, leaf);
2310                 BUG();
2311         }
2312         slot = path->slots[0];
2313         old_data = btrfs_item_end_nr(leaf, slot);
2314
2315         BUG_ON(slot < 0);
2316         if (slot >= nritems) {
2317                 btrfs_print_leaf(root, leaf);
2318                 printk("slot %d too large, nritems %d\n", slot, nritems);
2319                 BUG_ON(1);
2320         }
2321
2322         /*
2323          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2324          */
2325         /* first correct the data pointers */
2326         for (i = slot; i < nritems; i++) {
2327                 u32 ioff;
2328                 item = btrfs_item_nr(leaf, i);
2329
2330                 if (!leaf->map_token) {
2331                         map_extent_buffer(leaf, (unsigned long)item,
2332                                         sizeof(struct btrfs_item),
2333                                         &leaf->map_token, &leaf->kaddr,
2334                                         &leaf->map_start, &leaf->map_len,
2335                                         KM_USER1);
2336                 }
2337                 ioff = btrfs_item_offset(leaf, item);
2338                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2339         }
2340
2341         if (leaf->map_token) {
2342                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2343                 leaf->map_token = NULL;
2344         }
2345
2346         /* shift the data */
2347         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2348                       data_end - data_size, btrfs_leaf_data(leaf) +
2349                       data_end, old_data - data_end);
2350
2351         data_end = old_data;
2352         old_size = btrfs_item_size_nr(leaf, slot);
2353         item = btrfs_item_nr(leaf, slot);
2354         btrfs_set_item_size(leaf, item, old_size + data_size);
2355         btrfs_mark_buffer_dirty(leaf);
2356
2357         ret = 0;
2358         if (btrfs_leaf_free_space(root, leaf) < 0) {
2359                 btrfs_print_leaf(root, leaf);
2360                 BUG();
2361         }
2362         return ret;
2363 }
2364
2365 /*
2366  * Given a key and some data, insert an item into the tree.
2367  * This does all the path init required, making room in the tree if needed.
2368  */
2369 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2370                             struct btrfs_root *root,
2371                             struct btrfs_path *path,
2372                             struct btrfs_key *cpu_key, u32 *data_size,
2373                             int nr)
2374 {
2375         struct extent_buffer *leaf;
2376         struct btrfs_item *item;
2377         int ret = 0;
2378         int slot;
2379         int slot_orig;
2380         int i;
2381         u32 nritems;
2382         u32 total_size = 0;
2383         u32 total_data = 0;
2384         unsigned int data_end;
2385         struct btrfs_disk_key disk_key;
2386
2387         for (i = 0; i < nr; i++) {
2388                 total_data += data_size[i];
2389         }
2390
2391         /* create a root if there isn't one */
2392         if (!root->node)
2393                 BUG();
2394
2395         total_size = total_data + (nr - 1) * sizeof(struct btrfs_item);
2396         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2397         if (ret == 0) {
2398                 return -EEXIST;
2399         }
2400         if (ret < 0)
2401                 goto out;
2402
2403         slot_orig = path->slots[0];
2404         leaf = path->nodes[0];
2405
2406         nritems = btrfs_header_nritems(leaf);
2407         data_end = leaf_data_end(root, leaf);
2408
2409         if (btrfs_leaf_free_space(root, leaf) <
2410             sizeof(struct btrfs_item) + total_size) {
2411                 btrfs_print_leaf(root, leaf);
2412                 printk("not enough freespace need %u have %d\n",
2413                        total_size, btrfs_leaf_free_space(root, leaf));
2414                 BUG();
2415         }
2416
2417         slot = path->slots[0];
2418         BUG_ON(slot < 0);
2419
2420         if (slot != nritems) {
2421                 int i;
2422                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2423
2424                 if (old_data < data_end) {
2425                         btrfs_print_leaf(root, leaf);
2426                         printk("slot %d old_data %d data_end %d\n",
2427                                slot, old_data, data_end);
2428                         BUG_ON(1);
2429                 }
2430                 /*
2431                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2432                  */
2433                 /* first correct the data pointers */
2434                 WARN_ON(leaf->map_token);
2435                 for (i = slot; i < nritems; i++) {
2436                         u32 ioff;
2437
2438                         item = btrfs_item_nr(leaf, i);
2439                         if (!leaf->map_token) {
2440                                 map_extent_buffer(leaf, (unsigned long)item,
2441                                         sizeof(struct btrfs_item),
2442                                         &leaf->map_token, &leaf->kaddr,
2443                                         &leaf->map_start, &leaf->map_len,
2444                                         KM_USER1);
2445                         }
2446
2447                         ioff = btrfs_item_offset(leaf, item);
2448                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2449                 }
2450                 if (leaf->map_token) {
2451                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2452                         leaf->map_token = NULL;
2453                 }
2454
2455                 /* shift the items */
2456                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2457                               btrfs_item_nr_offset(slot),
2458                               (nritems - slot) * sizeof(struct btrfs_item));
2459
2460                 /* shift the data */
2461                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2462                               data_end - total_data, btrfs_leaf_data(leaf) +
2463                               data_end, old_data - data_end);
2464                 data_end = old_data;
2465         }
2466
2467         /* setup the item for the new data */
2468         for (i = 0; i < nr; i++) {
2469                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2470                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2471                 item = btrfs_item_nr(leaf, slot + i);
2472                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2473                 data_end -= data_size[i];
2474                 btrfs_set_item_size(leaf, item, data_size[i]);
2475         }
2476         btrfs_set_header_nritems(leaf, nritems + nr);
2477         btrfs_mark_buffer_dirty(leaf);
2478
2479         ret = 0;
2480         if (slot == 0) {
2481                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2482                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2483         }
2484
2485         if (btrfs_leaf_free_space(root, leaf) < 0) {
2486                 btrfs_print_leaf(root, leaf);
2487                 BUG();
2488         }
2489
2490 out:
2491         return ret;
2492 }
2493
2494 /*
2495  * Given a key and some data, insert an item into the tree.
2496  * This does all the path init required, making room in the tree if needed.
2497  */
2498 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2499                       *root, struct btrfs_key *cpu_key, void *data, u32
2500                       data_size)
2501 {
2502         int ret = 0;
2503         struct btrfs_path *path;
2504         struct extent_buffer *leaf;
2505         unsigned long ptr;
2506
2507         path = btrfs_alloc_path();
2508         BUG_ON(!path);
2509         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2510         if (!ret) {
2511                 leaf = path->nodes[0];
2512                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2513                 write_extent_buffer(leaf, data, ptr, data_size);
2514                 btrfs_mark_buffer_dirty(leaf);
2515         }
2516         btrfs_free_path(path);
2517         return ret;
2518 }
2519
2520 /*
2521  * delete the pointer from a given node.
2522  *
2523  * If the delete empties a node, the node is removed from the tree,
2524  * continuing all the way the root if required.  The root is converted into
2525  * a leaf if all the nodes are emptied.
2526  */
2527 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2528                    struct btrfs_path *path, int level, int slot)
2529 {
2530         struct extent_buffer *parent = path->nodes[level];
2531         u32 nritems;
2532         int ret = 0;
2533         int wret;
2534
2535         nritems = btrfs_header_nritems(parent);
2536         if (slot != nritems -1) {
2537                 memmove_extent_buffer(parent,
2538                               btrfs_node_key_ptr_offset(slot),
2539                               btrfs_node_key_ptr_offset(slot + 1),
2540                               sizeof(struct btrfs_key_ptr) *
2541                               (nritems - slot - 1));
2542         }
2543         nritems--;
2544         btrfs_set_header_nritems(parent, nritems);
2545         if (nritems == 0 && parent == root->node) {
2546                 BUG_ON(btrfs_header_level(root->node) != 1);
2547                 /* just turn the root into a leaf and break */
2548                 btrfs_set_header_level(root->node, 0);
2549         } else if (slot == 0) {
2550                 struct btrfs_disk_key disk_key;
2551
2552                 btrfs_node_key(parent, &disk_key, 0);
2553                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2554                 if (wret)
2555                         ret = wret;
2556         }
2557         btrfs_mark_buffer_dirty(parent);
2558         return ret;
2559 }
2560
2561 /*
2562  * delete the item at the leaf level in path.  If that empties
2563  * the leaf, remove it from the tree
2564  */
2565 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2566                     struct btrfs_path *path, int slot, int nr)
2567 {
2568         struct extent_buffer *leaf;
2569         struct btrfs_item *item;
2570         int last_off;
2571         int dsize = 0;
2572         int ret = 0;
2573         int wret;
2574         int i;
2575         u32 nritems;
2576
2577         leaf = path->nodes[0];
2578         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2579
2580         for (i = 0; i < nr; i++)
2581                 dsize += btrfs_item_size_nr(leaf, slot + i);
2582
2583         nritems = btrfs_header_nritems(leaf);
2584
2585         if (slot + nr != nritems) {
2586                 int i;
2587                 int data_end = leaf_data_end(root, leaf);
2588
2589                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2590                               data_end + dsize,
2591                               btrfs_leaf_data(leaf) + data_end,
2592                               last_off - data_end);
2593
2594                 for (i = slot + nr; i < nritems; i++) {
2595                         u32 ioff;
2596
2597                         item = btrfs_item_nr(leaf, i);
2598                         if (!leaf->map_token) {
2599                                 map_extent_buffer(leaf, (unsigned long)item,
2600                                         sizeof(struct btrfs_item),
2601                                         &leaf->map_token, &leaf->kaddr,
2602                                         &leaf->map_start, &leaf->map_len,
2603                                         KM_USER1);
2604                         }
2605                         ioff = btrfs_item_offset(leaf, item);
2606                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2607                 }
2608
2609                 if (leaf->map_token) {
2610                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2611                         leaf->map_token = NULL;
2612                 }
2613
2614                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2615                               btrfs_item_nr_offset(slot + nr),
2616                               sizeof(struct btrfs_item) *
2617                               (nritems - slot - nr));
2618         }
2619         btrfs_set_header_nritems(leaf, nritems - nr);
2620         nritems -= nr;
2621
2622         /* delete the leaf if we've emptied it */
2623         if (nritems == 0) {
2624                 if (leaf == root->node) {
2625                         btrfs_set_header_level(leaf, 0);
2626                 } else {
2627                         u64 root_gen = btrfs_header_generation(path->nodes[1]);
2628                         clean_tree_block(trans, root, leaf);
2629                         wait_on_tree_block_writeback(root, leaf);
2630                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
2631                         if (wret)
2632                                 ret = wret;
2633                         wret = btrfs_free_extent(trans, root,
2634                                          leaf->start, leaf->len,
2635                                          btrfs_header_owner(path->nodes[1]),
2636                                          root_gen, 0, 0, 1);
2637                         if (wret)
2638                                 ret = wret;
2639                 }
2640         } else {
2641                 int used = leaf_space_used(leaf, 0, nritems);
2642                 if (slot == 0) {
2643                         struct btrfs_disk_key disk_key;
2644
2645                         btrfs_item_key(leaf, &disk_key, 0);
2646                         wret = fixup_low_keys(trans, root, path,
2647                                               &disk_key, 1);
2648                         if (wret)
2649                                 ret = wret;
2650                 }
2651
2652                 /* delete the leaf if it is mostly empty */
2653                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2654                         /* push_leaf_left fixes the path.
2655                          * make sure the path still points to our leaf
2656                          * for possible call to del_ptr below
2657                          */
2658                         slot = path->slots[1];
2659                         extent_buffer_get(leaf);
2660
2661                         wret = push_leaf_left(trans, root, path, 1, 1);
2662                         if (wret < 0 && wret != -ENOSPC)
2663                                 ret = wret;
2664
2665                         if (path->nodes[0] == leaf &&
2666                             btrfs_header_nritems(leaf)) {
2667                                 wret = push_leaf_right(trans, root, path, 1, 1);
2668                                 if (wret < 0 && wret != -ENOSPC)
2669                                         ret = wret;
2670                         }
2671
2672                         if (btrfs_header_nritems(leaf) == 0) {
2673                                 u64 root_gen;
2674                                 u64 bytenr = leaf->start;
2675                                 u32 blocksize = leaf->len;
2676
2677                                 root_gen = btrfs_header_generation(
2678                                                            path->nodes[1]);
2679
2680                                 clean_tree_block(trans, root, leaf);
2681                                 wait_on_tree_block_writeback(root, leaf);
2682
2683                                 wret = del_ptr(trans, root, path, 1, slot);
2684                                 if (wret)
2685                                         ret = wret;
2686
2687                                 free_extent_buffer(leaf);
2688                                 wret = btrfs_free_extent(trans, root, bytenr,
2689                                              blocksize,
2690                                              btrfs_header_owner(path->nodes[1]),
2691                                              root_gen, 0, 0, 1);
2692                                 if (wret)
2693                                         ret = wret;
2694                         } else {
2695                                 btrfs_mark_buffer_dirty(leaf);
2696                                 free_extent_buffer(leaf);
2697                         }
2698                 } else {
2699                         btrfs_mark_buffer_dirty(leaf);
2700                 }
2701         }
2702         return ret;
2703 }
2704
2705 /*
2706  * walk up the tree as far as required to find the previous leaf.
2707  * returns 0 if it found something or 1 if there are no lesser leaves.
2708  * returns < 0 on io errors.
2709  */
2710 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2711 {
2712         u64 bytenr;
2713         int slot;
2714         int level = 1;
2715         struct extent_buffer *c;
2716         struct extent_buffer *next = NULL;
2717
2718         while(level < BTRFS_MAX_LEVEL) {
2719                 if (!path->nodes[level])
2720                         return 1;
2721
2722                 slot = path->slots[level];
2723                 c = path->nodes[level];
2724                 if (slot == 0) {
2725                         level++;
2726                         if (level == BTRFS_MAX_LEVEL)
2727                                 return 1;
2728                         continue;
2729                 }
2730                 slot--;
2731
2732                 bytenr = btrfs_node_blockptr(c, slot);
2733                 if (next)
2734                         free_extent_buffer(next);
2735
2736                 next = read_tree_block(root, bytenr,
2737                                        btrfs_level_size(root, level - 1));
2738                 break;
2739         }
2740         path->slots[level] = slot;
2741         while(1) {
2742                 level--;
2743                 c = path->nodes[level];
2744                 free_extent_buffer(c);
2745                 slot = btrfs_header_nritems(next);
2746                 if (slot != 0)
2747                         slot--;
2748                 path->nodes[level] = next;
2749                 path->slots[level] = slot;
2750                 if (!level)
2751                         break;
2752                 next = read_tree_block(root, btrfs_node_blockptr(next, slot),
2753                                        btrfs_level_size(root, level - 1));
2754         }
2755         return 0;
2756 }
2757
2758 /*
2759  * walk up the tree as far as required to find the next leaf.
2760  * returns 0 if it found something or 1 if there are no greater leaves.
2761  * returns < 0 on io errors.
2762  */
2763 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2764 {
2765         int slot;
2766         int level = 1;
2767         u64 bytenr;
2768         struct extent_buffer *c;
2769         struct extent_buffer *next = NULL;
2770
2771         while(level < BTRFS_MAX_LEVEL) {
2772                 if (!path->nodes[level])
2773                         return 1;
2774
2775                 slot = path->slots[level] + 1;
2776                 c = path->nodes[level];
2777                 if (slot >= btrfs_header_nritems(c)) {
2778                         level++;
2779                         if (level == BTRFS_MAX_LEVEL)
2780                                 return 1;
2781                         continue;
2782                 }
2783
2784                 bytenr = btrfs_node_blockptr(c, slot);
2785                 if (next)
2786                         free_extent_buffer(next);
2787
2788                 if (path->reada)
2789                         reada_for_search(root, path, level, slot, 0);
2790
2791                 next = read_tree_block(root, bytenr,
2792                                        btrfs_level_size(root, level -1));
2793                 break;
2794         }
2795         path->slots[level] = slot;
2796         while(1) {
2797                 level--;
2798                 c = path->nodes[level];
2799                 free_extent_buffer(c);
2800                 path->nodes[level] = next;
2801                 path->slots[level] = 0;
2802                 if (!level)
2803                         break;
2804                 if (path->reada)
2805                         reada_for_search(root, path, level, 0, 0);
2806                 next = read_tree_block(root, btrfs_node_blockptr(next, 0),
2807                                        btrfs_level_size(root, level - 1));
2808         }
2809         return 0;
2810 }
2811
2812 int btrfs_previous_item(struct btrfs_root *root,
2813                         struct btrfs_path *path, u64 min_objectid,
2814                         int type)
2815 {
2816         struct btrfs_key found_key;
2817         struct extent_buffer *leaf;
2818         int ret;
2819
2820         while(1) {
2821                 if (path->slots[0] == 0) {
2822                         ret = btrfs_prev_leaf(root, path);
2823                         if (ret != 0)
2824                                 return ret;
2825                 } else {
2826                         path->slots[0]--;
2827                 }
2828                 leaf = path->nodes[0];
2829                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2830                 if (found_key.type == type)
2831                         return 0;
2832         }
2833         return 1;
2834 }
2835