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                 /* once for the path */
 755                 free_extent_buffer(mid);
 756                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 757                                         root->root_key.objectid,
 758                                         btrfs_header_generation(mid), 0, 0, 1);
 759                 /* once for the root ptr */
 760                 free_extent_buffer(mid);
 761                 return ret;
 762         }
 763         if (btrfs_header_nritems(mid) >
 764             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 765                 return 0;
 766
 767         if (btrfs_header_nritems(mid) < 2)
 768                 err_on_enospc = 1;
 769
 770         left = read_node_slot(root, parent, pslot - 1);
 771         if (left) {
 772                 wret = btrfs_cow_block(trans, root, left,
 773                                        parent, pslot - 1, &left);
 774                 if (wret) {
 775                         ret = wret;
 776                         goto enospc;
 777                 }
 778         }
 779         right = read_node_slot(root, parent, pslot + 1);
 780         if (right) {
 781                 wret = btrfs_cow_block(trans, root, right,
 782                                        parent, pslot + 1, &right);
 783                 if (wret) {
 784                         ret = wret;
 785                         goto enospc;
 786                 }
 787         }
 788
 789         /* first, try to make some room in the middle buffer */
 790         if (left) {
 791                 orig_slot += btrfs_header_nritems(left);
 792                 wret = push_node_left(trans, root, left, mid);
 793                 if (wret < 0)
 794                         ret = wret;
 795                 if (btrfs_header_nritems(mid) < 2)
 796                         err_on_enospc = 1;
 797         }
 798
 799         /*
 800          * then try to empty the right most buffer into the middle
 801          */
 802         if (right) {
 803                 wret = push_node_left(trans, root, mid, right);
 804                 if (wret < 0 && wret != -ENOSPC)
 805                         ret = wret;
 806                 if (btrfs_header_nritems(right) == 0) {
 807                         u64 bytenr = right->start;
 808                         u64 generation = btrfs_header_generation(parent);
 809                         u32 blocksize = right->len;
 810
 811                         clean_tree_block(trans, root, right);
 812                         free_extent_buffer(right);
 813                         right = NULL;
 814                         wret = del_ptr(trans, root, path, level + 1, pslot +
 815                                        1);
 816                         if (wret)
 817                                 ret = wret;
 818                         wret = btrfs_free_extent(trans, root, bytenr,
 819                                                  blocksize,
 820                                                  btrfs_header_owner(parent),
 821                                                  generation, 0, 0, 1);
 822                         if (wret)
 823                                 ret = wret;
 824                 } else {
 825                         struct btrfs_disk_key right_key;
 826                         btrfs_node_key(right, &right_key, 0);
 827                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 828                         btrfs_mark_buffer_dirty(parent);
 829                 }
 830         }
 831         if (btrfs_header_nritems(mid) == 1) {
 832                 /*
 833                  * we're not allowed to leave a node with one item in the
 834                  * tree during a delete.  A deletion from lower in the tree
 835                  * could try to delete the only pointer in this node.
 836                  * So, pull some keys from the left.
 837                  * There has to be a left pointer at this point because
 838                  * otherwise we would have pulled some pointers from the
 839                  * right
 840                  */
 841                 BUG_ON(!left);
 842                 wret = balance_node_right(trans, root, mid, left);
 843                 if (wret < 0) {
 844                         ret = wret;
 845                         goto enospc;
 846                 }
 847                 BUG_ON(wret == 1);
 848         }
 849         if (btrfs_header_nritems(mid) == 0) {
 850                 /* we've managed to empty the middle node, drop it */
 851                 u64 root_gen = btrfs_header_generation(parent);
 852                 u64 bytenr = mid->start;
 853                 u32 blocksize = mid->len;
 854                 clean_tree_block(trans, root, mid);
 855                 free_extent_buffer(mid);
 856                 mid = NULL;
 857                 wret = del_ptr(trans, root, path, level + 1, pslot);
 858                 if (wret)
 859                         ret = wret;
 860                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 861                                          btrfs_header_owner(parent),
 862                                          root_gen, 0, 0, 1);
 863                 if (wret)
 864                         ret = wret;
 865         } else {
 866                 /* update the parent key to reflect our changes */
 867                 struct btrfs_disk_key mid_key;
 868                 btrfs_node_key(mid, &mid_key, 0);
 869                 btrfs_set_node_key(parent, &mid_key, pslot);
 870                 btrfs_mark_buffer_dirty(parent);
 871         }
 872
 873         /* update the path */
 874         if (left) {
 875                 if (btrfs_header_nritems(left) > orig_slot) {
 876                         extent_buffer_get(left);
 877                         path->nodes[level] = left;
 878                         path->slots[level + 1] -= 1;
 879                         path->slots[level] = orig_slot;
 880                         if (mid)
 881                                 free_extent_buffer(mid);
 882                 } else {
 883                         orig_slot -= btrfs_header_nritems(left);
 884                         path->slots[level] = orig_slot;
 885                 }
 886         }
 887         /* double check we haven't messed things up */
 888         check_block(root, path, level);
 889         if (orig_ptr !=
 890             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 891                 BUG();
 892 enospc:
 893         if (right)
 894                 free_extent_buffer(right);
 895         if (left)
 896                 free_extent_buffer(left);
 897         return ret;
 898 }
 899
 900 /* returns zero if the push worked, non-zero otherwise */
 901 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
 902                                           struct btrfs_root *root,
 903                                           struct btrfs_path *path, int level)
 904 {
 905         struct extent_buffer *right = NULL;
 906         struct extent_buffer *mid;
 907         struct extent_buffer *left = NULL;
 908         struct extent_buffer *parent = NULL;
 909         int ret = 0;
 910         int wret;
 911         int pslot;
 912         int orig_slot = path->slots[level];
 913         u64 orig_ptr;
 914
 915         if (level == 0)
 916                 return 1;
 917
 918         mid = path->nodes[level];
 919         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 920         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 921
 922         if (level < BTRFS_MAX_LEVEL - 1)
 923                 parent = path->nodes[level + 1];
 924         pslot = path->slots[level + 1];
 925
 926         if (!parent)
 927                 return 1;
 928
 929         left = read_node_slot(root, parent, pslot - 1);
 930
 931         /* first, try to make some room in the middle buffer */
 932         if (left) {
 933                 u32 left_nr;
 934                 left_nr = btrfs_header_nritems(left);
 935                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 936                         wret = 1;
 937                 } else {
 938                         ret = btrfs_cow_block(trans, root, left, parent,
 939                                               pslot - 1, &left);
 940                         if (ret)
 941                                 wret = 1;
 942                         else {
 943                                 wret = push_node_left(trans, root,
 944                                                       left, mid);
 945                         }
 946                 }
 947                 if (wret < 0)
 948                         ret = wret;
 949                 if (wret == 0) {
 950                         struct btrfs_disk_key disk_key;
 951                         orig_slot += left_nr;
 952                         btrfs_node_key(mid, &disk_key, 0);
 953                         btrfs_set_node_key(parent, &disk_key, pslot);
 954                         btrfs_mark_buffer_dirty(parent);
 955                         if (btrfs_header_nritems(left) > orig_slot) {
 956                                 path->nodes[level] = left;
 957                                 path->slots[level + 1] -= 1;
 958                                 path->slots[level] = orig_slot;
 959                                 free_extent_buffer(mid);
 960                         } else {
 961                                 orig_slot -=
 962                                         btrfs_header_nritems(left);
 963                                 path->slots[level] = orig_slot;
 964                                 free_extent_buffer(left);
 965                         }
 966                         return 0;
 967                 }
 968                 free_extent_buffer(left);
 969         }
 970         right= read_node_slot(root, parent, pslot + 1);
 971
 972         /*
 973          * then try to empty the right most buffer into the middle
 974          */
 975         if (right) {
 976                 u32 right_nr;
 977                 right_nr = btrfs_header_nritems(right);
 978                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 979                         wret = 1;
 980                 } else {
 981                         ret = btrfs_cow_block(trans, root, right,
 982                                               parent, pslot + 1,
 983                                               &right);
 984                         if (ret)
 985                                 wret = 1;
 986                         else {
 987                                 wret = balance_node_right(trans, root,
 988                                                           right, mid);
 989                         }
 990                 }
 991                 if (wret < 0)
 992                         ret = wret;
 993                 if (wret == 0) {
 994                         struct btrfs_disk_key disk_key;
 995
 996                         btrfs_node_key(right, &disk_key, 0);
 997                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
 998                         btrfs_mark_buffer_dirty(parent);
 999
1000                         if (btrfs_header_nritems(mid) <= orig_slot) {
1001                                 path->nodes[level] = right;
1002                                 path->slots[level + 1] += 1;
1003                                 path->slots[level] = orig_slot -
1004                                         btrfs_header_nritems(mid);
1005                                 free_extent_buffer(mid);
1006                         } else {
1007                                 free_extent_buffer(right);
1008                         }
1009                         return 0;
1010                 }
1011                 free_extent_buffer(right);
1012         }
1013         return 1;
1014 }
1015
1016 /*
1017  * readahead one full node of leaves
1018  */
1019 static void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
1020                              int level, int slot, u64 objectid)
1021 {
1022         struct extent_buffer *node;
1023         struct btrfs_disk_key disk_key;
1024         u32 nritems;
1025         u64 search;
1026         u64 lowest_read;
1027         u64 highest_read;
1028         u64 nread = 0;
1029         int direction = path->reada;
1030         struct extent_buffer *eb;
1031         u32 nr;
1032         u32 blocksize;
1033         u32 nscan = 0;
1034
1035         if (level != 1)
1036                 return;
1037
1038         if (!path->nodes[level])
1039                 return;
1040
1041         node = path->nodes[level];
1042         search = btrfs_node_blockptr(node, slot);
1043         blocksize = btrfs_level_size(root, level - 1);
1044         eb = btrfs_find_tree_block(root, search, blocksize);
1045         if (eb) {
1046                 free_extent_buffer(eb);
1047                 return;
1048         }
1049
1050         highest_read = search;
1051         lowest_read = search;
1052
1053         nritems = btrfs_header_nritems(node);
1054         nr = slot;
1055         while(1) {
1056                 if (direction < 0) {
1057                         if (nr == 0)
1058                                 break;
1059                         nr--;
1060                 } else if (direction > 0) {
1061                         nr++;
1062                         if (nr >= nritems)
1063                                 break;
1064                 }
1065                 if (path->reada < 0 && objectid) {
1066                         btrfs_node_key(node, &disk_key, nr);
1067                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1068                                 break;
1069                 }
1070                 search = btrfs_node_blockptr(node, nr);
1071                 if ((search >= lowest_read && search <= highest_read) ||
1072                     (search < lowest_read && lowest_read - search <= 32768) ||
1073                     (search > highest_read && search - highest_read <= 32768)) {
1074                         readahead_tree_block(root, search, blocksize);
1075                         nread += blocksize;
1076                 }
1077                 nscan++;
1078                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
1079                         break;
1080                 if(nread > (1024 * 1024) || nscan > 128)
1081                         break;
1082
1083                 if (search < lowest_read)
1084                         lowest_read = search;
1085                 if (search > highest_read)
1086                         highest_read = search;
1087         }
1088 }
1089 /*
1090  * look for key in the tree.  path is filled in with nodes along the way
1091  * if key is found, we return zero and you can find the item in the leaf
1092  * level of the path (level 0)
1093  *
1094  * If the key isn't found, the path points to the slot where it should
1095  * be inserted, and 1 is returned.  If there are other errors during the
1096  * search a negative error number is returned.
1097  *
1098  * if ins_len > 0, nodes and leaves will be split as we walk down the
1099  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1100  * possible)
1101  */
1102 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1103                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1104                       ins_len, int cow)
1105 {
1106         struct extent_buffer *b;
1107         u64 bytenr;
1108         u64 ptr_gen;
1109         int slot;
1110         int ret;
1111         int level;
1112         int should_reada = p->reada;
1113         u8 lowest_level = 0;
1114
1115         lowest_level = p->lowest_level;
1116         WARN_ON(lowest_level && ins_len);
1117         WARN_ON(p->nodes[0] != NULL);
1118         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
1119 again:
1120         b = root->node;
1121         extent_buffer_get(b);
1122         while (b) {
1123                 level = btrfs_header_level(b);
1124                 if (cow) {
1125                         int wret;
1126                         wret = btrfs_cow_block(trans, root, b,
1127                                                p->nodes[level + 1],
1128                                                p->slots[level + 1],
1129                                                &b);
1130                         if (wret) {
1131                                 free_extent_buffer(b);
1132                                 return wret;
1133                         }
1134                 }
1135                 BUG_ON(!cow && ins_len);
1136                 if (level != btrfs_header_level(b))
1137                         WARN_ON(1);
1138                 level = btrfs_header_level(b);
1139                 p->nodes[level] = b;
1140                 ret = check_block(root, p, level);
1141                 if (ret)
1142                         return -1;
1143                 ret = bin_search(b, key, level, &slot);
1144                 if (level != 0) {
1145                         if (ret && slot > 0)
1146                                 slot -= 1;
1147                         p->slots[level] = slot;
1148                         if (ins_len > 0 && btrfs_header_nritems(b) >=
1149                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1150                                 int sret = split_node(trans, root, p, level);
1151                                 BUG_ON(sret > 0);
1152                                 if (sret)
1153                                         return sret;
1154                                 b = p->nodes[level];
1155                                 slot = p->slots[level];
1156                         } else if (ins_len < 0) {
1157                                 int sret = balance_level(trans, root, p,
1158                                                          level);
1159                                 if (sret)
1160                                         return sret;
1161                                 b = p->nodes[level];
1162                                 if (!b) {
1163                                         btrfs_release_path(NULL, p);
1164                                         goto again;
1165                                 }
1166                                 slot = p->slots[level];
1167                                 BUG_ON(btrfs_header_nritems(b) == 1);
1168                         }
1169                         /* this is only true while dropping a snapshot */
1170                         if (level == lowest_level)
1171                                 break;
1172                         bytenr = btrfs_node_blockptr(b, slot);
1173                         ptr_gen = btrfs_node_ptr_generation(b, slot);
1174                         if (should_reada)
1175                                 reada_for_search(root, p, level, slot,
1176                                                  key->objectid);
1177                         b = read_tree_block(root, bytenr,
1178                                             btrfs_level_size(root, level - 1));
1179                         if (ptr_gen != btrfs_header_generation(b)) {
1180                                 printk("block %llu bad gen wanted %llu "
1181                                        "found %llu\n",
1182                                 (unsigned long long)b->start,
1183                                 (unsigned long long)ptr_gen,
1184                                 (unsigned long long)btrfs_header_generation(b));
1185                         }
1186                 } else {
1187                         p->slots[level] = slot;
1188                         if (ins_len > 0 && btrfs_leaf_free_space(root, b) <
1189                             sizeof(struct btrfs_item) + ins_len) {
1190                                 int sret = split_leaf(trans, root, key,
1191                                                       p, ins_len, ret == 0);
1192                                 BUG_ON(sret > 0);
1193                                 if (sret)
1194                                         return sret;
1195                         }
1196                         return ret;
1197                 }
1198         }
1199         return 1;
1200 }
1201
1202 /*
1203  * adjust the pointers going up the tree, starting at level
1204  * making sure the right key of each node is points to 'key'.
1205  * This is used after shifting pointers to the left, so it stops
1206  * fixing up pointers when a given leaf/node is not in slot 0 of the
1207  * higher levels
1208  *
1209  * If this fails to write a tree block, it returns -1, but continues
1210  * fixing up the blocks in ram so the tree is consistent.
1211  */
1212 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1213                           struct btrfs_root *root, struct btrfs_path *path,
1214                           struct btrfs_disk_key *key, int level)
1215 {
1216         int i;
1217         int ret = 0;
1218         struct extent_buffer *t;
1219
1220         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1221                 int tslot = path->slots[i];
1222                 if (!path->nodes[i])
1223                         break;
1224                 t = path->nodes[i];
1225                 btrfs_set_node_key(t, key, tslot);
1226                 btrfs_mark_buffer_dirty(path->nodes[i]);
1227                 if (tslot != 0)
1228                         break;
1229         }
1230         return ret;
1231 }
1232
1233 /*
1234  * try to push data from one node into the next node left in the
1235  * tree.
1236  *
1237  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1238  * error, and > 0 if there was no room in the left hand block.
1239  */
1240 static int push_node_left(struct btrfs_trans_handle *trans,
1241                           struct btrfs_root *root, struct extent_buffer *dst,
1242                           struct extent_buffer *src)
1243 {
1244         int push_items = 0;
1245         int src_nritems;
1246         int dst_nritems;
1247         int ret = 0;
1248
1249         src_nritems = btrfs_header_nritems(src);
1250         dst_nritems = btrfs_header_nritems(dst);
1251         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1252         WARN_ON(btrfs_header_generation(src) != trans->transid);
1253         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1254
1255         if (push_items <= 0) {
1256                 return 1;
1257         }
1258
1259         if (src_nritems < push_items)
1260                 push_items = src_nritems;
1261
1262         copy_extent_buffer(dst, src,
1263                            btrfs_node_key_ptr_offset(dst_nritems),
1264                            btrfs_node_key_ptr_offset(0),
1265                            push_items * sizeof(struct btrfs_key_ptr));
1266
1267         if (push_items < src_nritems) {
1268                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1269                                       btrfs_node_key_ptr_offset(push_items),
1270                                       (src_nritems - push_items) *
1271                                       sizeof(struct btrfs_key_ptr));
1272         }
1273         btrfs_set_header_nritems(src, src_nritems - push_items);
1274         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1275         btrfs_mark_buffer_dirty(src);
1276         btrfs_mark_buffer_dirty(dst);
1277         return ret;
1278 }
1279
1280 /*
1281  * try to push data from one node into the next node right in the
1282  * tree.
1283  *
1284  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1285  * error, and > 0 if there was no room in the right hand block.
1286  *
1287  * this will  only push up to 1/2 the contents of the left node over
1288  */
1289 static int balance_node_right(struct btrfs_trans_handle *trans,
1290                               struct btrfs_root *root,
1291                               struct extent_buffer *dst,
1292                               struct extent_buffer *src)
1293 {
1294         int push_items = 0;
1295         int max_push;
1296         int src_nritems;
1297         int dst_nritems;
1298         int ret = 0;
1299
1300         WARN_ON(btrfs_header_generation(src) != trans->transid);
1301         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1302
1303         src_nritems = btrfs_header_nritems(src);
1304         dst_nritems = btrfs_header_nritems(dst);
1305         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1306         if (push_items <= 0)
1307                 return 1;
1308
1309         max_push = src_nritems / 2 + 1;
1310         /* don't try to empty the node */
1311         if (max_push >= src_nritems)
1312                 return 1;
1313
1314         if (max_push < push_items)
1315                 push_items = max_push;
1316
1317         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1318                                       btrfs_node_key_ptr_offset(0),
1319                                       (dst_nritems) *
1320                                       sizeof(struct btrfs_key_ptr));
1321
1322         copy_extent_buffer(dst, src,
1323                            btrfs_node_key_ptr_offset(0),
1324                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1325                            push_items * sizeof(struct btrfs_key_ptr));
1326
1327         btrfs_set_header_nritems(src, src_nritems - push_items);
1328         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1329
1330         btrfs_mark_buffer_dirty(src);
1331         btrfs_mark_buffer_dirty(dst);
1332         return ret;
1333 }
1334
1335 /*
1336  * helper function to insert a new root level in the tree.
1337  * A new node is allocated, and a single item is inserted to
1338  * point to the existing root
1339  *
1340  * returns zero on success or < 0 on failure.
1341  */
1342 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1343                            struct btrfs_root *root,
1344                            struct btrfs_path *path, int level)
1345 {
1346         u64 root_gen;
1347         u64 lower_gen;
1348         struct extent_buffer *lower;
1349         struct extent_buffer *c;
1350         struct btrfs_disk_key lower_key;
1351
1352         BUG_ON(path->nodes[level]);
1353         BUG_ON(path->nodes[level-1] != root->node);
1354
1355         if (root->ref_cows)
1356                 root_gen = trans->transid;
1357         else
1358                 root_gen = 0;
1359
1360         lower = path->nodes[level-1];
1361         if (level == 1)
1362                 btrfs_item_key(lower, &lower_key, 0);
1363         else
1364                 btrfs_node_key(lower, &lower_key, 0);
1365
1366         c = __btrfs_alloc_free_block(trans, root, root->nodesize,
1367                                    root->root_key.objectid,
1368                                    root_gen, lower_key.objectid, level,
1369                                    root->node->start, 0);
1370         if (IS_ERR(c))
1371                 return PTR_ERR(c);
1372         memset_extent_buffer(c, 0, 0, root->nodesize);
1373         btrfs_set_header_nritems(c, 1);
1374         btrfs_set_header_level(c, level);
1375         btrfs_set_header_bytenr(c, c->start);
1376         btrfs_set_header_generation(c, trans->transid);
1377         btrfs_set_header_owner(c, root->root_key.objectid);
1378
1379         write_extent_buffer(c, root->fs_info->fsid,
1380                             (unsigned long)btrfs_header_fsid(c),
1381                             BTRFS_FSID_SIZE);
1382
1383         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1384                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
1385                             BTRFS_UUID_SIZE);
1386
1387         btrfs_set_node_key(c, &lower_key, 0);
1388         btrfs_set_node_blockptr(c, 0, lower->start);
1389         lower_gen = btrfs_header_generation(lower);
1390         WARN_ON(lower_gen == 0);
1391
1392         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1393
1394         btrfs_mark_buffer_dirty(c);
1395
1396         /* the super has an extra ref to root->node */
1397         free_extent_buffer(root->node);
1398         root->node = c;
1399         add_root_to_dirty_list(root);
1400         extent_buffer_get(c);
1401         path->nodes[level] = c;
1402         path->slots[level] = 0;
1403
1404         if (root->ref_cows && lower_gen != trans->transid) {
1405                 struct btrfs_path *back_path = btrfs_alloc_path();
1406                 int ret;
1407                 ret = btrfs_insert_extent_backref(trans,
1408                                                   root->fs_info->extent_root,
1409                                                   path, lower->start,
1410                                                   root->root_key.objectid,
1411                                                   trans->transid, 0, 0);
1412                 BUG_ON(ret);
1413                 btrfs_free_path(back_path);
1414         }
1415         return 0;
1416 }
1417
1418 /*
1419  * worker function to insert a single pointer in a node.
1420  * the node should have enough room for the pointer already
1421  *
1422  * slot and level indicate where you want the key to go, and
1423  * blocknr is the block the key points to.
1424  *
1425  * returns zero on success and < 0 on any error
1426  */
1427 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1428                       *root, struct btrfs_path *path, struct btrfs_disk_key
1429                       *key, u64 bytenr, int slot, int level)
1430 {
1431         struct extent_buffer *lower;
1432         int nritems;
1433
1434         BUG_ON(!path->nodes[level]);
1435         lower = path->nodes[level];
1436         nritems = btrfs_header_nritems(lower);
1437         if (slot > nritems)
1438                 BUG();
1439         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1440                 BUG();
1441         if (slot != nritems) {
1442                 memmove_extent_buffer(lower,
1443                               btrfs_node_key_ptr_offset(slot + 1),
1444                               btrfs_node_key_ptr_offset(slot),
1445                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1446         }
1447         btrfs_set_node_key(lower, key, slot);
1448         btrfs_set_node_blockptr(lower, slot, bytenr);
1449         WARN_ON(trans->transid == 0);
1450         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1451         btrfs_set_header_nritems(lower, nritems + 1);
1452         btrfs_mark_buffer_dirty(lower);
1453         return 0;
1454 }
1455
1456 /*
1457  * split the node at the specified level in path in two.
1458  * The path is corrected to point to the appropriate node after the split
1459  *
1460  * Before splitting this tries to make some room in the node by pushing
1461  * left and right, if either one works, it returns right away.
1462  *
1463  * returns 0 on success and < 0 on failure
1464  */
1465 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1466                       *root, struct btrfs_path *path, int level)
1467 {
1468         u64 root_gen;
1469         struct extent_buffer *c;
1470         struct extent_buffer *split;
1471         struct btrfs_disk_key disk_key;
1472         int mid;
1473         int ret;
1474         int wret;
1475         u32 c_nritems;
1476
1477         c = path->nodes[level];
1478         WARN_ON(btrfs_header_generation(c) != trans->transid);
1479         if (c == root->node) {
1480                 /* trying to split the root, lets make a new one */
1481                 ret = insert_new_root(trans, root, path, level + 1);
1482                 if (ret)
1483                         return ret;
1484         } else {
1485                 ret = push_nodes_for_insert(trans, root, path, level);
1486                 c = path->nodes[level];
1487                 if (!ret && btrfs_header_nritems(c) <
1488                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
1489                         return 0;
1490                 if (ret < 0)
1491                         return ret;
1492         }
1493
1494         c_nritems = btrfs_header_nritems(c);
1495         if (root->ref_cows)
1496                 root_gen = trans->transid;
1497         else
1498                 root_gen = 0;
1499
1500         btrfs_node_key(c, &disk_key, 0);
1501         split = __btrfs_alloc_free_block(trans, root, root->nodesize,
1502                                          root->root_key.objectid,
1503                                          root_gen,
1504                                          btrfs_disk_key_objectid(&disk_key),
1505                                          level, c->start, 0);
1506         if (IS_ERR(split))
1507                 return PTR_ERR(split);
1508
1509         btrfs_set_header_flags(split, btrfs_header_flags(c));
1510         btrfs_set_header_level(split, btrfs_header_level(c));
1511         btrfs_set_header_bytenr(split, split->start);
1512         btrfs_set_header_generation(split, trans->transid);
1513         btrfs_set_header_owner(split, root->root_key.objectid);
1514         btrfs_set_header_flags(split, 0);
1515         write_extent_buffer(split, root->fs_info->fsid,
1516                             (unsigned long)btrfs_header_fsid(split),
1517                             BTRFS_FSID_SIZE);
1518         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1519                             (unsigned long)btrfs_header_chunk_tree_uuid(split),
1520                             BTRFS_UUID_SIZE);
1521
1522         mid = (c_nritems + 1) / 2;
1523
1524         copy_extent_buffer(split, c,
1525                            btrfs_node_key_ptr_offset(0),
1526                            btrfs_node_key_ptr_offset(mid),
1527                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1528         btrfs_set_header_nritems(split, c_nritems - mid);
1529         btrfs_set_header_nritems(c, mid);
1530         ret = 0;
1531
1532         btrfs_mark_buffer_dirty(c);
1533         btrfs_mark_buffer_dirty(split);
1534
1535         btrfs_node_key(split, &disk_key, 0);
1536         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1537                           path->slots[level + 1] + 1,
1538                           level + 1);
1539         if (wret)
1540                 ret = wret;
1541
1542         if (path->slots[level] >= mid) {
1543                 path->slots[level] -= mid;
1544                 free_extent_buffer(c);
1545                 path->nodes[level] = split;
1546                 path->slots[level + 1] += 1;
1547         } else {
1548                 free_extent_buffer(split);
1549         }
1550         return ret;
1551 }
1552
1553 /*
1554  * how many bytes are required to store the items in a leaf.  start
1555  * and nr indicate which items in the leaf to check.  This totals up the
1556  * space used both by the item structs and the item data
1557  */
1558 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1559 {
1560         int data_len;
1561         int nritems = btrfs_header_nritems(l);
1562         int end = min(nritems, start + nr) - 1;
1563
1564         if (!nr)
1565                 return 0;
1566         data_len = btrfs_item_end_nr(l, start);
1567         data_len = data_len - btrfs_item_offset_nr(l, end);
1568         data_len += sizeof(struct btrfs_item) * nr;
1569         WARN_ON(data_len < 0);
1570         return data_len;
1571 }
1572
1573 /*
1574  * The space between the end of the leaf items and
1575  * the start of the leaf data.  IOW, how much room
1576  * the leaf has left for both items and data
1577  */
1578 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1579 {
1580         int nritems = btrfs_header_nritems(leaf);
1581         int ret;
1582         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1583         if (ret < 0) {
1584                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1585                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1586                        leaf_space_used(leaf, 0, nritems), nritems);
1587         }
1588         return ret;
1589 }
1590
1591 /*
1592  * push some data in the path leaf to the right, trying to free up at
1593  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1594  *
1595  * returns 1 if the push failed because the other node didn't have enough
1596  * room, 0 if everything worked out and < 0 if there were major errors.
1597  */
1598 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1599                            *root, struct btrfs_path *path, int data_size,
1600                            int empty)
1601 {
1602         struct extent_buffer *left = path->nodes[0];
1603         struct extent_buffer *right;
1604         struct extent_buffer *upper;
1605         struct btrfs_disk_key disk_key;
1606         int slot;
1607         u32 i;
1608         int free_space;
1609         int push_space = 0;
1610         int push_items = 0;
1611         struct btrfs_item *item;
1612         u32 left_nritems;
1613         u32 nr;
1614         u32 right_nritems;
1615         u32 data_end;
1616         u32 this_item_size;
1617         int ret;
1618
1619         slot = path->slots[1];
1620         if (!path->nodes[1]) {
1621                 return 1;
1622         }
1623         upper = path->nodes[1];
1624         if (slot >= btrfs_header_nritems(upper) - 1)
1625                 return 1;
1626
1627         right = read_tree_block(root, btrfs_node_blockptr(upper, slot + 1),
1628                                 root->leafsize);
1629         free_space = btrfs_leaf_free_space(root, right);
1630         if (free_space < data_size + sizeof(struct btrfs_item)) {
1631                 free_extent_buffer(right);
1632                 return 1;
1633         }
1634
1635         /* cow and double check */
1636         ret = btrfs_cow_block(trans, root, right, upper,
1637                               slot + 1, &right);
1638         if (ret) {
1639                 free_extent_buffer(right);
1640                 return 1;
1641         }
1642         free_space = btrfs_leaf_free_space(root, right);
1643         if (free_space < data_size + sizeof(struct btrfs_item)) {
1644                 free_extent_buffer(right);
1645                 return 1;
1646         }
1647
1648         left_nritems = btrfs_header_nritems(left);
1649         if (left_nritems == 0) {
1650                 free_extent_buffer(right);
1651                 return 1;
1652         }
1653
1654         if (empty)
1655                 nr = 0;
1656         else
1657                 nr = 1;
1658
1659         i = left_nritems - 1;
1660         while (i >= nr) {
1661                 item = btrfs_item_nr(left, i);
1662
1663                 if (path->slots[0] == i)
1664                         push_space += data_size + sizeof(*item);
1665
1666                 if (!left->map_token) {
1667                         map_extent_buffer(left, (unsigned long)item,
1668                                         sizeof(struct btrfs_item),
1669                                         &left->map_token, &left->kaddr,
1670                                         &left->map_start, &left->map_len,
1671                                         KM_USER1);
1672                 }
1673
1674                 this_item_size = btrfs_item_size(left, item);
1675                 if (this_item_size + sizeof(*item) + push_space > free_space)
1676                         break;
1677                 push_items++;
1678                 push_space += this_item_size + sizeof(*item);
1679                 if (i == 0)
1680                         break;
1681                 i--;
1682         }
1683         if (left->map_token) {
1684                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1685                 left->map_token = NULL;
1686         }
1687
1688         if (push_items == 0) {
1689                 free_extent_buffer(right);
1690                 return 1;
1691         }
1692
1693         if (!empty && push_items == left_nritems)
1694                 WARN_ON(1);
1695
1696         /* push left to right */
1697         right_nritems = btrfs_header_nritems(right);
1698
1699         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1700         push_space -= leaf_data_end(root, left);
1701
1702         /* make room in the right data area */
1703         data_end = leaf_data_end(root, right);
1704         memmove_extent_buffer(right,
1705                               btrfs_leaf_data(right) + data_end - push_space,
1706                               btrfs_leaf_data(right) + data_end,
1707                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1708
1709         /* copy from the left data area */
1710         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1711                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1712                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1713                      push_space);
1714
1715         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1716                               btrfs_item_nr_offset(0),
1717                               right_nritems * sizeof(struct btrfs_item));
1718
1719         /* copy the items from left to right */
1720         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1721                    btrfs_item_nr_offset(left_nritems - push_items),
1722                    push_items * sizeof(struct btrfs_item));
1723
1724         /* update the item pointers */
1725         right_nritems += push_items;
1726         btrfs_set_header_nritems(right, right_nritems);
1727         push_space = BTRFS_LEAF_DATA_SIZE(root);
1728         for (i = 0; i < right_nritems; i++) {
1729                 item = btrfs_item_nr(right, i);
1730                 if (!right->map_token) {
1731                         map_extent_buffer(right, (unsigned long)item,
1732                                         sizeof(struct btrfs_item),
1733                                         &right->map_token, &right->kaddr,
1734                                         &right->map_start, &right->map_len,
1735                                         KM_USER1);
1736                 }
1737                 push_space -= btrfs_item_size(right, item);
1738                 btrfs_set_item_offset(right, item, push_space);
1739         }
1740
1741         if (right->map_token) {
1742                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1743                 right->map_token = NULL;
1744         }
1745         left_nritems -= push_items;
1746         btrfs_set_header_nritems(left, left_nritems);
1747
1748         if (left_nritems)
1749                 btrfs_mark_buffer_dirty(left);
1750         btrfs_mark_buffer_dirty(right);
1751
1752         btrfs_item_key(right, &disk_key, 0);
1753         btrfs_set_node_key(upper, &disk_key, slot + 1);
1754         btrfs_mark_buffer_dirty(upper);
1755
1756         /* then fixup the leaf pointer in the path */
1757         if (path->slots[0] >= left_nritems) {
1758                 path->slots[0] -= left_nritems;
1759                 free_extent_buffer(path->nodes[0]);
1760                 path->nodes[0] = right;
1761                 path->slots[1] += 1;
1762         } else {
1763                 free_extent_buffer(right);
1764         }
1765         return 0;
1766 }
1767 /*
1768  * push some data in the path leaf to the left, trying to free up at
1769  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1770  */
1771 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1772                           *root, struct btrfs_path *path, int data_size,
1773                           int empty)
1774 {
1775         struct btrfs_disk_key disk_key;
1776         struct extent_buffer *right = path->nodes[0];
1777         struct extent_buffer *left;
1778         int slot;
1779         int i;
1780         int free_space;
1781         int push_space = 0;
1782         int push_items = 0;
1783         struct btrfs_item *item;
1784         u32 old_left_nritems;
1785         u32 right_nritems;
1786         u32 nr;
1787         int ret = 0;
1788         int wret;
1789         u32 this_item_size;
1790         u32 old_left_item_size;
1791
1792         slot = path->slots[1];
1793         if (slot == 0)
1794                 return 1;
1795         if (!path->nodes[1])
1796                 return 1;
1797
1798         right_nritems = btrfs_header_nritems(right);
1799         if (right_nritems == 0) {
1800                 return 1;
1801         }
1802
1803         left = read_tree_block(root, btrfs_node_blockptr(path->nodes[1],
1804                                slot - 1), root->leafsize);
1805         free_space = btrfs_leaf_free_space(root, left);
1806         if (free_space < data_size + sizeof(struct btrfs_item)) {
1807                 free_extent_buffer(left);
1808                 return 1;
1809         }
1810
1811         /* cow and double check */
1812         ret = btrfs_cow_block(trans, root, left,
1813                               path->nodes[1], slot - 1, &left);
1814         if (ret) {
1815                 /* we hit -ENOSPC, but it isn't fatal here */
1816                 free_extent_buffer(left);
1817                 return 1;
1818         }
1819
1820         free_space = btrfs_leaf_free_space(root, left);
1821         if (free_space < data_size + sizeof(struct btrfs_item)) {
1822                 free_extent_buffer(left);
1823                 return 1;
1824         }
1825
1826         if (empty)
1827                 nr = right_nritems;
1828         else
1829                 nr = right_nritems - 1;
1830
1831         for (i = 0; i < nr; i++) {
1832                 item = btrfs_item_nr(right, i);
1833                 if (!right->map_token) {
1834                         map_extent_buffer(right, (unsigned long)item,
1835                                         sizeof(struct btrfs_item),
1836                                         &right->map_token, &right->kaddr,
1837                                         &right->map_start, &right->map_len,
1838                                         KM_USER1);
1839                 }
1840
1841                 if (path->slots[0] == i)
1842                         push_space += data_size + sizeof(*item);
1843
1844                 this_item_size = btrfs_item_size(right, item);
1845                 if (this_item_size + sizeof(*item) + push_space > free_space)
1846                         break;
1847
1848                 push_items++;
1849                 push_space += this_item_size + sizeof(*item);
1850         }
1851
1852         if (right->map_token) {
1853                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1854                 right->map_token = NULL;
1855         }
1856
1857         if (push_items == 0) {
1858                 free_extent_buffer(left);
1859                 return 1;
1860         }
1861         if (!empty && push_items == btrfs_header_nritems(right))
1862                 WARN_ON(1);
1863
1864         /* push data from right to left */
1865         copy_extent_buffer(left, right,
1866                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1867                            btrfs_item_nr_offset(0),
1868                            push_items * sizeof(struct btrfs_item));
1869
1870         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1871                      btrfs_item_offset_nr(right, push_items -1);
1872
1873         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1874                      leaf_data_end(root, left) - push_space,
1875                      btrfs_leaf_data(right) +
1876                      btrfs_item_offset_nr(right, push_items - 1),
1877                      push_space);
1878         old_left_nritems = btrfs_header_nritems(left);
1879         BUG_ON(old_left_nritems < 0);
1880
1881         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1882         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1883                 u32 ioff;
1884
1885                 item = btrfs_item_nr(left, i);
1886                 if (!left->map_token) {
1887                         map_extent_buffer(left, (unsigned long)item,
1888                                         sizeof(struct btrfs_item),
1889                                         &left->map_token, &left->kaddr,
1890                                         &left->map_start, &left->map_len,
1891                                         KM_USER1);
1892                 }
1893
1894                 ioff = btrfs_item_offset(left, item);
1895                 btrfs_set_item_offset(left, item,
1896                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1897         }
1898         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1899         if (left->map_token) {
1900                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1901                 left->map_token = NULL;
1902         }
1903
1904         /* fixup right node */
1905         if (push_items > right_nritems) {
1906                 printk("push items %d nr %u\n", push_items, right_nritems);
1907                 WARN_ON(1);
1908         }
1909
1910         if (push_items < right_nritems) {
1911                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1912                                                   leaf_data_end(root, right);
1913                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
1914                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1915                                       btrfs_leaf_data(right) +
1916                                       leaf_data_end(root, right), push_space);
1917
1918                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1919                               btrfs_item_nr_offset(push_items),
1920                              (btrfs_header_nritems(right) - push_items) *
1921                              sizeof(struct btrfs_item));
1922         }
1923         right_nritems -= push_items;
1924         btrfs_set_header_nritems(right, right_nritems);
1925         push_space = BTRFS_LEAF_DATA_SIZE(root);
1926         for (i = 0; i < right_nritems; i++) {
1927                 item = btrfs_item_nr(right, i);
1928
1929                 if (!right->map_token) {
1930                         map_extent_buffer(right, (unsigned long)item,
1931                                         sizeof(struct btrfs_item),
1932                                         &right->map_token, &right->kaddr,
1933                                         &right->map_start, &right->map_len,
1934                                         KM_USER1);
1935                 }
1936
1937                 push_space = push_space - btrfs_item_size(right, item);
1938                 btrfs_set_item_offset(right, item, push_space);
1939         }
1940         if (right->map_token) {
1941                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1942                 right->map_token = NULL;
1943         }
1944
1945         btrfs_mark_buffer_dirty(left);
1946         if (right_nritems)
1947                 btrfs_mark_buffer_dirty(right);
1948
1949         btrfs_item_key(right, &disk_key, 0);
1950         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1951         if (wret)
1952                 ret = wret;
1953
1954         /* then fixup the leaf pointer in the path */
1955         if (path->slots[0] < push_items) {
1956                 path->slots[0] += old_left_nritems;
1957                 free_extent_buffer(path->nodes[0]);
1958                 path->nodes[0] = left;
1959                 path->slots[1] -= 1;
1960         } else {
1961                 free_extent_buffer(left);
1962                 path->slots[0] -= push_items;
1963         }
1964         BUG_ON(path->slots[0] < 0);
1965         return ret;
1966 }
1967
1968 /*
1969  * split the path's leaf in two, making sure there is at least data_size
1970  * available for the resulting leaf level of the path.
1971  *
1972  * returns 0 if all went well and < 0 on failure.
1973  */
1974 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
1975                       *root, struct btrfs_key *ins_key,
1976                       struct btrfs_path *path, int data_size, int extend)
1977 {
1978         u64 root_gen;
1979         struct extent_buffer *l;
1980         u32 nritems;
1981         int mid;
1982         int slot;
1983         struct extent_buffer *right;
1984         int space_needed = data_size + sizeof(struct btrfs_item);
1985         int data_copy_size;
1986         int rt_data_off;
1987         int i;
1988         int ret = 0;
1989         int wret;
1990         int double_split;
1991         int num_doubles = 0;
1992         struct btrfs_disk_key disk_key;
1993
1994         if (extend)
1995                 space_needed = data_size;
1996
1997         if (root->ref_cows)
1998                 root_gen = trans->transid;
1999         else
2000                 root_gen = 0;
2001
2002         /* first try to make some room by pushing left and right */
2003         if (ins_key->type != BTRFS_DIR_ITEM_KEY) {
2004                 wret = push_leaf_right(trans, root, path, data_size, 0);
2005                 if (wret < 0) {
2006                         return wret;
2007                 }
2008                 if (wret) {
2009                         wret = push_leaf_left(trans, root, path, data_size, 0);
2010                         if (wret < 0)
2011                                 return wret;
2012                 }
2013                 l = path->nodes[0];
2014
2015                 /* did the pushes work? */
2016                 if (btrfs_leaf_free_space(root, l) >= space_needed)
2017                         return 0;
2018         }
2019
2020         if (!path->nodes[1]) {
2021                 ret = insert_new_root(trans, root, path, 1);
2022                 if (ret)
2023                         return ret;
2024         }
2025 again:
2026         double_split = 0;
2027         l = path->nodes[0];
2028         slot = path->slots[0];
2029         nritems = btrfs_header_nritems(l);
2030         mid = (nritems + 1)/ 2;
2031
2032         btrfs_item_key(l, &disk_key, 0);
2033
2034         right = __btrfs_alloc_free_block(trans, root, root->leafsize,
2035                                          root->root_key.objectid,
2036                                          root_gen, disk_key.objectid, 0,
2037                                          l->start, 0);
2038         if (IS_ERR(right)) {
2039                 BUG_ON(1);
2040                 return PTR_ERR(right);
2041         }
2042
2043         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2044         btrfs_set_header_bytenr(right, right->start);
2045         btrfs_set_header_generation(right, trans->transid);
2046         btrfs_set_header_owner(right, root->root_key.objectid);
2047         btrfs_set_header_level(right, 0);
2048         write_extent_buffer(right, root->fs_info->fsid,
2049                             (unsigned long)btrfs_header_fsid(right),
2050                             BTRFS_FSID_SIZE);
2051
2052         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2053                             (unsigned long)btrfs_header_chunk_tree_uuid(right),
2054                             BTRFS_UUID_SIZE);
2055         if (mid <= slot) {
2056                 if (nritems == 1 ||
2057                     leaf_space_used(l, mid, nritems - mid) + space_needed >
2058                         BTRFS_LEAF_DATA_SIZE(root)) {
2059                         if (slot >= nritems) {
2060                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2061                                 btrfs_set_header_nritems(right, 0);
2062                                 wret = insert_ptr(trans, root, path,
2063                                                   &disk_key, right->start,
2064                                                   path->slots[1] + 1, 1);
2065                                 if (wret)
2066                                         ret = wret;
2067                                 free_extent_buffer(path->nodes[0]);
2068                                 path->nodes[0] = right;
2069                                 path->slots[0] = 0;
2070                                 path->slots[1] += 1;
2071                                 btrfs_mark_buffer_dirty(right);
2072                                 return ret;
2073                         }
2074                         mid = slot;
2075                         if (mid != nritems &&
2076                             leaf_space_used(l, mid, nritems - mid) +
2077                             space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2078                                 double_split = 1;
2079                         }
2080                 }
2081         } else {
2082                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
2083                         BTRFS_LEAF_DATA_SIZE(root)) {
2084                         if (!extend && slot == 0) {
2085                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2086                                 btrfs_set_header_nritems(right, 0);
2087                                 wret = insert_ptr(trans, root, path,
2088                                                   &disk_key,
2089                                                   right->start,
2090                                                   path->slots[1], 1);
2091                                 if (wret)
2092                                         ret = wret;
2093                                 free_extent_buffer(path->nodes[0]);
2094                                 path->nodes[0] = right;
2095                                 path->slots[0] = 0;
2096                                 if (path->slots[1] == 0) {
2097                                         wret = fixup_low_keys(trans, root,
2098                                                    path, &disk_key, 1);
2099                                         if (wret)
2100                                                 ret = wret;
2101                                 }
2102                                 btrfs_mark_buffer_dirty(right);
2103                                 return ret;
2104                         } else if (extend && slot == 0) {
2105                                 mid = 1;
2106                         } else {
2107                                 mid = slot;
2108                                 if (mid != nritems &&
2109                                     leaf_space_used(l, mid, nritems - mid) +
2110                                     space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2111                                         double_split = 1;
2112                                 }
2113                         }
2114                 }
2115         }
2116         nritems = nritems - mid;
2117         btrfs_set_header_nritems(right, nritems);
2118         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2119
2120         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2121                            btrfs_item_nr_offset(mid),
2122                            nritems * sizeof(struct btrfs_item));
2123
2124         copy_extent_buffer(right, l,
2125                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2126                      data_copy_size, btrfs_leaf_data(l) +
2127                      leaf_data_end(root, l), data_copy_size);
2128
2129         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2130                       btrfs_item_end_nr(l, mid);
2131
2132         for (i = 0; i < nritems; i++) {
2133                 struct btrfs_item *item = btrfs_item_nr(right, i);
2134                 u32 ioff;
2135
2136                 if (!right->map_token) {
2137                         map_extent_buffer(right, (unsigned long)item,
2138                                         sizeof(struct btrfs_item),
2139                                         &right->map_token, &right->kaddr,
2140                                         &right->map_start, &right->map_len,
2141                                         KM_USER1);
2142                 }
2143
2144                 ioff = btrfs_item_offset(right, item);
2145                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2146         }
2147
2148         if (right->map_token) {
2149                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2150                 right->map_token = NULL;
2151         }
2152
2153         btrfs_set_header_nritems(l, mid);
2154         ret = 0;
2155         btrfs_item_key(right, &disk_key, 0);
2156         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2157                           path->slots[1] + 1, 1);
2158         if (wret)
2159                 ret = wret;
2160
2161         btrfs_mark_buffer_dirty(right);
2162         btrfs_mark_buffer_dirty(l);
2163         BUG_ON(path->slots[0] != slot);
2164
2165         if (mid <= slot) {
2166                 free_extent_buffer(path->nodes[0]);
2167                 path->nodes[0] = right;
2168                 path->slots[0] -= mid;
2169                 path->slots[1] += 1;
2170         } else
2171                 free_extent_buffer(right);
2172
2173         BUG_ON(path->slots[0] < 0);
2174
2175         if (double_split) {
2176                 BUG_ON(num_doubles != 0);
2177                 num_doubles++;
2178                 goto again;
2179         }
2180         return ret;
2181 }
2182
2183 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2184                         struct btrfs_root *root,
2185                         struct btrfs_path *path,
2186                         u32 new_size, int from_end)
2187 {
2188         int ret = 0;
2189         int slot;
2190         int slot_orig;
2191         struct extent_buffer *leaf;
2192         struct btrfs_item *item;
2193         u32 nritems;
2194         unsigned int data_end;
2195         unsigned int old_data_start;
2196         unsigned int old_size;
2197         unsigned int size_diff;
2198         int i;
2199
2200         slot_orig = path->slots[0];
2201         leaf = path->nodes[0];
2202         slot = path->slots[0];
2203
2204         old_size = btrfs_item_size_nr(leaf, slot);
2205         if (old_size == new_size)
2206                 return 0;
2207
2208         nritems = btrfs_header_nritems(leaf);
2209         data_end = leaf_data_end(root, leaf);
2210
2211         old_data_start = btrfs_item_offset_nr(leaf, slot);
2212
2213         size_diff = old_size - new_size;
2214
2215         BUG_ON(slot < 0);
2216         BUG_ON(slot >= nritems);
2217
2218         /*
2219          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2220          */
2221         /* first correct the data pointers */
2222         for (i = slot; i < nritems; i++) {
2223                 u32 ioff;
2224                 item = btrfs_item_nr(leaf, i);
2225
2226                 if (!leaf->map_token) {
2227                         map_extent_buffer(leaf, (unsigned long)item,
2228                                         sizeof(struct btrfs_item),
2229                                         &leaf->map_token, &leaf->kaddr,
2230                                         &leaf->map_start, &leaf->map_len,
2231                                         KM_USER1);
2232                 }
2233
2234                 ioff = btrfs_item_offset(leaf, item);
2235                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2236         }
2237
2238         if (leaf->map_token) {
2239                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2240                 leaf->map_token = NULL;
2241         }
2242
2243         /* shift the data */
2244         if (from_end) {
2245                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2246                               data_end + size_diff, btrfs_leaf_data(leaf) +
2247                               data_end, old_data_start + new_size - data_end);
2248         } else {
2249                 struct btrfs_disk_key disk_key;
2250                 u64 offset;
2251
2252                 btrfs_item_key(leaf, &disk_key, slot);
2253
2254                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2255                         unsigned long ptr;
2256                         struct btrfs_file_extent_item *fi;
2257
2258                         fi = btrfs_item_ptr(leaf, slot,
2259                                             struct btrfs_file_extent_item);
2260                         fi = (struct btrfs_file_extent_item *)(
2261                              (unsigned long)fi - size_diff);
2262
2263                         if (btrfs_file_extent_type(leaf, fi) ==
2264                             BTRFS_FILE_EXTENT_INLINE) {
2265                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2266                                 memmove_extent_buffer(leaf, ptr,
2267                                         (unsigned long)fi,
2268                                         offsetof(struct btrfs_file_extent_item,
2269                                                  disk_bytenr));
2270                         }
2271                 }
2272
2273                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2274                               data_end + size_diff, btrfs_leaf_data(leaf) +
2275                               data_end, old_data_start - data_end);
2276
2277                 offset = btrfs_disk_key_offset(&disk_key);
2278                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2279                 btrfs_set_item_key(leaf, &disk_key, slot);
2280                 if (slot == 0)
2281                         fixup_low_keys(trans, root, path, &disk_key, 1);
2282         }
2283
2284         item = btrfs_item_nr(leaf, slot);
2285         btrfs_set_item_size(leaf, item, new_size);
2286         btrfs_mark_buffer_dirty(leaf);
2287
2288         ret = 0;
2289         if (btrfs_leaf_free_space(root, leaf) < 0) {
2290                 btrfs_print_leaf(root, leaf);
2291                 BUG();
2292         }
2293         return ret;
2294 }
2295
2296 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2297                       struct btrfs_root *root, struct btrfs_path *path,
2298                       u32 data_size)
2299 {
2300         int ret = 0;
2301         int slot;
2302         int slot_orig;
2303         struct extent_buffer *leaf;
2304         struct btrfs_item *item;
2305         u32 nritems;
2306         unsigned int data_end;
2307         unsigned int old_data;
2308         unsigned int old_size;
2309         int i;
2310
2311         slot_orig = path->slots[0];
2312         leaf = path->nodes[0];
2313
2314         nritems = btrfs_header_nritems(leaf);
2315         data_end = leaf_data_end(root, leaf);
2316
2317         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2318                 btrfs_print_leaf(root, leaf);
2319                 BUG();
2320         }
2321         slot = path->slots[0];
2322         old_data = btrfs_item_end_nr(leaf, slot);
2323
2324         BUG_ON(slot < 0);
2325         if (slot >= nritems) {
2326                 btrfs_print_leaf(root, leaf);
2327                 printk("slot %d too large, nritems %d\n", slot, nritems);
2328                 BUG_ON(1);
2329         }
2330
2331         /*
2332          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2333          */
2334         /* first correct the data pointers */
2335         for (i = slot; i < nritems; i++) {
2336                 u32 ioff;
2337                 item = btrfs_item_nr(leaf, i);
2338
2339                 if (!leaf->map_token) {
2340                         map_extent_buffer(leaf, (unsigned long)item,
2341                                         sizeof(struct btrfs_item),
2342                                         &leaf->map_token, &leaf->kaddr,
2343                                         &leaf->map_start, &leaf->map_len,
2344                                         KM_USER1);
2345                 }
2346                 ioff = btrfs_item_offset(leaf, item);
2347                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2348         }
2349
2350         if (leaf->map_token) {
2351                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2352                 leaf->map_token = NULL;
2353         }
2354
2355         /* shift the data */
2356         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2357                       data_end - data_size, btrfs_leaf_data(leaf) +
2358                       data_end, old_data - data_end);
2359
2360         data_end = old_data;
2361         old_size = btrfs_item_size_nr(leaf, slot);
2362         item = btrfs_item_nr(leaf, slot);
2363         btrfs_set_item_size(leaf, item, old_size + data_size);
2364         btrfs_mark_buffer_dirty(leaf);
2365
2366         ret = 0;
2367         if (btrfs_leaf_free_space(root, leaf) < 0) {
2368                 btrfs_print_leaf(root, leaf);
2369                 BUG();
2370         }
2371         return ret;
2372 }
2373
2374 /*
2375  * Given a key and some data, insert an item into the tree.
2376  * This does all the path init required, making room in the tree if needed.
2377  */
2378 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2379                             struct btrfs_root *root,
2380                             struct btrfs_path *path,
2381                             struct btrfs_key *cpu_key, u32 *data_size,
2382                             int nr)
2383 {
2384         struct extent_buffer *leaf;
2385         struct btrfs_item *item;
2386         int ret = 0;
2387         int slot;
2388         int slot_orig;
2389         int i;
2390         u32 nritems;
2391         u32 total_size = 0;
2392         u32 total_data = 0;
2393         unsigned int data_end;
2394         struct btrfs_disk_key disk_key;
2395
2396         for (i = 0; i < nr; i++) {
2397                 total_data += data_size[i];
2398         }
2399
2400         /* create a root if there isn't one */
2401         if (!root->node)
2402                 BUG();
2403
2404         total_size = total_data + (nr - 1) * sizeof(struct btrfs_item);
2405         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2406         if (ret == 0) {
2407                 return -EEXIST;
2408         }
2409         if (ret < 0)
2410                 goto out;
2411
2412         slot_orig = path->slots[0];
2413         leaf = path->nodes[0];
2414
2415         nritems = btrfs_header_nritems(leaf);
2416         data_end = leaf_data_end(root, leaf);
2417
2418         if (btrfs_leaf_free_space(root, leaf) <
2419             sizeof(struct btrfs_item) + total_size) {
2420                 btrfs_print_leaf(root, leaf);
2421                 printk("not enough freespace need %u have %d\n",
2422                        total_size, btrfs_leaf_free_space(root, leaf));
2423                 BUG();
2424         }
2425
2426         slot = path->slots[0];
2427         BUG_ON(slot < 0);
2428
2429         if (slot != nritems) {
2430                 int i;
2431                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2432
2433                 if (old_data < data_end) {
2434                         btrfs_print_leaf(root, leaf);
2435                         printk("slot %d old_data %d data_end %d\n",
2436                                slot, old_data, data_end);
2437                         BUG_ON(1);
2438                 }
2439                 /*
2440                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2441                  */
2442                 /* first correct the data pointers */
2443                 WARN_ON(leaf->map_token);
2444                 for (i = slot; i < nritems; i++) {
2445                         u32 ioff;
2446
2447                         item = btrfs_item_nr(leaf, i);
2448                         if (!leaf->map_token) {
2449                                 map_extent_buffer(leaf, (unsigned long)item,
2450                                         sizeof(struct btrfs_item),
2451                                         &leaf->map_token, &leaf->kaddr,
2452                                         &leaf->map_start, &leaf->map_len,
2453                                         KM_USER1);
2454                         }
2455
2456                         ioff = btrfs_item_offset(leaf, item);
2457                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2458                 }
2459                 if (leaf->map_token) {
2460                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2461                         leaf->map_token = NULL;
2462                 }
2463
2464                 /* shift the items */
2465                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2466                               btrfs_item_nr_offset(slot),
2467                               (nritems - slot) * sizeof(struct btrfs_item));
2468
2469                 /* shift the data */
2470                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2471                               data_end - total_data, btrfs_leaf_data(leaf) +
2472                               data_end, old_data - data_end);
2473                 data_end = old_data;
2474         }
2475
2476         /* setup the item for the new data */
2477         for (i = 0; i < nr; i++) {
2478                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2479                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2480                 item = btrfs_item_nr(leaf, slot + i);
2481                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2482                 data_end -= data_size[i];
2483                 btrfs_set_item_size(leaf, item, data_size[i]);
2484         }
2485         btrfs_set_header_nritems(leaf, nritems + nr);
2486         btrfs_mark_buffer_dirty(leaf);
2487
2488         ret = 0;
2489         if (slot == 0) {
2490                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2491                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2492         }
2493
2494         if (btrfs_leaf_free_space(root, leaf) < 0) {
2495                 btrfs_print_leaf(root, leaf);
2496                 BUG();
2497         }
2498
2499 out:
2500         return ret;
2501 }
2502
2503 /*
2504  * Given a key and some data, insert an item into the tree.
2505  * This does all the path init required, making room in the tree if needed.
2506  */
2507 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2508                       *root, struct btrfs_key *cpu_key, void *data, u32
2509                       data_size)
2510 {
2511         int ret = 0;
2512         struct btrfs_path *path;
2513         struct extent_buffer *leaf;
2514         unsigned long ptr;
2515
2516         path = btrfs_alloc_path();
2517         BUG_ON(!path);
2518         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2519         if (!ret) {
2520                 leaf = path->nodes[0];
2521                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2522                 write_extent_buffer(leaf, data, ptr, data_size);
2523                 btrfs_mark_buffer_dirty(leaf);
2524         }
2525         btrfs_free_path(path);
2526         return ret;
2527 }
2528
2529 /*
2530  * delete the pointer from a given node.
2531  *
2532  * If the delete empties a node, the node is removed from the tree,
2533  * continuing all the way the root if required.  The root is converted into
2534  * a leaf if all the nodes are emptied.
2535  */
2536 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2537                    struct btrfs_path *path, int level, int slot)
2538 {
2539         struct extent_buffer *parent = path->nodes[level];
2540         u32 nritems;
2541         int ret = 0;
2542         int wret;
2543
2544         nritems = btrfs_header_nritems(parent);
2545         if (slot != nritems -1) {
2546                 memmove_extent_buffer(parent,
2547                               btrfs_node_key_ptr_offset(slot),
2548                               btrfs_node_key_ptr_offset(slot + 1),
2549                               sizeof(struct btrfs_key_ptr) *
2550                               (nritems - slot - 1));
2551         }
2552         nritems--;
2553         btrfs_set_header_nritems(parent, nritems);
2554         if (nritems == 0 && parent == root->node) {
2555                 BUG_ON(btrfs_header_level(root->node) != 1);
2556                 /* just turn the root into a leaf and break */
2557                 btrfs_set_header_level(root->node, 0);
2558         } else if (slot == 0) {
2559                 struct btrfs_disk_key disk_key;
2560
2561                 btrfs_node_key(parent, &disk_key, 0);
2562                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2563                 if (wret)
2564                         ret = wret;
2565         }
2566         btrfs_mark_buffer_dirty(parent);
2567         return ret;
2568 }
2569
2570 /*
2571  * delete the item at the leaf level in path.  If that empties
2572  * the leaf, remove it from the tree
2573  */
2574 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2575                     struct btrfs_path *path, int slot, int nr)
2576 {
2577         struct extent_buffer *leaf;
2578         struct btrfs_item *item;
2579         int last_off;
2580         int dsize = 0;
2581         int ret = 0;
2582         int wret;
2583         int i;
2584         u32 nritems;
2585
2586         leaf = path->nodes[0];
2587         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2588
2589         for (i = 0; i < nr; i++)
2590                 dsize += btrfs_item_size_nr(leaf, slot + i);
2591
2592         nritems = btrfs_header_nritems(leaf);
2593
2594         if (slot + nr != nritems) {
2595                 int i;
2596                 int data_end = leaf_data_end(root, leaf);
2597
2598                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2599                               data_end + dsize,
2600                               btrfs_leaf_data(leaf) + data_end,
2601                               last_off - data_end);
2602
2603                 for (i = slot + nr; i < nritems; i++) {
2604                         u32 ioff;
2605
2606                         item = btrfs_item_nr(leaf, i);
2607                         if (!leaf->map_token) {
2608                                 map_extent_buffer(leaf, (unsigned long)item,
2609                                         sizeof(struct btrfs_item),
2610                                         &leaf->map_token, &leaf->kaddr,
2611                                         &leaf->map_start, &leaf->map_len,
2612                                         KM_USER1);
2613                         }
2614                         ioff = btrfs_item_offset(leaf, item);
2615                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2616                 }
2617
2618                 if (leaf->map_token) {
2619                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2620                         leaf->map_token = NULL;
2621                 }
2622
2623                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2624                               btrfs_item_nr_offset(slot + nr),
2625                               sizeof(struct btrfs_item) *
2626                               (nritems - slot - nr));
2627         }
2628         btrfs_set_header_nritems(leaf, nritems - nr);
2629         nritems -= nr;
2630
2631         /* delete the leaf if we've emptied it */
2632         if (nritems == 0) {
2633                 if (leaf == root->node) {
2634                         btrfs_set_header_level(leaf, 0);
2635                 } else {
2636                         u64 root_gen = btrfs_header_generation(path->nodes[1]);
2637                         clean_tree_block(trans, root, leaf);
2638                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
2639                         if (wret)
2640                                 ret = wret;
2641                         wret = btrfs_free_extent(trans, root,
2642                                          leaf->start, leaf->len,
2643                                          btrfs_header_owner(path->nodes[1]),
2644                                          root_gen, 0, 0, 1);
2645                         if (wret)
2646                                 ret = wret;
2647                 }
2648         } else {
2649                 int used = leaf_space_used(leaf, 0, nritems);
2650                 if (slot == 0) {
2651                         struct btrfs_disk_key disk_key;
2652
2653                         btrfs_item_key(leaf, &disk_key, 0);
2654                         wret = fixup_low_keys(trans, root, path,
2655                                               &disk_key, 1);
2656                         if (wret)
2657                                 ret = wret;
2658                 }
2659
2660                 /* delete the leaf if it is mostly empty */
2661                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2662                         /* push_leaf_left fixes the path.
2663                          * make sure the path still points to our leaf
2664                          * for possible call to del_ptr below
2665                          */
2666                         slot = path->slots[1];
2667                         extent_buffer_get(leaf);
2668
2669                         wret = push_leaf_left(trans, root, path, 1, 1);
2670                         if (wret < 0 && wret != -ENOSPC)
2671                                 ret = wret;
2672
2673                         if (path->nodes[0] == leaf &&
2674                             btrfs_header_nritems(leaf)) {
2675                                 wret = push_leaf_right(trans, root, path, 1, 1);
2676                                 if (wret < 0 && wret != -ENOSPC)
2677                                         ret = wret;
2678                         }
2679
2680                         if (btrfs_header_nritems(leaf) == 0) {
2681                                 u64 root_gen;
2682                                 u64 bytenr = leaf->start;
2683                                 u32 blocksize = leaf->len;
2684
2685                                 root_gen = btrfs_header_generation(
2686                                                            path->nodes[1]);
2687
2688                                 clean_tree_block(trans, root, leaf);
2689
2690                                 wret = del_ptr(trans, root, path, 1, slot);
2691                                 if (wret)
2692                                         ret = wret;
2693
2694                                 free_extent_buffer(leaf);
2695                                 wret = btrfs_free_extent(trans, root, bytenr,
2696                                              blocksize,
2697                                              btrfs_header_owner(path->nodes[1]),
2698                                              root_gen, 0, 0, 1);
2699                                 if (wret)
2700                                         ret = wret;
2701                         } else {
2702                                 btrfs_mark_buffer_dirty(leaf);
2703                                 free_extent_buffer(leaf);
2704                         }
2705                 } else {
2706                         btrfs_mark_buffer_dirty(leaf);
2707                 }
2708         }
2709         return ret;
2710 }
2711
2712 /*
2713  * walk up the tree as far as required to find the previous leaf.
2714  * returns 0 if it found something or 1 if there are no lesser leaves.
2715  * returns < 0 on io errors.
2716  */
2717 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2718 {
2719         u64 bytenr;
2720         int slot;
2721         int level = 1;
2722         struct extent_buffer *c;
2723         struct extent_buffer *next = NULL;
2724
2725         while(level < BTRFS_MAX_LEVEL) {
2726                 if (!path->nodes[level])
2727                         return 1;
2728
2729                 slot = path->slots[level];
2730                 c = path->nodes[level];
2731                 if (slot == 0) {
2732                         level++;
2733                         if (level == BTRFS_MAX_LEVEL)
2734                                 return 1;
2735                         continue;
2736                 }
2737                 slot--;
2738
2739                 bytenr = btrfs_node_blockptr(c, slot);
2740                 if (next)
2741                         free_extent_buffer(next);
2742
2743                 next = read_tree_block(root, bytenr,
2744                                        btrfs_level_size(root, level - 1));
2745                 break;
2746         }
2747         path->slots[level] = slot;
2748         while(1) {
2749                 level--;
2750                 c = path->nodes[level];
2751                 free_extent_buffer(c);
2752                 slot = btrfs_header_nritems(next);
2753                 if (slot != 0)
2754                         slot--;
2755                 path->nodes[level] = next;
2756                 path->slots[level] = slot;
2757                 if (!level)
2758                         break;
2759                 next = read_tree_block(root, btrfs_node_blockptr(next, slot),
2760                                        btrfs_level_size(root, level - 1));
2761         }
2762         return 0;
2763 }
2764
2765 /*
2766  * walk up the tree as far as required to find the next leaf.
2767  * returns 0 if it found something or 1 if there are no greater leaves.
2768  * returns < 0 on io errors.
2769  */
2770 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2771 {
2772         int slot;
2773         int level = 1;
2774         u64 bytenr;
2775         struct extent_buffer *c;
2776         struct extent_buffer *next = NULL;
2777
2778         while(level < BTRFS_MAX_LEVEL) {
2779                 if (!path->nodes[level])
2780                         return 1;
2781
2782                 slot = path->slots[level] + 1;
2783                 c = path->nodes[level];
2784                 if (slot >= btrfs_header_nritems(c)) {
2785                         level++;
2786                         if (level == BTRFS_MAX_LEVEL)
2787                                 return 1;
2788                         continue;
2789                 }
2790
2791                 bytenr = btrfs_node_blockptr(c, slot);
2792                 if (next)
2793                         free_extent_buffer(next);
2794
2795                 if (path->reada)
2796                         reada_for_search(root, path, level, slot, 0);
2797
2798                 next = read_tree_block(root, bytenr,
2799                                        btrfs_level_size(root, level -1));
2800                 break;
2801         }
2802         path->slots[level] = slot;
2803         while(1) {
2804                 level--;
2805                 c = path->nodes[level];
2806                 free_extent_buffer(c);
2807                 path->nodes[level] = next;
2808                 path->slots[level] = 0;
2809                 if (!level)
2810                         break;
2811                 if (path->reada)
2812                         reada_for_search(root, path, level, 0, 0);
2813                 next = read_tree_block(root, btrfs_node_blockptr(next, 0),
2814                                        btrfs_level_size(root, level - 1));
2815         }
2816         return 0;
2817 }
2818
2819 int btrfs_previous_item(struct btrfs_root *root,
2820                         struct btrfs_path *path, u64 min_objectid,
2821                         int type)
2822 {
2823         struct btrfs_key found_key;
2824         struct extent_buffer *leaf;
2825         int ret;
2826
2827         while(1) {
2828                 if (path->slots[0] == 0) {
2829                         ret = btrfs_prev_leaf(root, path);
2830                         if (ret != 0)
2831                                 return ret;
2832                 } else {
2833                         path->slots[0]--;
2834                 }
2835                 leaf = path->nodes[0];
2836                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2837                 if (found_key.type == type)
2838                         return 0;
2839         }
2840         return 1;
2841 }
2842