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         if (btrfs_header_level(path->nodes[level]) != level)
 574             printk("warning: bad level %Lu wanted %d found %d\n",
 575                    path->nodes[level]->start, level,
 576                    btrfs_header_level(path->nodes[level]));
 577         found_start = btrfs_header_bytenr(path->nodes[level]);
 578         if (found_start != path->nodes[level]->start) {
 579             printk("warning: bad bytentr %Lu found %Lu\n",
 580                    path->nodes[level]->start, found_start);
 581         }
 582 #if 0
 583         struct extent_buffer *buf = path->nodes[level];
 584
 585         if (memcmp_extent_buffer(buf, root->fs_info->fsid,
 586                                  (unsigned long)btrfs_header_fsid(buf),
 587                                  BTRFS_FSID_SIZE)) {
 588                 printk("warning bad block %Lu\n", buf->start);
 589                 return 1;
 590         }
 591 #endif
 592         if (level == 0)
 593                 return check_leaf(root, path, level);
 594         return check_node(root, path, level);
 595 }
 596
 597 /*
 598  * search for key in the extent_buffer.  The items start at offset p,
 599  * and they are item_size apart.  There are 'max' items in p.
 600  *
 601  * the slot in the array is returned via slot, and it points to
 602  * the place where you would insert key if it is not found in
 603  * the array.
 604  *
 605  * slot may point to max if the key is bigger than all of the keys
 606  */
 607 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 608                               int item_size, struct btrfs_key *key,
 609                               int max, int *slot)
 610 {
 611         int low = 0;
 612         int high = max;
 613         int mid;
 614         int ret;
 615         struct btrfs_disk_key *tmp = NULL;
 616         struct btrfs_disk_key unaligned;
 617         unsigned long offset;
 618         char *map_token = NULL;
 619         char *kaddr = NULL;
 620         unsigned long map_start = 0;
 621         unsigned long map_len = 0;
 622         int err;
 623
 624         while(low < high) {
 625                 mid = (low + high) / 2;
 626                 offset = p + mid * item_size;
 627
 628                 if (!map_token || offset < map_start ||
 629                     (offset + sizeof(struct btrfs_disk_key)) >
 630                     map_start + map_len) {
 631                         if (map_token) {
 632                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 633                                 map_token = NULL;
 634                         }
 635                         err = map_extent_buffer(eb, offset,
 636                                                 sizeof(struct btrfs_disk_key),
 637                                                 &map_token, &kaddr,
 638                                                 &map_start, &map_len, KM_USER0);
 639
 640                         if (!err) {
 641                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
 642                                                         map_start);
 643                         } else {
 644                                 read_extent_buffer(eb, &unaligned,
 645                                                    offset, sizeof(unaligned));
 646                                 tmp = &unaligned;
 647                         }
 648
 649                 } else {
 650                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
 651                                                         map_start);
 652                 }
 653                 ret = comp_keys(tmp, key);
 654
 655                 if (ret < 0)
 656                         low = mid + 1;
 657                 else if (ret > 0)
 658                         high = mid;
 659                 else {
 660                         *slot = mid;
 661                         if (map_token)
 662                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 663                         return 0;
 664                 }
 665         }
 666         *slot = low;
 667         if (map_token)
 668                 unmap_extent_buffer(eb, map_token, KM_USER0);
 669         return 1;
 670 }
 671
 672 /*
 673  * simple bin_search frontend that does the right thing for
 674  * leaves vs nodes
 675  */
 676 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 677                       int level, int *slot)
 678 {
 679         if (level == 0) {
 680                 return generic_bin_search(eb,
 681                                           offsetof(struct btrfs_leaf, items),
 682                                           sizeof(struct btrfs_item),
 683                                           key, btrfs_header_nritems(eb),
 684                                           slot);
 685         } else {
 686                 return generic_bin_search(eb,
 687                                           offsetof(struct btrfs_node, ptrs),
 688                                           sizeof(struct btrfs_key_ptr),
 689                                           key, btrfs_header_nritems(eb),
 690                                           slot);
 691         }
 692         return -1;
 693 }
 694
 695 static struct extent_buffer *read_node_slot(struct btrfs_root *root,
 696                                    struct extent_buffer *parent, int slot)
 697 {
 698         if (slot < 0)
 699                 return NULL;
 700         if (slot >= btrfs_header_nritems(parent))
 701                 return NULL;
 702         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 703                        btrfs_level_size(root, btrfs_header_level(parent) - 1));
 704 }
 705
 706 static int balance_level(struct btrfs_trans_handle *trans,
 707                          struct btrfs_root *root,
 708                          struct btrfs_path *path, int level)
 709 {
 710         struct extent_buffer *right = NULL;
 711         struct extent_buffer *mid;
 712         struct extent_buffer *left = NULL;
 713         struct extent_buffer *parent = NULL;
 714         int ret = 0;
 715         int wret;
 716         int pslot;
 717         int orig_slot = path->slots[level];
 718         int err_on_enospc = 0;
 719         u64 orig_ptr;
 720
 721         if (level == 0)
 722                 return 0;
 723
 724         mid = path->nodes[level];
 725         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 726
 727         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 728
 729         if (level < BTRFS_MAX_LEVEL - 1)
 730                 parent = path->nodes[level + 1];
 731         pslot = path->slots[level + 1];
 732
 733         /*
 734          * deal with the case where there is only one pointer in the root
 735          * by promoting the node below to a root
 736          */
 737         if (!parent) {
 738                 struct extent_buffer *child;
 739
 740                 if (btrfs_header_nritems(mid) != 1)
 741                         return 0;
 742
 743                 /* promote the child to a root */
 744                 child = read_node_slot(root, mid, 0);
 745                 BUG_ON(!child);
 746                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
 747                 BUG_ON(ret);
 748
 749                 root->node = child;
 750                 add_root_to_dirty_list(root);
 751                 path->nodes[level] = NULL;
 752                 clean_tree_block(trans, root, mid);
 753                 wait_on_tree_block_writeback(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                         wait_on_tree_block_writeback(root, right);
 813                         free_extent_buffer(right);
 814                         right = NULL;
 815                         wret = del_ptr(trans, root, path, level + 1, pslot +
 816                                        1);
 817                         if (wret)
 818                                 ret = wret;
 819                         wret = btrfs_free_extent(trans, root, bytenr,
 820                                                  blocksize,
 821                                                  btrfs_header_owner(parent),
 822                                                  generation, 0, 0, 1);
 823                         if (wret)
 824                                 ret = wret;
 825                 } else {
 826                         struct btrfs_disk_key right_key;
 827                         btrfs_node_key(right, &right_key, 0);
 828                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 829                         btrfs_mark_buffer_dirty(parent);
 830                 }
 831         }
 832         if (btrfs_header_nritems(mid) == 1) {
 833                 /*
 834                  * we're not allowed to leave a node with one item in the
 835                  * tree during a delete.  A deletion from lower in the tree
 836                  * could try to delete the only pointer in this node.
 837                  * So, pull some keys from the left.
 838                  * There has to be a left pointer at this point because
 839                  * otherwise we would have pulled some pointers from the
 840                  * right
 841                  */
 842                 BUG_ON(!left);
 843                 wret = balance_node_right(trans, root, mid, left);
 844                 if (wret < 0) {
 845                         ret = wret;
 846                         goto enospc;
 847                 }
 848                 BUG_ON(wret == 1);
 849         }
 850         if (btrfs_header_nritems(mid) == 0) {
 851                 /* we've managed to empty the middle node, drop it */
 852                 u64 root_gen = btrfs_header_generation(parent);
 853                 u64 bytenr = mid->start;
 854                 u32 blocksize = mid->len;
 855                 clean_tree_block(trans, root, mid);
 856                 wait_on_tree_block_writeback(root, mid);
 857                 free_extent_buffer(mid);
 858                 mid = NULL;
 859                 wret = del_ptr(trans, root, path, level + 1, pslot);
 860                 if (wret)
 861                         ret = wret;
 862                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 863                                          btrfs_header_owner(parent),
 864                                          root_gen, 0, 0, 1);
 865                 if (wret)
 866                         ret = wret;
 867         } else {
 868                 /* update the parent key to reflect our changes */
 869                 struct btrfs_disk_key mid_key;
 870                 btrfs_node_key(mid, &mid_key, 0);
 871                 btrfs_set_node_key(parent, &mid_key, pslot);
 872                 btrfs_mark_buffer_dirty(parent);
 873         }
 874
 875         /* update the path */
 876         if (left) {
 877                 if (btrfs_header_nritems(left) > orig_slot) {
 878                         extent_buffer_get(left);
 879                         path->nodes[level] = left;
 880                         path->slots[level + 1] -= 1;
 881                         path->slots[level] = orig_slot;
 882                         if (mid)
 883                                 free_extent_buffer(mid);
 884                 } else {
 885                         orig_slot -= btrfs_header_nritems(left);
 886                         path->slots[level] = orig_slot;
 887                 }
 888         }
 889         /* double check we haven't messed things up */
 890         check_block(root, path, level);
 891         if (orig_ptr !=
 892             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 893                 BUG();
 894 enospc:
 895         if (right)
 896                 free_extent_buffer(right);
 897         if (left)
 898                 free_extent_buffer(left);
 899         return ret;
 900 }
 901
 902 /* returns zero if the push worked, non-zero otherwise */
 903 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
 904                                           struct btrfs_root *root,
 905                                           struct btrfs_path *path, int level)
 906 {
 907         struct extent_buffer *right = NULL;
 908         struct extent_buffer *mid;
 909         struct extent_buffer *left = NULL;
 910         struct extent_buffer *parent = NULL;
 911         int ret = 0;
 912         int wret;
 913         int pslot;
 914         int orig_slot = path->slots[level];
 915         u64 orig_ptr;
 916
 917         if (level == 0)
 918                 return 1;
 919
 920         mid = path->nodes[level];
 921         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 922         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 923
 924         if (level < BTRFS_MAX_LEVEL - 1)
 925                 parent = path->nodes[level + 1];
 926         pslot = path->slots[level + 1];
 927
 928         if (!parent)
 929                 return 1;
 930
 931         left = read_node_slot(root, parent, pslot - 1);
 932
 933         /* first, try to make some room in the middle buffer */
 934         if (left) {
 935                 u32 left_nr;
 936                 left_nr = btrfs_header_nritems(left);
 937                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 938                         wret = 1;
 939                 } else {
 940                         ret = btrfs_cow_block(trans, root, left, parent,
 941                                               pslot - 1, &left);
 942                         if (ret)
 943                                 wret = 1;
 944                         else {
 945                                 wret = push_node_left(trans, root,
 946                                                       left, mid);
 947                         }
 948                 }
 949                 if (wret < 0)
 950                         ret = wret;
 951                 if (wret == 0) {
 952                         struct btrfs_disk_key disk_key;
 953                         orig_slot += left_nr;
 954                         btrfs_node_key(mid, &disk_key, 0);
 955                         btrfs_set_node_key(parent, &disk_key, pslot);
 956                         btrfs_mark_buffer_dirty(parent);
 957                         if (btrfs_header_nritems(left) > orig_slot) {
 958                                 path->nodes[level] = left;
 959                                 path->slots[level + 1] -= 1;
 960                                 path->slots[level] = orig_slot;
 961                                 free_extent_buffer(mid);
 962                         } else {
 963                                 orig_slot -=
 964                                         btrfs_header_nritems(left);
 965                                 path->slots[level] = orig_slot;
 966                                 free_extent_buffer(left);
 967                         }
 968                         return 0;
 969                 }
 970                 free_extent_buffer(left);
 971         }
 972         right= read_node_slot(root, parent, pslot + 1);
 973
 974         /*
 975          * then try to empty the right most buffer into the middle
 976          */
 977         if (right) {
 978                 u32 right_nr;
 979                 right_nr = btrfs_header_nritems(right);
 980                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 981                         wret = 1;
 982                 } else {
 983                         ret = btrfs_cow_block(trans, root, right,
 984                                               parent, pslot + 1,
 985                                               &right);
 986                         if (ret)
 987                                 wret = 1;
 988                         else {
 989                                 wret = balance_node_right(trans, root,
 990                                                           right, mid);
 991                         }
 992                 }
 993                 if (wret < 0)
 994                         ret = wret;
 995                 if (wret == 0) {
 996                         struct btrfs_disk_key disk_key;
 997
 998                         btrfs_node_key(right, &disk_key, 0);
 999                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
1000                         btrfs_mark_buffer_dirty(parent);
1001
1002                         if (btrfs_header_nritems(mid) <= orig_slot) {
1003                                 path->nodes[level] = right;
1004                                 path->slots[level + 1] += 1;
1005                                 path->slots[level] = orig_slot -
1006                                         btrfs_header_nritems(mid);
1007                                 free_extent_buffer(mid);
1008                         } else {
1009                                 free_extent_buffer(right);
1010                         }
1011                         return 0;
1012                 }
1013                 free_extent_buffer(right);
1014         }
1015         return 1;
1016 }
1017
1018 /*
1019  * readahead one full node of leaves
1020  */
1021 static void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
1022                              int level, int slot, u64 objectid)
1023 {
1024         struct extent_buffer *node;
1025         struct btrfs_disk_key disk_key;
1026         u32 nritems;
1027         u64 search;
1028         u64 lowest_read;
1029         u64 highest_read;
1030         u64 nread = 0;
1031         int direction = path->reada;
1032         struct extent_buffer *eb;
1033         u32 nr;
1034         u32 blocksize;
1035         u32 nscan = 0;
1036
1037         if (level != 1)
1038                 return;
1039
1040         if (!path->nodes[level])
1041                 return;
1042
1043         node = path->nodes[level];
1044         search = btrfs_node_blockptr(node, slot);
1045         blocksize = btrfs_level_size(root, level - 1);
1046         eb = btrfs_find_tree_block(root, search, blocksize);
1047         if (eb) {
1048                 free_extent_buffer(eb);
1049                 return;
1050         }
1051
1052         highest_read = search;
1053         lowest_read = search;
1054
1055         nritems = btrfs_header_nritems(node);
1056         nr = slot;
1057         while(1) {
1058                 if (direction < 0) {
1059                         if (nr == 0)
1060                                 break;
1061                         nr--;
1062                 } else if (direction > 0) {
1063                         nr++;
1064                         if (nr >= nritems)
1065                                 break;
1066                 }
1067                 if (path->reada < 0 && objectid) {
1068                         btrfs_node_key(node, &disk_key, nr);
1069                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1070                                 break;
1071                 }
1072                 search = btrfs_node_blockptr(node, nr);
1073                 if ((search >= lowest_read && search <= highest_read) ||
1074                     (search < lowest_read && lowest_read - search <= 32768) ||
1075                     (search > highest_read && search - highest_read <= 32768)) {
1076                         readahead_tree_block(root, search, blocksize);
1077                         nread += blocksize;
1078                 }
1079                 nscan++;
1080                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
1081                         break;
1082                 if(nread > (1024 * 1024) || nscan > 128)
1083                         break;
1084
1085                 if (search < lowest_read)
1086                         lowest_read = search;
1087                 if (search > highest_read)
1088                         highest_read = search;
1089         }
1090 }
1091 /*
1092  * look for key in the tree.  path is filled in with nodes along the way
1093  * if key is found, we return zero and you can find the item in the leaf
1094  * level of the path (level 0)
1095  *
1096  * If the key isn't found, the path points to the slot where it should
1097  * be inserted, and 1 is returned.  If there are other errors during the
1098  * search a negative error number is returned.
1099  *
1100  * if ins_len > 0, nodes and leaves will be split as we walk down the
1101  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1102  * possible)
1103  */
1104 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1105                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1106                       ins_len, int cow)
1107 {
1108         struct extent_buffer *b;
1109         u64 bytenr;
1110         u64 ptr_gen;
1111         int slot;
1112         int ret;
1113         int level;
1114         int should_reada = p->reada;
1115         u8 lowest_level = 0;
1116
1117         lowest_level = p->lowest_level;
1118         WARN_ON(lowest_level && ins_len);
1119         WARN_ON(p->nodes[0] != NULL);
1120         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
1121 again:
1122         b = root->node;
1123         extent_buffer_get(b);
1124         while (b) {
1125                 level = btrfs_header_level(b);
1126                 if (cow) {
1127                         int wret;
1128                         wret = btrfs_cow_block(trans, root, b,
1129                                                p->nodes[level + 1],
1130                                                p->slots[level + 1],
1131                                                &b);
1132                         if (wret) {
1133                                 free_extent_buffer(b);
1134                                 return wret;
1135                         }
1136                 }
1137                 BUG_ON(!cow && ins_len);
1138                 if (level != btrfs_header_level(b))
1139                         WARN_ON(1);
1140                 level = btrfs_header_level(b);
1141                 p->nodes[level] = b;
1142                 ret = check_block(root, p, level);
1143                 if (ret)
1144                         return -1;
1145                 ret = bin_search(b, key, level, &slot);
1146                 if (level != 0) {
1147                         if (ret && slot > 0)
1148                                 slot -= 1;
1149                         p->slots[level] = slot;
1150                         if (ins_len > 0 && btrfs_header_nritems(b) >=
1151                             BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1152                                 int sret = split_node(trans, root, p, level);
1153                                 BUG_ON(sret > 0);
1154                                 if (sret)
1155                                         return sret;
1156                                 b = p->nodes[level];
1157                                 slot = p->slots[level];
1158                         } else if (ins_len < 0) {
1159                                 int sret = balance_level(trans, root, p,
1160                                                          level);
1161                                 if (sret)
1162                                         return sret;
1163                                 b = p->nodes[level];
1164                                 if (!b) {
1165                                         btrfs_release_path(NULL, p);
1166                                         goto again;
1167                                 }
1168                                 slot = p->slots[level];
1169                                 BUG_ON(btrfs_header_nritems(b) == 1);
1170                         }
1171                         /* this is only true while dropping a snapshot */
1172                         if (level == lowest_level)
1173                                 break;
1174                         bytenr = btrfs_node_blockptr(b, slot);
1175                         ptr_gen = btrfs_node_ptr_generation(b, slot);
1176                         if (should_reada)
1177                                 reada_for_search(root, p, level, slot,
1178                                                  key->objectid);
1179                         b = read_tree_block(root, bytenr,
1180                                             btrfs_level_size(root, level - 1));
1181                         if (ptr_gen != btrfs_header_generation(b)) {
1182                                 printk("block %llu bad gen wanted %llu "
1183                                        "found %llu\n",
1184                                 (unsigned long long)b->start,
1185                                 (unsigned long long)ptr_gen,
1186                                 (unsigned long long)btrfs_header_generation(b));
1187                         }
1188                 } else {
1189                         p->slots[level] = slot;
1190                         if (ins_len > 0 && btrfs_leaf_free_space(root, b) <
1191                             sizeof(struct btrfs_item) + ins_len) {
1192                                 int sret = split_leaf(trans, root, key,
1193                                                       p, ins_len, ret == 0);
1194                                 BUG_ON(sret > 0);
1195                                 if (sret)
1196                                         return sret;
1197                         }
1198                         return ret;
1199                 }
1200         }
1201         return 1;
1202 }
1203
1204 /*
1205  * adjust the pointers going up the tree, starting at level
1206  * making sure the right key of each node is points to 'key'.
1207  * This is used after shifting pointers to the left, so it stops
1208  * fixing up pointers when a given leaf/node is not in slot 0 of the
1209  * higher levels
1210  *
1211  * If this fails to write a tree block, it returns -1, but continues
1212  * fixing up the blocks in ram so the tree is consistent.
1213  */
1214 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1215                           struct btrfs_root *root, struct btrfs_path *path,
1216                           struct btrfs_disk_key *key, int level)
1217 {
1218         int i;
1219         int ret = 0;
1220         struct extent_buffer *t;
1221
1222         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1223                 int tslot = path->slots[i];
1224                 if (!path->nodes[i])
1225                         break;
1226                 t = path->nodes[i];
1227                 btrfs_set_node_key(t, key, tslot);
1228                 btrfs_mark_buffer_dirty(path->nodes[i]);
1229                 if (tslot != 0)
1230                         break;
1231         }
1232         return ret;
1233 }
1234
1235 /*
1236  * try to push data from one node into the next node left in the
1237  * tree.
1238  *
1239  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1240  * error, and > 0 if there was no room in the left hand block.
1241  */
1242 static int push_node_left(struct btrfs_trans_handle *trans,
1243                           struct btrfs_root *root, struct extent_buffer *dst,
1244                           struct extent_buffer *src)
1245 {
1246         int push_items = 0;
1247         int src_nritems;
1248         int dst_nritems;
1249         int ret = 0;
1250
1251         src_nritems = btrfs_header_nritems(src);
1252         dst_nritems = btrfs_header_nritems(dst);
1253         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1254         WARN_ON(btrfs_header_generation(src) != trans->transid);
1255         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1256
1257         if (push_items <= 0) {
1258                 return 1;
1259         }
1260
1261         if (src_nritems < push_items)
1262                 push_items = src_nritems;
1263
1264         copy_extent_buffer(dst, src,
1265                            btrfs_node_key_ptr_offset(dst_nritems),
1266                            btrfs_node_key_ptr_offset(0),
1267                            push_items * sizeof(struct btrfs_key_ptr));
1268
1269         if (push_items < src_nritems) {
1270                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1271                                       btrfs_node_key_ptr_offset(push_items),
1272                                       (src_nritems - push_items) *
1273                                       sizeof(struct btrfs_key_ptr));
1274         }
1275         btrfs_set_header_nritems(src, src_nritems - push_items);
1276         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1277         btrfs_mark_buffer_dirty(src);
1278         btrfs_mark_buffer_dirty(dst);
1279         return ret;
1280 }
1281
1282 /*
1283  * try to push data from one node into the next node right in the
1284  * tree.
1285  *
1286  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1287  * error, and > 0 if there was no room in the right hand block.
1288  *
1289  * this will  only push up to 1/2 the contents of the left node over
1290  */
1291 static int balance_node_right(struct btrfs_trans_handle *trans,
1292                               struct btrfs_root *root,
1293                               struct extent_buffer *dst,
1294                               struct extent_buffer *src)
1295 {
1296         int push_items = 0;
1297         int max_push;
1298         int src_nritems;
1299         int dst_nritems;
1300         int ret = 0;
1301
1302         WARN_ON(btrfs_header_generation(src) != trans->transid);
1303         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1304
1305         src_nritems = btrfs_header_nritems(src);
1306         dst_nritems = btrfs_header_nritems(dst);
1307         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1308         if (push_items <= 0)
1309                 return 1;
1310
1311         max_push = src_nritems / 2 + 1;
1312         /* don't try to empty the node */
1313         if (max_push >= src_nritems)
1314                 return 1;
1315
1316         if (max_push < push_items)
1317                 push_items = max_push;
1318
1319         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1320                                       btrfs_node_key_ptr_offset(0),
1321                                       (dst_nritems) *
1322                                       sizeof(struct btrfs_key_ptr));
1323
1324         copy_extent_buffer(dst, src,
1325                            btrfs_node_key_ptr_offset(0),
1326                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1327                            push_items * sizeof(struct btrfs_key_ptr));
1328
1329         btrfs_set_header_nritems(src, src_nritems - push_items);
1330         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1331
1332         btrfs_mark_buffer_dirty(src);
1333         btrfs_mark_buffer_dirty(dst);
1334         return ret;
1335 }
1336
1337 /*
1338  * helper function to insert a new root level in the tree.
1339  * A new node is allocated, and a single item is inserted to
1340  * point to the existing root
1341  *
1342  * returns zero on success or < 0 on failure.
1343  */
1344 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1345                            struct btrfs_root *root,
1346                            struct btrfs_path *path, int level)
1347 {
1348         u64 root_gen;
1349         u64 lower_gen;
1350         struct extent_buffer *lower;
1351         struct extent_buffer *c;
1352         struct btrfs_disk_key lower_key;
1353
1354         BUG_ON(path->nodes[level]);
1355         BUG_ON(path->nodes[level-1] != root->node);
1356
1357         if (root->ref_cows)
1358                 root_gen = trans->transid;
1359         else
1360                 root_gen = 0;
1361
1362         lower = path->nodes[level-1];
1363         if (level == 1)
1364                 btrfs_item_key(lower, &lower_key, 0);
1365         else
1366                 btrfs_node_key(lower, &lower_key, 0);
1367
1368         c = __btrfs_alloc_free_block(trans, root, root->nodesize,
1369                                    root->root_key.objectid,
1370                                    root_gen, lower_key.objectid, level,
1371                                    root->node->start, 0);
1372         if (IS_ERR(c))
1373                 return PTR_ERR(c);
1374         memset_extent_buffer(c, 0, 0, root->nodesize);
1375         btrfs_set_header_nritems(c, 1);
1376         btrfs_set_header_level(c, level);
1377         btrfs_set_header_bytenr(c, c->start);
1378         btrfs_set_header_generation(c, trans->transid);
1379         btrfs_set_header_owner(c, root->root_key.objectid);
1380
1381         write_extent_buffer(c, root->fs_info->fsid,
1382                             (unsigned long)btrfs_header_fsid(c),
1383                             BTRFS_FSID_SIZE);
1384         btrfs_set_node_key(c, &lower_key, 0);
1385         btrfs_set_node_blockptr(c, 0, lower->start);
1386         lower_gen = btrfs_header_generation(lower);
1387         WARN_ON(lower_gen == 0);
1388
1389         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1390
1391         btrfs_mark_buffer_dirty(c);
1392
1393         /* the super has an extra ref to root->node */
1394         free_extent_buffer(root->node);
1395         root->node = c;
1396         add_root_to_dirty_list(root);
1397         extent_buffer_get(c);
1398         path->nodes[level] = c;
1399         path->slots[level] = 0;
1400
1401         if (root->ref_cows && lower_gen != trans->transid) {
1402                 struct btrfs_path *back_path = btrfs_alloc_path();
1403                 int ret;
1404                 ret = btrfs_insert_extent_backref(trans,
1405                                                   root->fs_info->extent_root,
1406                                                   path, lower->start,
1407                                                   root->root_key.objectid,
1408                                                   trans->transid, 0, 0);
1409                 BUG_ON(ret);
1410                 btrfs_free_path(back_path);
1411         }
1412         return 0;
1413 }
1414
1415 /*
1416  * worker function to insert a single pointer in a node.
1417  * the node should have enough room for the pointer already
1418  *
1419  * slot and level indicate where you want the key to go, and
1420  * blocknr is the block the key points to.
1421  *
1422  * returns zero on success and < 0 on any error
1423  */
1424 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1425                       *root, struct btrfs_path *path, struct btrfs_disk_key
1426                       *key, u64 bytenr, int slot, int level)
1427 {
1428         struct extent_buffer *lower;
1429         int nritems;
1430
1431         BUG_ON(!path->nodes[level]);
1432         lower = path->nodes[level];
1433         nritems = btrfs_header_nritems(lower);
1434         if (slot > nritems)
1435                 BUG();
1436         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1437                 BUG();
1438         if (slot != nritems) {
1439                 memmove_extent_buffer(lower,
1440                               btrfs_node_key_ptr_offset(slot + 1),
1441                               btrfs_node_key_ptr_offset(slot),
1442                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1443         }
1444         btrfs_set_node_key(lower, key, slot);
1445         btrfs_set_node_blockptr(lower, slot, bytenr);
1446         WARN_ON(trans->transid == 0);
1447         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1448         btrfs_set_header_nritems(lower, nritems + 1);
1449         btrfs_mark_buffer_dirty(lower);
1450         return 0;
1451 }
1452
1453 /*
1454  * split the node at the specified level in path in two.
1455  * The path is corrected to point to the appropriate node after the split
1456  *
1457  * Before splitting this tries to make some room in the node by pushing
1458  * left and right, if either one works, it returns right away.
1459  *
1460  * returns 0 on success and < 0 on failure
1461  */
1462 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1463                       *root, struct btrfs_path *path, int level)
1464 {
1465         u64 root_gen;
1466         struct extent_buffer *c;
1467         struct extent_buffer *split;
1468         struct btrfs_disk_key disk_key;
1469         int mid;
1470         int ret;
1471         int wret;
1472         u32 c_nritems;
1473
1474         c = path->nodes[level];
1475         WARN_ON(btrfs_header_generation(c) != trans->transid);
1476         if (c == root->node) {
1477                 /* trying to split the root, lets make a new one */
1478                 ret = insert_new_root(trans, root, path, level + 1);
1479                 if (ret)
1480                         return ret;
1481         } else {
1482                 ret = push_nodes_for_insert(trans, root, path, level);
1483                 c = path->nodes[level];
1484                 if (!ret && btrfs_header_nritems(c) <
1485                     BTRFS_NODEPTRS_PER_BLOCK(root) - 1)
1486                         return 0;
1487                 if (ret < 0)
1488                         return ret;
1489         }
1490
1491         c_nritems = btrfs_header_nritems(c);
1492         if (root->ref_cows)
1493                 root_gen = trans->transid;
1494         else
1495                 root_gen = 0;
1496
1497         btrfs_node_key(c, &disk_key, 0);
1498         split = __btrfs_alloc_free_block(trans, root, root->nodesize,
1499                                          root->root_key.objectid,
1500                                          root_gen,
1501                                          btrfs_disk_key_objectid(&disk_key),
1502                                          level, c->start, 0);
1503         if (IS_ERR(split))
1504                 return PTR_ERR(split);
1505
1506         btrfs_set_header_flags(split, btrfs_header_flags(c));
1507         btrfs_set_header_level(split, btrfs_header_level(c));
1508         btrfs_set_header_bytenr(split, split->start);
1509         btrfs_set_header_generation(split, trans->transid);
1510         btrfs_set_header_owner(split, root->root_key.objectid);
1511         btrfs_set_header_flags(split, 0);
1512         write_extent_buffer(split, root->fs_info->fsid,
1513                             (unsigned long)btrfs_header_fsid(split),
1514                             BTRFS_FSID_SIZE);
1515
1516         mid = (c_nritems + 1) / 2;
1517
1518         copy_extent_buffer(split, c,
1519                            btrfs_node_key_ptr_offset(0),
1520                            btrfs_node_key_ptr_offset(mid),
1521                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1522         btrfs_set_header_nritems(split, c_nritems - mid);
1523         btrfs_set_header_nritems(c, mid);
1524         ret = 0;
1525
1526         btrfs_mark_buffer_dirty(c);
1527         btrfs_mark_buffer_dirty(split);
1528
1529         btrfs_node_key(split, &disk_key, 0);
1530         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1531                           path->slots[level + 1] + 1,
1532                           level + 1);
1533         if (wret)
1534                 ret = wret;
1535
1536         if (path->slots[level] >= mid) {
1537                 path->slots[level] -= mid;
1538                 free_extent_buffer(c);
1539                 path->nodes[level] = split;
1540                 path->slots[level + 1] += 1;
1541         } else {
1542                 free_extent_buffer(split);
1543         }
1544         return ret;
1545 }
1546
1547 /*
1548  * how many bytes are required to store the items in a leaf.  start
1549  * and nr indicate which items in the leaf to check.  This totals up the
1550  * space used both by the item structs and the item data
1551  */
1552 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1553 {
1554         int data_len;
1555         int nritems = btrfs_header_nritems(l);
1556         int end = min(nritems, start + nr) - 1;
1557
1558         if (!nr)
1559                 return 0;
1560         data_len = btrfs_item_end_nr(l, start);
1561         data_len = data_len - btrfs_item_offset_nr(l, end);
1562         data_len += sizeof(struct btrfs_item) * nr;
1563         WARN_ON(data_len < 0);
1564         return data_len;
1565 }
1566
1567 /*
1568  * The space between the end of the leaf items and
1569  * the start of the leaf data.  IOW, how much room
1570  * the leaf has left for both items and data
1571  */
1572 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1573 {
1574         int nritems = btrfs_header_nritems(leaf);
1575         int ret;
1576         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1577         if (ret < 0) {
1578                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1579                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1580                        leaf_space_used(leaf, 0, nritems), nritems);
1581         }
1582         return ret;
1583 }
1584
1585 /*
1586  * push some data in the path leaf to the right, trying to free up at
1587  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1588  *
1589  * returns 1 if the push failed because the other node didn't have enough
1590  * room, 0 if everything worked out and < 0 if there were major errors.
1591  */
1592 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1593                            *root, struct btrfs_path *path, int data_size,
1594                            int empty)
1595 {
1596         struct extent_buffer *left = path->nodes[0];
1597         struct extent_buffer *right;
1598         struct extent_buffer *upper;
1599         struct btrfs_disk_key disk_key;
1600         int slot;
1601         u32 i;
1602         int free_space;
1603         int push_space = 0;
1604         int push_items = 0;
1605         struct btrfs_item *item;
1606         u32 left_nritems;
1607         u32 nr;
1608         u32 right_nritems;
1609         u32 data_end;
1610         u32 this_item_size;
1611         int ret;
1612
1613         slot = path->slots[1];
1614         if (!path->nodes[1]) {
1615                 return 1;
1616         }
1617         upper = path->nodes[1];
1618         if (slot >= btrfs_header_nritems(upper) - 1)
1619                 return 1;
1620
1621         right = read_tree_block(root, btrfs_node_blockptr(upper, slot + 1),
1622                                 root->leafsize);
1623         free_space = btrfs_leaf_free_space(root, right);
1624         if (free_space < data_size + sizeof(struct btrfs_item)) {
1625                 free_extent_buffer(right);
1626                 return 1;
1627         }
1628
1629         /* cow and double check */
1630         ret = btrfs_cow_block(trans, root, right, upper,
1631                               slot + 1, &right);
1632         if (ret) {
1633                 free_extent_buffer(right);
1634                 return 1;
1635         }
1636         free_space = btrfs_leaf_free_space(root, right);
1637         if (free_space < data_size + sizeof(struct btrfs_item)) {
1638                 free_extent_buffer(right);
1639                 return 1;
1640         }
1641
1642         left_nritems = btrfs_header_nritems(left);
1643         if (left_nritems == 0) {
1644                 free_extent_buffer(right);
1645                 return 1;
1646         }
1647
1648         if (empty)
1649                 nr = 0;
1650         else
1651                 nr = 1;
1652
1653         i = left_nritems - 1;
1654         while (i >= nr) {
1655                 item = btrfs_item_nr(left, i);
1656
1657                 if (path->slots[0] == i)
1658                         push_space += data_size + sizeof(*item);
1659
1660                 if (!left->map_token) {
1661                         map_extent_buffer(left, (unsigned long)item,
1662                                         sizeof(struct btrfs_item),
1663                                         &left->map_token, &left->kaddr,
1664                                         &left->map_start, &left->map_len,
1665                                         KM_USER1);
1666                 }
1667
1668                 this_item_size = btrfs_item_size(left, item);
1669                 if (this_item_size + sizeof(*item) + push_space > free_space)
1670                         break;
1671                 push_items++;
1672                 push_space += this_item_size + sizeof(*item);
1673                 if (i == 0)
1674                         break;
1675                 i--;
1676         }
1677         if (left->map_token) {
1678                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1679                 left->map_token = NULL;
1680         }
1681
1682         if (push_items == 0) {
1683                 free_extent_buffer(right);
1684                 return 1;
1685         }
1686
1687         if (!empty && push_items == left_nritems)
1688                 WARN_ON(1);
1689
1690         /* push left to right */
1691         right_nritems = btrfs_header_nritems(right);
1692
1693         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1694         push_space -= leaf_data_end(root, left);
1695
1696         /* make room in the right data area */
1697         data_end = leaf_data_end(root, right);
1698         memmove_extent_buffer(right,
1699                               btrfs_leaf_data(right) + data_end - push_space,
1700                               btrfs_leaf_data(right) + data_end,
1701                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1702
1703         /* copy from the left data area */
1704         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1705                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1706                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1707                      push_space);
1708
1709         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1710                               btrfs_item_nr_offset(0),
1711                               right_nritems * sizeof(struct btrfs_item));
1712
1713         /* copy the items from left to right */
1714         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1715                    btrfs_item_nr_offset(left_nritems - push_items),
1716                    push_items * sizeof(struct btrfs_item));
1717
1718         /* update the item pointers */
1719         right_nritems += push_items;
1720         btrfs_set_header_nritems(right, right_nritems);
1721         push_space = BTRFS_LEAF_DATA_SIZE(root);
1722         for (i = 0; i < right_nritems; i++) {
1723                 item = btrfs_item_nr(right, i);
1724                 if (!right->map_token) {
1725                         map_extent_buffer(right, (unsigned long)item,
1726                                         sizeof(struct btrfs_item),
1727                                         &right->map_token, &right->kaddr,
1728                                         &right->map_start, &right->map_len,
1729                                         KM_USER1);
1730                 }
1731                 push_space -= btrfs_item_size(right, item);
1732                 btrfs_set_item_offset(right, item, push_space);
1733         }
1734
1735         if (right->map_token) {
1736                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1737                 right->map_token = NULL;
1738         }
1739         left_nritems -= push_items;
1740         btrfs_set_header_nritems(left, left_nritems);
1741
1742         if (left_nritems)
1743                 btrfs_mark_buffer_dirty(left);
1744         btrfs_mark_buffer_dirty(right);
1745
1746         btrfs_item_key(right, &disk_key, 0);
1747         btrfs_set_node_key(upper, &disk_key, slot + 1);
1748         btrfs_mark_buffer_dirty(upper);
1749
1750         /* then fixup the leaf pointer in the path */
1751         if (path->slots[0] >= left_nritems) {
1752                 path->slots[0] -= left_nritems;
1753                 free_extent_buffer(path->nodes[0]);
1754                 path->nodes[0] = right;
1755                 path->slots[1] += 1;
1756         } else {
1757                 free_extent_buffer(right);
1758         }
1759         return 0;
1760 }
1761 /*
1762  * push some data in the path leaf to the left, trying to free up at
1763  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1764  */
1765 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1766                           *root, struct btrfs_path *path, int data_size,
1767                           int empty)
1768 {
1769         struct btrfs_disk_key disk_key;
1770         struct extent_buffer *right = path->nodes[0];
1771         struct extent_buffer *left;
1772         int slot;
1773         int i;
1774         int free_space;
1775         int push_space = 0;
1776         int push_items = 0;
1777         struct btrfs_item *item;
1778         u32 old_left_nritems;
1779         u32 right_nritems;
1780         u32 nr;
1781         int ret = 0;
1782         int wret;
1783         u32 this_item_size;
1784         u32 old_left_item_size;
1785
1786         slot = path->slots[1];
1787         if (slot == 0)
1788                 return 1;
1789         if (!path->nodes[1])
1790                 return 1;
1791
1792         right_nritems = btrfs_header_nritems(right);
1793         if (right_nritems == 0) {
1794                 return 1;
1795         }
1796
1797         left = read_tree_block(root, btrfs_node_blockptr(path->nodes[1],
1798                                slot - 1), root->leafsize);
1799         free_space = btrfs_leaf_free_space(root, left);
1800         if (free_space < data_size + sizeof(struct btrfs_item)) {
1801                 free_extent_buffer(left);
1802                 return 1;
1803         }
1804
1805         /* cow and double check */
1806         ret = btrfs_cow_block(trans, root, left,
1807                               path->nodes[1], slot - 1, &left);
1808         if (ret) {
1809                 /* we hit -ENOSPC, but it isn't fatal here */
1810                 free_extent_buffer(left);
1811                 return 1;
1812         }
1813
1814         free_space = btrfs_leaf_free_space(root, left);
1815         if (free_space < data_size + sizeof(struct btrfs_item)) {
1816                 free_extent_buffer(left);
1817                 return 1;
1818         }
1819
1820         if (empty)
1821                 nr = right_nritems;
1822         else
1823                 nr = right_nritems - 1;
1824
1825         for (i = 0; i < nr; i++) {
1826                 item = btrfs_item_nr(right, i);
1827                 if (!right->map_token) {
1828                         map_extent_buffer(right, (unsigned long)item,
1829                                         sizeof(struct btrfs_item),
1830                                         &right->map_token, &right->kaddr,
1831                                         &right->map_start, &right->map_len,
1832                                         KM_USER1);
1833                 }
1834
1835                 if (path->slots[0] == i)
1836                         push_space += data_size + sizeof(*item);
1837
1838                 this_item_size = btrfs_item_size(right, item);
1839                 if (this_item_size + sizeof(*item) + push_space > free_space)
1840                         break;
1841
1842                 push_items++;
1843                 push_space += this_item_size + sizeof(*item);
1844         }
1845
1846         if (right->map_token) {
1847                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1848                 right->map_token = NULL;
1849         }
1850
1851         if (push_items == 0) {
1852                 free_extent_buffer(left);
1853                 return 1;
1854         }
1855         if (!empty && push_items == btrfs_header_nritems(right))
1856                 WARN_ON(1);
1857
1858         /* push data from right to left */
1859         copy_extent_buffer(left, right,
1860                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1861                            btrfs_item_nr_offset(0),
1862                            push_items * sizeof(struct btrfs_item));
1863
1864         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1865                      btrfs_item_offset_nr(right, push_items -1);
1866
1867         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1868                      leaf_data_end(root, left) - push_space,
1869                      btrfs_leaf_data(right) +
1870                      btrfs_item_offset_nr(right, push_items - 1),
1871                      push_space);
1872         old_left_nritems = btrfs_header_nritems(left);
1873         BUG_ON(old_left_nritems < 0);
1874
1875         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1876         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1877                 u32 ioff;
1878
1879                 item = btrfs_item_nr(left, i);
1880                 if (!left->map_token) {
1881                         map_extent_buffer(left, (unsigned long)item,
1882                                         sizeof(struct btrfs_item),
1883                                         &left->map_token, &left->kaddr,
1884                                         &left->map_start, &left->map_len,
1885                                         KM_USER1);
1886                 }
1887
1888                 ioff = btrfs_item_offset(left, item);
1889                 btrfs_set_item_offset(left, item,
1890                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1891         }
1892         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1893         if (left->map_token) {
1894                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1895                 left->map_token = NULL;
1896         }
1897
1898         /* fixup right node */
1899         if (push_items > right_nritems) {
1900                 printk("push items %d nr %u\n", push_items, right_nritems);
1901                 WARN_ON(1);
1902         }
1903
1904         if (push_items < right_nritems) {
1905                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1906                                                   leaf_data_end(root, right);
1907                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
1908                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1909                                       btrfs_leaf_data(right) +
1910                                       leaf_data_end(root, right), push_space);
1911
1912                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1913                               btrfs_item_nr_offset(push_items),
1914                              (btrfs_header_nritems(right) - push_items) *
1915                              sizeof(struct btrfs_item));
1916         }
1917         right_nritems -= push_items;
1918         btrfs_set_header_nritems(right, right_nritems);
1919         push_space = BTRFS_LEAF_DATA_SIZE(root);
1920         for (i = 0; i < right_nritems; i++) {
1921                 item = btrfs_item_nr(right, i);
1922
1923                 if (!right->map_token) {
1924                         map_extent_buffer(right, (unsigned long)item,
1925                                         sizeof(struct btrfs_item),
1926                                         &right->map_token, &right->kaddr,
1927                                         &right->map_start, &right->map_len,
1928                                         KM_USER1);
1929                 }
1930
1931                 push_space = push_space - btrfs_item_size(right, item);
1932                 btrfs_set_item_offset(right, item, push_space);
1933         }
1934         if (right->map_token) {
1935                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1936                 right->map_token = NULL;
1937         }
1938
1939         btrfs_mark_buffer_dirty(left);
1940         if (right_nritems)
1941                 btrfs_mark_buffer_dirty(right);
1942
1943         btrfs_item_key(right, &disk_key, 0);
1944         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1945         if (wret)
1946                 ret = wret;
1947
1948         /* then fixup the leaf pointer in the path */
1949         if (path->slots[0] < push_items) {
1950                 path->slots[0] += old_left_nritems;
1951                 free_extent_buffer(path->nodes[0]);
1952                 path->nodes[0] = left;
1953                 path->slots[1] -= 1;
1954         } else {
1955                 free_extent_buffer(left);
1956                 path->slots[0] -= push_items;
1957         }
1958         BUG_ON(path->slots[0] < 0);
1959         return ret;
1960 }
1961
1962 /*
1963  * split the path's leaf in two, making sure there is at least data_size
1964  * available for the resulting leaf level of the path.
1965  *
1966  * returns 0 if all went well and < 0 on failure.
1967  */
1968 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
1969                       *root, struct btrfs_key *ins_key,
1970                       struct btrfs_path *path, int data_size, int extend)
1971 {
1972         u64 root_gen;
1973         struct extent_buffer *l;
1974         u32 nritems;
1975         int mid;
1976         int slot;
1977         struct extent_buffer *right;
1978         int space_needed = data_size + sizeof(struct btrfs_item);
1979         int data_copy_size;
1980         int rt_data_off;
1981         int i;
1982         int ret = 0;
1983         int wret;
1984         int double_split;
1985         int num_doubles = 0;
1986         struct btrfs_disk_key disk_key;
1987
1988         if (extend)
1989                 space_needed = data_size;
1990
1991         if (root->ref_cows)
1992                 root_gen = trans->transid;
1993         else
1994                 root_gen = 0;
1995
1996         /* first try to make some room by pushing left and right */
1997         if (ins_key->type != BTRFS_DIR_ITEM_KEY) {
1998                 wret = push_leaf_right(trans, root, path, data_size, 0);
1999                 if (wret < 0) {
2000                         return wret;
2001                 }
2002                 if (wret) {
2003                         wret = push_leaf_left(trans, root, path, data_size, 0);
2004                         if (wret < 0)
2005                                 return wret;
2006                 }
2007                 l = path->nodes[0];
2008
2009                 /* did the pushes work? */
2010                 if (btrfs_leaf_free_space(root, l) >= space_needed)
2011                         return 0;
2012         }
2013
2014         if (!path->nodes[1]) {
2015                 ret = insert_new_root(trans, root, path, 1);
2016                 if (ret)
2017                         return ret;
2018         }
2019 again:
2020         double_split = 0;
2021         l = path->nodes[0];
2022         slot = path->slots[0];
2023         nritems = btrfs_header_nritems(l);
2024         mid = (nritems + 1)/ 2;
2025
2026         btrfs_item_key(l, &disk_key, 0);
2027
2028         right = __btrfs_alloc_free_block(trans, root, root->leafsize,
2029                                          root->root_key.objectid,
2030                                          root_gen, disk_key.objectid, 0,
2031                                          l->start, 0);
2032         if (IS_ERR(right)) {
2033                 BUG_ON(1);
2034                 return PTR_ERR(right);
2035         }
2036
2037         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2038         btrfs_set_header_bytenr(right, right->start);
2039         btrfs_set_header_generation(right, trans->transid);
2040         btrfs_set_header_owner(right, root->root_key.objectid);
2041         btrfs_set_header_level(right, 0);
2042         write_extent_buffer(right, root->fs_info->fsid,
2043                             (unsigned long)btrfs_header_fsid(right),
2044                             BTRFS_FSID_SIZE);
2045         if (mid <= slot) {
2046                 if (nritems == 1 ||
2047                     leaf_space_used(l, mid, nritems - mid) + space_needed >
2048                         BTRFS_LEAF_DATA_SIZE(root)) {
2049                         if (slot >= nritems) {
2050                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2051                                 btrfs_set_header_nritems(right, 0);
2052                                 wret = insert_ptr(trans, root, path,
2053                                                   &disk_key, right->start,
2054                                                   path->slots[1] + 1, 1);
2055                                 if (wret)
2056                                         ret = wret;
2057                                 free_extent_buffer(path->nodes[0]);
2058                                 path->nodes[0] = right;
2059                                 path->slots[0] = 0;
2060                                 path->slots[1] += 1;
2061                                 btrfs_mark_buffer_dirty(right);
2062                                 return ret;
2063                         }
2064                         mid = slot;
2065                         if (mid != nritems &&
2066                             leaf_space_used(l, mid, nritems - mid) +
2067                             space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2068                                 double_split = 1;
2069                         }
2070                 }
2071         } else {
2072                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
2073                         BTRFS_LEAF_DATA_SIZE(root)) {
2074                         if (!extend && slot == 0) {
2075                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2076                                 btrfs_set_header_nritems(right, 0);
2077                                 wret = insert_ptr(trans, root, path,
2078                                                   &disk_key,
2079                                                   right->start,
2080                                                   path->slots[1], 1);
2081                                 if (wret)
2082                                         ret = wret;
2083                                 free_extent_buffer(path->nodes[0]);
2084                                 path->nodes[0] = right;
2085                                 path->slots[0] = 0;
2086                                 if (path->slots[1] == 0) {
2087                                         wret = fixup_low_keys(trans, root,
2088                                                    path, &disk_key, 1);
2089                                         if (wret)
2090                                                 ret = wret;
2091                                 }
2092                                 btrfs_mark_buffer_dirty(right);
2093                                 return ret;
2094                         } else if (extend && slot == 0) {
2095                                 mid = 1;
2096                         } else {
2097                                 mid = slot;
2098                                 if (mid != nritems &&
2099                                     leaf_space_used(l, mid, nritems - mid) +
2100                                     space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2101                                         double_split = 1;
2102                                 }
2103                         }
2104                 }
2105         }
2106         nritems = nritems - mid;
2107         btrfs_set_header_nritems(right, nritems);
2108         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2109
2110         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2111                            btrfs_item_nr_offset(mid),
2112                            nritems * sizeof(struct btrfs_item));
2113
2114         copy_extent_buffer(right, l,
2115                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2116                      data_copy_size, btrfs_leaf_data(l) +
2117                      leaf_data_end(root, l), data_copy_size);
2118
2119         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2120                       btrfs_item_end_nr(l, mid);
2121
2122         for (i = 0; i < nritems; i++) {
2123                 struct btrfs_item *item = btrfs_item_nr(right, i);
2124                 u32 ioff;
2125
2126                 if (!right->map_token) {
2127                         map_extent_buffer(right, (unsigned long)item,
2128                                         sizeof(struct btrfs_item),
2129                                         &right->map_token, &right->kaddr,
2130                                         &right->map_start, &right->map_len,
2131                                         KM_USER1);
2132                 }
2133
2134                 ioff = btrfs_item_offset(right, item);
2135                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2136         }
2137
2138         if (right->map_token) {
2139                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2140                 right->map_token = NULL;
2141         }
2142
2143         btrfs_set_header_nritems(l, mid);
2144         ret = 0;
2145         btrfs_item_key(right, &disk_key, 0);
2146         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2147                           path->slots[1] + 1, 1);
2148         if (wret)
2149                 ret = wret;
2150
2151         btrfs_mark_buffer_dirty(right);
2152         btrfs_mark_buffer_dirty(l);
2153         BUG_ON(path->slots[0] != slot);
2154
2155         if (mid <= slot) {
2156                 free_extent_buffer(path->nodes[0]);
2157                 path->nodes[0] = right;
2158                 path->slots[0] -= mid;
2159                 path->slots[1] += 1;
2160         } else
2161                 free_extent_buffer(right);
2162
2163         BUG_ON(path->slots[0] < 0);
2164
2165         if (double_split) {
2166                 BUG_ON(num_doubles != 0);
2167                 num_doubles++;
2168                 goto again;
2169         }
2170         return ret;
2171 }
2172
2173 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2174                         struct btrfs_root *root,
2175                         struct btrfs_path *path,
2176                         u32 new_size, int from_end)
2177 {
2178         int ret = 0;
2179         int slot;
2180         int slot_orig;
2181         struct extent_buffer *leaf;
2182         struct btrfs_item *item;
2183         u32 nritems;
2184         unsigned int data_end;
2185         unsigned int old_data_start;
2186         unsigned int old_size;
2187         unsigned int size_diff;
2188         int i;
2189
2190         slot_orig = path->slots[0];
2191         leaf = path->nodes[0];
2192         slot = path->slots[0];
2193
2194         old_size = btrfs_item_size_nr(leaf, slot);
2195         if (old_size == new_size)
2196                 return 0;
2197
2198         nritems = btrfs_header_nritems(leaf);
2199         data_end = leaf_data_end(root, leaf);
2200
2201         old_data_start = btrfs_item_offset_nr(leaf, slot);
2202
2203         size_diff = old_size - new_size;
2204
2205         BUG_ON(slot < 0);
2206         BUG_ON(slot >= nritems);
2207
2208         /*
2209          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2210          */
2211         /* first correct the data pointers */
2212         for (i = slot; i < nritems; i++) {
2213                 u32 ioff;
2214                 item = btrfs_item_nr(leaf, i);
2215
2216                 if (!leaf->map_token) {
2217                         map_extent_buffer(leaf, (unsigned long)item,
2218                                         sizeof(struct btrfs_item),
2219                                         &leaf->map_token, &leaf->kaddr,
2220                                         &leaf->map_start, &leaf->map_len,
2221                                         KM_USER1);
2222                 }
2223
2224                 ioff = btrfs_item_offset(leaf, item);
2225                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2226         }
2227
2228         if (leaf->map_token) {
2229                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2230                 leaf->map_token = NULL;
2231         }
2232
2233         /* shift the data */
2234         if (from_end) {
2235                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2236                               data_end + size_diff, btrfs_leaf_data(leaf) +
2237                               data_end, old_data_start + new_size - data_end);
2238         } else {
2239                 struct btrfs_disk_key disk_key;
2240                 u64 offset;
2241
2242                 btrfs_item_key(leaf, &disk_key, slot);
2243
2244                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2245                         unsigned long ptr;
2246                         struct btrfs_file_extent_item *fi;
2247
2248                         fi = btrfs_item_ptr(leaf, slot,
2249                                             struct btrfs_file_extent_item);
2250                         fi = (struct btrfs_file_extent_item *)(
2251                              (unsigned long)fi - size_diff);
2252
2253                         if (btrfs_file_extent_type(leaf, fi) ==
2254                             BTRFS_FILE_EXTENT_INLINE) {
2255                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2256                                 memmove_extent_buffer(leaf, ptr,
2257                                         (unsigned long)fi,
2258                                         offsetof(struct btrfs_file_extent_item,
2259                                                  disk_bytenr));
2260                         }
2261                 }
2262
2263                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2264                               data_end + size_diff, btrfs_leaf_data(leaf) +
2265                               data_end, old_data_start - data_end);
2266
2267                 offset = btrfs_disk_key_offset(&disk_key);
2268                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2269                 btrfs_set_item_key(leaf, &disk_key, slot);
2270                 if (slot == 0)
2271                         fixup_low_keys(trans, root, path, &disk_key, 1);
2272         }
2273
2274         item = btrfs_item_nr(leaf, slot);
2275         btrfs_set_item_size(leaf, item, new_size);
2276         btrfs_mark_buffer_dirty(leaf);
2277
2278         ret = 0;
2279         if (btrfs_leaf_free_space(root, leaf) < 0) {
2280                 btrfs_print_leaf(root, leaf);
2281                 BUG();
2282         }
2283         return ret;
2284 }
2285
2286 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2287                       struct btrfs_root *root, struct btrfs_path *path,
2288                       u32 data_size)
2289 {
2290         int ret = 0;
2291         int slot;
2292         int slot_orig;
2293         struct extent_buffer *leaf;
2294         struct btrfs_item *item;
2295         u32 nritems;
2296         unsigned int data_end;
2297         unsigned int old_data;
2298         unsigned int old_size;
2299         int i;
2300
2301         slot_orig = path->slots[0];
2302         leaf = path->nodes[0];
2303
2304         nritems = btrfs_header_nritems(leaf);
2305         data_end = leaf_data_end(root, leaf);
2306
2307         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2308                 btrfs_print_leaf(root, leaf);
2309                 BUG();
2310         }
2311         slot = path->slots[0];
2312         old_data = btrfs_item_end_nr(leaf, slot);
2313
2314         BUG_ON(slot < 0);
2315         if (slot >= nritems) {
2316                 btrfs_print_leaf(root, leaf);
2317                 printk("slot %d too large, nritems %d\n", slot, nritems);
2318                 BUG_ON(1);
2319         }
2320
2321         /*
2322          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2323          */
2324         /* first correct the data pointers */
2325         for (i = slot; i < nritems; i++) {
2326                 u32 ioff;
2327                 item = btrfs_item_nr(leaf, i);
2328
2329                 if (!leaf->map_token) {
2330                         map_extent_buffer(leaf, (unsigned long)item,
2331                                         sizeof(struct btrfs_item),
2332                                         &leaf->map_token, &leaf->kaddr,
2333                                         &leaf->map_start, &leaf->map_len,
2334                                         KM_USER1);
2335                 }
2336                 ioff = btrfs_item_offset(leaf, item);
2337                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2338         }
2339
2340         if (leaf->map_token) {
2341                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2342                 leaf->map_token = NULL;
2343         }
2344
2345         /* shift the data */
2346         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2347                       data_end - data_size, btrfs_leaf_data(leaf) +
2348                       data_end, old_data - data_end);
2349
2350         data_end = old_data;
2351         old_size = btrfs_item_size_nr(leaf, slot);
2352         item = btrfs_item_nr(leaf, slot);
2353         btrfs_set_item_size(leaf, item, old_size + data_size);
2354         btrfs_mark_buffer_dirty(leaf);
2355
2356         ret = 0;
2357         if (btrfs_leaf_free_space(root, leaf) < 0) {
2358                 btrfs_print_leaf(root, leaf);
2359                 BUG();
2360         }
2361         return ret;
2362 }
2363
2364 /*
2365  * Given a key and some data, insert an item into the tree.
2366  * This does all the path init required, making room in the tree if needed.
2367  */
2368 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2369                             struct btrfs_root *root,
2370                             struct btrfs_path *path,
2371                             struct btrfs_key *cpu_key, u32 *data_size,
2372                             int nr)
2373 {
2374         struct extent_buffer *leaf;
2375         struct btrfs_item *item;
2376         int ret = 0;
2377         int slot;
2378         int slot_orig;
2379         int i;
2380         u32 nritems;
2381         u32 total_size = 0;
2382         u32 total_data = 0;
2383         unsigned int data_end;
2384         struct btrfs_disk_key disk_key;
2385
2386         for (i = 0; i < nr; i++) {
2387                 total_data += data_size[i];
2388         }
2389
2390         /* create a root if there isn't one */
2391         if (!root->node)
2392                 BUG();
2393
2394         total_size = total_data + (nr - 1) * sizeof(struct btrfs_item);
2395         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2396         if (ret == 0) {
2397                 return -EEXIST;
2398         }
2399         if (ret < 0)
2400                 goto out;
2401
2402         slot_orig = path->slots[0];
2403         leaf = path->nodes[0];
2404
2405         nritems = btrfs_header_nritems(leaf);
2406         data_end = leaf_data_end(root, leaf);
2407
2408         if (btrfs_leaf_free_space(root, leaf) <
2409             sizeof(struct btrfs_item) + total_size) {
2410                 btrfs_print_leaf(root, leaf);
2411                 printk("not enough freespace need %u have %d\n",
2412                        total_size, btrfs_leaf_free_space(root, leaf));
2413                 BUG();
2414         }
2415
2416         slot = path->slots[0];
2417         BUG_ON(slot < 0);
2418
2419         if (slot != nritems) {
2420                 int i;
2421                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2422
2423                 if (old_data < data_end) {
2424                         btrfs_print_leaf(root, leaf);
2425                         printk("slot %d old_data %d data_end %d\n",
2426                                slot, old_data, data_end);
2427                         BUG_ON(1);
2428                 }
2429                 /*
2430                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2431                  */
2432                 /* first correct the data pointers */
2433                 WARN_ON(leaf->map_token);
2434                 for (i = slot; i < nritems; i++) {
2435                         u32 ioff;
2436
2437                         item = btrfs_item_nr(leaf, i);
2438                         if (!leaf->map_token) {
2439                                 map_extent_buffer(leaf, (unsigned long)item,
2440                                         sizeof(struct btrfs_item),
2441                                         &leaf->map_token, &leaf->kaddr,
2442                                         &leaf->map_start, &leaf->map_len,
2443                                         KM_USER1);
2444                         }
2445
2446                         ioff = btrfs_item_offset(leaf, item);
2447                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2448                 }
2449                 if (leaf->map_token) {
2450                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2451                         leaf->map_token = NULL;
2452                 }
2453
2454                 /* shift the items */
2455                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2456                               btrfs_item_nr_offset(slot),
2457                               (nritems - slot) * sizeof(struct btrfs_item));
2458
2459                 /* shift the data */
2460                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2461                               data_end - total_data, btrfs_leaf_data(leaf) +
2462                               data_end, old_data - data_end);
2463                 data_end = old_data;
2464         }
2465
2466         /* setup the item for the new data */
2467         for (i = 0; i < nr; i++) {
2468                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2469                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2470                 item = btrfs_item_nr(leaf, slot + i);
2471                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2472                 data_end -= data_size[i];
2473                 btrfs_set_item_size(leaf, item, data_size[i]);
2474         }
2475         btrfs_set_header_nritems(leaf, nritems + nr);
2476         btrfs_mark_buffer_dirty(leaf);
2477
2478         ret = 0;
2479         if (slot == 0) {
2480                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2481                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2482         }
2483
2484         if (btrfs_leaf_free_space(root, leaf) < 0) {
2485                 btrfs_print_leaf(root, leaf);
2486                 BUG();
2487         }
2488
2489 out:
2490         return ret;
2491 }
2492
2493 /*
2494  * Given a key and some data, insert an item into the tree.
2495  * This does all the path init required, making room in the tree if needed.
2496  */
2497 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2498                       *root, struct btrfs_key *cpu_key, void *data, u32
2499                       data_size)
2500 {
2501         int ret = 0;
2502         struct btrfs_path *path;
2503         struct extent_buffer *leaf;
2504         unsigned long ptr;
2505
2506         path = btrfs_alloc_path();
2507         BUG_ON(!path);
2508         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2509         if (!ret) {
2510                 leaf = path->nodes[0];
2511                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2512                 write_extent_buffer(leaf, data, ptr, data_size);
2513                 btrfs_mark_buffer_dirty(leaf);
2514         }
2515         btrfs_free_path(path);
2516         return ret;
2517 }
2518
2519 /*
2520  * delete the pointer from a given node.
2521  *
2522  * If the delete empties a node, the node is removed from the tree,
2523  * continuing all the way the root if required.  The root is converted into
2524  * a leaf if all the nodes are emptied.
2525  */
2526 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2527                    struct btrfs_path *path, int level, int slot)
2528 {
2529         struct extent_buffer *parent = path->nodes[level];
2530         u32 nritems;
2531         int ret = 0;
2532         int wret;
2533
2534         nritems = btrfs_header_nritems(parent);
2535         if (slot != nritems -1) {
2536                 memmove_extent_buffer(parent,
2537                               btrfs_node_key_ptr_offset(slot),
2538                               btrfs_node_key_ptr_offset(slot + 1),
2539                               sizeof(struct btrfs_key_ptr) *
2540                               (nritems - slot - 1));
2541         }
2542         nritems--;
2543         btrfs_set_header_nritems(parent, nritems);
2544         if (nritems == 0 && parent == root->node) {
2545                 BUG_ON(btrfs_header_level(root->node) != 1);
2546                 /* just turn the root into a leaf and break */
2547                 btrfs_set_header_level(root->node, 0);
2548         } else if (slot == 0) {
2549                 struct btrfs_disk_key disk_key;
2550
2551                 btrfs_node_key(parent, &disk_key, 0);
2552                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2553                 if (wret)
2554                         ret = wret;
2555         }
2556         btrfs_mark_buffer_dirty(parent);
2557         return ret;
2558 }
2559
2560 /*
2561  * delete the item at the leaf level in path.  If that empties
2562  * the leaf, remove it from the tree
2563  */
2564 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2565                     struct btrfs_path *path, int slot, int nr)
2566 {
2567         struct extent_buffer *leaf;
2568         struct btrfs_item *item;
2569         int last_off;
2570         int dsize = 0;
2571         int ret = 0;
2572         int wret;
2573         int i;
2574         u32 nritems;
2575
2576         leaf = path->nodes[0];
2577         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2578
2579         for (i = 0; i < nr; i++)
2580                 dsize += btrfs_item_size_nr(leaf, slot + i);
2581
2582         nritems = btrfs_header_nritems(leaf);
2583
2584         if (slot + nr != nritems) {
2585                 int i;
2586                 int data_end = leaf_data_end(root, leaf);
2587
2588                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2589                               data_end + dsize,
2590                               btrfs_leaf_data(leaf) + data_end,
2591                               last_off - data_end);
2592
2593                 for (i = slot + nr; i < nritems; i++) {
2594                         u32 ioff;
2595
2596                         item = btrfs_item_nr(leaf, i);
2597                         if (!leaf->map_token) {
2598                                 map_extent_buffer(leaf, (unsigned long)item,
2599                                         sizeof(struct btrfs_item),
2600                                         &leaf->map_token, &leaf->kaddr,
2601                                         &leaf->map_start, &leaf->map_len,
2602                                         KM_USER1);
2603                         }
2604                         ioff = btrfs_item_offset(leaf, item);
2605                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2606                 }
2607
2608                 if (leaf->map_token) {
2609                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2610                         leaf->map_token = NULL;
2611                 }
2612
2613                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2614                               btrfs_item_nr_offset(slot + nr),
2615                               sizeof(struct btrfs_item) *
2616                               (nritems - slot - nr));
2617         }
2618         btrfs_set_header_nritems(leaf, nritems - nr);
2619         nritems -= nr;
2620
2621         /* delete the leaf if we've emptied it */
2622         if (nritems == 0) {
2623                 if (leaf == root->node) {
2624                         btrfs_set_header_level(leaf, 0);
2625                 } else {
2626                         u64 root_gen = btrfs_header_generation(path->nodes[1]);
2627                         clean_tree_block(trans, root, leaf);
2628                         wait_on_tree_block_writeback(root, leaf);
2629                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
2630                         if (wret)
2631                                 ret = wret;
2632                         wret = btrfs_free_extent(trans, root,
2633                                          leaf->start, leaf->len,
2634                                          btrfs_header_owner(path->nodes[1]),
2635                                          root_gen, 0, 0, 1);
2636                         if (wret)
2637                                 ret = wret;
2638                 }
2639         } else {
2640                 int used = leaf_space_used(leaf, 0, nritems);
2641                 if (slot == 0) {
2642                         struct btrfs_disk_key disk_key;
2643
2644                         btrfs_item_key(leaf, &disk_key, 0);
2645                         wret = fixup_low_keys(trans, root, path,
2646                                               &disk_key, 1);
2647                         if (wret)
2648                                 ret = wret;
2649                 }
2650
2651                 /* delete the leaf if it is mostly empty */
2652                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2653                         /* push_leaf_left fixes the path.
2654                          * make sure the path still points to our leaf
2655                          * for possible call to del_ptr below
2656                          */
2657                         slot = path->slots[1];
2658                         extent_buffer_get(leaf);
2659
2660                         wret = push_leaf_left(trans, root, path, 1, 1);
2661                         if (wret < 0 && wret != -ENOSPC)
2662                                 ret = wret;
2663
2664                         if (path->nodes[0] == leaf &&
2665                             btrfs_header_nritems(leaf)) {
2666                                 wret = push_leaf_right(trans, root, path, 1, 1);
2667                                 if (wret < 0 && wret != -ENOSPC)
2668                                         ret = wret;
2669                         }
2670
2671                         if (btrfs_header_nritems(leaf) == 0) {
2672                                 u64 root_gen;
2673                                 u64 bytenr = leaf->start;
2674                                 u32 blocksize = leaf->len;
2675
2676                                 root_gen = btrfs_header_generation(
2677                                                            path->nodes[1]);
2678
2679                                 clean_tree_block(trans, root, leaf);
2680                                 wait_on_tree_block_writeback(root, leaf);
2681
2682                                 wret = del_ptr(trans, root, path, 1, slot);
2683                                 if (wret)
2684                                         ret = wret;
2685
2686                                 free_extent_buffer(leaf);
2687                                 wret = btrfs_free_extent(trans, root, bytenr,
2688                                              blocksize,
2689                                              btrfs_header_owner(path->nodes[1]),
2690                                              root_gen, 0, 0, 1);
2691                                 if (wret)
2692                                         ret = wret;
2693                         } else {
2694                                 btrfs_mark_buffer_dirty(leaf);
2695                                 free_extent_buffer(leaf);
2696                         }
2697                 } else {
2698                         btrfs_mark_buffer_dirty(leaf);
2699                 }
2700         }
2701         return ret;
2702 }
2703
2704 /*
2705  * walk up the tree as far as required to find the previous leaf.
2706  * returns 0 if it found something or 1 if there are no lesser leaves.
2707  * returns < 0 on io errors.
2708  */
2709 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2710 {
2711         u64 bytenr;
2712         int slot;
2713         int level = 1;
2714         struct extent_buffer *c;
2715         struct extent_buffer *next = NULL;
2716
2717         while(level < BTRFS_MAX_LEVEL) {
2718                 if (!path->nodes[level])
2719                         return 1;
2720
2721                 slot = path->slots[level];
2722                 c = path->nodes[level];
2723                 if (slot == 0) {
2724                         level++;
2725                         if (level == BTRFS_MAX_LEVEL)
2726                                 return 1;
2727                         continue;
2728                 }
2729                 slot--;
2730
2731                 bytenr = btrfs_node_blockptr(c, slot);
2732                 if (next)
2733                         free_extent_buffer(next);
2734
2735                 next = read_tree_block(root, bytenr,
2736                                        btrfs_level_size(root, level - 1));
2737                 break;
2738         }
2739         path->slots[level] = slot;
2740         while(1) {
2741                 level--;
2742                 c = path->nodes[level];
2743                 free_extent_buffer(c);
2744                 slot = btrfs_header_nritems(next);
2745                 if (slot != 0)
2746                         slot--;
2747                 path->nodes[level] = next;
2748                 path->slots[level] = slot;
2749                 if (!level)
2750                         break;
2751                 next = read_tree_block(root, btrfs_node_blockptr(next, slot),
2752                                        btrfs_level_size(root, level - 1));
2753         }
2754         return 0;
2755 }
2756
2757 /*
2758  * walk up the tree as far as required to find the next leaf.
2759  * returns 0 if it found something or 1 if there are no greater leaves.
2760  * returns < 0 on io errors.
2761  */
2762 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2763 {
2764         int slot;
2765         int level = 1;
2766         u64 bytenr;
2767         struct extent_buffer *c;
2768         struct extent_buffer *next = NULL;
2769
2770         while(level < BTRFS_MAX_LEVEL) {
2771                 if (!path->nodes[level])
2772                         return 1;
2773
2774                 slot = path->slots[level] + 1;
2775                 c = path->nodes[level];
2776                 if (slot >= btrfs_header_nritems(c)) {
2777                         level++;
2778                         if (level == BTRFS_MAX_LEVEL)
2779                                 return 1;
2780                         continue;
2781                 }
2782
2783                 bytenr = btrfs_node_blockptr(c, slot);
2784                 if (next)
2785                         free_extent_buffer(next);
2786
2787                 if (path->reada)
2788                         reada_for_search(root, path, level, slot, 0);
2789
2790                 next = read_tree_block(root, bytenr,
2791                                        btrfs_level_size(root, level -1));
2792                 break;
2793         }
2794         path->slots[level] = slot;
2795         while(1) {
2796                 level--;
2797                 c = path->nodes[level];
2798                 free_extent_buffer(c);
2799                 path->nodes[level] = next;
2800                 path->slots[level] = 0;
2801                 if (!level)
2802                         break;
2803                 if (path->reada)
2804                         reada_for_search(root, path, level, 0, 0);
2805                 next = read_tree_block(root, btrfs_node_blockptr(next, 0),
2806                                        btrfs_level_size(root, level - 1));
2807         }
2808         return 0;
2809 }
2810
2811 int btrfs_previous_item(struct btrfs_root *root,
2812                         struct btrfs_path *path, u64 min_objectid,
2813                         int type)
2814 {
2815         struct btrfs_key found_key;
2816         struct extent_buffer *leaf;
2817         int ret;
2818
2819         while(1) {
2820                 if (path->slots[0] == 0) {
2821                         ret = btrfs_prev_leaf(root, path);
2822                         if (ret != 0)
2823                                 return ret;
2824                 } else {
2825                         path->slots[0]--;
2826                 }
2827                 leaf = path->nodes[0];
2828                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2829                 if (found_key.type == type)
2830                         return 0;
2831         }
2832         return 1;
2833 }
2834