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