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, int empty);
  33 static int balance_node_right(struct btrfs_trans_handle *trans,
  34                               struct btrfs_root *root,
  35                               struct extent_buffer *dst_buf,
  36                               struct extent_buffer *src_buf);
  37 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  38                    struct btrfs_path *path, int level, int slot);
  39
  40 inline void btrfs_init_path(struct btrfs_path *p)
  41 {
  42         memset(p, 0, sizeof(*p));
  43 }
  44
  45 struct btrfs_path *btrfs_alloc_path(void)
  46 {
  47         struct btrfs_path *path;
  48         path = kmem_cache_alloc(btrfs_path_cachep, GFP_NOFS);
  49         if (path) {
  50                 btrfs_init_path(path);
  51                 path->reada = 1;
  52         }
  53         return path;
  54 }
  55
  56 void btrfs_free_path(struct btrfs_path *p)
  57 {
  58         btrfs_release_path(NULL, p);
  59         kmem_cache_free(btrfs_path_cachep, p);
  60 }
  61
  62 void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
  63 {
  64         int i;
  65         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  66                 if (!p->nodes[i])
  67                         break;
  68                 free_extent_buffer(p->nodes[i]);
  69         }
  70         memset(p, 0, sizeof(*p));
  71 }
  72
  73 static void add_root_to_dirty_list(struct btrfs_root *root)
  74 {
  75         if (root->track_dirty && list_empty(&root->dirty_list)) {
  76                 list_add(&root->dirty_list,
  77                          &root->fs_info->dirty_cowonly_roots);
  78         }
  79 }
  80
  81 int btrfs_copy_root(struct btrfs_trans_handle *trans,
  82                       struct btrfs_root *root,
  83                       struct extent_buffer *buf,
  84                       struct extent_buffer **cow_ret, u64 new_root_objectid)
  85 {
  86         struct extent_buffer *cow;
  87         u32 nritems;
  88         int ret = 0;
  89         int level;
  90         struct btrfs_key first_key;
  91         struct btrfs_root *new_root;
  92
  93         new_root = kmalloc(sizeof(*new_root), GFP_NOFS);
  94         if (!new_root)
  95                 return -ENOMEM;
  96
  97         memcpy(new_root, root, sizeof(*new_root));
  98         new_root->root_key.objectid = new_root_objectid;
  99
 100         WARN_ON(root->ref_cows && trans->transid !=
 101                 root->fs_info->running_transaction->transid);
 102         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 103
 104         level = btrfs_header_level(buf);
 105         nritems = btrfs_header_nritems(buf);
 106         if (nritems) {
 107                 if (level == 0)
 108                         btrfs_item_key_to_cpu(buf, &first_key, 0);
 109                 else
 110                         btrfs_node_key_to_cpu(buf, &first_key, 0);
 111         } else {
 112                 first_key.objectid = 0;
 113         }
 114         cow = __btrfs_alloc_free_block(trans, new_root, buf->len,
 115                                        new_root_objectid,
 116                                        trans->transid, first_key.objectid,
 117                                        level, buf->start, 0);
 118         if (IS_ERR(cow)) {
 119                 kfree(new_root);
 120                 return PTR_ERR(cow);
 121         }
 122
 123         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 124         btrfs_set_header_bytenr(cow, cow->start);
 125         btrfs_set_header_generation(cow, trans->transid);
 126         btrfs_set_header_owner(cow, new_root_objectid);
 127         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN);
 128
 129         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 130         ret = btrfs_inc_ref(trans, new_root, buf);
 131         kfree(new_root);
 132
 133         if (ret)
 134                 return ret;
 135
 136         btrfs_mark_buffer_dirty(cow);
 137         *cow_ret = cow;
 138         return 0;
 139 }
 140
 141 int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 142                              struct btrfs_root *root,
 143                              struct extent_buffer *buf,
 144                              struct extent_buffer *parent, int parent_slot,
 145                              struct extent_buffer **cow_ret,
 146                              u64 search_start, u64 empty_size)
 147 {
 148         u64 root_gen;
 149         struct extent_buffer *cow;
 150         u32 nritems;
 151         int ret = 0;
 152         int different_trans = 0;
 153         int level;
 154         struct btrfs_key first_key;
 155
 156         if (root->ref_cows) {
 157                 root_gen = trans->transid;
 158         } else {
 159                 root_gen = 0;
 160         }
 161         WARN_ON(root->ref_cows && trans->transid !=
 162                 root->fs_info->running_transaction->transid);
 163         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 164
 165         level = btrfs_header_level(buf);
 166         nritems = btrfs_header_nritems(buf);
 167         if (nritems) {
 168                 if (level == 0)
 169                         btrfs_item_key_to_cpu(buf, &first_key, 0);
 170                 else
 171                         btrfs_node_key_to_cpu(buf, &first_key, 0);
 172         } else {
 173                 first_key.objectid = 0;
 174         }
 175         cow = __btrfs_alloc_free_block(trans, root, buf->len,
 176                                      root->root_key.objectid,
 177                                      root_gen, first_key.objectid, level,
 178                                      search_start, empty_size);
 179         if (IS_ERR(cow))
 180                 return PTR_ERR(cow);
 181
 182         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 183         btrfs_set_header_bytenr(cow, cow->start);
 184         btrfs_set_header_generation(cow, trans->transid);
 185         btrfs_set_header_owner(cow, root->root_key.objectid);
 186         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN);
 187
 188         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 189         if (btrfs_header_generation(buf) != trans->transid) {
 190                 different_trans = 1;
 191                 ret = btrfs_inc_ref(trans, root, buf);
 192                 if (ret)
 193                         return ret;
 194         } else {
 195                 clean_tree_block(trans, root, buf);
 196         }
 197
 198         if (buf == root->node) {
 199                 root_gen = btrfs_header_generation(buf);
 200                 root->node = cow;
 201                 extent_buffer_get(cow);
 202                 if (buf != root->commit_root) {
 203                         btrfs_free_extent(trans, root, buf->start,
 204                                           buf->len, root->root_key.objectid,
 205                                           root_gen, 0, 0, 1);
 206                 }
 207                 free_extent_buffer(buf);
 208                 add_root_to_dirty_list(root);
 209         } else {
 210                 root_gen = btrfs_header_generation(parent);
 211                 btrfs_set_node_blockptr(parent, parent_slot,
 212                                         cow->start);
 213                 WARN_ON(trans->transid == 0);
 214                 btrfs_set_node_ptr_generation(parent, parent_slot,
 215                                               trans->transid);
 216                 btrfs_mark_buffer_dirty(parent);
 217                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 218                 btrfs_free_extent(trans, root, buf->start, buf->len,
 219                                   btrfs_header_owner(parent), root_gen,
 220                                   0, 0, 1);
 221         }
 222         free_extent_buffer(buf);
 223         btrfs_mark_buffer_dirty(cow);
 224         *cow_ret = cow;
 225         return 0;
 226 }
 227
 228 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 229                     struct btrfs_root *root, struct extent_buffer *buf,
 230                     struct extent_buffer *parent, int parent_slot,
 231                     struct extent_buffer **cow_ret)
 232 {
 233         u64 search_start;
 234         u64 header_trans;
 235         int ret;
 236
 237         if (trans->transaction != root->fs_info->running_transaction) {
 238                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 239                        root->fs_info->running_transaction->transid);
 240                 WARN_ON(1);
 241         }
 242         if (trans->transid != root->fs_info->generation) {
 243                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 244                        root->fs_info->generation);
 245                 WARN_ON(1);
 246         }
 247
 248         header_trans = btrfs_header_generation(buf);
 249         spin_lock(&root->fs_info->hash_lock);
 250         if (header_trans == trans->transid &&
 251             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
 252                 *cow_ret = buf;
 253                 spin_unlock(&root->fs_info->hash_lock);
 254                 return 0;
 255         }
 256         spin_unlock(&root->fs_info->hash_lock);
 257         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
 258         ret = __btrfs_cow_block(trans, root, buf, parent,
 259                                  parent_slot, cow_ret, search_start, 0);
 260         return ret;
 261 }
 262
 263 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 264 {
 265         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 266                 return 1;
 267         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 268                 return 1;
 269         return 0;
 270 }
 271
 272 /*
 273  * compare two keys in a memcmp fashion
 274  */
 275 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 276 {
 277         struct btrfs_key k1;
 278
 279         btrfs_disk_key_to_cpu(&k1, disk);
 280
 281         if (k1.objectid > k2->objectid)
 282                 return 1;
 283         if (k1.objectid < k2->objectid)
 284                 return -1;
 285         if (k1.type > k2->type)
 286                 return 1;
 287         if (k1.type < k2->type)
 288                 return -1;
 289         if (k1.offset > k2->offset)
 290                 return 1;
 291         if (k1.offset < k2->offset)
 292                 return -1;
 293         return 0;
 294 }
 295
 296
 297 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 298                        struct btrfs_root *root, struct extent_buffer *parent,
 299                        int start_slot, int cache_only, u64 *last_ret,
 300                        struct btrfs_key *progress)
 301 {
 302         struct extent_buffer *cur;
 303         struct extent_buffer *tmp;
 304         u64 blocknr;
 305         u64 search_start = *last_ret;
 306         u64 last_block = 0;
 307         u64 other;
 308         u32 parent_nritems;
 309         int end_slot;
 310         int i;
 311         int err = 0;
 312         int parent_level;
 313         int uptodate;
 314         u32 blocksize;
 315         int progress_passed = 0;
 316         struct btrfs_disk_key disk_key;
 317
 318         parent_level = btrfs_header_level(parent);
 319         if (cache_only && parent_level != 1)
 320                 return 0;
 321
 322         if (trans->transaction != root->fs_info->running_transaction) {
 323                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 324                        root->fs_info->running_transaction->transid);
 325                 WARN_ON(1);
 326         }
 327         if (trans->transid != root->fs_info->generation) {
 328                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 329                        root->fs_info->generation);
 330                 WARN_ON(1);
 331         }
 332
 333         parent_nritems = btrfs_header_nritems(parent);
 334         blocksize = btrfs_level_size(root, parent_level - 1);
 335         end_slot = parent_nritems;
 336
 337         if (parent_nritems == 1)
 338                 return 0;
 339
 340         for (i = start_slot; i < end_slot; i++) {
 341                 int close = 1;
 342
 343                 if (!parent->map_token) {
 344                         map_extent_buffer(parent,
 345                                         btrfs_node_key_ptr_offset(i),
 346                                         sizeof(struct btrfs_key_ptr),
 347                                         &parent->map_token, &parent->kaddr,
 348                                         &parent->map_start, &parent->map_len,
 349                                         KM_USER1);
 350                 }
 351                 btrfs_node_key(parent, &disk_key, i);
 352                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
 353                         continue;
 354
 355                 progress_passed = 1;
 356                 blocknr = btrfs_node_blockptr(parent, i);
 357                 if (last_block == 0)
 358                         last_block = blocknr;
 359
 360                 if (i > 0) {
 361                         other = btrfs_node_blockptr(parent, i - 1);
 362                         close = close_blocks(blocknr, other, blocksize);
 363                 }
 364                 if (close && i < end_slot - 2) {
 365                         other = btrfs_node_blockptr(parent, i + 1);
 366                         close = close_blocks(blocknr, other, blocksize);
 367                 }
 368                 if (close) {
 369                         last_block = blocknr;
 370                         continue;
 371                 }
 372                 if (parent->map_token) {
 373                         unmap_extent_buffer(parent, parent->map_token,
 374                                             KM_USER1);
 375                         parent->map_token = NULL;
 376                 }
 377
 378                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
 379                 if (cur)
 380                         uptodate = btrfs_buffer_uptodate(cur);
 381                 else
 382                         uptodate = 0;
 383                 if (!cur || !uptodate) {
 384                         if (cache_only) {
 385                                 free_extent_buffer(cur);
 386                                 continue;
 387                         }
 388                         if (!cur) {
 389                                 cur = read_tree_block(root, blocknr,
 390                                                          blocksize);
 391                         } else if (!uptodate) {
 392                                 btrfs_read_buffer(cur);
 393                         }
 394                 }
 395                 if (search_start == 0)
 396                         search_start = last_block;
 397
 398                 btrfs_verify_block_csum(root, cur);
 399                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 400                                         &tmp, search_start,
 401                                         min(16 * blocksize,
 402                                             (end_slot - i) * blocksize));
 403                 if (err) {
 404                         free_extent_buffer(cur);
 405                         break;
 406                 }
 407                 search_start = tmp->start;
 408                 last_block = tmp->start;
 409                 *last_ret = search_start;
 410                 if (parent_level == 1)
 411                         btrfs_clear_buffer_defrag(tmp);
 412                 free_extent_buffer(tmp);
 413         }
 414         if (parent->map_token) {
 415                 unmap_extent_buffer(parent, parent->map_token,
 416                                     KM_USER1);
 417                 parent->map_token = NULL;
 418         }
 419         return err;
 420 }
 421
 422 /*
 423  * The leaf data grows from end-to-front in the node.
 424  * this returns the address of the start of the last item,
 425  * which is the stop of the leaf data stack
 426  */
 427 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 428                                          struct extent_buffer *leaf)
 429 {
 430         u32 nr = btrfs_header_nritems(leaf);
 431         if (nr == 0)
 432                 return BTRFS_LEAF_DATA_SIZE(root);
 433         return btrfs_item_offset_nr(leaf, nr - 1);
 434 }
 435
 436 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
 437                       int level)
 438 {
 439         struct extent_buffer *parent = NULL;
 440         struct extent_buffer *node = path->nodes[level];
 441         struct btrfs_disk_key parent_key;
 442         struct btrfs_disk_key node_key;
 443         int parent_slot;
 444         int slot;
 445         struct btrfs_key cpukey;
 446         u32 nritems = btrfs_header_nritems(node);
 447
 448         if (path->nodes[level + 1])
 449                 parent = path->nodes[level + 1];
 450
 451         slot = path->slots[level];
 452         BUG_ON(nritems == 0);
 453         if (parent) {
 454                 parent_slot = path->slots[level + 1];
 455                 btrfs_node_key(parent, &parent_key, parent_slot);
 456                 btrfs_node_key(node, &node_key, 0);
 457                 BUG_ON(memcmp(&parent_key, &node_key,
 458                               sizeof(struct btrfs_disk_key)));
 459                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 460                        btrfs_header_bytenr(node));
 461         }
 462         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 463         if (slot != 0) {
 464                 btrfs_node_key_to_cpu(node, &cpukey, slot - 1);
 465                 btrfs_node_key(node, &node_key, slot);
 466                 BUG_ON(comp_keys(&node_key, &cpukey) <= 0);
 467         }
 468         if (slot < nritems - 1) {
 469                 btrfs_node_key_to_cpu(node, &cpukey, slot + 1);
 470                 btrfs_node_key(node, &node_key, slot);
 471                 BUG_ON(comp_keys(&node_key, &cpukey) >= 0);
 472         }
 473         return 0;
 474 }
 475
 476 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 477                       int level)
 478 {
 479         struct extent_buffer *leaf = path->nodes[level];
 480         struct extent_buffer *parent = NULL;
 481         int parent_slot;
 482         struct btrfs_key cpukey;
 483         struct btrfs_disk_key parent_key;
 484         struct btrfs_disk_key leaf_key;
 485         int slot = path->slots[0];
 486
 487         u32 nritems = btrfs_header_nritems(leaf);
 488
 489         if (path->nodes[level + 1])
 490                 parent = path->nodes[level + 1];
 491
 492         if (nritems == 0)
 493                 return 0;
 494
 495         if (parent) {
 496                 parent_slot = path->slots[level + 1];
 497                 btrfs_node_key(parent, &parent_key, parent_slot);
 498                 btrfs_item_key(leaf, &leaf_key, 0);
 499
 500                 BUG_ON(memcmp(&parent_key, &leaf_key,
 501                        sizeof(struct btrfs_disk_key)));
 502                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 503                        btrfs_header_bytenr(leaf));
 504         }
 505 #if 0
 506         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 507                 btrfs_item_key_to_cpu(leaf, &cpukey, i + 1);
 508                 btrfs_item_key(leaf, &leaf_key, i);
 509                 if (comp_keys(&leaf_key, &cpukey) >= 0) {
 510                         btrfs_print_leaf(root, leaf);
 511                         printk("slot %d offset bad key\n", i);
 512                         BUG_ON(1);
 513                 }
 514                 if (btrfs_item_offset_nr(leaf, i) !=
 515                         btrfs_item_end_nr(leaf, i + 1)) {
 516                         btrfs_print_leaf(root, leaf);
 517                         printk("slot %d offset bad\n", i);
 518                         BUG_ON(1);
 519                 }
 520                 if (i == 0) {
 521                         if (btrfs_item_offset_nr(leaf, i) +
 522                                btrfs_item_size_nr(leaf, i) !=
 523                                BTRFS_LEAF_DATA_SIZE(root)) {
 524                                 btrfs_print_leaf(root, leaf);
 525                                 printk("slot %d first offset bad\n", i);
 526                                 BUG_ON(1);
 527                         }
 528                 }
 529         }
 530         if (nritems > 0) {
 531                 if (btrfs_item_size_nr(leaf, nritems - 1) > 4096) {
 532                                 btrfs_print_leaf(root, leaf);
 533                                 printk("slot %d bad size \n", nritems - 1);
 534                                 BUG_ON(1);
 535                 }
 536         }
 537 #endif
 538         if (slot != 0 && slot < nritems - 1) {
 539                 btrfs_item_key(leaf, &leaf_key, slot);
 540                 btrfs_item_key_to_cpu(leaf, &cpukey, slot - 1);
 541                 if (comp_keys(&leaf_key, &cpukey) <= 0) {
 542                         btrfs_print_leaf(root, leaf);
 543                         printk("slot %d offset bad key\n", slot);
 544                         BUG_ON(1);
 545                 }
 546                 if (btrfs_item_offset_nr(leaf, slot - 1) !=
 547                        btrfs_item_end_nr(leaf, slot)) {
 548                         btrfs_print_leaf(root, leaf);
 549                         printk("slot %d offset bad\n", slot);
 550                         BUG_ON(1);
 551                 }
 552         }
 553         if (slot < nritems - 1) {
 554                 btrfs_item_key(leaf, &leaf_key, slot);
 555                 btrfs_item_key_to_cpu(leaf, &cpukey, slot + 1);
 556                 BUG_ON(comp_keys(&leaf_key, &cpukey) >= 0);
 557                 if (btrfs_item_offset_nr(leaf, slot) !=
 558                         btrfs_item_end_nr(leaf, slot + 1)) {
 559                         btrfs_print_leaf(root, leaf);
 560                         printk("slot %d offset bad\n", slot);
 561                         BUG_ON(1);
 562                 }
 563         }
 564         BUG_ON(btrfs_item_offset_nr(leaf, 0) +
 565                btrfs_item_size_nr(leaf, 0) != BTRFS_LEAF_DATA_SIZE(root));
 566         return 0;
 567 }
 568
 569 static int noinline check_block(struct btrfs_root *root,
 570                                 struct btrfs_path *path, int level)
 571 {
 572         u64 found_start;
 573         return 0;
 574         if (btrfs_header_level(path->nodes[level]) != level)
 575             printk("warning: bad level %Lu wanted %d found %d\n",
 576                    path->nodes[level]->start, level,
 577                    btrfs_header_level(path->nodes[level]));
 578         found_start = btrfs_header_bytenr(path->nodes[level]);
 579         if (found_start != path->nodes[level]->start) {
 580             printk("warning: bad bytentr %Lu found %Lu\n",
 581                    path->nodes[level]->start, found_start);
 582         }
 583 #if 0
 584         struct extent_buffer *buf = path->nodes[level];
 585
 586         if (memcmp_extent_buffer(buf, root->fs_info->fsid,
 587                                  (unsigned long)btrfs_header_fsid(buf),
 588                                  BTRFS_FSID_SIZE)) {
 589                 printk("warning bad block %Lu\n", buf->start);
 590                 return 1;
 591         }
 592 #endif
 593         if (level == 0)
 594                 return check_leaf(root, path, level);
 595         return check_node(root, path, level);
 596 }
 597
 598 /*
 599  * search for key in the extent_buffer.  The items start at offset p,
 600  * and they are item_size apart.  There are 'max' items in p.
 601  *
 602  * the slot in the array is returned via slot, and it points to
 603  * the place where you would insert key if it is not found in
 604  * the array.
 605  *
 606  * slot may point to max if the key is bigger than all of the keys
 607  */
 608 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 609                               int item_size, struct btrfs_key *key,
 610                               int max, int *slot)
 611 {
 612         int low = 0;
 613         int high = max;
 614         int mid;
 615         int ret;
 616         struct btrfs_disk_key *tmp = NULL;
 617         struct btrfs_disk_key unaligned;
 618         unsigned long offset;
 619         char *map_token = NULL;
 620         char *kaddr = NULL;
 621         unsigned long map_start = 0;
 622         unsigned long map_len = 0;
 623         int err;
 624
 625         while(low < high) {
 626                 mid = (low + high) / 2;
 627                 offset = p + mid * item_size;
 628
 629                 if (!map_token || offset < map_start ||
 630                     (offset + sizeof(struct btrfs_disk_key)) >
 631                     map_start + map_len) {
 632                         if (map_token) {
 633                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 634                                 map_token = NULL;
 635                         }
 636                         err = map_extent_buffer(eb, offset,
 637                                                 sizeof(struct btrfs_disk_key),
 638                                                 &map_token, &kaddr,
 639                                                 &map_start, &map_len, KM_USER0);
 640
 641                         if (!err) {
 642                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
 643                                                         map_start);
 644                         } else {
 645                                 read_extent_buffer(eb, &unaligned,
 646                                                    offset, sizeof(unaligned));
 647                                 tmp = &unaligned;
 648                         }
 649
 650                 } else {
 651                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
 652                                                         map_start);
 653                 }
 654                 ret = comp_keys(tmp, key);
 655
 656                 if (ret < 0)
 657                         low = mid + 1;
 658                 else if (ret > 0)
 659                         high = mid;
 660                 else {
 661                         *slot = mid;
 662                         if (map_token)
 663                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 664                         return 0;
 665                 }
 666         }
 667         *slot = low;
 668         if (map_token)
 669                 unmap_extent_buffer(eb, map_token, KM_USER0);
 670         return 1;
 671 }
 672
 673 /*
 674  * simple bin_search frontend that does the right thing for
 675  * leaves vs nodes
 676  */
 677 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 678                       int level, int *slot)
 679 {
 680         if (level == 0) {
 681                 return generic_bin_search(eb,
 682                                           offsetof(struct btrfs_leaf, items),
 683                                           sizeof(struct btrfs_item),
 684                                           key, btrfs_header_nritems(eb),
 685                                           slot);
 686         } else {
 687                 return generic_bin_search(eb,
 688                                           offsetof(struct btrfs_node, ptrs),
 689                                           sizeof(struct btrfs_key_ptr),
 690                                           key, btrfs_header_nritems(eb),
 691                                           slot);
 692         }
 693         return -1;
 694 }
 695
 696 static struct extent_buffer *read_node_slot(struct btrfs_root *root,
 697                                    struct extent_buffer *parent, int slot)
 698 {
 699         if (slot < 0)
 700                 return NULL;
 701         if (slot >= btrfs_header_nritems(parent))
 702                 return NULL;
 703         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 704                        btrfs_level_size(root, btrfs_header_level(parent) - 1));
 705 }
 706
 707 static int balance_level(struct btrfs_trans_handle *trans,
 708                          struct btrfs_root *root,
 709                          struct btrfs_path *path, int level)
 710 {
 711         struct extent_buffer *right = NULL;
 712         struct extent_buffer *mid;
 713         struct extent_buffer *left = NULL;
 714         struct extent_buffer *parent = NULL;
 715         int ret = 0;
 716         int wret;
 717         int pslot;
 718         int orig_slot = path->slots[level];
 719         int err_on_enospc = 0;
 720         u64 orig_ptr;
 721
 722         if (level == 0)
 723                 return 0;
 724
 725         mid = path->nodes[level];
 726         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 727
 728         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 729
 730         if (level < BTRFS_MAX_LEVEL - 1)
 731                 parent = path->nodes[level + 1];
 732         pslot = path->slots[level + 1];
 733
 734         /*
 735          * deal with the case where there is only one pointer in the root
 736          * by promoting the node below to a root
 737          */
 738         if (!parent) {
 739                 struct extent_buffer *child;
 740
 741                 if (btrfs_header_nritems(mid) != 1)
 742                         return 0;
 743
 744                 /* promote the child to a root */
 745                 child = read_node_slot(root, mid, 0);
 746                 BUG_ON(!child);
 747                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
 748                 BUG_ON(ret);
 749
 750                 root->node = child;
 751                 add_root_to_dirty_list(root);
 752                 path->nodes[level] = NULL;
 753                 clean_tree_block(trans, root, mid);
 754                 /* once for the path */
 755                 free_extent_buffer(mid);
 756                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 757                                         root->root_key.objectid,
 758                                         btrfs_header_generation(mid), 0, 0, 1);
 759                 /* once for the root ptr */
 760                 free_extent_buffer(mid);
 761                 return ret;
 762         }
 763         if (btrfs_header_nritems(mid) >
 764             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 765                 return 0;
 766
 767         if (btrfs_header_nritems(mid) < 2)
 768                 err_on_enospc = 1;
 769
 770         left = read_node_slot(root, parent, pslot - 1);
 771         if (left) {
 772                 wret = btrfs_cow_block(trans, root, left,
 773                                        parent, pslot - 1, &left);
 774                 if (wret) {
 775                         ret = wret;
 776                         goto enospc;
 777                 }
 778         }
 779         right = read_node_slot(root, parent, pslot + 1);
 780         if (right) {
 781                 wret = btrfs_cow_block(trans, root, right,
 782                                        parent, pslot + 1, &right);
 783                 if (wret) {
 784                         ret = wret;
 785                         goto enospc;
 786                 }
 787         }
 788
 789         /* first, try to make some room in the middle buffer */
 790         if (left) {
 791                 orig_slot += btrfs_header_nritems(left);
 792                 wret = push_node_left(trans, root, left, mid, 0);
 793                 if (wret < 0)
 794                         ret = wret;
 795                 if (btrfs_header_nritems(mid) < 2)
 796                         err_on_enospc = 1;
 797         }
 798
 799         /*
 800          * then try to empty the right most buffer into the middle
 801          */
 802         if (right) {
 803                 wret = push_node_left(trans, root, mid, right, 1);
 804                 if (wret < 0 && wret != -ENOSPC)
 805                         ret = wret;
 806                 if (btrfs_header_nritems(right) == 0) {
 807                         u64 bytenr = right->start;
 808                         u64 generation = btrfs_header_generation(parent);
 809                         u32 blocksize = right->len;
 810
 811                         clean_tree_block(trans, root, right);
 812                         free_extent_buffer(right);
 813                         right = NULL;
 814                         wret = del_ptr(trans, root, path, level + 1, pslot +
 815                                        1);
 816                         if (wret)
 817                                 ret = wret;
 818                         wret = btrfs_free_extent(trans, root, bytenr,
 819                                                  blocksize,
 820                                                  btrfs_header_owner(parent),
 821                                                  generation, 0, 0, 1);
 822                         if (wret)
 823                                 ret = wret;
 824                 } else {
 825                         struct btrfs_disk_key right_key;
 826                         btrfs_node_key(right, &right_key, 0);
 827                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 828                         btrfs_mark_buffer_dirty(parent);
 829                 }
 830         }
 831         if (btrfs_header_nritems(mid) == 1) {
 832                 /*
 833                  * we're not allowed to leave a node with one item in the
 834                  * tree during a delete.  A deletion from lower in the tree
 835                  * could try to delete the only pointer in this node.
 836                  * So, pull some keys from the left.
 837                  * There has to be a left pointer at this point because
 838                  * otherwise we would have pulled some pointers from the
 839                  * right
 840                  */
 841                 BUG_ON(!left);
 842                 wret = balance_node_right(trans, root, mid, left);
 843                 if (wret < 0) {
 844                         ret = wret;
 845                         goto enospc;
 846                 }
 847                 BUG_ON(wret == 1);
 848         }
 849         if (btrfs_header_nritems(mid) == 0) {
 850                 /* we've managed to empty the middle node, drop it */
 851                 u64 root_gen = btrfs_header_generation(parent);
 852                 u64 bytenr = mid->start;
 853                 u32 blocksize = mid->len;
 854                 clean_tree_block(trans, root, mid);
 855                 free_extent_buffer(mid);
 856                 mid = NULL;
 857                 wret = del_ptr(trans, root, path, level + 1, pslot);
 858                 if (wret)
 859                         ret = wret;
 860                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 861                                          btrfs_header_owner(parent),
 862                                          root_gen, 0, 0, 1);
 863                 if (wret)
 864                         ret = wret;
 865         } else {
 866                 /* update the parent key to reflect our changes */
 867                 struct btrfs_disk_key mid_key;
 868                 btrfs_node_key(mid, &mid_key, 0);
 869                 btrfs_set_node_key(parent, &mid_key, pslot);
 870                 btrfs_mark_buffer_dirty(parent);
 871         }
 872
 873         /* update the path */
 874         if (left) {
 875                 if (btrfs_header_nritems(left) > orig_slot) {
 876                         extent_buffer_get(left);
 877                         path->nodes[level] = left;
 878                         path->slots[level + 1] -= 1;
 879                         path->slots[level] = orig_slot;
 880                         if (mid)
 881                                 free_extent_buffer(mid);
 882                 } else {
 883                         orig_slot -= btrfs_header_nritems(left);
 884                         path->slots[level] = orig_slot;
 885                 }
 886         }
 887         /* double check we haven't messed things up */
 888         check_block(root, path, level);
 889         if (orig_ptr !=
 890             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 891                 BUG();
 892 enospc:
 893         if (right)
 894                 free_extent_buffer(right);
 895         if (left)
 896                 free_extent_buffer(left);
 897         return ret;
 898 }
 899
 900 /* returns zero if the push worked, non-zero otherwise */
 901 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
 902                                           struct btrfs_root *root,
 903                                           struct btrfs_path *path, int level)
 904 {
 905         struct extent_buffer *right = NULL;
 906         struct extent_buffer *mid;
 907         struct extent_buffer *left = NULL;
 908         struct extent_buffer *parent = NULL;
 909         int ret = 0;
 910         int wret;
 911         int pslot;
 912         int orig_slot = path->slots[level];
 913         u64 orig_ptr;
 914
 915         if (level == 0)
 916                 return 1;
 917
 918         mid = path->nodes[level];
 919         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 920         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 921
 922         if (level < BTRFS_MAX_LEVEL - 1)
 923                 parent = path->nodes[level + 1];
 924         pslot = path->slots[level + 1];
 925
 926         if (!parent)
 927                 return 1;
 928
 929         left = read_node_slot(root, parent, pslot - 1);
 930
 931         /* first, try to make some room in the middle buffer */
 932         if (left) {
 933                 u32 left_nr;
 934                 left_nr = btrfs_header_nritems(left);
 935                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 936                         wret = 1;
 937                 } else {
 938                         ret = btrfs_cow_block(trans, root, left, parent,
 939                                               pslot - 1, &left);
 940                         if (ret)
 941                                 wret = 1;
 942                         else {
 943                                 wret = push_node_left(trans, root,
 944                                                       left, mid, 0);
 945                         }
 946                 }
 947                 if (wret < 0)
 948                         ret = wret;
 949                 if (wret == 0) {
 950                         struct btrfs_disk_key disk_key;
 951                         orig_slot += left_nr;
 952                         btrfs_node_key(mid, &disk_key, 0);
 953                         btrfs_set_node_key(parent, &disk_key, pslot);
 954                         btrfs_mark_buffer_dirty(parent);
 955                         if (btrfs_header_nritems(left) > orig_slot) {
 956                                 path->nodes[level] = left;
 957                                 path->slots[level + 1] -= 1;
 958                                 path->slots[level] = orig_slot;
 959                                 free_extent_buffer(mid);
 960                         } else {
 961                                 orig_slot -=
 962                                         btrfs_header_nritems(left);
 963                                 path->slots[level] = orig_slot;
 964                                 free_extent_buffer(left);
 965                         }
 966                         return 0;
 967                 }
 968                 free_extent_buffer(left);
 969         }
 970         right= read_node_slot(root, parent, pslot + 1);
 971
 972         /*
 973          * then try to empty the right most buffer into the middle
 974          */
 975         if (right) {
 976                 u32 right_nr;
 977                 right_nr = btrfs_header_nritems(right);
 978                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 979                         wret = 1;
 980                 } else {
 981                         ret = btrfs_cow_block(trans, root, right,
 982                                               parent, pslot + 1,
 983                                               &right);
 984                         if (ret)
 985                                 wret = 1;
 986                         else {
 987                                 wret = balance_node_right(trans, root,
 988                                                           right, mid);
 989                         }
 990                 }
 991                 if (wret < 0)
 992                         ret = wret;
 993                 if (wret == 0) {
 994                         struct btrfs_disk_key disk_key;
 995
 996                         btrfs_node_key(right, &disk_key, 0);
 997                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
 998                         btrfs_mark_buffer_dirty(parent);
 999
1000                         if (btrfs_header_nritems(mid) <= orig_slot) {
1001                                 path->nodes[level] = right;
1002                                 path->slots[level + 1] += 1;
1003                                 path->slots[level] = orig_slot -
1004                                         btrfs_header_nritems(mid);
1005                                 free_extent_buffer(mid);
1006                         } else {
1007                                 free_extent_buffer(right);
1008                         }
1009                         return 0;
1010                 }
1011                 free_extent_buffer(right);
1012         }
1013         return 1;
1014 }
1015
1016 /*
1017  * readahead one full node of leaves
1018  */
1019 static void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
1020                              int level, int slot, u64 objectid)
1021 {
1022         struct extent_buffer *node;
1023         struct btrfs_disk_key disk_key;
1024         u32 nritems;
1025         u64 search;
1026         u64 lowest_read;
1027         u64 highest_read;
1028         u64 nread = 0;
1029         int direction = path->reada;
1030         struct extent_buffer *eb;
1031         u32 nr;
1032         u32 blocksize;
1033         u32 nscan = 0;
1034
1035         if (level != 1)
1036                 return;
1037
1038         if (!path->nodes[level])
1039                 return;
1040
1041         node = path->nodes[level];
1042         search = btrfs_node_blockptr(node, slot);
1043         blocksize = btrfs_level_size(root, level - 1);
1044         eb = btrfs_find_tree_block(root, search, blocksize);
1045         if (eb) {
1046                 free_extent_buffer(eb);
1047                 return;
1048         }
1049
1050         highest_read = search;
1051         lowest_read = search;
1052
1053         nritems = btrfs_header_nritems(node);
1054         nr = slot;
1055         while(1) {
1056                 if (direction < 0) {
1057                         if (nr == 0)
1058                                 break;
1059                         nr--;
1060                 } else if (direction > 0) {
1061                         nr++;
1062                         if (nr >= nritems)
1063                                 break;
1064                 }
1065                 if (path->reada < 0 && objectid) {
1066                         btrfs_node_key(node, &disk_key, nr);
1067                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1068                                 break;
1069                 }
1070                 search = btrfs_node_blockptr(node, nr);
1071                 if ((search >= lowest_read && search <= highest_read) ||
1072                     (search < lowest_read && lowest_read - search <= 32768) ||
1073                     (search > highest_read && search - highest_read <= 32768)) {
1074                         readahead_tree_block(root, search, blocksize);
1075                         nread += blocksize;
1076                 }
1077                 nscan++;
1078                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
1079                         break;
1080                 if(nread > (1024 * 1024) || nscan > 128)
1081                         break;
1082
1083                 if (search < lowest_read)
1084                         lowest_read = search;
1085                 if (search > highest_read)
1086                         highest_read = search;
1087         }
1088 }
1089 /*
1090  * look for key in the tree.  path is filled in with nodes along the way
1091  * if key is found, we return zero and you can find the item in the leaf
1092  * level of the path (level 0)
1093  *
1094  * If the key isn't found, the path points to the slot where it should
1095  * be inserted, and 1 is returned.  If there are other errors during the
1096  * search a negative error number is returned.
1097  *
1098  * if ins_len > 0, nodes and leaves will be split as we walk down the
1099  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1100  * possible)
1101  */
1102 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1103                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1104                       ins_len, int cow)
1105 {
1106         struct extent_buffer *b;
1107         u64 bytenr;
1108         u64 ptr_gen;
1109         int slot;
1110         int ret;
1111         int level;
1112         int should_reada = p->reada;
1113         u8 lowest_level = 0;
1114
1115         lowest_level = p->lowest_level;
1116         WARN_ON(lowest_level && ins_len);
1117         WARN_ON(p->nodes[0] != NULL);
1118         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
1119 again:
1120         b = root->node;
1121         extent_buffer_get(b);
1122         while (b) {
1123                 level = btrfs_header_level(b);
1124                 if (cow) {
1125                         int wret;
1126                         wret = btrfs_cow_block(trans, root, b,
1127                                                p->nodes[level + 1],
1128                                                p->slots[level + 1],
1129                                                &b);
1130                         if (wret) {
1131                                 free_extent_buffer(b);
1132                                 return wret;
1133                         }
1134                 }
1135                 BUG_ON(!cow && ins_len);
1136                 if (level != btrfs_header_level(b))
1137                         WARN_ON(1);
1138                 level = btrfs_header_level(b);
1139                 p->nodes[level] = b;
1140                 ret = check_block(root, p, level);
1141                 if (ret)
1142                         return -1;
1143                 ret = bin_search(b, key, level, &slot);
1144                 if (level != 0) {
1145                         if (ret && slot > 0)
1146                                 slot -= 1;
1147                         p->slots[level] = slot;
1148                         if (ins_len > 0 && btrfs_header_nritems(b) >=
1149                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1150                                 int sret = split_node(trans, root, p, level);
1151                                 BUG_ON(sret > 0);
1152                                 if (sret)
1153                                         return sret;
1154                                 b = p->nodes[level];
1155                                 slot = p->slots[level];
1156                         } else if (ins_len < 0) {
1157                                 int sret = balance_level(trans, root, p,
1158                                                          level);
1159                                 if (sret)
1160                                         return sret;
1161                                 b = p->nodes[level];
1162                                 if (!b) {
1163                                         btrfs_release_path(NULL, p);
1164                                         goto again;
1165                                 }
1166                                 slot = p->slots[level];
1167                                 BUG_ON(btrfs_header_nritems(b) == 1);
1168                         }
1169                         /* this is only true while dropping a snapshot */
1170                         if (level == lowest_level)
1171                                 break;
1172                         bytenr = btrfs_node_blockptr(b, slot);
1173                         ptr_gen = btrfs_node_ptr_generation(b, slot);
1174                         if (should_reada)
1175                                 reada_for_search(root, p, level, slot,
1176                                                  key->objectid);
1177                         b = read_tree_block(root, bytenr,
1178                                             btrfs_level_size(root, level - 1));
1179                         if (ptr_gen != btrfs_header_generation(b)) {
1180                                 printk("block %llu bad gen wanted %llu "
1181                                        "found %llu\n",
1182                                 (unsigned long long)b->start,
1183                                 (unsigned long long)ptr_gen,
1184                                 (unsigned long long)btrfs_header_generation(b));
1185                         }
1186                 } else {
1187                         p->slots[level] = slot;
1188                         if (ins_len > 0 && btrfs_leaf_free_space(root, b) <
1189                             sizeof(struct btrfs_item) + ins_len) {
1190                                 int sret = split_leaf(trans, root, key,
1191                                                       p, ins_len, ret == 0);
1192                                 BUG_ON(sret > 0);
1193                                 if (sret)
1194                                         return sret;
1195                         }
1196                         return ret;
1197                 }
1198         }
1199         return 1;
1200 }
1201
1202 /*
1203  * adjust the pointers going up the tree, starting at level
1204  * making sure the right key of each node is points to 'key'.
1205  * This is used after shifting pointers to the left, so it stops
1206  * fixing up pointers when a given leaf/node is not in slot 0 of the
1207  * higher levels
1208  *
1209  * If this fails to write a tree block, it returns -1, but continues
1210  * fixing up the blocks in ram so the tree is consistent.
1211  */
1212 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1213                           struct btrfs_root *root, struct btrfs_path *path,
1214                           struct btrfs_disk_key *key, int level)
1215 {
1216         int i;
1217         int ret = 0;
1218         struct extent_buffer *t;
1219
1220         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1221                 int tslot = path->slots[i];
1222                 if (!path->nodes[i])
1223                         break;
1224                 t = path->nodes[i];
1225                 btrfs_set_node_key(t, key, tslot);
1226                 btrfs_mark_buffer_dirty(path->nodes[i]);
1227                 if (tslot != 0)
1228                         break;
1229         }
1230         return ret;
1231 }
1232
1233 /*
1234  * try to push data from one node into the next node left in the
1235  * tree.
1236  *
1237  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1238  * error, and > 0 if there was no room in the left hand block.
1239  */
1240 static int push_node_left(struct btrfs_trans_handle *trans,
1241                           struct btrfs_root *root, struct extent_buffer *dst,
1242                           struct extent_buffer *src, int empty)
1243 {
1244         int push_items = 0;
1245         int src_nritems;
1246         int dst_nritems;
1247         int ret = 0;
1248
1249         src_nritems = btrfs_header_nritems(src);
1250         dst_nritems = btrfs_header_nritems(dst);
1251         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1252         WARN_ON(btrfs_header_generation(src) != trans->transid);
1253         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1254
1255         if (!empty && src_nritems <= 2)
1256                 return 1;
1257
1258         if (push_items <= 0) {
1259                 return 1;
1260         }
1261
1262         if (empty)
1263                 push_items = min(src_nritems, push_items);
1264         else
1265                 push_items = min(src_nritems - 2, push_items);
1266
1267         copy_extent_buffer(dst, src,
1268                            btrfs_node_key_ptr_offset(dst_nritems),
1269                            btrfs_node_key_ptr_offset(0),
1270                            push_items * sizeof(struct btrfs_key_ptr));
1271
1272         if (push_items < src_nritems) {
1273                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1274                                       btrfs_node_key_ptr_offset(push_items),
1275                                       (src_nritems - push_items) *
1276                                       sizeof(struct btrfs_key_ptr));
1277         }
1278         btrfs_set_header_nritems(src, src_nritems - push_items);
1279         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1280         btrfs_mark_buffer_dirty(src);
1281         btrfs_mark_buffer_dirty(dst);
1282         return ret;
1283 }
1284
1285 /*
1286  * try to push data from one node into the next node right in the
1287  * tree.
1288  *
1289  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1290  * error, and > 0 if there was no room in the right hand block.
1291  *
1292  * this will  only push up to 1/2 the contents of the left node over
1293  */
1294 static int balance_node_right(struct btrfs_trans_handle *trans,
1295                               struct btrfs_root *root,
1296                               struct extent_buffer *dst,
1297                               struct extent_buffer *src)
1298 {
1299         int push_items = 0;
1300         int max_push;
1301         int src_nritems;
1302         int dst_nritems;
1303         int ret = 0;
1304
1305         WARN_ON(btrfs_header_generation(src) != trans->transid);
1306         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1307
1308         src_nritems = btrfs_header_nritems(src);
1309         dst_nritems = btrfs_header_nritems(dst);
1310         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1311         if (push_items <= 0)
1312                 return 1;
1313
1314         max_push = src_nritems / 2 + 1;
1315         /* don't try to empty the node */
1316         if (max_push >= src_nritems)
1317                 return 1;
1318
1319         if (max_push < push_items)
1320                 push_items = max_push;
1321
1322         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1323                                       btrfs_node_key_ptr_offset(0),
1324                                       (dst_nritems) *
1325                                       sizeof(struct btrfs_key_ptr));
1326
1327         copy_extent_buffer(dst, src,
1328                            btrfs_node_key_ptr_offset(0),
1329                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1330                            push_items * sizeof(struct btrfs_key_ptr));
1331
1332         btrfs_set_header_nritems(src, src_nritems - push_items);
1333         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1334
1335         btrfs_mark_buffer_dirty(src);
1336         btrfs_mark_buffer_dirty(dst);
1337         return ret;
1338 }
1339
1340 /*
1341  * helper function to insert a new root level in the tree.
1342  * A new node is allocated, and a single item is inserted to
1343  * point to the existing root
1344  *
1345  * returns zero on success or < 0 on failure.
1346  */
1347 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1348                            struct btrfs_root *root,
1349                            struct btrfs_path *path, int level)
1350 {
1351         u64 root_gen;
1352         u64 lower_gen;
1353         struct extent_buffer *lower;
1354         struct extent_buffer *c;
1355         struct btrfs_disk_key lower_key;
1356
1357         BUG_ON(path->nodes[level]);
1358         BUG_ON(path->nodes[level-1] != root->node);
1359
1360         if (root->ref_cows)
1361                 root_gen = trans->transid;
1362         else
1363                 root_gen = 0;
1364
1365         lower = path->nodes[level-1];
1366         if (level == 1)
1367                 btrfs_item_key(lower, &lower_key, 0);
1368         else
1369                 btrfs_node_key(lower, &lower_key, 0);
1370
1371         c = __btrfs_alloc_free_block(trans, root, root->nodesize,
1372                                    root->root_key.objectid,
1373                                    root_gen, lower_key.objectid, level,
1374                                    root->node->start, 0);
1375         if (IS_ERR(c))
1376                 return PTR_ERR(c);
1377         memset_extent_buffer(c, 0, 0, root->nodesize);
1378         btrfs_set_header_nritems(c, 1);
1379         btrfs_set_header_level(c, level);
1380         btrfs_set_header_bytenr(c, c->start);
1381         btrfs_set_header_generation(c, trans->transid);
1382         btrfs_set_header_owner(c, root->root_key.objectid);
1383
1384         write_extent_buffer(c, root->fs_info->fsid,
1385                             (unsigned long)btrfs_header_fsid(c),
1386                             BTRFS_FSID_SIZE);
1387
1388         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1389                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
1390                             BTRFS_UUID_SIZE);
1391
1392         btrfs_set_node_key(c, &lower_key, 0);
1393         btrfs_set_node_blockptr(c, 0, lower->start);
1394         lower_gen = btrfs_header_generation(lower);
1395         WARN_ON(lower_gen == 0);
1396
1397         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1398
1399         btrfs_mark_buffer_dirty(c);
1400
1401         /* the super has an extra ref to root->node */
1402         free_extent_buffer(root->node);
1403         root->node = c;
1404         add_root_to_dirty_list(root);
1405         extent_buffer_get(c);
1406         path->nodes[level] = c;
1407         path->slots[level] = 0;
1408
1409         if (root->ref_cows && lower_gen != trans->transid) {
1410                 struct btrfs_path *back_path = btrfs_alloc_path();
1411                 int ret;
1412                 ret = btrfs_insert_extent_backref(trans,
1413                                                   root->fs_info->extent_root,
1414                                                   path, lower->start,
1415                                                   root->root_key.objectid,
1416                                                   trans->transid, 0, 0);
1417                 BUG_ON(ret);
1418                 btrfs_free_path(back_path);
1419         }
1420         return 0;
1421 }
1422
1423 /*
1424  * worker function to insert a single pointer in a node.
1425  * the node should have enough room for the pointer already
1426  *
1427  * slot and level indicate where you want the key to go, and
1428  * blocknr is the block the key points to.
1429  *
1430  * returns zero on success and < 0 on any error
1431  */
1432 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1433                       *root, struct btrfs_path *path, struct btrfs_disk_key
1434                       *key, u64 bytenr, int slot, int level)
1435 {
1436         struct extent_buffer *lower;
1437         int nritems;
1438
1439         BUG_ON(!path->nodes[level]);
1440         lower = path->nodes[level];
1441         nritems = btrfs_header_nritems(lower);
1442         if (slot > nritems)
1443                 BUG();
1444         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1445                 BUG();
1446         if (slot != nritems) {
1447                 memmove_extent_buffer(lower,
1448                               btrfs_node_key_ptr_offset(slot + 1),
1449                               btrfs_node_key_ptr_offset(slot),
1450                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1451         }
1452         btrfs_set_node_key(lower, key, slot);
1453         btrfs_set_node_blockptr(lower, slot, bytenr);
1454         WARN_ON(trans->transid == 0);
1455         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1456         btrfs_set_header_nritems(lower, nritems + 1);
1457         btrfs_mark_buffer_dirty(lower);
1458         return 0;
1459 }
1460
1461 /*
1462  * split the node at the specified level in path in two.
1463  * The path is corrected to point to the appropriate node after the split
1464  *
1465  * Before splitting this tries to make some room in the node by pushing
1466  * left and right, if either one works, it returns right away.
1467  *
1468  * returns 0 on success and < 0 on failure
1469  */
1470 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1471                       *root, struct btrfs_path *path, int level)
1472 {
1473         u64 root_gen;
1474         struct extent_buffer *c;
1475         struct extent_buffer *split;
1476         struct btrfs_disk_key disk_key;
1477         int mid;
1478         int ret;
1479         int wret;
1480         u32 c_nritems;
1481
1482         c = path->nodes[level];
1483         WARN_ON(btrfs_header_generation(c) != trans->transid);
1484         if (c == root->node) {
1485                 /* trying to split the root, lets make a new one */
1486                 ret = insert_new_root(trans, root, path, level + 1);
1487                 if (ret)
1488                         return ret;
1489         } else {
1490                 ret = push_nodes_for_insert(trans, root, path, level);
1491                 c = path->nodes[level];
1492                 if (!ret && btrfs_header_nritems(c) <
1493                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
1494                         return 0;
1495                 if (ret < 0)
1496                         return ret;
1497         }
1498
1499         c_nritems = btrfs_header_nritems(c);
1500         if (root->ref_cows)
1501                 root_gen = trans->transid;
1502         else
1503                 root_gen = 0;
1504
1505         btrfs_node_key(c, &disk_key, 0);
1506         split = __btrfs_alloc_free_block(trans, root, root->nodesize,
1507                                          root->root_key.objectid,
1508                                          root_gen,
1509                                          btrfs_disk_key_objectid(&disk_key),
1510                                          level, c->start, 0);
1511         if (IS_ERR(split))
1512                 return PTR_ERR(split);
1513
1514         btrfs_set_header_flags(split, btrfs_header_flags(c));
1515         btrfs_set_header_level(split, btrfs_header_level(c));
1516         btrfs_set_header_bytenr(split, split->start);
1517         btrfs_set_header_generation(split, trans->transid);
1518         btrfs_set_header_owner(split, root->root_key.objectid);
1519         btrfs_set_header_flags(split, 0);
1520         write_extent_buffer(split, root->fs_info->fsid,
1521                             (unsigned long)btrfs_header_fsid(split),
1522                             BTRFS_FSID_SIZE);
1523         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1524                             (unsigned long)btrfs_header_chunk_tree_uuid(split),
1525                             BTRFS_UUID_SIZE);
1526
1527         mid = (c_nritems + 1) / 2;
1528
1529         copy_extent_buffer(split, c,
1530                            btrfs_node_key_ptr_offset(0),
1531                            btrfs_node_key_ptr_offset(mid),
1532                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1533         btrfs_set_header_nritems(split, c_nritems - mid);
1534         btrfs_set_header_nritems(c, mid);
1535         ret = 0;
1536
1537         btrfs_mark_buffer_dirty(c);
1538         btrfs_mark_buffer_dirty(split);
1539
1540         btrfs_node_key(split, &disk_key, 0);
1541         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1542                           path->slots[level + 1] + 1,
1543                           level + 1);
1544         if (wret)
1545                 ret = wret;
1546
1547         if (path->slots[level] >= mid) {
1548                 path->slots[level] -= mid;
1549                 free_extent_buffer(c);
1550                 path->nodes[level] = split;
1551                 path->slots[level + 1] += 1;
1552         } else {
1553                 free_extent_buffer(split);
1554         }
1555         return ret;
1556 }
1557
1558 /*
1559  * how many bytes are required to store the items in a leaf.  start
1560  * and nr indicate which items in the leaf to check.  This totals up the
1561  * space used both by the item structs and the item data
1562  */
1563 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1564 {
1565         int data_len;
1566         int nritems = btrfs_header_nritems(l);
1567         int end = min(nritems, start + nr) - 1;
1568
1569         if (!nr)
1570                 return 0;
1571         data_len = btrfs_item_end_nr(l, start);
1572         data_len = data_len - btrfs_item_offset_nr(l, end);
1573         data_len += sizeof(struct btrfs_item) * nr;
1574         WARN_ON(data_len < 0);
1575         return data_len;
1576 }
1577
1578 /*
1579  * The space between the end of the leaf items and
1580  * the start of the leaf data.  IOW, how much room
1581  * the leaf has left for both items and data
1582  */
1583 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1584 {
1585         int nritems = btrfs_header_nritems(leaf);
1586         int ret;
1587         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1588         if (ret < 0) {
1589                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1590                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1591                        leaf_space_used(leaf, 0, nritems), nritems);
1592         }
1593         return ret;
1594 }
1595
1596 /*
1597  * push some data in the path leaf to the right, trying to free up at
1598  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1599  *
1600  * returns 1 if the push failed because the other node didn't have enough
1601  * room, 0 if everything worked out and < 0 if there were major errors.
1602  */
1603 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1604                            *root, struct btrfs_path *path, int data_size,
1605                            int empty)
1606 {
1607         struct extent_buffer *left = path->nodes[0];
1608         struct extent_buffer *right;
1609         struct extent_buffer *upper;
1610         struct btrfs_disk_key disk_key;
1611         int slot;
1612         u32 i;
1613         int free_space;
1614         int push_space = 0;
1615         int push_items = 0;
1616         struct btrfs_item *item;
1617         u32 left_nritems;
1618         u32 nr;
1619         u32 right_nritems;
1620         u32 data_end;
1621         u32 this_item_size;
1622         int ret;
1623
1624         slot = path->slots[1];
1625         if (!path->nodes[1]) {
1626                 return 1;
1627         }
1628         upper = path->nodes[1];
1629         if (slot >= btrfs_header_nritems(upper) - 1)
1630                 return 1;
1631
1632         right = read_tree_block(root, btrfs_node_blockptr(upper, slot + 1),
1633                                 root->leafsize);
1634         free_space = btrfs_leaf_free_space(root, right);
1635         if (free_space < data_size + sizeof(struct btrfs_item)) {
1636                 free_extent_buffer(right);
1637                 return 1;
1638         }
1639
1640         /* cow and double check */
1641         ret = btrfs_cow_block(trans, root, right, upper,
1642                               slot + 1, &right);
1643         if (ret) {
1644                 free_extent_buffer(right);
1645                 return 1;
1646         }
1647         free_space = btrfs_leaf_free_space(root, right);
1648         if (free_space < data_size + sizeof(struct btrfs_item)) {
1649                 free_extent_buffer(right);
1650                 return 1;
1651         }
1652
1653         left_nritems = btrfs_header_nritems(left);
1654         if (left_nritems == 0) {
1655                 free_extent_buffer(right);
1656                 return 1;
1657         }
1658
1659         if (empty)
1660                 nr = 0;
1661         else
1662                 nr = 1;
1663
1664         i = left_nritems - 1;
1665         while (i >= nr) {
1666                 item = btrfs_item_nr(left, i);
1667
1668                 if (path->slots[0] == i)
1669                         push_space += data_size + sizeof(*item);
1670
1671                 if (!left->map_token) {
1672                         map_extent_buffer(left, (unsigned long)item,
1673                                         sizeof(struct btrfs_item),
1674                                         &left->map_token, &left->kaddr,
1675                                         &left->map_start, &left->map_len,
1676                                         KM_USER1);
1677                 }
1678
1679                 this_item_size = btrfs_item_size(left, item);
1680                 if (this_item_size + sizeof(*item) + push_space > free_space)
1681                         break;
1682                 push_items++;
1683                 push_space += this_item_size + sizeof(*item);
1684                 if (i == 0)
1685                         break;
1686                 i--;
1687         }
1688         if (left->map_token) {
1689                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1690                 left->map_token = NULL;
1691         }
1692
1693         if (push_items == 0) {
1694                 free_extent_buffer(right);
1695                 return 1;
1696         }
1697
1698         if (!empty && push_items == left_nritems)
1699                 WARN_ON(1);
1700
1701         /* push left to right */
1702         right_nritems = btrfs_header_nritems(right);
1703
1704         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1705         push_space -= leaf_data_end(root, left);
1706
1707         /* make room in the right data area */
1708         data_end = leaf_data_end(root, right);
1709         memmove_extent_buffer(right,
1710                               btrfs_leaf_data(right) + data_end - push_space,
1711                               btrfs_leaf_data(right) + data_end,
1712                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1713
1714         /* copy from the left data area */
1715         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1716                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1717                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1718                      push_space);
1719
1720         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1721                               btrfs_item_nr_offset(0),
1722                               right_nritems * sizeof(struct btrfs_item));
1723
1724         /* copy the items from left to right */
1725         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1726                    btrfs_item_nr_offset(left_nritems - push_items),
1727                    push_items * sizeof(struct btrfs_item));
1728
1729         /* update the item pointers */
1730         right_nritems += push_items;
1731         btrfs_set_header_nritems(right, right_nritems);
1732         push_space = BTRFS_LEAF_DATA_SIZE(root);
1733         for (i = 0; i < right_nritems; i++) {
1734                 item = btrfs_item_nr(right, i);
1735                 if (!right->map_token) {
1736                         map_extent_buffer(right, (unsigned long)item,
1737                                         sizeof(struct btrfs_item),
1738                                         &right->map_token, &right->kaddr,
1739                                         &right->map_start, &right->map_len,
1740                                         KM_USER1);
1741                 }
1742                 push_space -= btrfs_item_size(right, item);
1743                 btrfs_set_item_offset(right, item, push_space);
1744         }
1745
1746         if (right->map_token) {
1747                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1748                 right->map_token = NULL;
1749         }
1750         left_nritems -= push_items;
1751         btrfs_set_header_nritems(left, left_nritems);
1752
1753         if (left_nritems)
1754                 btrfs_mark_buffer_dirty(left);
1755         btrfs_mark_buffer_dirty(right);
1756
1757         btrfs_item_key(right, &disk_key, 0);
1758         btrfs_set_node_key(upper, &disk_key, slot + 1);
1759         btrfs_mark_buffer_dirty(upper);
1760
1761         /* then fixup the leaf pointer in the path */
1762         if (path->slots[0] >= left_nritems) {
1763                 path->slots[0] -= left_nritems;
1764                 free_extent_buffer(path->nodes[0]);
1765                 path->nodes[0] = right;
1766                 path->slots[1] += 1;
1767         } else {
1768                 free_extent_buffer(right);
1769         }
1770         return 0;
1771 }
1772 /*
1773  * push some data in the path leaf to the left, trying to free up at
1774  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1775  */
1776 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1777                           *root, struct btrfs_path *path, int data_size,
1778                           int empty)
1779 {
1780         struct btrfs_disk_key disk_key;
1781         struct extent_buffer *right = path->nodes[0];
1782         struct extent_buffer *left;
1783         int slot;
1784         int i;
1785         int free_space;
1786         int push_space = 0;
1787         int push_items = 0;
1788         struct btrfs_item *item;
1789         u32 old_left_nritems;
1790         u32 right_nritems;
1791         u32 nr;
1792         int ret = 0;
1793         int wret;
1794         u32 this_item_size;
1795         u32 old_left_item_size;
1796
1797         slot = path->slots[1];
1798         if (slot == 0)
1799                 return 1;
1800         if (!path->nodes[1])
1801                 return 1;
1802
1803         right_nritems = btrfs_header_nritems(right);
1804         if (right_nritems == 0) {
1805                 return 1;
1806         }
1807
1808         left = read_tree_block(root, btrfs_node_blockptr(path->nodes[1],
1809                                slot - 1), root->leafsize);
1810         free_space = btrfs_leaf_free_space(root, left);
1811         if (free_space < data_size + sizeof(struct btrfs_item)) {
1812                 free_extent_buffer(left);
1813                 return 1;
1814         }
1815
1816         /* cow and double check */
1817         ret = btrfs_cow_block(trans, root, left,
1818                               path->nodes[1], slot - 1, &left);
1819         if (ret) {
1820                 /* we hit -ENOSPC, but it isn't fatal here */
1821                 free_extent_buffer(left);
1822                 return 1;
1823         }
1824
1825         free_space = btrfs_leaf_free_space(root, left);
1826         if (free_space < data_size + sizeof(struct btrfs_item)) {
1827                 free_extent_buffer(left);
1828                 return 1;
1829         }
1830
1831         if (empty)
1832                 nr = right_nritems;
1833         else
1834                 nr = right_nritems - 1;
1835
1836         for (i = 0; i < nr; i++) {
1837                 item = btrfs_item_nr(right, i);
1838                 if (!right->map_token) {
1839                         map_extent_buffer(right, (unsigned long)item,
1840                                         sizeof(struct btrfs_item),
1841                                         &right->map_token, &right->kaddr,
1842                                         &right->map_start, &right->map_len,
1843                                         KM_USER1);
1844                 }
1845
1846                 if (path->slots[0] == i)
1847                         push_space += data_size + sizeof(*item);
1848
1849                 this_item_size = btrfs_item_size(right, item);
1850                 if (this_item_size + sizeof(*item) + push_space > free_space)
1851                         break;
1852
1853                 push_items++;
1854                 push_space += this_item_size + sizeof(*item);
1855         }
1856
1857         if (right->map_token) {
1858                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1859                 right->map_token = NULL;
1860         }
1861
1862         if (push_items == 0) {
1863                 free_extent_buffer(left);
1864                 return 1;
1865         }
1866         if (!empty && push_items == btrfs_header_nritems(right))
1867                 WARN_ON(1);
1868
1869         /* push data from right to left */
1870         copy_extent_buffer(left, right,
1871                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1872                            btrfs_item_nr_offset(0),
1873                            push_items * sizeof(struct btrfs_item));
1874
1875         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1876                      btrfs_item_offset_nr(right, push_items -1);
1877
1878         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1879                      leaf_data_end(root, left) - push_space,
1880                      btrfs_leaf_data(right) +
1881                      btrfs_item_offset_nr(right, push_items - 1),
1882                      push_space);
1883         old_left_nritems = btrfs_header_nritems(left);
1884         BUG_ON(old_left_nritems < 0);
1885
1886         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1887         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1888                 u32 ioff;
1889
1890                 item = btrfs_item_nr(left, i);
1891                 if (!left->map_token) {
1892                         map_extent_buffer(left, (unsigned long)item,
1893                                         sizeof(struct btrfs_item),
1894                                         &left->map_token, &left->kaddr,
1895                                         &left->map_start, &left->map_len,
1896                                         KM_USER1);
1897                 }
1898
1899                 ioff = btrfs_item_offset(left, item);
1900                 btrfs_set_item_offset(left, item,
1901                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1902         }
1903         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1904         if (left->map_token) {
1905                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1906                 left->map_token = NULL;
1907         }
1908
1909         /* fixup right node */
1910         if (push_items > right_nritems) {
1911                 printk("push items %d nr %u\n", push_items, right_nritems);
1912                 WARN_ON(1);
1913         }
1914
1915         if (push_items < right_nritems) {
1916                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1917                                                   leaf_data_end(root, right);
1918                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
1919                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1920                                       btrfs_leaf_data(right) +
1921                                       leaf_data_end(root, right), push_space);
1922
1923                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1924                               btrfs_item_nr_offset(push_items),
1925                              (btrfs_header_nritems(right) - push_items) *
1926                              sizeof(struct btrfs_item));
1927         }
1928         right_nritems -= push_items;
1929         btrfs_set_header_nritems(right, right_nritems);
1930         push_space = BTRFS_LEAF_DATA_SIZE(root);
1931         for (i = 0; i < right_nritems; i++) {
1932                 item = btrfs_item_nr(right, i);
1933
1934                 if (!right->map_token) {
1935                         map_extent_buffer(right, (unsigned long)item,
1936                                         sizeof(struct btrfs_item),
1937                                         &right->map_token, &right->kaddr,
1938                                         &right->map_start, &right->map_len,
1939                                         KM_USER1);
1940                 }
1941
1942                 push_space = push_space - btrfs_item_size(right, item);
1943                 btrfs_set_item_offset(right, item, push_space);
1944         }
1945         if (right->map_token) {
1946                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1947                 right->map_token = NULL;
1948         }
1949
1950         btrfs_mark_buffer_dirty(left);
1951         if (right_nritems)
1952                 btrfs_mark_buffer_dirty(right);
1953
1954         btrfs_item_key(right, &disk_key, 0);
1955         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1956         if (wret)
1957                 ret = wret;
1958
1959         /* then fixup the leaf pointer in the path */
1960         if (path->slots[0] < push_items) {
1961                 path->slots[0] += old_left_nritems;
1962                 free_extent_buffer(path->nodes[0]);
1963                 path->nodes[0] = left;
1964                 path->slots[1] -= 1;
1965         } else {
1966                 free_extent_buffer(left);
1967                 path->slots[0] -= push_items;
1968         }
1969         BUG_ON(path->slots[0] < 0);
1970         return ret;
1971 }
1972
1973 /*
1974  * split the path's leaf in two, making sure there is at least data_size
1975  * available for the resulting leaf level of the path.
1976  *
1977  * returns 0 if all went well and < 0 on failure.
1978  */
1979 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
1980                       *root, struct btrfs_key *ins_key,
1981                       struct btrfs_path *path, int data_size, int extend)
1982 {
1983         u64 root_gen;
1984         struct extent_buffer *l;
1985         u32 nritems;
1986         int mid;
1987         int slot;
1988         struct extent_buffer *right;
1989         int space_needed = data_size + sizeof(struct btrfs_item);
1990         int data_copy_size;
1991         int rt_data_off;
1992         int i;
1993         int ret = 0;
1994         int wret;
1995         int double_split;
1996         int num_doubles = 0;
1997         struct btrfs_disk_key disk_key;
1998
1999         if (extend)
2000                 space_needed = data_size;
2001
2002         if (root->ref_cows)
2003                 root_gen = trans->transid;
2004         else
2005                 root_gen = 0;
2006
2007         /* first try to make some room by pushing left and right */
2008         if (ins_key->type != BTRFS_DIR_ITEM_KEY) {
2009                 wret = push_leaf_right(trans, root, path, data_size, 0);
2010                 if (wret < 0) {
2011                         return wret;
2012                 }
2013                 if (wret) {
2014                         wret = push_leaf_left(trans, root, path, data_size, 0);
2015                         if (wret < 0)
2016                                 return wret;
2017                 }
2018                 l = path->nodes[0];
2019
2020                 /* did the pushes work? */
2021                 if (btrfs_leaf_free_space(root, l) >= space_needed)
2022                         return 0;
2023         }
2024
2025         if (!path->nodes[1]) {
2026                 ret = insert_new_root(trans, root, path, 1);
2027                 if (ret)
2028                         return ret;
2029         }
2030 again:
2031         double_split = 0;
2032         l = path->nodes[0];
2033         slot = path->slots[0];
2034         nritems = btrfs_header_nritems(l);
2035         mid = (nritems + 1)/ 2;
2036
2037         btrfs_item_key(l, &disk_key, 0);
2038
2039         right = __btrfs_alloc_free_block(trans, root, root->leafsize,
2040                                          root->root_key.objectid,
2041                                          root_gen, disk_key.objectid, 0,
2042                                          l->start, 0);
2043         if (IS_ERR(right)) {
2044                 BUG_ON(1);
2045                 return PTR_ERR(right);
2046         }
2047
2048         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2049         btrfs_set_header_bytenr(right, right->start);
2050         btrfs_set_header_generation(right, trans->transid);
2051         btrfs_set_header_owner(right, root->root_key.objectid);
2052         btrfs_set_header_level(right, 0);
2053         write_extent_buffer(right, root->fs_info->fsid,
2054                             (unsigned long)btrfs_header_fsid(right),
2055                             BTRFS_FSID_SIZE);
2056
2057         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2058                             (unsigned long)btrfs_header_chunk_tree_uuid(right),
2059                             BTRFS_UUID_SIZE);
2060         if (mid <= slot) {
2061                 if (nritems == 1 ||
2062                     leaf_space_used(l, mid, nritems - mid) + space_needed >
2063                         BTRFS_LEAF_DATA_SIZE(root)) {
2064                         if (slot >= nritems) {
2065                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2066                                 btrfs_set_header_nritems(right, 0);
2067                                 wret = insert_ptr(trans, root, path,
2068                                                   &disk_key, right->start,
2069                                                   path->slots[1] + 1, 1);
2070                                 if (wret)
2071                                         ret = wret;
2072                                 free_extent_buffer(path->nodes[0]);
2073                                 path->nodes[0] = right;
2074                                 path->slots[0] = 0;
2075                                 path->slots[1] += 1;
2076                                 btrfs_mark_buffer_dirty(right);
2077                                 return ret;
2078                         }
2079                         mid = slot;
2080                         if (mid != nritems &&
2081                             leaf_space_used(l, mid, nritems - mid) +
2082                             space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2083                                 double_split = 1;
2084                         }
2085                 }
2086         } else {
2087                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
2088                         BTRFS_LEAF_DATA_SIZE(root)) {
2089                         if (!extend && slot == 0) {
2090                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2091                                 btrfs_set_header_nritems(right, 0);
2092                                 wret = insert_ptr(trans, root, path,
2093                                                   &disk_key,
2094                                                   right->start,
2095                                                   path->slots[1], 1);
2096                                 if (wret)
2097                                         ret = wret;
2098                                 free_extent_buffer(path->nodes[0]);
2099                                 path->nodes[0] = right;
2100                                 path->slots[0] = 0;
2101                                 if (path->slots[1] == 0) {
2102                                         wret = fixup_low_keys(trans, root,
2103                                                    path, &disk_key, 1);
2104                                         if (wret)
2105                                                 ret = wret;
2106                                 }
2107                                 btrfs_mark_buffer_dirty(right);
2108                                 return ret;
2109                         } else if (extend && slot == 0) {
2110                                 mid = 1;
2111                         } else {
2112                                 mid = slot;
2113                                 if (mid != nritems &&
2114                                     leaf_space_used(l, mid, nritems - mid) +
2115                                     space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2116                                         double_split = 1;
2117                                 }
2118                         }
2119                 }
2120         }
2121         nritems = nritems - mid;
2122         btrfs_set_header_nritems(right, nritems);
2123         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2124
2125         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2126                            btrfs_item_nr_offset(mid),
2127                            nritems * sizeof(struct btrfs_item));
2128
2129         copy_extent_buffer(right, l,
2130                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2131                      data_copy_size, btrfs_leaf_data(l) +
2132                      leaf_data_end(root, l), data_copy_size);
2133
2134         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2135                       btrfs_item_end_nr(l, mid);
2136
2137         for (i = 0; i < nritems; i++) {
2138                 struct btrfs_item *item = btrfs_item_nr(right, i);
2139                 u32 ioff;
2140
2141                 if (!right->map_token) {
2142                         map_extent_buffer(right, (unsigned long)item,
2143                                         sizeof(struct btrfs_item),
2144                                         &right->map_token, &right->kaddr,
2145                                         &right->map_start, &right->map_len,
2146                                         KM_USER1);
2147                 }
2148
2149                 ioff = btrfs_item_offset(right, item);
2150                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2151         }
2152
2153         if (right->map_token) {
2154                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2155                 right->map_token = NULL;
2156         }
2157
2158         btrfs_set_header_nritems(l, mid);
2159         ret = 0;
2160         btrfs_item_key(right, &disk_key, 0);
2161         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2162                           path->slots[1] + 1, 1);
2163         if (wret)
2164                 ret = wret;
2165
2166         btrfs_mark_buffer_dirty(right);
2167         btrfs_mark_buffer_dirty(l);
2168         BUG_ON(path->slots[0] != slot);
2169
2170         if (mid <= slot) {
2171                 free_extent_buffer(path->nodes[0]);
2172                 path->nodes[0] = right;
2173                 path->slots[0] -= mid;
2174                 path->slots[1] += 1;
2175         } else
2176                 free_extent_buffer(right);
2177
2178         BUG_ON(path->slots[0] < 0);
2179
2180         if (double_split) {
2181                 BUG_ON(num_doubles != 0);
2182                 num_doubles++;
2183                 goto again;
2184         }
2185         return ret;
2186 }
2187
2188 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2189                         struct btrfs_root *root,
2190                         struct btrfs_path *path,
2191                         u32 new_size, int from_end)
2192 {
2193         int ret = 0;
2194         int slot;
2195         int slot_orig;
2196         struct extent_buffer *leaf;
2197         struct btrfs_item *item;
2198         u32 nritems;
2199         unsigned int data_end;
2200         unsigned int old_data_start;
2201         unsigned int old_size;
2202         unsigned int size_diff;
2203         int i;
2204
2205         slot_orig = path->slots[0];
2206         leaf = path->nodes[0];
2207         slot = path->slots[0];
2208
2209         old_size = btrfs_item_size_nr(leaf, slot);
2210         if (old_size == new_size)
2211                 return 0;
2212
2213         nritems = btrfs_header_nritems(leaf);
2214         data_end = leaf_data_end(root, leaf);
2215
2216         old_data_start = btrfs_item_offset_nr(leaf, slot);
2217
2218         size_diff = old_size - new_size;
2219
2220         BUG_ON(slot < 0);
2221         BUG_ON(slot >= nritems);
2222
2223         /*
2224          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2225          */
2226         /* first correct the data pointers */
2227         for (i = slot; i < nritems; i++) {
2228                 u32 ioff;
2229                 item = btrfs_item_nr(leaf, i);
2230
2231                 if (!leaf->map_token) {
2232                         map_extent_buffer(leaf, (unsigned long)item,
2233                                         sizeof(struct btrfs_item),
2234                                         &leaf->map_token, &leaf->kaddr,
2235                                         &leaf->map_start, &leaf->map_len,
2236                                         KM_USER1);
2237                 }
2238
2239                 ioff = btrfs_item_offset(leaf, item);
2240                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2241         }
2242
2243         if (leaf->map_token) {
2244                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2245                 leaf->map_token = NULL;
2246         }
2247
2248         /* shift the data */
2249         if (from_end) {
2250                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2251                               data_end + size_diff, btrfs_leaf_data(leaf) +
2252                               data_end, old_data_start + new_size - data_end);
2253         } else {
2254                 struct btrfs_disk_key disk_key;
2255                 u64 offset;
2256
2257                 btrfs_item_key(leaf, &disk_key, slot);
2258
2259                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2260                         unsigned long ptr;
2261                         struct btrfs_file_extent_item *fi;
2262
2263                         fi = btrfs_item_ptr(leaf, slot,
2264                                             struct btrfs_file_extent_item);
2265                         fi = (struct btrfs_file_extent_item *)(
2266                              (unsigned long)fi - size_diff);
2267
2268                         if (btrfs_file_extent_type(leaf, fi) ==
2269                             BTRFS_FILE_EXTENT_INLINE) {
2270                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2271                                 memmove_extent_buffer(leaf, ptr,
2272                                         (unsigned long)fi,
2273                                         offsetof(struct btrfs_file_extent_item,
2274                                                  disk_bytenr));
2275                         }
2276                 }
2277
2278                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2279                               data_end + size_diff, btrfs_leaf_data(leaf) +
2280                               data_end, old_data_start - data_end);
2281
2282                 offset = btrfs_disk_key_offset(&disk_key);
2283                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2284                 btrfs_set_item_key(leaf, &disk_key, slot);
2285                 if (slot == 0)
2286                         fixup_low_keys(trans, root, path, &disk_key, 1);
2287         }
2288
2289         item = btrfs_item_nr(leaf, slot);
2290         btrfs_set_item_size(leaf, item, new_size);
2291         btrfs_mark_buffer_dirty(leaf);
2292
2293         ret = 0;
2294         if (btrfs_leaf_free_space(root, leaf) < 0) {
2295                 btrfs_print_leaf(root, leaf);
2296                 BUG();
2297         }
2298         return ret;
2299 }
2300
2301 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2302                       struct btrfs_root *root, struct btrfs_path *path,
2303                       u32 data_size)
2304 {
2305         int ret = 0;
2306         int slot;
2307         int slot_orig;
2308         struct extent_buffer *leaf;
2309         struct btrfs_item *item;
2310         u32 nritems;
2311         unsigned int data_end;
2312         unsigned int old_data;
2313         unsigned int old_size;
2314         int i;
2315
2316         slot_orig = path->slots[0];
2317         leaf = path->nodes[0];
2318
2319         nritems = btrfs_header_nritems(leaf);
2320         data_end = leaf_data_end(root, leaf);
2321
2322         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2323                 btrfs_print_leaf(root, leaf);
2324                 BUG();
2325         }
2326         slot = path->slots[0];
2327         old_data = btrfs_item_end_nr(leaf, slot);
2328
2329         BUG_ON(slot < 0);
2330         if (slot >= nritems) {
2331                 btrfs_print_leaf(root, leaf);
2332                 printk("slot %d too large, nritems %d\n", slot, nritems);
2333                 BUG_ON(1);
2334         }
2335
2336         /*
2337          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2338          */
2339         /* first correct the data pointers */
2340         for (i = slot; i < nritems; i++) {
2341                 u32 ioff;
2342                 item = btrfs_item_nr(leaf, i);
2343
2344                 if (!leaf->map_token) {
2345                         map_extent_buffer(leaf, (unsigned long)item,
2346                                         sizeof(struct btrfs_item),
2347                                         &leaf->map_token, &leaf->kaddr,
2348                                         &leaf->map_start, &leaf->map_len,
2349                                         KM_USER1);
2350                 }
2351                 ioff = btrfs_item_offset(leaf, item);
2352                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2353         }
2354
2355         if (leaf->map_token) {
2356                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2357                 leaf->map_token = NULL;
2358         }
2359
2360         /* shift the data */
2361         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2362                       data_end - data_size, btrfs_leaf_data(leaf) +
2363                       data_end, old_data - data_end);
2364
2365         data_end = old_data;
2366         old_size = btrfs_item_size_nr(leaf, slot);
2367         item = btrfs_item_nr(leaf, slot);
2368         btrfs_set_item_size(leaf, item, old_size + data_size);
2369         btrfs_mark_buffer_dirty(leaf);
2370
2371         ret = 0;
2372         if (btrfs_leaf_free_space(root, leaf) < 0) {
2373                 btrfs_print_leaf(root, leaf);
2374                 BUG();
2375         }
2376         return ret;
2377 }
2378
2379 /*
2380  * Given a key and some data, insert an item into the tree.
2381  * This does all the path init required, making room in the tree if needed.
2382  */
2383 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2384                             struct btrfs_root *root,
2385                             struct btrfs_path *path,
2386                             struct btrfs_key *cpu_key, u32 *data_size,
2387                             int nr)
2388 {
2389         struct extent_buffer *leaf;
2390         struct btrfs_item *item;
2391         int ret = 0;
2392         int slot;
2393         int slot_orig;
2394         int i;
2395         u32 nritems;
2396         u32 total_size = 0;
2397         u32 total_data = 0;
2398         unsigned int data_end;
2399         struct btrfs_disk_key disk_key;
2400
2401         for (i = 0; i < nr; i++) {
2402                 total_data += data_size[i];
2403         }
2404
2405         /* create a root if there isn't one */
2406         if (!root->node)
2407                 BUG();
2408
2409         total_size = total_data + (nr - 1) * sizeof(struct btrfs_item);
2410         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2411         if (ret == 0) {
2412                 return -EEXIST;
2413         }
2414         if (ret < 0)
2415                 goto out;
2416
2417         slot_orig = path->slots[0];
2418         leaf = path->nodes[0];
2419
2420         nritems = btrfs_header_nritems(leaf);
2421         data_end = leaf_data_end(root, leaf);
2422
2423         if (btrfs_leaf_free_space(root, leaf) <
2424             sizeof(struct btrfs_item) + total_size) {
2425                 btrfs_print_leaf(root, leaf);
2426                 printk("not enough freespace need %u have %d\n",
2427                        total_size, btrfs_leaf_free_space(root, leaf));
2428                 BUG();
2429         }
2430
2431         slot = path->slots[0];
2432         BUG_ON(slot < 0);
2433
2434         if (slot != nritems) {
2435                 int i;
2436                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2437
2438                 if (old_data < data_end) {
2439                         btrfs_print_leaf(root, leaf);
2440                         printk("slot %d old_data %d data_end %d\n",
2441                                slot, old_data, data_end);
2442                         BUG_ON(1);
2443                 }
2444                 /*
2445                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2446                  */
2447                 /* first correct the data pointers */
2448                 WARN_ON(leaf->map_token);
2449                 for (i = slot; i < nritems; i++) {
2450                         u32 ioff;
2451
2452                         item = btrfs_item_nr(leaf, i);
2453                         if (!leaf->map_token) {
2454                                 map_extent_buffer(leaf, (unsigned long)item,
2455                                         sizeof(struct btrfs_item),
2456                                         &leaf->map_token, &leaf->kaddr,
2457                                         &leaf->map_start, &leaf->map_len,
2458                                         KM_USER1);
2459                         }
2460
2461                         ioff = btrfs_item_offset(leaf, item);
2462                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2463                 }
2464                 if (leaf->map_token) {
2465                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2466                         leaf->map_token = NULL;
2467                 }
2468
2469                 /* shift the items */
2470                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2471                               btrfs_item_nr_offset(slot),
2472                               (nritems - slot) * sizeof(struct btrfs_item));
2473
2474                 /* shift the data */
2475                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2476                               data_end - total_data, btrfs_leaf_data(leaf) +
2477                               data_end, old_data - data_end);
2478                 data_end = old_data;
2479         }
2480
2481         /* setup the item for the new data */
2482         for (i = 0; i < nr; i++) {
2483                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2484                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2485                 item = btrfs_item_nr(leaf, slot + i);
2486                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2487                 data_end -= data_size[i];
2488                 btrfs_set_item_size(leaf, item, data_size[i]);
2489         }
2490         btrfs_set_header_nritems(leaf, nritems + nr);
2491         btrfs_mark_buffer_dirty(leaf);
2492
2493         ret = 0;
2494         if (slot == 0) {
2495                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2496                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2497         }
2498
2499         if (btrfs_leaf_free_space(root, leaf) < 0) {
2500                 btrfs_print_leaf(root, leaf);
2501                 BUG();
2502         }
2503
2504 out:
2505         return ret;
2506 }
2507
2508 /*
2509  * Given a key and some data, insert an item into the tree.
2510  * This does all the path init required, making room in the tree if needed.
2511  */
2512 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2513                       *root, struct btrfs_key *cpu_key, void *data, u32
2514                       data_size)
2515 {
2516         int ret = 0;
2517         struct btrfs_path *path;
2518         struct extent_buffer *leaf;
2519         unsigned long ptr;
2520
2521         path = btrfs_alloc_path();
2522         BUG_ON(!path);
2523         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2524         if (!ret) {
2525                 leaf = path->nodes[0];
2526                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2527                 write_extent_buffer(leaf, data, ptr, data_size);
2528                 btrfs_mark_buffer_dirty(leaf);
2529         }
2530         btrfs_free_path(path);
2531         return ret;
2532 }
2533
2534 /*
2535  * delete the pointer from a given node.
2536  *
2537  * If the delete empties a node, the node is removed from the tree,
2538  * continuing all the way the root if required.  The root is converted into
2539  * a leaf if all the nodes are emptied.
2540  */
2541 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2542                    struct btrfs_path *path, int level, int slot)
2543 {
2544         struct extent_buffer *parent = path->nodes[level];
2545         u32 nritems;
2546         int ret = 0;
2547         int wret;
2548
2549         nritems = btrfs_header_nritems(parent);
2550         if (slot != nritems -1) {
2551                 memmove_extent_buffer(parent,
2552                               btrfs_node_key_ptr_offset(slot),
2553                               btrfs_node_key_ptr_offset(slot + 1),
2554                               sizeof(struct btrfs_key_ptr) *
2555                               (nritems - slot - 1));
2556         }
2557         nritems--;
2558         btrfs_set_header_nritems(parent, nritems);
2559         if (nritems == 0 && parent == root->node) {
2560                 BUG_ON(btrfs_header_level(root->node) != 1);
2561                 /* just turn the root into a leaf and break */
2562                 btrfs_set_header_level(root->node, 0);
2563         } else if (slot == 0) {
2564                 struct btrfs_disk_key disk_key;
2565
2566                 btrfs_node_key(parent, &disk_key, 0);
2567                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2568                 if (wret)
2569                         ret = wret;
2570         }
2571         btrfs_mark_buffer_dirty(parent);
2572         return ret;
2573 }
2574
2575 /*
2576  * delete the item at the leaf level in path.  If that empties
2577  * the leaf, remove it from the tree
2578  */
2579 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2580                     struct btrfs_path *path, int slot, int nr)
2581 {
2582         struct extent_buffer *leaf;
2583         struct btrfs_item *item;
2584         int last_off;
2585         int dsize = 0;
2586         int ret = 0;
2587         int wret;
2588         int i;
2589         u32 nritems;
2590
2591         leaf = path->nodes[0];
2592         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2593
2594         for (i = 0; i < nr; i++)
2595                 dsize += btrfs_item_size_nr(leaf, slot + i);
2596
2597         nritems = btrfs_header_nritems(leaf);
2598
2599         if (slot + nr != nritems) {
2600                 int i;
2601                 int data_end = leaf_data_end(root, leaf);
2602
2603                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2604                               data_end + dsize,
2605                               btrfs_leaf_data(leaf) + data_end,
2606                               last_off - data_end);
2607
2608                 for (i = slot + nr; i < nritems; i++) {
2609                         u32 ioff;
2610
2611                         item = btrfs_item_nr(leaf, i);
2612                         if (!leaf->map_token) {
2613                                 map_extent_buffer(leaf, (unsigned long)item,
2614                                         sizeof(struct btrfs_item),
2615                                         &leaf->map_token, &leaf->kaddr,
2616                                         &leaf->map_start, &leaf->map_len,
2617                                         KM_USER1);
2618                         }
2619                         ioff = btrfs_item_offset(leaf, item);
2620                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2621                 }
2622
2623                 if (leaf->map_token) {
2624                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2625                         leaf->map_token = NULL;
2626                 }
2627
2628                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2629                               btrfs_item_nr_offset(slot + nr),
2630                               sizeof(struct btrfs_item) *
2631                               (nritems - slot - nr));
2632         }
2633         btrfs_set_header_nritems(leaf, nritems - nr);
2634         nritems -= nr;
2635
2636         /* delete the leaf if we've emptied it */
2637         if (nritems == 0) {
2638                 if (leaf == root->node) {
2639                         btrfs_set_header_level(leaf, 0);
2640                 } else {
2641                         u64 root_gen = btrfs_header_generation(path->nodes[1]);
2642                         clean_tree_block(trans, root, leaf);
2643                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
2644                         if (wret)
2645                                 ret = wret;
2646                         wret = btrfs_free_extent(trans, root,
2647                                          leaf->start, leaf->len,
2648                                          btrfs_header_owner(path->nodes[1]),
2649                                          root_gen, 0, 0, 1);
2650                         if (wret)
2651                                 ret = wret;
2652                 }
2653         } else {
2654                 int used = leaf_space_used(leaf, 0, nritems);
2655                 if (slot == 0) {
2656                         struct btrfs_disk_key disk_key;
2657
2658                         btrfs_item_key(leaf, &disk_key, 0);
2659                         wret = fixup_low_keys(trans, root, path,
2660                                               &disk_key, 1);
2661                         if (wret)
2662                                 ret = wret;
2663                 }
2664
2665                 /* delete the leaf if it is mostly empty */
2666                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2667                         /* push_leaf_left fixes the path.
2668                          * make sure the path still points to our leaf
2669                          * for possible call to del_ptr below
2670                          */
2671                         slot = path->slots[1];
2672                         extent_buffer_get(leaf);
2673
2674                         wret = push_leaf_left(trans, root, path, 1, 1);
2675                         if (wret < 0 && wret != -ENOSPC)
2676                                 ret = wret;
2677
2678                         if (path->nodes[0] == leaf &&
2679                             btrfs_header_nritems(leaf)) {
2680                                 wret = push_leaf_right(trans, root, path, 1, 1);
2681                                 if (wret < 0 && wret != -ENOSPC)
2682                                         ret = wret;
2683                         }
2684
2685                         if (btrfs_header_nritems(leaf) == 0) {
2686                                 u64 root_gen;
2687                                 u64 bytenr = leaf->start;
2688                                 u32 blocksize = leaf->len;
2689
2690                                 root_gen = btrfs_header_generation(
2691                                                            path->nodes[1]);
2692
2693                                 clean_tree_block(trans, root, leaf);
2694
2695                                 wret = del_ptr(trans, root, path, 1, slot);
2696                                 if (wret)
2697                                         ret = wret;
2698
2699                                 free_extent_buffer(leaf);
2700                                 wret = btrfs_free_extent(trans, root, bytenr,
2701                                              blocksize,
2702                                              btrfs_header_owner(path->nodes[1]),
2703                                              root_gen, 0, 0, 1);
2704                                 if (wret)
2705                                         ret = wret;
2706                         } else {
2707                                 btrfs_mark_buffer_dirty(leaf);
2708                                 free_extent_buffer(leaf);
2709                         }
2710                 } else {
2711                         btrfs_mark_buffer_dirty(leaf);
2712                 }
2713         }
2714         return ret;
2715 }
2716
2717 /*
2718  * walk up the tree as far as required to find the previous leaf.
2719  * returns 0 if it found something or 1 if there are no lesser leaves.
2720  * returns < 0 on io errors.
2721  */
2722 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2723 {
2724         u64 bytenr;
2725         int slot;
2726         int level = 1;
2727         struct extent_buffer *c;
2728         struct extent_buffer *next = NULL;
2729
2730         while(level < BTRFS_MAX_LEVEL) {
2731                 if (!path->nodes[level])
2732                         return 1;
2733
2734                 slot = path->slots[level];
2735                 c = path->nodes[level];
2736                 if (slot == 0) {
2737                         level++;
2738                         if (level == BTRFS_MAX_LEVEL)
2739                                 return 1;
2740                         continue;
2741                 }
2742                 slot--;
2743
2744                 bytenr = btrfs_node_blockptr(c, slot);
2745                 if (next)
2746                         free_extent_buffer(next);
2747
2748                 next = read_tree_block(root, bytenr,
2749                                        btrfs_level_size(root, level - 1));
2750                 break;
2751         }
2752         path->slots[level] = slot;
2753         while(1) {
2754                 level--;
2755                 c = path->nodes[level];
2756                 free_extent_buffer(c);
2757                 slot = btrfs_header_nritems(next);
2758                 if (slot != 0)
2759                         slot--;
2760                 path->nodes[level] = next;
2761                 path->slots[level] = slot;
2762                 if (!level)
2763                         break;
2764                 next = read_tree_block(root, btrfs_node_blockptr(next, slot),
2765                                        btrfs_level_size(root, level - 1));
2766         }
2767         return 0;
2768 }
2769
2770 /*
2771  * walk up the tree as far as required to find the next leaf.
2772  * returns 0 if it found something or 1 if there are no greater leaves.
2773  * returns < 0 on io errors.
2774  */
2775 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2776 {
2777         int slot;
2778         int level = 1;
2779         u64 bytenr;
2780         struct extent_buffer *c;
2781         struct extent_buffer *next = NULL;
2782
2783         while(level < BTRFS_MAX_LEVEL) {
2784                 if (!path->nodes[level])
2785                         return 1;
2786
2787                 slot = path->slots[level] + 1;
2788                 c = path->nodes[level];
2789                 if (slot >= btrfs_header_nritems(c)) {
2790                         level++;
2791                         if (level == BTRFS_MAX_LEVEL)
2792                                 return 1;
2793                         continue;
2794                 }
2795
2796                 bytenr = btrfs_node_blockptr(c, slot);
2797                 if (next)
2798                         free_extent_buffer(next);
2799
2800                 if (path->reada)
2801                         reada_for_search(root, path, level, slot, 0);
2802
2803                 next = read_tree_block(root, bytenr,
2804                                        btrfs_level_size(root, level -1));
2805                 break;
2806         }
2807         path->slots[level] = slot;
2808         while(1) {
2809                 level--;
2810                 c = path->nodes[level];
2811                 free_extent_buffer(c);
2812                 path->nodes[level] = next;
2813                 path->slots[level] = 0;
2814                 if (!level)
2815                         break;
2816                 if (path->reada)
2817                         reada_for_search(root, path, level, 0, 0);
2818                 next = read_tree_block(root, btrfs_node_blockptr(next, 0),
2819                                        btrfs_level_size(root, level - 1));
2820         }
2821         return 0;
2822 }
2823
2824 int btrfs_previous_item(struct btrfs_root *root,
2825                         struct btrfs_path *path, u64 min_objectid,
2826                         int type)
2827 {
2828         struct btrfs_key found_key;
2829         struct extent_buffer *leaf;
2830         int ret;
2831
2832         while(1) {
2833                 if (path->slots[0] == 0) {
2834                         ret = btrfs_prev_leaf(root, path);
2835                         if (ret != 0)
2836                                 return ret;
2837                 } else {
2838                         path->slots[0]--;
2839                 }
2840                 leaf = path->nodes[0];
2841                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2842                 if (found_key.type == type)
2843                         return 0;
2844         }
2845         return 1;
2846 }
2847