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, 1);
 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                 if (wret == 1) {
 848                         wret = push_node_left(trans, root, left, mid, 1);
 849                         if (wret < 0)
 850                                 ret = wret;
 851                 }
 852                 BUG_ON(wret == 1);
 853         }
 854         if (btrfs_header_nritems(mid) == 0) {
 855                 /* we've managed to empty the middle node, drop it */
 856                 u64 root_gen = btrfs_header_generation(parent);
 857                 u64 bytenr = mid->start;
 858                 u32 blocksize = mid->len;
 859                 clean_tree_block(trans, root, mid);
 860                 free_extent_buffer(mid);
 861                 mid = NULL;
 862                 wret = del_ptr(trans, root, path, level + 1, pslot);
 863                 if (wret)
 864                         ret = wret;
 865                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 866                                          btrfs_header_owner(parent),
 867                                          root_gen, 0, 0, 1);
 868                 if (wret)
 869                         ret = wret;
 870         } else {
 871                 /* update the parent key to reflect our changes */
 872                 struct btrfs_disk_key mid_key;
 873                 btrfs_node_key(mid, &mid_key, 0);
 874                 btrfs_set_node_key(parent, &mid_key, pslot);
 875                 btrfs_mark_buffer_dirty(parent);
 876         }
 877
 878         /* update the path */
 879         if (left) {
 880                 if (btrfs_header_nritems(left) > orig_slot) {
 881                         extent_buffer_get(left);
 882                         path->nodes[level] = left;
 883                         path->slots[level + 1] -= 1;
 884                         path->slots[level] = orig_slot;
 885                         if (mid)
 886                                 free_extent_buffer(mid);
 887                 } else {
 888                         orig_slot -= btrfs_header_nritems(left);
 889                         path->slots[level] = orig_slot;
 890                 }
 891         }
 892         /* double check we haven't messed things up */
 893         check_block(root, path, level);
 894         if (orig_ptr !=
 895             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 896                 BUG();
 897 enospc:
 898         if (right)
 899                 free_extent_buffer(right);
 900         if (left)
 901                 free_extent_buffer(left);
 902         return ret;
 903 }
 904
 905 /* returns zero if the push worked, non-zero otherwise */
 906 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
 907                                           struct btrfs_root *root,
 908                                           struct btrfs_path *path, int level)
 909 {
 910         struct extent_buffer *right = NULL;
 911         struct extent_buffer *mid;
 912         struct extent_buffer *left = NULL;
 913         struct extent_buffer *parent = NULL;
 914         int ret = 0;
 915         int wret;
 916         int pslot;
 917         int orig_slot = path->slots[level];
 918         u64 orig_ptr;
 919
 920         if (level == 0)
 921                 return 1;
 922
 923         mid = path->nodes[level];
 924         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 925         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 926
 927         if (level < BTRFS_MAX_LEVEL - 1)
 928                 parent = path->nodes[level + 1];
 929         pslot = path->slots[level + 1];
 930
 931         if (!parent)
 932                 return 1;
 933
 934         left = read_node_slot(root, parent, pslot - 1);
 935
 936         /* first, try to make some room in the middle buffer */
 937         if (left) {
 938                 u32 left_nr;
 939                 left_nr = btrfs_header_nritems(left);
 940                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 941                         wret = 1;
 942                 } else {
 943                         ret = btrfs_cow_block(trans, root, left, parent,
 944                                               pslot - 1, &left);
 945                         if (ret)
 946                                 wret = 1;
 947                         else {
 948                                 wret = push_node_left(trans, root,
 949                                                       left, mid, 0);
 950                         }
 951                 }
 952                 if (wret < 0)
 953                         ret = wret;
 954                 if (wret == 0) {
 955                         struct btrfs_disk_key disk_key;
 956                         orig_slot += left_nr;
 957                         btrfs_node_key(mid, &disk_key, 0);
 958                         btrfs_set_node_key(parent, &disk_key, pslot);
 959                         btrfs_mark_buffer_dirty(parent);
 960                         if (btrfs_header_nritems(left) > orig_slot) {
 961                                 path->nodes[level] = left;
 962                                 path->slots[level + 1] -= 1;
 963                                 path->slots[level] = orig_slot;
 964                                 free_extent_buffer(mid);
 965                         } else {
 966                                 orig_slot -=
 967                                         btrfs_header_nritems(left);
 968                                 path->slots[level] = orig_slot;
 969                                 free_extent_buffer(left);
 970                         }
 971                         return 0;
 972                 }
 973                 free_extent_buffer(left);
 974         }
 975         right= read_node_slot(root, parent, pslot + 1);
 976
 977         /*
 978          * then try to empty the right most buffer into the middle
 979          */
 980         if (right) {
 981                 u32 right_nr;
 982                 right_nr = btrfs_header_nritems(right);
 983                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 984                         wret = 1;
 985                 } else {
 986                         ret = btrfs_cow_block(trans, root, right,
 987                                               parent, pslot + 1,
 988                                               &right);
 989                         if (ret)
 990                                 wret = 1;
 991                         else {
 992                                 wret = balance_node_right(trans, root,
 993                                                           right, mid);
 994                         }
 995                 }
 996                 if (wret < 0)
 997                         ret = wret;
 998                 if (wret == 0) {
 999                         struct btrfs_disk_key disk_key;
1000
1001                         btrfs_node_key(right, &disk_key, 0);
1002                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
1003                         btrfs_mark_buffer_dirty(parent);
1004
1005                         if (btrfs_header_nritems(mid) <= orig_slot) {
1006                                 path->nodes[level] = right;
1007                                 path->slots[level + 1] += 1;
1008                                 path->slots[level] = orig_slot -
1009                                         btrfs_header_nritems(mid);
1010                                 free_extent_buffer(mid);
1011                         } else {
1012                                 free_extent_buffer(right);
1013                         }
1014                         return 0;
1015                 }
1016                 free_extent_buffer(right);
1017         }
1018         return 1;
1019 }
1020
1021 /*
1022  * readahead one full node of leaves
1023  */
1024 static void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
1025                              int level, int slot, u64 objectid)
1026 {
1027         struct extent_buffer *node;
1028         struct btrfs_disk_key disk_key;
1029         u32 nritems;
1030         u64 search;
1031         u64 lowest_read;
1032         u64 highest_read;
1033         u64 nread = 0;
1034         int direction = path->reada;
1035         struct extent_buffer *eb;
1036         u32 nr;
1037         u32 blocksize;
1038         u32 nscan = 0;
1039
1040         if (level != 1)
1041                 return;
1042
1043         if (!path->nodes[level])
1044                 return;
1045
1046         node = path->nodes[level];
1047         search = btrfs_node_blockptr(node, slot);
1048         blocksize = btrfs_level_size(root, level - 1);
1049         eb = btrfs_find_tree_block(root, search, blocksize);
1050         if (eb) {
1051                 free_extent_buffer(eb);
1052                 return;
1053         }
1054
1055         highest_read = search;
1056         lowest_read = search;
1057
1058         nritems = btrfs_header_nritems(node);
1059         nr = slot;
1060         while(1) {
1061                 if (direction < 0) {
1062                         if (nr == 0)
1063                                 break;
1064                         nr--;
1065                 } else if (direction > 0) {
1066                         nr++;
1067                         if (nr >= nritems)
1068                                 break;
1069                 }
1070                 if (path->reada < 0 && objectid) {
1071                         btrfs_node_key(node, &disk_key, nr);
1072                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1073                                 break;
1074                 }
1075                 search = btrfs_node_blockptr(node, nr);
1076                 if ((search >= lowest_read && search <= highest_read) ||
1077                     (search < lowest_read && lowest_read - search <= 32768) ||
1078                     (search > highest_read && search - highest_read <= 32768)) {
1079                         readahead_tree_block(root, search, blocksize);
1080                         nread += blocksize;
1081                 }
1082                 nscan++;
1083                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
1084                         break;
1085                 if(nread > (1024 * 1024) || nscan > 128)
1086                         break;
1087
1088                 if (search < lowest_read)
1089                         lowest_read = search;
1090                 if (search > highest_read)
1091                         highest_read = search;
1092         }
1093 }
1094 /*
1095  * look for key in the tree.  path is filled in with nodes along the way
1096  * if key is found, we return zero and you can find the item in the leaf
1097  * level of the path (level 0)
1098  *
1099  * If the key isn't found, the path points to the slot where it should
1100  * be inserted, and 1 is returned.  If there are other errors during the
1101  * search a negative error number is returned.
1102  *
1103  * if ins_len > 0, nodes and leaves will be split as we walk down the
1104  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1105  * possible)
1106  */
1107 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1108                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1109                       ins_len, int cow)
1110 {
1111         struct extent_buffer *b;
1112         u64 bytenr;
1113         u64 ptr_gen;
1114         int slot;
1115         int ret;
1116         int level;
1117         int should_reada = p->reada;
1118         u8 lowest_level = 0;
1119
1120         lowest_level = p->lowest_level;
1121         WARN_ON(lowest_level && ins_len);
1122         WARN_ON(p->nodes[0] != NULL);
1123         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
1124 again:
1125         b = root->node;
1126         extent_buffer_get(b);
1127         while (b) {
1128                 level = btrfs_header_level(b);
1129                 if (cow) {
1130                         int wret;
1131                         wret = btrfs_cow_block(trans, root, b,
1132                                                p->nodes[level + 1],
1133                                                p->slots[level + 1],
1134                                                &b);
1135                         if (wret) {
1136                                 free_extent_buffer(b);
1137                                 return wret;
1138                         }
1139                 }
1140                 BUG_ON(!cow && ins_len);
1141                 if (level != btrfs_header_level(b))
1142                         WARN_ON(1);
1143                 level = btrfs_header_level(b);
1144                 p->nodes[level] = b;
1145                 ret = check_block(root, p, level);
1146                 if (ret)
1147                         return -1;
1148                 ret = bin_search(b, key, level, &slot);
1149                 if (level != 0) {
1150                         if (ret && slot > 0)
1151                                 slot -= 1;
1152                         p->slots[level] = slot;
1153                         if (ins_len > 0 && btrfs_header_nritems(b) >=
1154                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1155                                 int sret = split_node(trans, root, p, level);
1156                                 BUG_ON(sret > 0);
1157                                 if (sret)
1158                                         return sret;
1159                                 b = p->nodes[level];
1160                                 slot = p->slots[level];
1161                         } else if (ins_len < 0) {
1162                                 int sret = balance_level(trans, root, p,
1163                                                          level);
1164                                 if (sret)
1165                                         return sret;
1166                                 b = p->nodes[level];
1167                                 if (!b) {
1168                                         btrfs_release_path(NULL, p);
1169                                         goto again;
1170                                 }
1171                                 slot = p->slots[level];
1172                                 BUG_ON(btrfs_header_nritems(b) == 1);
1173                         }
1174                         /* this is only true while dropping a snapshot */
1175                         if (level == lowest_level)
1176                                 break;
1177                         bytenr = btrfs_node_blockptr(b, slot);
1178                         ptr_gen = btrfs_node_ptr_generation(b, slot);
1179                         if (should_reada)
1180                                 reada_for_search(root, p, level, slot,
1181                                                  key->objectid);
1182                         b = read_tree_block(root, bytenr,
1183                                             btrfs_level_size(root, level - 1));
1184                         if (ptr_gen != btrfs_header_generation(b)) {
1185                                 printk("block %llu bad gen wanted %llu "
1186                                        "found %llu\n",
1187                                 (unsigned long long)b->start,
1188                                 (unsigned long long)ptr_gen,
1189                                 (unsigned long long)btrfs_header_generation(b));
1190                         }
1191                 } else {
1192                         p->slots[level] = slot;
1193                         if (ins_len > 0 && btrfs_leaf_free_space(root, b) <
1194                             sizeof(struct btrfs_item) + ins_len) {
1195                                 int sret = split_leaf(trans, root, key,
1196                                                       p, ins_len, ret == 0);
1197                                 BUG_ON(sret > 0);
1198                                 if (sret)
1199                                         return sret;
1200                         }
1201                         return ret;
1202                 }
1203         }
1204         return 1;
1205 }
1206
1207 /*
1208  * adjust the pointers going up the tree, starting at level
1209  * making sure the right key of each node is points to 'key'.
1210  * This is used after shifting pointers to the left, so it stops
1211  * fixing up pointers when a given leaf/node is not in slot 0 of the
1212  * higher levels
1213  *
1214  * If this fails to write a tree block, it returns -1, but continues
1215  * fixing up the blocks in ram so the tree is consistent.
1216  */
1217 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1218                           struct btrfs_root *root, struct btrfs_path *path,
1219                           struct btrfs_disk_key *key, int level)
1220 {
1221         int i;
1222         int ret = 0;
1223         struct extent_buffer *t;
1224
1225         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1226                 int tslot = path->slots[i];
1227                 if (!path->nodes[i])
1228                         break;
1229                 t = path->nodes[i];
1230                 btrfs_set_node_key(t, key, tslot);
1231                 btrfs_mark_buffer_dirty(path->nodes[i]);
1232                 if (tslot != 0)
1233                         break;
1234         }
1235         return ret;
1236 }
1237
1238 /*
1239  * try to push data from one node into the next node left in the
1240  * tree.
1241  *
1242  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1243  * error, and > 0 if there was no room in the left hand block.
1244  */
1245 static int push_node_left(struct btrfs_trans_handle *trans,
1246                           struct btrfs_root *root, struct extent_buffer *dst,
1247                           struct extent_buffer *src, int empty)
1248 {
1249         int push_items = 0;
1250         int src_nritems;
1251         int dst_nritems;
1252         int ret = 0;
1253
1254         src_nritems = btrfs_header_nritems(src);
1255         dst_nritems = btrfs_header_nritems(dst);
1256         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1257         WARN_ON(btrfs_header_generation(src) != trans->transid);
1258         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1259
1260         if (!empty && src_nritems <= 8)
1261                 return 1;
1262
1263         if (push_items <= 0) {
1264                 return 1;
1265         }
1266
1267         if (empty) {
1268                 push_items = min(src_nritems, push_items);
1269                 if (push_items < src_nritems) {
1270                         /* leave at least 8 pointers in the node if
1271                          * we aren't going to empty it
1272                          */
1273                         if (src_nritems - push_items < 8) {
1274                                 if (push_items <= 8)
1275                                         return 1;
1276                                 push_items -= 8;
1277                         }
1278                 }
1279         } else
1280                 push_items = min(src_nritems - 8, push_items);
1281
1282         copy_extent_buffer(dst, src,
1283                            btrfs_node_key_ptr_offset(dst_nritems),
1284                            btrfs_node_key_ptr_offset(0),
1285                            push_items * sizeof(struct btrfs_key_ptr));
1286
1287         if (push_items < src_nritems) {
1288                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1289                                       btrfs_node_key_ptr_offset(push_items),
1290                                       (src_nritems - push_items) *
1291                                       sizeof(struct btrfs_key_ptr));
1292         }
1293         btrfs_set_header_nritems(src, src_nritems - push_items);
1294         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1295         btrfs_mark_buffer_dirty(src);
1296         btrfs_mark_buffer_dirty(dst);
1297         return ret;
1298 }
1299
1300 /*
1301  * try to push data from one node into the next node right in the
1302  * tree.
1303  *
1304  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1305  * error, and > 0 if there was no room in the right hand block.
1306  *
1307  * this will  only push up to 1/2 the contents of the left node over
1308  */
1309 static int balance_node_right(struct btrfs_trans_handle *trans,
1310                               struct btrfs_root *root,
1311                               struct extent_buffer *dst,
1312                               struct extent_buffer *src)
1313 {
1314         int push_items = 0;
1315         int max_push;
1316         int src_nritems;
1317         int dst_nritems;
1318         int ret = 0;
1319
1320         WARN_ON(btrfs_header_generation(src) != trans->transid);
1321         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1322
1323         src_nritems = btrfs_header_nritems(src);
1324         dst_nritems = btrfs_header_nritems(dst);
1325         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1326         if (push_items <= 0) {
1327                 return 1;
1328         }
1329
1330         if (src_nritems < 4) {
1331                 return 1;
1332         }
1333
1334         max_push = src_nritems / 2 + 1;
1335         /* don't try to empty the node */
1336         if (max_push >= src_nritems) {
1337                 return 1;
1338         }
1339
1340         if (max_push < push_items)
1341                 push_items = max_push;
1342
1343         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1344                                       btrfs_node_key_ptr_offset(0),
1345                                       (dst_nritems) *
1346                                       sizeof(struct btrfs_key_ptr));
1347
1348         copy_extent_buffer(dst, src,
1349                            btrfs_node_key_ptr_offset(0),
1350                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1351                            push_items * sizeof(struct btrfs_key_ptr));
1352
1353         btrfs_set_header_nritems(src, src_nritems - push_items);
1354         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1355
1356         btrfs_mark_buffer_dirty(src);
1357         btrfs_mark_buffer_dirty(dst);
1358         return ret;
1359 }
1360
1361 /*
1362  * helper function to insert a new root level in the tree.
1363  * A new node is allocated, and a single item is inserted to
1364  * point to the existing root
1365  *
1366  * returns zero on success or < 0 on failure.
1367  */
1368 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1369                            struct btrfs_root *root,
1370                            struct btrfs_path *path, int level)
1371 {
1372         u64 root_gen;
1373         u64 lower_gen;
1374         struct extent_buffer *lower;
1375         struct extent_buffer *c;
1376         struct btrfs_disk_key lower_key;
1377
1378         BUG_ON(path->nodes[level]);
1379         BUG_ON(path->nodes[level-1] != root->node);
1380
1381         if (root->ref_cows)
1382                 root_gen = trans->transid;
1383         else
1384                 root_gen = 0;
1385
1386         lower = path->nodes[level-1];
1387         if (level == 1)
1388                 btrfs_item_key(lower, &lower_key, 0);
1389         else
1390                 btrfs_node_key(lower, &lower_key, 0);
1391
1392         c = __btrfs_alloc_free_block(trans, root, root->nodesize,
1393                                    root->root_key.objectid,
1394                                    root_gen, lower_key.objectid, level,
1395                                    root->node->start, 0);
1396         if (IS_ERR(c))
1397                 return PTR_ERR(c);
1398         memset_extent_buffer(c, 0, 0, root->nodesize);
1399         btrfs_set_header_nritems(c, 1);
1400         btrfs_set_header_level(c, level);
1401         btrfs_set_header_bytenr(c, c->start);
1402         btrfs_set_header_generation(c, trans->transid);
1403         btrfs_set_header_owner(c, root->root_key.objectid);
1404
1405         write_extent_buffer(c, root->fs_info->fsid,
1406                             (unsigned long)btrfs_header_fsid(c),
1407                             BTRFS_FSID_SIZE);
1408
1409         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1410                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
1411                             BTRFS_UUID_SIZE);
1412
1413         btrfs_set_node_key(c, &lower_key, 0);
1414         btrfs_set_node_blockptr(c, 0, lower->start);
1415         lower_gen = btrfs_header_generation(lower);
1416         WARN_ON(lower_gen == 0);
1417
1418         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1419
1420         btrfs_mark_buffer_dirty(c);
1421
1422         /* the super has an extra ref to root->node */
1423         free_extent_buffer(root->node);
1424         root->node = c;
1425         add_root_to_dirty_list(root);
1426         extent_buffer_get(c);
1427         path->nodes[level] = c;
1428         path->slots[level] = 0;
1429
1430         if (root->ref_cows && lower_gen != trans->transid) {
1431                 struct btrfs_path *back_path = btrfs_alloc_path();
1432                 int ret;
1433                 ret = btrfs_insert_extent_backref(trans,
1434                                                   root->fs_info->extent_root,
1435                                                   path, lower->start,
1436                                                   root->root_key.objectid,
1437                                                   trans->transid, 0, 0);
1438                 BUG_ON(ret);
1439                 btrfs_free_path(back_path);
1440         }
1441         return 0;
1442 }
1443
1444 /*
1445  * worker function to insert a single pointer in a node.
1446  * the node should have enough room for the pointer already
1447  *
1448  * slot and level indicate where you want the key to go, and
1449  * blocknr is the block the key points to.
1450  *
1451  * returns zero on success and < 0 on any error
1452  */
1453 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1454                       *root, struct btrfs_path *path, struct btrfs_disk_key
1455                       *key, u64 bytenr, int slot, int level)
1456 {
1457         struct extent_buffer *lower;
1458         int nritems;
1459
1460         BUG_ON(!path->nodes[level]);
1461         lower = path->nodes[level];
1462         nritems = btrfs_header_nritems(lower);
1463         if (slot > nritems)
1464                 BUG();
1465         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1466                 BUG();
1467         if (slot != nritems) {
1468                 memmove_extent_buffer(lower,
1469                               btrfs_node_key_ptr_offset(slot + 1),
1470                               btrfs_node_key_ptr_offset(slot),
1471                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1472         }
1473         btrfs_set_node_key(lower, key, slot);
1474         btrfs_set_node_blockptr(lower, slot, bytenr);
1475         WARN_ON(trans->transid == 0);
1476         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1477         btrfs_set_header_nritems(lower, nritems + 1);
1478         btrfs_mark_buffer_dirty(lower);
1479         return 0;
1480 }
1481
1482 /*
1483  * split the node at the specified level in path in two.
1484  * The path is corrected to point to the appropriate node after the split
1485  *
1486  * Before splitting this tries to make some room in the node by pushing
1487  * left and right, if either one works, it returns right away.
1488  *
1489  * returns 0 on success and < 0 on failure
1490  */
1491 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1492                       *root, struct btrfs_path *path, int level)
1493 {
1494         u64 root_gen;
1495         struct extent_buffer *c;
1496         struct extent_buffer *split;
1497         struct btrfs_disk_key disk_key;
1498         int mid;
1499         int ret;
1500         int wret;
1501         u32 c_nritems;
1502
1503         c = path->nodes[level];
1504         WARN_ON(btrfs_header_generation(c) != trans->transid);
1505         if (c == root->node) {
1506                 /* trying to split the root, lets make a new one */
1507                 ret = insert_new_root(trans, root, path, level + 1);
1508                 if (ret)
1509                         return ret;
1510         } else {
1511                 ret = push_nodes_for_insert(trans, root, path, level);
1512                 c = path->nodes[level];
1513                 if (!ret && btrfs_header_nritems(c) <
1514                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
1515                         return 0;
1516                 if (ret < 0)
1517                         return ret;
1518         }
1519
1520         c_nritems = btrfs_header_nritems(c);
1521         if (root->ref_cows)
1522                 root_gen = trans->transid;
1523         else
1524                 root_gen = 0;
1525
1526         btrfs_node_key(c, &disk_key, 0);
1527         split = __btrfs_alloc_free_block(trans, root, root->nodesize,
1528                                          root->root_key.objectid,
1529                                          root_gen,
1530                                          btrfs_disk_key_objectid(&disk_key),
1531                                          level, c->start, 0);
1532         if (IS_ERR(split))
1533                 return PTR_ERR(split);
1534
1535         btrfs_set_header_flags(split, btrfs_header_flags(c));
1536         btrfs_set_header_level(split, btrfs_header_level(c));
1537         btrfs_set_header_bytenr(split, split->start);
1538         btrfs_set_header_generation(split, trans->transid);
1539         btrfs_set_header_owner(split, root->root_key.objectid);
1540         btrfs_set_header_flags(split, 0);
1541         write_extent_buffer(split, root->fs_info->fsid,
1542                             (unsigned long)btrfs_header_fsid(split),
1543                             BTRFS_FSID_SIZE);
1544         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1545                             (unsigned long)btrfs_header_chunk_tree_uuid(split),
1546                             BTRFS_UUID_SIZE);
1547
1548         mid = (c_nritems + 1) / 2;
1549
1550         copy_extent_buffer(split, c,
1551                            btrfs_node_key_ptr_offset(0),
1552                            btrfs_node_key_ptr_offset(mid),
1553                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1554         btrfs_set_header_nritems(split, c_nritems - mid);
1555         btrfs_set_header_nritems(c, mid);
1556         ret = 0;
1557
1558         btrfs_mark_buffer_dirty(c);
1559         btrfs_mark_buffer_dirty(split);
1560
1561         btrfs_node_key(split, &disk_key, 0);
1562         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1563                           path->slots[level + 1] + 1,
1564                           level + 1);
1565         if (wret)
1566                 ret = wret;
1567
1568         if (path->slots[level] >= mid) {
1569                 path->slots[level] -= mid;
1570                 free_extent_buffer(c);
1571                 path->nodes[level] = split;
1572                 path->slots[level + 1] += 1;
1573         } else {
1574                 free_extent_buffer(split);
1575         }
1576         return ret;
1577 }
1578
1579 /*
1580  * how many bytes are required to store the items in a leaf.  start
1581  * and nr indicate which items in the leaf to check.  This totals up the
1582  * space used both by the item structs and the item data
1583  */
1584 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1585 {
1586         int data_len;
1587         int nritems = btrfs_header_nritems(l);
1588         int end = min(nritems, start + nr) - 1;
1589
1590         if (!nr)
1591                 return 0;
1592         data_len = btrfs_item_end_nr(l, start);
1593         data_len = data_len - btrfs_item_offset_nr(l, end);
1594         data_len += sizeof(struct btrfs_item) * nr;
1595         WARN_ON(data_len < 0);
1596         return data_len;
1597 }
1598
1599 /*
1600  * The space between the end of the leaf items and
1601  * the start of the leaf data.  IOW, how much room
1602  * the leaf has left for both items and data
1603  */
1604 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1605 {
1606         int nritems = btrfs_header_nritems(leaf);
1607         int ret;
1608         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1609         if (ret < 0) {
1610                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1611                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1612                        leaf_space_used(leaf, 0, nritems), nritems);
1613         }
1614         return ret;
1615 }
1616
1617 /*
1618  * push some data in the path leaf to the right, trying to free up at
1619  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1620  *
1621  * returns 1 if the push failed because the other node didn't have enough
1622  * room, 0 if everything worked out and < 0 if there were major errors.
1623  */
1624 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1625                            *root, struct btrfs_path *path, int data_size,
1626                            int empty)
1627 {
1628         struct extent_buffer *left = path->nodes[0];
1629         struct extent_buffer *right;
1630         struct extent_buffer *upper;
1631         struct btrfs_disk_key disk_key;
1632         int slot;
1633         u32 i;
1634         int free_space;
1635         int push_space = 0;
1636         int push_items = 0;
1637         struct btrfs_item *item;
1638         u32 left_nritems;
1639         u32 nr;
1640         u32 right_nritems;
1641         u32 data_end;
1642         u32 this_item_size;
1643         int ret;
1644
1645         slot = path->slots[1];
1646         if (!path->nodes[1]) {
1647                 return 1;
1648         }
1649         upper = path->nodes[1];
1650         if (slot >= btrfs_header_nritems(upper) - 1)
1651                 return 1;
1652
1653         right = read_tree_block(root, btrfs_node_blockptr(upper, slot + 1),
1654                                 root->leafsize);
1655         free_space = btrfs_leaf_free_space(root, right);
1656         if (free_space < data_size + sizeof(struct btrfs_item)) {
1657                 free_extent_buffer(right);
1658                 return 1;
1659         }
1660
1661         /* cow and double check */
1662         ret = btrfs_cow_block(trans, root, right, upper,
1663                               slot + 1, &right);
1664         if (ret) {
1665                 free_extent_buffer(right);
1666                 return 1;
1667         }
1668         free_space = btrfs_leaf_free_space(root, right);
1669         if (free_space < data_size + sizeof(struct btrfs_item)) {
1670                 free_extent_buffer(right);
1671                 return 1;
1672         }
1673
1674         left_nritems = btrfs_header_nritems(left);
1675         if (left_nritems == 0) {
1676                 free_extent_buffer(right);
1677                 return 1;
1678         }
1679
1680         if (empty)
1681                 nr = 0;
1682         else
1683                 nr = 1;
1684
1685         i = left_nritems - 1;
1686         while (i >= nr) {
1687                 item = btrfs_item_nr(left, i);
1688
1689                 if (path->slots[0] == i)
1690                         push_space += data_size + sizeof(*item);
1691
1692                 if (!left->map_token) {
1693                         map_extent_buffer(left, (unsigned long)item,
1694                                         sizeof(struct btrfs_item),
1695                                         &left->map_token, &left->kaddr,
1696                                         &left->map_start, &left->map_len,
1697                                         KM_USER1);
1698                 }
1699
1700                 this_item_size = btrfs_item_size(left, item);
1701                 if (this_item_size + sizeof(*item) + push_space > free_space)
1702                         break;
1703                 push_items++;
1704                 push_space += this_item_size + sizeof(*item);
1705                 if (i == 0)
1706                         break;
1707                 i--;
1708         }
1709         if (left->map_token) {
1710                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1711                 left->map_token = NULL;
1712         }
1713
1714         if (push_items == 0) {
1715                 free_extent_buffer(right);
1716                 return 1;
1717         }
1718
1719         if (!empty && push_items == left_nritems)
1720                 WARN_ON(1);
1721
1722         /* push left to right */
1723         right_nritems = btrfs_header_nritems(right);
1724
1725         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1726         push_space -= leaf_data_end(root, left);
1727
1728         /* make room in the right data area */
1729         data_end = leaf_data_end(root, right);
1730         memmove_extent_buffer(right,
1731                               btrfs_leaf_data(right) + data_end - push_space,
1732                               btrfs_leaf_data(right) + data_end,
1733                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1734
1735         /* copy from the left data area */
1736         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1737                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1738                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1739                      push_space);
1740
1741         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1742                               btrfs_item_nr_offset(0),
1743                               right_nritems * sizeof(struct btrfs_item));
1744
1745         /* copy the items from left to right */
1746         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1747                    btrfs_item_nr_offset(left_nritems - push_items),
1748                    push_items * sizeof(struct btrfs_item));
1749
1750         /* update the item pointers */
1751         right_nritems += push_items;
1752         btrfs_set_header_nritems(right, right_nritems);
1753         push_space = BTRFS_LEAF_DATA_SIZE(root);
1754         for (i = 0; i < right_nritems; i++) {
1755                 item = btrfs_item_nr(right, i);
1756                 if (!right->map_token) {
1757                         map_extent_buffer(right, (unsigned long)item,
1758                                         sizeof(struct btrfs_item),
1759                                         &right->map_token, &right->kaddr,
1760                                         &right->map_start, &right->map_len,
1761                                         KM_USER1);
1762                 }
1763                 push_space -= btrfs_item_size(right, item);
1764                 btrfs_set_item_offset(right, item, push_space);
1765         }
1766
1767         if (right->map_token) {
1768                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1769                 right->map_token = NULL;
1770         }
1771         left_nritems -= push_items;
1772         btrfs_set_header_nritems(left, left_nritems);
1773
1774         if (left_nritems)
1775                 btrfs_mark_buffer_dirty(left);
1776         btrfs_mark_buffer_dirty(right);
1777
1778         btrfs_item_key(right, &disk_key, 0);
1779         btrfs_set_node_key(upper, &disk_key, slot + 1);
1780         btrfs_mark_buffer_dirty(upper);
1781
1782         /* then fixup the leaf pointer in the path */
1783         if (path->slots[0] >= left_nritems) {
1784                 path->slots[0] -= left_nritems;
1785                 free_extent_buffer(path->nodes[0]);
1786                 path->nodes[0] = right;
1787                 path->slots[1] += 1;
1788         } else {
1789                 free_extent_buffer(right);
1790         }
1791         return 0;
1792 }
1793 /*
1794  * push some data in the path leaf to the left, trying to free up at
1795  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1796  */
1797 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1798                           *root, struct btrfs_path *path, int data_size,
1799                           int empty)
1800 {
1801         struct btrfs_disk_key disk_key;
1802         struct extent_buffer *right = path->nodes[0];
1803         struct extent_buffer *left;
1804         int slot;
1805         int i;
1806         int free_space;
1807         int push_space = 0;
1808         int push_items = 0;
1809         struct btrfs_item *item;
1810         u32 old_left_nritems;
1811         u32 right_nritems;
1812         u32 nr;
1813         int ret = 0;
1814         int wret;
1815         u32 this_item_size;
1816         u32 old_left_item_size;
1817
1818         slot = path->slots[1];
1819         if (slot == 0)
1820                 return 1;
1821         if (!path->nodes[1])
1822                 return 1;
1823
1824         right_nritems = btrfs_header_nritems(right);
1825         if (right_nritems == 0) {
1826                 return 1;
1827         }
1828
1829         left = read_tree_block(root, btrfs_node_blockptr(path->nodes[1],
1830                                slot - 1), root->leafsize);
1831         free_space = btrfs_leaf_free_space(root, left);
1832         if (free_space < data_size + sizeof(struct btrfs_item)) {
1833                 free_extent_buffer(left);
1834                 return 1;
1835         }
1836
1837         /* cow and double check */
1838         ret = btrfs_cow_block(trans, root, left,
1839                               path->nodes[1], slot - 1, &left);
1840         if (ret) {
1841                 /* we hit -ENOSPC, but it isn't fatal here */
1842                 free_extent_buffer(left);
1843                 return 1;
1844         }
1845
1846         free_space = btrfs_leaf_free_space(root, left);
1847         if (free_space < data_size + sizeof(struct btrfs_item)) {
1848                 free_extent_buffer(left);
1849                 return 1;
1850         }
1851
1852         if (empty)
1853                 nr = right_nritems;
1854         else
1855                 nr = right_nritems - 1;
1856
1857         for (i = 0; i < nr; i++) {
1858                 item = btrfs_item_nr(right, i);
1859                 if (!right->map_token) {
1860                         map_extent_buffer(right, (unsigned long)item,
1861                                         sizeof(struct btrfs_item),
1862                                         &right->map_token, &right->kaddr,
1863                                         &right->map_start, &right->map_len,
1864                                         KM_USER1);
1865                 }
1866
1867                 if (path->slots[0] == i)
1868                         push_space += data_size + sizeof(*item);
1869
1870                 this_item_size = btrfs_item_size(right, item);
1871                 if (this_item_size + sizeof(*item) + push_space > free_space)
1872                         break;
1873
1874                 push_items++;
1875                 push_space += this_item_size + sizeof(*item);
1876         }
1877
1878         if (right->map_token) {
1879                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1880                 right->map_token = NULL;
1881         }
1882
1883         if (push_items == 0) {
1884                 free_extent_buffer(left);
1885                 return 1;
1886         }
1887         if (!empty && push_items == btrfs_header_nritems(right))
1888                 WARN_ON(1);
1889
1890         /* push data from right to left */
1891         copy_extent_buffer(left, right,
1892                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1893                            btrfs_item_nr_offset(0),
1894                            push_items * sizeof(struct btrfs_item));
1895
1896         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1897                      btrfs_item_offset_nr(right, push_items -1);
1898
1899         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1900                      leaf_data_end(root, left) - push_space,
1901                      btrfs_leaf_data(right) +
1902                      btrfs_item_offset_nr(right, push_items - 1),
1903                      push_space);
1904         old_left_nritems = btrfs_header_nritems(left);
1905         BUG_ON(old_left_nritems < 0);
1906
1907         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1908         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1909                 u32 ioff;
1910
1911                 item = btrfs_item_nr(left, i);
1912                 if (!left->map_token) {
1913                         map_extent_buffer(left, (unsigned long)item,
1914                                         sizeof(struct btrfs_item),
1915                                         &left->map_token, &left->kaddr,
1916                                         &left->map_start, &left->map_len,
1917                                         KM_USER1);
1918                 }
1919
1920                 ioff = btrfs_item_offset(left, item);
1921                 btrfs_set_item_offset(left, item,
1922                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1923         }
1924         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1925         if (left->map_token) {
1926                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1927                 left->map_token = NULL;
1928         }
1929
1930         /* fixup right node */
1931         if (push_items > right_nritems) {
1932                 printk("push items %d nr %u\n", push_items, right_nritems);
1933                 WARN_ON(1);
1934         }
1935
1936         if (push_items < right_nritems) {
1937                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1938                                                   leaf_data_end(root, right);
1939                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
1940                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1941                                       btrfs_leaf_data(right) +
1942                                       leaf_data_end(root, right), push_space);
1943
1944                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1945                               btrfs_item_nr_offset(push_items),
1946                              (btrfs_header_nritems(right) - push_items) *
1947                              sizeof(struct btrfs_item));
1948         }
1949         right_nritems -= push_items;
1950         btrfs_set_header_nritems(right, right_nritems);
1951         push_space = BTRFS_LEAF_DATA_SIZE(root);
1952         for (i = 0; i < right_nritems; i++) {
1953                 item = btrfs_item_nr(right, i);
1954
1955                 if (!right->map_token) {
1956                         map_extent_buffer(right, (unsigned long)item,
1957                                         sizeof(struct btrfs_item),
1958                                         &right->map_token, &right->kaddr,
1959                                         &right->map_start, &right->map_len,
1960                                         KM_USER1);
1961                 }
1962
1963                 push_space = push_space - btrfs_item_size(right, item);
1964                 btrfs_set_item_offset(right, item, push_space);
1965         }
1966         if (right->map_token) {
1967                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1968                 right->map_token = NULL;
1969         }
1970
1971         btrfs_mark_buffer_dirty(left);
1972         if (right_nritems)
1973                 btrfs_mark_buffer_dirty(right);
1974
1975         btrfs_item_key(right, &disk_key, 0);
1976         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1977         if (wret)
1978                 ret = wret;
1979
1980         /* then fixup the leaf pointer in the path */
1981         if (path->slots[0] < push_items) {
1982                 path->slots[0] += old_left_nritems;
1983                 free_extent_buffer(path->nodes[0]);
1984                 path->nodes[0] = left;
1985                 path->slots[1] -= 1;
1986         } else {
1987                 free_extent_buffer(left);
1988                 path->slots[0] -= push_items;
1989         }
1990         BUG_ON(path->slots[0] < 0);
1991         return ret;
1992 }
1993
1994 /*
1995  * split the path's leaf in two, making sure there is at least data_size
1996  * available for the resulting leaf level of the path.
1997  *
1998  * returns 0 if all went well and < 0 on failure.
1999  */
2000 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
2001                       *root, struct btrfs_key *ins_key,
2002                       struct btrfs_path *path, int data_size, int extend)
2003 {
2004         u64 root_gen;
2005         struct extent_buffer *l;
2006         u32 nritems;
2007         int mid;
2008         int slot;
2009         struct extent_buffer *right;
2010         int space_needed = data_size + sizeof(struct btrfs_item);
2011         int data_copy_size;
2012         int rt_data_off;
2013         int i;
2014         int ret = 0;
2015         int wret;
2016         int double_split;
2017         int num_doubles = 0;
2018         struct btrfs_disk_key disk_key;
2019
2020         if (extend)
2021                 space_needed = data_size;
2022
2023         if (root->ref_cows)
2024                 root_gen = trans->transid;
2025         else
2026                 root_gen = 0;
2027
2028         /* first try to make some room by pushing left and right */
2029         if (ins_key->type != BTRFS_DIR_ITEM_KEY) {
2030                 wret = push_leaf_right(trans, root, path, data_size, 0);
2031                 if (wret < 0) {
2032                         return wret;
2033                 }
2034                 if (wret) {
2035                         wret = push_leaf_left(trans, root, path, data_size, 0);
2036                         if (wret < 0)
2037                                 return wret;
2038                 }
2039                 l = path->nodes[0];
2040
2041                 /* did the pushes work? */
2042                 if (btrfs_leaf_free_space(root, l) >= space_needed)
2043                         return 0;
2044         }
2045
2046         if (!path->nodes[1]) {
2047                 ret = insert_new_root(trans, root, path, 1);
2048                 if (ret)
2049                         return ret;
2050         }
2051 again:
2052         double_split = 0;
2053         l = path->nodes[0];
2054         slot = path->slots[0];
2055         nritems = btrfs_header_nritems(l);
2056         mid = (nritems + 1)/ 2;
2057
2058         btrfs_item_key(l, &disk_key, 0);
2059
2060         right = __btrfs_alloc_free_block(trans, root, root->leafsize,
2061                                          root->root_key.objectid,
2062                                          root_gen, disk_key.objectid, 0,
2063                                          l->start, 0);
2064         if (IS_ERR(right)) {
2065                 BUG_ON(1);
2066                 return PTR_ERR(right);
2067         }
2068
2069         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2070         btrfs_set_header_bytenr(right, right->start);
2071         btrfs_set_header_generation(right, trans->transid);
2072         btrfs_set_header_owner(right, root->root_key.objectid);
2073         btrfs_set_header_level(right, 0);
2074         write_extent_buffer(right, root->fs_info->fsid,
2075                             (unsigned long)btrfs_header_fsid(right),
2076                             BTRFS_FSID_SIZE);
2077
2078         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2079                             (unsigned long)btrfs_header_chunk_tree_uuid(right),
2080                             BTRFS_UUID_SIZE);
2081         if (mid <= slot) {
2082                 if (nritems == 1 ||
2083                     leaf_space_used(l, mid, nritems - mid) + space_needed >
2084                         BTRFS_LEAF_DATA_SIZE(root)) {
2085                         if (slot >= nritems) {
2086                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2087                                 btrfs_set_header_nritems(right, 0);
2088                                 wret = insert_ptr(trans, root, path,
2089                                                   &disk_key, right->start,
2090                                                   path->slots[1] + 1, 1);
2091                                 if (wret)
2092                                         ret = wret;
2093                                 free_extent_buffer(path->nodes[0]);
2094                                 path->nodes[0] = right;
2095                                 path->slots[0] = 0;
2096                                 path->slots[1] += 1;
2097                                 btrfs_mark_buffer_dirty(right);
2098                                 return ret;
2099                         }
2100                         mid = slot;
2101                         if (mid != nritems &&
2102                             leaf_space_used(l, mid, nritems - mid) +
2103                             space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2104                                 double_split = 1;
2105                         }
2106                 }
2107         } else {
2108                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
2109                         BTRFS_LEAF_DATA_SIZE(root)) {
2110                         if (!extend && slot == 0) {
2111                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2112                                 btrfs_set_header_nritems(right, 0);
2113                                 wret = insert_ptr(trans, root, path,
2114                                                   &disk_key,
2115                                                   right->start,
2116                                                   path->slots[1], 1);
2117                                 if (wret)
2118                                         ret = wret;
2119                                 free_extent_buffer(path->nodes[0]);
2120                                 path->nodes[0] = right;
2121                                 path->slots[0] = 0;
2122                                 if (path->slots[1] == 0) {
2123                                         wret = fixup_low_keys(trans, root,
2124                                                    path, &disk_key, 1);
2125                                         if (wret)
2126                                                 ret = wret;
2127                                 }
2128                                 btrfs_mark_buffer_dirty(right);
2129                                 return ret;
2130                         } else if (extend && slot == 0) {
2131                                 mid = 1;
2132                         } else {
2133                                 mid = slot;
2134                                 if (mid != nritems &&
2135                                     leaf_space_used(l, mid, nritems - mid) +
2136                                     space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2137                                         double_split = 1;
2138                                 }
2139                         }
2140                 }
2141         }
2142         nritems = nritems - mid;
2143         btrfs_set_header_nritems(right, nritems);
2144         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2145
2146         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2147                            btrfs_item_nr_offset(mid),
2148                            nritems * sizeof(struct btrfs_item));
2149
2150         copy_extent_buffer(right, l,
2151                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2152                      data_copy_size, btrfs_leaf_data(l) +
2153                      leaf_data_end(root, l), data_copy_size);
2154
2155         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2156                       btrfs_item_end_nr(l, mid);
2157
2158         for (i = 0; i < nritems; i++) {
2159                 struct btrfs_item *item = btrfs_item_nr(right, i);
2160                 u32 ioff;
2161
2162                 if (!right->map_token) {
2163                         map_extent_buffer(right, (unsigned long)item,
2164                                         sizeof(struct btrfs_item),
2165                                         &right->map_token, &right->kaddr,
2166                                         &right->map_start, &right->map_len,
2167                                         KM_USER1);
2168                 }
2169
2170                 ioff = btrfs_item_offset(right, item);
2171                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2172         }
2173
2174         if (right->map_token) {
2175                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2176                 right->map_token = NULL;
2177         }
2178
2179         btrfs_set_header_nritems(l, mid);
2180         ret = 0;
2181         btrfs_item_key(right, &disk_key, 0);
2182         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2183                           path->slots[1] + 1, 1);
2184         if (wret)
2185                 ret = wret;
2186
2187         btrfs_mark_buffer_dirty(right);
2188         btrfs_mark_buffer_dirty(l);
2189         BUG_ON(path->slots[0] != slot);
2190
2191         if (mid <= slot) {
2192                 free_extent_buffer(path->nodes[0]);
2193                 path->nodes[0] = right;
2194                 path->slots[0] -= mid;
2195                 path->slots[1] += 1;
2196         } else
2197                 free_extent_buffer(right);
2198
2199         BUG_ON(path->slots[0] < 0);
2200
2201         if (double_split) {
2202                 BUG_ON(num_doubles != 0);
2203                 num_doubles++;
2204                 goto again;
2205         }
2206         return ret;
2207 }
2208
2209 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2210                         struct btrfs_root *root,
2211                         struct btrfs_path *path,
2212                         u32 new_size, int from_end)
2213 {
2214         int ret = 0;
2215         int slot;
2216         int slot_orig;
2217         struct extent_buffer *leaf;
2218         struct btrfs_item *item;
2219         u32 nritems;
2220         unsigned int data_end;
2221         unsigned int old_data_start;
2222         unsigned int old_size;
2223         unsigned int size_diff;
2224         int i;
2225
2226         slot_orig = path->slots[0];
2227         leaf = path->nodes[0];
2228         slot = path->slots[0];
2229
2230         old_size = btrfs_item_size_nr(leaf, slot);
2231         if (old_size == new_size)
2232                 return 0;
2233
2234         nritems = btrfs_header_nritems(leaf);
2235         data_end = leaf_data_end(root, leaf);
2236
2237         old_data_start = btrfs_item_offset_nr(leaf, slot);
2238
2239         size_diff = old_size - new_size;
2240
2241         BUG_ON(slot < 0);
2242         BUG_ON(slot >= nritems);
2243
2244         /*
2245          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2246          */
2247         /* first correct the data pointers */
2248         for (i = slot; i < nritems; i++) {
2249                 u32 ioff;
2250                 item = btrfs_item_nr(leaf, i);
2251
2252                 if (!leaf->map_token) {
2253                         map_extent_buffer(leaf, (unsigned long)item,
2254                                         sizeof(struct btrfs_item),
2255                                         &leaf->map_token, &leaf->kaddr,
2256                                         &leaf->map_start, &leaf->map_len,
2257                                         KM_USER1);
2258                 }
2259
2260                 ioff = btrfs_item_offset(leaf, item);
2261                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2262         }
2263
2264         if (leaf->map_token) {
2265                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2266                 leaf->map_token = NULL;
2267         }
2268
2269         /* shift the data */
2270         if (from_end) {
2271                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2272                               data_end + size_diff, btrfs_leaf_data(leaf) +
2273                               data_end, old_data_start + new_size - data_end);
2274         } else {
2275                 struct btrfs_disk_key disk_key;
2276                 u64 offset;
2277
2278                 btrfs_item_key(leaf, &disk_key, slot);
2279
2280                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2281                         unsigned long ptr;
2282                         struct btrfs_file_extent_item *fi;
2283
2284                         fi = btrfs_item_ptr(leaf, slot,
2285                                             struct btrfs_file_extent_item);
2286                         fi = (struct btrfs_file_extent_item *)(
2287                              (unsigned long)fi - size_diff);
2288
2289                         if (btrfs_file_extent_type(leaf, fi) ==
2290                             BTRFS_FILE_EXTENT_INLINE) {
2291                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2292                                 memmove_extent_buffer(leaf, ptr,
2293                                         (unsigned long)fi,
2294                                         offsetof(struct btrfs_file_extent_item,
2295                                                  disk_bytenr));
2296                         }
2297                 }
2298
2299                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2300                               data_end + size_diff, btrfs_leaf_data(leaf) +
2301                               data_end, old_data_start - data_end);
2302
2303                 offset = btrfs_disk_key_offset(&disk_key);
2304                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2305                 btrfs_set_item_key(leaf, &disk_key, slot);
2306                 if (slot == 0)
2307                         fixup_low_keys(trans, root, path, &disk_key, 1);
2308         }
2309
2310         item = btrfs_item_nr(leaf, slot);
2311         btrfs_set_item_size(leaf, item, new_size);
2312         btrfs_mark_buffer_dirty(leaf);
2313
2314         ret = 0;
2315         if (btrfs_leaf_free_space(root, leaf) < 0) {
2316                 btrfs_print_leaf(root, leaf);
2317                 BUG();
2318         }
2319         return ret;
2320 }
2321
2322 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2323                       struct btrfs_root *root, struct btrfs_path *path,
2324                       u32 data_size)
2325 {
2326         int ret = 0;
2327         int slot;
2328         int slot_orig;
2329         struct extent_buffer *leaf;
2330         struct btrfs_item *item;
2331         u32 nritems;
2332         unsigned int data_end;
2333         unsigned int old_data;
2334         unsigned int old_size;
2335         int i;
2336
2337         slot_orig = path->slots[0];
2338         leaf = path->nodes[0];
2339
2340         nritems = btrfs_header_nritems(leaf);
2341         data_end = leaf_data_end(root, leaf);
2342
2343         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2344                 btrfs_print_leaf(root, leaf);
2345                 BUG();
2346         }
2347         slot = path->slots[0];
2348         old_data = btrfs_item_end_nr(leaf, slot);
2349
2350         BUG_ON(slot < 0);
2351         if (slot >= nritems) {
2352                 btrfs_print_leaf(root, leaf);
2353                 printk("slot %d too large, nritems %d\n", slot, nritems);
2354                 BUG_ON(1);
2355         }
2356
2357         /*
2358          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2359          */
2360         /* first correct the data pointers */
2361         for (i = slot; i < nritems; i++) {
2362                 u32 ioff;
2363                 item = btrfs_item_nr(leaf, i);
2364
2365                 if (!leaf->map_token) {
2366                         map_extent_buffer(leaf, (unsigned long)item,
2367                                         sizeof(struct btrfs_item),
2368                                         &leaf->map_token, &leaf->kaddr,
2369                                         &leaf->map_start, &leaf->map_len,
2370                                         KM_USER1);
2371                 }
2372                 ioff = btrfs_item_offset(leaf, item);
2373                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2374         }
2375
2376         if (leaf->map_token) {
2377                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2378                 leaf->map_token = NULL;
2379         }
2380
2381         /* shift the data */
2382         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2383                       data_end - data_size, btrfs_leaf_data(leaf) +
2384                       data_end, old_data - data_end);
2385
2386         data_end = old_data;
2387         old_size = btrfs_item_size_nr(leaf, slot);
2388         item = btrfs_item_nr(leaf, slot);
2389         btrfs_set_item_size(leaf, item, old_size + data_size);
2390         btrfs_mark_buffer_dirty(leaf);
2391
2392         ret = 0;
2393         if (btrfs_leaf_free_space(root, leaf) < 0) {
2394                 btrfs_print_leaf(root, leaf);
2395                 BUG();
2396         }
2397         return ret;
2398 }
2399
2400 /*
2401  * Given a key and some data, insert an item into the tree.
2402  * This does all the path init required, making room in the tree if needed.
2403  */
2404 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2405                             struct btrfs_root *root,
2406                             struct btrfs_path *path,
2407                             struct btrfs_key *cpu_key, u32 *data_size,
2408                             int nr)
2409 {
2410         struct extent_buffer *leaf;
2411         struct btrfs_item *item;
2412         int ret = 0;
2413         int slot;
2414         int slot_orig;
2415         int i;
2416         u32 nritems;
2417         u32 total_size = 0;
2418         u32 total_data = 0;
2419         unsigned int data_end;
2420         struct btrfs_disk_key disk_key;
2421
2422         for (i = 0; i < nr; i++) {
2423                 total_data += data_size[i];
2424         }
2425
2426         /* create a root if there isn't one */
2427         if (!root->node)
2428                 BUG();
2429
2430         total_size = total_data + (nr - 1) * sizeof(struct btrfs_item);
2431         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2432         if (ret == 0) {
2433                 return -EEXIST;
2434         }
2435         if (ret < 0)
2436                 goto out;
2437
2438         slot_orig = path->slots[0];
2439         leaf = path->nodes[0];
2440
2441         nritems = btrfs_header_nritems(leaf);
2442         data_end = leaf_data_end(root, leaf);
2443
2444         if (btrfs_leaf_free_space(root, leaf) <
2445             sizeof(struct btrfs_item) + total_size) {
2446                 btrfs_print_leaf(root, leaf);
2447                 printk("not enough freespace need %u have %d\n",
2448                        total_size, btrfs_leaf_free_space(root, leaf));
2449                 BUG();
2450         }
2451
2452         slot = path->slots[0];
2453         BUG_ON(slot < 0);
2454
2455         if (slot != nritems) {
2456                 int i;
2457                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2458
2459                 if (old_data < data_end) {
2460                         btrfs_print_leaf(root, leaf);
2461                         printk("slot %d old_data %d data_end %d\n",
2462                                slot, old_data, data_end);
2463                         BUG_ON(1);
2464                 }
2465                 /*
2466                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2467                  */
2468                 /* first correct the data pointers */
2469                 WARN_ON(leaf->map_token);
2470                 for (i = slot; i < nritems; i++) {
2471                         u32 ioff;
2472
2473                         item = btrfs_item_nr(leaf, i);
2474                         if (!leaf->map_token) {
2475                                 map_extent_buffer(leaf, (unsigned long)item,
2476                                         sizeof(struct btrfs_item),
2477                                         &leaf->map_token, &leaf->kaddr,
2478                                         &leaf->map_start, &leaf->map_len,
2479                                         KM_USER1);
2480                         }
2481
2482                         ioff = btrfs_item_offset(leaf, item);
2483                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2484                 }
2485                 if (leaf->map_token) {
2486                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2487                         leaf->map_token = NULL;
2488                 }
2489
2490                 /* shift the items */
2491                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2492                               btrfs_item_nr_offset(slot),
2493                               (nritems - slot) * sizeof(struct btrfs_item));
2494
2495                 /* shift the data */
2496                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2497                               data_end - total_data, btrfs_leaf_data(leaf) +
2498                               data_end, old_data - data_end);
2499                 data_end = old_data;
2500         }
2501
2502         /* setup the item for the new data */
2503         for (i = 0; i < nr; i++) {
2504                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2505                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2506                 item = btrfs_item_nr(leaf, slot + i);
2507                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2508                 data_end -= data_size[i];
2509                 btrfs_set_item_size(leaf, item, data_size[i]);
2510         }
2511         btrfs_set_header_nritems(leaf, nritems + nr);
2512         btrfs_mark_buffer_dirty(leaf);
2513
2514         ret = 0;
2515         if (slot == 0) {
2516                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2517                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2518         }
2519
2520         if (btrfs_leaf_free_space(root, leaf) < 0) {
2521                 btrfs_print_leaf(root, leaf);
2522                 BUG();
2523         }
2524
2525 out:
2526         return ret;
2527 }
2528
2529 /*
2530  * Given a key and some data, insert an item into the tree.
2531  * This does all the path init required, making room in the tree if needed.
2532  */
2533 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2534                       *root, struct btrfs_key *cpu_key, void *data, u32
2535                       data_size)
2536 {
2537         int ret = 0;
2538         struct btrfs_path *path;
2539         struct extent_buffer *leaf;
2540         unsigned long ptr;
2541
2542         path = btrfs_alloc_path();
2543         BUG_ON(!path);
2544         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2545         if (!ret) {
2546                 leaf = path->nodes[0];
2547                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2548                 write_extent_buffer(leaf, data, ptr, data_size);
2549                 btrfs_mark_buffer_dirty(leaf);
2550         }
2551         btrfs_free_path(path);
2552         return ret;
2553 }
2554
2555 /*
2556  * delete the pointer from a given node.
2557  *
2558  * If the delete empties a node, the node is removed from the tree,
2559  * continuing all the way the root if required.  The root is converted into
2560  * a leaf if all the nodes are emptied.
2561  */
2562 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2563                    struct btrfs_path *path, int level, int slot)
2564 {
2565         struct extent_buffer *parent = path->nodes[level];
2566         u32 nritems;
2567         int ret = 0;
2568         int wret;
2569
2570         nritems = btrfs_header_nritems(parent);
2571         if (slot != nritems -1) {
2572                 memmove_extent_buffer(parent,
2573                               btrfs_node_key_ptr_offset(slot),
2574                               btrfs_node_key_ptr_offset(slot + 1),
2575                               sizeof(struct btrfs_key_ptr) *
2576                               (nritems - slot - 1));
2577         }
2578         nritems--;
2579         btrfs_set_header_nritems(parent, nritems);
2580         if (nritems == 0 && parent == root->node) {
2581                 BUG_ON(btrfs_header_level(root->node) != 1);
2582                 /* just turn the root into a leaf and break */
2583                 btrfs_set_header_level(root->node, 0);
2584         } else if (slot == 0) {
2585                 struct btrfs_disk_key disk_key;
2586
2587                 btrfs_node_key(parent, &disk_key, 0);
2588                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2589                 if (wret)
2590                         ret = wret;
2591         }
2592         btrfs_mark_buffer_dirty(parent);
2593         return ret;
2594 }
2595
2596 /*
2597  * delete the item at the leaf level in path.  If that empties
2598  * the leaf, remove it from the tree
2599  */
2600 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2601                     struct btrfs_path *path, int slot, int nr)
2602 {
2603         struct extent_buffer *leaf;
2604         struct btrfs_item *item;
2605         int last_off;
2606         int dsize = 0;
2607         int ret = 0;
2608         int wret;
2609         int i;
2610         u32 nritems;
2611
2612         leaf = path->nodes[0];
2613         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2614
2615         for (i = 0; i < nr; i++)
2616                 dsize += btrfs_item_size_nr(leaf, slot + i);
2617
2618         nritems = btrfs_header_nritems(leaf);
2619
2620         if (slot + nr != nritems) {
2621                 int i;
2622                 int data_end = leaf_data_end(root, leaf);
2623
2624                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2625                               data_end + dsize,
2626                               btrfs_leaf_data(leaf) + data_end,
2627                               last_off - data_end);
2628
2629                 for (i = slot + nr; i < nritems; i++) {
2630                         u32 ioff;
2631
2632                         item = btrfs_item_nr(leaf, i);
2633                         if (!leaf->map_token) {
2634                                 map_extent_buffer(leaf, (unsigned long)item,
2635                                         sizeof(struct btrfs_item),
2636                                         &leaf->map_token, &leaf->kaddr,
2637                                         &leaf->map_start, &leaf->map_len,
2638                                         KM_USER1);
2639                         }
2640                         ioff = btrfs_item_offset(leaf, item);
2641                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2642                 }
2643
2644                 if (leaf->map_token) {
2645                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2646                         leaf->map_token = NULL;
2647                 }
2648
2649                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2650                               btrfs_item_nr_offset(slot + nr),
2651                               sizeof(struct btrfs_item) *
2652                               (nritems - slot - nr));
2653         }
2654         btrfs_set_header_nritems(leaf, nritems - nr);
2655         nritems -= nr;
2656
2657         /* delete the leaf if we've emptied it */
2658         if (nritems == 0) {
2659                 if (leaf == root->node) {
2660                         btrfs_set_header_level(leaf, 0);
2661                 } else {
2662                         u64 root_gen = btrfs_header_generation(path->nodes[1]);
2663                         clean_tree_block(trans, root, leaf);
2664                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
2665                         if (wret)
2666                                 ret = wret;
2667                         wret = btrfs_free_extent(trans, root,
2668                                          leaf->start, leaf->len,
2669                                          btrfs_header_owner(path->nodes[1]),
2670                                          root_gen, 0, 0, 1);
2671                         if (wret)
2672                                 ret = wret;
2673                 }
2674         } else {
2675                 int used = leaf_space_used(leaf, 0, nritems);
2676                 if (slot == 0) {
2677                         struct btrfs_disk_key disk_key;
2678
2679                         btrfs_item_key(leaf, &disk_key, 0);
2680                         wret = fixup_low_keys(trans, root, path,
2681                                               &disk_key, 1);
2682                         if (wret)
2683                                 ret = wret;
2684                 }
2685
2686                 /* delete the leaf if it is mostly empty */
2687                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2688                         /* push_leaf_left fixes the path.
2689                          * make sure the path still points to our leaf
2690                          * for possible call to del_ptr below
2691                          */
2692                         slot = path->slots[1];
2693                         extent_buffer_get(leaf);
2694
2695                         wret = push_leaf_left(trans, root, path, 1, 1);
2696                         if (wret < 0 && wret != -ENOSPC)
2697                                 ret = wret;
2698
2699                         if (path->nodes[0] == leaf &&
2700                             btrfs_header_nritems(leaf)) {
2701                                 wret = push_leaf_right(trans, root, path, 1, 1);
2702                                 if (wret < 0 && wret != -ENOSPC)
2703                                         ret = wret;
2704                         }
2705
2706                         if (btrfs_header_nritems(leaf) == 0) {
2707                                 u64 root_gen;
2708                                 u64 bytenr = leaf->start;
2709                                 u32 blocksize = leaf->len;
2710
2711                                 root_gen = btrfs_header_generation(
2712                                                            path->nodes[1]);
2713
2714                                 clean_tree_block(trans, root, leaf);
2715
2716                                 wret = del_ptr(trans, root, path, 1, slot);
2717                                 if (wret)
2718                                         ret = wret;
2719
2720                                 free_extent_buffer(leaf);
2721                                 wret = btrfs_free_extent(trans, root, bytenr,
2722                                              blocksize,
2723                                              btrfs_header_owner(path->nodes[1]),
2724                                              root_gen, 0, 0, 1);
2725                                 if (wret)
2726                                         ret = wret;
2727                         } else {
2728                                 btrfs_mark_buffer_dirty(leaf);
2729                                 free_extent_buffer(leaf);
2730                         }
2731                 } else {
2732                         btrfs_mark_buffer_dirty(leaf);
2733                 }
2734         }
2735         return ret;
2736 }
2737
2738 /*
2739  * walk up the tree as far as required to find the previous leaf.
2740  * returns 0 if it found something or 1 if there are no lesser leaves.
2741  * returns < 0 on io errors.
2742  */
2743 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2744 {
2745         u64 bytenr;
2746         int slot;
2747         int level = 1;
2748         struct extent_buffer *c;
2749         struct extent_buffer *next = NULL;
2750
2751         while(level < BTRFS_MAX_LEVEL) {
2752                 if (!path->nodes[level])
2753                         return 1;
2754
2755                 slot = path->slots[level];
2756                 c = path->nodes[level];
2757                 if (slot == 0) {
2758                         level++;
2759                         if (level == BTRFS_MAX_LEVEL)
2760                                 return 1;
2761                         continue;
2762                 }
2763                 slot--;
2764
2765                 bytenr = btrfs_node_blockptr(c, slot);
2766                 if (next)
2767                         free_extent_buffer(next);
2768
2769                 next = read_tree_block(root, bytenr,
2770                                        btrfs_level_size(root, level - 1));
2771                 break;
2772         }
2773         path->slots[level] = slot;
2774         while(1) {
2775                 level--;
2776                 c = path->nodes[level];
2777                 free_extent_buffer(c);
2778                 slot = btrfs_header_nritems(next);
2779                 if (slot != 0)
2780                         slot--;
2781                 path->nodes[level] = next;
2782                 path->slots[level] = slot;
2783                 if (!level)
2784                         break;
2785                 next = read_tree_block(root, btrfs_node_blockptr(next, slot),
2786                                        btrfs_level_size(root, level - 1));
2787         }
2788         return 0;
2789 }
2790
2791 /*
2792  * walk up the tree as far as required to find the next leaf.
2793  * returns 0 if it found something or 1 if there are no greater leaves.
2794  * returns < 0 on io errors.
2795  */
2796 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2797 {
2798         int slot;
2799         int level = 1;
2800         u64 bytenr;
2801         struct extent_buffer *c;
2802         struct extent_buffer *next = NULL;
2803
2804         while(level < BTRFS_MAX_LEVEL) {
2805                 if (!path->nodes[level])
2806                         return 1;
2807
2808                 slot = path->slots[level] + 1;
2809                 c = path->nodes[level];
2810                 if (slot >= btrfs_header_nritems(c)) {
2811                         level++;
2812                         if (level == BTRFS_MAX_LEVEL)
2813                                 return 1;
2814                         continue;
2815                 }
2816
2817                 bytenr = btrfs_node_blockptr(c, slot);
2818                 if (next)
2819                         free_extent_buffer(next);
2820
2821                 if (path->reada)
2822                         reada_for_search(root, path, level, slot, 0);
2823
2824                 next = read_tree_block(root, bytenr,
2825                                        btrfs_level_size(root, level -1));
2826                 break;
2827         }
2828         path->slots[level] = slot;
2829         while(1) {
2830                 level--;
2831                 c = path->nodes[level];
2832                 free_extent_buffer(c);
2833                 path->nodes[level] = next;
2834                 path->slots[level] = 0;
2835                 if (!level)
2836                         break;
2837                 if (path->reada)
2838                         reada_for_search(root, path, level, 0, 0);
2839                 next = read_tree_block(root, btrfs_node_blockptr(next, 0),
2840                                        btrfs_level_size(root, level - 1));
2841         }
2842         return 0;
2843 }
2844
2845 int btrfs_previous_item(struct btrfs_root *root,
2846                         struct btrfs_path *path, u64 min_objectid,
2847                         int type)
2848 {
2849         struct btrfs_key found_key;
2850         struct extent_buffer *leaf;
2851         int ret;
2852
2853         while(1) {
2854                 if (path->slots[0] == 0) {
2855                         ret = btrfs_prev_leaf(root, path);
2856                         if (ret != 0)
2857                                 return ret;
2858                 } else {
2859                         path->slots[0]--;
2860                 }
2861                 leaf = path->nodes[0];
2862                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2863                 if (found_key.type == type)
2864                         return 0;
2865         }
2866         return 1;
2867 }
2868