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 #include "locking.h"
  25
  26 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  27                       *root, struct btrfs_path *path, int level);
  28 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  29                       *root, struct btrfs_key *ins_key,
  30                       struct btrfs_path *path, int data_size, int extend);
  31 static int push_node_left(struct btrfs_trans_handle *trans,
  32                           struct btrfs_root *root, struct extent_buffer *dst,
  33                           struct extent_buffer *src, int empty);
  34 static int balance_node_right(struct btrfs_trans_handle *trans,
  35                               struct btrfs_root *root,
  36                               struct extent_buffer *dst_buf,
  37                               struct extent_buffer *src_buf);
  38 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  39                    struct btrfs_path *path, int level, int slot);
  40
  41 inline void btrfs_init_path(struct btrfs_path *p)
  42 {
  43         memset(p, 0, sizeof(*p));
  44 }
  45
  46 struct btrfs_path *btrfs_alloc_path(void)
  47 {
  48         struct btrfs_path *path;
  49         path = kmem_cache_alloc(btrfs_path_cachep, GFP_NOFS);
  50         if (path) {
  51                 btrfs_init_path(path);
  52                 path->reada = 1;
  53         }
  54         return path;
  55 }
  56
  57 void btrfs_free_path(struct btrfs_path *p)
  58 {
  59         btrfs_release_path(NULL, p);
  60         kmem_cache_free(btrfs_path_cachep, p);
  61 }
  62
  63 void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
  64 {
  65         int i;
  66
  67         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  68                 p->slots[i] = 0;
  69                 if (!p->nodes[i])
  70                         continue;
  71                 if (p->locks[i]) {
  72                         btrfs_tree_unlock(p->nodes[i]);
  73                         p->locks[i] = 0;
  74                 }
  75                 free_extent_buffer(p->nodes[i]);
  76                 p->nodes[i] = NULL;
  77         }
  78 }
  79
  80 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
  81 {
  82         struct extent_buffer *eb;
  83         spin_lock(&root->node_lock);
  84         eb = root->node;
  85         extent_buffer_get(eb);
  86         spin_unlock(&root->node_lock);
  87         return eb;
  88 }
  89
  90 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
  91 {
  92         struct extent_buffer *eb;
  93
  94         while(1) {
  95                 eb = btrfs_root_node(root);
  96                 btrfs_tree_lock(eb);
  97
  98                 spin_lock(&root->node_lock);
  99                 if (eb == root->node) {
 100                         spin_unlock(&root->node_lock);
 101                         break;
 102                 }
 103                 spin_unlock(&root->node_lock);
 104
 105                 btrfs_tree_unlock(eb);
 106                 free_extent_buffer(eb);
 107         }
 108         return eb;
 109 }
 110
 111 static void add_root_to_dirty_list(struct btrfs_root *root)
 112 {
 113         if (root->track_dirty && list_empty(&root->dirty_list)) {
 114                 list_add(&root->dirty_list,
 115                          &root->fs_info->dirty_cowonly_roots);
 116         }
 117 }
 118
 119 int btrfs_copy_root(struct btrfs_trans_handle *trans,
 120                       struct btrfs_root *root,
 121                       struct extent_buffer *buf,
 122                       struct extent_buffer **cow_ret, u64 new_root_objectid)
 123 {
 124         struct extent_buffer *cow;
 125         u32 nritems;
 126         int ret = 0;
 127         int level;
 128         struct btrfs_key first_key;
 129         struct btrfs_root *new_root;
 130
 131         new_root = kmalloc(sizeof(*new_root), GFP_NOFS);
 132         if (!new_root)
 133                 return -ENOMEM;
 134
 135         memcpy(new_root, root, sizeof(*new_root));
 136         new_root->root_key.objectid = new_root_objectid;
 137
 138         WARN_ON(root->ref_cows && trans->transid !=
 139                 root->fs_info->running_transaction->transid);
 140         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 141
 142         level = btrfs_header_level(buf);
 143         nritems = btrfs_header_nritems(buf);
 144         if (nritems) {
 145                 if (level == 0)
 146                         btrfs_item_key_to_cpu(buf, &first_key, 0);
 147                 else
 148                         btrfs_node_key_to_cpu(buf, &first_key, 0);
 149         } else {
 150                 first_key.objectid = 0;
 151         }
 152         cow = btrfs_alloc_free_block(trans, new_root, buf->len,
 153                                        new_root_objectid,
 154                                        trans->transid, first_key.objectid,
 155                                        level, buf->start, 0);
 156         if (IS_ERR(cow)) {
 157                 kfree(new_root);
 158                 return PTR_ERR(cow);
 159         }
 160
 161         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 162         btrfs_set_header_bytenr(cow, cow->start);
 163         btrfs_set_header_generation(cow, trans->transid);
 164         btrfs_set_header_owner(cow, new_root_objectid);
 165         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN);
 166
 167         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 168         ret = btrfs_inc_ref(trans, new_root, buf, 0);
 169         kfree(new_root);
 170
 171         if (ret)
 172                 return ret;
 173
 174         btrfs_mark_buffer_dirty(cow);
 175         *cow_ret = cow;
 176         return 0;
 177 }
 178
 179 int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 180                              struct btrfs_root *root,
 181                              struct extent_buffer *buf,
 182                              struct extent_buffer *parent, int parent_slot,
 183                              struct extent_buffer **cow_ret,
 184                              u64 search_start, u64 empty_size,
 185                              u64 prealloc_dest)
 186 {
 187         u64 root_gen;
 188         struct extent_buffer *cow;
 189         u32 nritems;
 190         int ret = 0;
 191         int different_trans = 0;
 192         int level;
 193         int unlock_orig = 0;
 194         struct btrfs_key first_key;
 195
 196         if (*cow_ret == buf)
 197                 unlock_orig = 1;
 198
 199         WARN_ON(!btrfs_tree_locked(buf));
 200
 201         if (root->ref_cows) {
 202                 root_gen = trans->transid;
 203         } else {
 204                 root_gen = 0;
 205         }
 206         WARN_ON(root->ref_cows && trans->transid !=
 207                 root->fs_info->running_transaction->transid);
 208         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 209
 210         level = btrfs_header_level(buf);
 211         nritems = btrfs_header_nritems(buf);
 212         if (nritems) {
 213                 if (level == 0)
 214                         btrfs_item_key_to_cpu(buf, &first_key, 0);
 215                 else
 216                         btrfs_node_key_to_cpu(buf, &first_key, 0);
 217         } else {
 218                 first_key.objectid = 0;
 219         }
 220         if (prealloc_dest) {
 221                 struct btrfs_key ins;
 222
 223                 ins.objectid = prealloc_dest;
 224                 ins.offset = buf->len;
 225                 ins.type = BTRFS_EXTENT_ITEM_KEY;
 226
 227                 ret = btrfs_alloc_reserved_extent(trans, root,
 228                                                   root->root_key.objectid,
 229                                                   root_gen, level,
 230                                                   first_key.objectid,
 231                                                   &ins);
 232                 BUG_ON(ret);
 233                 cow = btrfs_init_new_buffer(trans, root, prealloc_dest,
 234                                             buf->len);
 235         } else {
 236                 cow = btrfs_alloc_free_block(trans, root, buf->len,
 237                                              root->root_key.objectid,
 238                                              root_gen, first_key.objectid,
 239                                              level, search_start, empty_size);
 240         }
 241         if (IS_ERR(cow))
 242                 return PTR_ERR(cow);
 243
 244         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 245         btrfs_set_header_bytenr(cow, cow->start);
 246         btrfs_set_header_generation(cow, trans->transid);
 247         btrfs_set_header_owner(cow, root->root_key.objectid);
 248         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN);
 249
 250         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 251         if (btrfs_header_generation(buf) != trans->transid) {
 252                 different_trans = 1;
 253                 ret = btrfs_inc_ref(trans, root, buf, 1);
 254                 if (ret)
 255                         return ret;
 256         } else {
 257                 clean_tree_block(trans, root, buf);
 258         }
 259
 260         if (buf == root->node) {
 261                 WARN_ON(parent && parent != buf);
 262                 root_gen = btrfs_header_generation(buf);
 263
 264                 spin_lock(&root->node_lock);
 265                 root->node = cow;
 266                 extent_buffer_get(cow);
 267                 spin_unlock(&root->node_lock);
 268
 269                 if (buf != root->commit_root) {
 270                         btrfs_free_extent(trans, root, buf->start,
 271                                           buf->len, root->root_key.objectid,
 272                                           root_gen, 0, 0, 1);
 273                 }
 274                 free_extent_buffer(buf);
 275                 add_root_to_dirty_list(root);
 276         } else {
 277                 root_gen = btrfs_header_generation(parent);
 278                 btrfs_set_node_blockptr(parent, parent_slot,
 279                                         cow->start);
 280                 WARN_ON(trans->transid == 0);
 281                 btrfs_set_node_ptr_generation(parent, parent_slot,
 282                                               trans->transid);
 283                 btrfs_mark_buffer_dirty(parent);
 284                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 285                 btrfs_free_extent(trans, root, buf->start, buf->len,
 286                                   btrfs_header_owner(parent), root_gen,
 287                                   0, 0, 1);
 288         }
 289         if (unlock_orig)
 290                 btrfs_tree_unlock(buf);
 291         free_extent_buffer(buf);
 292         btrfs_mark_buffer_dirty(cow);
 293         *cow_ret = cow;
 294         return 0;
 295 }
 296
 297 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 298                     struct btrfs_root *root, struct extent_buffer *buf,
 299                     struct extent_buffer *parent, int parent_slot,
 300                     struct extent_buffer **cow_ret, u64 prealloc_dest)
 301 {
 302         u64 search_start;
 303         u64 header_trans;
 304         int ret;
 305
 306         if (trans->transaction != root->fs_info->running_transaction) {
 307                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 308                        root->fs_info->running_transaction->transid);
 309                 WARN_ON(1);
 310         }
 311         if (trans->transid != root->fs_info->generation) {
 312                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 313                        root->fs_info->generation);
 314                 WARN_ON(1);
 315         }
 316
 317         header_trans = btrfs_header_generation(buf);
 318         spin_lock(&root->fs_info->hash_lock);
 319         if (header_trans == trans->transid &&
 320             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
 321                 *cow_ret = buf;
 322                 spin_unlock(&root->fs_info->hash_lock);
 323                 WARN_ON(prealloc_dest);
 324                 return 0;
 325         }
 326         spin_unlock(&root->fs_info->hash_lock);
 327         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
 328         ret = __btrfs_cow_block(trans, root, buf, parent,
 329                                  parent_slot, cow_ret, search_start, 0,
 330                                  prealloc_dest);
 331         return ret;
 332 }
 333
 334 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 335 {
 336         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 337                 return 1;
 338         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 339                 return 1;
 340         return 0;
 341 }
 342
 343 /*
 344  * compare two keys in a memcmp fashion
 345  */
 346 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 347 {
 348         struct btrfs_key k1;
 349
 350         btrfs_disk_key_to_cpu(&k1, disk);
 351
 352         if (k1.objectid > k2->objectid)
 353                 return 1;
 354         if (k1.objectid < k2->objectid)
 355                 return -1;
 356         if (k1.type > k2->type)
 357                 return 1;
 358         if (k1.type < k2->type)
 359                 return -1;
 360         if (k1.offset > k2->offset)
 361                 return 1;
 362         if (k1.offset < k2->offset)
 363                 return -1;
 364         return 0;
 365 }
 366
 367
 368 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 369                        struct btrfs_root *root, struct extent_buffer *parent,
 370                        int start_slot, int cache_only, u64 *last_ret,
 371                        struct btrfs_key *progress)
 372 {
 373         struct extent_buffer *cur;
 374         u64 blocknr;
 375         u64 gen;
 376         u64 search_start = *last_ret;
 377         u64 last_block = 0;
 378         u64 other;
 379         u32 parent_nritems;
 380         int end_slot;
 381         int i;
 382         int err = 0;
 383         int parent_level;
 384         int uptodate;
 385         u32 blocksize;
 386         int progress_passed = 0;
 387         struct btrfs_disk_key disk_key;
 388
 389         parent_level = btrfs_header_level(parent);
 390         if (cache_only && parent_level != 1)
 391                 return 0;
 392
 393         if (trans->transaction != root->fs_info->running_transaction) {
 394                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 395                        root->fs_info->running_transaction->transid);
 396                 WARN_ON(1);
 397         }
 398         if (trans->transid != root->fs_info->generation) {
 399                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 400                        root->fs_info->generation);
 401                 WARN_ON(1);
 402         }
 403
 404         parent_nritems = btrfs_header_nritems(parent);
 405         blocksize = btrfs_level_size(root, parent_level - 1);
 406         end_slot = parent_nritems;
 407
 408         if (parent_nritems == 1)
 409                 return 0;
 410
 411         for (i = start_slot; i < end_slot; i++) {
 412                 int close = 1;
 413
 414                 if (!parent->map_token) {
 415                         map_extent_buffer(parent,
 416                                         btrfs_node_key_ptr_offset(i),
 417                                         sizeof(struct btrfs_key_ptr),
 418                                         &parent->map_token, &parent->kaddr,
 419                                         &parent->map_start, &parent->map_len,
 420                                         KM_USER1);
 421                 }
 422                 btrfs_node_key(parent, &disk_key, i);
 423                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
 424                         continue;
 425
 426                 progress_passed = 1;
 427                 blocknr = btrfs_node_blockptr(parent, i);
 428                 gen = btrfs_node_ptr_generation(parent, i);
 429                 if (last_block == 0)
 430                         last_block = blocknr;
 431
 432                 if (i > 0) {
 433                         other = btrfs_node_blockptr(parent, i - 1);
 434                         close = close_blocks(blocknr, other, blocksize);
 435                 }
 436                 if (!close && i < end_slot - 2) {
 437                         other = btrfs_node_blockptr(parent, i + 1);
 438                         close = close_blocks(blocknr, other, blocksize);
 439                 }
 440                 if (close) {
 441                         last_block = blocknr;
 442                         continue;
 443                 }
 444                 if (parent->map_token) {
 445                         unmap_extent_buffer(parent, parent->map_token,
 446                                             KM_USER1);
 447                         parent->map_token = NULL;
 448                 }
 449
 450                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
 451                 if (cur)
 452                         uptodate = btrfs_buffer_uptodate(cur, gen);
 453                 else
 454                         uptodate = 0;
 455                 if (!cur || !uptodate) {
 456                         if (cache_only) {
 457                                 free_extent_buffer(cur);
 458                                 continue;
 459                         }
 460                         if (!cur) {
 461                                 cur = read_tree_block(root, blocknr,
 462                                                          blocksize, gen);
 463                         } else if (!uptodate) {
 464                                 btrfs_read_buffer(cur, gen);
 465                         }
 466                 }
 467                 if (search_start == 0)
 468                         search_start = last_block;
 469
 470                 btrfs_tree_lock(cur);
 471                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 472                                         &cur, search_start,
 473                                         min(16 * blocksize,
 474                                             (end_slot - i) * blocksize), 0);
 475                 if (err) {
 476                         btrfs_tree_unlock(cur);
 477                         free_extent_buffer(cur);
 478                         break;
 479                 }
 480                 search_start = cur->start;
 481                 last_block = cur->start;
 482                 *last_ret = search_start;
 483                 btrfs_tree_unlock(cur);
 484                 free_extent_buffer(cur);
 485         }
 486         if (parent->map_token) {
 487                 unmap_extent_buffer(parent, parent->map_token,
 488                                     KM_USER1);
 489                 parent->map_token = NULL;
 490         }
 491         return err;
 492 }
 493
 494 /*
 495  * The leaf data grows from end-to-front in the node.
 496  * this returns the address of the start of the last item,
 497  * which is the stop of the leaf data stack
 498  */
 499 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 500                                          struct extent_buffer *leaf)
 501 {
 502         u32 nr = btrfs_header_nritems(leaf);
 503         if (nr == 0)
 504                 return BTRFS_LEAF_DATA_SIZE(root);
 505         return btrfs_item_offset_nr(leaf, nr - 1);
 506 }
 507
 508 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
 509                       int level)
 510 {
 511         struct extent_buffer *parent = NULL;
 512         struct extent_buffer *node = path->nodes[level];
 513         struct btrfs_disk_key parent_key;
 514         struct btrfs_disk_key node_key;
 515         int parent_slot;
 516         int slot;
 517         struct btrfs_key cpukey;
 518         u32 nritems = btrfs_header_nritems(node);
 519
 520         if (path->nodes[level + 1])
 521                 parent = path->nodes[level + 1];
 522
 523         slot = path->slots[level];
 524         BUG_ON(nritems == 0);
 525         if (parent) {
 526                 parent_slot = path->slots[level + 1];
 527                 btrfs_node_key(parent, &parent_key, parent_slot);
 528                 btrfs_node_key(node, &node_key, 0);
 529                 BUG_ON(memcmp(&parent_key, &node_key,
 530                               sizeof(struct btrfs_disk_key)));
 531                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 532                        btrfs_header_bytenr(node));
 533         }
 534         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 535         if (slot != 0) {
 536                 btrfs_node_key_to_cpu(node, &cpukey, slot - 1);
 537                 btrfs_node_key(node, &node_key, slot);
 538                 BUG_ON(comp_keys(&node_key, &cpukey) <= 0);
 539         }
 540         if (slot < nritems - 1) {
 541                 btrfs_node_key_to_cpu(node, &cpukey, slot + 1);
 542                 btrfs_node_key(node, &node_key, slot);
 543                 BUG_ON(comp_keys(&node_key, &cpukey) >= 0);
 544         }
 545         return 0;
 546 }
 547
 548 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 549                       int level)
 550 {
 551         struct extent_buffer *leaf = path->nodes[level];
 552         struct extent_buffer *parent = NULL;
 553         int parent_slot;
 554         struct btrfs_key cpukey;
 555         struct btrfs_disk_key parent_key;
 556         struct btrfs_disk_key leaf_key;
 557         int slot = path->slots[0];
 558
 559         u32 nritems = btrfs_header_nritems(leaf);
 560
 561         if (path->nodes[level + 1])
 562                 parent = path->nodes[level + 1];
 563
 564         if (nritems == 0)
 565                 return 0;
 566
 567         if (parent) {
 568                 parent_slot = path->slots[level + 1];
 569                 btrfs_node_key(parent, &parent_key, parent_slot);
 570                 btrfs_item_key(leaf, &leaf_key, 0);
 571
 572                 BUG_ON(memcmp(&parent_key, &leaf_key,
 573                        sizeof(struct btrfs_disk_key)));
 574                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 575                        btrfs_header_bytenr(leaf));
 576         }
 577 #if 0
 578         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 579                 btrfs_item_key_to_cpu(leaf, &cpukey, i + 1);
 580                 btrfs_item_key(leaf, &leaf_key, i);
 581                 if (comp_keys(&leaf_key, &cpukey) >= 0) {
 582                         btrfs_print_leaf(root, leaf);
 583                         printk("slot %d offset bad key\n", i);
 584                         BUG_ON(1);
 585                 }
 586                 if (btrfs_item_offset_nr(leaf, i) !=
 587                         btrfs_item_end_nr(leaf, i + 1)) {
 588                         btrfs_print_leaf(root, leaf);
 589                         printk("slot %d offset bad\n", i);
 590                         BUG_ON(1);
 591                 }
 592                 if (i == 0) {
 593                         if (btrfs_item_offset_nr(leaf, i) +
 594                                btrfs_item_size_nr(leaf, i) !=
 595                                BTRFS_LEAF_DATA_SIZE(root)) {
 596                                 btrfs_print_leaf(root, leaf);
 597                                 printk("slot %d first offset bad\n", i);
 598                                 BUG_ON(1);
 599                         }
 600                 }
 601         }
 602         if (nritems > 0) {
 603                 if (btrfs_item_size_nr(leaf, nritems - 1) > 4096) {
 604                                 btrfs_print_leaf(root, leaf);
 605                                 printk("slot %d bad size \n", nritems - 1);
 606                                 BUG_ON(1);
 607                 }
 608         }
 609 #endif
 610         if (slot != 0 && slot < nritems - 1) {
 611                 btrfs_item_key(leaf, &leaf_key, slot);
 612                 btrfs_item_key_to_cpu(leaf, &cpukey, slot - 1);
 613                 if (comp_keys(&leaf_key, &cpukey) <= 0) {
 614                         btrfs_print_leaf(root, leaf);
 615                         printk("slot %d offset bad key\n", slot);
 616                         BUG_ON(1);
 617                 }
 618                 if (btrfs_item_offset_nr(leaf, slot - 1) !=
 619                        btrfs_item_end_nr(leaf, slot)) {
 620                         btrfs_print_leaf(root, leaf);
 621                         printk("slot %d offset bad\n", slot);
 622                         BUG_ON(1);
 623                 }
 624         }
 625         if (slot < nritems - 1) {
 626                 btrfs_item_key(leaf, &leaf_key, slot);
 627                 btrfs_item_key_to_cpu(leaf, &cpukey, slot + 1);
 628                 BUG_ON(comp_keys(&leaf_key, &cpukey) >= 0);
 629                 if (btrfs_item_offset_nr(leaf, slot) !=
 630                         btrfs_item_end_nr(leaf, slot + 1)) {
 631                         btrfs_print_leaf(root, leaf);
 632                         printk("slot %d offset bad\n", slot);
 633                         BUG_ON(1);
 634                 }
 635         }
 636         BUG_ON(btrfs_item_offset_nr(leaf, 0) +
 637                btrfs_item_size_nr(leaf, 0) != BTRFS_LEAF_DATA_SIZE(root));
 638         return 0;
 639 }
 640
 641 static int noinline check_block(struct btrfs_root *root,
 642                                 struct btrfs_path *path, int level)
 643 {
 644         u64 found_start;
 645         return 0;
 646         if (btrfs_header_level(path->nodes[level]) != level)
 647             printk("warning: bad level %Lu wanted %d found %d\n",
 648                    path->nodes[level]->start, level,
 649                    btrfs_header_level(path->nodes[level]));
 650         found_start = btrfs_header_bytenr(path->nodes[level]);
 651         if (found_start != path->nodes[level]->start) {
 652             printk("warning: bad bytentr %Lu found %Lu\n",
 653                    path->nodes[level]->start, found_start);
 654         }
 655 #if 0
 656         struct extent_buffer *buf = path->nodes[level];
 657
 658         if (memcmp_extent_buffer(buf, root->fs_info->fsid,
 659                                  (unsigned long)btrfs_header_fsid(buf),
 660                                  BTRFS_FSID_SIZE)) {
 661                 printk("warning bad block %Lu\n", buf->start);
 662                 return 1;
 663         }
 664 #endif
 665         if (level == 0)
 666                 return check_leaf(root, path, level);
 667         return check_node(root, path, level);
 668 }
 669
 670 /*
 671  * search for key in the extent_buffer.  The items start at offset p,
 672  * and they are item_size apart.  There are 'max' items in p.
 673  *
 674  * the slot in the array is returned via slot, and it points to
 675  * the place where you would insert key if it is not found in
 676  * the array.
 677  *
 678  * slot may point to max if the key is bigger than all of the keys
 679  */
 680 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 681                               int item_size, struct btrfs_key *key,
 682                               int max, int *slot)
 683 {
 684         int low = 0;
 685         int high = max;
 686         int mid;
 687         int ret;
 688         struct btrfs_disk_key *tmp = NULL;
 689         struct btrfs_disk_key unaligned;
 690         unsigned long offset;
 691         char *map_token = NULL;
 692         char *kaddr = NULL;
 693         unsigned long map_start = 0;
 694         unsigned long map_len = 0;
 695         int err;
 696
 697         while(low < high) {
 698                 mid = (low + high) / 2;
 699                 offset = p + mid * item_size;
 700
 701                 if (!map_token || offset < map_start ||
 702                     (offset + sizeof(struct btrfs_disk_key)) >
 703                     map_start + map_len) {
 704                         if (map_token) {
 705                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 706                                 map_token = NULL;
 707                         }
 708                         err = map_extent_buffer(eb, offset,
 709                                                 sizeof(struct btrfs_disk_key),
 710                                                 &map_token, &kaddr,
 711                                                 &map_start, &map_len, KM_USER0);
 712
 713                         if (!err) {
 714                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
 715                                                         map_start);
 716                         } else {
 717                                 read_extent_buffer(eb, &unaligned,
 718                                                    offset, sizeof(unaligned));
 719                                 tmp = &unaligned;
 720                         }
 721
 722                 } else {
 723                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
 724                                                         map_start);
 725                 }
 726                 ret = comp_keys(tmp, key);
 727
 728                 if (ret < 0)
 729                         low = mid + 1;
 730                 else if (ret > 0)
 731                         high = mid;
 732                 else {
 733                         *slot = mid;
 734                         if (map_token)
 735                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 736                         return 0;
 737                 }
 738         }
 739         *slot = low;
 740         if (map_token)
 741                 unmap_extent_buffer(eb, map_token, KM_USER0);
 742         return 1;
 743 }
 744
 745 /*
 746  * simple bin_search frontend that does the right thing for
 747  * leaves vs nodes
 748  */
 749 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 750                       int level, int *slot)
 751 {
 752         if (level == 0) {
 753                 return generic_bin_search(eb,
 754                                           offsetof(struct btrfs_leaf, items),
 755                                           sizeof(struct btrfs_item),
 756                                           key, btrfs_header_nritems(eb),
 757                                           slot);
 758         } else {
 759                 return generic_bin_search(eb,
 760                                           offsetof(struct btrfs_node, ptrs),
 761                                           sizeof(struct btrfs_key_ptr),
 762                                           key, btrfs_header_nritems(eb),
 763                                           slot);
 764         }
 765         return -1;
 766 }
 767
 768 static struct extent_buffer *read_node_slot(struct btrfs_root *root,
 769                                    struct extent_buffer *parent, int slot)
 770 {
 771         int level = btrfs_header_level(parent);
 772         if (slot < 0)
 773                 return NULL;
 774         if (slot >= btrfs_header_nritems(parent))
 775                 return NULL;
 776
 777         BUG_ON(level == 0);
 778
 779         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 780                        btrfs_level_size(root, level - 1),
 781                        btrfs_node_ptr_generation(parent, slot));
 782 }
 783
 784 static int balance_level(struct btrfs_trans_handle *trans,
 785                          struct btrfs_root *root,
 786                          struct btrfs_path *path, int level)
 787 {
 788         struct extent_buffer *right = NULL;
 789         struct extent_buffer *mid;
 790         struct extent_buffer *left = NULL;
 791         struct extent_buffer *parent = NULL;
 792         int ret = 0;
 793         int wret;
 794         int pslot;
 795         int orig_slot = path->slots[level];
 796         int err_on_enospc = 0;
 797         u64 orig_ptr;
 798
 799         if (level == 0)
 800                 return 0;
 801
 802         mid = path->nodes[level];
 803         WARN_ON(!path->locks[level]);
 804         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 805
 806         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 807
 808         if (level < BTRFS_MAX_LEVEL - 1)
 809                 parent = path->nodes[level + 1];
 810         pslot = path->slots[level + 1];
 811
 812         /*
 813          * deal with the case where there is only one pointer in the root
 814          * by promoting the node below to a root
 815          */
 816         if (!parent) {
 817                 struct extent_buffer *child;
 818
 819                 if (btrfs_header_nritems(mid) != 1)
 820                         return 0;
 821
 822                 /* promote the child to a root */
 823                 child = read_node_slot(root, mid, 0);
 824                 btrfs_tree_lock(child);
 825                 BUG_ON(!child);
 826                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child, 0);
 827                 BUG_ON(ret);
 828
 829                 spin_lock(&root->node_lock);
 830                 root->node = child;
 831                 spin_unlock(&root->node_lock);
 832
 833                 add_root_to_dirty_list(root);
 834                 btrfs_tree_unlock(child);
 835                 path->locks[level] = 0;
 836                 path->nodes[level] = NULL;
 837                 clean_tree_block(trans, root, mid);
 838                 btrfs_tree_unlock(mid);
 839                 /* once for the path */
 840                 free_extent_buffer(mid);
 841                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 842                                         root->root_key.objectid,
 843                                         btrfs_header_generation(mid), 0, 0, 1);
 844                 /* once for the root ptr */
 845                 free_extent_buffer(mid);
 846                 return ret;
 847         }
 848         if (btrfs_header_nritems(mid) >
 849             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 850                 return 0;
 851
 852         if (btrfs_header_nritems(mid) < 2)
 853                 err_on_enospc = 1;
 854
 855         left = read_node_slot(root, parent, pslot - 1);
 856         if (left) {
 857                 btrfs_tree_lock(left);
 858                 wret = btrfs_cow_block(trans, root, left,
 859                                        parent, pslot - 1, &left, 0);
 860                 if (wret) {
 861                         ret = wret;
 862                         goto enospc;
 863                 }
 864         }
 865         right = read_node_slot(root, parent, pslot + 1);
 866         if (right) {
 867                 btrfs_tree_lock(right);
 868                 wret = btrfs_cow_block(trans, root, right,
 869                                        parent, pslot + 1, &right, 0);
 870                 if (wret) {
 871                         ret = wret;
 872                         goto enospc;
 873                 }
 874         }
 875
 876         /* first, try to make some room in the middle buffer */
 877         if (left) {
 878                 orig_slot += btrfs_header_nritems(left);
 879                 wret = push_node_left(trans, root, left, mid, 1);
 880                 if (wret < 0)
 881                         ret = wret;
 882                 if (btrfs_header_nritems(mid) < 2)
 883                         err_on_enospc = 1;
 884         }
 885
 886         /*
 887          * then try to empty the right most buffer into the middle
 888          */
 889         if (right) {
 890                 wret = push_node_left(trans, root, mid, right, 1);
 891                 if (wret < 0 && wret != -ENOSPC)
 892                         ret = wret;
 893                 if (btrfs_header_nritems(right) == 0) {
 894                         u64 bytenr = right->start;
 895                         u64 generation = btrfs_header_generation(parent);
 896                         u32 blocksize = right->len;
 897
 898                         clean_tree_block(trans, root, right);
 899                         btrfs_tree_unlock(right);
 900                         free_extent_buffer(right);
 901                         right = NULL;
 902                         wret = del_ptr(trans, root, path, level + 1, pslot +
 903                                        1);
 904                         if (wret)
 905                                 ret = wret;
 906                         wret = btrfs_free_extent(trans, root, bytenr,
 907                                                  blocksize,
 908                                                  btrfs_header_owner(parent),
 909                                                  generation, 0, 0, 1);
 910                         if (wret)
 911                                 ret = wret;
 912                 } else {
 913                         struct btrfs_disk_key right_key;
 914                         btrfs_node_key(right, &right_key, 0);
 915                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 916                         btrfs_mark_buffer_dirty(parent);
 917                 }
 918         }
 919         if (btrfs_header_nritems(mid) == 1) {
 920                 /*
 921                  * we're not allowed to leave a node with one item in the
 922                  * tree during a delete.  A deletion from lower in the tree
 923                  * could try to delete the only pointer in this node.
 924                  * So, pull some keys from the left.
 925                  * There has to be a left pointer at this point because
 926                  * otherwise we would have pulled some pointers from the
 927                  * right
 928                  */
 929                 BUG_ON(!left);
 930                 wret = balance_node_right(trans, root, mid, left);
 931                 if (wret < 0) {
 932                         ret = wret;
 933                         goto enospc;
 934                 }
 935                 if (wret == 1) {
 936                         wret = push_node_left(trans, root, left, mid, 1);
 937                         if (wret < 0)
 938                                 ret = wret;
 939                 }
 940                 BUG_ON(wret == 1);
 941         }
 942         if (btrfs_header_nritems(mid) == 0) {
 943                 /* we've managed to empty the middle node, drop it */
 944                 u64 root_gen = btrfs_header_generation(parent);
 945                 u64 bytenr = mid->start;
 946                 u32 blocksize = mid->len;
 947
 948                 clean_tree_block(trans, root, mid);
 949                 btrfs_tree_unlock(mid);
 950                 free_extent_buffer(mid);
 951                 mid = NULL;
 952                 wret = del_ptr(trans, root, path, level + 1, pslot);
 953                 if (wret)
 954                         ret = wret;
 955                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 956                                          btrfs_header_owner(parent),
 957                                          root_gen, 0, 0, 1);
 958                 if (wret)
 959                         ret = wret;
 960         } else {
 961                 /* update the parent key to reflect our changes */
 962                 struct btrfs_disk_key mid_key;
 963                 btrfs_node_key(mid, &mid_key, 0);
 964                 btrfs_set_node_key(parent, &mid_key, pslot);
 965                 btrfs_mark_buffer_dirty(parent);
 966         }
 967
 968         /* update the path */
 969         if (left) {
 970                 if (btrfs_header_nritems(left) > orig_slot) {
 971                         extent_buffer_get(left);
 972                         /* left was locked after cow */
 973                         path->nodes[level] = left;
 974                         path->slots[level + 1] -= 1;
 975                         path->slots[level] = orig_slot;
 976                         if (mid) {
 977                                 btrfs_tree_unlock(mid);
 978                                 free_extent_buffer(mid);
 979                         }
 980                 } else {
 981                         orig_slot -= btrfs_header_nritems(left);
 982                         path->slots[level] = orig_slot;
 983                 }
 984         }
 985         /* double check we haven't messed things up */
 986         check_block(root, path, level);
 987         if (orig_ptr !=
 988             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 989                 BUG();
 990 enospc:
 991         if (right) {
 992                 btrfs_tree_unlock(right);
 993                 free_extent_buffer(right);
 994         }
 995         if (left) {
 996                 if (path->nodes[level] != left)
 997                         btrfs_tree_unlock(left);
 998                 free_extent_buffer(left);
 999         }
1000         return ret;
1001 }
1002
1003 /* returns zero if the push worked, non-zero otherwise */
1004 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
1005                                           struct btrfs_root *root,
1006                                           struct btrfs_path *path, int level)
1007 {
1008         struct extent_buffer *right = NULL;
1009         struct extent_buffer *mid;
1010         struct extent_buffer *left = NULL;
1011         struct extent_buffer *parent = NULL;
1012         int ret = 0;
1013         int wret;
1014         int pslot;
1015         int orig_slot = path->slots[level];
1016         u64 orig_ptr;
1017
1018         if (level == 0)
1019                 return 1;
1020
1021         mid = path->nodes[level];
1022         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1023         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1024
1025         if (level < BTRFS_MAX_LEVEL - 1)
1026                 parent = path->nodes[level + 1];
1027         pslot = path->slots[level + 1];
1028
1029         if (!parent)
1030                 return 1;
1031
1032         left = read_node_slot(root, parent, pslot - 1);
1033
1034         /* first, try to make some room in the middle buffer */
1035         if (left) {
1036                 u32 left_nr;
1037
1038                 btrfs_tree_lock(left);
1039                 left_nr = btrfs_header_nritems(left);
1040                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1041                         wret = 1;
1042                 } else {
1043                         ret = btrfs_cow_block(trans, root, left, parent,
1044                                               pslot - 1, &left, 0);
1045                         if (ret)
1046                                 wret = 1;
1047                         else {
1048                                 wret = push_node_left(trans, root,
1049                                                       left, mid, 0);
1050                         }
1051                 }
1052                 if (wret < 0)
1053                         ret = wret;
1054                 if (wret == 0) {
1055                         struct btrfs_disk_key disk_key;
1056                         orig_slot += left_nr;
1057                         btrfs_node_key(mid, &disk_key, 0);
1058                         btrfs_set_node_key(parent, &disk_key, pslot);
1059                         btrfs_mark_buffer_dirty(parent);
1060                         if (btrfs_header_nritems(left) > orig_slot) {
1061                                 path->nodes[level] = left;
1062                                 path->slots[level + 1] -= 1;
1063                                 path->slots[level] = orig_slot;
1064                                 btrfs_tree_unlock(mid);
1065                                 free_extent_buffer(mid);
1066                         } else {
1067                                 orig_slot -=
1068                                         btrfs_header_nritems(left);
1069                                 path->slots[level] = orig_slot;
1070                                 btrfs_tree_unlock(left);
1071                                 free_extent_buffer(left);
1072                         }
1073                         return 0;
1074                 }
1075                 btrfs_tree_unlock(left);
1076                 free_extent_buffer(left);
1077         }
1078         right = read_node_slot(root, parent, pslot + 1);
1079
1080         /*
1081          * then try to empty the right most buffer into the middle
1082          */
1083         if (right) {
1084                 u32 right_nr;
1085                 btrfs_tree_lock(right);
1086                 right_nr = btrfs_header_nritems(right);
1087                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1088                         wret = 1;
1089                 } else {
1090                         ret = btrfs_cow_block(trans, root, right,
1091                                               parent, pslot + 1,
1092                                               &right, 0);
1093                         if (ret)
1094                                 wret = 1;
1095                         else {
1096                                 wret = balance_node_right(trans, root,
1097                                                           right, mid);
1098                         }
1099                 }
1100                 if (wret < 0)
1101                         ret = wret;
1102                 if (wret == 0) {
1103                         struct btrfs_disk_key disk_key;
1104
1105                         btrfs_node_key(right, &disk_key, 0);
1106                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
1107                         btrfs_mark_buffer_dirty(parent);
1108
1109                         if (btrfs_header_nritems(mid) <= orig_slot) {
1110                                 path->nodes[level] = right;
1111                                 path->slots[level + 1] += 1;
1112                                 path->slots[level] = orig_slot -
1113                                         btrfs_header_nritems(mid);
1114                                 btrfs_tree_unlock(mid);
1115                                 free_extent_buffer(mid);
1116                         } else {
1117                                 btrfs_tree_unlock(right);
1118                                 free_extent_buffer(right);
1119                         }
1120                         return 0;
1121                 }
1122                 btrfs_tree_unlock(right);
1123                 free_extent_buffer(right);
1124         }
1125         return 1;
1126 }
1127
1128 /*
1129  * readahead one full node of leaves
1130  */
1131 static void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
1132                              int level, int slot, u64 objectid)
1133 {
1134         struct extent_buffer *node;
1135         struct btrfs_disk_key disk_key;
1136         u32 nritems;
1137         u64 search;
1138         u64 lowest_read;
1139         u64 highest_read;
1140         u64 nread = 0;
1141         int direction = path->reada;
1142         struct extent_buffer *eb;
1143         u32 nr;
1144         u32 blocksize;
1145         u32 nscan = 0;
1146
1147         if (level != 1)
1148                 return;
1149
1150         if (!path->nodes[level])
1151                 return;
1152
1153         node = path->nodes[level];
1154
1155         search = btrfs_node_blockptr(node, slot);
1156         blocksize = btrfs_level_size(root, level - 1);
1157         eb = btrfs_find_tree_block(root, search, blocksize);
1158         if (eb) {
1159                 free_extent_buffer(eb);
1160                 return;
1161         }
1162
1163         highest_read = search;
1164         lowest_read = search;
1165
1166         nritems = btrfs_header_nritems(node);
1167         nr = slot;
1168         while(1) {
1169                 if (direction < 0) {
1170                         if (nr == 0)
1171                                 break;
1172                         nr--;
1173                 } else if (direction > 0) {
1174                         nr++;
1175                         if (nr >= nritems)
1176                                 break;
1177                 }
1178                 if (path->reada < 0 && objectid) {
1179                         btrfs_node_key(node, &disk_key, nr);
1180                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1181                                 break;
1182                 }
1183                 search = btrfs_node_blockptr(node, nr);
1184                 if ((search >= lowest_read && search <= highest_read) ||
1185                     (search < lowest_read && lowest_read - search <= 32768) ||
1186                     (search > highest_read && search - highest_read <= 32768)) {
1187                         readahead_tree_block(root, search, blocksize,
1188                                      btrfs_node_ptr_generation(node, nr));
1189                         nread += blocksize;
1190                 }
1191                 nscan++;
1192                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
1193                         break;
1194                 if(nread > (1024 * 1024) || nscan > 128)
1195                         break;
1196
1197                 if (search < lowest_read)
1198                         lowest_read = search;
1199                 if (search > highest_read)
1200                         highest_read = search;
1201         }
1202 }
1203
1204 static void unlock_up(struct btrfs_path *path, int level, int lowest_unlock)
1205 {
1206         int i;
1207         int skip_level = level;
1208         int no_skips = 0;
1209         struct extent_buffer *t;
1210
1211         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1212                 if (!path->nodes[i])
1213                         break;
1214                 if (!path->locks[i])
1215                         break;
1216                 if (!no_skips && path->slots[i] == 0) {
1217                         skip_level = i + 1;
1218                         continue;
1219                 }
1220                 if (!no_skips && path->keep_locks) {
1221                         u32 nritems;
1222                         t = path->nodes[i];
1223                         nritems = btrfs_header_nritems(t);
1224                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
1225                                 skip_level = i + 1;
1226                                 continue;
1227                         }
1228                 }
1229                 if (skip_level < i && i >= lowest_unlock)
1230                         no_skips = 1;
1231
1232                 t = path->nodes[i];
1233                 if (i >= lowest_unlock && i > skip_level && path->locks[i]) {
1234                         btrfs_tree_unlock(t);
1235                         path->locks[i] = 0;
1236                 }
1237         }
1238 }
1239
1240 /*
1241  * look for key in the tree.  path is filled in with nodes along the way
1242  * if key is found, we return zero and you can find the item in the leaf
1243  * level of the path (level 0)
1244  *
1245  * If the key isn't found, the path points to the slot where it should
1246  * be inserted, and 1 is returned.  If there are other errors during the
1247  * search a negative error number is returned.
1248  *
1249  * if ins_len > 0, nodes and leaves will be split as we walk down the
1250  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1251  * possible)
1252  */
1253 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1254                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1255                       ins_len, int cow)
1256 {
1257         struct extent_buffer *b;
1258         struct extent_buffer *tmp;
1259         int slot;
1260         int ret;
1261         int level;
1262         int should_reada = p->reada;
1263         int lowest_unlock = 1;
1264         int blocksize;
1265         u8 lowest_level = 0;
1266         u64 blocknr;
1267         u64 gen;
1268         struct btrfs_key prealloc_block;
1269
1270         lowest_level = p->lowest_level;
1271         WARN_ON(lowest_level && ins_len);
1272         WARN_ON(p->nodes[0] != NULL);
1273         WARN_ON(cow && root == root->fs_info->extent_root &&
1274                 !mutex_is_locked(&root->fs_info->alloc_mutex));
1275         if (ins_len < 0)
1276                 lowest_unlock = 2;
1277
1278         prealloc_block.objectid = 0;
1279
1280 again:
1281         if (p->skip_locking)
1282                 b = btrfs_root_node(root);
1283         else
1284                 b = btrfs_lock_root_node(root);
1285
1286         while (b) {
1287                 level = btrfs_header_level(b);
1288
1289                 /*
1290                  * setup the path here so we can release it under lock
1291                  * contention with the cow code
1292                  */
1293                 p->nodes[level] = b;
1294                 if (!p->skip_locking)
1295                         p->locks[level] = 1;
1296
1297                 if (cow) {
1298                         int wret;
1299
1300                         /* is a cow on this block not required */
1301                         spin_lock(&root->fs_info->hash_lock);
1302                         if (btrfs_header_generation(b) == trans->transid &&
1303                             !btrfs_header_flag(b, BTRFS_HEADER_FLAG_WRITTEN)) {
1304                                 spin_unlock(&root->fs_info->hash_lock);
1305                                 goto cow_done;
1306                         }
1307                         spin_unlock(&root->fs_info->hash_lock);
1308
1309                         /* ok, we have to cow, is our old prealloc the right
1310                          * size?
1311                          */
1312                         if (prealloc_block.objectid &&
1313                             prealloc_block.offset != b->len) {
1314                                 btrfs_free_reserved_extent(root,
1315                                            prealloc_block.objectid,
1316                                            prealloc_block.offset);
1317                                 prealloc_block.objectid = 0;
1318                         }
1319
1320                         /*
1321                          * for higher level blocks, try not to allocate blocks
1322                          * with the block and the parent locks held.
1323                          */
1324                         if (level > 1 && !prealloc_block.objectid &&
1325                             btrfs_path_lock_waiting(p, level)) {
1326                                 u32 size = b->len;
1327                                 u64 hint = b->start;
1328
1329                                 btrfs_release_path(root, p);
1330                                 ret = btrfs_reserve_extent(trans, root,
1331                                                            size, size, 0,
1332                                                            hint, (u64)-1,
1333                                                            &prealloc_block, 0);
1334                                 BUG_ON(ret);
1335                                 goto again;
1336                         }
1337
1338                         wret = btrfs_cow_block(trans, root, b,
1339                                                p->nodes[level + 1],
1340                                                p->slots[level + 1],
1341                                                &b, prealloc_block.objectid);
1342                         prealloc_block.objectid = 0;
1343                         if (wret) {
1344                                 free_extent_buffer(b);
1345                                 ret = wret;
1346                                 goto done;
1347                         }
1348                 }
1349 cow_done:
1350                 BUG_ON(!cow && ins_len);
1351                 if (level != btrfs_header_level(b))
1352                         WARN_ON(1);
1353                 level = btrfs_header_level(b);
1354
1355                 p->nodes[level] = b;
1356                 if (!p->skip_locking)
1357                         p->locks[level] = 1;
1358
1359                 ret = check_block(root, p, level);
1360                 if (ret) {
1361                         ret = -1;
1362                         goto done;
1363                 }
1364
1365                 ret = bin_search(b, key, level, &slot);
1366                 if (level != 0) {
1367                         if (ret && slot > 0)
1368                                 slot -= 1;
1369                         p->slots[level] = slot;
1370                         if (ins_len > 0 && btrfs_header_nritems(b) >=
1371                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1372                                 int sret = split_node(trans, root, p, level);
1373                                 BUG_ON(sret > 0);
1374                                 if (sret) {
1375                                         ret = sret;
1376                                         goto done;
1377                                 }
1378                                 b = p->nodes[level];
1379                                 slot = p->slots[level];
1380                         } else if (ins_len < 0) {
1381                                 int sret = balance_level(trans, root, p,
1382                                                          level);
1383                                 if (sret) {
1384                                         ret = sret;
1385                                         goto done;
1386                                 }
1387                                 b = p->nodes[level];
1388                                 if (!b) {
1389                                         btrfs_release_path(NULL, p);
1390                                         goto again;
1391                                 }
1392                                 slot = p->slots[level];
1393                                 BUG_ON(btrfs_header_nritems(b) == 1);
1394                         }
1395                         unlock_up(p, level, lowest_unlock);
1396
1397                         /* this is only true while dropping a snapshot */
1398                         if (level == lowest_level) {
1399                                 break;
1400                         }
1401
1402                         blocknr = btrfs_node_blockptr(b, slot);
1403                         gen = btrfs_node_ptr_generation(b, slot);
1404                         blocksize = btrfs_level_size(root, level - 1);
1405
1406                         tmp = btrfs_find_tree_block(root, blocknr, blocksize);
1407                         if (tmp && btrfs_buffer_uptodate(tmp, gen)) {
1408                                 b = tmp;
1409                         } else {
1410                                 /*
1411                                  * reduce lock contention at high levels
1412                                  * of the btree by dropping locks before
1413                                  * we read.
1414                                  */
1415                                 if (level > 1) {
1416                                         btrfs_release_path(NULL, p);
1417                                         if (tmp)
1418                                                 free_extent_buffer(tmp);
1419                                         if (should_reada)
1420                                                 reada_for_search(root, p,
1421                                                                  level, slot,
1422                                                                  key->objectid);
1423
1424                                         tmp = read_tree_block(root, blocknr,
1425                                                          blocksize, gen);
1426                                         if (tmp)
1427                                                 free_extent_buffer(tmp);
1428                                         goto again;
1429                                 } else {
1430                                         if (tmp)
1431                                                 free_extent_buffer(tmp);
1432                                         if (should_reada)
1433                                                 reada_for_search(root, p,
1434                                                                  level, slot,
1435                                                                  key->objectid);
1436                                         b = read_node_slot(root, b, slot);
1437                                 }
1438                         }
1439                         if (!p->skip_locking)
1440                                 btrfs_tree_lock(b);
1441                 } else {
1442                         p->slots[level] = slot;
1443                         if (ins_len > 0 && btrfs_leaf_free_space(root, b) <
1444                             sizeof(struct btrfs_item) + ins_len) {
1445                                 int sret = split_leaf(trans, root, key,
1446                                                       p, ins_len, ret == 0);
1447                                 BUG_ON(sret > 0);
1448                                 if (sret) {
1449                                         ret = sret;
1450                                         goto done;
1451                                 }
1452                         }
1453                         unlock_up(p, level, lowest_unlock);
1454                         goto done;
1455                 }
1456         }
1457         ret = 1;
1458 done:
1459         if (prealloc_block.objectid) {
1460                 btrfs_free_reserved_extent(root,
1461                            prealloc_block.objectid,
1462                            prealloc_block.offset);
1463         }
1464
1465         return ret;
1466 }
1467
1468 /*
1469  * adjust the pointers going up the tree, starting at level
1470  * making sure the right key of each node is points to 'key'.
1471  * This is used after shifting pointers to the left, so it stops
1472  * fixing up pointers when a given leaf/node is not in slot 0 of the
1473  * higher levels
1474  *
1475  * If this fails to write a tree block, it returns -1, but continues
1476  * fixing up the blocks in ram so the tree is consistent.
1477  */
1478 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1479                           struct btrfs_root *root, struct btrfs_path *path,
1480                           struct btrfs_disk_key *key, int level)
1481 {
1482         int i;
1483         int ret = 0;
1484         struct extent_buffer *t;
1485
1486         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1487                 int tslot = path->slots[i];
1488                 if (!path->nodes[i])
1489                         break;
1490                 t = path->nodes[i];
1491                 btrfs_set_node_key(t, key, tslot);
1492                 btrfs_mark_buffer_dirty(path->nodes[i]);
1493                 if (tslot != 0)
1494                         break;
1495         }
1496         return ret;
1497 }
1498
1499 /*
1500  * try to push data from one node into the next node left in the
1501  * tree.
1502  *
1503  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1504  * error, and > 0 if there was no room in the left hand block.
1505  */
1506 static int push_node_left(struct btrfs_trans_handle *trans,
1507                           struct btrfs_root *root, struct extent_buffer *dst,
1508                           struct extent_buffer *src, int empty)
1509 {
1510         int push_items = 0;
1511         int src_nritems;
1512         int dst_nritems;
1513         int ret = 0;
1514
1515         src_nritems = btrfs_header_nritems(src);
1516         dst_nritems = btrfs_header_nritems(dst);
1517         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1518         WARN_ON(btrfs_header_generation(src) != trans->transid);
1519         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1520
1521         if (!empty && src_nritems <= 8)
1522                 return 1;
1523
1524         if (push_items <= 0) {
1525                 return 1;
1526         }
1527
1528         if (empty) {
1529                 push_items = min(src_nritems, push_items);
1530                 if (push_items < src_nritems) {
1531                         /* leave at least 8 pointers in the node if
1532                          * we aren't going to empty it
1533                          */
1534                         if (src_nritems - push_items < 8) {
1535                                 if (push_items <= 8)
1536                                         return 1;
1537                                 push_items -= 8;
1538                         }
1539                 }
1540         } else
1541                 push_items = min(src_nritems - 8, push_items);
1542
1543         copy_extent_buffer(dst, src,
1544                            btrfs_node_key_ptr_offset(dst_nritems),
1545                            btrfs_node_key_ptr_offset(0),
1546                            push_items * sizeof(struct btrfs_key_ptr));
1547
1548         if (push_items < src_nritems) {
1549                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1550                                       btrfs_node_key_ptr_offset(push_items),
1551                                       (src_nritems - push_items) *
1552                                       sizeof(struct btrfs_key_ptr));
1553         }
1554         btrfs_set_header_nritems(src, src_nritems - push_items);
1555         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1556         btrfs_mark_buffer_dirty(src);
1557         btrfs_mark_buffer_dirty(dst);
1558         return ret;
1559 }
1560
1561 /*
1562  * try to push data from one node into the next node right in the
1563  * tree.
1564  *
1565  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1566  * error, and > 0 if there was no room in the right hand block.
1567  *
1568  * this will  only push up to 1/2 the contents of the left node over
1569  */
1570 static int balance_node_right(struct btrfs_trans_handle *trans,
1571                               struct btrfs_root *root,
1572                               struct extent_buffer *dst,
1573                               struct extent_buffer *src)
1574 {
1575         int push_items = 0;
1576         int max_push;
1577         int src_nritems;
1578         int dst_nritems;
1579         int ret = 0;
1580
1581         WARN_ON(btrfs_header_generation(src) != trans->transid);
1582         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1583
1584         src_nritems = btrfs_header_nritems(src);
1585         dst_nritems = btrfs_header_nritems(dst);
1586         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1587         if (push_items <= 0) {
1588                 return 1;
1589         }
1590
1591         if (src_nritems < 4) {
1592                 return 1;
1593         }
1594
1595         max_push = src_nritems / 2 + 1;
1596         /* don't try to empty the node */
1597         if (max_push >= src_nritems) {
1598                 return 1;
1599         }
1600
1601         if (max_push < push_items)
1602                 push_items = max_push;
1603
1604         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1605                                       btrfs_node_key_ptr_offset(0),
1606                                       (dst_nritems) *
1607                                       sizeof(struct btrfs_key_ptr));
1608
1609         copy_extent_buffer(dst, src,
1610                            btrfs_node_key_ptr_offset(0),
1611                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1612                            push_items * sizeof(struct btrfs_key_ptr));
1613
1614         btrfs_set_header_nritems(src, src_nritems - push_items);
1615         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1616
1617         btrfs_mark_buffer_dirty(src);
1618         btrfs_mark_buffer_dirty(dst);
1619         return ret;
1620 }
1621
1622 /*
1623  * helper function to insert a new root level in the tree.
1624  * A new node is allocated, and a single item is inserted to
1625  * point to the existing root
1626  *
1627  * returns zero on success or < 0 on failure.
1628  */
1629 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1630                            struct btrfs_root *root,
1631                            struct btrfs_path *path, int level)
1632 {
1633         u64 root_gen;
1634         u64 lower_gen;
1635         struct extent_buffer *lower;
1636         struct extent_buffer *c;
1637         struct extent_buffer *old;
1638         struct btrfs_disk_key lower_key;
1639
1640         BUG_ON(path->nodes[level]);
1641         BUG_ON(path->nodes[level-1] != root->node);
1642
1643         if (root->ref_cows)
1644                 root_gen = trans->transid;
1645         else
1646                 root_gen = 0;
1647
1648         lower = path->nodes[level-1];
1649         if (level == 1)
1650                 btrfs_item_key(lower, &lower_key, 0);
1651         else
1652                 btrfs_node_key(lower, &lower_key, 0);
1653
1654         c = btrfs_alloc_free_block(trans, root, root->nodesize,
1655                                    root->root_key.objectid,
1656                                    root_gen, lower_key.objectid, level,
1657                                    root->node->start, 0);
1658         if (IS_ERR(c))
1659                 return PTR_ERR(c);
1660
1661         memset_extent_buffer(c, 0, 0, root->nodesize);
1662         btrfs_set_header_nritems(c, 1);
1663         btrfs_set_header_level(c, level);
1664         btrfs_set_header_bytenr(c, c->start);
1665         btrfs_set_header_generation(c, trans->transid);
1666         btrfs_set_header_owner(c, root->root_key.objectid);
1667
1668         write_extent_buffer(c, root->fs_info->fsid,
1669                             (unsigned long)btrfs_header_fsid(c),
1670                             BTRFS_FSID_SIZE);
1671
1672         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1673                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
1674                             BTRFS_UUID_SIZE);
1675
1676         btrfs_set_node_key(c, &lower_key, 0);
1677         btrfs_set_node_blockptr(c, 0, lower->start);
1678         lower_gen = btrfs_header_generation(lower);
1679         WARN_ON(lower_gen == 0);
1680
1681         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1682
1683         btrfs_mark_buffer_dirty(c);
1684
1685         spin_lock(&root->node_lock);
1686         old = root->node;
1687         root->node = c;
1688         spin_unlock(&root->node_lock);
1689
1690         /* the super has an extra ref to root->node */
1691         free_extent_buffer(old);
1692
1693         add_root_to_dirty_list(root);
1694         extent_buffer_get(c);
1695         path->nodes[level] = c;
1696         path->locks[level] = 1;
1697         path->slots[level] = 0;
1698
1699         if (root->ref_cows && lower_gen != trans->transid) {
1700                 struct btrfs_path *back_path = btrfs_alloc_path();
1701                 int ret;
1702                 mutex_lock(&root->fs_info->alloc_mutex);
1703                 ret = btrfs_insert_extent_backref(trans,
1704                                                   root->fs_info->extent_root,
1705                                                   path, lower->start,
1706                                                   root->root_key.objectid,
1707                                                   trans->transid, 0, 0);
1708                 BUG_ON(ret);
1709                 mutex_unlock(&root->fs_info->alloc_mutex);
1710                 btrfs_free_path(back_path);
1711         }
1712         return 0;
1713 }
1714
1715 /*
1716  * worker function to insert a single pointer in a node.
1717  * the node should have enough room for the pointer already
1718  *
1719  * slot and level indicate where you want the key to go, and
1720  * blocknr is the block the key points to.
1721  *
1722  * returns zero on success and < 0 on any error
1723  */
1724 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1725                       *root, struct btrfs_path *path, struct btrfs_disk_key
1726                       *key, u64 bytenr, int slot, int level)
1727 {
1728         struct extent_buffer *lower;
1729         int nritems;
1730
1731         BUG_ON(!path->nodes[level]);
1732         lower = path->nodes[level];
1733         nritems = btrfs_header_nritems(lower);
1734         if (slot > nritems)
1735                 BUG();
1736         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1737                 BUG();
1738         if (slot != nritems) {
1739                 memmove_extent_buffer(lower,
1740                               btrfs_node_key_ptr_offset(slot + 1),
1741                               btrfs_node_key_ptr_offset(slot),
1742                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1743         }
1744         btrfs_set_node_key(lower, key, slot);
1745         btrfs_set_node_blockptr(lower, slot, bytenr);
1746         WARN_ON(trans->transid == 0);
1747         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1748         btrfs_set_header_nritems(lower, nritems + 1);
1749         btrfs_mark_buffer_dirty(lower);
1750         return 0;
1751 }
1752
1753 /*
1754  * split the node at the specified level in path in two.
1755  * The path is corrected to point to the appropriate node after the split
1756  *
1757  * Before splitting this tries to make some room in the node by pushing
1758  * left and right, if either one works, it returns right away.
1759  *
1760  * returns 0 on success and < 0 on failure
1761  */
1762 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1763                       *root, struct btrfs_path *path, int level)
1764 {
1765         u64 root_gen;
1766         struct extent_buffer *c;
1767         struct extent_buffer *split;
1768         struct btrfs_disk_key disk_key;
1769         int mid;
1770         int ret;
1771         int wret;
1772         u32 c_nritems;
1773
1774         c = path->nodes[level];
1775         WARN_ON(btrfs_header_generation(c) != trans->transid);
1776         if (c == root->node) {
1777                 /* trying to split the root, lets make a new one */
1778                 ret = insert_new_root(trans, root, path, level + 1);
1779                 if (ret)
1780                         return ret;
1781         } else {
1782                 ret = push_nodes_for_insert(trans, root, path, level);
1783                 c = path->nodes[level];
1784                 if (!ret && btrfs_header_nritems(c) <
1785                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
1786                         return 0;
1787                 if (ret < 0)
1788                         return ret;
1789         }
1790
1791         c_nritems = btrfs_header_nritems(c);
1792         if (root->ref_cows)
1793                 root_gen = trans->transid;
1794         else
1795                 root_gen = 0;
1796
1797         btrfs_node_key(c, &disk_key, 0);
1798         split = btrfs_alloc_free_block(trans, root, root->nodesize,
1799                                          root->root_key.objectid,
1800                                          root_gen,
1801                                          btrfs_disk_key_objectid(&disk_key),
1802                                          level, c->start, 0);
1803         if (IS_ERR(split))
1804                 return PTR_ERR(split);
1805
1806         btrfs_set_header_flags(split, btrfs_header_flags(c));
1807         btrfs_set_header_level(split, btrfs_header_level(c));
1808         btrfs_set_header_bytenr(split, split->start);
1809         btrfs_set_header_generation(split, trans->transid);
1810         btrfs_set_header_owner(split, root->root_key.objectid);
1811         btrfs_set_header_flags(split, 0);
1812         write_extent_buffer(split, root->fs_info->fsid,
1813                             (unsigned long)btrfs_header_fsid(split),
1814                             BTRFS_FSID_SIZE);
1815         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1816                             (unsigned long)btrfs_header_chunk_tree_uuid(split),
1817                             BTRFS_UUID_SIZE);
1818
1819         mid = (c_nritems + 1) / 2;
1820
1821         copy_extent_buffer(split, c,
1822                            btrfs_node_key_ptr_offset(0),
1823                            btrfs_node_key_ptr_offset(mid),
1824                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1825         btrfs_set_header_nritems(split, c_nritems - mid);
1826         btrfs_set_header_nritems(c, mid);
1827         ret = 0;
1828
1829         btrfs_mark_buffer_dirty(c);
1830         btrfs_mark_buffer_dirty(split);
1831
1832         btrfs_node_key(split, &disk_key, 0);
1833         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1834                           path->slots[level + 1] + 1,
1835                           level + 1);
1836         if (wret)
1837                 ret = wret;
1838
1839         if (path->slots[level] >= mid) {
1840                 path->slots[level] -= mid;
1841                 btrfs_tree_unlock(c);
1842                 free_extent_buffer(c);
1843                 path->nodes[level] = split;
1844                 path->slots[level + 1] += 1;
1845         } else {
1846                 btrfs_tree_unlock(split);
1847                 free_extent_buffer(split);
1848         }
1849         return ret;
1850 }
1851
1852 /*
1853  * how many bytes are required to store the items in a leaf.  start
1854  * and nr indicate which items in the leaf to check.  This totals up the
1855  * space used both by the item structs and the item data
1856  */
1857 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1858 {
1859         int data_len;
1860         int nritems = btrfs_header_nritems(l);
1861         int end = min(nritems, start + nr) - 1;
1862
1863         if (!nr)
1864                 return 0;
1865         data_len = btrfs_item_end_nr(l, start);
1866         data_len = data_len - btrfs_item_offset_nr(l, end);
1867         data_len += sizeof(struct btrfs_item) * nr;
1868         WARN_ON(data_len < 0);
1869         return data_len;
1870 }
1871
1872 /*
1873  * The space between the end of the leaf items and
1874  * the start of the leaf data.  IOW, how much room
1875  * the leaf has left for both items and data
1876  */
1877 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1878 {
1879         int nritems = btrfs_header_nritems(leaf);
1880         int ret;
1881         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1882         if (ret < 0) {
1883                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1884                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1885                        leaf_space_used(leaf, 0, nritems), nritems);
1886         }
1887         return ret;
1888 }
1889
1890 /*
1891  * push some data in the path leaf to the right, trying to free up at
1892  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1893  *
1894  * returns 1 if the push failed because the other node didn't have enough
1895  * room, 0 if everything worked out and < 0 if there were major errors.
1896  */
1897 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1898                            *root, struct btrfs_path *path, int data_size,
1899                            int empty)
1900 {
1901         struct extent_buffer *left = path->nodes[0];
1902         struct extent_buffer *right;
1903         struct extent_buffer *upper;
1904         struct btrfs_disk_key disk_key;
1905         int slot;
1906         u32 i;
1907         int free_space;
1908         int push_space = 0;
1909         int push_items = 0;
1910         struct btrfs_item *item;
1911         u32 left_nritems;
1912         u32 nr;
1913         u32 right_nritems;
1914         u32 data_end;
1915         u32 this_item_size;
1916         int ret;
1917
1918         slot = path->slots[1];
1919         if (!path->nodes[1]) {
1920                 return 1;
1921         }
1922         upper = path->nodes[1];
1923         if (slot >= btrfs_header_nritems(upper) - 1)
1924                 return 1;
1925
1926         WARN_ON(!btrfs_tree_locked(path->nodes[1]));
1927
1928         right = read_node_slot(root, upper, slot + 1);
1929         btrfs_tree_lock(right);
1930         free_space = btrfs_leaf_free_space(root, right);
1931         if (free_space < data_size + sizeof(struct btrfs_item))
1932                 goto out_unlock;
1933
1934         /* cow and double check */
1935         ret = btrfs_cow_block(trans, root, right, upper,
1936                               slot + 1, &right, 0);
1937         if (ret)
1938                 goto out_unlock;
1939
1940         free_space = btrfs_leaf_free_space(root, right);
1941         if (free_space < data_size + sizeof(struct btrfs_item))
1942                 goto out_unlock;
1943
1944         left_nritems = btrfs_header_nritems(left);
1945         if (left_nritems == 0)
1946                 goto out_unlock;
1947
1948         if (empty)
1949                 nr = 0;
1950         else
1951                 nr = 1;
1952
1953         i = left_nritems - 1;
1954         while (i >= nr) {
1955                 item = btrfs_item_nr(left, i);
1956
1957                 if (path->slots[0] == i)
1958                         push_space += data_size + sizeof(*item);
1959
1960                 if (!left->map_token) {
1961                         map_extent_buffer(left, (unsigned long)item,
1962                                         sizeof(struct btrfs_item),
1963                                         &left->map_token, &left->kaddr,
1964                                         &left->map_start, &left->map_len,
1965                                         KM_USER1);
1966                 }
1967
1968                 this_item_size = btrfs_item_size(left, item);
1969                 if (this_item_size + sizeof(*item) + push_space > free_space)
1970                         break;
1971                 push_items++;
1972                 push_space += this_item_size + sizeof(*item);
1973                 if (i == 0)
1974                         break;
1975                 i--;
1976         }
1977         if (left->map_token) {
1978                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1979                 left->map_token = NULL;
1980         }
1981
1982         if (push_items == 0)
1983                 goto out_unlock;
1984
1985         if (!empty && push_items == left_nritems)
1986                 WARN_ON(1);
1987
1988         /* push left to right */
1989         right_nritems = btrfs_header_nritems(right);
1990
1991         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1992         push_space -= leaf_data_end(root, left);
1993
1994         /* make room in the right data area */
1995         data_end = leaf_data_end(root, right);
1996         memmove_extent_buffer(right,
1997                               btrfs_leaf_data(right) + data_end - push_space,
1998                               btrfs_leaf_data(right) + data_end,
1999                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
2000
2001         /* copy from the left data area */
2002         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
2003                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
2004                      btrfs_leaf_data(left) + leaf_data_end(root, left),
2005                      push_space);
2006
2007         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
2008                               btrfs_item_nr_offset(0),
2009                               right_nritems * sizeof(struct btrfs_item));
2010
2011         /* copy the items from left to right */
2012         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
2013                    btrfs_item_nr_offset(left_nritems - push_items),
2014                    push_items * sizeof(struct btrfs_item));
2015
2016         /* update the item pointers */
2017         right_nritems += push_items;
2018         btrfs_set_header_nritems(right, right_nritems);
2019         push_space = BTRFS_LEAF_DATA_SIZE(root);
2020         for (i = 0; i < right_nritems; i++) {
2021                 item = btrfs_item_nr(right, i);
2022                 if (!right->map_token) {
2023                         map_extent_buffer(right, (unsigned long)item,
2024                                         sizeof(struct btrfs_item),
2025                                         &right->map_token, &right->kaddr,
2026                                         &right->map_start, &right->map_len,
2027                                         KM_USER1);
2028                 }
2029                 push_space -= btrfs_item_size(right, item);
2030                 btrfs_set_item_offset(right, item, push_space);
2031         }
2032
2033         if (right->map_token) {
2034                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2035                 right->map_token = NULL;
2036         }
2037         left_nritems -= push_items;
2038         btrfs_set_header_nritems(left, left_nritems);
2039
2040         if (left_nritems)
2041                 btrfs_mark_buffer_dirty(left);
2042         btrfs_mark_buffer_dirty(right);
2043
2044         btrfs_item_key(right, &disk_key, 0);
2045         btrfs_set_node_key(upper, &disk_key, slot + 1);
2046         btrfs_mark_buffer_dirty(upper);
2047
2048         /* then fixup the leaf pointer in the path */
2049         if (path->slots[0] >= left_nritems) {
2050                 path->slots[0] -= left_nritems;
2051                 if (btrfs_header_nritems(path->nodes[0]) == 0)
2052                         clean_tree_block(trans, root, path->nodes[0]);
2053                 btrfs_tree_unlock(path->nodes[0]);
2054                 free_extent_buffer(path->nodes[0]);
2055                 path->nodes[0] = right;
2056                 path->slots[1] += 1;
2057         } else {
2058                 btrfs_tree_unlock(right);
2059                 free_extent_buffer(right);
2060         }
2061         return 0;
2062
2063 out_unlock:
2064         btrfs_tree_unlock(right);
2065         free_extent_buffer(right);
2066         return 1;
2067 }
2068
2069 /*
2070  * push some data in the path leaf to the left, trying to free up at
2071  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
2072  */
2073 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
2074                           *root, struct btrfs_path *path, int data_size,
2075                           int empty)
2076 {
2077         struct btrfs_disk_key disk_key;
2078         struct extent_buffer *right = path->nodes[0];
2079         struct extent_buffer *left;
2080         int slot;
2081         int i;
2082         int free_space;
2083         int push_space = 0;
2084         int push_items = 0;
2085         struct btrfs_item *item;
2086         u32 old_left_nritems;
2087         u32 right_nritems;
2088         u32 nr;
2089         int ret = 0;
2090         int wret;
2091         u32 this_item_size;
2092         u32 old_left_item_size;
2093
2094         slot = path->slots[1];
2095         if (slot == 0)
2096                 return 1;
2097         if (!path->nodes[1])
2098                 return 1;
2099
2100         right_nritems = btrfs_header_nritems(right);
2101         if (right_nritems == 0) {
2102                 return 1;
2103         }
2104
2105         WARN_ON(!btrfs_tree_locked(path->nodes[1]));
2106
2107         left = read_node_slot(root, path->nodes[1], slot - 1);
2108         btrfs_tree_lock(left);
2109         free_space = btrfs_leaf_free_space(root, left);
2110         if (free_space < data_size + sizeof(struct btrfs_item)) {
2111                 ret = 1;
2112                 goto out;
2113         }
2114
2115         /* cow and double check */
2116         ret = btrfs_cow_block(trans, root, left,
2117                               path->nodes[1], slot - 1, &left, 0);
2118         if (ret) {
2119                 /* we hit -ENOSPC, but it isn't fatal here */
2120                 ret = 1;
2121                 goto out;
2122         }
2123
2124         free_space = btrfs_leaf_free_space(root, left);
2125         if (free_space < data_size + sizeof(struct btrfs_item)) {
2126                 ret = 1;
2127                 goto out;
2128         }
2129
2130         if (empty)
2131                 nr = right_nritems;
2132         else
2133                 nr = right_nritems - 1;
2134
2135         for (i = 0; i < nr; i++) {
2136                 item = btrfs_item_nr(right, i);
2137                 if (!right->map_token) {
2138                         map_extent_buffer(right, (unsigned long)item,
2139                                         sizeof(struct btrfs_item),
2140                                         &right->map_token, &right->kaddr,
2141                                         &right->map_start, &right->map_len,
2142                                         KM_USER1);
2143                 }
2144
2145                 if (path->slots[0] == i)
2146                         push_space += data_size + sizeof(*item);
2147
2148                 this_item_size = btrfs_item_size(right, item);
2149                 if (this_item_size + sizeof(*item) + push_space > free_space)
2150                         break;
2151
2152                 push_items++;
2153                 push_space += this_item_size + sizeof(*item);
2154         }
2155
2156         if (right->map_token) {
2157                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2158                 right->map_token = NULL;
2159         }
2160
2161         if (push_items == 0) {
2162                 ret = 1;
2163                 goto out;
2164         }
2165         if (!empty && push_items == btrfs_header_nritems(right))
2166                 WARN_ON(1);
2167
2168         /* push data from right to left */
2169         copy_extent_buffer(left, right,
2170                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
2171                            btrfs_item_nr_offset(0),
2172                            push_items * sizeof(struct btrfs_item));
2173
2174         push_space = BTRFS_LEAF_DATA_SIZE(root) -
2175                      btrfs_item_offset_nr(right, push_items -1);
2176
2177         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
2178                      leaf_data_end(root, left) - push_space,
2179                      btrfs_leaf_data(right) +
2180                      btrfs_item_offset_nr(right, push_items - 1),
2181                      push_space);
2182         old_left_nritems = btrfs_header_nritems(left);
2183         BUG_ON(old_left_nritems < 0);
2184
2185         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
2186         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
2187                 u32 ioff;
2188
2189                 item = btrfs_item_nr(left, i);
2190                 if (!left->map_token) {
2191                         map_extent_buffer(left, (unsigned long)item,
2192                                         sizeof(struct btrfs_item),
2193                                         &left->map_token, &left->kaddr,
2194                                         &left->map_start, &left->map_len,
2195                                         KM_USER1);
2196                 }
2197
2198                 ioff = btrfs_item_offset(left, item);
2199                 btrfs_set_item_offset(left, item,
2200                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
2201         }
2202         btrfs_set_header_nritems(left, old_left_nritems + push_items);
2203         if (left->map_token) {
2204                 unmap_extent_buffer(left, left->map_token, KM_USER1);
2205                 left->map_token = NULL;
2206         }
2207
2208         /* fixup right node */
2209         if (push_items > right_nritems) {
2210                 printk("push items %d nr %u\n", push_items, right_nritems);
2211                 WARN_ON(1);
2212         }
2213
2214         if (push_items < right_nritems) {
2215                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
2216                                                   leaf_data_end(root, right);
2217                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
2218                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
2219                                       btrfs_leaf_data(right) +
2220                                       leaf_data_end(root, right), push_space);
2221
2222                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
2223                               btrfs_item_nr_offset(push_items),
2224                              (btrfs_header_nritems(right) - push_items) *
2225                              sizeof(struct btrfs_item));
2226         }
2227         right_nritems -= push_items;
2228         btrfs_set_header_nritems(right, right_nritems);
2229         push_space = BTRFS_LEAF_DATA_SIZE(root);
2230         for (i = 0; i < right_nritems; i++) {
2231                 item = btrfs_item_nr(right, i);
2232
2233                 if (!right->map_token) {
2234                         map_extent_buffer(right, (unsigned long)item,
2235                                         sizeof(struct btrfs_item),
2236                                         &right->map_token, &right->kaddr,
2237                                         &right->map_start, &right->map_len,
2238                                         KM_USER1);
2239                 }
2240
2241                 push_space = push_space - btrfs_item_size(right, item);
2242                 btrfs_set_item_offset(right, item, push_space);
2243         }
2244         if (right->map_token) {
2245                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2246                 right->map_token = NULL;
2247         }
2248
2249         btrfs_mark_buffer_dirty(left);
2250         if (right_nritems)
2251                 btrfs_mark_buffer_dirty(right);
2252
2253         btrfs_item_key(right, &disk_key, 0);
2254         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
2255         if (wret)
2256                 ret = wret;
2257
2258         /* then fixup the leaf pointer in the path */
2259         if (path->slots[0] < push_items) {
2260                 path->slots[0] += old_left_nritems;
2261                 if (btrfs_header_nritems(path->nodes[0]) == 0)
2262                         clean_tree_block(trans, root, path->nodes[0]);
2263                 btrfs_tree_unlock(path->nodes[0]);
2264                 free_extent_buffer(path->nodes[0]);
2265                 path->nodes[0] = left;
2266                 path->slots[1] -= 1;
2267         } else {
2268                 btrfs_tree_unlock(left);
2269                 free_extent_buffer(left);
2270                 path->slots[0] -= push_items;
2271         }
2272         BUG_ON(path->slots[0] < 0);
2273         return ret;
2274 out:
2275         btrfs_tree_unlock(left);
2276         free_extent_buffer(left);
2277         return ret;
2278 }
2279
2280 /*
2281  * split the path's leaf in two, making sure there is at least data_size
2282  * available for the resulting leaf level of the path.
2283  *
2284  * returns 0 if all went well and < 0 on failure.
2285  */
2286 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
2287                       *root, struct btrfs_key *ins_key,
2288                       struct btrfs_path *path, int data_size, int extend)
2289 {
2290         u64 root_gen;
2291         struct extent_buffer *l;
2292         u32 nritems;
2293         int mid;
2294         int slot;
2295         struct extent_buffer *right;
2296         int space_needed = data_size + sizeof(struct btrfs_item);
2297         int data_copy_size;
2298         int rt_data_off;
2299         int i;
2300         int ret = 0;
2301         int wret;
2302         int double_split;
2303         int num_doubles = 0;
2304         struct btrfs_disk_key disk_key;
2305
2306         if (extend)
2307                 space_needed = data_size;
2308
2309         if (root->ref_cows)
2310                 root_gen = trans->transid;
2311         else
2312                 root_gen = 0;
2313
2314         /* first try to make some room by pushing left and right */
2315         if (ins_key->type != BTRFS_DIR_ITEM_KEY) {
2316                 wret = push_leaf_right(trans, root, path, data_size, 0);
2317                 if (wret < 0) {
2318                         return wret;
2319                 }
2320                 if (wret) {
2321                         wret = push_leaf_left(trans, root, path, data_size, 0);
2322                         if (wret < 0)
2323                                 return wret;
2324                 }
2325                 l = path->nodes[0];
2326
2327                 /* did the pushes work? */
2328                 if (btrfs_leaf_free_space(root, l) >= space_needed)
2329                         return 0;
2330         }
2331
2332         if (!path->nodes[1]) {
2333                 ret = insert_new_root(trans, root, path, 1);
2334                 if (ret)
2335                         return ret;
2336         }
2337 again:
2338         double_split = 0;
2339         l = path->nodes[0];
2340         slot = path->slots[0];
2341         nritems = btrfs_header_nritems(l);
2342         mid = (nritems + 1)/ 2;
2343
2344         btrfs_item_key(l, &disk_key, 0);
2345
2346         right = btrfs_alloc_free_block(trans, root, root->leafsize,
2347                                          root->root_key.objectid,
2348                                          root_gen, disk_key.objectid, 0,
2349                                          l->start, 0);
2350         if (IS_ERR(right)) {
2351                 BUG_ON(1);
2352                 return PTR_ERR(right);
2353         }
2354
2355         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2356         btrfs_set_header_bytenr(right, right->start);
2357         btrfs_set_header_generation(right, trans->transid);
2358         btrfs_set_header_owner(right, root->root_key.objectid);
2359         btrfs_set_header_level(right, 0);
2360         write_extent_buffer(right, root->fs_info->fsid,
2361                             (unsigned long)btrfs_header_fsid(right),
2362                             BTRFS_FSID_SIZE);
2363
2364         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2365                             (unsigned long)btrfs_header_chunk_tree_uuid(right),
2366                             BTRFS_UUID_SIZE);
2367         if (mid <= slot) {
2368                 if (nritems == 1 ||
2369                     leaf_space_used(l, mid, nritems - mid) + space_needed >
2370                         BTRFS_LEAF_DATA_SIZE(root)) {
2371                         if (slot >= nritems) {
2372                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2373                                 btrfs_set_header_nritems(right, 0);
2374                                 wret = insert_ptr(trans, root, path,
2375                                                   &disk_key, right->start,
2376                                                   path->slots[1] + 1, 1);
2377                                 if (wret)
2378                                         ret = wret;
2379
2380                                 btrfs_tree_unlock(path->nodes[0]);
2381                                 free_extent_buffer(path->nodes[0]);
2382                                 path->nodes[0] = right;
2383                                 path->slots[0] = 0;
2384                                 path->slots[1] += 1;
2385                                 btrfs_mark_buffer_dirty(right);
2386                                 return ret;
2387                         }
2388                         mid = slot;
2389                         if (mid != nritems &&
2390                             leaf_space_used(l, mid, nritems - mid) +
2391                             space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2392                                 double_split = 1;
2393                         }
2394                 }
2395         } else {
2396                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
2397                         BTRFS_LEAF_DATA_SIZE(root)) {
2398                         if (!extend && slot == 0) {
2399                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2400                                 btrfs_set_header_nritems(right, 0);
2401                                 wret = insert_ptr(trans, root, path,
2402                                                   &disk_key,
2403                                                   right->start,
2404                                                   path->slots[1], 1);
2405                                 if (wret)
2406                                         ret = wret;
2407                                 btrfs_tree_unlock(path->nodes[0]);
2408                                 free_extent_buffer(path->nodes[0]);
2409                                 path->nodes[0] = right;
2410                                 path->slots[0] = 0;
2411                                 if (path->slots[1] == 0) {
2412                                         wret = fixup_low_keys(trans, root,
2413                                                    path, &disk_key, 1);
2414                                         if (wret)
2415                                                 ret = wret;
2416                                 }
2417                                 btrfs_mark_buffer_dirty(right);
2418                                 return ret;
2419                         } else if (extend && slot == 0) {
2420                                 mid = 1;
2421                         } else {
2422                                 mid = slot;
2423                                 if (mid != nritems &&
2424                                     leaf_space_used(l, mid, nritems - mid) +
2425                                     space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2426                                         double_split = 1;
2427                                 }
2428                         }
2429                 }
2430         }
2431         nritems = nritems - mid;
2432         btrfs_set_header_nritems(right, nritems);
2433         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2434
2435         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2436                            btrfs_item_nr_offset(mid),
2437                            nritems * sizeof(struct btrfs_item));
2438
2439         copy_extent_buffer(right, l,
2440                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2441                      data_copy_size, btrfs_leaf_data(l) +
2442                      leaf_data_end(root, l), data_copy_size);
2443
2444         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2445                       btrfs_item_end_nr(l, mid);
2446
2447         for (i = 0; i < nritems; i++) {
2448                 struct btrfs_item *item = btrfs_item_nr(right, i);
2449                 u32 ioff;
2450
2451                 if (!right->map_token) {
2452                         map_extent_buffer(right, (unsigned long)item,
2453                                         sizeof(struct btrfs_item),
2454                                         &right->map_token, &right->kaddr,
2455                                         &right->map_start, &right->map_len,
2456                                         KM_USER1);
2457                 }
2458
2459                 ioff = btrfs_item_offset(right, item);
2460                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2461         }
2462
2463         if (right->map_token) {
2464                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2465                 right->map_token = NULL;
2466         }
2467
2468         btrfs_set_header_nritems(l, mid);
2469         ret = 0;
2470         btrfs_item_key(right, &disk_key, 0);
2471         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2472                           path->slots[1] + 1, 1);
2473         if (wret)
2474                 ret = wret;
2475
2476         btrfs_mark_buffer_dirty(right);
2477         btrfs_mark_buffer_dirty(l);
2478         BUG_ON(path->slots[0] != slot);
2479
2480         if (mid <= slot) {
2481                 btrfs_tree_unlock(path->nodes[0]);
2482                 free_extent_buffer(path->nodes[0]);
2483                 path->nodes[0] = right;
2484                 path->slots[0] -= mid;
2485                 path->slots[1] += 1;
2486         } else {
2487                 btrfs_tree_unlock(right);
2488                 free_extent_buffer(right);
2489         }
2490
2491         BUG_ON(path->slots[0] < 0);
2492
2493         if (double_split) {
2494                 BUG_ON(num_doubles != 0);
2495                 num_doubles++;
2496                 goto again;
2497         }
2498         return ret;
2499 }
2500
2501 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2502                         struct btrfs_root *root,
2503                         struct btrfs_path *path,
2504                         u32 new_size, int from_end)
2505 {
2506         int ret = 0;
2507         int slot;
2508         int slot_orig;
2509         struct extent_buffer *leaf;
2510         struct btrfs_item *item;
2511         u32 nritems;
2512         unsigned int data_end;
2513         unsigned int old_data_start;
2514         unsigned int old_size;
2515         unsigned int size_diff;
2516         int i;
2517
2518         slot_orig = path->slots[0];
2519         leaf = path->nodes[0];
2520         slot = path->slots[0];
2521
2522         old_size = btrfs_item_size_nr(leaf, slot);
2523         if (old_size == new_size)
2524                 return 0;
2525
2526         nritems = btrfs_header_nritems(leaf);
2527         data_end = leaf_data_end(root, leaf);
2528
2529         old_data_start = btrfs_item_offset_nr(leaf, slot);
2530
2531         size_diff = old_size - new_size;
2532
2533         BUG_ON(slot < 0);
2534         BUG_ON(slot >= nritems);
2535
2536         /*
2537          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2538          */
2539         /* first correct the data pointers */
2540         for (i = slot; i < nritems; i++) {
2541                 u32 ioff;
2542                 item = btrfs_item_nr(leaf, i);
2543
2544                 if (!leaf->map_token) {
2545                         map_extent_buffer(leaf, (unsigned long)item,
2546                                         sizeof(struct btrfs_item),
2547                                         &leaf->map_token, &leaf->kaddr,
2548                                         &leaf->map_start, &leaf->map_len,
2549                                         KM_USER1);
2550                 }
2551
2552                 ioff = btrfs_item_offset(leaf, item);
2553                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2554         }
2555
2556         if (leaf->map_token) {
2557                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2558                 leaf->map_token = NULL;
2559         }
2560
2561         /* shift the data */
2562         if (from_end) {
2563                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2564                               data_end + size_diff, btrfs_leaf_data(leaf) +
2565                               data_end, old_data_start + new_size - data_end);
2566         } else {
2567                 struct btrfs_disk_key disk_key;
2568                 u64 offset;
2569
2570                 btrfs_item_key(leaf, &disk_key, slot);
2571
2572                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2573                         unsigned long ptr;
2574                         struct btrfs_file_extent_item *fi;
2575
2576                         fi = btrfs_item_ptr(leaf, slot,
2577                                             struct btrfs_file_extent_item);
2578                         fi = (struct btrfs_file_extent_item *)(
2579                              (unsigned long)fi - size_diff);
2580
2581                         if (btrfs_file_extent_type(leaf, fi) ==
2582                             BTRFS_FILE_EXTENT_INLINE) {
2583                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2584                                 memmove_extent_buffer(leaf, ptr,
2585                                         (unsigned long)fi,
2586                                         offsetof(struct btrfs_file_extent_item,
2587                                                  disk_bytenr));
2588                         }
2589                 }
2590
2591                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2592                               data_end + size_diff, btrfs_leaf_data(leaf) +
2593                               data_end, old_data_start - data_end);
2594
2595                 offset = btrfs_disk_key_offset(&disk_key);
2596                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2597                 btrfs_set_item_key(leaf, &disk_key, slot);
2598                 if (slot == 0)
2599                         fixup_low_keys(trans, root, path, &disk_key, 1);
2600         }
2601
2602         item = btrfs_item_nr(leaf, slot);
2603         btrfs_set_item_size(leaf, item, new_size);
2604         btrfs_mark_buffer_dirty(leaf);
2605
2606         ret = 0;
2607         if (btrfs_leaf_free_space(root, leaf) < 0) {
2608                 btrfs_print_leaf(root, leaf);
2609                 BUG();
2610         }
2611         return ret;
2612 }
2613
2614 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2615                       struct btrfs_root *root, struct btrfs_path *path,
2616                       u32 data_size)
2617 {
2618         int ret = 0;
2619         int slot;
2620         int slot_orig;
2621         struct extent_buffer *leaf;
2622         struct btrfs_item *item;
2623         u32 nritems;
2624         unsigned int data_end;
2625         unsigned int old_data;
2626         unsigned int old_size;
2627         int i;
2628
2629         slot_orig = path->slots[0];
2630         leaf = path->nodes[0];
2631
2632         nritems = btrfs_header_nritems(leaf);
2633         data_end = leaf_data_end(root, leaf);
2634
2635         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2636                 btrfs_print_leaf(root, leaf);
2637                 BUG();
2638         }
2639         slot = path->slots[0];
2640         old_data = btrfs_item_end_nr(leaf, slot);
2641
2642         BUG_ON(slot < 0);
2643         if (slot >= nritems) {
2644                 btrfs_print_leaf(root, leaf);
2645                 printk("slot %d too large, nritems %d\n", slot, nritems);
2646                 BUG_ON(1);
2647         }
2648
2649         /*
2650          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2651          */
2652         /* first correct the data pointers */
2653         for (i = slot; i < nritems; i++) {
2654                 u32 ioff;
2655                 item = btrfs_item_nr(leaf, i);
2656
2657                 if (!leaf->map_token) {
2658                         map_extent_buffer(leaf, (unsigned long)item,
2659                                         sizeof(struct btrfs_item),
2660                                         &leaf->map_token, &leaf->kaddr,
2661                                         &leaf->map_start, &leaf->map_len,
2662                                         KM_USER1);
2663                 }
2664                 ioff = btrfs_item_offset(leaf, item);
2665                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2666         }
2667
2668         if (leaf->map_token) {
2669                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2670                 leaf->map_token = NULL;
2671         }
2672
2673         /* shift the data */
2674         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2675                       data_end - data_size, btrfs_leaf_data(leaf) +
2676                       data_end, old_data - data_end);
2677
2678         data_end = old_data;
2679         old_size = btrfs_item_size_nr(leaf, slot);
2680         item = btrfs_item_nr(leaf, slot);
2681         btrfs_set_item_size(leaf, item, old_size + data_size);
2682         btrfs_mark_buffer_dirty(leaf);
2683
2684         ret = 0;
2685         if (btrfs_leaf_free_space(root, leaf) < 0) {
2686                 btrfs_print_leaf(root, leaf);
2687                 BUG();
2688         }
2689         return ret;
2690 }
2691
2692 /*
2693  * Given a key and some data, insert an item into the tree.
2694  * This does all the path init required, making room in the tree if needed.
2695  */
2696 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2697                             struct btrfs_root *root,
2698                             struct btrfs_path *path,
2699                             struct btrfs_key *cpu_key, u32 *data_size,
2700                             int nr)
2701 {
2702         struct extent_buffer *leaf;
2703         struct btrfs_item *item;
2704         int ret = 0;
2705         int slot;
2706         int slot_orig;
2707         int i;
2708         u32 nritems;
2709         u32 total_size = 0;
2710         u32 total_data = 0;
2711         unsigned int data_end;
2712         struct btrfs_disk_key disk_key;
2713
2714         for (i = 0; i < nr; i++) {
2715                 total_data += data_size[i];
2716         }
2717
2718         total_size = total_data + (nr * sizeof(struct btrfs_item));
2719         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2720         if (ret == 0) {
2721                 return -EEXIST;
2722         }
2723         if (ret < 0)
2724                 goto out;
2725
2726         slot_orig = path->slots[0];
2727         leaf = path->nodes[0];
2728
2729         nritems = btrfs_header_nritems(leaf);
2730         data_end = leaf_data_end(root, leaf);
2731
2732         if (btrfs_leaf_free_space(root, leaf) <
2733             sizeof(struct btrfs_item) + total_size) {
2734                 btrfs_print_leaf(root, leaf);
2735                 printk("not enough freespace need %u have %d\n",
2736                        total_size, btrfs_leaf_free_space(root, leaf));
2737                 BUG();
2738         }
2739
2740         slot = path->slots[0];
2741         BUG_ON(slot < 0);
2742
2743         if (slot != nritems) {
2744                 int i;
2745                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2746
2747                 if (old_data < data_end) {
2748                         btrfs_print_leaf(root, leaf);
2749                         printk("slot %d old_data %d data_end %d\n",
2750                                slot, old_data, data_end);
2751                         BUG_ON(1);
2752                 }
2753                 /*
2754                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2755                  */
2756                 /* first correct the data pointers */
2757                 WARN_ON(leaf->map_token);
2758                 for (i = slot; i < nritems; i++) {
2759                         u32 ioff;
2760
2761                         item = btrfs_item_nr(leaf, i);
2762                         if (!leaf->map_token) {
2763                                 map_extent_buffer(leaf, (unsigned long)item,
2764                                         sizeof(struct btrfs_item),
2765                                         &leaf->map_token, &leaf->kaddr,
2766                                         &leaf->map_start, &leaf->map_len,
2767                                         KM_USER1);
2768                         }
2769
2770                         ioff = btrfs_item_offset(leaf, item);
2771                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2772                 }
2773                 if (leaf->map_token) {
2774                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2775                         leaf->map_token = NULL;
2776                 }
2777
2778                 /* shift the items */
2779                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2780                               btrfs_item_nr_offset(slot),
2781                               (nritems - slot) * sizeof(struct btrfs_item));
2782
2783                 /* shift the data */
2784                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2785                               data_end - total_data, btrfs_leaf_data(leaf) +
2786                               data_end, old_data - data_end);
2787                 data_end = old_data;
2788         }
2789
2790         /* setup the item for the new data */
2791         for (i = 0; i < nr; i++) {
2792                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2793                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2794                 item = btrfs_item_nr(leaf, slot + i);
2795                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2796                 data_end -= data_size[i];
2797                 btrfs_set_item_size(leaf, item, data_size[i]);
2798         }
2799         btrfs_set_header_nritems(leaf, nritems + nr);
2800         btrfs_mark_buffer_dirty(leaf);
2801
2802         ret = 0;
2803         if (slot == 0) {
2804                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2805                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2806         }
2807
2808         if (btrfs_leaf_free_space(root, leaf) < 0) {
2809                 btrfs_print_leaf(root, leaf);
2810                 BUG();
2811         }
2812 out:
2813         return ret;
2814 }
2815
2816 /*
2817  * Given a key and some data, insert an item into the tree.
2818  * This does all the path init required, making room in the tree if needed.
2819  */
2820 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2821                       *root, struct btrfs_key *cpu_key, void *data, u32
2822                       data_size)
2823 {
2824         int ret = 0;
2825         struct btrfs_path *path;
2826         struct extent_buffer *leaf;
2827         unsigned long ptr;
2828
2829         path = btrfs_alloc_path();
2830         BUG_ON(!path);
2831         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2832         if (!ret) {
2833                 leaf = path->nodes[0];
2834                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2835                 write_extent_buffer(leaf, data, ptr, data_size);
2836                 btrfs_mark_buffer_dirty(leaf);
2837         }
2838         btrfs_free_path(path);
2839         return ret;
2840 }
2841
2842 /*
2843  * delete the pointer from a given node.
2844  *
2845  * If the delete empties a node, the node is removed from the tree,
2846  * continuing all the way the root if required.  The root is converted into
2847  * a leaf if all the nodes are emptied.
2848  */
2849 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2850                    struct btrfs_path *path, int level, int slot)
2851 {
2852         struct extent_buffer *parent = path->nodes[level];
2853         u32 nritems;
2854         int ret = 0;
2855         int wret;
2856
2857         nritems = btrfs_header_nritems(parent);
2858         if (slot != nritems -1) {
2859                 memmove_extent_buffer(parent,
2860                               btrfs_node_key_ptr_offset(slot),
2861                               btrfs_node_key_ptr_offset(slot + 1),
2862                               sizeof(struct btrfs_key_ptr) *
2863                               (nritems - slot - 1));
2864         }
2865         nritems--;
2866         btrfs_set_header_nritems(parent, nritems);
2867         if (nritems == 0 && parent == root->node) {
2868                 BUG_ON(btrfs_header_level(root->node) != 1);
2869                 /* just turn the root into a leaf and break */
2870                 btrfs_set_header_level(root->node, 0);
2871         } else if (slot == 0) {
2872                 struct btrfs_disk_key disk_key;
2873
2874                 btrfs_node_key(parent, &disk_key, 0);
2875                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2876                 if (wret)
2877                         ret = wret;
2878         }
2879         btrfs_mark_buffer_dirty(parent);
2880         return ret;
2881 }
2882
2883 /*
2884  * delete the item at the leaf level in path.  If that empties
2885  * the leaf, remove it from the tree
2886  */
2887 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2888                     struct btrfs_path *path, int slot, int nr)
2889 {
2890         struct extent_buffer *leaf;
2891         struct btrfs_item *item;
2892         int last_off;
2893         int dsize = 0;
2894         int ret = 0;
2895         int wret;
2896         int i;
2897         u32 nritems;
2898
2899         leaf = path->nodes[0];
2900         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2901
2902         for (i = 0; i < nr; i++)
2903                 dsize += btrfs_item_size_nr(leaf, slot + i);
2904
2905         nritems = btrfs_header_nritems(leaf);
2906
2907         if (slot + nr != nritems) {
2908                 int i;
2909                 int data_end = leaf_data_end(root, leaf);
2910
2911                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2912                               data_end + dsize,
2913                               btrfs_leaf_data(leaf) + data_end,
2914                               last_off - data_end);
2915
2916                 for (i = slot + nr; i < nritems; i++) {
2917                         u32 ioff;
2918
2919                         item = btrfs_item_nr(leaf, i);
2920                         if (!leaf->map_token) {
2921                                 map_extent_buffer(leaf, (unsigned long)item,
2922                                         sizeof(struct btrfs_item),
2923                                         &leaf->map_token, &leaf->kaddr,
2924                                         &leaf->map_start, &leaf->map_len,
2925                                         KM_USER1);
2926                         }
2927                         ioff = btrfs_item_offset(leaf, item);
2928                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2929                 }
2930
2931                 if (leaf->map_token) {
2932                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2933                         leaf->map_token = NULL;
2934                 }
2935
2936                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2937                               btrfs_item_nr_offset(slot + nr),
2938                               sizeof(struct btrfs_item) *
2939                               (nritems - slot - nr));
2940         }
2941         btrfs_set_header_nritems(leaf, nritems - nr);
2942         nritems -= nr;
2943
2944         /* delete the leaf if we've emptied it */
2945         if (nritems == 0) {
2946                 if (leaf == root->node) {
2947                         btrfs_set_header_level(leaf, 0);
2948                 } else {
2949                         u64 root_gen = btrfs_header_generation(path->nodes[1]);
2950                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
2951                         if (wret)
2952                                 ret = wret;
2953                         wret = btrfs_free_extent(trans, root,
2954                                          leaf->start, leaf->len,
2955                                          btrfs_header_owner(path->nodes[1]),
2956                                          root_gen, 0, 0, 1);
2957                         if (wret)
2958                                 ret = wret;
2959                 }
2960         } else {
2961                 int used = leaf_space_used(leaf, 0, nritems);
2962                 if (slot == 0) {
2963                         struct btrfs_disk_key disk_key;
2964
2965                         btrfs_item_key(leaf, &disk_key, 0);
2966                         wret = fixup_low_keys(trans, root, path,
2967                                               &disk_key, 1);
2968                         if (wret)
2969                                 ret = wret;
2970                 }
2971
2972                 /* delete the leaf if it is mostly empty */
2973                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2974                         /* push_leaf_left fixes the path.
2975                          * make sure the path still points to our leaf
2976                          * for possible call to del_ptr below
2977                          */
2978                         slot = path->slots[1];
2979                         extent_buffer_get(leaf);
2980
2981                         wret = push_leaf_left(trans, root, path, 1, 1);
2982                         if (wret < 0 && wret != -ENOSPC)
2983                                 ret = wret;
2984
2985                         if (path->nodes[0] == leaf &&
2986                             btrfs_header_nritems(leaf)) {
2987                                 wret = push_leaf_right(trans, root, path, 1, 1);
2988                                 if (wret < 0 && wret != -ENOSPC)
2989                                         ret = wret;
2990                         }
2991
2992                         if (btrfs_header_nritems(leaf) == 0) {
2993                                 u64 root_gen;
2994                                 u64 bytenr = leaf->start;
2995                                 u32 blocksize = leaf->len;
2996
2997                                 root_gen = btrfs_header_generation(
2998                                                            path->nodes[1]);
2999
3000                                 wret = del_ptr(trans, root, path, 1, slot);
3001                                 if (wret)
3002                                         ret = wret;
3003
3004                                 free_extent_buffer(leaf);
3005                                 wret = btrfs_free_extent(trans, root, bytenr,
3006                                              blocksize,
3007                                              btrfs_header_owner(path->nodes[1]),
3008                                              root_gen, 0, 0, 1);
3009                                 if (wret)
3010                                         ret = wret;
3011                         } else {
3012                                 /* if we're still in the path, make sure
3013                                  * we're dirty.  Otherwise, one of the
3014                                  * push_leaf functions must have already
3015                                  * dirtied this buffer
3016                                  */
3017                                 if (path->nodes[0] == leaf)
3018                                         btrfs_mark_buffer_dirty(leaf);
3019                                 free_extent_buffer(leaf);
3020                         }
3021                 } else {
3022                         btrfs_mark_buffer_dirty(leaf);
3023                 }
3024         }
3025         return ret;
3026 }
3027
3028 /*
3029  * search the tree again to find a leaf with lesser keys
3030  * returns 0 if it found something or 1 if there are no lesser leaves.
3031  * returns < 0 on io errors.
3032  */
3033 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
3034 {
3035         struct btrfs_key key;
3036         struct btrfs_disk_key found_key;
3037         int ret;
3038
3039         btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
3040
3041         if (key.offset > 0)
3042                 key.offset--;
3043         else if (key.type > 0)
3044                 key.type--;
3045         else if (key.objectid > 0)
3046                 key.objectid--;
3047         else
3048                 return 1;
3049
3050         btrfs_release_path(root, path);
3051         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3052         if (ret < 0)
3053                 return ret;
3054         btrfs_item_key(path->nodes[0], &found_key, 0);
3055         ret = comp_keys(&found_key, &key);
3056         if (ret < 0)
3057                 return 0;
3058         return 1;
3059 }
3060
3061 /*
3062  * A helper function to walk down the tree starting at min_key, and looking
3063  * for nodes or leaves that are either in cache or have a minimum
3064  * transaction id.  This is used by the btree defrag code, but could
3065  * also be used to search for blocks that have changed since a given
3066  * transaction id.
3067  *
3068  * This does not cow, but it does stuff the starting key it finds back
3069  * into min_key, so you can call btrfs_search_slot with cow=1 on the
3070  * key and get a writable path.
3071  *
3072  * This does lock as it descends, and path->keep_locks should be set
3073  * to 1 by the caller.
3074  *
3075  * This honors path->lowest_level to prevent descent past a given level
3076  * of the tree.
3077  *
3078  * returns zero if something useful was found, < 0 on error and 1 if there
3079  * was nothing in the tree that matched the search criteria.
3080  */
3081 int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
3082                          struct btrfs_path *path, int cache_only,
3083                          u64 min_trans)
3084 {
3085         struct extent_buffer *cur;
3086         struct btrfs_key found_key;
3087         int slot;
3088         int sret;
3089         u32 nritems;
3090         int level;
3091         int ret = 1;
3092
3093 again:
3094         cur = btrfs_lock_root_node(root);
3095         level = btrfs_header_level(cur);
3096         path->nodes[level] = cur;
3097         path->locks[level] = 1;
3098
3099         if (btrfs_header_generation(cur) < min_trans) {
3100                 ret = 1;
3101                 goto out;
3102         }
3103         while(1) {
3104                 nritems = btrfs_header_nritems(cur);
3105                 level = btrfs_header_level(cur);
3106                 sret = bin_search(cur, min_key, level, &slot);
3107
3108                 /* at level = 0, we're done, setup the path and exit */
3109                 if (level == 0) {
3110                         ret = 0;
3111                         path->slots[level] = slot;
3112                         btrfs_item_key_to_cpu(cur, &found_key, slot);
3113                         goto out;
3114                 }
3115                 if (sret && slot > 0)
3116                         slot--;
3117                 /*
3118                  * check this node pointer against the cache_only and
3119                  * min_trans parameters.  If it isn't in cache or is too
3120                  * old, skip to the next one.
3121                  */
3122                 while(slot < nritems) {
3123                         u64 blockptr;
3124                         u64 gen;
3125                         struct extent_buffer *tmp;
3126                         blockptr = btrfs_node_blockptr(cur, slot);
3127                         gen = btrfs_node_ptr_generation(cur, slot);
3128                         if (gen < min_trans) {
3129                                 slot++;
3130                                 continue;
3131                         }
3132                         if (!cache_only)
3133                                 break;
3134
3135                         tmp = btrfs_find_tree_block(root, blockptr,
3136                                             btrfs_level_size(root, level - 1));
3137
3138                         if (tmp && btrfs_buffer_uptodate(tmp, gen)) {
3139                                 free_extent_buffer(tmp);
3140                                 break;
3141                         }
3142                         if (tmp)
3143                                 free_extent_buffer(tmp);
3144                         slot++;
3145                 }
3146                 /*
3147                  * we didn't find a candidate key in this node, walk forward
3148                  * and find another one
3149                  */
3150                 if (slot >= nritems) {
3151                         ret = btrfs_find_next_key(root, path, min_key, level,
3152                                                   cache_only, min_trans);
3153                         if (ret == 0) {
3154                                 btrfs_release_path(root, path);
3155                                 goto again;
3156                         } else {
3157                                 goto out;
3158                         }
3159                 }
3160                 /* save our key for returning back */
3161                 btrfs_node_key_to_cpu(cur, &found_key, slot);
3162                 path->slots[level] = slot;
3163                 if (level == path->lowest_level) {
3164                         ret = 0;
3165                         unlock_up(path, level, 1);
3166                         goto out;
3167                 }
3168                 cur = read_node_slot(root, cur, slot);
3169
3170                 btrfs_tree_lock(cur);
3171                 path->locks[level - 1] = 1;
3172                 path->nodes[level - 1] = cur;
3173                 unlock_up(path, level, 1);
3174         }
3175 out:
3176         if (ret == 0)
3177                 memcpy(min_key, &found_key, sizeof(found_key));
3178         return ret;
3179 }
3180
3181 /*
3182  * this is similar to btrfs_next_leaf, but does not try to preserve
3183  * and fixup the path.  It looks for and returns the next key in the
3184  * tree based on the current path and the cache_only and min_trans
3185  * parameters.
3186  *
3187  * 0 is returned if another key is found, < 0 if there are any errors
3188  * and 1 is returned if there are no higher keys in the tree
3189  *
3190  * path->keep_locks should be set to 1 on the search made before
3191  * calling this function.
3192  */
3193 int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
3194                         struct btrfs_key *key, int lowest_level,
3195                         int cache_only, u64 min_trans)
3196 {
3197         int level = lowest_level;
3198         int slot;
3199         struct extent_buffer *c;
3200
3201         while(level < BTRFS_MAX_LEVEL) {
3202                 if (!path->nodes[level])
3203                         return 1;
3204
3205                 slot = path->slots[level] + 1;
3206                 c = path->nodes[level];
3207 next:
3208                 if (slot >= btrfs_header_nritems(c)) {
3209                         level++;
3210                         if (level == BTRFS_MAX_LEVEL) {
3211                                 return 1;
3212                         }
3213                         continue;
3214                 }
3215                 if (level == 0)
3216                         btrfs_item_key_to_cpu(c, key, slot);
3217                 else {
3218                         u64 blockptr = btrfs_node_blockptr(c, slot);
3219                         u64 gen = btrfs_node_ptr_generation(c, slot);
3220
3221                         if (cache_only) {
3222                                 struct extent_buffer *cur;
3223                                 cur = btrfs_find_tree_block(root, blockptr,
3224                                             btrfs_level_size(root, level - 1));
3225                                 if (!cur || !btrfs_buffer_uptodate(cur, gen)) {
3226                                         slot++;
3227                                         if (cur)
3228                                                 free_extent_buffer(cur);
3229                                         goto next;
3230                                 }
3231                                 free_extent_buffer(cur);
3232                         }
3233                         if (gen < min_trans) {
3234                                 slot++;
3235                                 goto next;
3236                         }
3237                         btrfs_node_key_to_cpu(c, key, slot);
3238                 }
3239                 return 0;
3240         }
3241         return 1;
3242 }
3243
3244 /*
3245  * search the tree again to find a leaf with greater keys
3246  * returns 0 if it found something or 1 if there are no greater leaves.
3247  * returns < 0 on io errors.
3248  */
3249 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
3250 {
3251         int slot;
3252         int level = 1;
3253         struct extent_buffer *c;
3254         struct extent_buffer *next = NULL;
3255         struct btrfs_key key;
3256         u32 nritems;
3257         int ret;
3258
3259         nritems = btrfs_header_nritems(path->nodes[0]);
3260         if (nritems == 0) {
3261                 return 1;
3262         }
3263
3264         btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
3265
3266         btrfs_release_path(root, path);
3267         path->keep_locks = 1;
3268         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3269         path->keep_locks = 0;
3270
3271         if (ret < 0)
3272                 return ret;
3273
3274         nritems = btrfs_header_nritems(path->nodes[0]);
3275         /*
3276          * by releasing the path above we dropped all our locks.  A balance
3277          * could have added more items next to the key that used to be
3278          * at the very end of the block.  So, check again here and
3279          * advance the path if there are now more items available.
3280          */
3281         if (nritems > 0 && path->slots[0] < nritems - 1) {
3282                 path->slots[0]++;
3283                 goto done;
3284         }
3285
3286         while(level < BTRFS_MAX_LEVEL) {
3287                 if (!path->nodes[level])
3288                         return 1;
3289
3290                 slot = path->slots[level] + 1;
3291                 c = path->nodes[level];
3292                 if (slot >= btrfs_header_nritems(c)) {
3293                         level++;
3294                         if (level == BTRFS_MAX_LEVEL) {
3295                                 return 1;
3296                         }
3297                         continue;
3298                 }
3299
3300                 if (next) {
3301                         btrfs_tree_unlock(next);
3302                         free_extent_buffer(next);
3303                 }
3304
3305                 if (level == 1 && (path->locks[1] || path->skip_locking) &&
3306                     path->reada)
3307                         reada_for_search(root, path, level, slot, 0);
3308
3309                 next = read_node_slot(root, c, slot);
3310                 if (!path->skip_locking) {
3311                         WARN_ON(!btrfs_tree_locked(c));
3312                         btrfs_tree_lock(next);
3313                 }
3314                 break;
3315         }
3316         path->slots[level] = slot;
3317         while(1) {
3318                 level--;
3319                 c = path->nodes[level];
3320                 if (path->locks[level])
3321                         btrfs_tree_unlock(c);
3322                 free_extent_buffer(c);
3323                 path->nodes[level] = next;
3324                 path->slots[level] = 0;
3325                 if (!path->skip_locking)
3326                         path->locks[level] = 1;
3327                 if (!level)
3328                         break;
3329                 if (level == 1 && path->locks[1] && path->reada)
3330                         reada_for_search(root, path, level, slot, 0);
3331                 next = read_node_slot(root, next, 0);
3332                 if (!path->skip_locking) {
3333                         WARN_ON(!btrfs_tree_locked(path->nodes[level]));
3334                         btrfs_tree_lock(next);
3335                 }
3336         }
3337 done:
3338         unlock_up(path, 0, 1);
3339         return 0;
3340 }
3341
3342 /*
3343  * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
3344  * searching until it gets past min_objectid or finds an item of 'type'
3345  *
3346  * returns 0 if something is found, 1 if nothing was found and < 0 on error
3347  */
3348 int btrfs_previous_item(struct btrfs_root *root,
3349                         struct btrfs_path *path, u64 min_objectid,
3350                         int type)
3351 {
3352         struct btrfs_key found_key;
3353         struct extent_buffer *leaf;
3354         int ret;
3355
3356         while(1) {
3357                 if (path->slots[0] == 0) {
3358                         ret = btrfs_prev_leaf(root, path);
3359                         if (ret != 0)
3360                                 return ret;
3361                 } else {
3362                         path->slots[0]--;
3363                 }
3364                 leaf = path->nodes[0];
3365                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3366                 if (found_key.type == type)
3367                         return 0;
3368         }
3369         return 1;
3370 }