SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/btrfs/ctree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/sched.h>
  20 #include "ctree.h"
  21 #include "disk-io.h"
  22 #include "transaction.h"
  23 #include "print-tree.h"
  24
  25 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  26                       *root, struct btrfs_path *path, int level);
  27 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  28                       *root, struct btrfs_key *ins_key,
  29                       struct btrfs_path *path, int data_size, int extend);
  30 static int push_node_left(struct btrfs_trans_handle *trans,
  31                           struct btrfs_root *root, struct extent_buffer *dst,
  32                           struct extent_buffer *src);
  33 static int balance_node_right(struct btrfs_trans_handle *trans,
  34                               struct btrfs_root *root,
  35                               struct extent_buffer *dst_buf,
  36                               struct extent_buffer *src_buf);
  37 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  38                    struct btrfs_path *path, int level, int slot);
  39
  40 inline void btrfs_init_path(struct btrfs_path *p)
  41 {
  42         memset(p, 0, sizeof(*p));
  43 }
  44
  45 struct btrfs_path *btrfs_alloc_path(void)
  46 {
  47         struct btrfs_path *path;
  48         path = kmem_cache_alloc(btrfs_path_cachep, GFP_NOFS);
  49         if (path) {
  50                 btrfs_init_path(path);
  51                 path->reada = 1;
  52         }
  53         return path;
  54 }
  55
  56 void btrfs_free_path(struct btrfs_path *p)
  57 {
  58         btrfs_release_path(NULL, p);
  59         kmem_cache_free(btrfs_path_cachep, p);
  60 }
  61
  62 void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
  63 {
  64         int i;
  65         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  66                 if (!p->nodes[i])
  67                         break;
  68                 free_extent_buffer(p->nodes[i]);
  69         }
  70         memset(p, 0, sizeof(*p));
  71 }
  72
  73 static void add_root_to_dirty_list(struct btrfs_root *root)
  74 {
  75         if (root->track_dirty && list_empty(&root->dirty_list)) {
  76                 list_add(&root->dirty_list,
  77                          &root->fs_info->dirty_cowonly_roots);
  78         }
  79 }
  80
  81 int btrfs_copy_root(struct btrfs_trans_handle *trans,
  82                       struct btrfs_root *root,
  83                       struct extent_buffer *buf,
  84                       struct extent_buffer **cow_ret, u64 new_root_objectid)
  85 {
  86         struct extent_buffer *cow;
  87         u32 nritems;
  88         int ret = 0;
  89         int level;
  90         struct btrfs_key first_key;
  91         struct btrfs_root *new_root;
  92
  93         new_root = kmalloc(sizeof(*new_root), GFP_NOFS);
  94         if (!new_root)
  95                 return -ENOMEM;
  96
  97         memcpy(new_root, root, sizeof(*new_root));
  98         new_root->root_key.objectid = new_root_objectid;
  99
 100         WARN_ON(root->ref_cows && trans->transid !=
 101                 root->fs_info->running_transaction->transid);
 102         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 103
 104         level = btrfs_header_level(buf);
 105         nritems = btrfs_header_nritems(buf);
 106         if (nritems) {
 107                 if (level == 0)
 108                         btrfs_item_key_to_cpu(buf, &first_key, 0);
 109                 else
 110                         btrfs_node_key_to_cpu(buf, &first_key, 0);
 111         } else {
 112                 first_key.objectid = 0;
 113         }
 114         cow = __btrfs_alloc_free_block(trans, new_root, buf->len,
 115                                        new_root_objectid,
 116                                        trans->transid, first_key.objectid,
 117                                        level, buf->start, 0);
 118         if (IS_ERR(cow)) {
 119                 kfree(new_root);
 120                 return PTR_ERR(cow);
 121         }
 122
 123         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 124         btrfs_set_header_bytenr(cow, cow->start);
 125         btrfs_set_header_generation(cow, trans->transid);
 126         btrfs_set_header_owner(cow, new_root_objectid);
 127
 128         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 129         ret = btrfs_inc_ref(trans, new_root, buf);
 130         kfree(new_root);
 131
 132         if (ret)
 133                 return ret;
 134
 135         btrfs_mark_buffer_dirty(cow);
 136         *cow_ret = cow;
 137         return 0;
 138 }
 139
 140 int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 141                              struct btrfs_root *root,
 142                              struct extent_buffer *buf,
 143                              struct extent_buffer *parent, int parent_slot,
 144                              struct extent_buffer **cow_ret,
 145                              u64 search_start, u64 empty_size)
 146 {
 147         u64 root_gen;
 148         struct extent_buffer *cow;
 149         u32 nritems;
 150         int ret = 0;
 151         int different_trans = 0;
 152         int level;
 153         struct btrfs_key first_key;
 154
 155         if (root->ref_cows) {
 156                 root_gen = trans->transid;
 157         } else {
 158                 root_gen = 0;
 159         }
 160
 161         WARN_ON(root->ref_cows && trans->transid !=
 162                 root->fs_info->running_transaction->transid);
 163         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 164
 165         level = btrfs_header_level(buf);
 166         nritems = btrfs_header_nritems(buf);
 167         if (nritems) {
 168                 if (level == 0)
 169                         btrfs_item_key_to_cpu(buf, &first_key, 0);
 170                 else
 171                         btrfs_node_key_to_cpu(buf, &first_key, 0);
 172         } else {
 173                 first_key.objectid = 0;
 174         }
 175         cow = __btrfs_alloc_free_block(trans, root, buf->len,
 176                                      root->root_key.objectid,
 177                                      root_gen, first_key.objectid, level,
 178                                      search_start, empty_size);
 179         if (IS_ERR(cow))
 180                 return PTR_ERR(cow);
 181
 182         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 183         btrfs_set_header_bytenr(cow, cow->start);
 184         btrfs_set_header_generation(cow, trans->transid);
 185         btrfs_set_header_owner(cow, root->root_key.objectid);
 186
 187         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 188         if (btrfs_header_generation(buf) != trans->transid) {
 189                 different_trans = 1;
 190                 ret = btrfs_inc_ref(trans, root, buf);
 191                 if (ret)
 192                         return ret;
 193         } else {
 194                 clean_tree_block(trans, root, buf);
 195         }
 196
 197         if (buf == root->node) {
 198                 root_gen = btrfs_header_generation(buf);
 199                 root->node = cow;
 200                 extent_buffer_get(cow);
 201                 if (buf != root->commit_root) {
 202                         btrfs_free_extent(trans, root, buf->start,
 203                                           buf->len, root->root_key.objectid,
 204                                           root_gen, 0, 0, 1);
 205                 }
 206                 free_extent_buffer(buf);
 207                 add_root_to_dirty_list(root);
 208         } else {
 209                 root_gen = btrfs_header_generation(parent);
 210                 btrfs_set_node_blockptr(parent, parent_slot,
 211                                         cow->start);
 212                 WARN_ON(trans->transid == 0);
 213                 btrfs_set_node_ptr_generation(parent, parent_slot,
 214                                               trans->transid);
 215                 btrfs_mark_buffer_dirty(parent);
 216                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 217                 btrfs_free_extent(trans, root, buf->start, buf->len,
 218                                   btrfs_header_owner(parent), root_gen,
 219                                   0, 0, 1);
 220         }
 221         free_extent_buffer(buf);
 222         btrfs_mark_buffer_dirty(cow);
 223         *cow_ret = cow;
 224         return 0;
 225 }
 226
 227 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 228                     struct btrfs_root *root, struct extent_buffer *buf,
 229                     struct extent_buffer *parent, int parent_slot,
 230                     struct extent_buffer **cow_ret)
 231 {
 232         u64 search_start;
 233         u64 header_trans;
 234         int ret;
 235
 236         if (trans->transaction != root->fs_info->running_transaction) {
 237                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 238                        root->fs_info->running_transaction->transid);
 239                 WARN_ON(1);
 240         }
 241         if (trans->transid != root->fs_info->generation) {
 242                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 243                        root->fs_info->generation);
 244                 WARN_ON(1);
 245         }
 246
 247         header_trans = btrfs_header_generation(buf);
 248         if (header_trans == trans->transid) {
 249                 *cow_ret = buf;
 250                 return 0;
 251         }
 252
 253         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
 254         ret = __btrfs_cow_block(trans, root, buf, parent,
 255                                  parent_slot, cow_ret, search_start, 0);
 256         return ret;
 257 }
 258
 259 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 260 {
 261         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 262                 return 1;
 263         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 264                 return 1;
 265         return 0;
 266 }
 267
 268 /*
 269  * compare two keys in a memcmp fashion
 270  */
 271 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 272 {
 273         struct btrfs_key k1;
 274
 275         btrfs_disk_key_to_cpu(&k1, disk);
 276
 277         if (k1.objectid > k2->objectid)
 278                 return 1;
 279         if (k1.objectid < k2->objectid)
 280                 return -1;
 281         if (k1.type > k2->type)
 282                 return 1;
 283         if (k1.type < k2->type)
 284                 return -1;
 285         if (k1.offset > k2->offset)
 286                 return 1;
 287         if (k1.offset < k2->offset)
 288                 return -1;
 289         return 0;
 290 }
 291
 292
 293 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 294                        struct btrfs_root *root, struct extent_buffer *parent,
 295                        int start_slot, int cache_only, u64 *last_ret,
 296                        struct btrfs_key *progress)
 297 {
 298         struct extent_buffer *cur;
 299         struct extent_buffer *tmp;
 300         u64 blocknr;
 301         u64 search_start = *last_ret;
 302         u64 last_block = 0;
 303         u64 other;
 304         u32 parent_nritems;
 305         int end_slot;
 306         int i;
 307         int err = 0;
 308         int parent_level;
 309         int uptodate;
 310         u32 blocksize;
 311         int progress_passed = 0;
 312         struct btrfs_disk_key disk_key;
 313
 314         parent_level = btrfs_header_level(parent);
 315         if (cache_only && parent_level != 1)
 316                 return 0;
 317
 318         if (trans->transaction != root->fs_info->running_transaction) {
 319                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 320                        root->fs_info->running_transaction->transid);
 321                 WARN_ON(1);
 322         }
 323         if (trans->transid != root->fs_info->generation) {
 324                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 325                        root->fs_info->generation);
 326                 WARN_ON(1);
 327         }
 328
 329         parent_nritems = btrfs_header_nritems(parent);
 330         blocksize = btrfs_level_size(root, parent_level - 1);
 331         end_slot = parent_nritems;
 332
 333         if (parent_nritems == 1)
 334                 return 0;
 335
 336         for (i = start_slot; i < end_slot; i++) {
 337                 int close = 1;
 338
 339                 if (!parent->map_token) {
 340                         map_extent_buffer(parent,
 341                                         btrfs_node_key_ptr_offset(i),
 342                                         sizeof(struct btrfs_key_ptr),
 343                                         &parent->map_token, &parent->kaddr,
 344                                         &parent->map_start, &parent->map_len,
 345                                         KM_USER1);
 346                 }
 347                 btrfs_node_key(parent, &disk_key, i);
 348                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
 349                         continue;
 350
 351                 progress_passed = 1;
 352                 blocknr = btrfs_node_blockptr(parent, i);
 353                 if (last_block == 0)
 354                         last_block = blocknr;
 355
 356                 if (i > 0) {
 357                         other = btrfs_node_blockptr(parent, i - 1);
 358                         close = close_blocks(blocknr, other, blocksize);
 359                 }
 360                 if (close && i < end_slot - 2) {
 361                         other = btrfs_node_blockptr(parent, i + 1);
 362                         close = close_blocks(blocknr, other, blocksize);
 363                 }
 364                 if (close) {
 365                         last_block = blocknr;
 366                         continue;
 367                 }
 368                 if (parent->map_token) {
 369                         unmap_extent_buffer(parent, parent->map_token,
 370                                             KM_USER1);
 371                         parent->map_token = NULL;
 372                 }
 373
 374                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
 375                 if (cur)
 376                         uptodate = btrfs_buffer_uptodate(cur);
 377                 else
 378                         uptodate = 0;
 379                 if (!cur || !uptodate) {
 380                         if (cache_only) {
 381                                 free_extent_buffer(cur);
 382                                 continue;
 383                         }
 384                         if (!cur) {
 385                                 cur = read_tree_block(root, blocknr,
 386                                                          blocksize);
 387                         } else if (!uptodate) {
 388                                 btrfs_read_buffer(cur);
 389                         }
 390                 }
 391                 if (search_start == 0)
 392                         search_start = last_block;
 393
 394                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 395                                         &tmp, search_start,
 396                                         min(16 * blocksize,
 397                                             (end_slot - i) * blocksize));
 398                 if (err) {
 399                         free_extent_buffer(cur);
 400                         break;
 401                 }
 402                 search_start = tmp->start;
 403                 last_block = tmp->start;
 404                 *last_ret = search_start;
 405                 if (parent_level == 1)
 406                         btrfs_clear_buffer_defrag(tmp);
 407                 free_extent_buffer(tmp);
 408         }
 409         if (parent->map_token) {
 410                 unmap_extent_buffer(parent, parent->map_token,
 411                                     KM_USER1);
 412                 parent->map_token = NULL;
 413         }
 414         return err;
 415 }
 416
 417 /*
 418  * The leaf data grows from end-to-front in the node.
 419  * this returns the address of the start of the last item,
 420  * which is the stop of the leaf data stack
 421  */
 422 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 423                                          struct extent_buffer *leaf)
 424 {
 425         u32 nr = btrfs_header_nritems(leaf);
 426         if (nr == 0)
 427                 return BTRFS_LEAF_DATA_SIZE(root);
 428         return btrfs_item_offset_nr(leaf, nr - 1);
 429 }
 430
 431 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
 432                       int level)
 433 {
 434         struct extent_buffer *parent = NULL;
 435         struct extent_buffer *node = path->nodes[level];
 436         struct btrfs_disk_key parent_key;
 437         struct btrfs_disk_key node_key;
 438         int parent_slot;
 439         int slot;
 440         struct btrfs_key cpukey;
 441         u32 nritems = btrfs_header_nritems(node);
 442
 443         if (path->nodes[level + 1])
 444                 parent = path->nodes[level + 1];
 445
 446         slot = path->slots[level];
 447         BUG_ON(nritems == 0);
 448         if (parent) {
 449                 parent_slot = path->slots[level + 1];
 450                 btrfs_node_key(parent, &parent_key, parent_slot);
 451                 btrfs_node_key(node, &node_key, 0);
 452                 BUG_ON(memcmp(&parent_key, &node_key,
 453                               sizeof(struct btrfs_disk_key)));
 454                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 455                        btrfs_header_bytenr(node));
 456         }
 457         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 458         if (slot != 0) {
 459                 btrfs_node_key_to_cpu(node, &cpukey, slot - 1);
 460                 btrfs_node_key(node, &node_key, slot);
 461                 BUG_ON(comp_keys(&node_key, &cpukey) <= 0);
 462         }
 463         if (slot < nritems - 1) {
 464                 btrfs_node_key_to_cpu(node, &cpukey, slot + 1);
 465                 btrfs_node_key(node, &node_key, slot);
 466                 BUG_ON(comp_keys(&node_key, &cpukey) >= 0);
 467         }
 468         return 0;
 469 }
 470
 471 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 472                       int level)
 473 {
 474         struct extent_buffer *leaf = path->nodes[level];
 475         struct extent_buffer *parent = NULL;
 476         int parent_slot;
 477         struct btrfs_key cpukey;
 478         struct btrfs_disk_key parent_key;
 479         struct btrfs_disk_key leaf_key;
 480         int slot = path->slots[0];
 481
 482         u32 nritems = btrfs_header_nritems(leaf);
 483
 484         if (path->nodes[level + 1])
 485                 parent = path->nodes[level + 1];
 486
 487         if (nritems == 0)
 488                 return 0;
 489
 490         if (parent) {
 491                 parent_slot = path->slots[level + 1];
 492                 btrfs_node_key(parent, &parent_key, parent_slot);
 493                 btrfs_item_key(leaf, &leaf_key, 0);
 494
 495                 BUG_ON(memcmp(&parent_key, &leaf_key,
 496                        sizeof(struct btrfs_disk_key)));
 497                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 498                        btrfs_header_bytenr(leaf));
 499         }
 500 #if 0
 501         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 502                 btrfs_item_key_to_cpu(leaf, &cpukey, i + 1);
 503                 btrfs_item_key(leaf, &leaf_key, i);
 504                 if (comp_keys(&leaf_key, &cpukey) >= 0) {
 505                         btrfs_print_leaf(root, leaf);
 506                         printk("slot %d offset bad key\n", i);
 507                         BUG_ON(1);
 508                 }
 509                 if (btrfs_item_offset_nr(leaf, i) !=
 510                         btrfs_item_end_nr(leaf, i + 1)) {
 511                         btrfs_print_leaf(root, leaf);
 512                         printk("slot %d offset bad\n", i);
 513                         BUG_ON(1);
 514                 }
 515                 if (i == 0) {
 516                         if (btrfs_item_offset_nr(leaf, i) +
 517                                btrfs_item_size_nr(leaf, i) !=
 518                                BTRFS_LEAF_DATA_SIZE(root)) {
 519                                 btrfs_print_leaf(root, leaf);
 520                                 printk("slot %d first offset bad\n", i);
 521                                 BUG_ON(1);
 522                         }
 523                 }
 524         }
 525         if (nritems > 0) {
 526                 if (btrfs_item_size_nr(leaf, nritems - 1) > 4096) {
 527                                 btrfs_print_leaf(root, leaf);
 528                                 printk("slot %d bad size \n", nritems - 1);
 529                                 BUG_ON(1);
 530                 }
 531         }
 532 #endif
 533         if (slot != 0 && slot < nritems - 1) {
 534                 btrfs_item_key(leaf, &leaf_key, slot);
 535                 btrfs_item_key_to_cpu(leaf, &cpukey, slot - 1);
 536                 if (comp_keys(&leaf_key, &cpukey) <= 0) {
 537                         btrfs_print_leaf(root, leaf);
 538                         printk("slot %d offset bad key\n", slot);
 539                         BUG_ON(1);
 540                 }
 541                 if (btrfs_item_offset_nr(leaf, slot - 1) !=
 542                        btrfs_item_end_nr(leaf, slot)) {
 543                         btrfs_print_leaf(root, leaf);
 544                         printk("slot %d offset bad\n", slot);
 545                         BUG_ON(1);
 546                 }
 547         }
 548         if (slot < nritems - 1) {
 549                 btrfs_item_key(leaf, &leaf_key, slot);
 550                 btrfs_item_key_to_cpu(leaf, &cpukey, slot + 1);
 551                 BUG_ON(comp_keys(&leaf_key, &cpukey) >= 0);
 552                 if (btrfs_item_offset_nr(leaf, slot) !=
 553                         btrfs_item_end_nr(leaf, slot + 1)) {
 554                         btrfs_print_leaf(root, leaf);
 555                         printk("slot %d offset bad\n", slot);
 556                         BUG_ON(1);
 557                 }
 558         }
 559         BUG_ON(btrfs_item_offset_nr(leaf, 0) +
 560                btrfs_item_size_nr(leaf, 0) != BTRFS_LEAF_DATA_SIZE(root));
 561         return 0;
 562 }
 563
 564 static int noinline check_block(struct btrfs_root *root,
 565                                 struct btrfs_path *path, int level)
 566 {
 567         return 0;
 568 #if 0
 569         struct extent_buffer *buf = path->nodes[level];
 570
 571         if (memcmp_extent_buffer(buf, root->fs_info->fsid,
 572                                  (unsigned long)btrfs_header_fsid(buf),
 573                                  BTRFS_FSID_SIZE)) {
 574                 printk("warning bad block %Lu\n", buf->start);
 575                 return 1;
 576         }
 577 #endif
 578         if (level == 0)
 579                 return check_leaf(root, path, level);
 580         return check_node(root, path, level);
 581 }
 582
 583 /*
 584  * search for key in the extent_buffer.  The items start at offset p,
 585  * and they are item_size apart.  There are 'max' items in p.
 586  *
 587  * the slot in the array is returned via slot, and it points to
 588  * the place where you would insert key if it is not found in
 589  * the array.
 590  *
 591  * slot may point to max if the key is bigger than all of the keys
 592  */
 593 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 594                               int item_size, struct btrfs_key *key,
 595                               int max, int *slot)
 596 {
 597         int low = 0;
 598         int high = max;
 599         int mid;
 600         int ret;
 601         struct btrfs_disk_key *tmp = NULL;
 602         struct btrfs_disk_key unaligned;
 603         unsigned long offset;
 604         char *map_token = NULL;
 605         char *kaddr = NULL;
 606         unsigned long map_start = 0;
 607         unsigned long map_len = 0;
 608         int err;
 609
 610         while(low < high) {
 611                 mid = (low + high) / 2;
 612                 offset = p + mid * item_size;
 613
 614                 if (!map_token || offset < map_start ||
 615                     (offset + sizeof(struct btrfs_disk_key)) >
 616                     map_start + map_len) {
 617                         if (map_token) {
 618                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 619                                 map_token = NULL;
 620                         }
 621                         err = map_extent_buffer(eb, offset,
 622                                                 sizeof(struct btrfs_disk_key),
 623                                                 &map_token, &kaddr,
 624                                                 &map_start, &map_len, KM_USER0);
 625
 626                         if (!err) {
 627                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
 628                                                         map_start);
 629                         } else {
 630                                 read_extent_buffer(eb, &unaligned,
 631                                                    offset, sizeof(unaligned));
 632                                 tmp = &unaligned;
 633                         }
 634
 635                 } else {
 636                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
 637                                                         map_start);
 638                 }
 639                 ret = comp_keys(tmp, key);
 640
 641                 if (ret < 0)
 642                         low = mid + 1;
 643                 else if (ret > 0)
 644                         high = mid;
 645                 else {
 646                         *slot = mid;
 647                         if (map_token)
 648                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 649                         return 0;
 650                 }
 651         }
 652         *slot = low;
 653         if (map_token)
 654                 unmap_extent_buffer(eb, map_token, KM_USER0);
 655         return 1;
 656 }
 657
 658 /*
 659  * simple bin_search frontend that does the right thing for
 660  * leaves vs nodes
 661  */
 662 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 663                       int level, int *slot)
 664 {
 665         if (level == 0) {
 666                 return generic_bin_search(eb,
 667                                           offsetof(struct btrfs_leaf, items),
 668                                           sizeof(struct btrfs_item),
 669                                           key, btrfs_header_nritems(eb),
 670                                           slot);
 671         } else {
 672                 return generic_bin_search(eb,
 673                                           offsetof(struct btrfs_node, ptrs),
 674                                           sizeof(struct btrfs_key_ptr),
 675                                           key, btrfs_header_nritems(eb),
 676                                           slot);
 677         }
 678         return -1;
 679 }
 680
 681 static struct extent_buffer *read_node_slot(struct btrfs_root *root,
 682                                    struct extent_buffer *parent, int slot)
 683 {
 684         if (slot < 0)
 685                 return NULL;
 686         if (slot >= btrfs_header_nritems(parent))
 687                 return NULL;
 688         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 689                        btrfs_level_size(root, btrfs_header_level(parent) - 1));
 690 }
 691
 692 static int balance_level(struct btrfs_trans_handle *trans,
 693                          struct btrfs_root *root,
 694                          struct btrfs_path *path, int level)
 695 {
 696         struct extent_buffer *right = NULL;
 697         struct extent_buffer *mid;
 698         struct extent_buffer *left = NULL;
 699         struct extent_buffer *parent = NULL;
 700         int ret = 0;
 701         int wret;
 702         int pslot;
 703         int orig_slot = path->slots[level];
 704         int err_on_enospc = 0;
 705         u64 orig_ptr;
 706
 707         if (level == 0)
 708                 return 0;
 709
 710         mid = path->nodes[level];
 711         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 712
 713         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 714
 715         if (level < BTRFS_MAX_LEVEL - 1)
 716                 parent = path->nodes[level + 1];
 717         pslot = path->slots[level + 1];
 718
 719         /*
 720          * deal with the case where there is only one pointer in the root
 721          * by promoting the node below to a root
 722          */
 723         if (!parent) {
 724                 struct extent_buffer *child;
 725
 726                 if (btrfs_header_nritems(mid) != 1)
 727                         return 0;
 728
 729                 /* promote the child to a root */
 730                 child = read_node_slot(root, mid, 0);
 731                 BUG_ON(!child);
 732                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
 733                 BUG_ON(ret);
 734
 735                 root->node = child;
 736                 add_root_to_dirty_list(root);
 737                 path->nodes[level] = NULL;
 738                 clean_tree_block(trans, root, mid);
 739                 wait_on_tree_block_writeback(root, mid);
 740                 /* once for the path */
 741                 free_extent_buffer(mid);
 742                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 743                                         root->root_key.objectid,
 744                                         btrfs_header_generation(mid), 0, 0, 1);
 745                 /* once for the root ptr */
 746                 free_extent_buffer(mid);
 747                 return ret;
 748         }
 749         if (btrfs_header_nritems(mid) >
 750             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 751                 return 0;
 752
 753         if (btrfs_header_nritems(mid) < 2)
 754                 err_on_enospc = 1;
 755
 756         left = read_node_slot(root, parent, pslot - 1);
 757         if (left) {
 758                 wret = btrfs_cow_block(trans, root, left,
 759                                        parent, pslot - 1, &left);
 760                 if (wret) {
 761                         ret = wret;
 762                         goto enospc;
 763                 }
 764         }
 765         right = read_node_slot(root, parent, pslot + 1);
 766         if (right) {
 767                 wret = btrfs_cow_block(trans, root, right,
 768                                        parent, pslot + 1, &right);
 769                 if (wret) {
 770                         ret = wret;
 771                         goto enospc;
 772                 }
 773         }
 774
 775         /* first, try to make some room in the middle buffer */
 776         if (left) {
 777                 orig_slot += btrfs_header_nritems(left);
 778                 wret = push_node_left(trans, root, left, mid);
 779                 if (wret < 0)
 780                         ret = wret;
 781                 if (btrfs_header_nritems(mid) < 2)
 782                         err_on_enospc = 1;
 783         }
 784
 785         /*
 786          * then try to empty the right most buffer into the middle
 787          */
 788         if (right) {
 789                 wret = push_node_left(trans, root, mid, right);
 790                 if (wret < 0 && wret != -ENOSPC)
 791                         ret = wret;
 792                 if (btrfs_header_nritems(right) == 0) {
 793                         u64 bytenr = right->start;
 794                         u64 generation = btrfs_header_generation(parent);
 795                         u32 blocksize = right->len;
 796
 797                         clean_tree_block(trans, root, right);
 798                         wait_on_tree_block_writeback(root, right);
 799                         free_extent_buffer(right);
 800                         right = NULL;
 801                         wret = del_ptr(trans, root, path, level + 1, pslot +
 802                                        1);
 803                         if (wret)
 804                                 ret = wret;
 805                         wret = btrfs_free_extent(trans, root, bytenr,
 806                                                  blocksize,
 807                                                  btrfs_header_owner(parent),
 808                                                  generation, 0, 0, 1);
 809                         if (wret)
 810                                 ret = wret;
 811                 } else {
 812                         struct btrfs_disk_key right_key;
 813                         btrfs_node_key(right, &right_key, 0);
 814                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 815                         btrfs_mark_buffer_dirty(parent);
 816                 }
 817         }
 818         if (btrfs_header_nritems(mid) == 1) {
 819                 /*
 820                  * we're not allowed to leave a node with one item in the
 821                  * tree during a delete.  A deletion from lower in the tree
 822                  * could try to delete the only pointer in this node.
 823                  * So, pull some keys from the left.
 824                  * There has to be a left pointer at this point because
 825                  * otherwise we would have pulled some pointers from the
 826                  * right
 827                  */
 828                 BUG_ON(!left);
 829                 wret = balance_node_right(trans, root, mid, left);
 830                 if (wret < 0) {
 831                         ret = wret;
 832                         goto enospc;
 833                 }
 834                 BUG_ON(wret == 1);
 835         }
 836         if (btrfs_header_nritems(mid) == 0) {
 837                 /* we've managed to empty the middle node, drop it */
 838                 u64 root_gen = btrfs_header_generation(parent);
 839                 u64 bytenr = mid->start;
 840                 u32 blocksize = mid->len;
 841                 clean_tree_block(trans, root, mid);
 842                 wait_on_tree_block_writeback(root, mid);
 843                 free_extent_buffer(mid);
 844                 mid = NULL;
 845                 wret = del_ptr(trans, root, path, level + 1, pslot);
 846                 if (wret)
 847                         ret = wret;
 848                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 849                                          btrfs_header_owner(parent),
 850                                          root_gen, 0, 0, 1);
 851                 if (wret)
 852                         ret = wret;
 853         } else {
 854                 /* update the parent key to reflect our changes */
 855                 struct btrfs_disk_key mid_key;
 856                 btrfs_node_key(mid, &mid_key, 0);
 857                 btrfs_set_node_key(parent, &mid_key, pslot);
 858                 btrfs_mark_buffer_dirty(parent);
 859         }
 860
 861         /* update the path */
 862         if (left) {
 863                 if (btrfs_header_nritems(left) > orig_slot) {
 864                         extent_buffer_get(left);
 865                         path->nodes[level] = left;
 866                         path->slots[level + 1] -= 1;
 867                         path->slots[level] = orig_slot;
 868                         if (mid)
 869                                 free_extent_buffer(mid);
 870                 } else {
 871                         orig_slot -= btrfs_header_nritems(left);
 872                         path->slots[level] = orig_slot;
 873                 }
 874         }
 875         /* double check we haven't messed things up */
 876         check_block(root, path, level);
 877         if (orig_ptr !=
 878             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 879                 BUG();
 880 enospc:
 881         if (right)
 882                 free_extent_buffer(right);
 883         if (left)
 884                 free_extent_buffer(left);
 885         return ret;
 886 }
 887
 888 /* returns zero if the push worked, non-zero otherwise */
 889 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
 890                                           struct btrfs_root *root,
 891                                           struct btrfs_path *path, int level)
 892 {
 893         struct extent_buffer *right = NULL;
 894         struct extent_buffer *mid;
 895         struct extent_buffer *left = NULL;
 896         struct extent_buffer *parent = NULL;
 897         int ret = 0;
 898         int wret;
 899         int pslot;
 900         int orig_slot = path->slots[level];
 901         u64 orig_ptr;
 902
 903         if (level == 0)
 904                 return 1;
 905
 906         mid = path->nodes[level];
 907         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 908         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 909
 910         if (level < BTRFS_MAX_LEVEL - 1)
 911                 parent = path->nodes[level + 1];
 912         pslot = path->slots[level + 1];
 913
 914         if (!parent)
 915                 return 1;
 916
 917         left = read_node_slot(root, parent, pslot - 1);
 918
 919         /* first, try to make some room in the middle buffer */
 920         if (left) {
 921                 u32 left_nr;
 922                 left_nr = btrfs_header_nritems(left);
 923                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 924                         wret = 1;
 925                 } else {
 926                         ret = btrfs_cow_block(trans, root, left, parent,
 927                                               pslot - 1, &left);
 928                         if (ret)
 929                                 wret = 1;
 930                         else {
 931                                 wret = push_node_left(trans, root,
 932                                                       left, mid);
 933                         }
 934                 }
 935                 if (wret < 0)
 936                         ret = wret;
 937                 if (wret == 0) {
 938                         struct btrfs_disk_key disk_key;
 939                         orig_slot += left_nr;
 940                         btrfs_node_key(mid, &disk_key, 0);
 941                         btrfs_set_node_key(parent, &disk_key, pslot);
 942                         btrfs_mark_buffer_dirty(parent);
 943                         if (btrfs_header_nritems(left) > orig_slot) {
 944                                 path->nodes[level] = left;
 945                                 path->slots[level + 1] -= 1;
 946                                 path->slots[level] = orig_slot;
 947                                 free_extent_buffer(mid);
 948                         } else {
 949                                 orig_slot -=
 950                                         btrfs_header_nritems(left);
 951                                 path->slots[level] = orig_slot;
 952                                 free_extent_buffer(left);
 953                         }
 954                         return 0;
 955                 }
 956                 free_extent_buffer(left);
 957         }
 958         right= read_node_slot(root, parent, pslot + 1);
 959
 960         /*
 961          * then try to empty the right most buffer into the middle
 962          */
 963         if (right) {
 964                 u32 right_nr;
 965                 right_nr = btrfs_header_nritems(right);
 966                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 967                         wret = 1;
 968                 } else {
 969                         ret = btrfs_cow_block(trans, root, right,
 970                                               parent, pslot + 1,
 971                                               &right);
 972                         if (ret)
 973                                 wret = 1;
 974                         else {
 975                                 wret = balance_node_right(trans, root,
 976                                                           right, mid);
 977                         }
 978                 }
 979                 if (wret < 0)
 980                         ret = wret;
 981                 if (wret == 0) {
 982                         struct btrfs_disk_key disk_key;
 983
 984                         btrfs_node_key(right, &disk_key, 0);
 985                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
 986                         btrfs_mark_buffer_dirty(parent);
 987
 988                         if (btrfs_header_nritems(mid) <= orig_slot) {
 989                                 path->nodes[level] = right;
 990                                 path->slots[level + 1] += 1;
 991                                 path->slots[level] = orig_slot -
 992                                         btrfs_header_nritems(mid);
 993                                 free_extent_buffer(mid);
 994                         } else {
 995                                 free_extent_buffer(right);
 996                         }
 997                         return 0;
 998                 }
 999                 free_extent_buffer(right);
1000         }
1001         return 1;
1002 }
1003
1004 /*
1005  * readahead one full node of leaves
1006  */
1007 static void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
1008                              int level, int slot, u64 objectid)
1009 {
1010         struct extent_buffer *node;
1011         struct btrfs_disk_key disk_key;
1012         u32 nritems;
1013         u64 search;
1014         u64 lowest_read;
1015         u64 highest_read;
1016         u64 nread = 0;
1017         int direction = path->reada;
1018         struct extent_buffer *eb;
1019         u32 nr;
1020         u32 blocksize;
1021         u32 nscan = 0;
1022
1023         if (level != 1)
1024                 return;
1025
1026         if (!path->nodes[level])
1027                 return;
1028
1029         node = path->nodes[level];
1030         search = btrfs_node_blockptr(node, slot);
1031         blocksize = btrfs_level_size(root, level - 1);
1032         eb = btrfs_find_tree_block(root, search, blocksize);
1033         if (eb) {
1034                 free_extent_buffer(eb);
1035                 return;
1036         }
1037
1038         highest_read = search;
1039         lowest_read = search;
1040
1041         nritems = btrfs_header_nritems(node);
1042         nr = slot;
1043         while(1) {
1044                 if (direction < 0) {
1045                         if (nr == 0)
1046                                 break;
1047                         nr--;
1048                 } else if (direction > 0) {
1049                         nr++;
1050                         if (nr >= nritems)
1051                                 break;
1052                 }
1053                 if (path->reada < 0 && objectid) {
1054                         btrfs_node_key(node, &disk_key, nr);
1055                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1056                                 break;
1057                 }
1058                 search = btrfs_node_blockptr(node, nr);
1059                 if ((search >= lowest_read && search <= highest_read) ||
1060                     (search < lowest_read && lowest_read - search <= 32768) ||
1061                     (search > highest_read && search - highest_read <= 32768)) {
1062                         readahead_tree_block(root, search, blocksize);
1063                         nread += blocksize;
1064                 }
1065                 nscan++;
1066                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
1067                         break;
1068                 if(nread > (1024 * 1024) || nscan > 128)
1069                         break;
1070
1071                 if (search < lowest_read)
1072                         lowest_read = search;
1073                 if (search > highest_read)
1074                         highest_read = search;
1075         }
1076 }
1077 /*
1078  * look for key in the tree.  path is filled in with nodes along the way
1079  * if key is found, we return zero and you can find the item in the leaf
1080  * level of the path (level 0)
1081  *
1082  * If the key isn't found, the path points to the slot where it should
1083  * be inserted, and 1 is returned.  If there are other errors during the
1084  * search a negative error number is returned.
1085  *
1086  * if ins_len > 0, nodes and leaves will be split as we walk down the
1087  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1088  * possible)
1089  */
1090 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1091                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1092                       ins_len, int cow)
1093 {
1094         struct extent_buffer *b;
1095         u64 bytenr;
1096         u64 ptr_gen;
1097         int slot;
1098         int ret;
1099         int level;
1100         int should_reada = p->reada;
1101         u8 lowest_level = 0;
1102
1103         lowest_level = p->lowest_level;
1104         WARN_ON(lowest_level && ins_len);
1105         WARN_ON(p->nodes[0] != NULL);
1106         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
1107 again:
1108         b = root->node;
1109         extent_buffer_get(b);
1110         while (b) {
1111                 level = btrfs_header_level(b);
1112                 if (cow) {
1113                         int wret;
1114                         wret = btrfs_cow_block(trans, root, b,
1115                                                p->nodes[level + 1],
1116                                                p->slots[level + 1],
1117                                                &b);
1118                         if (wret) {
1119                                 free_extent_buffer(b);
1120                                 return wret;
1121                         }
1122                 }
1123                 BUG_ON(!cow && ins_len);
1124                 if (level != btrfs_header_level(b))
1125                         WARN_ON(1);
1126                 level = btrfs_header_level(b);
1127                 p->nodes[level] = b;
1128                 ret = check_block(root, p, level);
1129                 if (ret)
1130                         return -1;
1131                 ret = bin_search(b, key, level, &slot);
1132                 if (level != 0) {
1133                         if (ret && slot > 0)
1134                                 slot -= 1;
1135                         p->slots[level] = slot;
1136                         if (ins_len > 0 && btrfs_header_nritems(b) >=
1137                             BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1138                                 int sret = split_node(trans, root, p, level);
1139                                 BUG_ON(sret > 0);
1140                                 if (sret)
1141                                         return sret;
1142                                 b = p->nodes[level];
1143                                 slot = p->slots[level];
1144                         } else if (ins_len < 0) {
1145                                 int sret = balance_level(trans, root, p,
1146                                                          level);
1147                                 if (sret)
1148                                         return sret;
1149                                 b = p->nodes[level];
1150                                 if (!b) {
1151                                         btrfs_release_path(NULL, p);
1152                                         goto again;
1153                                 }
1154                                 slot = p->slots[level];
1155                                 BUG_ON(btrfs_header_nritems(b) == 1);
1156                         }
1157                         /* this is only true while dropping a snapshot */
1158                         if (level == lowest_level)
1159                                 break;
1160                         bytenr = btrfs_node_blockptr(b, slot);
1161                         ptr_gen = btrfs_node_ptr_generation(b, slot);
1162                         if (should_reada)
1163                                 reada_for_search(root, p, level, slot,
1164                                                  key->objectid);
1165                         b = read_tree_block(root, bytenr,
1166                                             btrfs_level_size(root, level - 1));
1167                         if (ptr_gen != btrfs_header_generation(b)) {
1168                                 printk("block %llu bad gen wanted %llu "
1169                                        "found %llu\n",
1170                                 (unsigned long long)b->start,
1171                                 (unsigned long long)ptr_gen,
1172                                 (unsigned long long)btrfs_header_generation(b));
1173                         }
1174                 } else {
1175                         p->slots[level] = slot;
1176                         if (ins_len > 0 && btrfs_leaf_free_space(root, b) <
1177                             sizeof(struct btrfs_item) + ins_len) {
1178                                 int sret = split_leaf(trans, root, key,
1179                                                       p, ins_len, ret == 0);
1180                                 BUG_ON(sret > 0);
1181                                 if (sret)
1182                                         return sret;
1183                         }
1184                         return ret;
1185                 }
1186         }
1187         return 1;
1188 }
1189
1190 /*
1191  * adjust the pointers going up the tree, starting at level
1192  * making sure the right key of each node is points to 'key'.
1193  * This is used after shifting pointers to the left, so it stops
1194  * fixing up pointers when a given leaf/node is not in slot 0 of the
1195  * higher levels
1196  *
1197  * If this fails to write a tree block, it returns -1, but continues
1198  * fixing up the blocks in ram so the tree is consistent.
1199  */
1200 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1201                           struct btrfs_root *root, struct btrfs_path *path,
1202                           struct btrfs_disk_key *key, int level)
1203 {
1204         int i;
1205         int ret = 0;
1206         struct extent_buffer *t;
1207
1208         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1209                 int tslot = path->slots[i];
1210                 if (!path->nodes[i])
1211                         break;
1212                 t = path->nodes[i];
1213                 btrfs_set_node_key(t, key, tslot);
1214                 btrfs_mark_buffer_dirty(path->nodes[i]);
1215                 if (tslot != 0)
1216                         break;
1217         }
1218         return ret;
1219 }
1220
1221 /*
1222  * try to push data from one node into the next node left in the
1223  * tree.
1224  *
1225  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1226  * error, and > 0 if there was no room in the left hand block.
1227  */
1228 static int push_node_left(struct btrfs_trans_handle *trans,
1229                           struct btrfs_root *root, struct extent_buffer *dst,
1230                           struct extent_buffer *src)
1231 {
1232         int push_items = 0;
1233         int src_nritems;
1234         int dst_nritems;
1235         int ret = 0;
1236
1237         src_nritems = btrfs_header_nritems(src);
1238         dst_nritems = btrfs_header_nritems(dst);
1239         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1240         WARN_ON(btrfs_header_generation(src) != trans->transid);
1241         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1242
1243         if (push_items <= 0) {
1244                 return 1;
1245         }
1246
1247         if (src_nritems < push_items)
1248                 push_items = src_nritems;
1249
1250         copy_extent_buffer(dst, src,
1251                            btrfs_node_key_ptr_offset(dst_nritems),
1252                            btrfs_node_key_ptr_offset(0),
1253                            push_items * sizeof(struct btrfs_key_ptr));
1254
1255         if (push_items < src_nritems) {
1256                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1257                                       btrfs_node_key_ptr_offset(push_items),
1258                                       (src_nritems - push_items) *
1259                                       sizeof(struct btrfs_key_ptr));
1260         }
1261         btrfs_set_header_nritems(src, src_nritems - push_items);
1262         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1263         btrfs_mark_buffer_dirty(src);
1264         btrfs_mark_buffer_dirty(dst);
1265         return ret;
1266 }
1267
1268 /*
1269  * try to push data from one node into the next node right in the
1270  * tree.
1271  *
1272  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1273  * error, and > 0 if there was no room in the right hand block.
1274  *
1275  * this will  only push up to 1/2 the contents of the left node over
1276  */
1277 static int balance_node_right(struct btrfs_trans_handle *trans,
1278                               struct btrfs_root *root,
1279                               struct extent_buffer *dst,
1280                               struct extent_buffer *src)
1281 {
1282         int push_items = 0;
1283         int max_push;
1284         int src_nritems;
1285         int dst_nritems;
1286         int ret = 0;
1287
1288         WARN_ON(btrfs_header_generation(src) != trans->transid);
1289         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1290
1291         src_nritems = btrfs_header_nritems(src);
1292         dst_nritems = btrfs_header_nritems(dst);
1293         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1294         if (push_items <= 0)
1295                 return 1;
1296
1297         max_push = src_nritems / 2 + 1;
1298         /* don't try to empty the node */
1299         if (max_push >= src_nritems)
1300                 return 1;
1301
1302         if (max_push < push_items)
1303                 push_items = max_push;
1304
1305         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1306                                       btrfs_node_key_ptr_offset(0),
1307                                       (dst_nritems) *
1308                                       sizeof(struct btrfs_key_ptr));
1309
1310         copy_extent_buffer(dst, src,
1311                            btrfs_node_key_ptr_offset(0),
1312                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1313                            push_items * sizeof(struct btrfs_key_ptr));
1314
1315         btrfs_set_header_nritems(src, src_nritems - push_items);
1316         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1317
1318         btrfs_mark_buffer_dirty(src);
1319         btrfs_mark_buffer_dirty(dst);
1320         return ret;
1321 }
1322
1323 /*
1324  * helper function to insert a new root level in the tree.
1325  * A new node is allocated, and a single item is inserted to
1326  * point to the existing root
1327  *
1328  * returns zero on success or < 0 on failure.
1329  */
1330 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1331                            struct btrfs_root *root,
1332                            struct btrfs_path *path, int level)
1333 {
1334         u64 root_gen;
1335         u64 lower_gen;
1336         struct extent_buffer *lower;
1337         struct extent_buffer *c;
1338         struct btrfs_disk_key lower_key;
1339
1340         BUG_ON(path->nodes[level]);
1341         BUG_ON(path->nodes[level-1] != root->node);
1342
1343         if (root->ref_cows)
1344                 root_gen = trans->transid;
1345         else
1346                 root_gen = 0;
1347
1348         lower = path->nodes[level-1];
1349         if (level == 1)
1350                 btrfs_item_key(lower, &lower_key, 0);
1351         else
1352                 btrfs_node_key(lower, &lower_key, 0);
1353
1354         c = __btrfs_alloc_free_block(trans, root, root->nodesize,
1355                                    root->root_key.objectid,
1356                                    root_gen, lower_key.objectid, level,
1357                                    root->node->start, 0);
1358         if (IS_ERR(c))
1359                 return PTR_ERR(c);
1360         memset_extent_buffer(c, 0, 0, root->nodesize);
1361         btrfs_set_header_nritems(c, 1);
1362         btrfs_set_header_level(c, level);
1363         btrfs_set_header_bytenr(c, c->start);
1364         btrfs_set_header_generation(c, trans->transid);
1365         btrfs_set_header_owner(c, root->root_key.objectid);
1366
1367         write_extent_buffer(c, root->fs_info->fsid,
1368                             (unsigned long)btrfs_header_fsid(c),
1369                             BTRFS_FSID_SIZE);
1370         btrfs_set_node_key(c, &lower_key, 0);
1371         btrfs_set_node_blockptr(c, 0, lower->start);
1372         lower_gen = btrfs_header_generation(lower);
1373         WARN_ON(lower_gen == 0);
1374
1375         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1376
1377         btrfs_mark_buffer_dirty(c);
1378
1379         /* the super has an extra ref to root->node */
1380         free_extent_buffer(root->node);
1381         root->node = c;
1382         add_root_to_dirty_list(root);
1383         extent_buffer_get(c);
1384         path->nodes[level] = c;
1385         path->slots[level] = 0;
1386
1387         if (root->ref_cows && lower_gen != trans->transid) {
1388                 struct btrfs_path *back_path = btrfs_alloc_path();
1389                 int ret;
1390                 ret = btrfs_insert_extent_backref(trans,
1391                                                   root->fs_info->extent_root,
1392                                                   path, lower->start,
1393                                                   root->root_key.objectid,
1394                                                   trans->transid, 0, 0);
1395                 BUG_ON(ret);
1396                 btrfs_free_path(back_path);
1397         }
1398         return 0;
1399 }
1400
1401 /*
1402  * worker function to insert a single pointer in a node.
1403  * the node should have enough room for the pointer already
1404  *
1405  * slot and level indicate where you want the key to go, and
1406  * blocknr is the block the key points to.
1407  *
1408  * returns zero on success and < 0 on any error
1409  */
1410 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1411                       *root, struct btrfs_path *path, struct btrfs_disk_key
1412                       *key, u64 bytenr, int slot, int level)
1413 {
1414         struct extent_buffer *lower;
1415         int nritems;
1416
1417         BUG_ON(!path->nodes[level]);
1418         lower = path->nodes[level];
1419         nritems = btrfs_header_nritems(lower);
1420         if (slot > nritems)
1421                 BUG();
1422         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1423                 BUG();
1424         if (slot != nritems) {
1425                 memmove_extent_buffer(lower,
1426                               btrfs_node_key_ptr_offset(slot + 1),
1427                               btrfs_node_key_ptr_offset(slot),
1428                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1429         }
1430         btrfs_set_node_key(lower, key, slot);
1431         btrfs_set_node_blockptr(lower, slot, bytenr);
1432         WARN_ON(trans->transid == 0);
1433         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1434         btrfs_set_header_nritems(lower, nritems + 1);
1435         btrfs_mark_buffer_dirty(lower);
1436         return 0;
1437 }
1438
1439 /*
1440  * split the node at the specified level in path in two.
1441  * The path is corrected to point to the appropriate node after the split
1442  *
1443  * Before splitting this tries to make some room in the node by pushing
1444  * left and right, if either one works, it returns right away.
1445  *
1446  * returns 0 on success and < 0 on failure
1447  */
1448 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1449                       *root, struct btrfs_path *path, int level)
1450 {
1451         u64 root_gen;
1452         struct extent_buffer *c;
1453         struct extent_buffer *split;
1454         struct btrfs_disk_key disk_key;
1455         int mid;
1456         int ret;
1457         int wret;
1458         u32 c_nritems;
1459
1460         c = path->nodes[level];
1461         WARN_ON(btrfs_header_generation(c) != trans->transid);
1462         if (c == root->node) {
1463                 /* trying to split the root, lets make a new one */
1464                 ret = insert_new_root(trans, root, path, level + 1);
1465                 if (ret)
1466                         return ret;
1467         } else {
1468                 ret = push_nodes_for_insert(trans, root, path, level);
1469                 c = path->nodes[level];
1470                 if (!ret && btrfs_header_nritems(c) <
1471                     BTRFS_NODEPTRS_PER_BLOCK(root) - 1)
1472                         return 0;
1473                 if (ret < 0)
1474                         return ret;
1475         }
1476
1477         c_nritems = btrfs_header_nritems(c);
1478         if (root->ref_cows)
1479                 root_gen = trans->transid;
1480         else
1481                 root_gen = 0;
1482
1483         btrfs_node_key(c, &disk_key, 0);
1484         split = __btrfs_alloc_free_block(trans, root, root->nodesize,
1485                                          root->root_key.objectid,
1486                                          root_gen,
1487                                          btrfs_disk_key_objectid(&disk_key),
1488                                          level, c->start, 0);
1489         if (IS_ERR(split))
1490                 return PTR_ERR(split);
1491
1492         btrfs_set_header_flags(split, btrfs_header_flags(c));
1493         btrfs_set_header_level(split, btrfs_header_level(c));
1494         btrfs_set_header_bytenr(split, split->start);
1495         btrfs_set_header_generation(split, trans->transid);
1496         btrfs_set_header_owner(split, root->root_key.objectid);
1497         write_extent_buffer(split, root->fs_info->fsid,
1498                             (unsigned long)btrfs_header_fsid(split),
1499                             BTRFS_FSID_SIZE);
1500
1501         mid = (c_nritems + 1) / 2;
1502
1503         copy_extent_buffer(split, c,
1504                            btrfs_node_key_ptr_offset(0),
1505                            btrfs_node_key_ptr_offset(mid),
1506                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1507         btrfs_set_header_nritems(split, c_nritems - mid);
1508         btrfs_set_header_nritems(c, mid);
1509         ret = 0;
1510
1511         btrfs_mark_buffer_dirty(c);
1512         btrfs_mark_buffer_dirty(split);
1513
1514         btrfs_node_key(split, &disk_key, 0);
1515         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1516                           path->slots[level + 1] + 1,
1517                           level + 1);
1518         if (wret)
1519                 ret = wret;
1520
1521         if (path->slots[level] >= mid) {
1522                 path->slots[level] -= mid;
1523                 free_extent_buffer(c);
1524                 path->nodes[level] = split;
1525                 path->slots[level + 1] += 1;
1526         } else {
1527                 free_extent_buffer(split);
1528         }
1529         return ret;
1530 }
1531
1532 /*
1533  * how many bytes are required to store the items in a leaf.  start
1534  * and nr indicate which items in the leaf to check.  This totals up the
1535  * space used both by the item structs and the item data
1536  */
1537 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1538 {
1539         int data_len;
1540         int nritems = btrfs_header_nritems(l);
1541         int end = min(nritems, start + nr) - 1;
1542
1543         if (!nr)
1544                 return 0;
1545         data_len = btrfs_item_end_nr(l, start);
1546         data_len = data_len - btrfs_item_offset_nr(l, end);
1547         data_len += sizeof(struct btrfs_item) * nr;
1548         WARN_ON(data_len < 0);
1549         return data_len;
1550 }
1551
1552 /*
1553  * The space between the end of the leaf items and
1554  * the start of the leaf data.  IOW, how much room
1555  * the leaf has left for both items and data
1556  */
1557 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1558 {
1559         int nritems = btrfs_header_nritems(leaf);
1560         int ret;
1561         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1562         if (ret < 0) {
1563                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1564                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1565                        leaf_space_used(leaf, 0, nritems), nritems);
1566         }
1567         return ret;
1568 }
1569
1570 /*
1571  * push some data in the path leaf to the right, trying to free up at
1572  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1573  *
1574  * returns 1 if the push failed because the other node didn't have enough
1575  * room, 0 if everything worked out and < 0 if there were major errors.
1576  */
1577 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1578                            *root, struct btrfs_path *path, int data_size,
1579                            int empty)
1580 {
1581         struct extent_buffer *left = path->nodes[0];
1582         struct extent_buffer *right;
1583         struct extent_buffer *upper;
1584         struct btrfs_disk_key disk_key;
1585         int slot;
1586         u32 i;
1587         int free_space;
1588         int push_space = 0;
1589         int push_items = 0;
1590         struct btrfs_item *item;
1591         u32 left_nritems;
1592         u32 nr;
1593         u32 right_nritems;
1594         u32 data_end;
1595         u32 this_item_size;
1596         int ret;
1597
1598         slot = path->slots[1];
1599         if (!path->nodes[1]) {
1600                 return 1;
1601         }
1602         upper = path->nodes[1];
1603         if (slot >= btrfs_header_nritems(upper) - 1)
1604                 return 1;
1605
1606         right = read_tree_block(root, btrfs_node_blockptr(upper, slot + 1),
1607                                 root->leafsize);
1608         free_space = btrfs_leaf_free_space(root, right);
1609         if (free_space < data_size + sizeof(struct btrfs_item)) {
1610                 free_extent_buffer(right);
1611                 return 1;
1612         }
1613
1614         /* cow and double check */
1615         ret = btrfs_cow_block(trans, root, right, upper,
1616                               slot + 1, &right);
1617         if (ret) {
1618                 free_extent_buffer(right);
1619                 return 1;
1620         }
1621         free_space = btrfs_leaf_free_space(root, right);
1622         if (free_space < data_size + sizeof(struct btrfs_item)) {
1623                 free_extent_buffer(right);
1624                 return 1;
1625         }
1626
1627         left_nritems = btrfs_header_nritems(left);
1628         if (left_nritems == 0) {
1629                 free_extent_buffer(right);
1630                 return 1;
1631         }
1632
1633         if (empty)
1634                 nr = 0;
1635         else
1636                 nr = 1;
1637
1638         i = left_nritems - 1;
1639         while (i >= nr) {
1640                 item = btrfs_item_nr(left, i);
1641
1642                 if (path->slots[0] == i)
1643                         push_space += data_size + sizeof(*item);
1644
1645                 if (!left->map_token) {
1646                         map_extent_buffer(left, (unsigned long)item,
1647                                         sizeof(struct btrfs_item),
1648                                         &left->map_token, &left->kaddr,
1649                                         &left->map_start, &left->map_len,
1650                                         KM_USER1);
1651                 }
1652
1653                 this_item_size = btrfs_item_size(left, item);
1654                 if (this_item_size + sizeof(*item) + push_space > free_space)
1655                         break;
1656                 push_items++;
1657                 push_space += this_item_size + sizeof(*item);
1658                 if (i == 0)
1659                         break;
1660                 i--;
1661         }
1662         if (left->map_token) {
1663                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1664                 left->map_token = NULL;
1665         }
1666
1667         if (push_items == 0) {
1668                 free_extent_buffer(right);
1669                 return 1;
1670         }
1671
1672         if (!empty && push_items == left_nritems)
1673                 WARN_ON(1);
1674
1675         /* push left to right */
1676         right_nritems = btrfs_header_nritems(right);
1677
1678         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1679         push_space -= leaf_data_end(root, left);
1680
1681         /* make room in the right data area */
1682         data_end = leaf_data_end(root, right);
1683         memmove_extent_buffer(right,
1684                               btrfs_leaf_data(right) + data_end - push_space,
1685                               btrfs_leaf_data(right) + data_end,
1686                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1687
1688         /* copy from the left data area */
1689         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1690                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1691                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1692                      push_space);
1693
1694         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1695                               btrfs_item_nr_offset(0),
1696                               right_nritems * sizeof(struct btrfs_item));
1697
1698         /* copy the items from left to right */
1699         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1700                    btrfs_item_nr_offset(left_nritems - push_items),
1701                    push_items * sizeof(struct btrfs_item));
1702
1703         /* update the item pointers */
1704         right_nritems += push_items;
1705         btrfs_set_header_nritems(right, right_nritems);
1706         push_space = BTRFS_LEAF_DATA_SIZE(root);
1707         for (i = 0; i < right_nritems; i++) {
1708                 item = btrfs_item_nr(right, i);
1709                 if (!right->map_token) {
1710                         map_extent_buffer(right, (unsigned long)item,
1711                                         sizeof(struct btrfs_item),
1712                                         &right->map_token, &right->kaddr,
1713                                         &right->map_start, &right->map_len,
1714                                         KM_USER1);
1715                 }
1716                 push_space -= btrfs_item_size(right, item);
1717                 btrfs_set_item_offset(right, item, push_space);
1718         }
1719
1720         if (right->map_token) {
1721                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1722                 right->map_token = NULL;
1723         }
1724         left_nritems -= push_items;
1725         btrfs_set_header_nritems(left, left_nritems);
1726
1727         if (left_nritems)
1728                 btrfs_mark_buffer_dirty(left);
1729         btrfs_mark_buffer_dirty(right);
1730
1731         btrfs_item_key(right, &disk_key, 0);
1732         btrfs_set_node_key(upper, &disk_key, slot + 1);
1733         btrfs_mark_buffer_dirty(upper);
1734
1735         /* then fixup the leaf pointer in the path */
1736         if (path->slots[0] >= left_nritems) {
1737                 path->slots[0] -= left_nritems;
1738                 free_extent_buffer(path->nodes[0]);
1739                 path->nodes[0] = right;
1740                 path->slots[1] += 1;
1741         } else {
1742                 free_extent_buffer(right);
1743         }
1744         return 0;
1745 }
1746 /*
1747  * push some data in the path leaf to the left, trying to free up at
1748  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1749  */
1750 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1751                           *root, struct btrfs_path *path, int data_size,
1752                           int empty)
1753 {
1754         struct btrfs_disk_key disk_key;
1755         struct extent_buffer *right = path->nodes[0];
1756         struct extent_buffer *left;
1757         int slot;
1758         int i;
1759         int free_space;
1760         int push_space = 0;
1761         int push_items = 0;
1762         struct btrfs_item *item;
1763         u32 old_left_nritems;
1764         u32 right_nritems;
1765         u32 nr;
1766         int ret = 0;
1767         int wret;
1768         u32 this_item_size;
1769         u32 old_left_item_size;
1770
1771         slot = path->slots[1];
1772         if (slot == 0)
1773                 return 1;
1774         if (!path->nodes[1])
1775                 return 1;
1776
1777         right_nritems = btrfs_header_nritems(right);
1778         if (right_nritems == 0) {
1779                 return 1;
1780         }
1781
1782         left = read_tree_block(root, btrfs_node_blockptr(path->nodes[1],
1783                                slot - 1), root->leafsize);
1784         free_space = btrfs_leaf_free_space(root, left);
1785         if (free_space < data_size + sizeof(struct btrfs_item)) {
1786                 free_extent_buffer(left);
1787                 return 1;
1788         }
1789
1790         /* cow and double check */
1791         ret = btrfs_cow_block(trans, root, left,
1792                               path->nodes[1], slot - 1, &left);
1793         if (ret) {
1794                 /* we hit -ENOSPC, but it isn't fatal here */
1795                 free_extent_buffer(left);
1796                 return 1;
1797         }
1798
1799         free_space = btrfs_leaf_free_space(root, left);
1800         if (free_space < data_size + sizeof(struct btrfs_item)) {
1801                 free_extent_buffer(left);
1802                 return 1;
1803         }
1804
1805         if (empty)
1806                 nr = right_nritems;
1807         else
1808                 nr = right_nritems - 1;
1809
1810         for (i = 0; i < nr; i++) {
1811                 item = btrfs_item_nr(right, i);
1812                 if (!right->map_token) {
1813                         map_extent_buffer(right, (unsigned long)item,
1814                                         sizeof(struct btrfs_item),
1815                                         &right->map_token, &right->kaddr,
1816                                         &right->map_start, &right->map_len,
1817                                         KM_USER1);
1818                 }
1819
1820                 if (path->slots[0] == i)
1821                         push_space += data_size + sizeof(*item);
1822
1823                 this_item_size = btrfs_item_size(right, item);
1824                 if (this_item_size + sizeof(*item) + push_space > free_space)
1825                         break;
1826
1827                 push_items++;
1828                 push_space += this_item_size + sizeof(*item);
1829         }
1830
1831         if (right->map_token) {
1832                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1833                 right->map_token = NULL;
1834         }
1835
1836         if (push_items == 0) {
1837                 free_extent_buffer(left);
1838                 return 1;
1839         }
1840         if (!empty && push_items == btrfs_header_nritems(right))
1841                 WARN_ON(1);
1842
1843         /* push data from right to left */
1844         copy_extent_buffer(left, right,
1845                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1846                            btrfs_item_nr_offset(0),
1847                            push_items * sizeof(struct btrfs_item));
1848
1849         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1850                      btrfs_item_offset_nr(right, push_items -1);
1851
1852         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1853                      leaf_data_end(root, left) - push_space,
1854                      btrfs_leaf_data(right) +
1855                      btrfs_item_offset_nr(right, push_items - 1),
1856                      push_space);
1857         old_left_nritems = btrfs_header_nritems(left);
1858         BUG_ON(old_left_nritems < 0);
1859
1860         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1861         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1862                 u32 ioff;
1863
1864                 item = btrfs_item_nr(left, i);
1865                 if (!left->map_token) {
1866                         map_extent_buffer(left, (unsigned long)item,
1867                                         sizeof(struct btrfs_item),
1868                                         &left->map_token, &left->kaddr,
1869                                         &left->map_start, &left->map_len,
1870                                         KM_USER1);
1871                 }
1872
1873                 ioff = btrfs_item_offset(left, item);
1874                 btrfs_set_item_offset(left, item,
1875                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1876         }
1877         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1878         if (left->map_token) {
1879                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1880                 left->map_token = NULL;
1881         }
1882
1883         /* fixup right node */
1884         if (push_items > right_nritems) {
1885                 printk("push items %d nr %u\n", push_items, right_nritems);
1886                 WARN_ON(1);
1887         }
1888
1889         if (push_items < right_nritems) {
1890                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1891                                                   leaf_data_end(root, right);
1892                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
1893                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1894                                       btrfs_leaf_data(right) +
1895                                       leaf_data_end(root, right), push_space);
1896
1897                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1898                               btrfs_item_nr_offset(push_items),
1899                              (btrfs_header_nritems(right) - push_items) *
1900                              sizeof(struct btrfs_item));
1901         }
1902         right_nritems -= push_items;
1903         btrfs_set_header_nritems(right, right_nritems);
1904         push_space = BTRFS_LEAF_DATA_SIZE(root);
1905         for (i = 0; i < right_nritems; i++) {
1906                 item = btrfs_item_nr(right, i);
1907
1908                 if (!right->map_token) {
1909                         map_extent_buffer(right, (unsigned long)item,
1910                                         sizeof(struct btrfs_item),
1911                                         &right->map_token, &right->kaddr,
1912                                         &right->map_start, &right->map_len,
1913                                         KM_USER1);
1914                 }
1915
1916                 push_space = push_space - btrfs_item_size(right, item);
1917                 btrfs_set_item_offset(right, item, push_space);
1918         }
1919         if (right->map_token) {
1920                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1921                 right->map_token = NULL;
1922         }
1923
1924         btrfs_mark_buffer_dirty(left);
1925         if (right_nritems)
1926                 btrfs_mark_buffer_dirty(right);
1927
1928         btrfs_item_key(right, &disk_key, 0);
1929         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1930         if (wret)
1931                 ret = wret;
1932
1933         /* then fixup the leaf pointer in the path */
1934         if (path->slots[0] < push_items) {
1935                 path->slots[0] += old_left_nritems;
1936                 free_extent_buffer(path->nodes[0]);
1937                 path->nodes[0] = left;
1938                 path->slots[1] -= 1;
1939         } else {
1940                 free_extent_buffer(left);
1941                 path->slots[0] -= push_items;
1942         }
1943         BUG_ON(path->slots[0] < 0);
1944         return ret;
1945 }
1946
1947 /*
1948  * split the path's leaf in two, making sure there is at least data_size
1949  * available for the resulting leaf level of the path.
1950  *
1951  * returns 0 if all went well and < 0 on failure.
1952  */
1953 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
1954                       *root, struct btrfs_key *ins_key,
1955                       struct btrfs_path *path, int data_size, int extend)
1956 {
1957         u64 root_gen;
1958         struct extent_buffer *l;
1959         u32 nritems;
1960         int mid;
1961         int slot;
1962         struct extent_buffer *right;
1963         int space_needed = data_size + sizeof(struct btrfs_item);
1964         int data_copy_size;
1965         int rt_data_off;
1966         int i;
1967         int ret = 0;
1968         int wret;
1969         int double_split;
1970         int num_doubles = 0;
1971         struct btrfs_disk_key disk_key;
1972
1973         if (extend)
1974                 space_needed = data_size;
1975
1976         if (root->ref_cows)
1977                 root_gen = trans->transid;
1978         else
1979                 root_gen = 0;
1980
1981         /* first try to make some room by pushing left and right */
1982         if (ins_key->type != BTRFS_DIR_ITEM_KEY) {
1983                 wret = push_leaf_right(trans, root, path, data_size, 0);
1984                 if (wret < 0) {
1985                         return wret;
1986                 }
1987                 if (wret) {
1988                         wret = push_leaf_left(trans, root, path, data_size, 0);
1989                         if (wret < 0)
1990                                 return wret;
1991                 }
1992                 l = path->nodes[0];
1993
1994                 /* did the pushes work? */
1995                 if (btrfs_leaf_free_space(root, l) >= space_needed)
1996                         return 0;
1997         }
1998
1999         if (!path->nodes[1]) {
2000                 ret = insert_new_root(trans, root, path, 1);
2001                 if (ret)
2002                         return ret;
2003         }
2004 again:
2005         double_split = 0;
2006         l = path->nodes[0];
2007         slot = path->slots[0];
2008         nritems = btrfs_header_nritems(l);
2009         mid = (nritems + 1)/ 2;
2010
2011         btrfs_item_key(l, &disk_key, 0);
2012
2013         right = __btrfs_alloc_free_block(trans, root, root->leafsize,
2014                                          root->root_key.objectid,
2015                                          root_gen, disk_key.objectid, 0,
2016                                          l->start, 0);
2017         if (IS_ERR(right))
2018                 return PTR_ERR(right);
2019
2020         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2021         btrfs_set_header_bytenr(right, right->start);
2022         btrfs_set_header_generation(right, trans->transid);
2023         btrfs_set_header_owner(right, root->root_key.objectid);
2024         btrfs_set_header_level(right, 0);
2025         write_extent_buffer(right, root->fs_info->fsid,
2026                             (unsigned long)btrfs_header_fsid(right),
2027                             BTRFS_FSID_SIZE);
2028         if (mid <= slot) {
2029                 if (nritems == 1 ||
2030                     leaf_space_used(l, mid, nritems - mid) + space_needed >
2031                         BTRFS_LEAF_DATA_SIZE(root)) {
2032                         if (slot >= nritems) {
2033                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2034                                 btrfs_set_header_nritems(right, 0);
2035                                 wret = insert_ptr(trans, root, path,
2036                                                   &disk_key, right->start,
2037                                                   path->slots[1] + 1, 1);
2038                                 if (wret)
2039                                         ret = wret;
2040                                 free_extent_buffer(path->nodes[0]);
2041                                 path->nodes[0] = right;
2042                                 path->slots[0] = 0;
2043                                 path->slots[1] += 1;
2044                                 return ret;
2045                         }
2046                         mid = slot;
2047                         if (mid != nritems &&
2048                             leaf_space_used(l, mid, nritems - mid) +
2049                             space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2050                                 double_split = 1;
2051                         }
2052                 }
2053         } else {
2054                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
2055                         BTRFS_LEAF_DATA_SIZE(root)) {
2056                         if (!extend && slot == 0) {
2057                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2058                                 btrfs_set_header_nritems(right, 0);
2059                                 wret = insert_ptr(trans, root, path,
2060                                                   &disk_key,
2061                                                   right->start,
2062                                                   path->slots[1], 1);
2063                                 if (wret)
2064                                         ret = wret;
2065                                 free_extent_buffer(path->nodes[0]);
2066                                 path->nodes[0] = right;
2067                                 path->slots[0] = 0;
2068                                 if (path->slots[1] == 0) {
2069                                         wret = fixup_low_keys(trans, root,
2070                                                    path, &disk_key, 1);
2071                                         if (wret)
2072                                                 ret = wret;
2073                                 }
2074                                 return ret;
2075                         } else if (extend && slot == 0) {
2076                                 mid = 1;
2077                         } else {
2078                                 mid = slot;
2079                                 if (mid != nritems &&
2080                                     leaf_space_used(l, mid, nritems - mid) +
2081                                     space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2082                                         double_split = 1;
2083                                 }
2084                         }
2085                 }
2086         }
2087         nritems = nritems - mid;
2088         btrfs_set_header_nritems(right, nritems);
2089         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2090
2091         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2092                            btrfs_item_nr_offset(mid),
2093                            nritems * sizeof(struct btrfs_item));
2094
2095         copy_extent_buffer(right, l,
2096                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2097                      data_copy_size, btrfs_leaf_data(l) +
2098                      leaf_data_end(root, l), data_copy_size);
2099
2100         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2101                       btrfs_item_end_nr(l, mid);
2102
2103         for (i = 0; i < nritems; i++) {
2104                 struct btrfs_item *item = btrfs_item_nr(right, i);
2105                 u32 ioff;
2106
2107                 if (!right->map_token) {
2108                         map_extent_buffer(right, (unsigned long)item,
2109                                         sizeof(struct btrfs_item),
2110                                         &right->map_token, &right->kaddr,
2111                                         &right->map_start, &right->map_len,
2112                                         KM_USER1);
2113                 }
2114
2115                 ioff = btrfs_item_offset(right, item);
2116                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2117         }
2118
2119         if (right->map_token) {
2120                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2121                 right->map_token = NULL;
2122         }
2123
2124         btrfs_set_header_nritems(l, mid);
2125         ret = 0;
2126         btrfs_item_key(right, &disk_key, 0);
2127         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2128                           path->slots[1] + 1, 1);
2129         if (wret)
2130                 ret = wret;
2131
2132         btrfs_mark_buffer_dirty(right);
2133         btrfs_mark_buffer_dirty(l);
2134         BUG_ON(path->slots[0] != slot);
2135
2136         if (mid <= slot) {
2137                 free_extent_buffer(path->nodes[0]);
2138                 path->nodes[0] = right;
2139                 path->slots[0] -= mid;
2140                 path->slots[1] += 1;
2141         } else
2142                 free_extent_buffer(right);
2143
2144         BUG_ON(path->slots[0] < 0);
2145
2146         if (double_split) {
2147                 BUG_ON(num_doubles != 0);
2148                 num_doubles++;
2149                 goto again;
2150         }
2151         return ret;
2152 }
2153
2154 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2155                         struct btrfs_root *root,
2156                         struct btrfs_path *path,
2157                         u32 new_size, int from_end)
2158 {
2159         int ret = 0;
2160         int slot;
2161         int slot_orig;
2162         struct extent_buffer *leaf;
2163         struct btrfs_item *item;
2164         u32 nritems;
2165         unsigned int data_end;
2166         unsigned int old_data_start;
2167         unsigned int old_size;
2168         unsigned int size_diff;
2169         int i;
2170
2171         slot_orig = path->slots[0];
2172         leaf = path->nodes[0];
2173         slot = path->slots[0];
2174
2175         old_size = btrfs_item_size_nr(leaf, slot);
2176         if (old_size == new_size)
2177                 return 0;
2178
2179         nritems = btrfs_header_nritems(leaf);
2180         data_end = leaf_data_end(root, leaf);
2181
2182         old_data_start = btrfs_item_offset_nr(leaf, slot);
2183
2184         size_diff = old_size - new_size;
2185
2186         BUG_ON(slot < 0);
2187         BUG_ON(slot >= nritems);
2188
2189         /*
2190          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2191          */
2192         /* first correct the data pointers */
2193         for (i = slot; i < nritems; i++) {
2194                 u32 ioff;
2195                 item = btrfs_item_nr(leaf, i);
2196
2197                 if (!leaf->map_token) {
2198                         map_extent_buffer(leaf, (unsigned long)item,
2199                                         sizeof(struct btrfs_item),
2200                                         &leaf->map_token, &leaf->kaddr,
2201                                         &leaf->map_start, &leaf->map_len,
2202                                         KM_USER1);
2203                 }
2204
2205                 ioff = btrfs_item_offset(leaf, item);
2206                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2207         }
2208
2209         if (leaf->map_token) {
2210                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2211                 leaf->map_token = NULL;
2212         }
2213
2214         /* shift the data */
2215         if (from_end) {
2216                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2217                               data_end + size_diff, btrfs_leaf_data(leaf) +
2218                               data_end, old_data_start + new_size - data_end);
2219         } else {
2220                 struct btrfs_disk_key disk_key;
2221                 u64 offset;
2222
2223                 btrfs_item_key(leaf, &disk_key, slot);
2224
2225                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2226                         unsigned long ptr;
2227                         struct btrfs_file_extent_item *fi;
2228
2229                         fi = btrfs_item_ptr(leaf, slot,
2230                                             struct btrfs_file_extent_item);
2231                         fi = (struct btrfs_file_extent_item *)(
2232                              (unsigned long)fi - size_diff);
2233
2234                         if (btrfs_file_extent_type(leaf, fi) ==
2235                             BTRFS_FILE_EXTENT_INLINE) {
2236                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2237                                 memmove_extent_buffer(leaf, ptr,
2238                                         (unsigned long)fi,
2239                                         offsetof(struct btrfs_file_extent_item,
2240                                                  disk_bytenr));
2241                         }
2242                 }
2243
2244                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2245                               data_end + size_diff, btrfs_leaf_data(leaf) +
2246                               data_end, old_data_start - data_end);
2247
2248                 offset = btrfs_disk_key_offset(&disk_key);
2249                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2250                 btrfs_set_item_key(leaf, &disk_key, slot);
2251                 if (slot == 0)
2252                         fixup_low_keys(trans, root, path, &disk_key, 1);
2253         }
2254
2255         item = btrfs_item_nr(leaf, slot);
2256         btrfs_set_item_size(leaf, item, new_size);
2257         btrfs_mark_buffer_dirty(leaf);
2258
2259         ret = 0;
2260         if (btrfs_leaf_free_space(root, leaf) < 0) {
2261                 btrfs_print_leaf(root, leaf);
2262                 BUG();
2263         }
2264         return ret;
2265 }
2266
2267 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2268                       struct btrfs_root *root, struct btrfs_path *path,
2269                       u32 data_size)
2270 {
2271         int ret = 0;
2272         int slot;
2273         int slot_orig;
2274         struct extent_buffer *leaf;
2275         struct btrfs_item *item;
2276         u32 nritems;
2277         unsigned int data_end;
2278         unsigned int old_data;
2279         unsigned int old_size;
2280         int i;
2281
2282         slot_orig = path->slots[0];
2283         leaf = path->nodes[0];
2284
2285         nritems = btrfs_header_nritems(leaf);
2286         data_end = leaf_data_end(root, leaf);
2287
2288         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2289                 btrfs_print_leaf(root, leaf);
2290                 BUG();
2291         }
2292         slot = path->slots[0];
2293         old_data = btrfs_item_end_nr(leaf, slot);
2294
2295         BUG_ON(slot < 0);
2296         if (slot >= nritems) {
2297                 btrfs_print_leaf(root, leaf);
2298                 printk("slot %d too large, nritems %d\n", slot, nritems);
2299                 BUG_ON(1);
2300         }
2301
2302         /*
2303          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2304          */
2305         /* first correct the data pointers */
2306         for (i = slot; i < nritems; i++) {
2307                 u32 ioff;
2308                 item = btrfs_item_nr(leaf, i);
2309
2310                 if (!leaf->map_token) {
2311                         map_extent_buffer(leaf, (unsigned long)item,
2312                                         sizeof(struct btrfs_item),
2313                                         &leaf->map_token, &leaf->kaddr,
2314                                         &leaf->map_start, &leaf->map_len,
2315                                         KM_USER1);
2316                 }
2317                 ioff = btrfs_item_offset(leaf, item);
2318                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2319         }
2320
2321         if (leaf->map_token) {
2322                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2323                 leaf->map_token = NULL;
2324         }
2325
2326         /* shift the data */
2327         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2328                       data_end - data_size, btrfs_leaf_data(leaf) +
2329                       data_end, old_data - data_end);
2330
2331         data_end = old_data;
2332         old_size = btrfs_item_size_nr(leaf, slot);
2333         item = btrfs_item_nr(leaf, slot);
2334         btrfs_set_item_size(leaf, item, old_size + data_size);
2335         btrfs_mark_buffer_dirty(leaf);
2336
2337         ret = 0;
2338         if (btrfs_leaf_free_space(root, leaf) < 0) {
2339                 btrfs_print_leaf(root, leaf);
2340                 BUG();
2341         }
2342         return ret;
2343 }
2344
2345 /*
2346  * Given a key and some data, insert an item into the tree.
2347  * This does all the path init required, making room in the tree if needed.
2348  */
2349 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2350                             struct btrfs_root *root,
2351                             struct btrfs_path *path,
2352                             struct btrfs_key *cpu_key, u32 *data_size,
2353                             int nr)
2354 {
2355         struct extent_buffer *leaf;
2356         struct btrfs_item *item;
2357         int ret = 0;
2358         int slot;
2359         int slot_orig;
2360         int i;
2361         u32 nritems;
2362         u32 total_size = 0;
2363         u32 total_data = 0;
2364         unsigned int data_end;
2365         struct btrfs_disk_key disk_key;
2366
2367         for (i = 0; i < nr; i++) {
2368                 total_data += data_size[i];
2369         }
2370
2371         /* create a root if there isn't one */
2372         if (!root->node)
2373                 BUG();
2374
2375         total_size = total_data + (nr - 1) * sizeof(struct btrfs_item);
2376         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2377         if (ret == 0) {
2378                 return -EEXIST;
2379         }
2380         if (ret < 0)
2381                 goto out;
2382
2383         slot_orig = path->slots[0];
2384         leaf = path->nodes[0];
2385
2386         nritems = btrfs_header_nritems(leaf);
2387         data_end = leaf_data_end(root, leaf);
2388
2389         if (btrfs_leaf_free_space(root, leaf) <
2390             sizeof(struct btrfs_item) + total_size) {
2391                 btrfs_print_leaf(root, leaf);
2392                 printk("not enough freespace need %u have %d\n",
2393                        total_size, btrfs_leaf_free_space(root, leaf));
2394                 BUG();
2395         }
2396
2397         slot = path->slots[0];
2398         BUG_ON(slot < 0);
2399
2400         if (slot != nritems) {
2401                 int i;
2402                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2403
2404                 if (old_data < data_end) {
2405                         btrfs_print_leaf(root, leaf);
2406                         printk("slot %d old_data %d data_end %d\n",
2407                                slot, old_data, data_end);
2408                         BUG_ON(1);
2409                 }
2410                 /*
2411                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2412                  */
2413                 /* first correct the data pointers */
2414                 WARN_ON(leaf->map_token);
2415                 for (i = slot; i < nritems; i++) {
2416                         u32 ioff;
2417
2418                         item = btrfs_item_nr(leaf, i);
2419                         if (!leaf->map_token) {
2420                                 map_extent_buffer(leaf, (unsigned long)item,
2421                                         sizeof(struct btrfs_item),
2422                                         &leaf->map_token, &leaf->kaddr,
2423                                         &leaf->map_start, &leaf->map_len,
2424                                         KM_USER1);
2425                         }
2426
2427                         ioff = btrfs_item_offset(leaf, item);
2428                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2429                 }
2430                 if (leaf->map_token) {
2431                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2432                         leaf->map_token = NULL;
2433                 }
2434
2435                 /* shift the items */
2436                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2437                               btrfs_item_nr_offset(slot),
2438                               (nritems - slot) * sizeof(struct btrfs_item));
2439
2440                 /* shift the data */
2441                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2442                               data_end - total_data, btrfs_leaf_data(leaf) +
2443                               data_end, old_data - data_end);
2444                 data_end = old_data;
2445         }
2446
2447         /* setup the item for the new data */
2448         for (i = 0; i < nr; i++) {
2449                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2450                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2451                 item = btrfs_item_nr(leaf, slot + i);
2452                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2453                 data_end -= data_size[i];
2454                 btrfs_set_item_size(leaf, item, data_size[i]);
2455         }
2456         btrfs_set_header_nritems(leaf, nritems + nr);
2457         btrfs_mark_buffer_dirty(leaf);
2458
2459         ret = 0;
2460         if (slot == 0) {
2461                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2462                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2463         }
2464
2465         if (btrfs_leaf_free_space(root, leaf) < 0) {
2466                 btrfs_print_leaf(root, leaf);
2467                 BUG();
2468         }
2469
2470 out:
2471         return ret;
2472 }
2473
2474 /*
2475  * Given a key and some data, insert an item into the tree.
2476  * This does all the path init required, making room in the tree if needed.
2477  */
2478 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2479                       *root, struct btrfs_key *cpu_key, void *data, u32
2480                       data_size)
2481 {
2482         int ret = 0;
2483         struct btrfs_path *path;
2484         struct extent_buffer *leaf;
2485         unsigned long ptr;
2486
2487         path = btrfs_alloc_path();
2488         BUG_ON(!path);
2489         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2490         if (!ret) {
2491                 leaf = path->nodes[0];
2492                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2493                 write_extent_buffer(leaf, data, ptr, data_size);
2494                 btrfs_mark_buffer_dirty(leaf);
2495         }
2496         btrfs_free_path(path);
2497         return ret;
2498 }
2499
2500 /*
2501  * delete the pointer from a given node.
2502  *
2503  * If the delete empties a node, the node is removed from the tree,
2504  * continuing all the way the root if required.  The root is converted into
2505  * a leaf if all the nodes are emptied.
2506  */
2507 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2508                    struct btrfs_path *path, int level, int slot)
2509 {
2510         struct extent_buffer *parent = path->nodes[level];
2511         u32 nritems;
2512         int ret = 0;
2513         int wret;
2514
2515         nritems = btrfs_header_nritems(parent);
2516         if (slot != nritems -1) {
2517                 memmove_extent_buffer(parent,
2518                               btrfs_node_key_ptr_offset(slot),
2519                               btrfs_node_key_ptr_offset(slot + 1),
2520                               sizeof(struct btrfs_key_ptr) *
2521                               (nritems - slot - 1));
2522         }
2523         nritems--;
2524         btrfs_set_header_nritems(parent, nritems);
2525         if (nritems == 0 && parent == root->node) {
2526                 BUG_ON(btrfs_header_level(root->node) != 1);
2527                 /* just turn the root into a leaf and break */
2528                 btrfs_set_header_level(root->node, 0);
2529         } else if (slot == 0) {
2530                 struct btrfs_disk_key disk_key;
2531
2532                 btrfs_node_key(parent, &disk_key, 0);
2533                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2534                 if (wret)
2535                         ret = wret;
2536         }
2537         btrfs_mark_buffer_dirty(parent);
2538         return ret;
2539 }
2540
2541 /*
2542  * delete the item at the leaf level in path.  If that empties
2543  * the leaf, remove it from the tree
2544  */
2545 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2546                     struct btrfs_path *path, int slot, int nr)
2547 {
2548         struct extent_buffer *leaf;
2549         struct btrfs_item *item;
2550         int last_off;
2551         int dsize = 0;
2552         int ret = 0;
2553         int wret;
2554         int i;
2555         u32 nritems;
2556
2557         leaf = path->nodes[0];
2558         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2559
2560         for (i = 0; i < nr; i++)
2561                 dsize += btrfs_item_size_nr(leaf, slot + i);
2562
2563         nritems = btrfs_header_nritems(leaf);
2564
2565         if (slot + nr != nritems) {
2566                 int i;
2567                 int data_end = leaf_data_end(root, leaf);
2568
2569                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2570                               data_end + dsize,
2571                               btrfs_leaf_data(leaf) + data_end,
2572                               last_off - data_end);
2573
2574                 for (i = slot + nr; i < nritems; i++) {
2575                         u32 ioff;
2576
2577                         item = btrfs_item_nr(leaf, i);
2578                         if (!leaf->map_token) {
2579                                 map_extent_buffer(leaf, (unsigned long)item,
2580                                         sizeof(struct btrfs_item),
2581                                         &leaf->map_token, &leaf->kaddr,
2582                                         &leaf->map_start, &leaf->map_len,
2583                                         KM_USER1);
2584                         }
2585                         ioff = btrfs_item_offset(leaf, item);
2586                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2587                 }
2588
2589                 if (leaf->map_token) {
2590                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2591                         leaf->map_token = NULL;
2592                 }
2593
2594                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2595                               btrfs_item_nr_offset(slot + nr),
2596                               sizeof(struct btrfs_item) *
2597                               (nritems - slot - nr));
2598         }
2599         btrfs_set_header_nritems(leaf, nritems - nr);
2600         nritems -= nr;
2601
2602         /* delete the leaf if we've emptied it */
2603         if (nritems == 0) {
2604                 if (leaf == root->node) {
2605                         btrfs_set_header_level(leaf, 0);
2606                 } else {
2607                         u64 root_gen = btrfs_header_generation(path->nodes[1]);
2608                         clean_tree_block(trans, root, leaf);
2609                         wait_on_tree_block_writeback(root, leaf);
2610                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
2611                         if (wret)
2612                                 ret = wret;
2613                         wret = btrfs_free_extent(trans, root,
2614                                          leaf->start, leaf->len,
2615                                          btrfs_header_owner(path->nodes[1]),
2616                                          root_gen, 0, 0, 1);
2617                         if (wret)
2618                                 ret = wret;
2619                 }
2620         } else {
2621                 int used = leaf_space_used(leaf, 0, nritems);
2622                 if (slot == 0) {
2623                         struct btrfs_disk_key disk_key;
2624
2625                         btrfs_item_key(leaf, &disk_key, 0);
2626                         wret = fixup_low_keys(trans, root, path,
2627                                               &disk_key, 1);
2628                         if (wret)
2629                                 ret = wret;
2630                 }
2631
2632                 /* delete the leaf if it is mostly empty */
2633                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2634                         /* push_leaf_left fixes the path.
2635                          * make sure the path still points to our leaf
2636                          * for possible call to del_ptr below
2637                          */
2638                         slot = path->slots[1];
2639                         extent_buffer_get(leaf);
2640
2641                         wret = push_leaf_left(trans, root, path, 1, 1);
2642                         if (wret < 0 && wret != -ENOSPC)
2643                                 ret = wret;
2644
2645                         if (path->nodes[0] == leaf &&
2646                             btrfs_header_nritems(leaf)) {
2647                                 wret = push_leaf_right(trans, root, path, 1, 1);
2648                                 if (wret < 0 && wret != -ENOSPC)
2649                                         ret = wret;
2650                         }
2651
2652                         if (btrfs_header_nritems(leaf) == 0) {
2653                                 u64 root_gen;
2654                                 u64 bytenr = leaf->start;
2655                                 u32 blocksize = leaf->len;
2656
2657                                 root_gen = btrfs_header_generation(
2658                                                            path->nodes[1]);
2659
2660                                 clean_tree_block(trans, root, leaf);
2661                                 wait_on_tree_block_writeback(root, leaf);
2662
2663                                 wret = del_ptr(trans, root, path, 1, slot);
2664                                 if (wret)
2665                                         ret = wret;
2666
2667                                 free_extent_buffer(leaf);
2668                                 wret = btrfs_free_extent(trans, root, bytenr,
2669                                              blocksize,
2670                                              btrfs_header_owner(path->nodes[1]),
2671                                              root_gen, 0, 0, 1);
2672                                 if (wret)
2673                                         ret = wret;
2674                         } else {
2675                                 btrfs_mark_buffer_dirty(leaf);
2676                                 free_extent_buffer(leaf);
2677                         }
2678                 } else {
2679                         btrfs_mark_buffer_dirty(leaf);
2680                 }
2681         }
2682         return ret;
2683 }
2684
2685 /*
2686  * walk up the tree as far as required to find the previous leaf.
2687  * returns 0 if it found something or 1 if there are no lesser leaves.
2688  * returns < 0 on io errors.
2689  */
2690 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2691 {
2692         u64 bytenr;
2693         int slot;
2694         int level = 1;
2695         struct extent_buffer *c;
2696         struct extent_buffer *next = NULL;
2697
2698         while(level < BTRFS_MAX_LEVEL) {
2699                 if (!path->nodes[level])
2700                         return 1;
2701
2702                 slot = path->slots[level];
2703                 c = path->nodes[level];
2704                 if (slot == 0) {
2705                         level++;
2706                         if (level == BTRFS_MAX_LEVEL)
2707                                 return 1;
2708                         continue;
2709                 }
2710                 slot--;
2711
2712                 bytenr = btrfs_node_blockptr(c, slot);
2713                 if (next)
2714                         free_extent_buffer(next);
2715
2716                 next = read_tree_block(root, bytenr,
2717                                        btrfs_level_size(root, level - 1));
2718                 break;
2719         }
2720         path->slots[level] = slot;
2721         while(1) {
2722                 level--;
2723                 c = path->nodes[level];
2724                 free_extent_buffer(c);
2725                 slot = btrfs_header_nritems(next);
2726                 if (slot != 0)
2727                         slot--;
2728                 path->nodes[level] = next;
2729                 path->slots[level] = slot;
2730                 if (!level)
2731                         break;
2732                 next = read_tree_block(root, btrfs_node_blockptr(next, slot),
2733                                        btrfs_level_size(root, level - 1));
2734         }
2735         return 0;
2736 }
2737
2738 /*
2739  * walk up the tree as far as required to find the next leaf.
2740  * returns 0 if it found something or 1 if there are no greater leaves.
2741  * returns < 0 on io errors.
2742  */
2743 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2744 {
2745         int slot;
2746         int level = 1;
2747         u64 bytenr;
2748         struct extent_buffer *c;
2749         struct extent_buffer *next = NULL;
2750
2751         while(level < BTRFS_MAX_LEVEL) {
2752                 if (!path->nodes[level])
2753                         return 1;
2754
2755                 slot = path->slots[level] + 1;
2756                 c = path->nodes[level];
2757                 if (slot >= btrfs_header_nritems(c)) {
2758                         level++;
2759                         if (level == BTRFS_MAX_LEVEL)
2760                                 return 1;
2761                         continue;
2762                 }
2763
2764                 bytenr = btrfs_node_blockptr(c, slot);
2765                 if (next)
2766                         free_extent_buffer(next);
2767
2768                 if (path->reada)
2769                         reada_for_search(root, path, level, slot, 0);
2770
2771                 next = read_tree_block(root, bytenr,
2772                                        btrfs_level_size(root, level -1));
2773                 break;
2774         }
2775         path->slots[level] = slot;
2776         while(1) {
2777                 level--;
2778                 c = path->nodes[level];
2779                 free_extent_buffer(c);
2780                 path->nodes[level] = next;
2781                 path->slots[level] = 0;
2782                 if (!level)
2783                         break;
2784                 if (path->reada)
2785                         reada_for_search(root, path, level, 0, 0);
2786                 next = read_tree_block(root, btrfs_node_blockptr(next, 0),
2787                                        btrfs_level_size(root, level - 1));
2788         }
2789         return 0;
2790 }
2791
2792 int btrfs_previous_item(struct btrfs_root *root,
2793                         struct btrfs_path *path, u64 min_objectid,
2794                         int type)
2795 {
2796         struct btrfs_key found_key;
2797         struct extent_buffer *leaf;
2798         int ret;
2799
2800         while(1) {
2801                 if (path->slots[0] == 0) {
2802                         ret = btrfs_prev_leaf(root, path);
2803                         if (ret != 0)
2804                                 return ret;
2805                 } else {
2806                         path->slots[0]--;
2807                 }
2808                 leaf = path->nodes[0];
2809                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2810                 if (found_key.type == type)
2811                         return 0;
2812         }
2813         return 1;
2814 }
2815