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 #include "ctree.h"
  19 #include "disk-io.h"
  20 #include "transaction.h"
  21 #include "print-tree.h"
  22 #include "repair.h"
  23
  24 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  25                       *root, struct btrfs_path *path, int level);
  26 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  27                       *root, struct btrfs_key *ins_key,
  28                       struct btrfs_path *path, int data_size, int extend);
  29 static int push_node_left(struct btrfs_trans_handle *trans,
  30                           struct btrfs_root *root, struct extent_buffer *dst,
  31                           struct extent_buffer *src, int empty);
  32 static int balance_node_right(struct btrfs_trans_handle *trans,
  33                               struct btrfs_root *root,
  34                               struct extent_buffer *dst_buf,
  35                               struct extent_buffer *src_buf);
  36
  37 inline void btrfs_init_path(struct btrfs_path *p)
  38 {
  39         memset(p, 0, sizeof(*p));
  40 }
  41
  42 struct btrfs_path *btrfs_alloc_path(void)
  43 {
  44         struct btrfs_path *path;
  45         path = kmalloc(sizeof(struct btrfs_path), GFP_NOFS);
  46         if (path) {
  47                 btrfs_init_path(path);
  48                 path->reada = 0;
  49         }
  50         return path;
  51 }
  52
  53 void btrfs_free_path(struct btrfs_path *p)
  54 {
  55         btrfs_release_path(NULL, p);
  56         kfree(p);
  57 }
  58
  59 void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
  60 {
  61         int i;
  62         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  63                 if (!p->nodes[i])
  64                         continue;
  65                 free_extent_buffer(p->nodes[i]);
  66         }
  67         memset(p, 0, sizeof(*p));
  68 }
  69
  70 void add_root_to_dirty_list(struct btrfs_root *root)
  71 {
  72         if (root->track_dirty && list_empty(&root->dirty_list)) {
  73                 list_add(&root->dirty_list,
  74                          &root->fs_info->dirty_cowonly_roots);
  75         }
  76 }
  77
  78 int btrfs_copy_root(struct btrfs_trans_handle *trans,
  79                       struct btrfs_root *root,
  80                       struct extent_buffer *buf,
  81                       struct extent_buffer **cow_ret, u64 new_root_objectid)
  82 {
  83         struct extent_buffer *cow;
  84         int ret = 0;
  85         int level;
  86         struct btrfs_root *new_root;
  87         struct btrfs_disk_key disk_key;
  88
  89         new_root = kmalloc(sizeof(*new_root), GFP_NOFS);
  90         if (!new_root)
  91                 return -ENOMEM;
  92
  93         memcpy(new_root, root, sizeof(*new_root));
  94         new_root->root_key.objectid = new_root_objectid;
  95
  96         WARN_ON(root->ref_cows && trans->transid !=
  97                 root->fs_info->running_transaction->transid);
  98         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
  99
 100         level = btrfs_header_level(buf);
 101         if (level == 0)
 102                 btrfs_item_key(buf, &disk_key, 0);
 103         else
 104                 btrfs_node_key(buf, &disk_key, 0);
 105         cow = btrfs_alloc_free_block(trans, new_root, buf->len,
 106                                      new_root_objectid, &disk_key,
 107                                      level, buf->start, 0);
 108         if (IS_ERR(cow)) {
 109                 kfree(new_root);
 110                 return PTR_ERR(cow);
 111         }
 112
 113         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 114         btrfs_set_header_bytenr(cow, cow->start);
 115         btrfs_set_header_generation(cow, trans->transid);
 116         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 117         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 118                                      BTRFS_HEADER_FLAG_RELOC);
 119         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 120                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 121         else
 122                 btrfs_set_header_owner(cow, new_root_objectid);
 123
 124         write_extent_buffer(cow, root->fs_info->fsid,
 125                             (unsigned long)btrfs_header_fsid(cow),
 126                             BTRFS_FSID_SIZE);
 127
 128         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 129         ret = btrfs_inc_ref(trans, new_root, cow, 0);
 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 /*
 141  * check if the tree block can be shared by multiple trees
 142  */
 143 int btrfs_block_can_be_shared(struct btrfs_root *root,
 144                               struct extent_buffer *buf)
 145 {
 146         /*
 147          * Tree blocks not in refernece counted trees and tree roots
 148          * are never shared. If a block was allocated after the last
 149          * snapshot and the block was not allocated by tree relocation,
 150          * we know the block is not shared.
 151          */
 152         if (root->ref_cows &&
 153             buf != root->node && buf != root->commit_root &&
 154             (btrfs_header_generation(buf) <=
 155              btrfs_root_last_snapshot(&root->root_item) ||
 156              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 157                 return 1;
 158 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 159         if (root->ref_cows &&
 160             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 161                 return 1;
 162 #endif
 163         return 0;
 164 }
 165
 166 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
 167                                        struct btrfs_root *root,
 168                                        struct extent_buffer *buf,
 169                                        struct extent_buffer *cow)
 170 {
 171         u64 refs;
 172         u64 owner;
 173         u64 flags;
 174         u64 new_flags = 0;
 175         int ret;
 176
 177         /*
 178          * Backrefs update rules:
 179          *
 180          * Always use full backrefs for extent pointers in tree block
 181          * allocated by tree relocation.
 182          *
 183          * If a shared tree block is no longer referenced by its owner
 184          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
 185          * use full backrefs for extent pointers in tree block.
 186          *
 187          * If a tree block is been relocating
 188          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
 189          * use full backrefs for extent pointers in tree block.
 190          * The reason for this is some operations (such as drop tree)
 191          * are only allowed for blocks use full backrefs.
 192          */
 193
 194         if (btrfs_block_can_be_shared(root, buf)) {
 195                 ret = btrfs_lookup_extent_info(trans, root, buf->start,
 196                                                btrfs_header_level(buf), 1,
 197                                                &refs, &flags);
 198                 BUG_ON(ret);
 199                 BUG_ON(refs == 0);
 200         } else {
 201                 refs = 1;
 202                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 203                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 204                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 205                 else
 206                         flags = 0;
 207         }
 208
 209         owner = btrfs_header_owner(buf);
 210         BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) &&
 211                owner == BTRFS_TREE_RELOC_OBJECTID);
 212
 213         if (refs > 1) {
 214                 if ((owner == root->root_key.objectid ||
 215                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
 216                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
 217                         ret = btrfs_inc_ref(trans, root, buf, 1);
 218                         BUG_ON(ret);
 219
 220                         if (root->root_key.objectid ==
 221                             BTRFS_TREE_RELOC_OBJECTID) {
 222                                 ret = btrfs_dec_ref(trans, root, buf, 0);
 223                                 BUG_ON(ret);
 224                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 225                                 BUG_ON(ret);
 226                         }
 227                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
 228                 } else {
 229
 230                         if (root->root_key.objectid ==
 231                             BTRFS_TREE_RELOC_OBJECTID)
 232                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 233                         else
 234                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 235                         BUG_ON(ret);
 236                 }
 237                 if (new_flags != 0) {
 238                         ret = btrfs_set_block_flags(trans, root, buf->start,
 239                                                     btrfs_header_level(buf),
 240                                                     new_flags);
 241                         BUG_ON(ret);
 242                 }
 243         } else {
 244                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
 245                         if (root->root_key.objectid ==
 246                             BTRFS_TREE_RELOC_OBJECTID)
 247                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 248                         else
 249                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 250                         BUG_ON(ret);
 251                         ret = btrfs_dec_ref(trans, root, buf, 1);
 252                         BUG_ON(ret);
 253                 }
 254                 clean_tree_block(trans, root, buf);
 255         }
 256         return 0;
 257 }
 258
 259 int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 260                              struct btrfs_root *root,
 261                              struct extent_buffer *buf,
 262                              struct extent_buffer *parent, int parent_slot,
 263                              struct extent_buffer **cow_ret,
 264                              u64 search_start, u64 empty_size)
 265 {
 266         struct extent_buffer *cow;
 267         struct btrfs_disk_key disk_key;
 268         int level;
 269
 270         WARN_ON(root->ref_cows && trans->transid !=
 271                 root->fs_info->running_transaction->transid);
 272         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 273
 274         level = btrfs_header_level(buf);
 275
 276         if (level == 0)
 277                 btrfs_item_key(buf, &disk_key, 0);
 278         else
 279                 btrfs_node_key(buf, &disk_key, 0);
 280
 281         cow = btrfs_alloc_free_block(trans, root, buf->len,
 282                                      root->root_key.objectid, &disk_key,
 283                                      level, search_start, empty_size);
 284         if (IS_ERR(cow))
 285                 return PTR_ERR(cow);
 286
 287         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 288         btrfs_set_header_bytenr(cow, cow->start);
 289         btrfs_set_header_generation(cow, trans->transid);
 290         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 291         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 292                                      BTRFS_HEADER_FLAG_RELOC);
 293         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 294                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 295         else
 296                 btrfs_set_header_owner(cow, root->root_key.objectid);
 297
 298         write_extent_buffer(cow, root->fs_info->fsid,
 299                             (unsigned long)btrfs_header_fsid(cow),
 300                             BTRFS_FSID_SIZE);
 301
 302         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 303
 304         update_ref_for_cow(trans, root, buf, cow);
 305
 306         if (buf == root->node) {
 307                 root->node = cow;
 308                 extent_buffer_get(cow);
 309
 310                 btrfs_free_extent(trans, root, buf->start, buf->len,
 311                                   0, root->root_key.objectid, level, 0);
 312                 free_extent_buffer(buf);
 313                 add_root_to_dirty_list(root);
 314         } else {
 315                 btrfs_set_node_blockptr(parent, parent_slot,
 316                                         cow->start);
 317                 WARN_ON(trans->transid == 0);
 318                 btrfs_set_node_ptr_generation(parent, parent_slot,
 319                                               trans->transid);
 320                 btrfs_mark_buffer_dirty(parent);
 321                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 322
 323                 btrfs_free_extent(trans, root, buf->start, buf->len,
 324                                   0, root->root_key.objectid, level, 1);
 325         }
 326         free_extent_buffer(buf);
 327         btrfs_mark_buffer_dirty(cow);
 328         *cow_ret = cow;
 329         return 0;
 330 }
 331
 332 static inline int should_cow_block(struct btrfs_trans_handle *trans,
 333                                    struct btrfs_root *root,
 334                                    struct extent_buffer *buf)
 335 {
 336         if (btrfs_header_generation(buf) == trans->transid &&
 337             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
 338             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
 339               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 340                 return 0;
 341         return 1;
 342 }
 343
 344 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 345                     struct btrfs_root *root, struct extent_buffer *buf,
 346                     struct extent_buffer *parent, int parent_slot,
 347                     struct extent_buffer **cow_ret)
 348 {
 349         u64 search_start;
 350         int ret;
 351         /*
 352         if (trans->transaction != root->fs_info->running_transaction) {
 353                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 354                        root->fs_info->running_transaction->transid);
 355                 WARN_ON(1);
 356         }
 357         */
 358         if (trans->transid != root->fs_info->generation) {
 359                 printk(KERN_CRIT "trans %llu running %llu\n",
 360                         (unsigned long long)trans->transid,
 361                         (unsigned long long)root->fs_info->generation);
 362                 WARN_ON(1);
 363         }
 364         if (!should_cow_block(trans, root, buf)) {
 365                 *cow_ret = buf;
 366                 return 0;
 367         }
 368
 369         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
 370         ret = __btrfs_cow_block(trans, root, buf, parent,
 371                                  parent_slot, cow_ret, search_start, 0);
 372         return ret;
 373 }
 374
 375 /*
 376 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 377 {
 378         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 379                 return 1;
 380         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 381                 return 1;
 382         return 0;
 383 }
 384 */
 385
 386 /*
 387  * compare two keys in a memcmp fashion
 388  */
 389 int btrfs_comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 390 {
 391         struct btrfs_key k1;
 392
 393         btrfs_disk_key_to_cpu(&k1, disk);
 394
 395         if (k1.objectid > k2->objectid)
 396                 return 1;
 397         if (k1.objectid < k2->objectid)
 398                 return -1;
 399         if (k1.type > k2->type)
 400                 return 1;
 401         if (k1.type < k2->type)
 402                 return -1;
 403         if (k1.offset > k2->offset)
 404                 return 1;
 405         if (k1.offset < k2->offset)
 406                 return -1;
 407         return 0;
 408 }
 409
 410
 411 #if 0
 412 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 413                        struct btrfs_root *root, struct extent_buffer *parent,
 414                        int start_slot, int cache_only, u64 *last_ret,
 415                        struct btrfs_key *progress)
 416 {
 417         struct extent_buffer *cur;
 418         struct extent_buffer *tmp;
 419         u64 blocknr;
 420         u64 gen;
 421         u64 search_start = *last_ret;
 422         u64 last_block = 0;
 423         u64 other;
 424         u32 parent_nritems;
 425         int end_slot;
 426         int i;
 427         int err = 0;
 428         int parent_level;
 429         int uptodate;
 430         u32 blocksize;
 431         int progress_passed = 0;
 432         struct btrfs_disk_key disk_key;
 433
 434         parent_level = btrfs_header_level(parent);
 435         if (cache_only && parent_level != 1)
 436                 return 0;
 437
 438         if (trans->transaction != root->fs_info->running_transaction) {
 439                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 440                        root->fs_info->running_transaction->transid);
 441                 WARN_ON(1);
 442         }
 443         if (trans->transid != root->fs_info->generation) {
 444                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 445                        root->fs_info->generation);
 446                 WARN_ON(1);
 447         }
 448
 449         parent_nritems = btrfs_header_nritems(parent);
 450         blocksize = btrfs_level_size(root, parent_level - 1);
 451         end_slot = parent_nritems;
 452
 453         if (parent_nritems == 1)
 454                 return 0;
 455
 456         for (i = start_slot; i < end_slot; i++) {
 457                 int close = 1;
 458
 459                 if (!parent->map_token) {
 460                         map_extent_buffer(parent,
 461                                         btrfs_node_key_ptr_offset(i),
 462                                         sizeof(struct btrfs_key_ptr),
 463                                         &parent->map_token, &parent->kaddr,
 464                                         &parent->map_start, &parent->map_len,
 465                                         KM_USER1);
 466                 }
 467                 btrfs_node_key(parent, &disk_key, i);
 468                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
 469                         continue;
 470
 471                 progress_passed = 1;
 472                 blocknr = btrfs_node_blockptr(parent, i);
 473                 gen = btrfs_node_ptr_generation(parent, i);
 474                 if (last_block == 0)
 475                         last_block = blocknr;
 476
 477                 if (i > 0) {
 478                         other = btrfs_node_blockptr(parent, i - 1);
 479                         close = close_blocks(blocknr, other, blocksize);
 480                 }
 481                 if (close && i < end_slot - 2) {
 482                         other = btrfs_node_blockptr(parent, i + 1);
 483                         close = close_blocks(blocknr, other, blocksize);
 484                 }
 485                 if (close) {
 486                         last_block = blocknr;
 487                         continue;
 488                 }
 489                 if (parent->map_token) {
 490                         unmap_extent_buffer(parent, parent->map_token,
 491                                             KM_USER1);
 492                         parent->map_token = NULL;
 493                 }
 494
 495                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
 496                 if (cur)
 497                         uptodate = btrfs_buffer_uptodate(cur, gen);
 498                 else
 499                         uptodate = 0;
 500                 if (!cur || !uptodate) {
 501                         if (cache_only) {
 502                                 free_extent_buffer(cur);
 503                                 continue;
 504                         }
 505                         if (!cur) {
 506                                 cur = read_tree_block(root, blocknr,
 507                                                          blocksize, gen);
 508                         } else if (!uptodate) {
 509                                 btrfs_read_buffer(cur, gen);
 510                         }
 511                 }
 512                 if (search_start == 0)
 513                         search_start = last_block;
 514
 515                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 516                                         &tmp, search_start,
 517                                         min(16 * blocksize,
 518                                             (end_slot - i) * blocksize));
 519                 if (err) {
 520                         free_extent_buffer(cur);
 521                         break;
 522                 }
 523                 search_start = tmp->start;
 524                 last_block = tmp->start;
 525                 *last_ret = search_start;
 526                 if (parent_level == 1)
 527                         btrfs_clear_buffer_defrag(tmp);
 528                 free_extent_buffer(tmp);
 529         }
 530         if (parent->map_token) {
 531                 unmap_extent_buffer(parent, parent->map_token,
 532                                     KM_USER1);
 533                 parent->map_token = NULL;
 534         }
 535         return err;
 536 }
 537 #endif
 538
 539 /*
 540  * The leaf data grows from end-to-front in the node.
 541  * this returns the address of the start of the last item,
 542  * which is the stop of the leaf data stack
 543  */
 544 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 545                                          struct extent_buffer *leaf)
 546 {
 547         u32 nr = btrfs_header_nritems(leaf);
 548         if (nr == 0)
 549                 return BTRFS_LEAF_DATA_SIZE(root);
 550         return btrfs_item_offset_nr(leaf, nr - 1);
 551 }
 552
 553 int btrfs_check_node(struct btrfs_root *root,
 554                       struct btrfs_disk_key *parent_key,
 555                       struct extent_buffer *buf)
 556 {
 557         int i;
 558         struct btrfs_key cpukey;
 559         struct btrfs_disk_key key;
 560         u32 nritems = btrfs_header_nritems(buf);
 561
 562         if (nritems == 0 || nritems > BTRFS_NODEPTRS_PER_BLOCK(root))
 563                 goto fail;
 564
 565         if (parent_key && parent_key->type) {
 566                 btrfs_node_key(buf, &key, 0);
 567                 if (memcmp(parent_key, &key, sizeof(key)))
 568                         goto fail;
 569         }
 570         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 571                 btrfs_node_key(buf, &key, i);
 572                 btrfs_node_key_to_cpu(buf, &cpukey, i + 1);
 573                 if (btrfs_comp_keys(&key, &cpukey) >= 0)
 574                         goto fail;
 575         }
 576         return 0;
 577 fail:
 578         if (btrfs_header_owner(buf) == BTRFS_EXTENT_TREE_OBJECTID) {
 579                 if (parent_key)
 580                         btrfs_disk_key_to_cpu(&cpukey, parent_key);
 581                 else
 582                         btrfs_node_key_to_cpu(buf, &cpukey, 0);
 583                 btrfs_add_corrupt_extent_record(root->fs_info, &cpukey,
 584                                                 buf->start, buf->len,
 585                                                 btrfs_header_level(buf));
 586         }
 587         return -EIO;
 588 }
 589
 590 int btrfs_check_leaf(struct btrfs_root *root,
 591                       struct btrfs_disk_key *parent_key,
 592                       struct extent_buffer *buf)
 593 {
 594         int i;
 595         struct btrfs_key cpukey;
 596         struct btrfs_disk_key key;
 597         u32 nritems = btrfs_header_nritems(buf);
 598
 599         if (nritems * sizeof(struct btrfs_item) > buf->len)  {
 600                 fprintf(stderr, "invalid number of items %llu\n",
 601                         (unsigned long long)buf->start);
 602                 goto fail;
 603         }
 604
 605         if (btrfs_header_level(buf) != 0) {
 606                 fprintf(stderr, "leaf is not a leaf %llu\n",
 607                        (unsigned long long)btrfs_header_bytenr(buf));
 608                 goto fail;
 609         }
 610         if (btrfs_leaf_free_space(root, buf) < 0) {
 611                 fprintf(stderr, "leaf free space incorrect %llu %d\n",
 612                         (unsigned long long)btrfs_header_bytenr(buf),
 613                         btrfs_leaf_free_space(root, buf));
 614                 goto fail;
 615         }
 616
 617         if (nritems == 0)
 618                 return 0;
 619
 620         btrfs_item_key(buf, &key, 0);
 621         if (parent_key && parent_key->type &&
 622             memcmp(parent_key, &key, sizeof(key))) {
 623                 fprintf(stderr, "leaf parent key incorrect %llu\n",
 624                        (unsigned long long)btrfs_header_bytenr(buf));
 625                 goto fail;
 626         }
 627         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 628                 btrfs_item_key(buf, &key, i);
 629                 btrfs_item_key_to_cpu(buf, &cpukey, i + 1);
 630                 if (btrfs_comp_keys(&key, &cpukey) >= 0) {
 631                         fprintf(stderr, "bad key ordering %d %d\n", i, i+1);
 632                         goto fail;
 633                 }
 634                 if (btrfs_item_offset_nr(buf, i) !=
 635                         btrfs_item_end_nr(buf, i + 1)) {
 636                         fprintf(stderr, "incorrect offsets %u %u\n",
 637                                 btrfs_item_offset_nr(buf, i),
 638                                 btrfs_item_end_nr(buf, i + 1));
 639                         goto fail;
 640                 }
 641                 if (i == 0 && btrfs_item_end_nr(buf, i) !=
 642                     BTRFS_LEAF_DATA_SIZE(root)) {
 643                         fprintf(stderr, "bad item end %u wanted %u\n",
 644                                 btrfs_item_end_nr(buf, i),
 645                                 (unsigned)BTRFS_LEAF_DATA_SIZE(root));
 646                         goto fail;
 647                 }
 648         }
 649         return 0;
 650 fail:
 651         if (btrfs_header_owner(buf) == BTRFS_EXTENT_TREE_OBJECTID) {
 652                 if (parent_key)
 653                         btrfs_disk_key_to_cpu(&cpukey, parent_key);
 654                 else
 655                         btrfs_item_key_to_cpu(buf, &cpukey, 0);
 656
 657                 btrfs_add_corrupt_extent_record(root->fs_info, &cpukey,
 658                                                 buf->start, buf->len, 0);
 659         }
 660         return -EIO;
 661 }
 662
 663 static int noinline check_block(struct btrfs_root *root,
 664                                 struct btrfs_path *path, int level)
 665 {
 666         struct btrfs_disk_key key;
 667         struct btrfs_disk_key *key_ptr = NULL;
 668         struct extent_buffer *parent;
 669
 670         if (path->nodes[level + 1]) {
 671                 parent = path->nodes[level + 1];
 672                 btrfs_node_key(parent, &key, path->slots[level + 1]);
 673                 key_ptr = &key;
 674         }
 675         if (level == 0)
 676                 return btrfs_check_leaf(root, key_ptr, path->nodes[0]);
 677         return btrfs_check_node(root, key_ptr, path->nodes[level]);
 678 }
 679
 680 /*
 681  * search for key in the extent_buffer.  The items start at offset p,
 682  * and they are item_size apart.  There are 'max' items in p.
 683  *
 684  * the slot in the array is returned via slot, and it points to
 685  * the place where you would insert key if it is not found in
 686  * the array.
 687  *
 688  * slot may point to max if the key is bigger than all of the keys
 689  */
 690 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 691                               int item_size, struct btrfs_key *key,
 692                               int max, int *slot)
 693 {
 694         int low = 0;
 695         int high = max;
 696         int mid;
 697         int ret;
 698         unsigned long offset;
 699         struct btrfs_disk_key *tmp;
 700
 701         while(low < high) {
 702                 mid = (low + high) / 2;
 703                 offset = p + mid * item_size;
 704
 705                 tmp = (struct btrfs_disk_key *)(eb->data + offset);
 706                 ret = btrfs_comp_keys(tmp, key);
 707
 708                 if (ret < 0)
 709                         low = mid + 1;
 710                 else if (ret > 0)
 711                         high = mid;
 712                 else {
 713                         *slot = mid;
 714                         return 0;
 715                 }
 716         }
 717         *slot = low;
 718         return 1;
 719 }
 720
 721 /*
 722  * simple bin_search frontend that does the right thing for
 723  * leaves vs nodes
 724  */
 725 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 726                       int level, int *slot)
 727 {
 728         if (level == 0) {
 729                 return generic_bin_search(eb,
 730                                           offsetof(struct btrfs_leaf, items),
 731                                           sizeof(struct btrfs_item),
 732                                           key, btrfs_header_nritems(eb),
 733                                           slot);
 734         } else {
 735                 return generic_bin_search(eb,
 736                                           offsetof(struct btrfs_node, ptrs),
 737                                           sizeof(struct btrfs_key_ptr),
 738                                           key, btrfs_header_nritems(eb),
 739                                           slot);
 740         }
 741         return -1;
 742 }
 743
 744 struct extent_buffer *read_node_slot(struct btrfs_root *root,
 745                                    struct extent_buffer *parent, int slot)
 746 {
 747         int level = btrfs_header_level(parent);
 748         if (slot < 0)
 749                 return NULL;
 750         if (slot >= btrfs_header_nritems(parent))
 751                 return NULL;
 752
 753         if (level == 0)
 754                 return NULL;
 755
 756         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 757                        btrfs_level_size(root, level - 1),
 758                        btrfs_node_ptr_generation(parent, slot));
 759 }
 760
 761 static int balance_level(struct btrfs_trans_handle *trans,
 762                          struct btrfs_root *root,
 763                          struct btrfs_path *path, int level)
 764 {
 765         struct extent_buffer *right = NULL;
 766         struct extent_buffer *mid;
 767         struct extent_buffer *left = NULL;
 768         struct extent_buffer *parent = NULL;
 769         int ret = 0;
 770         int wret;
 771         int pslot;
 772         int orig_slot = path->slots[level];
 773         u64 orig_ptr;
 774
 775         if (level == 0)
 776                 return 0;
 777
 778         mid = path->nodes[level];
 779         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 780
 781         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 782
 783         if (level < BTRFS_MAX_LEVEL - 1) {
 784                 parent = path->nodes[level + 1];
 785                 pslot = path->slots[level + 1];
 786         }
 787
 788         /*
 789          * deal with the case where there is only one pointer in the root
 790          * by promoting the node below to a root
 791          */
 792         if (!parent) {
 793                 struct extent_buffer *child;
 794
 795                 if (btrfs_header_nritems(mid) != 1)
 796                         return 0;
 797
 798                 /* promote the child to a root */
 799                 child = read_node_slot(root, mid, 0);
 800                 BUG_ON(!child);
 801                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
 802                 BUG_ON(ret);
 803
 804                 root->node = child;
 805                 add_root_to_dirty_list(root);
 806                 path->nodes[level] = NULL;
 807                 clean_tree_block(trans, root, mid);
 808                 wait_on_tree_block_writeback(root, mid);
 809                 /* once for the path */
 810                 free_extent_buffer(mid);
 811
 812                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 813                                         0, root->root_key.objectid,
 814                                         level, 1);
 815                 /* once for the root ptr */
 816                 free_extent_buffer(mid);
 817                 return ret;
 818         }
 819         if (btrfs_header_nritems(mid) >
 820             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 821                 return 0;
 822
 823         left = read_node_slot(root, parent, pslot - 1);
 824         if (left) {
 825                 wret = btrfs_cow_block(trans, root, left,
 826                                        parent, pslot - 1, &left);
 827                 if (wret) {
 828                         ret = wret;
 829                         goto enospc;
 830                 }
 831         }
 832         right = read_node_slot(root, parent, pslot + 1);
 833         if (right) {
 834                 wret = btrfs_cow_block(trans, root, right,
 835                                        parent, pslot + 1, &right);
 836                 if (wret) {
 837                         ret = wret;
 838                         goto enospc;
 839                 }
 840         }
 841
 842         /* first, try to make some room in the middle buffer */
 843         if (left) {
 844                 orig_slot += btrfs_header_nritems(left);
 845                 wret = push_node_left(trans, root, left, mid, 1);
 846                 if (wret < 0)
 847                         ret = wret;
 848         }
 849
 850         /*
 851          * then try to empty the right most buffer into the middle
 852          */
 853         if (right) {
 854                 wret = push_node_left(trans, root, mid, right, 1);
 855                 if (wret < 0 && wret != -ENOSPC)
 856                         ret = wret;
 857                 if (btrfs_header_nritems(right) == 0) {
 858                         u64 bytenr = right->start;
 859                         u32 blocksize = right->len;
 860
 861                         clean_tree_block(trans, root, right);
 862                         wait_on_tree_block_writeback(root, right);
 863                         free_extent_buffer(right);
 864                         right = NULL;
 865                         wret = btrfs_del_ptr(trans, root, path,
 866                                              level + 1, pslot + 1);
 867                         if (wret)
 868                                 ret = wret;
 869                         wret = btrfs_free_extent(trans, root, bytenr,
 870                                                  blocksize, 0,
 871                                                  root->root_key.objectid,
 872                                                  level, 0);
 873                         if (wret)
 874                                 ret = wret;
 875                 } else {
 876                         struct btrfs_disk_key right_key;
 877                         btrfs_node_key(right, &right_key, 0);
 878                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 879                         btrfs_mark_buffer_dirty(parent);
 880                 }
 881         }
 882         if (btrfs_header_nritems(mid) == 1) {
 883                 /*
 884                  * we're not allowed to leave a node with one item in the
 885                  * tree during a delete.  A deletion from lower in the tree
 886                  * could try to delete the only pointer in this node.
 887                  * So, pull some keys from the left.
 888                  * There has to be a left pointer at this point because
 889                  * otherwise we would have pulled some pointers from the
 890                  * right
 891                  */
 892                 BUG_ON(!left);
 893                 wret = balance_node_right(trans, root, mid, left);
 894                 if (wret < 0) {
 895                         ret = wret;
 896                         goto enospc;
 897                 }
 898                 if (wret == 1) {
 899                         wret = push_node_left(trans, root, left, mid, 1);
 900                         if (wret < 0)
 901                                 ret = wret;
 902                 }
 903                 BUG_ON(wret == 1);
 904         }
 905         if (btrfs_header_nritems(mid) == 0) {
 906                 /* we've managed to empty the middle node, drop it */
 907                 u64 bytenr = mid->start;
 908                 u32 blocksize = mid->len;
 909                 clean_tree_block(trans, root, mid);
 910                 wait_on_tree_block_writeback(root, mid);
 911                 free_extent_buffer(mid);
 912                 mid = NULL;
 913                 wret = btrfs_del_ptr(trans, root, path, level + 1, pslot);
 914                 if (wret)
 915                         ret = wret;
 916                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 917                                          0, root->root_key.objectid,
 918                                          level, 0);
 919                 if (wret)
 920                         ret = wret;
 921         } else {
 922                 /* update the parent key to reflect our changes */
 923                 struct btrfs_disk_key mid_key;
 924                 btrfs_node_key(mid, &mid_key, 0);
 925                 btrfs_set_node_key(parent, &mid_key, pslot);
 926                 btrfs_mark_buffer_dirty(parent);
 927         }
 928
 929         /* update the path */
 930         if (left) {
 931                 if (btrfs_header_nritems(left) > orig_slot) {
 932                         extent_buffer_get(left);
 933                         path->nodes[level] = left;
 934                         path->slots[level + 1] -= 1;
 935                         path->slots[level] = orig_slot;
 936                         if (mid)
 937                                 free_extent_buffer(mid);
 938                 } else {
 939                         orig_slot -= btrfs_header_nritems(left);
 940                         path->slots[level] = orig_slot;
 941                 }
 942         }
 943         /* double check we haven't messed things up */
 944         check_block(root, path, level);
 945         if (orig_ptr !=
 946             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 947                 BUG();
 948 enospc:
 949         if (right)
 950                 free_extent_buffer(right);
 951         if (left)
 952                 free_extent_buffer(left);
 953         return ret;
 954 }
 955
 956 /* returns zero if the push worked, non-zero otherwise */
 957 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
 958                                           struct btrfs_root *root,
 959                                           struct btrfs_path *path, int level)
 960 {
 961         struct extent_buffer *right = NULL;
 962         struct extent_buffer *mid;
 963         struct extent_buffer *left = NULL;
 964         struct extent_buffer *parent = NULL;
 965         int ret = 0;
 966         int wret;
 967         int pslot;
 968         int orig_slot = path->slots[level];
 969
 970         if (level == 0)
 971                 return 1;
 972
 973         mid = path->nodes[level];
 974         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 975
 976         if (level < BTRFS_MAX_LEVEL - 1) {
 977                 parent = path->nodes[level + 1];
 978                 pslot = path->slots[level + 1];
 979         }
 980
 981         if (!parent)
 982                 return 1;
 983
 984         left = read_node_slot(root, parent, pslot - 1);
 985
 986         /* first, try to make some room in the middle buffer */
 987         if (left) {
 988                 u32 left_nr;
 989                 left_nr = btrfs_header_nritems(left);
 990                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 991                         wret = 1;
 992                 } else {
 993                         ret = btrfs_cow_block(trans, root, left, parent,
 994                                               pslot - 1, &left);
 995                         if (ret)
 996                                 wret = 1;
 997                         else {
 998                                 wret = push_node_left(trans, root,
 999                                                       left, mid, 0);
1000                         }
1001                 }
1002                 if (wret < 0)
1003                         ret = wret;
1004                 if (wret == 0) {
1005                         struct btrfs_disk_key disk_key;
1006                         orig_slot += left_nr;
1007                         btrfs_node_key(mid, &disk_key, 0);
1008                         btrfs_set_node_key(parent, &disk_key, pslot);
1009                         btrfs_mark_buffer_dirty(parent);
1010                         if (btrfs_header_nritems(left) > orig_slot) {
1011                                 path->nodes[level] = left;
1012                                 path->slots[level + 1] -= 1;
1013                                 path->slots[level] = orig_slot;
1014                                 free_extent_buffer(mid);
1015                         } else {
1016                                 orig_slot -=
1017                                         btrfs_header_nritems(left);
1018                                 path->slots[level] = orig_slot;
1019                                 free_extent_buffer(left);
1020                         }
1021                         return 0;
1022                 }
1023                 free_extent_buffer(left);
1024         }
1025         right= read_node_slot(root, parent, pslot + 1);
1026
1027         /*
1028          * then try to empty the right most buffer into the middle
1029          */
1030         if (right) {
1031                 u32 right_nr;
1032                 right_nr = btrfs_header_nritems(right);
1033                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1034                         wret = 1;
1035                 } else {
1036                         ret = btrfs_cow_block(trans, root, right,
1037                                               parent, pslot + 1,
1038                                               &right);
1039                         if (ret)
1040                                 wret = 1;
1041                         else {
1042                                 wret = balance_node_right(trans, root,
1043                                                           right, mid);
1044                         }
1045                 }
1046                 if (wret < 0)
1047                         ret = wret;
1048                 if (wret == 0) {
1049                         struct btrfs_disk_key disk_key;
1050
1051                         btrfs_node_key(right, &disk_key, 0);
1052                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
1053                         btrfs_mark_buffer_dirty(parent);
1054
1055                         if (btrfs_header_nritems(mid) <= orig_slot) {
1056                                 path->nodes[level] = right;
1057                                 path->slots[level + 1] += 1;
1058                                 path->slots[level] = orig_slot -
1059                                         btrfs_header_nritems(mid);
1060                                 free_extent_buffer(mid);
1061                         } else {
1062                                 free_extent_buffer(right);
1063                         }
1064                         return 0;
1065                 }
1066                 free_extent_buffer(right);
1067         }
1068         return 1;
1069 }
1070
1071 /*
1072  * readahead one full node of leaves
1073  */
1074 void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
1075                              int level, int slot, u64 objectid)
1076 {
1077         struct extent_buffer *node;
1078         struct btrfs_disk_key disk_key;
1079         u32 nritems;
1080         u64 search;
1081         u64 lowest_read;
1082         u64 highest_read;
1083         u64 nread = 0;
1084         int direction = path->reada;
1085         struct extent_buffer *eb;
1086         u32 nr;
1087         u32 blocksize;
1088         u32 nscan = 0;
1089
1090         if (level != 1)
1091                 return;
1092
1093         if (!path->nodes[level])
1094                 return;
1095
1096         node = path->nodes[level];
1097         search = btrfs_node_blockptr(node, slot);
1098         blocksize = btrfs_level_size(root, level - 1);
1099         eb = btrfs_find_tree_block(root, search, blocksize);
1100         if (eb) {
1101                 free_extent_buffer(eb);
1102                 return;
1103         }
1104
1105         highest_read = search;
1106         lowest_read = search;
1107
1108         nritems = btrfs_header_nritems(node);
1109         nr = slot;
1110         while(1) {
1111                 if (direction < 0) {
1112                         if (nr == 0)
1113                                 break;
1114                         nr--;
1115                 } else if (direction > 0) {
1116                         nr++;
1117                         if (nr >= nritems)
1118                                 break;
1119                 }
1120                 if (path->reada < 0 && objectid) {
1121                         btrfs_node_key(node, &disk_key, nr);
1122                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1123                                 break;
1124                 }
1125                 search = btrfs_node_blockptr(node, nr);
1126                 if ((search >= lowest_read && search <= highest_read) ||
1127                     (search < lowest_read && lowest_read - search <= 32768) ||
1128                     (search > highest_read && search - highest_read <= 32768)) {
1129                         readahead_tree_block(root, search, blocksize,
1130                                      btrfs_node_ptr_generation(node, nr));
1131                         nread += blocksize;
1132                 }
1133                 nscan++;
1134                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
1135                         break;
1136                 if(nread > (1024 * 1024) || nscan > 128)
1137                         break;
1138
1139                 if (search < lowest_read)
1140                         lowest_read = search;
1141                 if (search > highest_read)
1142                         highest_read = search;
1143         }
1144 }
1145
1146 /*
1147  * look for key in the tree.  path is filled in with nodes along the way
1148  * if key is found, we return zero and you can find the item in the leaf
1149  * level of the path (level 0)
1150  *
1151  * If the key isn't found, the path points to the slot where it should
1152  * be inserted, and 1 is returned.  If there are other errors during the
1153  * search a negative error number is returned.
1154  *
1155  * if ins_len > 0, nodes and leaves will be split as we walk down the
1156  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1157  * possible)
1158  */
1159 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1160                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1161                       ins_len, int cow)
1162 {
1163         struct extent_buffer *b;
1164         int slot;
1165         int ret;
1166         int level;
1167         int should_reada = p->reada;
1168         u8 lowest_level = 0;
1169
1170         lowest_level = p->lowest_level;
1171         WARN_ON(lowest_level && ins_len > 0);
1172         WARN_ON(p->nodes[0] != NULL);
1173         /*
1174         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
1175         */
1176 again:
1177         b = root->node;
1178         extent_buffer_get(b);
1179         while (b) {
1180                 level = btrfs_header_level(b);
1181                 if (cow) {
1182                         int wret;
1183                         wret = btrfs_cow_block(trans, root, b,
1184                                                p->nodes[level + 1],
1185                                                p->slots[level + 1],
1186                                                &b);
1187                         if (wret) {
1188                                 free_extent_buffer(b);
1189                                 return wret;
1190                         }
1191                 }
1192                 BUG_ON(!cow && ins_len);
1193                 if (level != btrfs_header_level(b))
1194                         WARN_ON(1);
1195                 level = btrfs_header_level(b);
1196                 p->nodes[level] = b;
1197                 ret = check_block(root, p, level);
1198                 if (ret)
1199                         return -1;
1200                 ret = bin_search(b, key, level, &slot);
1201                 if (level != 0) {
1202                         if (ret && slot > 0)
1203                                 slot -= 1;
1204                         p->slots[level] = slot;
1205                         if ((p->search_for_split || ins_len > 0) &&
1206                             btrfs_header_nritems(b) >=
1207                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1208                                 int sret = split_node(trans, root, p, level);
1209                                 BUG_ON(sret > 0);
1210                                 if (sret)
1211                                         return sret;
1212                                 b = p->nodes[level];
1213                                 slot = p->slots[level];
1214                         } else if (ins_len < 0) {
1215                                 int sret = balance_level(trans, root, p,
1216                                                          level);
1217                                 if (sret)
1218                                         return sret;
1219                                 b = p->nodes[level];
1220                                 if (!b) {
1221                                         btrfs_release_path(NULL, p);
1222                                         goto again;
1223                                 }
1224                                 slot = p->slots[level];
1225                                 BUG_ON(btrfs_header_nritems(b) == 1);
1226                         }
1227                         /* this is only true while dropping a snapshot */
1228                         if (level == lowest_level)
1229                                 break;
1230
1231                         if (should_reada)
1232                                 reada_for_search(root, p, level, slot,
1233                                                  key->objectid);
1234
1235                         b = read_node_slot(root, b, slot);
1236                         if (!extent_buffer_uptodate(b))
1237                                 return -EIO;
1238                 } else {
1239                         p->slots[level] = slot;
1240                         if (ins_len > 0 &&
1241                             ins_len > btrfs_leaf_free_space(root, b)) {
1242                                 int sret = split_leaf(trans, root, key,
1243                                                       p, ins_len, ret == 0);
1244                                 BUG_ON(sret > 0);
1245                                 if (sret)
1246                                         return sret;
1247                         }
1248                         return ret;
1249                 }
1250         }
1251         return 1;
1252 }
1253
1254 /*
1255  * adjust the pointers going up the tree, starting at level
1256  * making sure the right key of each node is points to 'key'.
1257  * This is used after shifting pointers to the left, so it stops
1258  * fixing up pointers when a given leaf/node is not in slot 0 of the
1259  * higher levels
1260  *
1261  * If this fails to write a tree block, it returns -1, but continues
1262  * fixing up the blocks in ram so the tree is consistent.
1263  */
1264 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1265                           struct btrfs_root *root, struct btrfs_path *path,
1266                           struct btrfs_disk_key *key, int level)
1267 {
1268         int i;
1269         int ret = 0;
1270         struct extent_buffer *t;
1271
1272         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1273                 int tslot = path->slots[i];
1274                 if (!path->nodes[i])
1275                         break;
1276                 t = path->nodes[i];
1277                 btrfs_set_node_key(t, key, tslot);
1278                 btrfs_mark_buffer_dirty(path->nodes[i]);
1279                 if (tslot != 0)
1280                         break;
1281         }
1282         return ret;
1283 }
1284
1285 /*
1286  * update item key.
1287  *
1288  * This function isn't completely safe. It's the caller's responsibility
1289  * that the new key won't break the order
1290  */
1291 int btrfs_set_item_key_safe(struct btrfs_trans_handle *trans,
1292                             struct btrfs_root *root, struct btrfs_path *path,
1293                             struct btrfs_key *new_key)
1294 {
1295         struct btrfs_disk_key disk_key;
1296         struct extent_buffer *eb;
1297         int slot;
1298
1299         eb = path->nodes[0];
1300         slot = path->slots[0];
1301         if (slot > 0) {
1302                 btrfs_item_key(eb, &disk_key, slot - 1);
1303                 if (btrfs_comp_keys(&disk_key, new_key) >= 0)
1304                         return -1;
1305         }
1306         if (slot < btrfs_header_nritems(eb) - 1) {
1307                 btrfs_item_key(eb, &disk_key, slot + 1);
1308                 if (btrfs_comp_keys(&disk_key, new_key) <= 0)
1309                         return -1;
1310         }
1311
1312         btrfs_cpu_key_to_disk(&disk_key, new_key);
1313         btrfs_set_item_key(eb, &disk_key, slot);
1314         btrfs_mark_buffer_dirty(eb);
1315         if (slot == 0)
1316                 fixup_low_keys(trans, root, path, &disk_key, 1);
1317         return 0;
1318 }
1319
1320 /*
1321  * try to push data from one node into the next node left in the
1322  * tree.
1323  *
1324  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1325  * error, and > 0 if there was no room in the left hand block.
1326  */
1327 static int push_node_left(struct btrfs_trans_handle *trans,
1328                           struct btrfs_root *root, struct extent_buffer *dst,
1329                           struct extent_buffer *src, int empty)
1330 {
1331         int push_items = 0;
1332         int src_nritems;
1333         int dst_nritems;
1334         int ret = 0;
1335
1336         src_nritems = btrfs_header_nritems(src);
1337         dst_nritems = btrfs_header_nritems(dst);
1338         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1339         WARN_ON(btrfs_header_generation(src) != trans->transid);
1340         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1341
1342         if (!empty && src_nritems <= 8)
1343                 return 1;
1344
1345         if (push_items <= 0) {
1346                 return 1;
1347         }
1348
1349         if (empty) {
1350                 push_items = min(src_nritems, push_items);
1351                 if (push_items < src_nritems) {
1352                         /* leave at least 8 pointers in the node if
1353                          * we aren't going to empty it
1354                          */
1355                         if (src_nritems - push_items < 8) {
1356                                 if (push_items <= 8)
1357                                         return 1;
1358                                 push_items -= 8;
1359                         }
1360                 }
1361         } else
1362                 push_items = min(src_nritems - 8, push_items);
1363
1364         copy_extent_buffer(dst, src,
1365                            btrfs_node_key_ptr_offset(dst_nritems),
1366                            btrfs_node_key_ptr_offset(0),
1367                            push_items * sizeof(struct btrfs_key_ptr));
1368
1369         if (push_items < src_nritems) {
1370                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1371                                       btrfs_node_key_ptr_offset(push_items),
1372                                       (src_nritems - push_items) *
1373                                       sizeof(struct btrfs_key_ptr));
1374         }
1375         btrfs_set_header_nritems(src, src_nritems - push_items);
1376         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1377         btrfs_mark_buffer_dirty(src);
1378         btrfs_mark_buffer_dirty(dst);
1379
1380         return ret;
1381 }
1382
1383 /*
1384  * try to push data from one node into the next node right in the
1385  * tree.
1386  *
1387  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1388  * error, and > 0 if there was no room in the right hand block.
1389  *
1390  * this will  only push up to 1/2 the contents of the left node over
1391  */
1392 static int balance_node_right(struct btrfs_trans_handle *trans,
1393                               struct btrfs_root *root,
1394                               struct extent_buffer *dst,
1395                               struct extent_buffer *src)
1396 {
1397         int push_items = 0;
1398         int max_push;
1399         int src_nritems;
1400         int dst_nritems;
1401         int ret = 0;
1402
1403         WARN_ON(btrfs_header_generation(src) != trans->transid);
1404         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1405
1406         src_nritems = btrfs_header_nritems(src);
1407         dst_nritems = btrfs_header_nritems(dst);
1408         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1409         if (push_items <= 0) {
1410                 return 1;
1411         }
1412
1413         if (src_nritems < 4) {
1414                 return 1;
1415         }
1416
1417         max_push = src_nritems / 2 + 1;
1418         /* don't try to empty the node */
1419         if (max_push >= src_nritems) {
1420                 return 1;
1421         }
1422
1423         if (max_push < push_items)
1424                 push_items = max_push;
1425
1426         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1427                                       btrfs_node_key_ptr_offset(0),
1428                                       (dst_nritems) *
1429                                       sizeof(struct btrfs_key_ptr));
1430
1431         copy_extent_buffer(dst, src,
1432                            btrfs_node_key_ptr_offset(0),
1433                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1434                            push_items * sizeof(struct btrfs_key_ptr));
1435
1436         btrfs_set_header_nritems(src, src_nritems - push_items);
1437         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1438
1439         btrfs_mark_buffer_dirty(src);
1440         btrfs_mark_buffer_dirty(dst);
1441
1442         return ret;
1443 }
1444
1445 /*
1446  * helper function to insert a new root level in the tree.
1447  * A new node is allocated, and a single item is inserted to
1448  * point to the existing root
1449  *
1450  * returns zero on success or < 0 on failure.
1451  */
1452 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1453                            struct btrfs_root *root,
1454                            struct btrfs_path *path, int level)
1455 {
1456         u64 lower_gen;
1457         struct extent_buffer *lower;
1458         struct extent_buffer *c;
1459         struct extent_buffer *old;
1460         struct btrfs_disk_key lower_key;
1461
1462         BUG_ON(path->nodes[level]);
1463         BUG_ON(path->nodes[level-1] != root->node);
1464
1465         lower = path->nodes[level-1];
1466         if (level == 1)
1467                 btrfs_item_key(lower, &lower_key, 0);
1468         else
1469                 btrfs_node_key(lower, &lower_key, 0);
1470
1471         c = btrfs_alloc_free_block(trans, root, root->nodesize,
1472                                    root->root_key.objectid, &lower_key,
1473                                    level, root->node->start, 0);
1474
1475         if (IS_ERR(c))
1476                 return PTR_ERR(c);
1477
1478         memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
1479         btrfs_set_header_nritems(c, 1);
1480         btrfs_set_header_level(c, level);
1481         btrfs_set_header_bytenr(c, c->start);
1482         btrfs_set_header_generation(c, trans->transid);
1483         btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
1484         btrfs_set_header_owner(c, root->root_key.objectid);
1485
1486         write_extent_buffer(c, root->fs_info->fsid,
1487                             (unsigned long)btrfs_header_fsid(c),
1488                             BTRFS_FSID_SIZE);
1489
1490         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1491                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
1492                             BTRFS_UUID_SIZE);
1493
1494         btrfs_set_node_key(c, &lower_key, 0);
1495         btrfs_set_node_blockptr(c, 0, lower->start);
1496         lower_gen = btrfs_header_generation(lower);
1497         WARN_ON(lower_gen != trans->transid);
1498
1499         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1500
1501         btrfs_mark_buffer_dirty(c);
1502
1503         old = root->node;
1504         root->node = c;
1505
1506         /* the super has an extra ref to root->node */
1507         free_extent_buffer(old);
1508
1509         add_root_to_dirty_list(root);
1510         extent_buffer_get(c);
1511         path->nodes[level] = c;
1512         path->slots[level] = 0;
1513         return 0;
1514 }
1515
1516 /*
1517  * worker function to insert a single pointer in a node.
1518  * the node should have enough room for the pointer already
1519  *
1520  * slot and level indicate where you want the key to go, and
1521  * blocknr is the block the key points to.
1522  *
1523  * returns zero on success and < 0 on any error
1524  */
1525 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1526                       *root, struct btrfs_path *path, struct btrfs_disk_key
1527                       *key, u64 bytenr, int slot, int level)
1528 {
1529         struct extent_buffer *lower;
1530         int nritems;
1531
1532         BUG_ON(!path->nodes[level]);
1533         lower = path->nodes[level];
1534         nritems = btrfs_header_nritems(lower);
1535         if (slot > nritems)
1536                 BUG();
1537         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1538                 BUG();
1539         if (slot != nritems) {
1540                 memmove_extent_buffer(lower,
1541                               btrfs_node_key_ptr_offset(slot + 1),
1542                               btrfs_node_key_ptr_offset(slot),
1543                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1544         }
1545         btrfs_set_node_key(lower, key, slot);
1546         btrfs_set_node_blockptr(lower, slot, bytenr);
1547         WARN_ON(trans->transid == 0);
1548         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1549         btrfs_set_header_nritems(lower, nritems + 1);
1550         btrfs_mark_buffer_dirty(lower);
1551         return 0;
1552 }
1553
1554 /*
1555  * split the node at the specified level in path in two.
1556  * The path is corrected to point to the appropriate node after the split
1557  *
1558  * Before splitting this tries to make some room in the node by pushing
1559  * left and right, if either one works, it returns right away.
1560  *
1561  * returns 0 on success and < 0 on failure
1562  */
1563 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1564                       *root, struct btrfs_path *path, int level)
1565 {
1566         struct extent_buffer *c;
1567         struct extent_buffer *split;
1568         struct btrfs_disk_key disk_key;
1569         int mid;
1570         int ret;
1571         int wret;
1572         u32 c_nritems;
1573
1574         c = path->nodes[level];
1575         WARN_ON(btrfs_header_generation(c) != trans->transid);
1576         if (c == root->node) {
1577                 /* trying to split the root, lets make a new one */
1578                 ret = insert_new_root(trans, root, path, level + 1);
1579                 if (ret)
1580                         return ret;
1581         } else {
1582                 ret = push_nodes_for_insert(trans, root, path, level);
1583                 c = path->nodes[level];
1584                 if (!ret && btrfs_header_nritems(c) <
1585                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
1586                         return 0;
1587                 if (ret < 0)
1588                         return ret;
1589         }
1590
1591         c_nritems = btrfs_header_nritems(c);
1592         mid = (c_nritems + 1) / 2;
1593         btrfs_node_key(c, &disk_key, mid);
1594
1595         split = btrfs_alloc_free_block(trans, root, root->nodesize,
1596                                         root->root_key.objectid,
1597                                         &disk_key, level, c->start, 0);
1598         if (IS_ERR(split))
1599                 return PTR_ERR(split);
1600
1601         memset_extent_buffer(split, 0, 0, sizeof(struct btrfs_header));
1602         btrfs_set_header_level(split, btrfs_header_level(c));
1603         btrfs_set_header_bytenr(split, split->start);
1604         btrfs_set_header_generation(split, trans->transid);
1605         btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
1606         btrfs_set_header_owner(split, root->root_key.objectid);
1607         write_extent_buffer(split, root->fs_info->fsid,
1608                             (unsigned long)btrfs_header_fsid(split),
1609                             BTRFS_FSID_SIZE);
1610         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1611                             (unsigned long)btrfs_header_chunk_tree_uuid(split),
1612                             BTRFS_UUID_SIZE);
1613
1614
1615         copy_extent_buffer(split, c,
1616                            btrfs_node_key_ptr_offset(0),
1617                            btrfs_node_key_ptr_offset(mid),
1618                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1619         btrfs_set_header_nritems(split, c_nritems - mid);
1620         btrfs_set_header_nritems(c, mid);
1621         ret = 0;
1622
1623         btrfs_mark_buffer_dirty(c);
1624         btrfs_mark_buffer_dirty(split);
1625
1626         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1627                           path->slots[level + 1] + 1,
1628                           level + 1);
1629         if (wret)
1630                 ret = wret;
1631
1632         if (path->slots[level] >= mid) {
1633                 path->slots[level] -= mid;
1634                 free_extent_buffer(c);
1635                 path->nodes[level] = split;
1636                 path->slots[level + 1] += 1;
1637         } else {
1638                 free_extent_buffer(split);
1639         }
1640         return ret;
1641 }
1642
1643 /*
1644  * how many bytes are required to store the items in a leaf.  start
1645  * and nr indicate which items in the leaf to check.  This totals up the
1646  * space used both by the item structs and the item data
1647  */
1648 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1649 {
1650         int data_len;
1651         int nritems = btrfs_header_nritems(l);
1652         int end = min(nritems, start + nr) - 1;
1653
1654         if (!nr)
1655                 return 0;
1656         data_len = btrfs_item_end_nr(l, start);
1657         data_len = data_len - btrfs_item_offset_nr(l, end);
1658         data_len += sizeof(struct btrfs_item) * nr;
1659         WARN_ON(data_len < 0);
1660         return data_len;
1661 }
1662
1663 /*
1664  * The space between the end of the leaf items and
1665  * the start of the leaf data.  IOW, how much room
1666  * the leaf has left for both items and data
1667  */
1668 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1669 {
1670         int nritems = btrfs_header_nritems(leaf);
1671         int ret;
1672         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1673         if (ret < 0) {
1674                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1675                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1676                        leaf_space_used(leaf, 0, nritems), nritems);
1677         }
1678         return ret;
1679 }
1680
1681 /*
1682  * push some data in the path leaf to the right, trying to free up at
1683  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1684  *
1685  * returns 1 if the push failed because the other node didn't have enough
1686  * room, 0 if everything worked out and < 0 if there were major errors.
1687  */
1688 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1689                            *root, struct btrfs_path *path, int data_size,
1690                            int empty)
1691 {
1692         struct extent_buffer *left = path->nodes[0];
1693         struct extent_buffer *right;
1694         struct extent_buffer *upper;
1695         struct btrfs_disk_key disk_key;
1696         int slot;
1697         u32 i;
1698         int free_space;
1699         int push_space = 0;
1700         int push_items = 0;
1701         struct btrfs_item *item;
1702         u32 left_nritems;
1703         u32 nr;
1704         u32 right_nritems;
1705         u32 data_end;
1706         u32 this_item_size;
1707         int ret;
1708
1709         slot = path->slots[1];
1710         if (!path->nodes[1]) {
1711                 return 1;
1712         }
1713         upper = path->nodes[1];
1714         if (slot >= btrfs_header_nritems(upper) - 1)
1715                 return 1;
1716
1717         right = read_node_slot(root, upper, slot + 1);
1718         free_space = btrfs_leaf_free_space(root, right);
1719         if (free_space < data_size) {
1720                 free_extent_buffer(right);
1721                 return 1;
1722         }
1723
1724         /* cow and double check */
1725         ret = btrfs_cow_block(trans, root, right, upper,
1726                               slot + 1, &right);
1727         if (ret) {
1728                 free_extent_buffer(right);
1729                 return 1;
1730         }
1731         free_space = btrfs_leaf_free_space(root, right);
1732         if (free_space < data_size) {
1733                 free_extent_buffer(right);
1734                 return 1;
1735         }
1736
1737         left_nritems = btrfs_header_nritems(left);
1738         if (left_nritems == 0) {
1739                 free_extent_buffer(right);
1740                 return 1;
1741         }
1742
1743         if (empty)
1744                 nr = 0;
1745         else
1746                 nr = 1;
1747
1748         i = left_nritems - 1;
1749         while (i >= nr) {
1750                 item = btrfs_item_nr(left, i);
1751
1752                 if (path->slots[0] == i)
1753                         push_space += data_size + sizeof(*item);
1754
1755                 this_item_size = btrfs_item_size(left, item);
1756                 if (this_item_size + sizeof(*item) + push_space > free_space)
1757                         break;
1758                 push_items++;
1759                 push_space += this_item_size + sizeof(*item);
1760                 if (i == 0)
1761                         break;
1762                 i--;
1763         }
1764
1765         if (push_items == 0) {
1766                 free_extent_buffer(right);
1767                 return 1;
1768         }
1769
1770         if (!empty && push_items == left_nritems)
1771                 WARN_ON(1);
1772
1773         /* push left to right */
1774         right_nritems = btrfs_header_nritems(right);
1775
1776         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1777         push_space -= leaf_data_end(root, left);
1778
1779         /* make room in the right data area */
1780         data_end = leaf_data_end(root, right);
1781         memmove_extent_buffer(right,
1782                               btrfs_leaf_data(right) + data_end - push_space,
1783                               btrfs_leaf_data(right) + data_end,
1784                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1785
1786         /* copy from the left data area */
1787         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1788                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1789                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1790                      push_space);
1791
1792         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1793                               btrfs_item_nr_offset(0),
1794                               right_nritems * sizeof(struct btrfs_item));
1795
1796         /* copy the items from left to right */
1797         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1798                    btrfs_item_nr_offset(left_nritems - push_items),
1799                    push_items * sizeof(struct btrfs_item));
1800
1801         /* update the item pointers */
1802         right_nritems += push_items;
1803         btrfs_set_header_nritems(right, right_nritems);
1804         push_space = BTRFS_LEAF_DATA_SIZE(root);
1805         for (i = 0; i < right_nritems; i++) {
1806                 item = btrfs_item_nr(right, i);
1807                 push_space -= btrfs_item_size(right, item);
1808                 btrfs_set_item_offset(right, item, push_space);
1809         }
1810
1811         left_nritems -= push_items;
1812         btrfs_set_header_nritems(left, left_nritems);
1813
1814         if (left_nritems)
1815                 btrfs_mark_buffer_dirty(left);
1816         btrfs_mark_buffer_dirty(right);
1817
1818         btrfs_item_key(right, &disk_key, 0);
1819         btrfs_set_node_key(upper, &disk_key, slot + 1);
1820         btrfs_mark_buffer_dirty(upper);
1821
1822         /* then fixup the leaf pointer in the path */
1823         if (path->slots[0] >= left_nritems) {
1824                 path->slots[0] -= left_nritems;
1825                 free_extent_buffer(path->nodes[0]);
1826                 path->nodes[0] = right;
1827                 path->slots[1] += 1;
1828         } else {
1829                 free_extent_buffer(right);
1830         }
1831         return 0;
1832 }
1833 /*
1834  * push some data in the path leaf to the left, trying to free up at
1835  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1836  */
1837 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1838                           *root, struct btrfs_path *path, int data_size,
1839                           int empty)
1840 {
1841         struct btrfs_disk_key disk_key;
1842         struct extent_buffer *right = path->nodes[0];
1843         struct extent_buffer *left;
1844         int slot;
1845         int i;
1846         int free_space;
1847         int push_space = 0;
1848         int push_items = 0;
1849         struct btrfs_item *item;
1850         u32 old_left_nritems;
1851         u32 right_nritems;
1852         u32 nr;
1853         int ret = 0;
1854         int wret;
1855         u32 this_item_size;
1856         u32 old_left_item_size;
1857
1858         slot = path->slots[1];
1859         if (slot == 0)
1860                 return 1;
1861         if (!path->nodes[1])
1862                 return 1;
1863
1864         right_nritems = btrfs_header_nritems(right);
1865         if (right_nritems == 0) {
1866                 return 1;
1867         }
1868
1869         left = read_node_slot(root, path->nodes[1], slot - 1);
1870         free_space = btrfs_leaf_free_space(root, left);
1871         if (free_space < data_size) {
1872                 free_extent_buffer(left);
1873                 return 1;
1874         }
1875
1876         /* cow and double check */
1877         ret = btrfs_cow_block(trans, root, left,
1878                               path->nodes[1], slot - 1, &left);
1879         if (ret) {
1880                 /* we hit -ENOSPC, but it isn't fatal here */
1881                 free_extent_buffer(left);
1882                 return 1;
1883         }
1884
1885         free_space = btrfs_leaf_free_space(root, left);
1886         if (free_space < data_size) {
1887                 free_extent_buffer(left);
1888                 return 1;
1889         }
1890
1891         if (empty)
1892                 nr = right_nritems;
1893         else
1894                 nr = right_nritems - 1;
1895
1896         for (i = 0; i < nr; i++) {
1897                 item = btrfs_item_nr(right, i);
1898
1899                 if (path->slots[0] == i)
1900                         push_space += data_size + sizeof(*item);
1901
1902                 this_item_size = btrfs_item_size(right, item);
1903                 if (this_item_size + sizeof(*item) + push_space > free_space)
1904                         break;
1905
1906                 push_items++;
1907                 push_space += this_item_size + sizeof(*item);
1908         }
1909
1910         if (push_items == 0) {
1911                 free_extent_buffer(left);
1912                 return 1;
1913         }
1914         if (!empty && push_items == btrfs_header_nritems(right))
1915                 WARN_ON(1);
1916
1917         /* push data from right to left */
1918         copy_extent_buffer(left, right,
1919                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1920                            btrfs_item_nr_offset(0),
1921                            push_items * sizeof(struct btrfs_item));
1922
1923         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1924                      btrfs_item_offset_nr(right, push_items -1);
1925
1926         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1927                      leaf_data_end(root, left) - push_space,
1928                      btrfs_leaf_data(right) +
1929                      btrfs_item_offset_nr(right, push_items - 1),
1930                      push_space);
1931         old_left_nritems = btrfs_header_nritems(left);
1932         BUG_ON(old_left_nritems == 0);
1933
1934         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1935         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1936                 u32 ioff;
1937
1938                 item = btrfs_item_nr(left, i);
1939                 ioff = btrfs_item_offset(left, item);
1940                 btrfs_set_item_offset(left, item,
1941                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1942         }
1943         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1944
1945         /* fixup right node */
1946         if (push_items > right_nritems) {
1947                 printk("push items %d nr %u\n", push_items, right_nritems);
1948                 WARN_ON(1);
1949         }
1950
1951         if (push_items < right_nritems) {
1952                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1953                                                   leaf_data_end(root, right);
1954                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
1955                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1956                                       btrfs_leaf_data(right) +
1957                                       leaf_data_end(root, right), push_space);
1958
1959                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1960                               btrfs_item_nr_offset(push_items),
1961                              (btrfs_header_nritems(right) - push_items) *
1962                              sizeof(struct btrfs_item));
1963         }
1964         right_nritems -= push_items;
1965         btrfs_set_header_nritems(right, right_nritems);
1966         push_space = BTRFS_LEAF_DATA_SIZE(root);
1967         for (i = 0; i < right_nritems; i++) {
1968                 item = btrfs_item_nr(right, i);
1969                 push_space = push_space - btrfs_item_size(right, item);
1970                 btrfs_set_item_offset(right, item, push_space);
1971         }
1972
1973         btrfs_mark_buffer_dirty(left);
1974         if (right_nritems)
1975                 btrfs_mark_buffer_dirty(right);
1976
1977         btrfs_item_key(right, &disk_key, 0);
1978         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1979         if (wret)
1980                 ret = wret;
1981
1982         /* then fixup the leaf pointer in the path */
1983         if (path->slots[0] < push_items) {
1984                 path->slots[0] += old_left_nritems;
1985                 free_extent_buffer(path->nodes[0]);
1986                 path->nodes[0] = left;
1987                 path->slots[1] -= 1;
1988         } else {
1989                 free_extent_buffer(left);
1990                 path->slots[0] -= push_items;
1991         }
1992         BUG_ON(path->slots[0] < 0);
1993         return ret;
1994 }
1995
1996 /*
1997  * split the path's leaf in two, making sure there is at least data_size
1998  * available for the resulting leaf level of the path.
1999  *
2000  * returns 0 if all went well and < 0 on failure.
2001  */
2002 static noinline int copy_for_split(struct btrfs_trans_handle *trans,
2003                                struct btrfs_root *root,
2004                                struct btrfs_path *path,
2005                                struct extent_buffer *l,
2006                                struct extent_buffer *right,
2007                                int slot, int mid, int nritems)
2008 {
2009         int data_copy_size;
2010         int rt_data_off;
2011         int i;
2012         int ret = 0;
2013         int wret;
2014         struct btrfs_disk_key disk_key;
2015
2016         nritems = nritems - mid;
2017         btrfs_set_header_nritems(right, nritems);
2018         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2019
2020         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2021                            btrfs_item_nr_offset(mid),
2022                            nritems * sizeof(struct btrfs_item));
2023
2024         copy_extent_buffer(right, l,
2025                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2026                      data_copy_size, btrfs_leaf_data(l) +
2027                      leaf_data_end(root, l), data_copy_size);
2028
2029         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2030                       btrfs_item_end_nr(l, mid);
2031
2032         for (i = 0; i < nritems; i++) {
2033                 struct btrfs_item *item = btrfs_item_nr(right, i);
2034                 u32 ioff = btrfs_item_offset(right, item);
2035                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2036         }
2037
2038         btrfs_set_header_nritems(l, mid);
2039         ret = 0;
2040         btrfs_item_key(right, &disk_key, 0);
2041         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2042                           path->slots[1] + 1, 1);
2043         if (wret)
2044                 ret = wret;
2045
2046         btrfs_mark_buffer_dirty(right);
2047         btrfs_mark_buffer_dirty(l);
2048         BUG_ON(path->slots[0] != slot);
2049
2050         if (mid <= slot) {
2051                 free_extent_buffer(path->nodes[0]);
2052                 path->nodes[0] = right;
2053                 path->slots[0] -= mid;
2054                 path->slots[1] += 1;
2055         } else {
2056                 free_extent_buffer(right);
2057         }
2058
2059         BUG_ON(path->slots[0] < 0);
2060
2061         return ret;
2062 }
2063
2064 /*
2065  * split the path's leaf in two, making sure there is at least data_size
2066  * available for the resulting leaf level of the path.
2067  *
2068  * returns 0 if all went well and < 0 on failure.
2069  */
2070 static noinline int split_leaf(struct btrfs_trans_handle *trans,
2071                                struct btrfs_root *root,
2072                                struct btrfs_key *ins_key,
2073                                struct btrfs_path *path, int data_size,
2074                                int extend)
2075 {
2076         struct btrfs_disk_key disk_key;
2077         struct extent_buffer *l;
2078         u32 nritems;
2079         int mid;
2080         int slot;
2081         struct extent_buffer *right;
2082         int ret = 0;
2083         int wret;
2084         int split;
2085         int num_doubles = 0;
2086
2087         /* first try to make some room by pushing left and right */
2088         if (data_size && ins_key->type != BTRFS_DIR_ITEM_KEY) {
2089                 wret = push_leaf_right(trans, root, path, data_size, 0);
2090                 if (wret < 0)
2091                         return wret;
2092                 if (wret) {
2093                         wret = push_leaf_left(trans, root, path, data_size, 0);
2094                         if (wret < 0)
2095                                 return wret;
2096                 }
2097                 l = path->nodes[0];
2098
2099                 /* did the pushes work? */
2100                 if (btrfs_leaf_free_space(root, l) >= data_size)
2101                         return 0;
2102         }
2103
2104         if (!path->nodes[1]) {
2105                 ret = insert_new_root(trans, root, path, 1);
2106                 if (ret)
2107                         return ret;
2108         }
2109 again:
2110         split = 1;
2111         l = path->nodes[0];
2112         slot = path->slots[0];
2113         nritems = btrfs_header_nritems(l);
2114         mid = (nritems + 1) / 2;
2115
2116         if (mid <= slot) {
2117                 if (nritems == 1 ||
2118                     leaf_space_used(l, mid, nritems - mid) + data_size >
2119                         BTRFS_LEAF_DATA_SIZE(root)) {
2120                         if (slot >= nritems) {
2121                                 split = 0;
2122                         } else {
2123                                 mid = slot;
2124                                 if (mid != nritems &&
2125                                     leaf_space_used(l, mid, nritems - mid) +
2126                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2127                                         split = 2;
2128                                 }
2129                         }
2130                 }
2131         } else {
2132                 if (leaf_space_used(l, 0, mid) + data_size >
2133                         BTRFS_LEAF_DATA_SIZE(root)) {
2134                         if (!extend && data_size && slot == 0) {
2135                                 split = 0;
2136                         } else if ((extend || !data_size) && slot == 0) {
2137                                 mid = 1;
2138                         } else {
2139                                 mid = slot;
2140                                 if (mid != nritems &&
2141                                     leaf_space_used(l, mid, nritems - mid) +
2142                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2143                                         split = 2 ;
2144                                 }
2145                         }
2146                 }
2147         }
2148
2149         if (split == 0)
2150                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2151         else
2152                 btrfs_item_key(l, &disk_key, mid);
2153
2154         right = btrfs_alloc_free_block(trans, root, root->leafsize,
2155                                         root->root_key.objectid,
2156                                         &disk_key, 0, l->start, 0);
2157         if (IS_ERR(right)) {
2158                 BUG_ON(1);
2159                 return PTR_ERR(right);
2160         }
2161
2162         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2163         btrfs_set_header_bytenr(right, right->start);
2164         btrfs_set_header_generation(right, trans->transid);
2165         btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
2166         btrfs_set_header_owner(right, root->root_key.objectid);
2167         btrfs_set_header_level(right, 0);
2168         write_extent_buffer(right, root->fs_info->fsid,
2169                             (unsigned long)btrfs_header_fsid(right),
2170                             BTRFS_FSID_SIZE);
2171
2172         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2173                             (unsigned long)btrfs_header_chunk_tree_uuid(right),
2174                             BTRFS_UUID_SIZE);
2175
2176         if (split == 0) {
2177                 if (mid <= slot) {
2178                         btrfs_set_header_nritems(right, 0);
2179                         wret = insert_ptr(trans, root, path,
2180                                           &disk_key, right->start,
2181                                           path->slots[1] + 1, 1);
2182                         if (wret)
2183                                 ret = wret;
2184
2185                         free_extent_buffer(path->nodes[0]);
2186                         path->nodes[0] = right;
2187                         path->slots[0] = 0;
2188                         path->slots[1] += 1;
2189                 } else {
2190                         btrfs_set_header_nritems(right, 0);
2191                         wret = insert_ptr(trans, root, path,
2192                                           &disk_key,
2193                                           right->start,
2194                                           path->slots[1], 1);
2195                         if (wret)
2196                                 ret = wret;
2197                         free_extent_buffer(path->nodes[0]);
2198                         path->nodes[0] = right;
2199                         path->slots[0] = 0;
2200                         if (path->slots[1] == 0) {
2201                                 wret = fixup_low_keys(trans, root,
2202                                                 path, &disk_key, 1);
2203                                 if (wret)
2204                                         ret = wret;
2205                         }
2206                 }
2207                 btrfs_mark_buffer_dirty(right);
2208                 return ret;
2209         }
2210
2211         ret = copy_for_split(trans, root, path, l, right, slot, mid, nritems);
2212         BUG_ON(ret);
2213
2214         if (split == 2) {
2215                 BUG_ON(num_doubles != 0);
2216                 num_doubles++;
2217                 goto again;
2218         }
2219
2220         return ret;
2221 }
2222
2223 /*
2224  * This function splits a single item into two items,
2225  * giving 'new_key' to the new item and splitting the
2226  * old one at split_offset (from the start of the item).
2227  *
2228  * The path may be released by this operation.  After
2229  * the split, the path is pointing to the old item.  The
2230  * new item is going to be in the same node as the old one.
2231  *
2232  * Note, the item being split must be smaller enough to live alone on
2233  * a tree block with room for one extra struct btrfs_item
2234  *
2235  * This allows us to split the item in place, keeping a lock on the
2236  * leaf the entire time.
2237  */
2238 int btrfs_split_item(struct btrfs_trans_handle *trans,
2239                      struct btrfs_root *root,
2240                      struct btrfs_path *path,
2241                      struct btrfs_key *new_key,
2242                      unsigned long split_offset)
2243 {
2244         u32 item_size;
2245         struct extent_buffer *leaf;
2246         struct btrfs_key orig_key;
2247         struct btrfs_item *item;
2248         struct btrfs_item *new_item;
2249         int ret = 0;
2250         int slot;
2251         u32 nritems;
2252         u32 orig_offset;
2253         struct btrfs_disk_key disk_key;
2254         char *buf;
2255
2256         leaf = path->nodes[0];
2257         btrfs_item_key_to_cpu(leaf, &orig_key, path->slots[0]);
2258         if (btrfs_leaf_free_space(root, leaf) >= sizeof(struct btrfs_item))
2259                 goto split;
2260
2261         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2262         btrfs_release_path(root, path);
2263
2264         path->search_for_split = 1;
2265
2266         ret = btrfs_search_slot(trans, root, &orig_key, path, 0, 1);
2267         path->search_for_split = 0;
2268
2269         /* if our item isn't there or got smaller, return now */
2270         if (ret != 0 || item_size != btrfs_item_size_nr(path->nodes[0],
2271                                                         path->slots[0])) {
2272                 return -EAGAIN;
2273         }
2274
2275         ret = split_leaf(trans, root, &orig_key, path, 0, 0);
2276         BUG_ON(ret);
2277
2278         BUG_ON(btrfs_leaf_free_space(root, leaf) < sizeof(struct btrfs_item));
2279         leaf = path->nodes[0];
2280
2281 split:
2282         item = btrfs_item_nr(leaf, path->slots[0]);
2283         orig_offset = btrfs_item_offset(leaf, item);
2284         item_size = btrfs_item_size(leaf, item);
2285
2286
2287         buf = kmalloc(item_size, GFP_NOFS);
2288         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
2289                             path->slots[0]), item_size);
2290         slot = path->slots[0] + 1;
2291         leaf = path->nodes[0];
2292
2293         nritems = btrfs_header_nritems(leaf);
2294
2295         if (slot != nritems) {
2296                 /* shift the items */
2297                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
2298                               btrfs_item_nr_offset(slot),
2299                               (nritems - slot) * sizeof(struct btrfs_item));
2300
2301         }
2302
2303         btrfs_cpu_key_to_disk(&disk_key, new_key);
2304         btrfs_set_item_key(leaf, &disk_key, slot);
2305
2306         new_item = btrfs_item_nr(leaf, slot);
2307
2308         btrfs_set_item_offset(leaf, new_item, orig_offset);
2309         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
2310
2311         btrfs_set_item_offset(leaf, item,
2312                               orig_offset + item_size - split_offset);
2313         btrfs_set_item_size(leaf, item, split_offset);
2314
2315         btrfs_set_header_nritems(leaf, nritems + 1);
2316
2317         /* write the data for the start of the original item */
2318         write_extent_buffer(leaf, buf,
2319                             btrfs_item_ptr_offset(leaf, path->slots[0]),
2320                             split_offset);
2321
2322         /* write the data for the new item */
2323         write_extent_buffer(leaf, buf + split_offset,
2324                             btrfs_item_ptr_offset(leaf, slot),
2325                             item_size - split_offset);
2326         btrfs_mark_buffer_dirty(leaf);
2327
2328         ret = 0;
2329         if (btrfs_leaf_free_space(root, leaf) < 0) {
2330                 btrfs_print_leaf(root, leaf);
2331                 BUG();
2332         }
2333         kfree(buf);
2334         return ret;
2335 }
2336
2337 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2338                         struct btrfs_root *root,
2339                         struct btrfs_path *path,
2340                         u32 new_size, int from_end)
2341 {
2342         int ret = 0;
2343         int slot;
2344         struct extent_buffer *leaf;
2345         struct btrfs_item *item;
2346         u32 nritems;
2347         unsigned int data_end;
2348         unsigned int old_data_start;
2349         unsigned int old_size;
2350         unsigned int size_diff;
2351         int i;
2352
2353         leaf = path->nodes[0];
2354         slot = path->slots[0];
2355
2356         old_size = btrfs_item_size_nr(leaf, slot);
2357         if (old_size == new_size)
2358                 return 0;
2359
2360         nritems = btrfs_header_nritems(leaf);
2361         data_end = leaf_data_end(root, leaf);
2362
2363         old_data_start = btrfs_item_offset_nr(leaf, slot);
2364
2365         size_diff = old_size - new_size;
2366
2367         BUG_ON(slot < 0);
2368         BUG_ON(slot >= nritems);
2369
2370         /*
2371          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2372          */
2373         /* first correct the data pointers */
2374         for (i = slot; i < nritems; i++) {
2375                 u32 ioff;
2376                 item = btrfs_item_nr(leaf, i);
2377                 ioff = btrfs_item_offset(leaf, item);
2378                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2379         }
2380
2381         /* shift the data */
2382         if (from_end) {
2383                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2384                               data_end + size_diff, btrfs_leaf_data(leaf) +
2385                               data_end, old_data_start + new_size - data_end);
2386         } else {
2387                 struct btrfs_disk_key disk_key;
2388                 u64 offset;
2389
2390                 btrfs_item_key(leaf, &disk_key, slot);
2391
2392                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2393                         unsigned long ptr;
2394                         struct btrfs_file_extent_item *fi;
2395
2396                         fi = btrfs_item_ptr(leaf, slot,
2397                                             struct btrfs_file_extent_item);
2398                         fi = (struct btrfs_file_extent_item *)(
2399                              (unsigned long)fi - size_diff);
2400
2401                         if (btrfs_file_extent_type(leaf, fi) ==
2402                             BTRFS_FILE_EXTENT_INLINE) {
2403                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2404                                 memmove_extent_buffer(leaf, ptr,
2405                                         (unsigned long)fi,
2406                                         offsetof(struct btrfs_file_extent_item,
2407                                                  disk_bytenr));
2408                         }
2409                 }
2410
2411                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2412                               data_end + size_diff, btrfs_leaf_data(leaf) +
2413                               data_end, old_data_start - data_end);
2414
2415                 offset = btrfs_disk_key_offset(&disk_key);
2416                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2417                 btrfs_set_item_key(leaf, &disk_key, slot);
2418                 if (slot == 0)
2419                         fixup_low_keys(trans, root, path, &disk_key, 1);
2420         }
2421
2422         item = btrfs_item_nr(leaf, slot);
2423         btrfs_set_item_size(leaf, item, new_size);
2424         btrfs_mark_buffer_dirty(leaf);
2425
2426         ret = 0;
2427         if (btrfs_leaf_free_space(root, leaf) < 0) {
2428                 btrfs_print_leaf(root, leaf);
2429                 BUG();
2430         }
2431         return ret;
2432 }
2433
2434 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2435                       struct btrfs_root *root, struct btrfs_path *path,
2436                       u32 data_size)
2437 {
2438         int ret = 0;
2439         int slot;
2440         struct extent_buffer *leaf;
2441         struct btrfs_item *item;
2442         u32 nritems;
2443         unsigned int data_end;
2444         unsigned int old_data;
2445         unsigned int old_size;
2446         int i;
2447
2448         leaf = path->nodes[0];
2449
2450         nritems = btrfs_header_nritems(leaf);
2451         data_end = leaf_data_end(root, leaf);
2452
2453         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2454                 btrfs_print_leaf(root, leaf);
2455                 BUG();
2456         }
2457         slot = path->slots[0];
2458         old_data = btrfs_item_end_nr(leaf, slot);
2459
2460         BUG_ON(slot < 0);
2461         if (slot >= nritems) {
2462                 btrfs_print_leaf(root, leaf);
2463                 printk("slot %d too large, nritems %d\n", slot, nritems);
2464                 BUG_ON(1);
2465         }
2466
2467         /*
2468          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2469          */
2470         /* first correct the data pointers */
2471         for (i = slot; i < nritems; i++) {
2472                 u32 ioff;
2473                 item = btrfs_item_nr(leaf, i);
2474                 ioff = btrfs_item_offset(leaf, item);
2475                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2476         }
2477
2478         /* shift the data */
2479         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2480                       data_end - data_size, btrfs_leaf_data(leaf) +
2481                       data_end, old_data - data_end);
2482
2483         data_end = old_data;
2484         old_size = btrfs_item_size_nr(leaf, slot);
2485         item = btrfs_item_nr(leaf, slot);
2486         btrfs_set_item_size(leaf, item, old_size + data_size);
2487         btrfs_mark_buffer_dirty(leaf);
2488
2489         ret = 0;
2490         if (btrfs_leaf_free_space(root, leaf) < 0) {
2491                 btrfs_print_leaf(root, leaf);
2492                 BUG();
2493         }
2494         return ret;
2495 }
2496
2497 /*
2498  * Given a key and some data, insert an item into the tree.
2499  * This does all the path init required, making room in the tree if needed.
2500  */
2501 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2502                             struct btrfs_root *root,
2503                             struct btrfs_path *path,
2504                             struct btrfs_key *cpu_key, u32 *data_size,
2505                             int nr)
2506 {
2507         struct extent_buffer *leaf;
2508         struct btrfs_item *item;
2509         int ret = 0;
2510         int slot;
2511         int i;
2512         u32 nritems;
2513         u32 total_size = 0;
2514         u32 total_data = 0;
2515         unsigned int data_end;
2516         struct btrfs_disk_key disk_key;
2517
2518         for (i = 0; i < nr; i++) {
2519                 total_data += data_size[i];
2520         }
2521
2522         /* create a root if there isn't one */
2523         if (!root->node)
2524                 BUG();
2525
2526         total_size = total_data + nr * sizeof(struct btrfs_item);
2527         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2528         if (ret == 0) {
2529                 return -EEXIST;
2530         }
2531         if (ret < 0)
2532                 goto out;
2533
2534         leaf = path->nodes[0];
2535
2536         nritems = btrfs_header_nritems(leaf);
2537         data_end = leaf_data_end(root, leaf);
2538
2539         if (btrfs_leaf_free_space(root, leaf) < total_size) {
2540                 btrfs_print_leaf(root, leaf);
2541                 printk("not enough freespace need %u have %d\n",
2542                        total_size, btrfs_leaf_free_space(root, leaf));
2543                 BUG();
2544         }
2545
2546         slot = path->slots[0];
2547         BUG_ON(slot < 0);
2548
2549         if (slot != nritems) {
2550                 int i;
2551                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2552
2553                 if (old_data < data_end) {
2554                         btrfs_print_leaf(root, leaf);
2555                         printk("slot %d old_data %d data_end %d\n",
2556                                slot, old_data, data_end);
2557                         BUG_ON(1);
2558                 }
2559                 /*
2560                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2561                  */
2562                 /* first correct the data pointers */
2563                 for (i = slot; i < nritems; i++) {
2564                         u32 ioff;
2565
2566                         item = btrfs_item_nr(leaf, i);
2567                         ioff = btrfs_item_offset(leaf, item);
2568                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2569                 }
2570
2571                 /* shift the items */
2572                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2573                               btrfs_item_nr_offset(slot),
2574                               (nritems - slot) * sizeof(struct btrfs_item));
2575
2576                 /* shift the data */
2577                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2578                               data_end - total_data, btrfs_leaf_data(leaf) +
2579                               data_end, old_data - data_end);
2580                 data_end = old_data;
2581         }
2582
2583         /* setup the item for the new data */
2584         for (i = 0; i < nr; i++) {
2585                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2586                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2587                 item = btrfs_item_nr(leaf, slot + i);
2588                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2589                 data_end -= data_size[i];
2590                 btrfs_set_item_size(leaf, item, data_size[i]);
2591         }
2592         btrfs_set_header_nritems(leaf, nritems + nr);
2593         btrfs_mark_buffer_dirty(leaf);
2594
2595         ret = 0;
2596         if (slot == 0) {
2597                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2598                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2599         }
2600
2601         if (btrfs_leaf_free_space(root, leaf) < 0) {
2602                 btrfs_print_leaf(root, leaf);
2603                 BUG();
2604         }
2605
2606 out:
2607         return ret;
2608 }
2609
2610 /*
2611  * Given a key and some data, insert an item into the tree.
2612  * This does all the path init required, making room in the tree if needed.
2613  */
2614 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2615                       *root, struct btrfs_key *cpu_key, void *data, u32
2616                       data_size)
2617 {
2618         int ret = 0;
2619         struct btrfs_path *path;
2620         struct extent_buffer *leaf;
2621         unsigned long ptr;
2622
2623         path = btrfs_alloc_path();
2624         BUG_ON(!path);
2625         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2626         if (!ret) {
2627                 leaf = path->nodes[0];
2628                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2629                 write_extent_buffer(leaf, data, ptr, data_size);
2630                 btrfs_mark_buffer_dirty(leaf);
2631         }
2632         btrfs_free_path(path);
2633         return ret;
2634 }
2635
2636 /*
2637  * delete the pointer from a given node.
2638  *
2639  * If the delete empties a node, the node is removed from the tree,
2640  * continuing all the way the root if required.  The root is converted into
2641  * a leaf if all the nodes are emptied.
2642  */
2643 int btrfs_del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2644                    struct btrfs_path *path, int level, int slot)
2645 {
2646         struct extent_buffer *parent = path->nodes[level];
2647         u32 nritems;
2648         int ret = 0;
2649         int wret;
2650
2651         nritems = btrfs_header_nritems(parent);
2652         if (slot != nritems -1) {
2653                 memmove_extent_buffer(parent,
2654                               btrfs_node_key_ptr_offset(slot),
2655                               btrfs_node_key_ptr_offset(slot + 1),
2656                               sizeof(struct btrfs_key_ptr) *
2657                               (nritems - slot - 1));
2658         }
2659         nritems--;
2660         btrfs_set_header_nritems(parent, nritems);
2661         if (nritems == 0 && parent == root->node) {
2662                 BUG_ON(btrfs_header_level(root->node) != 1);
2663                 /* just turn the root into a leaf and break */
2664                 btrfs_set_header_level(root->node, 0);
2665         } else if (slot == 0) {
2666                 struct btrfs_disk_key disk_key;
2667
2668                 btrfs_node_key(parent, &disk_key, 0);
2669                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2670                 if (wret)
2671                         ret = wret;
2672         }
2673         btrfs_mark_buffer_dirty(parent);
2674         return ret;
2675 }
2676
2677 /*
2678  * a helper function to delete the leaf pointed to by path->slots[1] and
2679  * path->nodes[1].
2680  *
2681  * This deletes the pointer in path->nodes[1] and frees the leaf
2682  * block extent.  zero is returned if it all worked out, < 0 otherwise.
2683  *
2684  * The path must have already been setup for deleting the leaf, including
2685  * all the proper balancing.  path->nodes[1] must be locked.
2686  */
2687 static noinline int btrfs_del_leaf(struct btrfs_trans_handle *trans,
2688                                    struct btrfs_root *root,
2689                                    struct btrfs_path *path,
2690                                    struct extent_buffer *leaf)
2691 {
2692         int ret;
2693
2694         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
2695         ret = btrfs_del_ptr(trans, root, path, 1, path->slots[1]);
2696         if (ret)
2697                 return ret;
2698
2699         ret = btrfs_free_extent(trans, root, leaf->start, leaf->len,
2700                                 0, root->root_key.objectid, 0, 0);
2701         return ret;
2702 }
2703
2704 /*
2705  * delete the item at the leaf level in path.  If that empties
2706  * the leaf, remove it from the tree
2707  */
2708 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2709                     struct btrfs_path *path, int slot, int nr)
2710 {
2711         struct extent_buffer *leaf;
2712         struct btrfs_item *item;
2713         int last_off;
2714         int dsize = 0;
2715         int ret = 0;
2716         int wret;
2717         int i;
2718         u32 nritems;
2719
2720         leaf = path->nodes[0];
2721         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2722
2723         for (i = 0; i < nr; i++)
2724                 dsize += btrfs_item_size_nr(leaf, slot + i);
2725
2726         nritems = btrfs_header_nritems(leaf);
2727
2728         if (slot + nr != nritems) {
2729                 int i;
2730                 int data_end = leaf_data_end(root, leaf);
2731
2732                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2733                               data_end + dsize,
2734                               btrfs_leaf_data(leaf) + data_end,
2735                               last_off - data_end);
2736
2737                 for (i = slot + nr; i < nritems; i++) {
2738                         u32 ioff;
2739
2740                         item = btrfs_item_nr(leaf, i);
2741                         ioff = btrfs_item_offset(leaf, item);
2742                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2743                 }
2744
2745                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2746                               btrfs_item_nr_offset(slot + nr),
2747                               sizeof(struct btrfs_item) *
2748                               (nritems - slot - nr));
2749         }
2750         btrfs_set_header_nritems(leaf, nritems - nr);
2751         nritems -= nr;
2752
2753         /* delete the leaf if we've emptied it */
2754         if (nritems == 0) {
2755                 if (leaf == root->node) {
2756                         btrfs_set_header_level(leaf, 0);
2757                 } else {
2758                         clean_tree_block(trans, root, leaf);
2759                         wait_on_tree_block_writeback(root, leaf);
2760
2761                         wret = btrfs_del_leaf(trans, root, path, leaf);
2762                         BUG_ON(ret);
2763                         if (wret)
2764                                 ret = wret;
2765                 }
2766         } else {
2767                 int used = leaf_space_used(leaf, 0, nritems);
2768                 if (slot == 0) {
2769                         struct btrfs_disk_key disk_key;
2770
2771                         btrfs_item_key(leaf, &disk_key, 0);
2772                         wret = fixup_low_keys(trans, root, path,
2773                                               &disk_key, 1);
2774                         if (wret)
2775                                 ret = wret;
2776                 }
2777
2778                 /* delete the leaf if it is mostly empty */
2779                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2780                         /* push_leaf_left fixes the path.
2781                          * make sure the path still points to our leaf
2782                          * for possible call to del_ptr below
2783                          */
2784                         slot = path->slots[1];
2785                         extent_buffer_get(leaf);
2786
2787                         wret = push_leaf_left(trans, root, path, 1, 1);
2788                         if (wret < 0 && wret != -ENOSPC)
2789                                 ret = wret;
2790
2791                         if (path->nodes[0] == leaf &&
2792                             btrfs_header_nritems(leaf)) {
2793                                 wret = push_leaf_right(trans, root, path, 1, 1);
2794                                 if (wret < 0 && wret != -ENOSPC)
2795                                         ret = wret;
2796                         }
2797
2798                         if (btrfs_header_nritems(leaf) == 0) {
2799                                 clean_tree_block(trans, root, leaf);
2800                                 wait_on_tree_block_writeback(root, leaf);
2801
2802                                 path->slots[1] = slot;
2803                                 ret = btrfs_del_leaf(trans, root, path, leaf);
2804                                 BUG_ON(ret);
2805                                 free_extent_buffer(leaf);
2806
2807                         } else {
2808                                 btrfs_mark_buffer_dirty(leaf);
2809                                 free_extent_buffer(leaf);
2810                         }
2811                 } else {
2812                         btrfs_mark_buffer_dirty(leaf);
2813                 }
2814         }
2815         return ret;
2816 }
2817
2818 /*
2819  * walk up the tree as far as required to find the previous leaf.
2820  * returns 0 if it found something or 1 if there are no lesser leaves.
2821  * returns < 0 on io errors.
2822  */
2823 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2824 {
2825         int slot;
2826         int level = 1;
2827         struct extent_buffer *c;
2828         struct extent_buffer *next = NULL;
2829
2830         while(level < BTRFS_MAX_LEVEL) {
2831                 if (!path->nodes[level])
2832                         return 1;
2833
2834                 slot = path->slots[level];
2835                 c = path->nodes[level];
2836                 if (slot == 0) {
2837                         level++;
2838                         if (level == BTRFS_MAX_LEVEL)
2839                                 return 1;
2840                         continue;
2841                 }
2842                 slot--;
2843
2844                 next = read_node_slot(root, c, slot);
2845                 break;
2846         }
2847         path->slots[level] = slot;
2848         while(1) {
2849                 level--;
2850                 c = path->nodes[level];
2851                 free_extent_buffer(c);
2852                 slot = btrfs_header_nritems(next);
2853                 if (slot != 0)
2854                         slot--;
2855                 path->nodes[level] = next;
2856                 path->slots[level] = slot;
2857                 if (!level)
2858                         break;
2859                 next = read_node_slot(root, next, slot);
2860         }
2861         return 0;
2862 }
2863
2864 /*
2865  * walk up the tree as far as required to find the next leaf.
2866  * returns 0 if it found something or 1 if there are no greater leaves.
2867  * returns < 0 on io errors.
2868  */
2869 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2870 {
2871         int slot;
2872         int level = 1;
2873         struct extent_buffer *c;
2874         struct extent_buffer *next = NULL;
2875
2876         while(level < BTRFS_MAX_LEVEL) {
2877                 if (!path->nodes[level])
2878                         return 1;
2879
2880                 slot = path->slots[level] + 1;
2881                 c = path->nodes[level];
2882                 if (slot >= btrfs_header_nritems(c)) {
2883                         level++;
2884                         if (level == BTRFS_MAX_LEVEL)
2885                                 return 1;
2886                         continue;
2887                 }
2888
2889                 if (path->reada)
2890                         reada_for_search(root, path, level, slot, 0);
2891
2892                 next = read_node_slot(root, c, slot);
2893                 if (!next)
2894                         return -EIO;
2895                 break;
2896         }
2897         path->slots[level] = slot;
2898         while(1) {
2899                 level--;
2900                 c = path->nodes[level];
2901                 free_extent_buffer(c);
2902                 path->nodes[level] = next;
2903                 path->slots[level] = 0;
2904                 if (!level)
2905                         break;
2906                 if (path->reada)
2907                         reada_for_search(root, path, level, 0, 0);
2908                 next = read_node_slot(root, next, 0);
2909                 if (!next)
2910                         return -EIO;
2911         }
2912         return 0;
2913 }
2914
2915 int btrfs_previous_item(struct btrfs_root *root,
2916                         struct btrfs_path *path, u64 min_objectid,
2917                         int type)
2918 {
2919         struct btrfs_key found_key;
2920         struct extent_buffer *leaf;
2921         int ret;
2922
2923         while(1) {
2924                 if (path->slots[0] == 0) {
2925                         ret = btrfs_prev_leaf(root, path);
2926                         if (ret != 0)
2927                                 return ret;
2928                 } else {
2929                         path->slots[0]--;
2930                 }
2931                 leaf = path->nodes[0];
2932                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2933                 if (found_key.type == type)
2934                         return 0;
2935         }
2936         return 1;
2937 }
2938