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 static 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 int btrfs_fsck_reinit_root(struct btrfs_trans_handle *trans,
 141                       struct btrfs_root *root)
 142 {
 143         struct extent_buffer *c;
 144         struct extent_buffer *old = root->node;
 145         int level;
 146         struct btrfs_disk_key disk_key = {0,0,0};
 147
 148         level = 0;
 149
 150         c = btrfs_alloc_free_block(trans, root,
 151                                    btrfs_level_size(root, 0),
 152                                    root->root_key.objectid,
 153                                    &disk_key, level, 0, 0);
 154         if (IS_ERR(c)) {
 155                 c = old;
 156                 extent_buffer_get(c);
 157         }
 158
 159         memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
 160         btrfs_set_header_level(c, level);
 161         btrfs_set_header_bytenr(c, c->start);
 162         btrfs_set_header_generation(c, trans->transid);
 163         btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
 164         btrfs_set_header_owner(c, root->root_key.objectid);
 165
 166         write_extent_buffer(c, root->fs_info->fsid,
 167                             (unsigned long)btrfs_header_fsid(c),
 168                             BTRFS_FSID_SIZE);
 169
 170         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
 171                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
 172                             BTRFS_UUID_SIZE);
 173
 174         btrfs_mark_buffer_dirty(c);
 175
 176         free_extent_buffer(old);
 177         root->node = c;
 178         add_root_to_dirty_list(root);
 179         return 0;
 180 }
 181
 182 /*
 183  * check if the tree block can be shared by multiple trees
 184  */
 185 int btrfs_block_can_be_shared(struct btrfs_root *root,
 186                               struct extent_buffer *buf)
 187 {
 188         /*
 189          * Tree blocks not in refernece counted trees and tree roots
 190          * are never shared. If a block was allocated after the last
 191          * snapshot and the block was not allocated by tree relocation,
 192          * we know the block is not shared.
 193          */
 194         if (root->ref_cows &&
 195             buf != root->node && buf != root->commit_root &&
 196             (btrfs_header_generation(buf) <=
 197              btrfs_root_last_snapshot(&root->root_item) ||
 198              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 199                 return 1;
 200 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 201         if (root->ref_cows &&
 202             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 203                 return 1;
 204 #endif
 205         return 0;
 206 }
 207
 208 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
 209                                        struct btrfs_root *root,
 210                                        struct extent_buffer *buf,
 211                                        struct extent_buffer *cow)
 212 {
 213         u64 refs;
 214         u64 owner;
 215         u64 flags;
 216         u64 new_flags = 0;
 217         int ret;
 218
 219         /*
 220          * Backrefs update rules:
 221          *
 222          * Always use full backrefs for extent pointers in tree block
 223          * allocated by tree relocation.
 224          *
 225          * If a shared tree block is no longer referenced by its owner
 226          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
 227          * use full backrefs for extent pointers in tree block.
 228          *
 229          * If a tree block is been relocating
 230          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
 231          * use full backrefs for extent pointers in tree block.
 232          * The reason for this is some operations (such as drop tree)
 233          * are only allowed for blocks use full backrefs.
 234          */
 235
 236         if (btrfs_block_can_be_shared(root, buf)) {
 237                 ret = btrfs_lookup_extent_info(trans, root, buf->start,
 238                                                btrfs_header_level(buf), 1,
 239                                                &refs, &flags);
 240                 BUG_ON(ret);
 241                 BUG_ON(refs == 0);
 242         } else {
 243                 refs = 1;
 244                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 245                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 246                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 247                 else
 248                         flags = 0;
 249         }
 250
 251         owner = btrfs_header_owner(buf);
 252         BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) &&
 253                owner == BTRFS_TREE_RELOC_OBJECTID);
 254
 255         if (refs > 1) {
 256                 if ((owner == root->root_key.objectid ||
 257                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
 258                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
 259                         ret = btrfs_inc_ref(trans, root, buf, 1);
 260                         BUG_ON(ret);
 261
 262                         if (root->root_key.objectid ==
 263                             BTRFS_TREE_RELOC_OBJECTID) {
 264                                 ret = btrfs_dec_ref(trans, root, buf, 0);
 265                                 BUG_ON(ret);
 266                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 267                                 BUG_ON(ret);
 268                         }
 269                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
 270                 } else {
 271
 272                         if (root->root_key.objectid ==
 273                             BTRFS_TREE_RELOC_OBJECTID)
 274                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 275                         else
 276                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 277                         BUG_ON(ret);
 278                 }
 279                 if (new_flags != 0) {
 280                         ret = btrfs_set_block_flags(trans, root, buf->start,
 281                                                     btrfs_header_level(buf),
 282                                                     new_flags);
 283                         BUG_ON(ret);
 284                 }
 285         } else {
 286                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
 287                         if (root->root_key.objectid ==
 288                             BTRFS_TREE_RELOC_OBJECTID)
 289                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 290                         else
 291                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 292                         BUG_ON(ret);
 293                         ret = btrfs_dec_ref(trans, root, buf, 1);
 294                         BUG_ON(ret);
 295                 }
 296                 clean_tree_block(trans, root, buf);
 297         }
 298         return 0;
 299 }
 300
 301 int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 302                              struct btrfs_root *root,
 303                              struct extent_buffer *buf,
 304                              struct extent_buffer *parent, int parent_slot,
 305                              struct extent_buffer **cow_ret,
 306                              u64 search_start, u64 empty_size)
 307 {
 308         struct extent_buffer *cow;
 309         struct btrfs_disk_key disk_key;
 310         int level;
 311
 312         WARN_ON(root->ref_cows && trans->transid !=
 313                 root->fs_info->running_transaction->transid);
 314         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 315
 316         level = btrfs_header_level(buf);
 317
 318         if (level == 0)
 319                 btrfs_item_key(buf, &disk_key, 0);
 320         else
 321                 btrfs_node_key(buf, &disk_key, 0);
 322
 323         cow = btrfs_alloc_free_block(trans, root, buf->len,
 324                                      root->root_key.objectid, &disk_key,
 325                                      level, search_start, empty_size);
 326         if (IS_ERR(cow))
 327                 return PTR_ERR(cow);
 328
 329         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 330         btrfs_set_header_bytenr(cow, cow->start);
 331         btrfs_set_header_generation(cow, trans->transid);
 332         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 333         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 334                                      BTRFS_HEADER_FLAG_RELOC);
 335         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 336                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 337         else
 338                 btrfs_set_header_owner(cow, root->root_key.objectid);
 339
 340         write_extent_buffer(cow, root->fs_info->fsid,
 341                             (unsigned long)btrfs_header_fsid(cow),
 342                             BTRFS_FSID_SIZE);
 343
 344         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 345
 346         update_ref_for_cow(trans, root, buf, cow);
 347
 348         if (buf == root->node) {
 349                 root->node = cow;
 350                 extent_buffer_get(cow);
 351
 352                 btrfs_free_extent(trans, root, buf->start, buf->len,
 353                                   0, root->root_key.objectid, level, 0);
 354                 free_extent_buffer(buf);
 355                 add_root_to_dirty_list(root);
 356         } else {
 357                 btrfs_set_node_blockptr(parent, parent_slot,
 358                                         cow->start);
 359                 WARN_ON(trans->transid == 0);
 360                 btrfs_set_node_ptr_generation(parent, parent_slot,
 361                                               trans->transid);
 362                 btrfs_mark_buffer_dirty(parent);
 363                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 364
 365                 btrfs_free_extent(trans, root, buf->start, buf->len,
 366                                   0, root->root_key.objectid, level, 1);
 367         }
 368         free_extent_buffer(buf);
 369         btrfs_mark_buffer_dirty(cow);
 370         *cow_ret = cow;
 371         return 0;
 372 }
 373
 374 static inline int should_cow_block(struct btrfs_trans_handle *trans,
 375                                    struct btrfs_root *root,
 376                                    struct extent_buffer *buf)
 377 {
 378         if (btrfs_header_generation(buf) == trans->transid &&
 379             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
 380             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
 381               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 382                 return 0;
 383         return 1;
 384 }
 385
 386 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 387                     struct btrfs_root *root, struct extent_buffer *buf,
 388                     struct extent_buffer *parent, int parent_slot,
 389                     struct extent_buffer **cow_ret)
 390 {
 391         u64 search_start;
 392         int ret;
 393         /*
 394         if (trans->transaction != root->fs_info->running_transaction) {
 395                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 396                        root->fs_info->running_transaction->transid);
 397                 WARN_ON(1);
 398         }
 399         */
 400         if (trans->transid != root->fs_info->generation) {
 401                 printk(KERN_CRIT "trans %llu running %llu\n",
 402                         (unsigned long long)trans->transid,
 403                         (unsigned long long)root->fs_info->generation);
 404                 WARN_ON(1);
 405         }
 406         if (!should_cow_block(trans, root, buf)) {
 407                 *cow_ret = buf;
 408                 return 0;
 409         }
 410
 411         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
 412         ret = __btrfs_cow_block(trans, root, buf, parent,
 413                                  parent_slot, cow_ret, search_start, 0);
 414         return ret;
 415 }
 416
 417 /*
 418 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 419 {
 420         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 421                 return 1;
 422         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 423                 return 1;
 424         return 0;
 425 }
 426 */
 427
 428 /*
 429  * compare two keys in a memcmp fashion
 430  */
 431 int btrfs_comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 432 {
 433         struct btrfs_key k1;
 434
 435         btrfs_disk_key_to_cpu(&k1, disk);
 436
 437         if (k1.objectid > k2->objectid)
 438                 return 1;
 439         if (k1.objectid < k2->objectid)
 440                 return -1;
 441         if (k1.type > k2->type)
 442                 return 1;
 443         if (k1.type < k2->type)
 444                 return -1;
 445         if (k1.offset > k2->offset)
 446                 return 1;
 447         if (k1.offset < k2->offset)
 448                 return -1;
 449         return 0;
 450 }
 451
 452
 453 #if 0
 454 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 455                        struct btrfs_root *root, struct extent_buffer *parent,
 456                        int start_slot, int cache_only, u64 *last_ret,
 457                        struct btrfs_key *progress)
 458 {
 459         struct extent_buffer *cur;
 460         struct extent_buffer *tmp;
 461         u64 blocknr;
 462         u64 gen;
 463         u64 search_start = *last_ret;
 464         u64 last_block = 0;
 465         u64 other;
 466         u32 parent_nritems;
 467         int end_slot;
 468         int i;
 469         int err = 0;
 470         int parent_level;
 471         int uptodate;
 472         u32 blocksize;
 473         int progress_passed = 0;
 474         struct btrfs_disk_key disk_key;
 475
 476         parent_level = btrfs_header_level(parent);
 477         if (cache_only && parent_level != 1)
 478                 return 0;
 479
 480         if (trans->transaction != root->fs_info->running_transaction) {
 481                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 482                        root->fs_info->running_transaction->transid);
 483                 WARN_ON(1);
 484         }
 485         if (trans->transid != root->fs_info->generation) {
 486                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 487                        root->fs_info->generation);
 488                 WARN_ON(1);
 489         }
 490
 491         parent_nritems = btrfs_header_nritems(parent);
 492         blocksize = btrfs_level_size(root, parent_level - 1);
 493         end_slot = parent_nritems;
 494
 495         if (parent_nritems == 1)
 496                 return 0;
 497
 498         for (i = start_slot; i < end_slot; i++) {
 499                 int close = 1;
 500
 501                 if (!parent->map_token) {
 502                         map_extent_buffer(parent,
 503                                         btrfs_node_key_ptr_offset(i),
 504                                         sizeof(struct btrfs_key_ptr),
 505                                         &parent->map_token, &parent->kaddr,
 506                                         &parent->map_start, &parent->map_len,
 507                                         KM_USER1);
 508                 }
 509                 btrfs_node_key(parent, &disk_key, i);
 510                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
 511                         continue;
 512
 513                 progress_passed = 1;
 514                 blocknr = btrfs_node_blockptr(parent, i);
 515                 gen = btrfs_node_ptr_generation(parent, i);
 516                 if (last_block == 0)
 517                         last_block = blocknr;
 518
 519                 if (i > 0) {
 520                         other = btrfs_node_blockptr(parent, i - 1);
 521                         close = close_blocks(blocknr, other, blocksize);
 522                 }
 523                 if (close && i < end_slot - 2) {
 524                         other = btrfs_node_blockptr(parent, i + 1);
 525                         close = close_blocks(blocknr, other, blocksize);
 526                 }
 527                 if (close) {
 528                         last_block = blocknr;
 529                         continue;
 530                 }
 531                 if (parent->map_token) {
 532                         unmap_extent_buffer(parent, parent->map_token,
 533                                             KM_USER1);
 534                         parent->map_token = NULL;
 535                 }
 536
 537                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
 538                 if (cur)
 539                         uptodate = btrfs_buffer_uptodate(cur, gen);
 540                 else
 541                         uptodate = 0;
 542                 if (!cur || !uptodate) {
 543                         if (cache_only) {
 544                                 free_extent_buffer(cur);
 545                                 continue;
 546                         }
 547                         if (!cur) {
 548                                 cur = read_tree_block(root, blocknr,
 549                                                          blocksize, gen);
 550                         } else if (!uptodate) {
 551                                 btrfs_read_buffer(cur, gen);
 552                         }
 553                 }
 554                 if (search_start == 0)
 555                         search_start = last_block;
 556
 557                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 558                                         &tmp, search_start,
 559                                         min(16 * blocksize,
 560                                             (end_slot - i) * blocksize));
 561                 if (err) {
 562                         free_extent_buffer(cur);
 563                         break;
 564                 }
 565                 search_start = tmp->start;
 566                 last_block = tmp->start;
 567                 *last_ret = search_start;
 568                 if (parent_level == 1)
 569                         btrfs_clear_buffer_defrag(tmp);
 570                 free_extent_buffer(tmp);
 571         }
 572         if (parent->map_token) {
 573                 unmap_extent_buffer(parent, parent->map_token,
 574                                     KM_USER1);
 575                 parent->map_token = NULL;
 576         }
 577         return err;
 578 }
 579 #endif
 580
 581 /*
 582  * The leaf data grows from end-to-front in the node.
 583  * this returns the address of the start of the last item,
 584  * which is the stop of the leaf data stack
 585  */
 586 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 587                                          struct extent_buffer *leaf)
 588 {
 589         u32 nr = btrfs_header_nritems(leaf);
 590         if (nr == 0)
 591                 return BTRFS_LEAF_DATA_SIZE(root);
 592         return btrfs_item_offset_nr(leaf, nr - 1);
 593 }
 594
 595 int btrfs_check_node(struct btrfs_root *root,
 596                       struct btrfs_disk_key *parent_key,
 597                       struct extent_buffer *buf)
 598 {
 599         int i;
 600         struct btrfs_key cpukey;
 601         struct btrfs_disk_key key;
 602         u32 nritems = btrfs_header_nritems(buf);
 603
 604         if (nritems == 0 || nritems > BTRFS_NODEPTRS_PER_BLOCK(root))
 605                 goto fail;
 606
 607         if (parent_key && parent_key->type) {
 608                 btrfs_node_key(buf, &key, 0);
 609                 if (memcmp(parent_key, &key, sizeof(key)))
 610                         goto fail;
 611         }
 612         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 613                 btrfs_node_key(buf, &key, i);
 614                 btrfs_node_key_to_cpu(buf, &cpukey, i + 1);
 615                 if (btrfs_comp_keys(&key, &cpukey) >= 0)
 616                         goto fail;
 617         }
 618         return 0;
 619 fail:
 620         if (btrfs_header_owner(buf) == BTRFS_EXTENT_TREE_OBJECTID) {
 621                 if (parent_key)
 622                         btrfs_disk_key_to_cpu(&cpukey, parent_key);
 623                 else
 624                         btrfs_node_key_to_cpu(buf, &cpukey, 0);
 625                 btrfs_add_corrupt_extent_record(root->fs_info, &cpukey,
 626                                                 buf->start, buf->len,
 627                                                 btrfs_header_level(buf));
 628         }
 629         return -EIO;
 630 }
 631
 632 int btrfs_check_leaf(struct btrfs_root *root,
 633                       struct btrfs_disk_key *parent_key,
 634                       struct extent_buffer *buf)
 635 {
 636         int i;
 637         struct btrfs_key cpukey;
 638         struct btrfs_disk_key key;
 639         u32 nritems = btrfs_header_nritems(buf);
 640
 641         if (btrfs_header_level(buf) != 0) {
 642                 fprintf(stderr, "leaf is not a leaf %llu\n",
 643                        (unsigned long long)btrfs_header_bytenr(buf));
 644                 goto fail;
 645         }
 646         if (btrfs_leaf_free_space(root, buf) < 0) {
 647                 fprintf(stderr, "leaf free space incorrect %llu %d\n",
 648                         (unsigned long long)btrfs_header_bytenr(buf),
 649                         btrfs_leaf_free_space(root, buf));
 650                 goto fail;
 651         }
 652
 653         if (nritems == 0)
 654                 return 0;
 655
 656         btrfs_item_key(buf, &key, 0);
 657         if (parent_key && parent_key->type &&
 658             memcmp(parent_key, &key, sizeof(key))) {
 659                 fprintf(stderr, "leaf parent key incorrect %llu\n",
 660                        (unsigned long long)btrfs_header_bytenr(buf));
 661                 goto fail;
 662         }
 663         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 664                 btrfs_item_key(buf, &key, i);
 665                 btrfs_item_key_to_cpu(buf, &cpukey, i + 1);
 666                 if (btrfs_comp_keys(&key, &cpukey) >= 0) {
 667                         fprintf(stderr, "bad key ordering %d %d\n", i, i+1);
 668                         goto fail;
 669                 }
 670                 if (btrfs_item_offset_nr(buf, i) !=
 671                         btrfs_item_end_nr(buf, i + 1)) {
 672                         fprintf(stderr, "incorrect offsets %u %u\n",
 673                                 btrfs_item_offset_nr(buf, i),
 674                                 btrfs_item_end_nr(buf, i + 1));
 675                         goto fail;
 676                 }
 677                 if (i == 0 && btrfs_item_end_nr(buf, i) !=
 678                     BTRFS_LEAF_DATA_SIZE(root)) {
 679                         fprintf(stderr, "bad item end %u wanted %u\n",
 680                                 btrfs_item_end_nr(buf, i),
 681                                 (unsigned)BTRFS_LEAF_DATA_SIZE(root));
 682                         goto fail;
 683                 }
 684         }
 685         return 0;
 686 fail:
 687         if (btrfs_header_owner(buf) == BTRFS_EXTENT_TREE_OBJECTID) {
 688                 if (parent_key)
 689                         btrfs_disk_key_to_cpu(&cpukey, parent_key);
 690                 else
 691                         btrfs_item_key_to_cpu(buf, &cpukey, 0);
 692
 693                 btrfs_add_corrupt_extent_record(root->fs_info, &cpukey,
 694                                                 buf->start, buf->len, 0);
 695         }
 696         return -EIO;
 697 }
 698
 699 static int noinline check_block(struct btrfs_root *root,
 700                                 struct btrfs_path *path, int level)
 701 {
 702         struct btrfs_disk_key key;
 703         struct btrfs_disk_key *key_ptr = NULL;
 704         struct extent_buffer *parent;
 705
 706         if (path->nodes[level + 1]) {
 707                 parent = path->nodes[level + 1];
 708                 btrfs_node_key(parent, &key, path->slots[level + 1]);
 709                 key_ptr = &key;
 710         }
 711         if (level == 0)
 712                 return btrfs_check_leaf(root, key_ptr, path->nodes[0]);
 713         return btrfs_check_node(root, key_ptr, path->nodes[level]);
 714 }
 715
 716 /*
 717  * search for key in the extent_buffer.  The items start at offset p,
 718  * and they are item_size apart.  There are 'max' items in p.
 719  *
 720  * the slot in the array is returned via slot, and it points to
 721  * the place where you would insert key if it is not found in
 722  * the array.
 723  *
 724  * slot may point to max if the key is bigger than all of the keys
 725  */
 726 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 727                               int item_size, struct btrfs_key *key,
 728                               int max, int *slot)
 729 {
 730         int low = 0;
 731         int high = max;
 732         int mid;
 733         int ret;
 734         unsigned long offset;
 735         struct btrfs_disk_key *tmp;
 736
 737         while(low < high) {
 738                 mid = (low + high) / 2;
 739                 offset = p + mid * item_size;
 740
 741                 tmp = (struct btrfs_disk_key *)(eb->data + offset);
 742                 ret = btrfs_comp_keys(tmp, key);
 743
 744                 if (ret < 0)
 745                         low = mid + 1;
 746                 else if (ret > 0)
 747                         high = mid;
 748                 else {
 749                         *slot = mid;
 750                         return 0;
 751                 }
 752         }
 753         *slot = low;
 754         return 1;
 755 }
 756
 757 /*
 758  * simple bin_search frontend that does the right thing for
 759  * leaves vs nodes
 760  */
 761 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 762                       int level, int *slot)
 763 {
 764         if (level == 0) {
 765                 return generic_bin_search(eb,
 766                                           offsetof(struct btrfs_leaf, items),
 767                                           sizeof(struct btrfs_item),
 768                                           key, btrfs_header_nritems(eb),
 769                                           slot);
 770         } else {
 771                 return generic_bin_search(eb,
 772                                           offsetof(struct btrfs_node, ptrs),
 773                                           sizeof(struct btrfs_key_ptr),
 774                                           key, btrfs_header_nritems(eb),
 775                                           slot);
 776         }
 777         return -1;
 778 }
 779
 780 struct extent_buffer *read_node_slot(struct btrfs_root *root,
 781                                    struct extent_buffer *parent, int slot)
 782 {
 783         int level = btrfs_header_level(parent);
 784         if (slot < 0)
 785                 return NULL;
 786         if (slot >= btrfs_header_nritems(parent))
 787                 return NULL;
 788
 789         BUG_ON(level == 0);
 790
 791         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 792                        btrfs_level_size(root, level - 1),
 793                        btrfs_node_ptr_generation(parent, slot));
 794 }
 795
 796 static int balance_level(struct btrfs_trans_handle *trans,
 797                          struct btrfs_root *root,
 798                          struct btrfs_path *path, int level)
 799 {
 800         struct extent_buffer *right = NULL;
 801         struct extent_buffer *mid;
 802         struct extent_buffer *left = NULL;
 803         struct extent_buffer *parent = NULL;
 804         int ret = 0;
 805         int wret;
 806         int pslot;
 807         int orig_slot = path->slots[level];
 808         u64 orig_ptr;
 809
 810         if (level == 0)
 811                 return 0;
 812
 813         mid = path->nodes[level];
 814         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 815
 816         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 817
 818         if (level < BTRFS_MAX_LEVEL - 1)
 819                 parent = path->nodes[level + 1];
 820         pslot = path->slots[level + 1];
 821
 822         /*
 823          * deal with the case where there is only one pointer in the root
 824          * by promoting the node below to a root
 825          */
 826         if (!parent) {
 827                 struct extent_buffer *child;
 828
 829                 if (btrfs_header_nritems(mid) != 1)
 830                         return 0;
 831
 832                 /* promote the child to a root */
 833                 child = read_node_slot(root, mid, 0);
 834                 BUG_ON(!child);
 835                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
 836                 BUG_ON(ret);
 837
 838                 root->node = child;
 839                 add_root_to_dirty_list(root);
 840                 path->nodes[level] = NULL;
 841                 clean_tree_block(trans, root, mid);
 842                 wait_on_tree_block_writeback(root, mid);
 843                 /* once for the path */
 844                 free_extent_buffer(mid);
 845
 846                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 847                                         0, root->root_key.objectid,
 848                                         level, 1);
 849                 /* once for the root ptr */
 850                 free_extent_buffer(mid);
 851                 return ret;
 852         }
 853         if (btrfs_header_nritems(mid) >
 854             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 855                 return 0;
 856
 857         left = read_node_slot(root, parent, pslot - 1);
 858         if (left) {
 859                 wret = btrfs_cow_block(trans, root, left,
 860                                        parent, pslot - 1, &left);
 861                 if (wret) {
 862                         ret = wret;
 863                         goto enospc;
 864                 }
 865         }
 866         right = read_node_slot(root, parent, pslot + 1);
 867         if (right) {
 868                 wret = btrfs_cow_block(trans, root, right,
 869                                        parent, pslot + 1, &right);
 870                 if (wret) {
 871                         ret = wret;
 872                         goto enospc;
 873                 }
 874         }
 875
 876         /* first, try to make some room in the middle buffer */
 877         if (left) {
 878                 orig_slot += btrfs_header_nritems(left);
 879                 wret = push_node_left(trans, root, left, mid, 1);
 880                 if (wret < 0)
 881                         ret = wret;
 882         }
 883
 884         /*
 885          * then try to empty the right most buffer into the middle
 886          */
 887         if (right) {
 888                 wret = push_node_left(trans, root, mid, right, 1);
 889                 if (wret < 0 && wret != -ENOSPC)
 890                         ret = wret;
 891                 if (btrfs_header_nritems(right) == 0) {
 892                         u64 bytenr = right->start;
 893                         u32 blocksize = right->len;
 894
 895                         clean_tree_block(trans, root, right);
 896                         wait_on_tree_block_writeback(root, right);
 897                         free_extent_buffer(right);
 898                         right = NULL;
 899                         wret = btrfs_del_ptr(trans, root, path,
 900                                              level + 1, pslot + 1);
 901                         if (wret)
 902                                 ret = wret;
 903                         wret = btrfs_free_extent(trans, root, bytenr,
 904                                                  blocksize, 0,
 905                                                  root->root_key.objectid,
 906                                                  level, 0);
 907                         if (wret)
 908                                 ret = wret;
 909                 } else {
 910                         struct btrfs_disk_key right_key;
 911                         btrfs_node_key(right, &right_key, 0);
 912                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 913                         btrfs_mark_buffer_dirty(parent);
 914                 }
 915         }
 916         if (btrfs_header_nritems(mid) == 1) {
 917                 /*
 918                  * we're not allowed to leave a node with one item in the
 919                  * tree during a delete.  A deletion from lower in the tree
 920                  * could try to delete the only pointer in this node.
 921                  * So, pull some keys from the left.
 922                  * There has to be a left pointer at this point because
 923                  * otherwise we would have pulled some pointers from the
 924                  * right
 925                  */
 926                 BUG_ON(!left);
 927                 wret = balance_node_right(trans, root, mid, left);
 928                 if (wret < 0) {
 929                         ret = wret;
 930                         goto enospc;
 931                 }
 932                 if (wret == 1) {
 933                         wret = push_node_left(trans, root, left, mid, 1);
 934                         if (wret < 0)
 935                                 ret = wret;
 936                 }
 937                 BUG_ON(wret == 1);
 938         }
 939         if (btrfs_header_nritems(mid) == 0) {
 940                 /* we've managed to empty the middle node, drop it */
 941                 u64 bytenr = mid->start;
 942                 u32 blocksize = mid->len;
 943                 clean_tree_block(trans, root, mid);
 944                 wait_on_tree_block_writeback(root, mid);
 945                 free_extent_buffer(mid);
 946                 mid = NULL;
 947                 wret = btrfs_del_ptr(trans, root, path, level + 1, pslot);
 948                 if (wret)
 949                         ret = wret;
 950                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 951                                          0, root->root_key.objectid,
 952                                          level, 0);
 953                 if (wret)
 954                         ret = wret;
 955         } else {
 956                 /* update the parent key to reflect our changes */
 957                 struct btrfs_disk_key mid_key;
 958                 btrfs_node_key(mid, &mid_key, 0);
 959                 btrfs_set_node_key(parent, &mid_key, pslot);
 960                 btrfs_mark_buffer_dirty(parent);
 961         }
 962
 963         /* update the path */
 964         if (left) {
 965                 if (btrfs_header_nritems(left) > orig_slot) {
 966                         extent_buffer_get(left);
 967                         path->nodes[level] = left;
 968                         path->slots[level + 1] -= 1;
 969                         path->slots[level] = orig_slot;
 970                         if (mid)
 971                                 free_extent_buffer(mid);
 972                 } else {
 973                         orig_slot -= btrfs_header_nritems(left);
 974                         path->slots[level] = orig_slot;
 975                 }
 976         }
 977         /* double check we haven't messed things up */
 978         check_block(root, path, level);
 979         if (orig_ptr !=
 980             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 981                 BUG();
 982 enospc:
 983         if (right)
 984                 free_extent_buffer(right);
 985         if (left)
 986                 free_extent_buffer(left);
 987         return ret;
 988 }
 989
 990 /* returns zero if the push worked, non-zero otherwise */
 991 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
 992                                           struct btrfs_root *root,
 993                                           struct btrfs_path *path, int level)
 994 {
 995         struct extent_buffer *right = NULL;
 996         struct extent_buffer *mid;
 997         struct extent_buffer *left = NULL;
 998         struct extent_buffer *parent = NULL;
 999         int ret = 0;
1000         int wret;
1001         int pslot;
1002         int orig_slot = path->slots[level];
1003
1004         if (level == 0)
1005                 return 1;
1006
1007         mid = path->nodes[level];
1008         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1009
1010         if (level < BTRFS_MAX_LEVEL - 1)
1011                 parent = path->nodes[level + 1];
1012         pslot = path->slots[level + 1];
1013
1014         if (!parent)
1015                 return 1;
1016
1017         left = read_node_slot(root, parent, pslot - 1);
1018
1019         /* first, try to make some room in the middle buffer */
1020         if (left) {
1021                 u32 left_nr;
1022                 left_nr = btrfs_header_nritems(left);
1023                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1024                         wret = 1;
1025                 } else {
1026                         ret = btrfs_cow_block(trans, root, left, parent,
1027                                               pslot - 1, &left);
1028                         if (ret)
1029                                 wret = 1;
1030                         else {
1031                                 wret = push_node_left(trans, root,
1032                                                       left, mid, 0);
1033                         }
1034                 }
1035                 if (wret < 0)
1036                         ret = wret;
1037                 if (wret == 0) {
1038                         struct btrfs_disk_key disk_key;
1039                         orig_slot += left_nr;
1040                         btrfs_node_key(mid, &disk_key, 0);
1041                         btrfs_set_node_key(parent, &disk_key, pslot);
1042                         btrfs_mark_buffer_dirty(parent);
1043                         if (btrfs_header_nritems(left) > orig_slot) {
1044                                 path->nodes[level] = left;
1045                                 path->slots[level + 1] -= 1;
1046                                 path->slots[level] = orig_slot;
1047                                 free_extent_buffer(mid);
1048                         } else {
1049                                 orig_slot -=
1050                                         btrfs_header_nritems(left);
1051                                 path->slots[level] = orig_slot;
1052                                 free_extent_buffer(left);
1053                         }
1054                         return 0;
1055                 }
1056                 free_extent_buffer(left);
1057         }
1058         right= read_node_slot(root, parent, pslot + 1);
1059
1060         /*
1061          * then try to empty the right most buffer into the middle
1062          */
1063         if (right) {
1064                 u32 right_nr;
1065                 right_nr = btrfs_header_nritems(right);
1066                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1067                         wret = 1;
1068                 } else {
1069                         ret = btrfs_cow_block(trans, root, right,
1070                                               parent, pslot + 1,
1071                                               &right);
1072                         if (ret)
1073                                 wret = 1;
1074                         else {
1075                                 wret = balance_node_right(trans, root,
1076                                                           right, mid);
1077                         }
1078                 }
1079                 if (wret < 0)
1080                         ret = wret;
1081                 if (wret == 0) {
1082                         struct btrfs_disk_key disk_key;
1083
1084                         btrfs_node_key(right, &disk_key, 0);
1085                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
1086                         btrfs_mark_buffer_dirty(parent);
1087
1088                         if (btrfs_header_nritems(mid) <= orig_slot) {
1089                                 path->nodes[level] = right;
1090                                 path->slots[level + 1] += 1;
1091                                 path->slots[level] = orig_slot -
1092                                         btrfs_header_nritems(mid);
1093                                 free_extent_buffer(mid);
1094                         } else {
1095                                 free_extent_buffer(right);
1096                         }
1097                         return 0;
1098                 }
1099                 free_extent_buffer(right);
1100         }
1101         return 1;
1102 }
1103
1104 /*
1105  * readahead one full node of leaves
1106  */
1107 void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
1108                              int level, int slot, u64 objectid)
1109 {
1110         struct extent_buffer *node;
1111         struct btrfs_disk_key disk_key;
1112         u32 nritems;
1113         u64 search;
1114         u64 lowest_read;
1115         u64 highest_read;
1116         u64 nread = 0;
1117         int direction = path->reada;
1118         struct extent_buffer *eb;
1119         u32 nr;
1120         u32 blocksize;
1121         u32 nscan = 0;
1122
1123         if (level != 1)
1124                 return;
1125
1126         if (!path->nodes[level])
1127                 return;
1128
1129         node = path->nodes[level];
1130         search = btrfs_node_blockptr(node, slot);
1131         blocksize = btrfs_level_size(root, level - 1);
1132         eb = btrfs_find_tree_block(root, search, blocksize);
1133         if (eb) {
1134                 free_extent_buffer(eb);
1135                 return;
1136         }
1137
1138         highest_read = search;
1139         lowest_read = search;
1140
1141         nritems = btrfs_header_nritems(node);
1142         nr = slot;
1143         while(1) {
1144                 if (direction < 0) {
1145                         if (nr == 0)
1146                                 break;
1147                         nr--;
1148                 } else if (direction > 0) {
1149                         nr++;
1150                         if (nr >= nritems)
1151                                 break;
1152                 }
1153                 if (path->reada < 0 && objectid) {
1154                         btrfs_node_key(node, &disk_key, nr);
1155                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1156                                 break;
1157                 }
1158                 search = btrfs_node_blockptr(node, nr);
1159                 if ((search >= lowest_read && search <= highest_read) ||
1160                     (search < lowest_read && lowest_read - search <= 32768) ||
1161                     (search > highest_read && search - highest_read <= 32768)) {
1162                         readahead_tree_block(root, search, blocksize,
1163                                      btrfs_node_ptr_generation(node, nr));
1164                         nread += blocksize;
1165                 }
1166                 nscan++;
1167                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
1168                         break;
1169                 if(nread > (1024 * 1024) || nscan > 128)
1170                         break;
1171
1172                 if (search < lowest_read)
1173                         lowest_read = search;
1174                 if (search > highest_read)
1175                         highest_read = search;
1176         }
1177 }
1178
1179 /*
1180  * look for key in the tree.  path is filled in with nodes along the way
1181  * if key is found, we return zero and you can find the item in the leaf
1182  * level of the path (level 0)
1183  *
1184  * If the key isn't found, the path points to the slot where it should
1185  * be inserted, and 1 is returned.  If there are other errors during the
1186  * search a negative error number is returned.
1187  *
1188  * if ins_len > 0, nodes and leaves will be split as we walk down the
1189  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1190  * possible)
1191  */
1192 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1193                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1194                       ins_len, int cow)
1195 {
1196         struct extent_buffer *b;
1197         int slot;
1198         int ret;
1199         int level;
1200         int should_reada = p->reada;
1201         u8 lowest_level = 0;
1202
1203         lowest_level = p->lowest_level;
1204         WARN_ON(lowest_level && ins_len > 0);
1205         WARN_ON(p->nodes[0] != NULL);
1206         /*
1207         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
1208         */
1209 again:
1210         b = root->node;
1211         extent_buffer_get(b);
1212         while (b) {
1213                 level = btrfs_header_level(b);
1214                 if (cow) {
1215                         int wret;
1216                         wret = btrfs_cow_block(trans, root, b,
1217                                                p->nodes[level + 1],
1218                                                p->slots[level + 1],
1219                                                &b);
1220                         if (wret) {
1221                                 free_extent_buffer(b);
1222                                 return wret;
1223                         }
1224                 }
1225                 BUG_ON(!cow && ins_len);
1226                 if (level != btrfs_header_level(b))
1227                         WARN_ON(1);
1228                 level = btrfs_header_level(b);
1229                 p->nodes[level] = b;
1230                 ret = check_block(root, p, level);
1231                 if (ret)
1232                         return -1;
1233                 ret = bin_search(b, key, level, &slot);
1234                 if (level != 0) {
1235                         if (ret && slot > 0)
1236                                 slot -= 1;
1237                         p->slots[level] = slot;
1238                         if ((p->search_for_split || ins_len > 0) &&
1239                             btrfs_header_nritems(b) >=
1240                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1241                                 int sret = split_node(trans, root, p, level);
1242                                 BUG_ON(sret > 0);
1243                                 if (sret)
1244                                         return sret;
1245                                 b = p->nodes[level];
1246                                 slot = p->slots[level];
1247                         } else if (ins_len < 0) {
1248                                 int sret = balance_level(trans, root, p,
1249                                                          level);
1250                                 if (sret)
1251                                         return sret;
1252                                 b = p->nodes[level];
1253                                 if (!b) {
1254                                         btrfs_release_path(NULL, p);
1255                                         goto again;
1256                                 }
1257                                 slot = p->slots[level];
1258                                 BUG_ON(btrfs_header_nritems(b) == 1);
1259                         }
1260                         /* this is only true while dropping a snapshot */
1261                         if (level == lowest_level)
1262                                 break;
1263
1264                         if (should_reada)
1265                                 reada_for_search(root, p, level, slot,
1266                                                  key->objectid);
1267
1268                         b = read_node_slot(root, b, slot);
1269                         if (!extent_buffer_uptodate(b))
1270                                 return -EIO;
1271                 } else {
1272                         p->slots[level] = slot;
1273                         if (ins_len > 0 &&
1274                             ins_len > btrfs_leaf_free_space(root, b)) {
1275                                 int sret = split_leaf(trans, root, key,
1276                                                       p, ins_len, ret == 0);
1277                                 BUG_ON(sret > 0);
1278                                 if (sret)
1279                                         return sret;
1280                         }
1281                         return ret;
1282                 }
1283         }
1284         return 1;
1285 }
1286
1287 /*
1288  * adjust the pointers going up the tree, starting at level
1289  * making sure the right key of each node is points to 'key'.
1290  * This is used after shifting pointers to the left, so it stops
1291  * fixing up pointers when a given leaf/node is not in slot 0 of the
1292  * higher levels
1293  *
1294  * If this fails to write a tree block, it returns -1, but continues
1295  * fixing up the blocks in ram so the tree is consistent.
1296  */
1297 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1298                           struct btrfs_root *root, struct btrfs_path *path,
1299                           struct btrfs_disk_key *key, int level)
1300 {
1301         int i;
1302         int ret = 0;
1303         struct extent_buffer *t;
1304
1305         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1306                 int tslot = path->slots[i];
1307                 if (!path->nodes[i])
1308                         break;
1309                 t = path->nodes[i];
1310                 btrfs_set_node_key(t, key, tslot);
1311                 btrfs_mark_buffer_dirty(path->nodes[i]);
1312                 if (tslot != 0)
1313                         break;
1314         }
1315         return ret;
1316 }
1317
1318 /*
1319  * update item key.
1320  *
1321  * This function isn't completely safe. It's the caller's responsibility
1322  * that the new key won't break the order
1323  */
1324 int btrfs_set_item_key_safe(struct btrfs_trans_handle *trans,
1325                             struct btrfs_root *root, struct btrfs_path *path,
1326                             struct btrfs_key *new_key)
1327 {
1328         struct btrfs_disk_key disk_key;
1329         struct extent_buffer *eb;
1330         int slot;
1331
1332         eb = path->nodes[0];
1333         slot = path->slots[0];
1334         if (slot > 0) {
1335                 btrfs_item_key(eb, &disk_key, slot - 1);
1336                 if (btrfs_comp_keys(&disk_key, new_key) >= 0)
1337                         return -1;
1338         }
1339         if (slot < btrfs_header_nritems(eb) - 1) {
1340                 btrfs_item_key(eb, &disk_key, slot + 1);
1341                 if (btrfs_comp_keys(&disk_key, new_key) <= 0)
1342                         return -1;
1343         }
1344
1345         btrfs_cpu_key_to_disk(&disk_key, new_key);
1346         btrfs_set_item_key(eb, &disk_key, slot);
1347         btrfs_mark_buffer_dirty(eb);
1348         if (slot == 0)
1349                 fixup_low_keys(trans, root, path, &disk_key, 1);
1350         return 0;
1351 }
1352
1353 /*
1354  * try to push data from one node into the next node left in the
1355  * tree.
1356  *
1357  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1358  * error, and > 0 if there was no room in the left hand block.
1359  */
1360 static int push_node_left(struct btrfs_trans_handle *trans,
1361                           struct btrfs_root *root, struct extent_buffer *dst,
1362                           struct extent_buffer *src, int empty)
1363 {
1364         int push_items = 0;
1365         int src_nritems;
1366         int dst_nritems;
1367         int ret = 0;
1368
1369         src_nritems = btrfs_header_nritems(src);
1370         dst_nritems = btrfs_header_nritems(dst);
1371         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1372         WARN_ON(btrfs_header_generation(src) != trans->transid);
1373         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1374
1375         if (!empty && src_nritems <= 8)
1376                 return 1;
1377
1378         if (push_items <= 0) {
1379                 return 1;
1380         }
1381
1382         if (empty) {
1383                 push_items = min(src_nritems, push_items);
1384                 if (push_items < src_nritems) {
1385                         /* leave at least 8 pointers in the node if
1386                          * we aren't going to empty it
1387                          */
1388                         if (src_nritems - push_items < 8) {
1389                                 if (push_items <= 8)
1390                                         return 1;
1391                                 push_items -= 8;
1392                         }
1393                 }
1394         } else
1395                 push_items = min(src_nritems - 8, push_items);
1396
1397         copy_extent_buffer(dst, src,
1398                            btrfs_node_key_ptr_offset(dst_nritems),
1399                            btrfs_node_key_ptr_offset(0),
1400                            push_items * sizeof(struct btrfs_key_ptr));
1401
1402         if (push_items < src_nritems) {
1403                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1404                                       btrfs_node_key_ptr_offset(push_items),
1405                                       (src_nritems - push_items) *
1406                                       sizeof(struct btrfs_key_ptr));
1407         }
1408         btrfs_set_header_nritems(src, src_nritems - push_items);
1409         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1410         btrfs_mark_buffer_dirty(src);
1411         btrfs_mark_buffer_dirty(dst);
1412
1413         return ret;
1414 }
1415
1416 /*
1417  * try to push data from one node into the next node right in the
1418  * tree.
1419  *
1420  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1421  * error, and > 0 if there was no room in the right hand block.
1422  *
1423  * this will  only push up to 1/2 the contents of the left node over
1424  */
1425 static int balance_node_right(struct btrfs_trans_handle *trans,
1426                               struct btrfs_root *root,
1427                               struct extent_buffer *dst,
1428                               struct extent_buffer *src)
1429 {
1430         int push_items = 0;
1431         int max_push;
1432         int src_nritems;
1433         int dst_nritems;
1434         int ret = 0;
1435
1436         WARN_ON(btrfs_header_generation(src) != trans->transid);
1437         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1438
1439         src_nritems = btrfs_header_nritems(src);
1440         dst_nritems = btrfs_header_nritems(dst);
1441         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1442         if (push_items <= 0) {
1443                 return 1;
1444         }
1445
1446         if (src_nritems < 4) {
1447                 return 1;
1448         }
1449
1450         max_push = src_nritems / 2 + 1;
1451         /* don't try to empty the node */
1452         if (max_push >= src_nritems) {
1453                 return 1;
1454         }
1455
1456         if (max_push < push_items)
1457                 push_items = max_push;
1458
1459         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1460                                       btrfs_node_key_ptr_offset(0),
1461                                       (dst_nritems) *
1462                                       sizeof(struct btrfs_key_ptr));
1463
1464         copy_extent_buffer(dst, src,
1465                            btrfs_node_key_ptr_offset(0),
1466                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1467                            push_items * sizeof(struct btrfs_key_ptr));
1468
1469         btrfs_set_header_nritems(src, src_nritems - push_items);
1470         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1471
1472         btrfs_mark_buffer_dirty(src);
1473         btrfs_mark_buffer_dirty(dst);
1474
1475         return ret;
1476 }
1477
1478 /*
1479  * helper function to insert a new root level in the tree.
1480  * A new node is allocated, and a single item is inserted to
1481  * point to the existing root
1482  *
1483  * returns zero on success or < 0 on failure.
1484  */
1485 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1486                            struct btrfs_root *root,
1487                            struct btrfs_path *path, int level)
1488 {
1489         u64 lower_gen;
1490         struct extent_buffer *lower;
1491         struct extent_buffer *c;
1492         struct extent_buffer *old;
1493         struct btrfs_disk_key lower_key;
1494
1495         BUG_ON(path->nodes[level]);
1496         BUG_ON(path->nodes[level-1] != root->node);
1497
1498         lower = path->nodes[level-1];
1499         if (level == 1)
1500                 btrfs_item_key(lower, &lower_key, 0);
1501         else
1502                 btrfs_node_key(lower, &lower_key, 0);
1503
1504         c = btrfs_alloc_free_block(trans, root, root->nodesize,
1505                                    root->root_key.objectid, &lower_key,
1506                                    level, root->node->start, 0);
1507
1508         if (IS_ERR(c))
1509                 return PTR_ERR(c);
1510
1511         memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
1512         btrfs_set_header_nritems(c, 1);
1513         btrfs_set_header_level(c, level);
1514         btrfs_set_header_bytenr(c, c->start);
1515         btrfs_set_header_generation(c, trans->transid);
1516         btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
1517         btrfs_set_header_owner(c, root->root_key.objectid);
1518
1519         write_extent_buffer(c, root->fs_info->fsid,
1520                             (unsigned long)btrfs_header_fsid(c),
1521                             BTRFS_FSID_SIZE);
1522
1523         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1524                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
1525                             BTRFS_UUID_SIZE);
1526
1527         btrfs_set_node_key(c, &lower_key, 0);
1528         btrfs_set_node_blockptr(c, 0, lower->start);
1529         lower_gen = btrfs_header_generation(lower);
1530         WARN_ON(lower_gen != trans->transid);
1531
1532         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1533
1534         btrfs_mark_buffer_dirty(c);
1535
1536         old = root->node;
1537         root->node = c;
1538
1539         /* the super has an extra ref to root->node */
1540         free_extent_buffer(old);
1541
1542         add_root_to_dirty_list(root);
1543         extent_buffer_get(c);
1544         path->nodes[level] = c;
1545         path->slots[level] = 0;
1546         return 0;
1547 }
1548
1549 /*
1550  * worker function to insert a single pointer in a node.
1551  * the node should have enough room for the pointer already
1552  *
1553  * slot and level indicate where you want the key to go, and
1554  * blocknr is the block the key points to.
1555  *
1556  * returns zero on success and < 0 on any error
1557  */
1558 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1559                       *root, struct btrfs_path *path, struct btrfs_disk_key
1560                       *key, u64 bytenr, int slot, int level)
1561 {
1562         struct extent_buffer *lower;
1563         int nritems;
1564
1565         BUG_ON(!path->nodes[level]);
1566         lower = path->nodes[level];
1567         nritems = btrfs_header_nritems(lower);
1568         if (slot > nritems)
1569                 BUG();
1570         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1571                 BUG();
1572         if (slot != nritems) {
1573                 memmove_extent_buffer(lower,
1574                               btrfs_node_key_ptr_offset(slot + 1),
1575                               btrfs_node_key_ptr_offset(slot),
1576                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1577         }
1578         btrfs_set_node_key(lower, key, slot);
1579         btrfs_set_node_blockptr(lower, slot, bytenr);
1580         WARN_ON(trans->transid == 0);
1581         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1582         btrfs_set_header_nritems(lower, nritems + 1);
1583         btrfs_mark_buffer_dirty(lower);
1584         return 0;
1585 }
1586
1587 /*
1588  * split the node at the specified level in path in two.
1589  * The path is corrected to point to the appropriate node after the split
1590  *
1591  * Before splitting this tries to make some room in the node by pushing
1592  * left and right, if either one works, it returns right away.
1593  *
1594  * returns 0 on success and < 0 on failure
1595  */
1596 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1597                       *root, struct btrfs_path *path, int level)
1598 {
1599         struct extent_buffer *c;
1600         struct extent_buffer *split;
1601         struct btrfs_disk_key disk_key;
1602         int mid;
1603         int ret;
1604         int wret;
1605         u32 c_nritems;
1606
1607         c = path->nodes[level];
1608         WARN_ON(btrfs_header_generation(c) != trans->transid);
1609         if (c == root->node) {
1610                 /* trying to split the root, lets make a new one */
1611                 ret = insert_new_root(trans, root, path, level + 1);
1612                 if (ret)
1613                         return ret;
1614         } else {
1615                 ret = push_nodes_for_insert(trans, root, path, level);
1616                 c = path->nodes[level];
1617                 if (!ret && btrfs_header_nritems(c) <
1618                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
1619                         return 0;
1620                 if (ret < 0)
1621                         return ret;
1622         }
1623
1624         c_nritems = btrfs_header_nritems(c);
1625         mid = (c_nritems + 1) / 2;
1626         btrfs_node_key(c, &disk_key, mid);
1627
1628         split = btrfs_alloc_free_block(trans, root, root->nodesize,
1629                                         root->root_key.objectid,
1630                                         &disk_key, level, c->start, 0);
1631         if (IS_ERR(split))
1632                 return PTR_ERR(split);
1633
1634         memset_extent_buffer(split, 0, 0, sizeof(struct btrfs_header));
1635         btrfs_set_header_level(split, btrfs_header_level(c));
1636         btrfs_set_header_bytenr(split, split->start);
1637         btrfs_set_header_generation(split, trans->transid);
1638         btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
1639         btrfs_set_header_owner(split, root->root_key.objectid);
1640         write_extent_buffer(split, root->fs_info->fsid,
1641                             (unsigned long)btrfs_header_fsid(split),
1642                             BTRFS_FSID_SIZE);
1643         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1644                             (unsigned long)btrfs_header_chunk_tree_uuid(split),
1645                             BTRFS_UUID_SIZE);
1646
1647
1648         copy_extent_buffer(split, c,
1649                            btrfs_node_key_ptr_offset(0),
1650                            btrfs_node_key_ptr_offset(mid),
1651                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1652         btrfs_set_header_nritems(split, c_nritems - mid);
1653         btrfs_set_header_nritems(c, mid);
1654         ret = 0;
1655
1656         btrfs_mark_buffer_dirty(c);
1657         btrfs_mark_buffer_dirty(split);
1658
1659         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1660                           path->slots[level + 1] + 1,
1661                           level + 1);
1662         if (wret)
1663                 ret = wret;
1664
1665         if (path->slots[level] >= mid) {
1666                 path->slots[level] -= mid;
1667                 free_extent_buffer(c);
1668                 path->nodes[level] = split;
1669                 path->slots[level + 1] += 1;
1670         } else {
1671                 free_extent_buffer(split);
1672         }
1673         return ret;
1674 }
1675
1676 /*
1677  * how many bytes are required to store the items in a leaf.  start
1678  * and nr indicate which items in the leaf to check.  This totals up the
1679  * space used both by the item structs and the item data
1680  */
1681 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1682 {
1683         int data_len;
1684         int nritems = btrfs_header_nritems(l);
1685         int end = min(nritems, start + nr) - 1;
1686
1687         if (!nr)
1688                 return 0;
1689         data_len = btrfs_item_end_nr(l, start);
1690         data_len = data_len - btrfs_item_offset_nr(l, end);
1691         data_len += sizeof(struct btrfs_item) * nr;
1692         WARN_ON(data_len < 0);
1693         return data_len;
1694 }
1695
1696 /*
1697  * The space between the end of the leaf items and
1698  * the start of the leaf data.  IOW, how much room
1699  * the leaf has left for both items and data
1700  */
1701 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1702 {
1703         int nritems = btrfs_header_nritems(leaf);
1704         int ret;
1705         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1706         if (ret < 0) {
1707                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1708                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1709                        leaf_space_used(leaf, 0, nritems), nritems);
1710         }
1711         return ret;
1712 }
1713
1714 /*
1715  * push some data in the path leaf to the right, trying to free up at
1716  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1717  *
1718  * returns 1 if the push failed because the other node didn't have enough
1719  * room, 0 if everything worked out and < 0 if there were major errors.
1720  */
1721 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1722                            *root, struct btrfs_path *path, int data_size,
1723                            int empty)
1724 {
1725         struct extent_buffer *left = path->nodes[0];
1726         struct extent_buffer *right;
1727         struct extent_buffer *upper;
1728         struct btrfs_disk_key disk_key;
1729         int slot;
1730         u32 i;
1731         int free_space;
1732         int push_space = 0;
1733         int push_items = 0;
1734         struct btrfs_item *item;
1735         u32 left_nritems;
1736         u32 nr;
1737         u32 right_nritems;
1738         u32 data_end;
1739         u32 this_item_size;
1740         int ret;
1741
1742         slot = path->slots[1];
1743         if (!path->nodes[1]) {
1744                 return 1;
1745         }
1746         upper = path->nodes[1];
1747         if (slot >= btrfs_header_nritems(upper) - 1)
1748                 return 1;
1749
1750         right = read_node_slot(root, upper, slot + 1);
1751         free_space = btrfs_leaf_free_space(root, right);
1752         if (free_space < data_size) {
1753                 free_extent_buffer(right);
1754                 return 1;
1755         }
1756
1757         /* cow and double check */
1758         ret = btrfs_cow_block(trans, root, right, upper,
1759                               slot + 1, &right);
1760         if (ret) {
1761                 free_extent_buffer(right);
1762                 return 1;
1763         }
1764         free_space = btrfs_leaf_free_space(root, right);
1765         if (free_space < data_size) {
1766                 free_extent_buffer(right);
1767                 return 1;
1768         }
1769
1770         left_nritems = btrfs_header_nritems(left);
1771         if (left_nritems == 0) {
1772                 free_extent_buffer(right);
1773                 return 1;
1774         }
1775
1776         if (empty)
1777                 nr = 0;
1778         else
1779                 nr = 1;
1780
1781         i = left_nritems - 1;
1782         while (i >= nr) {
1783                 item = btrfs_item_nr(left, i);
1784
1785                 if (path->slots[0] == i)
1786                         push_space += data_size + sizeof(*item);
1787
1788                 this_item_size = btrfs_item_size(left, item);
1789                 if (this_item_size + sizeof(*item) + push_space > free_space)
1790                         break;
1791                 push_items++;
1792                 push_space += this_item_size + sizeof(*item);
1793                 if (i == 0)
1794                         break;
1795                 i--;
1796         }
1797
1798         if (push_items == 0) {
1799                 free_extent_buffer(right);
1800                 return 1;
1801         }
1802
1803         if (!empty && push_items == left_nritems)
1804                 WARN_ON(1);
1805
1806         /* push left to right */
1807         right_nritems = btrfs_header_nritems(right);
1808
1809         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1810         push_space -= leaf_data_end(root, left);
1811
1812         /* make room in the right data area */
1813         data_end = leaf_data_end(root, right);
1814         memmove_extent_buffer(right,
1815                               btrfs_leaf_data(right) + data_end - push_space,
1816                               btrfs_leaf_data(right) + data_end,
1817                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1818
1819         /* copy from the left data area */
1820         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1821                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1822                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1823                      push_space);
1824
1825         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1826                               btrfs_item_nr_offset(0),
1827                               right_nritems * sizeof(struct btrfs_item));
1828
1829         /* copy the items from left to right */
1830         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1831                    btrfs_item_nr_offset(left_nritems - push_items),
1832                    push_items * sizeof(struct btrfs_item));
1833
1834         /* update the item pointers */
1835         right_nritems += push_items;
1836         btrfs_set_header_nritems(right, right_nritems);
1837         push_space = BTRFS_LEAF_DATA_SIZE(root);
1838         for (i = 0; i < right_nritems; i++) {
1839                 item = btrfs_item_nr(right, i);
1840                 push_space -= btrfs_item_size(right, item);
1841                 btrfs_set_item_offset(right, item, push_space);
1842         }
1843
1844         left_nritems -= push_items;
1845         btrfs_set_header_nritems(left, left_nritems);
1846
1847         if (left_nritems)
1848                 btrfs_mark_buffer_dirty(left);
1849         btrfs_mark_buffer_dirty(right);
1850
1851         btrfs_item_key(right, &disk_key, 0);
1852         btrfs_set_node_key(upper, &disk_key, slot + 1);
1853         btrfs_mark_buffer_dirty(upper);
1854
1855         /* then fixup the leaf pointer in the path */
1856         if (path->slots[0] >= left_nritems) {
1857                 path->slots[0] -= left_nritems;
1858                 free_extent_buffer(path->nodes[0]);
1859                 path->nodes[0] = right;
1860                 path->slots[1] += 1;
1861         } else {
1862                 free_extent_buffer(right);
1863         }
1864         return 0;
1865 }
1866 /*
1867  * push some data in the path leaf to the left, trying to free up at
1868  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1869  */
1870 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1871                           *root, struct btrfs_path *path, int data_size,
1872                           int empty)
1873 {
1874         struct btrfs_disk_key disk_key;
1875         struct extent_buffer *right = path->nodes[0];
1876         struct extent_buffer *left;
1877         int slot;
1878         int i;
1879         int free_space;
1880         int push_space = 0;
1881         int push_items = 0;
1882         struct btrfs_item *item;
1883         u32 old_left_nritems;
1884         u32 right_nritems;
1885         u32 nr;
1886         int ret = 0;
1887         int wret;
1888         u32 this_item_size;
1889         u32 old_left_item_size;
1890
1891         slot = path->slots[1];
1892         if (slot == 0)
1893                 return 1;
1894         if (!path->nodes[1])
1895                 return 1;
1896
1897         right_nritems = btrfs_header_nritems(right);
1898         if (right_nritems == 0) {
1899                 return 1;
1900         }
1901
1902         left = read_node_slot(root, path->nodes[1], slot - 1);
1903         free_space = btrfs_leaf_free_space(root, left);
1904         if (free_space < data_size) {
1905                 free_extent_buffer(left);
1906                 return 1;
1907         }
1908
1909         /* cow and double check */
1910         ret = btrfs_cow_block(trans, root, left,
1911                               path->nodes[1], slot - 1, &left);
1912         if (ret) {
1913                 /* we hit -ENOSPC, but it isn't fatal here */
1914                 free_extent_buffer(left);
1915                 return 1;
1916         }
1917
1918         free_space = btrfs_leaf_free_space(root, left);
1919         if (free_space < data_size) {
1920                 free_extent_buffer(left);
1921                 return 1;
1922         }
1923
1924         if (empty)
1925                 nr = right_nritems;
1926         else
1927                 nr = right_nritems - 1;
1928
1929         for (i = 0; i < nr; i++) {
1930                 item = btrfs_item_nr(right, i);
1931
1932                 if (path->slots[0] == i)
1933                         push_space += data_size + sizeof(*item);
1934
1935                 this_item_size = btrfs_item_size(right, item);
1936                 if (this_item_size + sizeof(*item) + push_space > free_space)
1937                         break;
1938
1939                 push_items++;
1940                 push_space += this_item_size + sizeof(*item);
1941         }
1942
1943         if (push_items == 0) {
1944                 free_extent_buffer(left);
1945                 return 1;
1946         }
1947         if (!empty && push_items == btrfs_header_nritems(right))
1948                 WARN_ON(1);
1949
1950         /* push data from right to left */
1951         copy_extent_buffer(left, right,
1952                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1953                            btrfs_item_nr_offset(0),
1954                            push_items * sizeof(struct btrfs_item));
1955
1956         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1957                      btrfs_item_offset_nr(right, push_items -1);
1958
1959         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1960                      leaf_data_end(root, left) - push_space,
1961                      btrfs_leaf_data(right) +
1962                      btrfs_item_offset_nr(right, push_items - 1),
1963                      push_space);
1964         old_left_nritems = btrfs_header_nritems(left);
1965         BUG_ON(old_left_nritems == 0);
1966
1967         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1968         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1969                 u32 ioff;
1970
1971                 item = btrfs_item_nr(left, i);
1972                 ioff = btrfs_item_offset(left, item);
1973                 btrfs_set_item_offset(left, item,
1974                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1975         }
1976         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1977
1978         /* fixup right node */
1979         if (push_items > right_nritems) {
1980                 printk("push items %d nr %u\n", push_items, right_nritems);
1981                 WARN_ON(1);
1982         }
1983
1984         if (push_items < right_nritems) {
1985                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1986                                                   leaf_data_end(root, right);
1987                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
1988                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1989                                       btrfs_leaf_data(right) +
1990                                       leaf_data_end(root, right), push_space);
1991
1992                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1993                               btrfs_item_nr_offset(push_items),
1994                              (btrfs_header_nritems(right) - push_items) *
1995                              sizeof(struct btrfs_item));
1996         }
1997         right_nritems -= push_items;
1998         btrfs_set_header_nritems(right, right_nritems);
1999         push_space = BTRFS_LEAF_DATA_SIZE(root);
2000         for (i = 0; i < right_nritems; i++) {
2001                 item = btrfs_item_nr(right, i);
2002                 push_space = push_space - btrfs_item_size(right, item);
2003                 btrfs_set_item_offset(right, item, push_space);
2004         }
2005
2006         btrfs_mark_buffer_dirty(left);
2007         if (right_nritems)
2008                 btrfs_mark_buffer_dirty(right);
2009
2010         btrfs_item_key(right, &disk_key, 0);
2011         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
2012         if (wret)
2013                 ret = wret;
2014
2015         /* then fixup the leaf pointer in the path */
2016         if (path->slots[0] < push_items) {
2017                 path->slots[0] += old_left_nritems;
2018                 free_extent_buffer(path->nodes[0]);
2019                 path->nodes[0] = left;
2020                 path->slots[1] -= 1;
2021         } else {
2022                 free_extent_buffer(left);
2023                 path->slots[0] -= push_items;
2024         }
2025         BUG_ON(path->slots[0] < 0);
2026         return ret;
2027 }
2028
2029 /*
2030  * split the path's leaf in two, making sure there is at least data_size
2031  * available for the resulting leaf level of the path.
2032  *
2033  * returns 0 if all went well and < 0 on failure.
2034  */
2035 static noinline int copy_for_split(struct btrfs_trans_handle *trans,
2036                                struct btrfs_root *root,
2037                                struct btrfs_path *path,
2038                                struct extent_buffer *l,
2039                                struct extent_buffer *right,
2040                                int slot, int mid, int nritems)
2041 {
2042         int data_copy_size;
2043         int rt_data_off;
2044         int i;
2045         int ret = 0;
2046         int wret;
2047         struct btrfs_disk_key disk_key;
2048
2049         nritems = nritems - mid;
2050         btrfs_set_header_nritems(right, nritems);
2051         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2052
2053         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2054                            btrfs_item_nr_offset(mid),
2055                            nritems * sizeof(struct btrfs_item));
2056
2057         copy_extent_buffer(right, l,
2058                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2059                      data_copy_size, btrfs_leaf_data(l) +
2060                      leaf_data_end(root, l), data_copy_size);
2061
2062         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2063                       btrfs_item_end_nr(l, mid);
2064
2065         for (i = 0; i < nritems; i++) {
2066                 struct btrfs_item *item = btrfs_item_nr(right, i);
2067                 u32 ioff = btrfs_item_offset(right, item);
2068                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2069         }
2070
2071         btrfs_set_header_nritems(l, mid);
2072         ret = 0;
2073         btrfs_item_key(right, &disk_key, 0);
2074         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2075                           path->slots[1] + 1, 1);
2076         if (wret)
2077                 ret = wret;
2078
2079         btrfs_mark_buffer_dirty(right);
2080         btrfs_mark_buffer_dirty(l);
2081         BUG_ON(path->slots[0] != slot);
2082
2083         if (mid <= slot) {
2084                 free_extent_buffer(path->nodes[0]);
2085                 path->nodes[0] = right;
2086                 path->slots[0] -= mid;
2087                 path->slots[1] += 1;
2088         } else {
2089                 free_extent_buffer(right);
2090         }
2091
2092         BUG_ON(path->slots[0] < 0);
2093
2094         return ret;
2095 }
2096
2097 /*
2098  * split the path's leaf in two, making sure there is at least data_size
2099  * available for the resulting leaf level of the path.
2100  *
2101  * returns 0 if all went well and < 0 on failure.
2102  */
2103 static noinline int split_leaf(struct btrfs_trans_handle *trans,
2104                                struct btrfs_root *root,
2105                                struct btrfs_key *ins_key,
2106                                struct btrfs_path *path, int data_size,
2107                                int extend)
2108 {
2109         struct btrfs_disk_key disk_key;
2110         struct extent_buffer *l;
2111         u32 nritems;
2112         int mid;
2113         int slot;
2114         struct extent_buffer *right;
2115         int ret = 0;
2116         int wret;
2117         int split;
2118         int num_doubles = 0;
2119
2120         /* first try to make some room by pushing left and right */
2121         if (data_size && ins_key->type != BTRFS_DIR_ITEM_KEY) {
2122                 wret = push_leaf_right(trans, root, path, data_size, 0);
2123                 if (wret < 0)
2124                         return wret;
2125                 if (wret) {
2126                         wret = push_leaf_left(trans, root, path, data_size, 0);
2127                         if (wret < 0)
2128                                 return wret;
2129                 }
2130                 l = path->nodes[0];
2131
2132                 /* did the pushes work? */
2133                 if (btrfs_leaf_free_space(root, l) >= data_size)
2134                         return 0;
2135         }
2136
2137         if (!path->nodes[1]) {
2138                 ret = insert_new_root(trans, root, path, 1);
2139                 if (ret)
2140                         return ret;
2141         }
2142 again:
2143         split = 1;
2144         l = path->nodes[0];
2145         slot = path->slots[0];
2146         nritems = btrfs_header_nritems(l);
2147         mid = (nritems + 1) / 2;
2148
2149         if (mid <= slot) {
2150                 if (nritems == 1 ||
2151                     leaf_space_used(l, mid, nritems - mid) + data_size >
2152                         BTRFS_LEAF_DATA_SIZE(root)) {
2153                         if (slot >= nritems) {
2154                                 split = 0;
2155                         } else {
2156                                 mid = slot;
2157                                 if (mid != nritems &&
2158                                     leaf_space_used(l, mid, nritems - mid) +
2159                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2160                                         split = 2;
2161                                 }
2162                         }
2163                 }
2164         } else {
2165                 if (leaf_space_used(l, 0, mid) + data_size >
2166                         BTRFS_LEAF_DATA_SIZE(root)) {
2167                         if (!extend && data_size && slot == 0) {
2168                                 split = 0;
2169                         } else if ((extend || !data_size) && slot == 0) {
2170                                 mid = 1;
2171                         } else {
2172                                 mid = slot;
2173                                 if (mid != nritems &&
2174                                     leaf_space_used(l, mid, nritems - mid) +
2175                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2176                                         split = 2 ;
2177                                 }
2178                         }
2179                 }
2180         }
2181
2182         if (split == 0)
2183                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2184         else
2185                 btrfs_item_key(l, &disk_key, mid);
2186
2187         right = btrfs_alloc_free_block(trans, root, root->leafsize,
2188                                         root->root_key.objectid,
2189                                         &disk_key, 0, l->start, 0);
2190         if (IS_ERR(right)) {
2191                 BUG_ON(1);
2192                 return PTR_ERR(right);
2193         }
2194
2195         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2196         btrfs_set_header_bytenr(right, right->start);
2197         btrfs_set_header_generation(right, trans->transid);
2198         btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
2199         btrfs_set_header_owner(right, root->root_key.objectid);
2200         btrfs_set_header_level(right, 0);
2201         write_extent_buffer(right, root->fs_info->fsid,
2202                             (unsigned long)btrfs_header_fsid(right),
2203                             BTRFS_FSID_SIZE);
2204
2205         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2206                             (unsigned long)btrfs_header_chunk_tree_uuid(right),
2207                             BTRFS_UUID_SIZE);
2208
2209         if (split == 0) {
2210                 if (mid <= slot) {
2211                         btrfs_set_header_nritems(right, 0);
2212                         wret = insert_ptr(trans, root, path,
2213                                           &disk_key, right->start,
2214                                           path->slots[1] + 1, 1);
2215                         if (wret)
2216                                 ret = wret;
2217
2218                         free_extent_buffer(path->nodes[0]);
2219                         path->nodes[0] = right;
2220                         path->slots[0] = 0;
2221                         path->slots[1] += 1;
2222                 } else {
2223                         btrfs_set_header_nritems(right, 0);
2224                         wret = insert_ptr(trans, root, path,
2225                                           &disk_key,
2226                                           right->start,
2227                                           path->slots[1], 1);
2228                         if (wret)
2229                                 ret = wret;
2230                         free_extent_buffer(path->nodes[0]);
2231                         path->nodes[0] = right;
2232                         path->slots[0] = 0;
2233                         if (path->slots[1] == 0) {
2234                                 wret = fixup_low_keys(trans, root,
2235                                                 path, &disk_key, 1);
2236                                 if (wret)
2237                                         ret = wret;
2238                         }
2239                 }
2240                 btrfs_mark_buffer_dirty(right);
2241                 return ret;
2242         }
2243
2244         ret = copy_for_split(trans, root, path, l, right, slot, mid, nritems);
2245         BUG_ON(ret);
2246
2247         if (split == 2) {
2248                 BUG_ON(num_doubles != 0);
2249                 num_doubles++;
2250                 goto again;
2251         }
2252
2253         return ret;
2254 }
2255
2256 /*
2257  * This function splits a single item into two items,
2258  * giving 'new_key' to the new item and splitting the
2259  * old one at split_offset (from the start of the item).
2260  *
2261  * The path may be released by this operation.  After
2262  * the split, the path is pointing to the old item.  The
2263  * new item is going to be in the same node as the old one.
2264  *
2265  * Note, the item being split must be smaller enough to live alone on
2266  * a tree block with room for one extra struct btrfs_item
2267  *
2268  * This allows us to split the item in place, keeping a lock on the
2269  * leaf the entire time.
2270  */
2271 int btrfs_split_item(struct btrfs_trans_handle *trans,
2272                      struct btrfs_root *root,
2273                      struct btrfs_path *path,
2274                      struct btrfs_key *new_key,
2275                      unsigned long split_offset)
2276 {
2277         u32 item_size;
2278         struct extent_buffer *leaf;
2279         struct btrfs_key orig_key;
2280         struct btrfs_item *item;
2281         struct btrfs_item *new_item;
2282         int ret = 0;
2283         int slot;
2284         u32 nritems;
2285         u32 orig_offset;
2286         struct btrfs_disk_key disk_key;
2287         char *buf;
2288
2289         leaf = path->nodes[0];
2290         btrfs_item_key_to_cpu(leaf, &orig_key, path->slots[0]);
2291         if (btrfs_leaf_free_space(root, leaf) >= sizeof(struct btrfs_item))
2292                 goto split;
2293
2294         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2295         btrfs_release_path(root, path);
2296
2297         path->search_for_split = 1;
2298
2299         ret = btrfs_search_slot(trans, root, &orig_key, path, 0, 1);
2300         path->search_for_split = 0;
2301
2302         /* if our item isn't there or got smaller, return now */
2303         if (ret != 0 || item_size != btrfs_item_size_nr(path->nodes[0],
2304                                                         path->slots[0])) {
2305                 return -EAGAIN;
2306         }
2307
2308         ret = split_leaf(trans, root, &orig_key, path, 0, 0);
2309         BUG_ON(ret);
2310
2311         BUG_ON(btrfs_leaf_free_space(root, leaf) < sizeof(struct btrfs_item));
2312         leaf = path->nodes[0];
2313
2314 split:
2315         item = btrfs_item_nr(leaf, path->slots[0]);
2316         orig_offset = btrfs_item_offset(leaf, item);
2317         item_size = btrfs_item_size(leaf, item);
2318
2319
2320         buf = kmalloc(item_size, GFP_NOFS);
2321         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
2322                             path->slots[0]), item_size);
2323         slot = path->slots[0] + 1;
2324         leaf = path->nodes[0];
2325
2326         nritems = btrfs_header_nritems(leaf);
2327
2328         if (slot != nritems) {
2329                 /* shift the items */
2330                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
2331                               btrfs_item_nr_offset(slot),
2332                               (nritems - slot) * sizeof(struct btrfs_item));
2333
2334         }
2335
2336         btrfs_cpu_key_to_disk(&disk_key, new_key);
2337         btrfs_set_item_key(leaf, &disk_key, slot);
2338
2339         new_item = btrfs_item_nr(leaf, slot);
2340
2341         btrfs_set_item_offset(leaf, new_item, orig_offset);
2342         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
2343
2344         btrfs_set_item_offset(leaf, item,
2345                               orig_offset + item_size - split_offset);
2346         btrfs_set_item_size(leaf, item, split_offset);
2347
2348         btrfs_set_header_nritems(leaf, nritems + 1);
2349
2350         /* write the data for the start of the original item */
2351         write_extent_buffer(leaf, buf,
2352                             btrfs_item_ptr_offset(leaf, path->slots[0]),
2353                             split_offset);
2354
2355         /* write the data for the new item */
2356         write_extent_buffer(leaf, buf + split_offset,
2357                             btrfs_item_ptr_offset(leaf, slot),
2358                             item_size - split_offset);
2359         btrfs_mark_buffer_dirty(leaf);
2360
2361         ret = 0;
2362         if (btrfs_leaf_free_space(root, leaf) < 0) {
2363                 btrfs_print_leaf(root, leaf);
2364                 BUG();
2365         }
2366         kfree(buf);
2367         return ret;
2368 }
2369
2370 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2371                         struct btrfs_root *root,
2372                         struct btrfs_path *path,
2373                         u32 new_size, int from_end)
2374 {
2375         int ret = 0;
2376         int slot;
2377         struct extent_buffer *leaf;
2378         struct btrfs_item *item;
2379         u32 nritems;
2380         unsigned int data_end;
2381         unsigned int old_data_start;
2382         unsigned int old_size;
2383         unsigned int size_diff;
2384         int i;
2385
2386         leaf = path->nodes[0];
2387         slot = path->slots[0];
2388
2389         old_size = btrfs_item_size_nr(leaf, slot);
2390         if (old_size == new_size)
2391                 return 0;
2392
2393         nritems = btrfs_header_nritems(leaf);
2394         data_end = leaf_data_end(root, leaf);
2395
2396         old_data_start = btrfs_item_offset_nr(leaf, slot);
2397
2398         size_diff = old_size - new_size;
2399
2400         BUG_ON(slot < 0);
2401         BUG_ON(slot >= nritems);
2402
2403         /*
2404          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2405          */
2406         /* first correct the data pointers */
2407         for (i = slot; i < nritems; i++) {
2408                 u32 ioff;
2409                 item = btrfs_item_nr(leaf, i);
2410                 ioff = btrfs_item_offset(leaf, item);
2411                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2412         }
2413
2414         /* shift the data */
2415         if (from_end) {
2416                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2417                               data_end + size_diff, btrfs_leaf_data(leaf) +
2418                               data_end, old_data_start + new_size - data_end);
2419         } else {
2420                 struct btrfs_disk_key disk_key;
2421                 u64 offset;
2422
2423                 btrfs_item_key(leaf, &disk_key, slot);
2424
2425                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2426                         unsigned long ptr;
2427                         struct btrfs_file_extent_item *fi;
2428
2429                         fi = btrfs_item_ptr(leaf, slot,
2430                                             struct btrfs_file_extent_item);
2431                         fi = (struct btrfs_file_extent_item *)(
2432                              (unsigned long)fi - size_diff);
2433
2434                         if (btrfs_file_extent_type(leaf, fi) ==
2435                             BTRFS_FILE_EXTENT_INLINE) {
2436                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2437                                 memmove_extent_buffer(leaf, ptr,
2438                                         (unsigned long)fi,
2439                                         offsetof(struct btrfs_file_extent_item,
2440                                                  disk_bytenr));
2441                         }
2442                 }
2443
2444                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2445                               data_end + size_diff, btrfs_leaf_data(leaf) +
2446                               data_end, old_data_start - data_end);
2447
2448                 offset = btrfs_disk_key_offset(&disk_key);
2449                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2450                 btrfs_set_item_key(leaf, &disk_key, slot);
2451                 if (slot == 0)
2452                         fixup_low_keys(trans, root, path, &disk_key, 1);
2453         }
2454
2455         item = btrfs_item_nr(leaf, slot);
2456         btrfs_set_item_size(leaf, item, new_size);
2457         btrfs_mark_buffer_dirty(leaf);
2458
2459         ret = 0;
2460         if (btrfs_leaf_free_space(root, leaf) < 0) {
2461                 btrfs_print_leaf(root, leaf);
2462                 BUG();
2463         }
2464         return ret;
2465 }
2466
2467 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2468                       struct btrfs_root *root, struct btrfs_path *path,
2469                       u32 data_size)
2470 {
2471         int ret = 0;
2472         int slot;
2473         struct extent_buffer *leaf;
2474         struct btrfs_item *item;
2475         u32 nritems;
2476         unsigned int data_end;
2477         unsigned int old_data;
2478         unsigned int old_size;
2479         int i;
2480
2481         leaf = path->nodes[0];
2482
2483         nritems = btrfs_header_nritems(leaf);
2484         data_end = leaf_data_end(root, leaf);
2485
2486         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2487                 btrfs_print_leaf(root, leaf);
2488                 BUG();
2489         }
2490         slot = path->slots[0];
2491         old_data = btrfs_item_end_nr(leaf, slot);
2492
2493         BUG_ON(slot < 0);
2494         if (slot >= nritems) {
2495                 btrfs_print_leaf(root, leaf);
2496                 printk("slot %d too large, nritems %d\n", slot, nritems);
2497                 BUG_ON(1);
2498         }
2499
2500         /*
2501          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2502          */
2503         /* first correct the data pointers */
2504         for (i = slot; i < nritems; i++) {
2505                 u32 ioff;
2506                 item = btrfs_item_nr(leaf, i);
2507                 ioff = btrfs_item_offset(leaf, item);
2508                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2509         }
2510
2511         /* shift the data */
2512         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2513                       data_end - data_size, btrfs_leaf_data(leaf) +
2514                       data_end, old_data - data_end);
2515
2516         data_end = old_data;
2517         old_size = btrfs_item_size_nr(leaf, slot);
2518         item = btrfs_item_nr(leaf, slot);
2519         btrfs_set_item_size(leaf, item, old_size + data_size);
2520         btrfs_mark_buffer_dirty(leaf);
2521
2522         ret = 0;
2523         if (btrfs_leaf_free_space(root, leaf) < 0) {
2524                 btrfs_print_leaf(root, leaf);
2525                 BUG();
2526         }
2527         return ret;
2528 }
2529
2530 /*
2531  * Given a key and some data, insert an item into the tree.
2532  * This does all the path init required, making room in the tree if needed.
2533  */
2534 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2535                             struct btrfs_root *root,
2536                             struct btrfs_path *path,
2537                             struct btrfs_key *cpu_key, u32 *data_size,
2538                             int nr)
2539 {
2540         struct extent_buffer *leaf;
2541         struct btrfs_item *item;
2542         int ret = 0;
2543         int slot;
2544         int i;
2545         u32 nritems;
2546         u32 total_size = 0;
2547         u32 total_data = 0;
2548         unsigned int data_end;
2549         struct btrfs_disk_key disk_key;
2550
2551         for (i = 0; i < nr; i++) {
2552                 total_data += data_size[i];
2553         }
2554
2555         /* create a root if there isn't one */
2556         if (!root->node)
2557                 BUG();
2558
2559         total_size = total_data + nr * sizeof(struct btrfs_item);
2560         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2561         if (ret == 0) {
2562                 return -EEXIST;
2563         }
2564         if (ret < 0)
2565                 goto out;
2566
2567         leaf = path->nodes[0];
2568
2569         nritems = btrfs_header_nritems(leaf);
2570         data_end = leaf_data_end(root, leaf);
2571
2572         if (btrfs_leaf_free_space(root, leaf) < total_size) {
2573                 btrfs_print_leaf(root, leaf);
2574                 printk("not enough freespace need %u have %d\n",
2575                        total_size, btrfs_leaf_free_space(root, leaf));
2576                 BUG();
2577         }
2578
2579         slot = path->slots[0];
2580         BUG_ON(slot < 0);
2581
2582         if (slot != nritems) {
2583                 int i;
2584                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2585
2586                 if (old_data < data_end) {
2587                         btrfs_print_leaf(root, leaf);
2588                         printk("slot %d old_data %d data_end %d\n",
2589                                slot, old_data, data_end);
2590                         BUG_ON(1);
2591                 }
2592                 /*
2593                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2594                  */
2595                 /* first correct the data pointers */
2596                 for (i = slot; i < nritems; i++) {
2597                         u32 ioff;
2598
2599                         item = btrfs_item_nr(leaf, i);
2600                         ioff = btrfs_item_offset(leaf, item);
2601                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2602                 }
2603
2604                 /* shift the items */
2605                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2606                               btrfs_item_nr_offset(slot),
2607                               (nritems - slot) * sizeof(struct btrfs_item));
2608
2609                 /* shift the data */
2610                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2611                               data_end - total_data, btrfs_leaf_data(leaf) +
2612                               data_end, old_data - data_end);
2613                 data_end = old_data;
2614         }
2615
2616         /* setup the item for the new data */
2617         for (i = 0; i < nr; i++) {
2618                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2619                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2620                 item = btrfs_item_nr(leaf, slot + i);
2621                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2622                 data_end -= data_size[i];
2623                 btrfs_set_item_size(leaf, item, data_size[i]);
2624         }
2625         btrfs_set_header_nritems(leaf, nritems + nr);
2626         btrfs_mark_buffer_dirty(leaf);
2627
2628         ret = 0;
2629         if (slot == 0) {
2630                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2631                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2632         }
2633
2634         if (btrfs_leaf_free_space(root, leaf) < 0) {
2635                 btrfs_print_leaf(root, leaf);
2636                 BUG();
2637         }
2638
2639 out:
2640         return ret;
2641 }
2642
2643 /*
2644  * Given a key and some data, insert an item into the tree.
2645  * This does all the path init required, making room in the tree if needed.
2646  */
2647 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2648                       *root, struct btrfs_key *cpu_key, void *data, u32
2649                       data_size)
2650 {
2651         int ret = 0;
2652         struct btrfs_path *path;
2653         struct extent_buffer *leaf;
2654         unsigned long ptr;
2655
2656         path = btrfs_alloc_path();
2657         BUG_ON(!path);
2658         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2659         if (!ret) {
2660                 leaf = path->nodes[0];
2661                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2662                 write_extent_buffer(leaf, data, ptr, data_size);
2663                 btrfs_mark_buffer_dirty(leaf);
2664         }
2665         btrfs_free_path(path);
2666         return ret;
2667 }
2668
2669 /*
2670  * delete the pointer from a given node.
2671  *
2672  * If the delete empties a node, the node is removed from the tree,
2673  * continuing all the way the root if required.  The root is converted into
2674  * a leaf if all the nodes are emptied.
2675  */
2676 int btrfs_del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2677                    struct btrfs_path *path, int level, int slot)
2678 {
2679         struct extent_buffer *parent = path->nodes[level];
2680         u32 nritems;
2681         int ret = 0;
2682         int wret;
2683
2684         nritems = btrfs_header_nritems(parent);
2685         if (slot != nritems -1) {
2686                 memmove_extent_buffer(parent,
2687                               btrfs_node_key_ptr_offset(slot),
2688                               btrfs_node_key_ptr_offset(slot + 1),
2689                               sizeof(struct btrfs_key_ptr) *
2690                               (nritems - slot - 1));
2691         }
2692         nritems--;
2693         btrfs_set_header_nritems(parent, nritems);
2694         if (nritems == 0 && parent == root->node) {
2695                 BUG_ON(btrfs_header_level(root->node) != 1);
2696                 /* just turn the root into a leaf and break */
2697                 btrfs_set_header_level(root->node, 0);
2698         } else if (slot == 0) {
2699                 struct btrfs_disk_key disk_key;
2700
2701                 btrfs_node_key(parent, &disk_key, 0);
2702                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2703                 if (wret)
2704                         ret = wret;
2705         }
2706         btrfs_mark_buffer_dirty(parent);
2707         return ret;
2708 }
2709
2710 /*
2711  * a helper function to delete the leaf pointed to by path->slots[1] and
2712  * path->nodes[1].
2713  *
2714  * This deletes the pointer in path->nodes[1] and frees the leaf
2715  * block extent.  zero is returned if it all worked out, < 0 otherwise.
2716  *
2717  * The path must have already been setup for deleting the leaf, including
2718  * all the proper balancing.  path->nodes[1] must be locked.
2719  */
2720 static noinline int btrfs_del_leaf(struct btrfs_trans_handle *trans,
2721                                    struct btrfs_root *root,
2722                                    struct btrfs_path *path,
2723                                    struct extent_buffer *leaf)
2724 {
2725         int ret;
2726
2727         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
2728         ret = btrfs_del_ptr(trans, root, path, 1, path->slots[1]);
2729         if (ret)
2730                 return ret;
2731
2732         ret = btrfs_free_extent(trans, root, leaf->start, leaf->len,
2733                                 0, root->root_key.objectid, 0, 0);
2734         return ret;
2735 }
2736
2737 /*
2738  * delete the item at the leaf level in path.  If that empties
2739  * the leaf, remove it from the tree
2740  */
2741 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2742                     struct btrfs_path *path, int slot, int nr)
2743 {
2744         struct extent_buffer *leaf;
2745         struct btrfs_item *item;
2746         int last_off;
2747         int dsize = 0;
2748         int ret = 0;
2749         int wret;
2750         int i;
2751         u32 nritems;
2752
2753         leaf = path->nodes[0];
2754         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2755
2756         for (i = 0; i < nr; i++)
2757                 dsize += btrfs_item_size_nr(leaf, slot + i);
2758
2759         nritems = btrfs_header_nritems(leaf);
2760
2761         if (slot + nr != nritems) {
2762                 int i;
2763                 int data_end = leaf_data_end(root, leaf);
2764
2765                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2766                               data_end + dsize,
2767                               btrfs_leaf_data(leaf) + data_end,
2768                               last_off - data_end);
2769
2770                 for (i = slot + nr; i < nritems; i++) {
2771                         u32 ioff;
2772
2773                         item = btrfs_item_nr(leaf, i);
2774                         ioff = btrfs_item_offset(leaf, item);
2775                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2776                 }
2777
2778                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2779                               btrfs_item_nr_offset(slot + nr),
2780                               sizeof(struct btrfs_item) *
2781                               (nritems - slot - nr));
2782         }
2783         btrfs_set_header_nritems(leaf, nritems - nr);
2784         nritems -= nr;
2785
2786         /* delete the leaf if we've emptied it */
2787         if (nritems == 0) {
2788                 if (leaf == root->node) {
2789                         btrfs_set_header_level(leaf, 0);
2790                 } else {
2791                         clean_tree_block(trans, root, leaf);
2792                         wait_on_tree_block_writeback(root, leaf);
2793
2794                         wret = btrfs_del_leaf(trans, root, path, leaf);
2795                         BUG_ON(ret);
2796                         if (wret)
2797                                 ret = wret;
2798                 }
2799         } else {
2800                 int used = leaf_space_used(leaf, 0, nritems);
2801                 if (slot == 0) {
2802                         struct btrfs_disk_key disk_key;
2803
2804                         btrfs_item_key(leaf, &disk_key, 0);
2805                         wret = fixup_low_keys(trans, root, path,
2806                                               &disk_key, 1);
2807                         if (wret)
2808                                 ret = wret;
2809                 }
2810
2811                 /* delete the leaf if it is mostly empty */
2812                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2813                         /* push_leaf_left fixes the path.
2814                          * make sure the path still points to our leaf
2815                          * for possible call to del_ptr below
2816                          */
2817                         slot = path->slots[1];
2818                         extent_buffer_get(leaf);
2819
2820                         wret = push_leaf_left(trans, root, path, 1, 1);
2821                         if (wret < 0 && wret != -ENOSPC)
2822                                 ret = wret;
2823
2824                         if (path->nodes[0] == leaf &&
2825                             btrfs_header_nritems(leaf)) {
2826                                 wret = push_leaf_right(trans, root, path, 1, 1);
2827                                 if (wret < 0 && wret != -ENOSPC)
2828                                         ret = wret;
2829                         }
2830
2831                         if (btrfs_header_nritems(leaf) == 0) {
2832                                 clean_tree_block(trans, root, leaf);
2833                                 wait_on_tree_block_writeback(root, leaf);
2834
2835                                 path->slots[1] = slot;
2836                                 ret = btrfs_del_leaf(trans, root, path, leaf);
2837                                 BUG_ON(ret);
2838                                 free_extent_buffer(leaf);
2839
2840                         } else {
2841                                 btrfs_mark_buffer_dirty(leaf);
2842                                 free_extent_buffer(leaf);
2843                         }
2844                 } else {
2845                         btrfs_mark_buffer_dirty(leaf);
2846                 }
2847         }
2848         return ret;
2849 }
2850
2851 /*
2852  * walk up the tree as far as required to find the previous leaf.
2853  * returns 0 if it found something or 1 if there are no lesser leaves.
2854  * returns < 0 on io errors.
2855  */
2856 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2857 {
2858         int slot;
2859         int level = 1;
2860         struct extent_buffer *c;
2861         struct extent_buffer *next = NULL;
2862
2863         while(level < BTRFS_MAX_LEVEL) {
2864                 if (!path->nodes[level])
2865                         return 1;
2866
2867                 slot = path->slots[level];
2868                 c = path->nodes[level];
2869                 if (slot == 0) {
2870                         level++;
2871                         if (level == BTRFS_MAX_LEVEL)
2872                                 return 1;
2873                         continue;
2874                 }
2875                 slot--;
2876
2877                 next = read_node_slot(root, c, slot);
2878                 break;
2879         }
2880         path->slots[level] = slot;
2881         while(1) {
2882                 level--;
2883                 c = path->nodes[level];
2884                 free_extent_buffer(c);
2885                 slot = btrfs_header_nritems(next);
2886                 if (slot != 0)
2887                         slot--;
2888                 path->nodes[level] = next;
2889                 path->slots[level] = slot;
2890                 if (!level)
2891                         break;
2892                 next = read_node_slot(root, next, slot);
2893         }
2894         return 0;
2895 }
2896
2897 /*
2898  * walk up the tree as far as required to find the next leaf.
2899  * returns 0 if it found something or 1 if there are no greater leaves.
2900  * returns < 0 on io errors.
2901  */
2902 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2903 {
2904         int slot;
2905         int level = 1;
2906         struct extent_buffer *c;
2907         struct extent_buffer *next = NULL;
2908
2909         while(level < BTRFS_MAX_LEVEL) {
2910                 if (!path->nodes[level])
2911                         return 1;
2912
2913                 slot = path->slots[level] + 1;
2914                 c = path->nodes[level];
2915                 if (slot >= btrfs_header_nritems(c)) {
2916                         level++;
2917                         if (level == BTRFS_MAX_LEVEL)
2918                                 return 1;
2919                         continue;
2920                 }
2921
2922                 if (path->reada)
2923                         reada_for_search(root, path, level, slot, 0);
2924
2925                 next = read_node_slot(root, c, slot);
2926                 if (!next)
2927                         return -EIO;
2928                 break;
2929         }
2930         path->slots[level] = slot;
2931         while(1) {
2932                 level--;
2933                 c = path->nodes[level];
2934                 free_extent_buffer(c);
2935                 path->nodes[level] = next;
2936                 path->slots[level] = 0;
2937                 if (!level)
2938                         break;
2939                 if (path->reada)
2940                         reada_for_search(root, path, level, 0, 0);
2941                 next = read_node_slot(root, next, 0);
2942                 if (!next)
2943                         return -EIO;
2944         }
2945         return 0;
2946 }
2947
2948 int btrfs_previous_item(struct btrfs_root *root,
2949                         struct btrfs_path *path, u64 min_objectid,
2950                         int type)
2951 {
2952         struct btrfs_key found_key;
2953         struct extent_buffer *leaf;
2954         int ret;
2955
2956         while(1) {
2957                 if (path->slots[0] == 0) {
2958                         ret = btrfs_prev_leaf(root, path);
2959                         if (ret != 0)
2960                                 return ret;
2961                 } else {
2962                         path->slots[0]--;
2963                 }
2964                 leaf = path->nodes[0];
2965                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2966                 if (found_key.type == type)
2967                         return 0;
2968         }
2969         return 1;
2970 }
2971