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 #include "internal.h"
  24 #include "sizes.h"
  25
  26 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  27                       *root, struct btrfs_path *path, int level);
  28 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  29                       *root, struct btrfs_key *ins_key,
  30                       struct btrfs_path *path, int data_size, int extend);
  31 static int push_node_left(struct btrfs_trans_handle *trans,
  32                           struct btrfs_root *root, struct extent_buffer *dst,
  33                           struct extent_buffer *src, int empty);
  34 static int balance_node_right(struct btrfs_trans_handle *trans,
  35                               struct btrfs_root *root,
  36                               struct extent_buffer *dst_buf,
  37                               struct extent_buffer *src_buf);
  38
  39 inline void btrfs_init_path(struct btrfs_path *p)
  40 {
  41         memset(p, 0, sizeof(*p));
  42 }
  43
  44 struct btrfs_path *btrfs_alloc_path(void)
  45 {
  46         struct btrfs_path *path;
  47         path = kzalloc(sizeof(struct btrfs_path), GFP_NOFS);
  48         return path;
  49 }
  50
  51 void btrfs_free_path(struct btrfs_path *p)
  52 {
  53         if (!p)
  54                 return;
  55         btrfs_release_path(p);
  56         kfree(p);
  57 }
  58
  59 void btrfs_release_path(struct btrfs_path *p)
  60 {
  61         int i;
  62         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  63                 if (!p->nodes[i])
  64                         continue;
  65                 free_extent_buffer(p->nodes[i]);
  66         }
  67         memset(p, 0, sizeof(*p));
  68 }
  69
  70 void add_root_to_dirty_list(struct btrfs_root *root)
  71 {
  72         if (root->track_dirty && list_empty(&root->dirty_list)) {
  73                 list_add(&root->dirty_list,
  74                          &root->fs_info->dirty_cowonly_roots);
  75         }
  76 }
  77
  78 int btrfs_copy_root(struct btrfs_trans_handle *trans,
  79                       struct btrfs_root *root,
  80                       struct extent_buffer *buf,
  81                       struct extent_buffer **cow_ret, u64 new_root_objectid)
  82 {
  83         struct extent_buffer *cow;
  84         int ret = 0;
  85         int level;
  86         struct btrfs_root *new_root;
  87         struct btrfs_disk_key disk_key;
  88
  89         new_root = kmalloc(sizeof(*new_root), GFP_NOFS);
  90         if (!new_root)
  91                 return -ENOMEM;
  92
  93         memcpy(new_root, root, sizeof(*new_root));
  94         new_root->root_key.objectid = new_root_objectid;
  95
  96         WARN_ON(root->ref_cows && trans->transid !=
  97                 root->fs_info->running_transaction->transid);
  98         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
  99
 100         level = btrfs_header_level(buf);
 101         if (level == 0)
 102                 btrfs_item_key(buf, &disk_key, 0);
 103         else
 104                 btrfs_node_key(buf, &disk_key, 0);
 105         cow = btrfs_alloc_free_block(trans, new_root, buf->len,
 106                                      new_root_objectid, &disk_key,
 107                                      level, buf->start, 0);
 108         if (IS_ERR(cow)) {
 109                 kfree(new_root);
 110                 return PTR_ERR(cow);
 111         }
 112
 113         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 114         btrfs_set_header_bytenr(cow, cow->start);
 115         btrfs_set_header_generation(cow, trans->transid);
 116         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 117         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 118                                      BTRFS_HEADER_FLAG_RELOC);
 119         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 120                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 121         else
 122                 btrfs_set_header_owner(cow, new_root_objectid);
 123
 124         write_extent_buffer(cow, root->fs_info->fsid,
 125                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
 126
 127         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 128         ret = btrfs_inc_ref(trans, new_root, cow, 0);
 129         kfree(new_root);
 130
 131         if (ret)
 132                 return ret;
 133
 134         btrfs_mark_buffer_dirty(cow);
 135         *cow_ret = cow;
 136         return 0;
 137 }
 138
 139 /*
 140  * check if the tree block can be shared by multiple trees
 141  */
 142 static int btrfs_block_can_be_shared(struct btrfs_root *root,
 143                                      struct extent_buffer *buf)
 144 {
 145         /*
 146          * Tree blocks not in reference counted trees and tree roots
 147          * are never shared. If a block was allocated after the last
 148          * snapshot and the block was not allocated by tree relocation,
 149          * we know the block is not shared.
 150          */
 151         if (root->ref_cows &&
 152             buf != root->node && buf != root->commit_root &&
 153             (btrfs_header_generation(buf) <=
 154              btrfs_root_last_snapshot(&root->root_item) ||
 155              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 156                 return 1;
 157 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 158         if (root->ref_cows &&
 159             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 160                 return 1;
 161 #endif
 162         return 0;
 163 }
 164
 165 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
 166                                        struct btrfs_root *root,
 167                                        struct extent_buffer *buf,
 168                                        struct extent_buffer *cow)
 169 {
 170         u64 refs;
 171         u64 owner;
 172         u64 flags;
 173         u64 new_flags = 0;
 174         int ret;
 175
 176         /*
 177          * Backrefs update rules:
 178          *
 179          * Always use full backrefs for extent pointers in tree block
 180          * allocated by tree relocation.
 181          *
 182          * If a shared tree block is no longer referenced by its owner
 183          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
 184          * use full backrefs for extent pointers in tree block.
 185          *
 186          * If a tree block is been relocating
 187          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
 188          * use full backrefs for extent pointers in tree block.
 189          * The reason for this is some operations (such as drop tree)
 190          * are only allowed for blocks use full backrefs.
 191          */
 192
 193         if (btrfs_block_can_be_shared(root, buf)) {
 194                 ret = btrfs_lookup_extent_info(trans, root, buf->start,
 195                                                btrfs_header_level(buf), 1,
 196                                                &refs, &flags);
 197                 BUG_ON(ret);
 198                 BUG_ON(refs == 0);
 199         } else {
 200                 refs = 1;
 201                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 202                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 203                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 204                 else
 205                         flags = 0;
 206         }
 207
 208         owner = btrfs_header_owner(buf);
 209         BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) &&
 210                owner == BTRFS_TREE_RELOC_OBJECTID);
 211
 212         if (refs > 1) {
 213                 if ((owner == root->root_key.objectid ||
 214                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
 215                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
 216                         ret = btrfs_inc_ref(trans, root, buf, 1);
 217                         BUG_ON(ret);
 218
 219                         if (root->root_key.objectid ==
 220                             BTRFS_TREE_RELOC_OBJECTID) {
 221                                 ret = btrfs_dec_ref(trans, root, buf, 0);
 222                                 BUG_ON(ret);
 223                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 224                                 BUG_ON(ret);
 225                         }
 226                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
 227                 } else {
 228
 229                         if (root->root_key.objectid ==
 230                             BTRFS_TREE_RELOC_OBJECTID)
 231                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 232                         else
 233                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 234                         BUG_ON(ret);
 235                 }
 236                 if (new_flags != 0) {
 237                         ret = btrfs_set_block_flags(trans, root, buf->start,
 238                                                     btrfs_header_level(buf),
 239                                                     new_flags);
 240                         BUG_ON(ret);
 241                 }
 242         } else {
 243                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
 244                         if (root->root_key.objectid ==
 245                             BTRFS_TREE_RELOC_OBJECTID)
 246                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 247                         else
 248                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 249                         BUG_ON(ret);
 250                         ret = btrfs_dec_ref(trans, root, buf, 1);
 251                         BUG_ON(ret);
 252                 }
 253                 clean_tree_block(trans, root, buf);
 254         }
 255         return 0;
 256 }
 257
 258 int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 259                              struct btrfs_root *root,
 260                              struct extent_buffer *buf,
 261                              struct extent_buffer *parent, int parent_slot,
 262                              struct extent_buffer **cow_ret,
 263                              u64 search_start, u64 empty_size)
 264 {
 265         struct extent_buffer *cow;
 266         struct btrfs_disk_key disk_key;
 267         int level;
 268
 269         WARN_ON(root->ref_cows && trans->transid !=
 270                 root->fs_info->running_transaction->transid);
 271         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 272
 273         level = btrfs_header_level(buf);
 274
 275         if (level == 0)
 276                 btrfs_item_key(buf, &disk_key, 0);
 277         else
 278                 btrfs_node_key(buf, &disk_key, 0);
 279
 280         cow = btrfs_alloc_free_block(trans, root, buf->len,
 281                                      root->root_key.objectid, &disk_key,
 282                                      level, search_start, empty_size);
 283         if (IS_ERR(cow))
 284                 return PTR_ERR(cow);
 285
 286         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 287         btrfs_set_header_bytenr(cow, cow->start);
 288         btrfs_set_header_generation(cow, trans->transid);
 289         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 290         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 291                                      BTRFS_HEADER_FLAG_RELOC);
 292         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 293                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 294         else
 295                 btrfs_set_header_owner(cow, root->root_key.objectid);
 296
 297         write_extent_buffer(cow, root->fs_info->fsid,
 298                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
 299
 300         WARN_ON(!(buf->flags & EXTENT_BAD_TRANSID) &&
 301                 btrfs_header_generation(buf) > trans->transid);
 302
 303         update_ref_for_cow(trans, root, buf, cow);
 304
 305         if (buf == root->node) {
 306                 root->node = cow;
 307                 extent_buffer_get(cow);
 308
 309                 btrfs_free_extent(trans, root, buf->start, buf->len,
 310                                   0, root->root_key.objectid, level, 0);
 311                 free_extent_buffer(buf);
 312                 add_root_to_dirty_list(root);
 313         } else {
 314                 btrfs_set_node_blockptr(parent, parent_slot,
 315                                         cow->start);
 316                 WARN_ON(trans->transid == 0);
 317                 btrfs_set_node_ptr_generation(parent, parent_slot,
 318                                               trans->transid);
 319                 btrfs_mark_buffer_dirty(parent);
 320                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 321
 322                 btrfs_free_extent(trans, root, buf->start, buf->len,
 323                                   0, root->root_key.objectid, level, 1);
 324         }
 325         if (!list_empty(&buf->recow)) {
 326                 list_del_init(&buf->recow);
 327                 free_extent_buffer(buf);
 328         }
 329         free_extent_buffer(buf);
 330         btrfs_mark_buffer_dirty(cow);
 331         *cow_ret = cow;
 332         return 0;
 333 }
 334
 335 static inline int should_cow_block(struct btrfs_trans_handle *trans,
 336                                    struct btrfs_root *root,
 337                                    struct extent_buffer *buf)
 338 {
 339         if (btrfs_header_generation(buf) == trans->transid &&
 340             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
 341             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
 342               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 343                 return 0;
 344         return 1;
 345 }
 346
 347 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 348                     struct btrfs_root *root, struct extent_buffer *buf,
 349                     struct extent_buffer *parent, int parent_slot,
 350                     struct extent_buffer **cow_ret)
 351 {
 352         u64 search_start;
 353         int ret;
 354         /*
 355         if (trans->transaction != root->fs_info->running_transaction) {
 356                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 357                        root->fs_info->running_transaction->transid);
 358                 WARN_ON(1);
 359         }
 360         */
 361         if (trans->transid != root->fs_info->generation) {
 362                 printk(KERN_CRIT "trans %llu running %llu\n",
 363                         (unsigned long long)trans->transid,
 364                         (unsigned long long)root->fs_info->generation);
 365                 WARN_ON(1);
 366         }
 367         if (!should_cow_block(trans, root, buf)) {
 368                 *cow_ret = buf;
 369                 return 0;
 370         }
 371
 372         search_start = buf->start & ~((u64)SZ_1G - 1);
 373         ret = __btrfs_cow_block(trans, root, buf, parent,
 374                                  parent_slot, cow_ret, search_start, 0);
 375         return ret;
 376 }
 377
 378 int btrfs_comp_cpu_keys(struct btrfs_key *k1, struct btrfs_key *k2)
 379 {
 380         if (k1->objectid > k2->objectid)
 381                 return 1;
 382         if (k1->objectid < k2->objectid)
 383                 return -1;
 384         if (k1->type > k2->type)
 385                 return 1;
 386         if (k1->type < k2->type)
 387                 return -1;
 388         if (k1->offset > k2->offset)
 389                 return 1;
 390         if (k1->offset < k2->offset)
 391                 return -1;
 392         return 0;
 393 }
 394
 395 /*
 396  * compare two keys in a memcmp fashion
 397  */
 398 static int btrfs_comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 399 {
 400         struct btrfs_key k1;
 401
 402         btrfs_disk_key_to_cpu(&k1, disk);
 403         return btrfs_comp_cpu_keys(&k1, k2);
 404 }
 405
 406 /*
 407  * The leaf data grows from end-to-front in the node.
 408  * this returns the address of the start of the last item,
 409  * which is the stop of the leaf data stack
 410  */
 411 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 412                                          struct extent_buffer *leaf)
 413 {
 414         u32 nr = btrfs_header_nritems(leaf);
 415         if (nr == 0)
 416                 return BTRFS_LEAF_DATA_SIZE(root);
 417         return btrfs_item_offset_nr(leaf, nr - 1);
 418 }
 419
 420 enum btrfs_tree_block_status
 421 btrfs_check_node(struct btrfs_root *root, struct btrfs_disk_key *parent_key,
 422                  struct extent_buffer *buf)
 423 {
 424         int i;
 425         struct btrfs_key cpukey;
 426         struct btrfs_disk_key key;
 427         u32 nritems = btrfs_header_nritems(buf);
 428         enum btrfs_tree_block_status ret = BTRFS_TREE_BLOCK_INVALID_NRITEMS;
 429
 430         if (nritems == 0 || nritems > BTRFS_NODEPTRS_PER_BLOCK(root))
 431                 goto fail;
 432
 433         ret = BTRFS_TREE_BLOCK_INVALID_PARENT_KEY;
 434         if (parent_key && parent_key->type) {
 435                 btrfs_node_key(buf, &key, 0);
 436                 if (memcmp(parent_key, &key, sizeof(key)))
 437                         goto fail;
 438         }
 439         ret = BTRFS_TREE_BLOCK_BAD_KEY_ORDER;
 440         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 441                 btrfs_node_key(buf, &key, i);
 442                 btrfs_node_key_to_cpu(buf, &cpukey, i + 1);
 443                 if (btrfs_comp_keys(&key, &cpukey) >= 0)
 444                         goto fail;
 445         }
 446         return BTRFS_TREE_BLOCK_CLEAN;
 447 fail:
 448         if (btrfs_header_owner(buf) == BTRFS_EXTENT_TREE_OBJECTID) {
 449                 if (parent_key)
 450                         btrfs_disk_key_to_cpu(&cpukey, parent_key);
 451                 else
 452                         btrfs_node_key_to_cpu(buf, &cpukey, 0);
 453                 btrfs_add_corrupt_extent_record(root->fs_info, &cpukey,
 454                                                 buf->start, buf->len,
 455                                                 btrfs_header_level(buf));
 456         }
 457         return ret;
 458 }
 459
 460 enum btrfs_tree_block_status
 461 btrfs_check_leaf(struct btrfs_root *root, struct btrfs_disk_key *parent_key,
 462                  struct extent_buffer *buf)
 463 {
 464         int i;
 465         struct btrfs_key cpukey;
 466         struct btrfs_disk_key key;
 467         u32 nritems = btrfs_header_nritems(buf);
 468         enum btrfs_tree_block_status ret = BTRFS_TREE_BLOCK_INVALID_NRITEMS;
 469
 470         if (nritems * sizeof(struct btrfs_item) > buf->len)  {
 471                 fprintf(stderr, "invalid number of items %llu\n",
 472                         (unsigned long long)buf->start);
 473                 goto fail;
 474         }
 475
 476         if (btrfs_header_level(buf) != 0) {
 477                 ret = BTRFS_TREE_BLOCK_INVALID_LEVEL;
 478                 fprintf(stderr, "leaf is not a leaf %llu\n",
 479                        (unsigned long long)btrfs_header_bytenr(buf));
 480                 goto fail;
 481         }
 482         if (btrfs_leaf_free_space(root, buf) < 0) {
 483                 ret = BTRFS_TREE_BLOCK_INVALID_FREE_SPACE;
 484                 fprintf(stderr, "leaf free space incorrect %llu %d\n",
 485                         (unsigned long long)btrfs_header_bytenr(buf),
 486                         btrfs_leaf_free_space(root, buf));
 487                 goto fail;
 488         }
 489
 490         if (nritems == 0)
 491                 return BTRFS_TREE_BLOCK_CLEAN;
 492
 493         btrfs_item_key(buf, &key, 0);
 494         if (parent_key && parent_key->type &&
 495             memcmp(parent_key, &key, sizeof(key))) {
 496                 ret = BTRFS_TREE_BLOCK_INVALID_PARENT_KEY;
 497                 fprintf(stderr, "leaf parent key incorrect %llu\n",
 498                        (unsigned long long)btrfs_header_bytenr(buf));
 499                 goto fail;
 500         }
 501         for (i = 0; nritems > 1 && i < nritems - 1; i++) {
 502                 btrfs_item_key(buf, &key, i);
 503                 btrfs_item_key_to_cpu(buf, &cpukey, i + 1);
 504                 if (btrfs_comp_keys(&key, &cpukey) >= 0) {
 505                         ret = BTRFS_TREE_BLOCK_BAD_KEY_ORDER;
 506                         fprintf(stderr, "bad key ordering %d %d\n", i, i+1);
 507                         goto fail;
 508                 }
 509                 if (btrfs_item_offset_nr(buf, i) !=
 510                         btrfs_item_end_nr(buf, i + 1)) {
 511                         ret = BTRFS_TREE_BLOCK_INVALID_OFFSETS;
 512                         fprintf(stderr, "incorrect offsets %u %u\n",
 513                                 btrfs_item_offset_nr(buf, i),
 514                                 btrfs_item_end_nr(buf, i + 1));
 515                         goto fail;
 516                 }
 517                 if (i == 0 && btrfs_item_end_nr(buf, i) !=
 518                     BTRFS_LEAF_DATA_SIZE(root)) {
 519                         ret = BTRFS_TREE_BLOCK_INVALID_OFFSETS;
 520                         fprintf(stderr, "bad item end %u wanted %u\n",
 521                                 btrfs_item_end_nr(buf, i),
 522                                 (unsigned)BTRFS_LEAF_DATA_SIZE(root));
 523                         goto fail;
 524                 }
 525         }
 526
 527         for (i = 0; i < nritems; i++) {
 528                 if (btrfs_item_end_nr(buf, i) > BTRFS_LEAF_DATA_SIZE(root)) {
 529                         btrfs_item_key(buf, &key, 0);
 530                         btrfs_print_key(&key);
 531                         fflush(stdout);
 532                         ret = BTRFS_TREE_BLOCK_INVALID_OFFSETS;
 533                         fprintf(stderr, "slot end outside of leaf %llu > %llu\n",
 534                                 (unsigned long long)btrfs_item_end_nr(buf, i),
 535                                 (unsigned long long)BTRFS_LEAF_DATA_SIZE(root));
 536                         goto fail;
 537                 }
 538         }
 539
 540         return BTRFS_TREE_BLOCK_CLEAN;
 541 fail:
 542         if (btrfs_header_owner(buf) == BTRFS_EXTENT_TREE_OBJECTID) {
 543                 if (parent_key)
 544                         btrfs_disk_key_to_cpu(&cpukey, parent_key);
 545                 else
 546                         btrfs_item_key_to_cpu(buf, &cpukey, 0);
 547
 548                 btrfs_add_corrupt_extent_record(root->fs_info, &cpukey,
 549                                                 buf->start, buf->len, 0);
 550         }
 551         return ret;
 552 }
 553
 554 static int noinline check_block(struct btrfs_root *root,
 555                                 struct btrfs_path *path, int level)
 556 {
 557         struct btrfs_disk_key key;
 558         struct btrfs_disk_key *key_ptr = NULL;
 559         struct extent_buffer *parent;
 560         enum btrfs_tree_block_status ret;
 561
 562         if (path->skip_check_block)
 563                 return 0;
 564         if (path->nodes[level + 1]) {
 565                 parent = path->nodes[level + 1];
 566                 btrfs_node_key(parent, &key, path->slots[level + 1]);
 567                 key_ptr = &key;
 568         }
 569         if (level == 0)
 570                 ret =  btrfs_check_leaf(root, key_ptr, path->nodes[0]);
 571         else
 572                 ret = btrfs_check_node(root, key_ptr, path->nodes[level]);
 573         if (ret == BTRFS_TREE_BLOCK_CLEAN)
 574                 return 0;
 575         return -EIO;
 576 }
 577
 578 /*
 579  * search for key in the extent_buffer.  The items start at offset p,
 580  * and they are item_size apart.  There are 'max' items in p.
 581  *
 582  * the slot in the array is returned via slot, and it points to
 583  * the place where you would insert key if it is not found in
 584  * the array.
 585  *
 586  * slot may point to max if the key is bigger than all of the keys
 587  */
 588 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 589                               int item_size, struct btrfs_key *key,
 590                               int max, int *slot)
 591 {
 592         int low = 0;
 593         int high = max;
 594         int mid;
 595         int ret;
 596         unsigned long offset;
 597         struct btrfs_disk_key *tmp;
 598
 599         while(low < high) {
 600                 mid = (low + high) / 2;
 601                 offset = p + mid * item_size;
 602
 603                 tmp = (struct btrfs_disk_key *)(eb->data + offset);
 604                 ret = btrfs_comp_keys(tmp, key);
 605
 606                 if (ret < 0)
 607                         low = mid + 1;
 608                 else if (ret > 0)
 609                         high = mid;
 610                 else {
 611                         *slot = mid;
 612                         return 0;
 613                 }
 614         }
 615         *slot = low;
 616         return 1;
 617 }
 618
 619 /*
 620  * simple bin_search frontend that does the right thing for
 621  * leaves vs nodes
 622  */
 623 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 624                       int level, int *slot)
 625 {
 626         if (level == 0)
 627                 return generic_bin_search(eb,
 628                                           offsetof(struct btrfs_leaf, items),
 629                                           sizeof(struct btrfs_item),
 630                                           key, btrfs_header_nritems(eb),
 631                                           slot);
 632         else
 633                 return generic_bin_search(eb,
 634                                           offsetof(struct btrfs_node, ptrs),
 635                                           sizeof(struct btrfs_key_ptr),
 636                                           key, btrfs_header_nritems(eb),
 637                                           slot);
 638 }
 639
 640 struct extent_buffer *read_node_slot(struct btrfs_fs_info *fs_info,
 641                                    struct extent_buffer *parent, int slot)
 642 {
 643         int level = btrfs_header_level(parent);
 644         if (slot < 0)
 645                 return NULL;
 646         if (slot >= btrfs_header_nritems(parent))
 647                 return NULL;
 648
 649         if (level == 0)
 650                 return NULL;
 651
 652         return read_tree_block(fs_info, btrfs_node_blockptr(parent, slot),
 653                        fs_info->nodesize,
 654                        btrfs_node_ptr_generation(parent, slot));
 655 }
 656
 657 static int balance_level(struct btrfs_trans_handle *trans,
 658                          struct btrfs_root *root,
 659                          struct btrfs_path *path, int level)
 660 {
 661         struct extent_buffer *right = NULL;
 662         struct extent_buffer *mid;
 663         struct extent_buffer *left = NULL;
 664         struct extent_buffer *parent = NULL;
 665         struct btrfs_fs_info *fs_info = root->fs_info;
 666         int ret = 0;
 667         int wret;
 668         int pslot;
 669         int orig_slot = path->slots[level];
 670         u64 orig_ptr;
 671
 672         if (level == 0)
 673                 return 0;
 674
 675         mid = path->nodes[level];
 676         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 677
 678         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 679
 680         if (level < BTRFS_MAX_LEVEL - 1) {
 681                 parent = path->nodes[level + 1];
 682                 pslot = path->slots[level + 1];
 683         }
 684
 685         /*
 686          * deal with the case where there is only one pointer in the root
 687          * by promoting the node below to a root
 688          */
 689         if (!parent) {
 690                 struct extent_buffer *child;
 691
 692                 if (btrfs_header_nritems(mid) != 1)
 693                         return 0;
 694
 695                 /* promote the child to a root */
 696                 child = read_node_slot(fs_info, mid, 0);
 697                 BUG_ON(!extent_buffer_uptodate(child));
 698                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
 699                 BUG_ON(ret);
 700
 701                 root->node = child;
 702                 add_root_to_dirty_list(root);
 703                 path->nodes[level] = NULL;
 704                 clean_tree_block(trans, root, mid);
 705                 /* once for the path */
 706                 free_extent_buffer(mid);
 707
 708                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 709                                         0, root->root_key.objectid,
 710                                         level, 1);
 711                 /* once for the root ptr */
 712                 free_extent_buffer(mid);
 713                 return ret;
 714         }
 715         if (btrfs_header_nritems(mid) >
 716             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 717                 return 0;
 718
 719         left = read_node_slot(fs_info, parent, pslot - 1);
 720         if (extent_buffer_uptodate(left)) {
 721                 wret = btrfs_cow_block(trans, root, left,
 722                                        parent, pslot - 1, &left);
 723                 if (wret) {
 724                         ret = wret;
 725                         goto enospc;
 726                 }
 727         }
 728         right = read_node_slot(fs_info, parent, pslot + 1);
 729         if (extent_buffer_uptodate(right)) {
 730                 wret = btrfs_cow_block(trans, root, right,
 731                                        parent, pslot + 1, &right);
 732                 if (wret) {
 733                         ret = wret;
 734                         goto enospc;
 735                 }
 736         }
 737
 738         /* first, try to make some room in the middle buffer */
 739         if (left) {
 740                 orig_slot += btrfs_header_nritems(left);
 741                 wret = push_node_left(trans, root, left, mid, 1);
 742                 if (wret < 0)
 743                         ret = wret;
 744         }
 745
 746         /*
 747          * then try to empty the right most buffer into the middle
 748          */
 749         if (right) {
 750                 wret = push_node_left(trans, root, mid, right, 1);
 751                 if (wret < 0 && wret != -ENOSPC)
 752                         ret = wret;
 753                 if (btrfs_header_nritems(right) == 0) {
 754                         u64 bytenr = right->start;
 755                         u32 blocksize = right->len;
 756
 757                         clean_tree_block(trans, root, right);
 758                         free_extent_buffer(right);
 759                         right = NULL;
 760                         wret = btrfs_del_ptr(root, path, level + 1, pslot + 1);
 761                         if (wret)
 762                                 ret = wret;
 763                         wret = btrfs_free_extent(trans, root, bytenr,
 764                                                  blocksize, 0,
 765                                                  root->root_key.objectid,
 766                                                  level, 0);
 767                         if (wret)
 768                                 ret = wret;
 769                 } else {
 770                         struct btrfs_disk_key right_key;
 771                         btrfs_node_key(right, &right_key, 0);
 772                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 773                         btrfs_mark_buffer_dirty(parent);
 774                 }
 775         }
 776         if (btrfs_header_nritems(mid) == 1) {
 777                 /*
 778                  * we're not allowed to leave a node with one item in the
 779                  * tree during a delete.  A deletion from lower in the tree
 780                  * could try to delete the only pointer in this node.
 781                  * So, pull some keys from the left.
 782                  * There has to be a left pointer at this point because
 783                  * otherwise we would have pulled some pointers from the
 784                  * right
 785                  */
 786                 BUG_ON(!left);
 787                 wret = balance_node_right(trans, root, mid, left);
 788                 if (wret < 0) {
 789                         ret = wret;
 790                         goto enospc;
 791                 }
 792                 if (wret == 1) {
 793                         wret = push_node_left(trans, root, left, mid, 1);
 794                         if (wret < 0)
 795                                 ret = wret;
 796                 }
 797                 BUG_ON(wret == 1);
 798         }
 799         if (btrfs_header_nritems(mid) == 0) {
 800                 /* we've managed to empty the middle node, drop it */
 801                 u64 bytenr = mid->start;
 802                 u32 blocksize = mid->len;
 803                 clean_tree_block(trans, root, mid);
 804                 free_extent_buffer(mid);
 805                 mid = NULL;
 806                 wret = btrfs_del_ptr(root, path, level + 1, pslot);
 807                 if (wret)
 808                         ret = wret;
 809                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 810                                          0, root->root_key.objectid,
 811                                          level, 0);
 812                 if (wret)
 813                         ret = wret;
 814         } else {
 815                 /* update the parent key to reflect our changes */
 816                 struct btrfs_disk_key mid_key;
 817                 btrfs_node_key(mid, &mid_key, 0);
 818                 btrfs_set_node_key(parent, &mid_key, pslot);
 819                 btrfs_mark_buffer_dirty(parent);
 820         }
 821
 822         /* update the path */
 823         if (left) {
 824                 if (btrfs_header_nritems(left) > orig_slot) {
 825                         extent_buffer_get(left);
 826                         path->nodes[level] = left;
 827                         path->slots[level + 1] -= 1;
 828                         path->slots[level] = orig_slot;
 829                         if (mid)
 830                                 free_extent_buffer(mid);
 831                 } else {
 832                         orig_slot -= btrfs_header_nritems(left);
 833                         path->slots[level] = orig_slot;
 834                 }
 835         }
 836         /* double check we haven't messed things up */
 837         check_block(root, path, level);
 838         if (orig_ptr !=
 839             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 840                 BUG();
 841 enospc:
 842         if (right)
 843                 free_extent_buffer(right);
 844         if (left)
 845                 free_extent_buffer(left);
 846         return ret;
 847 }
 848
 849 /* returns zero if the push worked, non-zero otherwise */
 850 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
 851                                           struct btrfs_root *root,
 852                                           struct btrfs_path *path, int level)
 853 {
 854         struct extent_buffer *right = NULL;
 855         struct extent_buffer *mid;
 856         struct extent_buffer *left = NULL;
 857         struct extent_buffer *parent = NULL;
 858         struct btrfs_fs_info *fs_info = root->fs_info;
 859         int ret = 0;
 860         int wret;
 861         int pslot;
 862         int orig_slot = path->slots[level];
 863
 864         if (level == 0)
 865                 return 1;
 866
 867         mid = path->nodes[level];
 868         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 869
 870         if (level < BTRFS_MAX_LEVEL - 1) {
 871                 parent = path->nodes[level + 1];
 872                 pslot = path->slots[level + 1];
 873         }
 874
 875         if (!parent)
 876                 return 1;
 877
 878         left = read_node_slot(fs_info, parent, pslot - 1);
 879
 880         /* first, try to make some room in the middle buffer */
 881         if (extent_buffer_uptodate(left)) {
 882                 u32 left_nr;
 883                 left_nr = btrfs_header_nritems(left);
 884                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 885                         wret = 1;
 886                 } else {
 887                         ret = btrfs_cow_block(trans, root, left, parent,
 888                                               pslot - 1, &left);
 889                         if (ret)
 890                                 wret = 1;
 891                         else {
 892                                 wret = push_node_left(trans, root,
 893                                                       left, mid, 0);
 894                         }
 895                 }
 896                 if (wret < 0)
 897                         ret = wret;
 898                 if (wret == 0) {
 899                         struct btrfs_disk_key disk_key;
 900                         orig_slot += left_nr;
 901                         btrfs_node_key(mid, &disk_key, 0);
 902                         btrfs_set_node_key(parent, &disk_key, pslot);
 903                         btrfs_mark_buffer_dirty(parent);
 904                         if (btrfs_header_nritems(left) > orig_slot) {
 905                                 path->nodes[level] = left;
 906                                 path->slots[level + 1] -= 1;
 907                                 path->slots[level] = orig_slot;
 908                                 free_extent_buffer(mid);
 909                         } else {
 910                                 orig_slot -=
 911                                         btrfs_header_nritems(left);
 912                                 path->slots[level] = orig_slot;
 913                                 free_extent_buffer(left);
 914                         }
 915                         return 0;
 916                 }
 917                 free_extent_buffer(left);
 918         }
 919         right= read_node_slot(fs_info, parent, pslot + 1);
 920
 921         /*
 922          * then try to empty the right most buffer into the middle
 923          */
 924         if (extent_buffer_uptodate(right)) {
 925                 u32 right_nr;
 926                 right_nr = btrfs_header_nritems(right);
 927                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 928                         wret = 1;
 929                 } else {
 930                         ret = btrfs_cow_block(trans, root, right,
 931                                               parent, pslot + 1,
 932                                               &right);
 933                         if (ret)
 934                                 wret = 1;
 935                         else {
 936                                 wret = balance_node_right(trans, root,
 937                                                           right, mid);
 938                         }
 939                 }
 940                 if (wret < 0)
 941                         ret = wret;
 942                 if (wret == 0) {
 943                         struct btrfs_disk_key disk_key;
 944
 945                         btrfs_node_key(right, &disk_key, 0);
 946                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
 947                         btrfs_mark_buffer_dirty(parent);
 948
 949                         if (btrfs_header_nritems(mid) <= orig_slot) {
 950                                 path->nodes[level] = right;
 951                                 path->slots[level + 1] += 1;
 952                                 path->slots[level] = orig_slot -
 953                                         btrfs_header_nritems(mid);
 954                                 free_extent_buffer(mid);
 955                         } else {
 956                                 free_extent_buffer(right);
 957                         }
 958                         return 0;
 959                 }
 960                 free_extent_buffer(right);
 961         }
 962         return 1;
 963 }
 964
 965 /*
 966  * readahead one full node of leaves
 967  */
 968 void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
 969                              int level, int slot, u64 objectid)
 970 {
 971         struct btrfs_fs_info *fs_info = root->fs_info;
 972         struct extent_buffer *node;
 973         struct btrfs_disk_key disk_key;
 974         u32 nritems;
 975         u64 search;
 976         u64 lowest_read;
 977         u64 highest_read;
 978         u64 nread = 0;
 979         int direction = path->reada;
 980         struct extent_buffer *eb;
 981         u32 nr;
 982         u32 blocksize;
 983         u32 nscan = 0;
 984
 985         if (level != 1)
 986                 return;
 987
 988         if (!path->nodes[level])
 989                 return;
 990
 991         node = path->nodes[level];
 992         search = btrfs_node_blockptr(node, slot);
 993         blocksize = fs_info->nodesize;
 994         eb = btrfs_find_tree_block(fs_info, search, blocksize);
 995         if (eb) {
 996                 free_extent_buffer(eb);
 997                 return;
 998         }
 999
1000         highest_read = search;
1001         lowest_read = search;
1002
1003         nritems = btrfs_header_nritems(node);
1004         nr = slot;
1005         while(1) {
1006                 if (direction < 0) {
1007                         if (nr == 0)
1008                                 break;
1009                         nr--;
1010                 } else if (direction > 0) {
1011                         nr++;
1012                         if (nr >= nritems)
1013                                 break;
1014                 }
1015                 if (path->reada < 0 && objectid) {
1016                         btrfs_node_key(node, &disk_key, nr);
1017                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1018                                 break;
1019                 }
1020                 search = btrfs_node_blockptr(node, nr);
1021                 if ((search >= lowest_read && search <= highest_read) ||
1022                     (search < lowest_read && lowest_read - search <= 32768) ||
1023                     (search > highest_read && search - highest_read <= 32768)) {
1024                         readahead_tree_block(fs_info, search, blocksize,
1025                                      btrfs_node_ptr_generation(node, nr));
1026                         nread += blocksize;
1027                 }
1028                 nscan++;
1029                 if (path->reada < 2 && (nread > SZ_256K || nscan > 32))
1030                         break;
1031                 if(nread > SZ_1M || nscan > 128)
1032                         break;
1033
1034                 if (search < lowest_read)
1035                         lowest_read = search;
1036                 if (search > highest_read)
1037                         highest_read = search;
1038         }
1039 }
1040
1041 int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *found_path,
1042                 u64 iobjectid, u64 ioff, u8 key_type,
1043                 struct btrfs_key *found_key)
1044 {
1045         int ret;
1046         struct btrfs_key key;
1047         struct extent_buffer *eb;
1048         struct btrfs_path *path;
1049
1050         key.type = key_type;
1051         key.objectid = iobjectid;
1052         key.offset = ioff;
1053
1054         if (found_path == NULL) {
1055                 path = btrfs_alloc_path();
1056                 if (!path)
1057                         return -ENOMEM;
1058         } else
1059                 path = found_path;
1060
1061         ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
1062         if ((ret < 0) || (found_key == NULL))
1063                 goto out;
1064
1065         eb = path->nodes[0];
1066         if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
1067                 ret = btrfs_next_leaf(fs_root, path);
1068                 if (ret)
1069                         goto out;
1070                 eb = path->nodes[0];
1071         }
1072
1073         btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
1074         if (found_key->type != key.type ||
1075                         found_key->objectid != key.objectid) {
1076                 ret = 1;
1077                 goto out;
1078         }
1079
1080 out:
1081         if (path != found_path)
1082                 btrfs_free_path(path);
1083         return ret;
1084 }
1085
1086 /*
1087  * look for key in the tree.  path is filled in with nodes along the way
1088  * if key is found, we return zero and you can find the item in the leaf
1089  * level of the path (level 0)
1090  *
1091  * If the key isn't found, the path points to the slot where it should
1092  * be inserted, and 1 is returned.  If there are other errors during the
1093  * search a negative error number is returned.
1094  *
1095  * if ins_len > 0, nodes and leaves will be split as we walk down the
1096  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1097  * possible)
1098  */
1099 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1100                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1101                       ins_len, int cow)
1102 {
1103         struct extent_buffer *b;
1104         int slot;
1105         int ret;
1106         int level;
1107         int should_reada = p->reada;
1108         struct btrfs_fs_info *fs_info = root->fs_info;
1109         u8 lowest_level = 0;
1110
1111         lowest_level = p->lowest_level;
1112         WARN_ON(lowest_level && ins_len > 0);
1113         WARN_ON(p->nodes[0] != NULL);
1114         /*
1115         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
1116         */
1117 again:
1118         b = root->node;
1119         extent_buffer_get(b);
1120         while (b) {
1121                 level = btrfs_header_level(b);
1122                 if (cow) {
1123                         int wret;
1124                         wret = btrfs_cow_block(trans, root, b,
1125                                                p->nodes[level + 1],
1126                                                p->slots[level + 1],
1127                                                &b);
1128                         if (wret) {
1129                                 free_extent_buffer(b);
1130                                 return wret;
1131                         }
1132                 }
1133                 BUG_ON(!cow && ins_len);
1134                 if (level != btrfs_header_level(b))
1135                         WARN_ON(1);
1136                 level = btrfs_header_level(b);
1137                 p->nodes[level] = b;
1138                 ret = check_block(root, p, level);
1139                 if (ret)
1140                         return -1;
1141                 ret = bin_search(b, key, level, &slot);
1142                 if (level != 0) {
1143                         if (ret && slot > 0)
1144                                 slot -= 1;
1145                         p->slots[level] = slot;
1146                         if ((p->search_for_split || ins_len > 0) &&
1147                             btrfs_header_nritems(b) >=
1148                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1149                                 int sret = split_node(trans, root, p, level);
1150                                 BUG_ON(sret > 0);
1151                                 if (sret)
1152                                         return sret;
1153                                 b = p->nodes[level];
1154                                 slot = p->slots[level];
1155                         } else if (ins_len < 0) {
1156                                 int sret = balance_level(trans, root, p,
1157                                                          level);
1158                                 if (sret)
1159                                         return sret;
1160                                 b = p->nodes[level];
1161                                 if (!b) {
1162                                         btrfs_release_path(p);
1163                                         goto again;
1164                                 }
1165                                 slot = p->slots[level];
1166                                 BUG_ON(btrfs_header_nritems(b) == 1);
1167                         }
1168                         /* this is only true while dropping a snapshot */
1169                         if (level == lowest_level)
1170                                 break;
1171
1172                         if (should_reada)
1173                                 reada_for_search(root, p, level, slot,
1174                                                  key->objectid);
1175
1176                         b = read_node_slot(fs_info, b, slot);
1177                         if (!extent_buffer_uptodate(b))
1178                                 return -EIO;
1179                 } else {
1180                         p->slots[level] = slot;
1181                         if (ins_len > 0 &&
1182                             ins_len > btrfs_leaf_free_space(root, b)) {
1183                                 int sret = split_leaf(trans, root, key,
1184                                                       p, ins_len, ret == 0);
1185                                 BUG_ON(sret > 0);
1186                                 if (sret)
1187                                         return sret;
1188                         }
1189                         return ret;
1190                 }
1191         }
1192         return 1;
1193 }
1194
1195 /*
1196  * adjust the pointers going up the tree, starting at level
1197  * making sure the right key of each node is points to 'key'.
1198  * This is used after shifting pointers to the left, so it stops
1199  * fixing up pointers when a given leaf/node is not in slot 0 of the
1200  * higher levels
1201  */
1202 void btrfs_fixup_low_keys(struct btrfs_root *root, struct btrfs_path *path,
1203                           struct btrfs_disk_key *key, int level)
1204 {
1205         int i;
1206         struct extent_buffer *t;
1207
1208         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1209                 int tslot = path->slots[i];
1210                 if (!path->nodes[i])
1211                         break;
1212                 t = path->nodes[i];
1213                 btrfs_set_node_key(t, key, tslot);
1214                 btrfs_mark_buffer_dirty(path->nodes[i]);
1215                 if (tslot != 0)
1216                         break;
1217         }
1218 }
1219
1220 /*
1221  * update item key.
1222  *
1223  * This function isn't completely safe. It's the caller's responsibility
1224  * that the new key won't break the order
1225  */
1226 int btrfs_set_item_key_safe(struct btrfs_root *root, struct btrfs_path *path,
1227                             struct btrfs_key *new_key)
1228 {
1229         struct btrfs_disk_key disk_key;
1230         struct extent_buffer *eb;
1231         int slot;
1232
1233         eb = path->nodes[0];
1234         slot = path->slots[0];
1235         if (slot > 0) {
1236                 btrfs_item_key(eb, &disk_key, slot - 1);
1237                 if (btrfs_comp_keys(&disk_key, new_key) >= 0)
1238                         return -1;
1239         }
1240         if (slot < btrfs_header_nritems(eb) - 1) {
1241                 btrfs_item_key(eb, &disk_key, slot + 1);
1242                 if (btrfs_comp_keys(&disk_key, new_key) <= 0)
1243                         return -1;
1244         }
1245
1246         btrfs_cpu_key_to_disk(&disk_key, new_key);
1247         btrfs_set_item_key(eb, &disk_key, slot);
1248         btrfs_mark_buffer_dirty(eb);
1249         if (slot == 0)
1250                 btrfs_fixup_low_keys(root, path, &disk_key, 1);
1251         return 0;
1252 }
1253
1254 /*
1255  * update an item key without the safety checks.  This is meant to be called by
1256  * fsck only.
1257  */
1258 void btrfs_set_item_key_unsafe(struct btrfs_root *root,
1259                                struct btrfs_path *path,
1260                                struct btrfs_key *new_key)
1261 {
1262         struct btrfs_disk_key disk_key;
1263         struct extent_buffer *eb;
1264         int slot;
1265
1266         eb = path->nodes[0];
1267         slot = path->slots[0];
1268
1269         btrfs_cpu_key_to_disk(&disk_key, new_key);
1270         btrfs_set_item_key(eb, &disk_key, slot);
1271         btrfs_mark_buffer_dirty(eb);
1272         if (slot == 0)
1273                 btrfs_fixup_low_keys(root, path, &disk_key, 1);
1274 }
1275
1276 /*
1277  * try to push data from one node into the next node left in the
1278  * tree.
1279  *
1280  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1281  * error, and > 0 if there was no room in the left hand block.
1282  */
1283 static int push_node_left(struct btrfs_trans_handle *trans,
1284                           struct btrfs_root *root, struct extent_buffer *dst,
1285                           struct extent_buffer *src, int empty)
1286 {
1287         int push_items = 0;
1288         int src_nritems;
1289         int dst_nritems;
1290         int ret = 0;
1291
1292         src_nritems = btrfs_header_nritems(src);
1293         dst_nritems = btrfs_header_nritems(dst);
1294         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1295         WARN_ON(btrfs_header_generation(src) != trans->transid);
1296         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1297
1298         if (!empty && src_nritems <= 8)
1299                 return 1;
1300
1301         if (push_items <= 0) {
1302                 return 1;
1303         }
1304
1305         if (empty) {
1306                 push_items = min(src_nritems, push_items);
1307                 if (push_items < src_nritems) {
1308                         /* leave at least 8 pointers in the node if
1309                          * we aren't going to empty it
1310                          */
1311                         if (src_nritems - push_items < 8) {
1312                                 if (push_items <= 8)
1313                                         return 1;
1314                                 push_items -= 8;
1315                         }
1316                 }
1317         } else
1318                 push_items = min(src_nritems - 8, push_items);
1319
1320         copy_extent_buffer(dst, src,
1321                            btrfs_node_key_ptr_offset(dst_nritems),
1322                            btrfs_node_key_ptr_offset(0),
1323                            push_items * sizeof(struct btrfs_key_ptr));
1324
1325         if (push_items < src_nritems) {
1326                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1327                                       btrfs_node_key_ptr_offset(push_items),
1328                                       (src_nritems - push_items) *
1329                                       sizeof(struct btrfs_key_ptr));
1330         }
1331         btrfs_set_header_nritems(src, src_nritems - push_items);
1332         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1333         btrfs_mark_buffer_dirty(src);
1334         btrfs_mark_buffer_dirty(dst);
1335
1336         return ret;
1337 }
1338
1339 /*
1340  * try to push data from one node into the next node right in the
1341  * tree.
1342  *
1343  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1344  * error, and > 0 if there was no room in the right hand block.
1345  *
1346  * this will  only push up to 1/2 the contents of the left node over
1347  */
1348 static int balance_node_right(struct btrfs_trans_handle *trans,
1349                               struct btrfs_root *root,
1350                               struct extent_buffer *dst,
1351                               struct extent_buffer *src)
1352 {
1353         int push_items = 0;
1354         int max_push;
1355         int src_nritems;
1356         int dst_nritems;
1357         int ret = 0;
1358
1359         WARN_ON(btrfs_header_generation(src) != trans->transid);
1360         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1361
1362         src_nritems = btrfs_header_nritems(src);
1363         dst_nritems = btrfs_header_nritems(dst);
1364         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1365         if (push_items <= 0) {
1366                 return 1;
1367         }
1368
1369         if (src_nritems < 4) {
1370                 return 1;
1371         }
1372
1373         max_push = src_nritems / 2 + 1;
1374         /* don't try to empty the node */
1375         if (max_push >= src_nritems) {
1376                 return 1;
1377         }
1378
1379         if (max_push < push_items)
1380                 push_items = max_push;
1381
1382         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1383                                       btrfs_node_key_ptr_offset(0),
1384                                       (dst_nritems) *
1385                                       sizeof(struct btrfs_key_ptr));
1386
1387         copy_extent_buffer(dst, src,
1388                            btrfs_node_key_ptr_offset(0),
1389                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1390                            push_items * sizeof(struct btrfs_key_ptr));
1391
1392         btrfs_set_header_nritems(src, src_nritems - push_items);
1393         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1394
1395         btrfs_mark_buffer_dirty(src);
1396         btrfs_mark_buffer_dirty(dst);
1397
1398         return ret;
1399 }
1400
1401 /*
1402  * helper function to insert a new root level in the tree.
1403  * A new node is allocated, and a single item is inserted to
1404  * point to the existing root
1405  *
1406  * returns zero on success or < 0 on failure.
1407  */
1408 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1409                            struct btrfs_root *root,
1410                            struct btrfs_path *path, int level)
1411 {
1412         u64 lower_gen;
1413         struct extent_buffer *lower;
1414         struct extent_buffer *c;
1415         struct extent_buffer *old;
1416         struct btrfs_disk_key lower_key;
1417
1418         BUG_ON(path->nodes[level]);
1419         BUG_ON(path->nodes[level-1] != root->node);
1420
1421         lower = path->nodes[level-1];
1422         if (level == 1)
1423                 btrfs_item_key(lower, &lower_key, 0);
1424         else
1425                 btrfs_node_key(lower, &lower_key, 0);
1426
1427         c = btrfs_alloc_free_block(trans, root, root->fs_info->nodesize,
1428                                    root->root_key.objectid, &lower_key,
1429                                    level, root->node->start, 0);
1430
1431         if (IS_ERR(c))
1432                 return PTR_ERR(c);
1433
1434         memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
1435         btrfs_set_header_nritems(c, 1);
1436         btrfs_set_header_level(c, level);
1437         btrfs_set_header_bytenr(c, c->start);
1438         btrfs_set_header_generation(c, trans->transid);
1439         btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
1440         btrfs_set_header_owner(c, root->root_key.objectid);
1441
1442         write_extent_buffer(c, root->fs_info->fsid,
1443                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
1444
1445         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1446                             btrfs_header_chunk_tree_uuid(c),
1447                             BTRFS_UUID_SIZE);
1448
1449         btrfs_set_node_key(c, &lower_key, 0);
1450         btrfs_set_node_blockptr(c, 0, lower->start);
1451         lower_gen = btrfs_header_generation(lower);
1452         WARN_ON(lower_gen != trans->transid);
1453
1454         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1455
1456         btrfs_mark_buffer_dirty(c);
1457
1458         old = root->node;
1459         root->node = c;
1460
1461         /* the super has an extra ref to root->node */
1462         free_extent_buffer(old);
1463
1464         add_root_to_dirty_list(root);
1465         extent_buffer_get(c);
1466         path->nodes[level] = c;
1467         path->slots[level] = 0;
1468         return 0;
1469 }
1470
1471 /*
1472  * worker function to insert a single pointer in a node.
1473  * the node should have enough room for the pointer already
1474  *
1475  * slot and level indicate where you want the key to go, and
1476  * blocknr is the block the key points to.
1477  *
1478  * returns zero on success and < 0 on any error
1479  */
1480 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1481                       *root, struct btrfs_path *path, struct btrfs_disk_key
1482                       *key, u64 bytenr, int slot, int level)
1483 {
1484         struct extent_buffer *lower;
1485         int nritems;
1486
1487         BUG_ON(!path->nodes[level]);
1488         lower = path->nodes[level];
1489         nritems = btrfs_header_nritems(lower);
1490         if (slot > nritems)
1491                 BUG();
1492         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1493                 BUG();
1494         if (slot < nritems) {
1495                 /* shift the items */
1496                 memmove_extent_buffer(lower,
1497                               btrfs_node_key_ptr_offset(slot + 1),
1498                               btrfs_node_key_ptr_offset(slot),
1499                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1500         }
1501         btrfs_set_node_key(lower, key, slot);
1502         btrfs_set_node_blockptr(lower, slot, bytenr);
1503         WARN_ON(trans->transid == 0);
1504         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1505         btrfs_set_header_nritems(lower, nritems + 1);
1506         btrfs_mark_buffer_dirty(lower);
1507         return 0;
1508 }
1509
1510 /*
1511  * split the node at the specified level in path in two.
1512  * The path is corrected to point to the appropriate node after the split
1513  *
1514  * Before splitting this tries to make some room in the node by pushing
1515  * left and right, if either one works, it returns right away.
1516  *
1517  * returns 0 on success and < 0 on failure
1518  */
1519 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1520                       *root, struct btrfs_path *path, int level)
1521 {
1522         struct extent_buffer *c;
1523         struct extent_buffer *split;
1524         struct btrfs_disk_key disk_key;
1525         int mid;
1526         int ret;
1527         int wret;
1528         u32 c_nritems;
1529
1530         c = path->nodes[level];
1531         WARN_ON(btrfs_header_generation(c) != trans->transid);
1532         if (c == root->node) {
1533                 /* trying to split the root, lets make a new one */
1534                 ret = insert_new_root(trans, root, path, level + 1);
1535                 if (ret)
1536                         return ret;
1537         } else {
1538                 ret = push_nodes_for_insert(trans, root, path, level);
1539                 c = path->nodes[level];
1540                 if (!ret && btrfs_header_nritems(c) <
1541                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
1542                         return 0;
1543                 if (ret < 0)
1544                         return ret;
1545         }
1546
1547         c_nritems = btrfs_header_nritems(c);
1548         mid = (c_nritems + 1) / 2;
1549         btrfs_node_key(c, &disk_key, mid);
1550
1551         split = btrfs_alloc_free_block(trans, root, root->fs_info->nodesize,
1552                                         root->root_key.objectid,
1553                                         &disk_key, level, c->start, 0);
1554         if (IS_ERR(split))
1555                 return PTR_ERR(split);
1556
1557         memset_extent_buffer(split, 0, 0, sizeof(struct btrfs_header));
1558         btrfs_set_header_level(split, btrfs_header_level(c));
1559         btrfs_set_header_bytenr(split, split->start);
1560         btrfs_set_header_generation(split, trans->transid);
1561         btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
1562         btrfs_set_header_owner(split, root->root_key.objectid);
1563         write_extent_buffer(split, root->fs_info->fsid,
1564                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
1565         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1566                             btrfs_header_chunk_tree_uuid(split),
1567                             BTRFS_UUID_SIZE);
1568
1569
1570         copy_extent_buffer(split, c,
1571                            btrfs_node_key_ptr_offset(0),
1572                            btrfs_node_key_ptr_offset(mid),
1573                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1574         btrfs_set_header_nritems(split, c_nritems - mid);
1575         btrfs_set_header_nritems(c, mid);
1576         ret = 0;
1577
1578         btrfs_mark_buffer_dirty(c);
1579         btrfs_mark_buffer_dirty(split);
1580
1581         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1582                           path->slots[level + 1] + 1,
1583                           level + 1);
1584         if (wret)
1585                 ret = wret;
1586
1587         if (path->slots[level] >= mid) {
1588                 path->slots[level] -= mid;
1589                 free_extent_buffer(c);
1590                 path->nodes[level] = split;
1591                 path->slots[level + 1] += 1;
1592         } else {
1593                 free_extent_buffer(split);
1594         }
1595         return ret;
1596 }
1597
1598 /*
1599  * how many bytes are required to store the items in a leaf.  start
1600  * and nr indicate which items in the leaf to check.  This totals up the
1601  * space used both by the item structs and the item data
1602  */
1603 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1604 {
1605         int data_len;
1606         int nritems = btrfs_header_nritems(l);
1607         int end = min(nritems, start + nr) - 1;
1608
1609         if (!nr)
1610                 return 0;
1611         data_len = btrfs_item_end_nr(l, start);
1612         data_len = data_len - btrfs_item_offset_nr(l, end);
1613         data_len += sizeof(struct btrfs_item) * nr;
1614         WARN_ON(data_len < 0);
1615         return data_len;
1616 }
1617
1618 /*
1619  * The space between the end of the leaf items and
1620  * the start of the leaf data.  IOW, how much room
1621  * the leaf has left for both items and data
1622  */
1623 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1624 {
1625         u32 nodesize = (root ? BTRFS_LEAF_DATA_SIZE(root) : leaf->len);
1626         int nritems = btrfs_header_nritems(leaf);
1627         int ret;
1628         ret = nodesize - leaf_space_used(leaf, 0, nritems);
1629         if (ret < 0) {
1630                 printk("leaf free space ret %d, leaf data size %u, used %d nritems %d\n",
1631                        ret, nodesize, leaf_space_used(leaf, 0, nritems),
1632                        nritems);
1633         }
1634         return ret;
1635 }
1636
1637 /*
1638  * push some data in the path leaf to the right, trying to free up at
1639  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1640  *
1641  * returns 1 if the push failed because the other node didn't have enough
1642  * room, 0 if everything worked out and < 0 if there were major errors.
1643  */
1644 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1645                            *root, struct btrfs_path *path, int data_size,
1646                            int empty)
1647 {
1648         struct extent_buffer *left = path->nodes[0];
1649         struct extent_buffer *right;
1650         struct extent_buffer *upper;
1651         struct btrfs_disk_key disk_key;
1652         struct btrfs_fs_info *fs_info = root->fs_info;
1653         int slot;
1654         u32 i;
1655         int free_space;
1656         int push_space = 0;
1657         int push_items = 0;
1658         struct btrfs_item *item;
1659         u32 left_nritems;
1660         u32 nr;
1661         u32 right_nritems;
1662         u32 data_end;
1663         u32 this_item_size;
1664         int ret;
1665
1666         slot = path->slots[1];
1667         if (!path->nodes[1]) {
1668                 return 1;
1669         }
1670         upper = path->nodes[1];
1671         if (slot >= btrfs_header_nritems(upper) - 1)
1672                 return 1;
1673
1674         right = read_node_slot(fs_info, upper, slot + 1);
1675         if (!extent_buffer_uptodate(right)) {
1676                 if (IS_ERR(right))
1677                         return PTR_ERR(right);
1678                 return -EIO;
1679         }
1680         free_space = btrfs_leaf_free_space(root, right);
1681         if (free_space < data_size) {
1682                 free_extent_buffer(right);
1683                 return 1;
1684         }
1685
1686         /* cow and double check */
1687         ret = btrfs_cow_block(trans, root, right, upper,
1688                               slot + 1, &right);
1689         if (ret) {
1690                 free_extent_buffer(right);
1691                 return 1;
1692         }
1693         free_space = btrfs_leaf_free_space(root, right);
1694         if (free_space < data_size) {
1695                 free_extent_buffer(right);
1696                 return 1;
1697         }
1698
1699         left_nritems = btrfs_header_nritems(left);
1700         if (left_nritems == 0) {
1701                 free_extent_buffer(right);
1702                 return 1;
1703         }
1704
1705         if (empty)
1706                 nr = 0;
1707         else
1708                 nr = 1;
1709
1710         i = left_nritems - 1;
1711         while (i >= nr) {
1712                 item = btrfs_item_nr(i);
1713
1714                 if (path->slots[0] == i)
1715                         push_space += data_size + sizeof(*item);
1716
1717                 this_item_size = btrfs_item_size(left, item);
1718                 if (this_item_size + sizeof(*item) + push_space > free_space)
1719                         break;
1720                 push_items++;
1721                 push_space += this_item_size + sizeof(*item);
1722                 if (i == 0)
1723                         break;
1724                 i--;
1725         }
1726
1727         if (push_items == 0) {
1728                 free_extent_buffer(right);
1729                 return 1;
1730         }
1731
1732         if (!empty && push_items == left_nritems)
1733                 WARN_ON(1);
1734
1735         /* push left to right */
1736         right_nritems = btrfs_header_nritems(right);
1737
1738         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1739         push_space -= leaf_data_end(root, left);
1740
1741         /* make room in the right data area */
1742         data_end = leaf_data_end(root, right);
1743         memmove_extent_buffer(right,
1744                               btrfs_leaf_data(right) + data_end - push_space,
1745                               btrfs_leaf_data(right) + data_end,
1746                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1747
1748         /* copy from the left data area */
1749         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1750                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1751                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1752                      push_space);
1753
1754         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1755                               btrfs_item_nr_offset(0),
1756                               right_nritems * sizeof(struct btrfs_item));
1757
1758         /* copy the items from left to right */
1759         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1760                    btrfs_item_nr_offset(left_nritems - push_items),
1761                    push_items * sizeof(struct btrfs_item));
1762
1763         /* update the item pointers */
1764         right_nritems += push_items;
1765         btrfs_set_header_nritems(right, right_nritems);
1766         push_space = BTRFS_LEAF_DATA_SIZE(root);
1767         for (i = 0; i < right_nritems; i++) {
1768                 item = btrfs_item_nr(i);
1769                 push_space -= btrfs_item_size(right, item);
1770                 btrfs_set_item_offset(right, item, push_space);
1771         }
1772
1773         left_nritems -= push_items;
1774         btrfs_set_header_nritems(left, left_nritems);
1775
1776         if (left_nritems)
1777                 btrfs_mark_buffer_dirty(left);
1778         btrfs_mark_buffer_dirty(right);
1779
1780         btrfs_item_key(right, &disk_key, 0);
1781         btrfs_set_node_key(upper, &disk_key, slot + 1);
1782         btrfs_mark_buffer_dirty(upper);
1783
1784         /* then fixup the leaf pointer in the path */
1785         if (path->slots[0] >= left_nritems) {
1786                 path->slots[0] -= left_nritems;
1787                 free_extent_buffer(path->nodes[0]);
1788                 path->nodes[0] = right;
1789                 path->slots[1] += 1;
1790         } else {
1791                 free_extent_buffer(right);
1792         }
1793         return 0;
1794 }
1795 /*
1796  * push some data in the path leaf to the left, trying to free up at
1797  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1798  */
1799 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1800                           *root, struct btrfs_path *path, int data_size,
1801                           int empty)
1802 {
1803         struct btrfs_disk_key disk_key;
1804         struct extent_buffer *right = path->nodes[0];
1805         struct extent_buffer *left;
1806         struct btrfs_fs_info *fs_info = root->fs_info;
1807         int slot;
1808         int i;
1809         int free_space;
1810         int push_space = 0;
1811         int push_items = 0;
1812         struct btrfs_item *item;
1813         u32 old_left_nritems;
1814         u32 right_nritems;
1815         u32 nr;
1816         int ret = 0;
1817         u32 this_item_size;
1818         u32 old_left_item_size;
1819
1820         slot = path->slots[1];
1821         if (slot == 0)
1822                 return 1;
1823         if (!path->nodes[1])
1824                 return 1;
1825
1826         right_nritems = btrfs_header_nritems(right);
1827         if (right_nritems == 0) {
1828                 return 1;
1829         }
1830
1831         left = read_node_slot(fs_info, path->nodes[1], slot - 1);
1832         free_space = btrfs_leaf_free_space(root, left);
1833         if (free_space < data_size) {
1834                 free_extent_buffer(left);
1835                 return 1;
1836         }
1837
1838         /* cow and double check */
1839         ret = btrfs_cow_block(trans, root, left,
1840                               path->nodes[1], slot - 1, &left);
1841         if (ret) {
1842                 /* we hit -ENOSPC, but it isn't fatal here */
1843                 free_extent_buffer(left);
1844                 return 1;
1845         }
1846
1847         free_space = btrfs_leaf_free_space(root, left);
1848         if (free_space < data_size) {
1849                 free_extent_buffer(left);
1850                 return 1;
1851         }
1852
1853         if (empty)
1854                 nr = right_nritems;
1855         else
1856                 nr = right_nritems - 1;
1857
1858         for (i = 0; i < nr; i++) {
1859                 item = btrfs_item_nr(i);
1860
1861                 if (path->slots[0] == i)
1862                         push_space += data_size + sizeof(*item);
1863
1864                 this_item_size = btrfs_item_size(right, item);
1865                 if (this_item_size + sizeof(*item) + push_space > free_space)
1866                         break;
1867
1868                 push_items++;
1869                 push_space += this_item_size + sizeof(*item);
1870         }
1871
1872         if (push_items == 0) {
1873                 free_extent_buffer(left);
1874                 return 1;
1875         }
1876         if (!empty && push_items == btrfs_header_nritems(right))
1877                 WARN_ON(1);
1878
1879         /* push data from right to left */
1880         copy_extent_buffer(left, right,
1881                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1882                            btrfs_item_nr_offset(0),
1883                            push_items * sizeof(struct btrfs_item));
1884
1885         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1886                      btrfs_item_offset_nr(right, push_items -1);
1887
1888         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1889                      leaf_data_end(root, left) - push_space,
1890                      btrfs_leaf_data(right) +
1891                      btrfs_item_offset_nr(right, push_items - 1),
1892                      push_space);
1893         old_left_nritems = btrfs_header_nritems(left);
1894         BUG_ON(old_left_nritems == 0);
1895
1896         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1897         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1898                 u32 ioff;
1899
1900                 item = btrfs_item_nr(i);
1901                 ioff = btrfs_item_offset(left, item);
1902                 btrfs_set_item_offset(left, item,
1903                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1904         }
1905         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1906
1907         /* fixup right node */
1908         if (push_items > right_nritems) {
1909                 printk("push items %d nr %u\n", push_items, right_nritems);
1910                 WARN_ON(1);
1911         }
1912
1913         if (push_items < right_nritems) {
1914                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1915                                                   leaf_data_end(root, right);
1916                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
1917                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1918                                       btrfs_leaf_data(right) +
1919                                       leaf_data_end(root, right), push_space);
1920
1921                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1922                               btrfs_item_nr_offset(push_items),
1923                              (btrfs_header_nritems(right) - push_items) *
1924                              sizeof(struct btrfs_item));
1925         }
1926         right_nritems -= push_items;
1927         btrfs_set_header_nritems(right, right_nritems);
1928         push_space = BTRFS_LEAF_DATA_SIZE(root);
1929         for (i = 0; i < right_nritems; i++) {
1930                 item = btrfs_item_nr(i);
1931                 push_space = push_space - btrfs_item_size(right, item);
1932                 btrfs_set_item_offset(right, item, push_space);
1933         }
1934
1935         btrfs_mark_buffer_dirty(left);
1936         if (right_nritems)
1937                 btrfs_mark_buffer_dirty(right);
1938
1939         btrfs_item_key(right, &disk_key, 0);
1940         btrfs_fixup_low_keys(root, path, &disk_key, 1);
1941
1942         /* then fixup the leaf pointer in the path */
1943         if (path->slots[0] < push_items) {
1944                 path->slots[0] += old_left_nritems;
1945                 free_extent_buffer(path->nodes[0]);
1946                 path->nodes[0] = left;
1947                 path->slots[1] -= 1;
1948         } else {
1949                 free_extent_buffer(left);
1950                 path->slots[0] -= push_items;
1951         }
1952         BUG_ON(path->slots[0] < 0);
1953         return ret;
1954 }
1955
1956 /*
1957  * split the path's leaf in two, making sure there is at least data_size
1958  * available for the resulting leaf level of the path.
1959  *
1960  * returns 0 if all went well and < 0 on failure.
1961  */
1962 static noinline int copy_for_split(struct btrfs_trans_handle *trans,
1963                                struct btrfs_root *root,
1964                                struct btrfs_path *path,
1965                                struct extent_buffer *l,
1966                                struct extent_buffer *right,
1967                                int slot, int mid, int nritems)
1968 {
1969         int data_copy_size;
1970         int rt_data_off;
1971         int i;
1972         int ret = 0;
1973         int wret;
1974         struct btrfs_disk_key disk_key;
1975
1976         nritems = nritems - mid;
1977         btrfs_set_header_nritems(right, nritems);
1978         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
1979
1980         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
1981                            btrfs_item_nr_offset(mid),
1982                            nritems * sizeof(struct btrfs_item));
1983
1984         copy_extent_buffer(right, l,
1985                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
1986                      data_copy_size, btrfs_leaf_data(l) +
1987                      leaf_data_end(root, l), data_copy_size);
1988
1989         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
1990                       btrfs_item_end_nr(l, mid);
1991
1992         for (i = 0; i < nritems; i++) {
1993                 struct btrfs_item *item = btrfs_item_nr(i);
1994                 u32 ioff = btrfs_item_offset(right, item);
1995                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
1996         }
1997
1998         btrfs_set_header_nritems(l, mid);
1999         ret = 0;
2000         btrfs_item_key(right, &disk_key, 0);
2001         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2002                           path->slots[1] + 1, 1);
2003         if (wret)
2004                 ret = wret;
2005
2006         btrfs_mark_buffer_dirty(right);
2007         btrfs_mark_buffer_dirty(l);
2008         BUG_ON(path->slots[0] != slot);
2009
2010         if (mid <= slot) {
2011                 free_extent_buffer(path->nodes[0]);
2012                 path->nodes[0] = right;
2013                 path->slots[0] -= mid;
2014                 path->slots[1] += 1;
2015         } else {
2016                 free_extent_buffer(right);
2017         }
2018
2019         BUG_ON(path->slots[0] < 0);
2020
2021         return ret;
2022 }
2023
2024 /*
2025  * split the path's leaf in two, making sure there is at least data_size
2026  * available for the resulting leaf level of the path.
2027  *
2028  * returns 0 if all went well and < 0 on failure.
2029  */
2030 static noinline int split_leaf(struct btrfs_trans_handle *trans,
2031                                struct btrfs_root *root,
2032                                struct btrfs_key *ins_key,
2033                                struct btrfs_path *path, int data_size,
2034                                int extend)
2035 {
2036         struct btrfs_disk_key disk_key;
2037         struct extent_buffer *l;
2038         u32 nritems;
2039         int mid;
2040         int slot;
2041         struct extent_buffer *right;
2042         int ret = 0;
2043         int wret;
2044         int split;
2045         int num_doubles = 0;
2046
2047         l = path->nodes[0];
2048         slot = path->slots[0];
2049         if (extend && data_size + btrfs_item_size_nr(l, slot) +
2050             sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(root))
2051                 return -EOVERFLOW;
2052
2053         /* first try to make some room by pushing left and right */
2054         if (data_size && ins_key->type != BTRFS_DIR_ITEM_KEY) {
2055                 wret = push_leaf_right(trans, root, path, data_size, 0);
2056                 if (wret < 0)
2057                         return wret;
2058                 if (wret) {
2059                         wret = push_leaf_left(trans, root, path, data_size, 0);
2060                         if (wret < 0)
2061                                 return wret;
2062                 }
2063                 l = path->nodes[0];
2064
2065                 /* did the pushes work? */
2066                 if (btrfs_leaf_free_space(root, l) >= data_size)
2067                         return 0;
2068         }
2069
2070         if (!path->nodes[1]) {
2071                 ret = insert_new_root(trans, root, path, 1);
2072                 if (ret)
2073                         return ret;
2074         }
2075 again:
2076         split = 1;
2077         l = path->nodes[0];
2078         slot = path->slots[0];
2079         nritems = btrfs_header_nritems(l);
2080         mid = (nritems + 1) / 2;
2081
2082         if (mid <= slot) {
2083                 if (nritems == 1 ||
2084                     leaf_space_used(l, mid, nritems - mid) + data_size >
2085                         BTRFS_LEAF_DATA_SIZE(root)) {
2086                         if (slot >= nritems) {
2087                                 split = 0;
2088                         } else {
2089                                 mid = slot;
2090                                 if (mid != nritems &&
2091                                     leaf_space_used(l, mid, nritems - mid) +
2092                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2093                                         split = 2;
2094                                 }
2095                         }
2096                 }
2097         } else {
2098                 if (leaf_space_used(l, 0, mid) + data_size >
2099                         BTRFS_LEAF_DATA_SIZE(root)) {
2100                         if (!extend && data_size && slot == 0) {
2101                                 split = 0;
2102                         } else if ((extend || !data_size) && slot == 0) {
2103                                 mid = 1;
2104                         } else {
2105                                 mid = slot;
2106                                 if (mid != nritems &&
2107                                     leaf_space_used(l, mid, nritems - mid) +
2108                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2109                                         split = 2 ;
2110                                 }
2111                         }
2112                 }
2113         }
2114
2115         if (split == 0)
2116                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2117         else
2118                 btrfs_item_key(l, &disk_key, mid);
2119
2120         right = btrfs_alloc_free_block(trans, root, root->fs_info->nodesize,
2121                                         root->root_key.objectid,
2122                                         &disk_key, 0, l->start, 0);
2123         if (IS_ERR(right)) {
2124                 BUG_ON(1);
2125                 return PTR_ERR(right);
2126         }
2127
2128         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2129         btrfs_set_header_bytenr(right, right->start);
2130         btrfs_set_header_generation(right, trans->transid);
2131         btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
2132         btrfs_set_header_owner(right, root->root_key.objectid);
2133         btrfs_set_header_level(right, 0);
2134         write_extent_buffer(right, root->fs_info->fsid,
2135                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
2136
2137         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2138                             btrfs_header_chunk_tree_uuid(right),
2139                             BTRFS_UUID_SIZE);
2140
2141         if (split == 0) {
2142                 if (mid <= slot) {
2143                         btrfs_set_header_nritems(right, 0);
2144                         wret = insert_ptr(trans, root, path,
2145                                           &disk_key, right->start,
2146                                           path->slots[1] + 1, 1);
2147                         if (wret)
2148                                 ret = wret;
2149
2150                         free_extent_buffer(path->nodes[0]);
2151                         path->nodes[0] = right;
2152                         path->slots[0] = 0;
2153                         path->slots[1] += 1;
2154                 } else {
2155                         btrfs_set_header_nritems(right, 0);
2156                         wret = insert_ptr(trans, root, path,
2157                                           &disk_key,
2158                                           right->start,
2159                                           path->slots[1], 1);
2160                         if (wret)
2161                                 ret = wret;
2162                         free_extent_buffer(path->nodes[0]);
2163                         path->nodes[0] = right;
2164                         path->slots[0] = 0;
2165                         if (path->slots[1] == 0) {
2166                                 btrfs_fixup_low_keys(root, path,
2167                                                      &disk_key, 1);
2168                         }
2169                 }
2170                 btrfs_mark_buffer_dirty(right);
2171                 return ret;
2172         }
2173
2174         ret = copy_for_split(trans, root, path, l, right, slot, mid, nritems);
2175         BUG_ON(ret);
2176
2177         if (split == 2) {
2178                 BUG_ON(num_doubles != 0);
2179                 num_doubles++;
2180                 goto again;
2181         }
2182
2183         return ret;
2184 }
2185
2186 /*
2187  * This function splits a single item into two items,
2188  * giving 'new_key' to the new item and splitting the
2189  * old one at split_offset (from the start of the item).
2190  *
2191  * The path may be released by this operation.  After
2192  * the split, the path is pointing to the old item.  The
2193  * new item is going to be in the same node as the old one.
2194  *
2195  * Note, the item being split must be smaller enough to live alone on
2196  * a tree block with room for one extra struct btrfs_item
2197  *
2198  * This allows us to split the item in place, keeping a lock on the
2199  * leaf the entire time.
2200  */
2201 int btrfs_split_item(struct btrfs_trans_handle *trans,
2202                      struct btrfs_root *root,
2203                      struct btrfs_path *path,
2204                      struct btrfs_key *new_key,
2205                      unsigned long split_offset)
2206 {
2207         u32 item_size;
2208         struct extent_buffer *leaf;
2209         struct btrfs_key orig_key;
2210         struct btrfs_item *item;
2211         struct btrfs_item *new_item;
2212         int ret = 0;
2213         int slot;
2214         u32 nritems;
2215         u32 orig_offset;
2216         struct btrfs_disk_key disk_key;
2217         char *buf;
2218
2219         leaf = path->nodes[0];
2220         btrfs_item_key_to_cpu(leaf, &orig_key, path->slots[0]);
2221         if (btrfs_leaf_free_space(root, leaf) >= sizeof(struct btrfs_item))
2222                 goto split;
2223
2224         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2225         btrfs_release_path(path);
2226
2227         path->search_for_split = 1;
2228
2229         ret = btrfs_search_slot(trans, root, &orig_key, path, 0, 1);
2230         path->search_for_split = 0;
2231
2232         /* if our item isn't there or got smaller, return now */
2233         if (ret != 0 || item_size != btrfs_item_size_nr(path->nodes[0],
2234                                                         path->slots[0])) {
2235                 return -EAGAIN;
2236         }
2237
2238         ret = split_leaf(trans, root, &orig_key, path, 0, 0);
2239         BUG_ON(ret);
2240
2241         BUG_ON(btrfs_leaf_free_space(root, leaf) < sizeof(struct btrfs_item));
2242         leaf = path->nodes[0];
2243
2244 split:
2245         item = btrfs_item_nr(path->slots[0]);
2246         orig_offset = btrfs_item_offset(leaf, item);
2247         item_size = btrfs_item_size(leaf, item);
2248
2249
2250         buf = kmalloc(item_size, GFP_NOFS);
2251         BUG_ON(!buf);
2252         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
2253                             path->slots[0]), item_size);
2254         slot = path->slots[0] + 1;
2255         leaf = path->nodes[0];
2256
2257         nritems = btrfs_header_nritems(leaf);
2258
2259         if (slot < nritems) {
2260                 /* shift the items */
2261                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
2262                               btrfs_item_nr_offset(slot),
2263                               (nritems - slot) * sizeof(struct btrfs_item));
2264
2265         }
2266
2267         btrfs_cpu_key_to_disk(&disk_key, new_key);
2268         btrfs_set_item_key(leaf, &disk_key, slot);
2269
2270         new_item = btrfs_item_nr(slot);
2271
2272         btrfs_set_item_offset(leaf, new_item, orig_offset);
2273         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
2274
2275         btrfs_set_item_offset(leaf, item,
2276                               orig_offset + item_size - split_offset);
2277         btrfs_set_item_size(leaf, item, split_offset);
2278
2279         btrfs_set_header_nritems(leaf, nritems + 1);
2280
2281         /* write the data for the start of the original item */
2282         write_extent_buffer(leaf, buf,
2283                             btrfs_item_ptr_offset(leaf, path->slots[0]),
2284                             split_offset);
2285
2286         /* write the data for the new item */
2287         write_extent_buffer(leaf, buf + split_offset,
2288                             btrfs_item_ptr_offset(leaf, slot),
2289                             item_size - split_offset);
2290         btrfs_mark_buffer_dirty(leaf);
2291
2292         ret = 0;
2293         if (btrfs_leaf_free_space(root, leaf) < 0) {
2294                 btrfs_print_leaf(root, leaf);
2295                 BUG();
2296         }
2297         kfree(buf);
2298         return ret;
2299 }
2300
2301 int btrfs_truncate_item(struct btrfs_root *root, struct btrfs_path *path,
2302                         u32 new_size, int from_end)
2303 {
2304         int ret = 0;
2305         int slot;
2306         struct extent_buffer *leaf;
2307         struct btrfs_item *item;
2308         u32 nritems;
2309         unsigned int data_end;
2310         unsigned int old_data_start;
2311         unsigned int old_size;
2312         unsigned int size_diff;
2313         int i;
2314
2315         leaf = path->nodes[0];
2316         slot = path->slots[0];
2317
2318         old_size = btrfs_item_size_nr(leaf, slot);
2319         if (old_size == new_size)
2320                 return 0;
2321
2322         nritems = btrfs_header_nritems(leaf);
2323         data_end = leaf_data_end(root, leaf);
2324
2325         old_data_start = btrfs_item_offset_nr(leaf, slot);
2326
2327         size_diff = old_size - new_size;
2328
2329         BUG_ON(slot < 0);
2330         BUG_ON(slot >= nritems);
2331
2332         /*
2333          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2334          */
2335         /* first correct the data pointers */
2336         for (i = slot; i < nritems; i++) {
2337                 u32 ioff;
2338                 item = btrfs_item_nr(i);
2339                 ioff = btrfs_item_offset(leaf, item);
2340                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2341         }
2342
2343         /* shift the data */
2344         if (from_end) {
2345                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2346                               data_end + size_diff, btrfs_leaf_data(leaf) +
2347                               data_end, old_data_start + new_size - data_end);
2348         } else {
2349                 struct btrfs_disk_key disk_key;
2350                 u64 offset;
2351
2352                 btrfs_item_key(leaf, &disk_key, slot);
2353
2354                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2355                         unsigned long ptr;
2356                         struct btrfs_file_extent_item *fi;
2357
2358                         fi = btrfs_item_ptr(leaf, slot,
2359                                             struct btrfs_file_extent_item);
2360                         fi = (struct btrfs_file_extent_item *)(
2361                              (unsigned long)fi - size_diff);
2362
2363                         if (btrfs_file_extent_type(leaf, fi) ==
2364                             BTRFS_FILE_EXTENT_INLINE) {
2365                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2366                                 memmove_extent_buffer(leaf, ptr,
2367                                         (unsigned long)fi,
2368                                         offsetof(struct btrfs_file_extent_item,
2369                                                  disk_bytenr));
2370                         }
2371                 }
2372
2373                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2374                               data_end + size_diff, btrfs_leaf_data(leaf) +
2375                               data_end, old_data_start - data_end);
2376
2377                 offset = btrfs_disk_key_offset(&disk_key);
2378                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2379                 btrfs_set_item_key(leaf, &disk_key, slot);
2380                 if (slot == 0)
2381                         btrfs_fixup_low_keys(root, path, &disk_key, 1);
2382         }
2383
2384         item = btrfs_item_nr(slot);
2385         btrfs_set_item_size(leaf, item, new_size);
2386         btrfs_mark_buffer_dirty(leaf);
2387
2388         ret = 0;
2389         if (btrfs_leaf_free_space(root, leaf) < 0) {
2390                 btrfs_print_leaf(root, leaf);
2391                 BUG();
2392         }
2393         return ret;
2394 }
2395
2396 int btrfs_extend_item(struct btrfs_root *root, struct btrfs_path *path,
2397                       u32 data_size)
2398 {
2399         int ret = 0;
2400         int slot;
2401         struct extent_buffer *leaf;
2402         struct btrfs_item *item;
2403         u32 nritems;
2404         unsigned int data_end;
2405         unsigned int old_data;
2406         unsigned int old_size;
2407         int i;
2408
2409         leaf = path->nodes[0];
2410
2411         nritems = btrfs_header_nritems(leaf);
2412         data_end = leaf_data_end(root, leaf);
2413
2414         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2415                 btrfs_print_leaf(root, leaf);
2416                 BUG();
2417         }
2418         slot = path->slots[0];
2419         old_data = btrfs_item_end_nr(leaf, slot);
2420
2421         BUG_ON(slot < 0);
2422         if (slot >= nritems) {
2423                 btrfs_print_leaf(root, leaf);
2424                 printk("slot %d too large, nritems %d\n", slot, nritems);
2425                 BUG_ON(1);
2426         }
2427
2428         /*
2429          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2430          */
2431         /* first correct the data pointers */
2432         for (i = slot; i < nritems; i++) {
2433                 u32 ioff;
2434                 item = btrfs_item_nr(i);
2435                 ioff = btrfs_item_offset(leaf, item);
2436                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2437         }
2438
2439         /* shift the data */
2440         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2441                       data_end - data_size, btrfs_leaf_data(leaf) +
2442                       data_end, old_data - data_end);
2443
2444         data_end = old_data;
2445         old_size = btrfs_item_size_nr(leaf, slot);
2446         item = btrfs_item_nr(slot);
2447         btrfs_set_item_size(leaf, item, old_size + data_size);
2448         btrfs_mark_buffer_dirty(leaf);
2449
2450         ret = 0;
2451         if (btrfs_leaf_free_space(root, leaf) < 0) {
2452                 btrfs_print_leaf(root, leaf);
2453                 BUG();
2454         }
2455         return ret;
2456 }
2457
2458 /*
2459  * Given a key and some data, insert an item into the tree.
2460  * This does all the path init required, making room in the tree if needed.
2461  */
2462 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2463                             struct btrfs_root *root,
2464                             struct btrfs_path *path,
2465                             struct btrfs_key *cpu_key, u32 *data_size,
2466                             int nr)
2467 {
2468         struct extent_buffer *leaf;
2469         struct btrfs_item *item;
2470         int ret = 0;
2471         int slot;
2472         int i;
2473         u32 nritems;
2474         u32 total_size = 0;
2475         u32 total_data = 0;
2476         unsigned int data_end;
2477         struct btrfs_disk_key disk_key;
2478
2479         for (i = 0; i < nr; i++) {
2480                 total_data += data_size[i];
2481         }
2482
2483         /* create a root if there isn't one */
2484         if (!root->node)
2485                 BUG();
2486
2487         total_size = total_data + nr * sizeof(struct btrfs_item);
2488         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2489         if (ret == 0) {
2490                 return -EEXIST;
2491         }
2492         if (ret < 0)
2493                 goto out;
2494
2495         leaf = path->nodes[0];
2496
2497         nritems = btrfs_header_nritems(leaf);
2498         data_end = leaf_data_end(root, leaf);
2499
2500         if (btrfs_leaf_free_space(root, leaf) < total_size) {
2501                 btrfs_print_leaf(root, leaf);
2502                 printk("not enough freespace need %u have %d\n",
2503                        total_size, btrfs_leaf_free_space(root, leaf));
2504                 BUG();
2505         }
2506
2507         slot = path->slots[0];
2508         BUG_ON(slot < 0);
2509
2510         if (slot < nritems) {
2511                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2512
2513                 if (old_data < data_end) {
2514                         btrfs_print_leaf(root, leaf);
2515                         printk("slot %d old_data %d data_end %d\n",
2516                                slot, old_data, data_end);
2517                         BUG_ON(1);
2518                 }
2519                 /*
2520                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2521                  */
2522                 /* first correct the data pointers */
2523                 for (i = slot; i < nritems; i++) {
2524                         u32 ioff;
2525
2526                         item = btrfs_item_nr(i);
2527                         ioff = btrfs_item_offset(leaf, item);
2528                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2529                 }
2530
2531                 /* shift the items */
2532                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2533                               btrfs_item_nr_offset(slot),
2534                               (nritems - slot) * sizeof(struct btrfs_item));
2535
2536                 /* shift the data */
2537                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2538                               data_end - total_data, btrfs_leaf_data(leaf) +
2539                               data_end, old_data - data_end);
2540                 data_end = old_data;
2541         }
2542
2543         /* setup the item for the new data */
2544         for (i = 0; i < nr; i++) {
2545                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2546                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2547                 item = btrfs_item_nr(slot + i);
2548                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2549                 data_end -= data_size[i];
2550                 btrfs_set_item_size(leaf, item, data_size[i]);
2551         }
2552         btrfs_set_header_nritems(leaf, nritems + nr);
2553         btrfs_mark_buffer_dirty(leaf);
2554
2555         ret = 0;
2556         if (slot == 0) {
2557                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2558                 btrfs_fixup_low_keys(root, path, &disk_key, 1);
2559         }
2560
2561         if (btrfs_leaf_free_space(root, leaf) < 0) {
2562                 btrfs_print_leaf(root, leaf);
2563                 BUG();
2564         }
2565
2566 out:
2567         return ret;
2568 }
2569
2570 /*
2571  * Given a key and some data, insert an item into the tree.
2572  * This does all the path init required, making room in the tree if needed.
2573  */
2574 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2575                       *root, struct btrfs_key *cpu_key, void *data, u32
2576                       data_size)
2577 {
2578         int ret = 0;
2579         struct btrfs_path *path;
2580         struct extent_buffer *leaf;
2581         unsigned long ptr;
2582
2583         path = btrfs_alloc_path();
2584         if (!path)
2585                 return -ENOMEM;
2586
2587         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2588         if (!ret) {
2589                 leaf = path->nodes[0];
2590                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2591                 write_extent_buffer(leaf, data, ptr, data_size);
2592                 btrfs_mark_buffer_dirty(leaf);
2593         }
2594         btrfs_free_path(path);
2595         return ret;
2596 }
2597
2598 /*
2599  * delete the pointer from a given node.
2600  *
2601  * If the delete empties a node, the node is removed from the tree,
2602  * continuing all the way the root if required.  The root is converted into
2603  * a leaf if all the nodes are emptied.
2604  */
2605 int btrfs_del_ptr(struct btrfs_root *root, struct btrfs_path *path,
2606                 int level, int slot)
2607 {
2608         struct extent_buffer *parent = path->nodes[level];
2609         u32 nritems;
2610         int ret = 0;
2611
2612         nritems = btrfs_header_nritems(parent);
2613         if (slot < nritems - 1) {
2614                 /* shift the items */
2615                 memmove_extent_buffer(parent,
2616                               btrfs_node_key_ptr_offset(slot),
2617                               btrfs_node_key_ptr_offset(slot + 1),
2618                               sizeof(struct btrfs_key_ptr) *
2619                               (nritems - slot - 1));
2620         }
2621         nritems--;
2622         btrfs_set_header_nritems(parent, nritems);
2623         if (nritems == 0 && parent == root->node) {
2624                 BUG_ON(btrfs_header_level(root->node) != 1);
2625                 /* just turn the root into a leaf and break */
2626                 btrfs_set_header_level(root->node, 0);
2627         } else if (slot == 0) {
2628                 struct btrfs_disk_key disk_key;
2629
2630                 btrfs_node_key(parent, &disk_key, 0);
2631                 btrfs_fixup_low_keys(root, path, &disk_key, level + 1);
2632         }
2633         btrfs_mark_buffer_dirty(parent);
2634         return ret;
2635 }
2636
2637 /*
2638  * a helper function to delete the leaf pointed to by path->slots[1] and
2639  * path->nodes[1].
2640  *
2641  * This deletes the pointer in path->nodes[1] and frees the leaf
2642  * block extent.  zero is returned if it all worked out, < 0 otherwise.
2643  *
2644  * The path must have already been setup for deleting the leaf, including
2645  * all the proper balancing.  path->nodes[1] must be locked.
2646  */
2647 static noinline int btrfs_del_leaf(struct btrfs_trans_handle *trans,
2648                                    struct btrfs_root *root,
2649                                    struct btrfs_path *path,
2650                                    struct extent_buffer *leaf)
2651 {
2652         int ret;
2653
2654         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
2655         ret = btrfs_del_ptr(root, path, 1, path->slots[1]);
2656         if (ret)
2657                 return ret;
2658
2659         ret = btrfs_free_extent(trans, root, leaf->start, leaf->len,
2660                                 0, root->root_key.objectid, 0, 0);
2661         return ret;
2662 }
2663
2664 /*
2665  * delete the item at the leaf level in path.  If that empties
2666  * the leaf, remove it from the tree
2667  */
2668 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2669                     struct btrfs_path *path, int slot, int nr)
2670 {
2671         struct extent_buffer *leaf;
2672         struct btrfs_item *item;
2673         int last_off;
2674         int dsize = 0;
2675         int ret = 0;
2676         int wret;
2677         int i;
2678         u32 nritems;
2679
2680         leaf = path->nodes[0];
2681         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2682
2683         for (i = 0; i < nr; i++)
2684                 dsize += btrfs_item_size_nr(leaf, slot + i);
2685
2686         nritems = btrfs_header_nritems(leaf);
2687
2688         if (slot + nr != nritems) {
2689                 int data_end = leaf_data_end(root, leaf);
2690
2691                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2692                               data_end + dsize,
2693                               btrfs_leaf_data(leaf) + data_end,
2694                               last_off - data_end);
2695
2696                 for (i = slot + nr; i < nritems; i++) {
2697                         u32 ioff;
2698
2699                         item = btrfs_item_nr(i);
2700                         ioff = btrfs_item_offset(leaf, item);
2701                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2702                 }
2703
2704                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2705                               btrfs_item_nr_offset(slot + nr),
2706                               sizeof(struct btrfs_item) *
2707                               (nritems - slot - nr));
2708         }
2709         btrfs_set_header_nritems(leaf, nritems - nr);
2710         nritems -= nr;
2711
2712         /* delete the leaf if we've emptied it */
2713         if (nritems == 0) {
2714                 if (leaf == root->node) {
2715                         btrfs_set_header_level(leaf, 0);
2716                 } else {
2717                         clean_tree_block(trans, root, leaf);
2718                         wret = btrfs_del_leaf(trans, root, path, leaf);
2719                         BUG_ON(ret);
2720                         if (wret)
2721                                 ret = wret;
2722                 }
2723         } else {
2724                 int used = leaf_space_used(leaf, 0, nritems);
2725                 if (slot == 0) {
2726                         struct btrfs_disk_key disk_key;
2727
2728                         btrfs_item_key(leaf, &disk_key, 0);
2729                         btrfs_fixup_low_keys(root, path, &disk_key, 1);
2730                 }
2731
2732                 /* delete the leaf if it is mostly empty */
2733                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2734                         /* push_leaf_left fixes the path.
2735                          * make sure the path still points to our leaf
2736                          * for possible call to del_ptr below
2737                          */
2738                         slot = path->slots[1];
2739                         extent_buffer_get(leaf);
2740
2741                         wret = push_leaf_left(trans, root, path, 1, 1);
2742                         if (wret < 0 && wret != -ENOSPC)
2743                                 ret = wret;
2744
2745                         if (path->nodes[0] == leaf &&
2746                             btrfs_header_nritems(leaf)) {
2747                                 wret = push_leaf_right(trans, root, path, 1, 1);
2748                                 if (wret < 0 && wret != -ENOSPC)
2749                                         ret = wret;
2750                         }
2751
2752                         if (btrfs_header_nritems(leaf) == 0) {
2753                                 clean_tree_block(trans, root, leaf);
2754                                 path->slots[1] = slot;
2755                                 ret = btrfs_del_leaf(trans, root, path, leaf);
2756                                 BUG_ON(ret);
2757                                 free_extent_buffer(leaf);
2758
2759                         } else {
2760                                 btrfs_mark_buffer_dirty(leaf);
2761                                 free_extent_buffer(leaf);
2762                         }
2763                 } else {
2764                         btrfs_mark_buffer_dirty(leaf);
2765                 }
2766         }
2767         return ret;
2768 }
2769
2770 /*
2771  * walk up the tree as far as required to find the previous leaf.
2772  * returns 0 if it found something or 1 if there are no lesser leaves.
2773  * returns < 0 on io errors.
2774  */
2775 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2776 {
2777         int slot;
2778         int level = 1;
2779         struct extent_buffer *c;
2780         struct extent_buffer *next = NULL;
2781         struct btrfs_fs_info *fs_info = root->fs_info;
2782
2783         while(level < BTRFS_MAX_LEVEL) {
2784                 if (!path->nodes[level])
2785                         return 1;
2786
2787                 slot = path->slots[level];
2788                 c = path->nodes[level];
2789                 if (slot == 0) {
2790                         level++;
2791                         if (level == BTRFS_MAX_LEVEL)
2792                                 return 1;
2793                         continue;
2794                 }
2795                 slot--;
2796
2797                 next = read_node_slot(fs_info, c, slot);
2798                 if (!extent_buffer_uptodate(next)) {
2799                         if (IS_ERR(next))
2800                                 return PTR_ERR(next);
2801                         return -EIO;
2802                 }
2803                 break;
2804         }
2805         path->slots[level] = slot;
2806         while(1) {
2807                 level--;
2808                 c = path->nodes[level];
2809                 free_extent_buffer(c);
2810                 slot = btrfs_header_nritems(next);
2811                 if (slot != 0)
2812                         slot--;
2813                 path->nodes[level] = next;
2814                 path->slots[level] = slot;
2815                 if (!level)
2816                         break;
2817                 next = read_node_slot(fs_info, next, slot);
2818                 if (!extent_buffer_uptodate(next)) {
2819                         if (IS_ERR(next))
2820                                 return PTR_ERR(next);
2821                         return -EIO;
2822                 }
2823         }
2824         return 0;
2825 }
2826
2827 /*
2828  * walk up the tree as far as required to find the next leaf.
2829  * returns 0 if it found something or 1 if there are no greater leaves.
2830  * returns < 0 on io errors.
2831  */
2832 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2833 {
2834         int slot;
2835         int level = 1;
2836         struct extent_buffer *c;
2837         struct extent_buffer *next = NULL;
2838         struct btrfs_fs_info *fs_info = root->fs_info;
2839
2840         while(level < BTRFS_MAX_LEVEL) {
2841                 if (!path->nodes[level])
2842                         return 1;
2843
2844                 slot = path->slots[level] + 1;
2845                 c = path->nodes[level];
2846                 if (slot >= btrfs_header_nritems(c)) {
2847                         level++;
2848                         if (level == BTRFS_MAX_LEVEL)
2849                                 return 1;
2850                         continue;
2851                 }
2852
2853                 if (path->reada)
2854                         reada_for_search(root, path, level, slot, 0);
2855
2856                 next = read_node_slot(fs_info, c, slot);
2857                 if (!extent_buffer_uptodate(next))
2858                         return -EIO;
2859                 break;
2860         }
2861         path->slots[level] = slot;
2862         while(1) {
2863                 level--;
2864                 c = path->nodes[level];
2865                 free_extent_buffer(c);
2866                 path->nodes[level] = next;
2867                 path->slots[level] = 0;
2868                 if (!level)
2869                         break;
2870                 if (path->reada)
2871                         reada_for_search(root, path, level, 0, 0);
2872                 next = read_node_slot(fs_info, next, 0);
2873                 if (!extent_buffer_uptodate(next))
2874                         return -EIO;
2875         }
2876         return 0;
2877 }
2878
2879 int btrfs_previous_item(struct btrfs_root *root,
2880                         struct btrfs_path *path, u64 min_objectid,
2881                         int type)
2882 {
2883         struct btrfs_key found_key;
2884         struct extent_buffer *leaf;
2885         u32 nritems;
2886         int ret;
2887
2888         while(1) {
2889                 if (path->slots[0] == 0) {
2890                         ret = btrfs_prev_leaf(root, path);
2891                         if (ret != 0)
2892                                 return ret;
2893                 } else {
2894                         path->slots[0]--;
2895                 }
2896                 leaf = path->nodes[0];
2897                 nritems = btrfs_header_nritems(leaf);
2898                 if (nritems == 0)
2899                         return 1;
2900                 if (path->slots[0] == nritems)
2901                         path->slots[0]--;
2902
2903                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2904                 if (found_key.objectid < min_objectid)
2905                         break;
2906                 if (found_key.type == type)
2907                         return 0;
2908                 if (found_key.objectid == min_objectid &&
2909                     found_key.type < type)
2910                         break;
2911         }
2912         return 1;
2913 }
2914
2915 /*
2916  * search in extent tree to find a previous Metadata/Data extent item with
2917  * min objecitd.
2918  *
2919  * returns 0 if something is found, 1 if nothing was found and < 0 on error
2920  */
2921 int btrfs_previous_extent_item(struct btrfs_root *root,
2922                         struct btrfs_path *path, u64 min_objectid)
2923 {
2924         struct btrfs_key found_key;
2925         struct extent_buffer *leaf;
2926         u32 nritems;
2927         int ret;
2928
2929         while (1) {
2930                 if (path->slots[0] == 0) {
2931                         ret = btrfs_prev_leaf(root, path);
2932                         if (ret != 0)
2933                                 return ret;
2934                 } else {
2935                         path->slots[0]--;
2936                 }
2937                 leaf = path->nodes[0];
2938                 nritems = btrfs_header_nritems(leaf);
2939                 if (nritems == 0)
2940                         return 1;
2941                 if (path->slots[0] == nritems)
2942                         path->slots[0]--;
2943
2944                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2945                 if (found_key.objectid < min_objectid)
2946                         break;
2947                 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
2948                     found_key.type == BTRFS_METADATA_ITEM_KEY)
2949                         return 0;
2950                 if (found_key.objectid == min_objectid &&
2951                     found_key.type < BTRFS_EXTENT_ITEM_KEY)
2952                         break;
2953         }
2954         return 1;
2955 }
2956
2957 /*
2958  * Search in extent tree to found next meta/data extent
2959  * Caller needs to check for no-hole or skinny metadata features.
2960  */
2961 int btrfs_next_extent_item(struct btrfs_root *root,
2962                         struct btrfs_path *path, u64 max_objectid)
2963 {
2964         struct btrfs_key found_key;
2965         int ret;
2966
2967         while (1) {
2968                 ret = btrfs_next_item(root, path);
2969                 if (ret)
2970                         return ret;
2971                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2972                                       path->slots[0]);
2973                 if (found_key.objectid > max_objectid)
2974                         return 1;
2975                 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
2976                     found_key.type == BTRFS_METADATA_ITEM_KEY)
2977                 return 0;
2978         }
2979 }