ctree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include "ctree.h"
  19 #include "disk-io.h"
  20 #include "transaction.h"
  21 #include "print-tree.h"
  22 #include "repair.h"
  23
  24 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  25                       *root, struct btrfs_path *path, int level);
  26 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  27                       *root, struct btrfs_key *ins_key,
  28                       struct btrfs_path *path, int data_size, int extend);
  29 static int push_node_left(struct btrfs_trans_handle *trans,
  30                           struct btrfs_root *root, struct extent_buffer *dst,
  31                           struct extent_buffer *src, int empty);
  32 static int balance_node_right(struct btrfs_trans_handle *trans,
  33                               struct btrfs_root *root,
  34                               struct extent_buffer *dst_buf,
  35                               struct extent_buffer *src_buf);
  36
  37 inline void btrfs_init_path(struct btrfs_path *p)
  38 {
  39         memset(p, 0, sizeof(*p));
  40 }
  41
  42 struct btrfs_path *btrfs_alloc_path(void)
  43 {
  44         struct btrfs_path *path;
  45         path = kzalloc(sizeof(struct btrfs_path), GFP_NOFS);
  46         return path;
  47 }
  48
  49 void btrfs_free_path(struct btrfs_path *p)
  50 {
  51         btrfs_release_path(p);
  52         kfree(p);
  53 }
  54
  55 void btrfs_release_path(struct btrfs_path *p)
  56 {
  57         int i;
  58         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  59                 if (!p->nodes[i])
  60                         continue;
  61                 free_extent_buffer(p->nodes[i]);
  62         }
  63         memset(p, 0, sizeof(*p));
  64 }
  65
  66 void add_root_to_dirty_list(struct btrfs_root *root)
  67 {
  68         if (root->track_dirty && list_empty(&root->dirty_list)) {
  69                 list_add(&root->dirty_list,
  70                          &root->fs_info->dirty_cowonly_roots);
  71         }
  72 }
  73
  74 int btrfs_copy_root(struct btrfs_trans_handle *trans,
  75                       struct btrfs_root *root,
  76                       struct extent_buffer *buf,
  77                       struct extent_buffer **cow_ret, u64 new_root_objectid)
  78 {
  79         struct extent_buffer *cow;
  80         int ret = 0;
  81         int level;
  82         struct btrfs_root *new_root;
  83         struct btrfs_disk_key disk_key;
  84
  85         new_root = kmalloc(sizeof(*new_root), GFP_NOFS);
  86         if (!new_root)
  87                 return -ENOMEM;
  88
  89         memcpy(new_root, root, sizeof(*new_root));
  90         new_root->root_key.objectid = new_root_objectid;
  91
  92         WARN_ON(root->ref_cows && trans->transid !=
  93                 root->fs_info->running_transaction->transid);
  94         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
  95
  96         level = btrfs_header_level(buf);
  97         if (level == 0)
  98                 btrfs_item_key(buf, &disk_key, 0);
  99         else
 100                 btrfs_node_key(buf, &disk_key, 0);
 101         cow = btrfs_alloc_free_block(trans, new_root, buf->len,
 102                                      new_root_objectid, &disk_key,
 103                                      level, buf->start, 0);
 104         if (IS_ERR(cow)) {
 105                 kfree(new_root);
 106                 return PTR_ERR(cow);
 107         }
 108
 109         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 110         btrfs_set_header_bytenr(cow, cow->start);
 111         btrfs_set_header_generation(cow, trans->transid);
 112         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 113         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 114                                      BTRFS_HEADER_FLAG_RELOC);
 115         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 116                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 117         else
 118                 btrfs_set_header_owner(cow, new_root_objectid);
 119
 120         write_extent_buffer(cow, root->fs_info->fsid,
 121                             (unsigned long)btrfs_header_fsid(cow),
 122                             BTRFS_FSID_SIZE);
 123
 124         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 125         ret = btrfs_inc_ref(trans, new_root, cow, 0);
 126         kfree(new_root);
 127
 128         if (ret)
 129                 return ret;
 130
 131         btrfs_mark_buffer_dirty(cow);
 132         *cow_ret = cow;
 133         return 0;
 134 }
 135
 136 /*
 137  * check if the tree block can be shared by multiple trees
 138  */
 139 int btrfs_block_can_be_shared(struct btrfs_root *root,
 140                               struct extent_buffer *buf)
 141 {
 142         /*
 143          * Tree blocks not in refernece counted trees and tree roots
 144          * are never shared. If a block was allocated after the last
 145          * snapshot and the block was not allocated by tree relocation,
 146          * we know the block is not shared.
 147          */
 148         if (root->ref_cows &&
 149             buf != root->node && buf != root->commit_root &&
 150             (btrfs_header_generation(buf) <=
 151              btrfs_root_last_snapshot(&root->root_item) ||
 152              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 153                 return 1;
 154 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 155         if (root->ref_cows &&
 156             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 157                 return 1;
 158 #endif
 159         return 0;
 160 }
 161
 162 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
 163                                        struct btrfs_root *root,
 164                                        struct extent_buffer *buf,
 165                                        struct extent_buffer *cow)
 166 {
 167         u64 refs;
 168         u64 owner;
 169         u64 flags;
 170         u64 new_flags = 0;
 171         int ret;
 172
 173         /*
 174          * Backrefs update rules:
 175          *
 176          * Always use full backrefs for extent pointers in tree block
 177          * allocated by tree relocation.
 178          *
 179          * If a shared tree block is no longer referenced by its owner
 180          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
 181          * use full backrefs for extent pointers in tree block.
 182          *
 183          * If a tree block is been relocating
 184          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
 185          * use full backrefs for extent pointers in tree block.
 186          * The reason for this is some operations (such as drop tree)
 187          * are only allowed for blocks use full backrefs.
 188          */
 189
 190         if (btrfs_block_can_be_shared(root, buf)) {
 191                 ret = btrfs_lookup_extent_info(trans, root, buf->start,
 192                                                btrfs_header_level(buf), 1,
 193                                                &refs, &flags);
 194                 BUG_ON(ret);
 195                 BUG_ON(refs == 0);
 196         } else {
 197                 refs = 1;
 198                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 199                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 200                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 201                 else
 202                         flags = 0;
 203         }
 204
 205         owner = btrfs_header_owner(buf);
 206         BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) &&
 207                owner == BTRFS_TREE_RELOC_OBJECTID);
 208
 209         if (refs > 1) {
 210                 if ((owner == root->root_key.objectid ||
 211                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
 212                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
 213                         ret = btrfs_inc_ref(trans, root, buf, 1);
 214                         BUG_ON(ret);
 215
 216                         if (root->root_key.objectid ==
 217                             BTRFS_TREE_RELOC_OBJECTID) {
 218                                 ret = btrfs_dec_ref(trans, root, buf, 0);
 219                                 BUG_ON(ret);
 220                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 221                                 BUG_ON(ret);
 222                         }
 223                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
 224                 } else {
 225
 226                         if (root->root_key.objectid ==
 227                             BTRFS_TREE_RELOC_OBJECTID)
 228                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 229                         else
 230                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 231                         BUG_ON(ret);
 232                 }
 233                 if (new_flags != 0) {
 234                         ret = btrfs_set_block_flags(trans, root, buf->start,
 235                                                     btrfs_header_level(buf),
 236                                                     new_flags);
 237                         BUG_ON(ret);
 238                 }
 239         } else {
 240                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
 241                         if (root->root_key.objectid ==
 242                             BTRFS_TREE_RELOC_OBJECTID)
 243                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 244                         else
 245                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 246                         BUG_ON(ret);
 247                         ret = btrfs_dec_ref(trans, root, buf, 1);
 248                         BUG_ON(ret);
 249                 }
 250                 clean_tree_block(trans, root, buf);
 251         }
 252         return 0;
 253 }
 254
 255 int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 256                              struct btrfs_root *root,
 257                              struct extent_buffer *buf,
 258                              struct extent_buffer *parent, int parent_slot,
 259                              struct extent_buffer **cow_ret,
 260                              u64 search_start, u64 empty_size)
 261 {
 262         struct extent_buffer *cow;
 263         struct btrfs_disk_key disk_key;
 264         int level;
 265
 266         WARN_ON(root->ref_cows && trans->transid !=
 267                 root->fs_info->running_transaction->transid);
 268         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 269
 270         level = btrfs_header_level(buf);
 271
 272         if (level == 0)
 273                 btrfs_item_key(buf, &disk_key, 0);
 274         else
 275                 btrfs_node_key(buf, &disk_key, 0);
 276
 277         cow = btrfs_alloc_free_block(trans, root, buf->len,
 278                                      root->root_key.objectid, &disk_key,
 279                                      level, search_start, empty_size);
 280         if (IS_ERR(cow))
 281                 return PTR_ERR(cow);
 282
 283         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 284         btrfs_set_header_bytenr(cow, cow->start);
 285         btrfs_set_header_generation(cow, trans->transid);
 286         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 287         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 288                                      BTRFS_HEADER_FLAG_RELOC);
 289         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 290                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 291         else
 292                 btrfs_set_header_owner(cow, root->root_key.objectid);
 293
 294         write_extent_buffer(cow, root->fs_info->fsid,
 295                             (unsigned long)btrfs_header_fsid(cow),
 296                             BTRFS_FSID_SIZE);
 297
 298         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 299
 300         update_ref_for_cow(trans, root, buf, cow);
 301
 302         if (buf == root->node) {
 303                 root->node = cow;
 304                 extent_buffer_get(cow);
 305
 306                 btrfs_free_extent(trans, root, buf->start, buf->len,
 307                                   0, root->root_key.objectid, level, 0);
 308                 free_extent_buffer(buf);
 309                 add_root_to_dirty_list(root);
 310         } else {
 311                 btrfs_set_node_blockptr(parent, parent_slot,
 312                                         cow->start);
 313                 WARN_ON(trans->transid == 0);
 314                 btrfs_set_node_ptr_generation(parent, parent_slot,
 315                                               trans->transid);
 316                 btrfs_mark_buffer_dirty(parent);
 317                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 318
 319                 btrfs_free_extent(trans, root, buf->start, buf->len,
 320                                   0, root->root_key.objectid, level, 1);
 321         }
 322         free_extent_buffer(buf);
 323         btrfs_mark_buffer_dirty(cow);
 324         *cow_ret = cow;
 325         return 0;
 326 }
 327
 328 static inline int should_cow_block(struct btrfs_trans_handle *trans,
 329                                    struct btrfs_root *root,
 330                                    struct extent_buffer *buf)
 331 {
 332         if (btrfs_header_generation(buf) == trans->transid &&
 333             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
 334             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
 335               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 336                 return 0;
 337         return 1;
 338 }
 339
 340 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 341                     struct btrfs_root *root, struct extent_buffer *buf,
 342                     struct extent_buffer *parent, int parent_slot,
 343                     struct extent_buffer **cow_ret)
 344 {
 345         u64 search_start;
 346         int ret;
 347         /*
 348         if (trans->transaction != root->fs_info->running_transaction) {
 349                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 350                        root->fs_info->running_transaction->transid);
 351                 WARN_ON(1);
 352         }
 353         */
 354         if (trans->transid != root->fs_info->generation) {
 355                 printk(KERN_CRIT "trans %llu running %llu\n",
 356                         (unsigned long long)trans->transid,
 357                         (unsigned long long)root->fs_info->generation);
 358                 WARN_ON(1);
 359         }
 360         if (!should_cow_block(trans, root, buf)) {
 361                 *cow_ret = buf;
 362                 return 0;
 363         }
 364
 365         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
 366         ret = __btrfs_cow_block(trans, root, buf, parent,
 367                                  parent_slot, cow_ret, search_start, 0);
 368         return ret;
 369 }
 370
 371 /*
 372  * compare two keys in a memcmp fashion
 373  */
 374 static int btrfs_comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 375 {
 376         struct btrfs_key k1;
 377
 378         btrfs_disk_key_to_cpu(&k1, disk);
 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  * The leaf data grows from end-to-front in the node.
 397  * this returns the address of the start of the last item,
 398  * which is the stop of the leaf data stack
 399  */
 400 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 401                                          struct extent_buffer *leaf)
 402 {
 403         u32 nr = btrfs_header_nritems(leaf);
 404         if (nr == 0)
 405                 return BTRFS_LEAF_DATA_SIZE(root);
 406         return btrfs_item_offset_nr(leaf, nr - 1);
 407 }
 408
 409 int btrfs_check_node(struct btrfs_root *root,
 410                       struct btrfs_disk_key *parent_key,
 411                       struct extent_buffer *buf)
 412 {
 413         int i;
 414         struct btrfs_key cpukey;
 415         struct btrfs_disk_key key;
 416         u32 nritems = btrfs_header_nritems(buf);
 417
 418         if (nritems == 0 || nritems > BTRFS_NODEPTRS_PER_BLOCK(root))
 419                 goto fail;
 420
 421         if (parent_key && parent_key->type) {
 422                 btrfs_node_key(buf, &key, 0);
 423                 if (memcmp(parent_key, &key, sizeof(key)))
 424                         goto fail;
 425         }
 426         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 427                 btrfs_node_key(buf, &key, i);
 428                 btrfs_node_key_to_cpu(buf, &cpukey, i + 1);
 429                 if (btrfs_comp_keys(&key, &cpukey) >= 0)
 430                         goto fail;
 431         }
 432         return 0;
 433 fail:
 434         if (btrfs_header_owner(buf) == BTRFS_EXTENT_TREE_OBJECTID) {
 435                 if (parent_key)
 436                         btrfs_disk_key_to_cpu(&cpukey, parent_key);
 437                 else
 438                         btrfs_node_key_to_cpu(buf, &cpukey, 0);
 439                 btrfs_add_corrupt_extent_record(root->fs_info, &cpukey,
 440                                                 buf->start, buf->len,
 441                                                 btrfs_header_level(buf));
 442         }
 443         return -EIO;
 444 }
 445
 446 int btrfs_check_leaf(struct btrfs_root *root,
 447                       struct btrfs_disk_key *parent_key,
 448                       struct extent_buffer *buf)
 449 {
 450         int i;
 451         struct btrfs_key cpukey;
 452         struct btrfs_disk_key key;
 453         u32 nritems = btrfs_header_nritems(buf);
 454
 455         if (nritems * sizeof(struct btrfs_item) > buf->len)  {
 456                 fprintf(stderr, "invalid number of items %llu\n",
 457                         (unsigned long long)buf->start);
 458                 goto fail;
 459         }
 460
 461         if (btrfs_header_level(buf) != 0) {
 462                 fprintf(stderr, "leaf is not a leaf %llu\n",
 463                        (unsigned long long)btrfs_header_bytenr(buf));
 464                 goto fail;
 465         }
 466         if (btrfs_leaf_free_space(root, buf) < 0) {
 467                 fprintf(stderr, "leaf free space incorrect %llu %d\n",
 468                         (unsigned long long)btrfs_header_bytenr(buf),
 469                         btrfs_leaf_free_space(root, buf));
 470                 goto fail;
 471         }
 472
 473         if (nritems == 0)
 474                 return 0;
 475
 476         btrfs_item_key(buf, &key, 0);
 477         if (parent_key && parent_key->type &&
 478             memcmp(parent_key, &key, sizeof(key))) {
 479                 fprintf(stderr, "leaf parent key incorrect %llu\n",
 480                        (unsigned long long)btrfs_header_bytenr(buf));
 481                 goto fail;
 482         }
 483         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 484                 btrfs_item_key(buf, &key, i);
 485                 btrfs_item_key_to_cpu(buf, &cpukey, i + 1);
 486                 if (btrfs_comp_keys(&key, &cpukey) >= 0) {
 487                         fprintf(stderr, "bad key ordering %d %d\n", i, i+1);
 488                         goto fail;
 489                 }
 490                 if (btrfs_item_offset_nr(buf, i) !=
 491                         btrfs_item_end_nr(buf, i + 1)) {
 492                         fprintf(stderr, "incorrect offsets %u %u\n",
 493                                 btrfs_item_offset_nr(buf, i),
 494                                 btrfs_item_end_nr(buf, i + 1));
 495                         goto fail;
 496                 }
 497                 if (i == 0 && btrfs_item_end_nr(buf, i) !=
 498                     BTRFS_LEAF_DATA_SIZE(root)) {
 499                         fprintf(stderr, "bad item end %u wanted %u\n",
 500                                 btrfs_item_end_nr(buf, i),
 501                                 (unsigned)BTRFS_LEAF_DATA_SIZE(root));
 502                         goto fail;
 503                 }
 504         }
 505         return 0;
 506 fail:
 507         if (btrfs_header_owner(buf) == BTRFS_EXTENT_TREE_OBJECTID) {
 508                 if (parent_key)
 509                         btrfs_disk_key_to_cpu(&cpukey, parent_key);
 510                 else
 511                         btrfs_item_key_to_cpu(buf, &cpukey, 0);
 512
 513                 btrfs_add_corrupt_extent_record(root->fs_info, &cpukey,
 514                                                 buf->start, buf->len, 0);
 515         }
 516         return -EIO;
 517 }
 518
 519 static int noinline check_block(struct btrfs_root *root,
 520                                 struct btrfs_path *path, int level)
 521 {
 522         struct btrfs_disk_key key;
 523         struct btrfs_disk_key *key_ptr = NULL;
 524         struct extent_buffer *parent;
 525
 526         if (path->nodes[level + 1]) {
 527                 parent = path->nodes[level + 1];
 528                 btrfs_node_key(parent, &key, path->slots[level + 1]);
 529                 key_ptr = &key;
 530         }
 531         if (level == 0)
 532                 return btrfs_check_leaf(root, key_ptr, path->nodes[0]);
 533         return btrfs_check_node(root, key_ptr, path->nodes[level]);
 534 }
 535
 536 /*
 537  * search for key in the extent_buffer.  The items start at offset p,
 538  * and they are item_size apart.  There are 'max' items in p.
 539  *
 540  * the slot in the array is returned via slot, and it points to
 541  * the place where you would insert key if it is not found in
 542  * the array.
 543  *
 544  * slot may point to max if the key is bigger than all of the keys
 545  */
 546 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 547                               int item_size, struct btrfs_key *key,
 548                               int max, int *slot)
 549 {
 550         int low = 0;
 551         int high = max;
 552         int mid;
 553         int ret;
 554         unsigned long offset;
 555         struct btrfs_disk_key *tmp;
 556
 557         while(low < high) {
 558                 mid = (low + high) / 2;
 559                 offset = p + mid * item_size;
 560
 561                 tmp = (struct btrfs_disk_key *)(eb->data + offset);
 562                 ret = btrfs_comp_keys(tmp, key);
 563
 564                 if (ret < 0)
 565                         low = mid + 1;
 566                 else if (ret > 0)
 567                         high = mid;
 568                 else {
 569                         *slot = mid;
 570                         return 0;
 571                 }
 572         }
 573         *slot = low;
 574         return 1;
 575 }
 576
 577 /*
 578  * simple bin_search frontend that does the right thing for
 579  * leaves vs nodes
 580  */
 581 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 582                       int level, int *slot)
 583 {
 584         if (level == 0)
 585                 return generic_bin_search(eb,
 586                                           offsetof(struct btrfs_leaf, items),
 587                                           sizeof(struct btrfs_item),
 588                                           key, btrfs_header_nritems(eb),
 589                                           slot);
 590         else
 591                 return generic_bin_search(eb,
 592                                           offsetof(struct btrfs_node, ptrs),
 593                                           sizeof(struct btrfs_key_ptr),
 594                                           key, btrfs_header_nritems(eb),
 595                                           slot);
 596 }
 597
 598 struct extent_buffer *read_node_slot(struct btrfs_root *root,
 599                                    struct extent_buffer *parent, int slot)
 600 {
 601         int level = btrfs_header_level(parent);
 602         if (slot < 0)
 603                 return NULL;
 604         if (slot >= btrfs_header_nritems(parent))
 605                 return NULL;
 606
 607         if (level == 0)
 608                 return NULL;
 609
 610         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 611                        btrfs_level_size(root, level - 1),
 612                        btrfs_node_ptr_generation(parent, slot));
 613 }
 614
 615 static int balance_level(struct btrfs_trans_handle *trans,
 616                          struct btrfs_root *root,
 617                          struct btrfs_path *path, int level)
 618 {
 619         struct extent_buffer *right = NULL;
 620         struct extent_buffer *mid;
 621         struct extent_buffer *left = NULL;
 622         struct extent_buffer *parent = NULL;
 623         int ret = 0;
 624         int wret;
 625         int pslot;
 626         int orig_slot = path->slots[level];
 627         u64 orig_ptr;
 628
 629         if (level == 0)
 630                 return 0;
 631
 632         mid = path->nodes[level];
 633         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 634
 635         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 636
 637         if (level < BTRFS_MAX_LEVEL - 1) {
 638                 parent = path->nodes[level + 1];
 639                 pslot = path->slots[level + 1];
 640         }
 641
 642         /*
 643          * deal with the case where there is only one pointer in the root
 644          * by promoting the node below to a root
 645          */
 646         if (!parent) {
 647                 struct extent_buffer *child;
 648
 649                 if (btrfs_header_nritems(mid) != 1)
 650                         return 0;
 651
 652                 /* promote the child to a root */
 653                 child = read_node_slot(root, mid, 0);
 654                 BUG_ON(!child);
 655                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
 656                 BUG_ON(ret);
 657
 658                 root->node = child;
 659                 add_root_to_dirty_list(root);
 660                 path->nodes[level] = NULL;
 661                 clean_tree_block(trans, root, mid);
 662                 wait_on_tree_block_writeback(root, mid);
 663                 /* once for the path */
 664                 free_extent_buffer(mid);
 665
 666                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 667                                         0, root->root_key.objectid,
 668                                         level, 1);
 669                 /* once for the root ptr */
 670                 free_extent_buffer(mid);
 671                 return ret;
 672         }
 673         if (btrfs_header_nritems(mid) >
 674             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 675                 return 0;
 676
 677         left = read_node_slot(root, parent, pslot - 1);
 678         if (left) {
 679                 wret = btrfs_cow_block(trans, root, left,
 680                                        parent, pslot - 1, &left);
 681                 if (wret) {
 682                         ret = wret;
 683                         goto enospc;
 684                 }
 685         }
 686         right = read_node_slot(root, parent, pslot + 1);
 687         if (right) {
 688                 wret = btrfs_cow_block(trans, root, right,
 689                                        parent, pslot + 1, &right);
 690                 if (wret) {
 691                         ret = wret;
 692                         goto enospc;
 693                 }
 694         }
 695
 696         /* first, try to make some room in the middle buffer */
 697         if (left) {
 698                 orig_slot += btrfs_header_nritems(left);
 699                 wret = push_node_left(trans, root, left, mid, 1);
 700                 if (wret < 0)
 701                         ret = wret;
 702         }
 703
 704         /*
 705          * then try to empty the right most buffer into the middle
 706          */
 707         if (right) {
 708                 wret = push_node_left(trans, root, mid, right, 1);
 709                 if (wret < 0 && wret != -ENOSPC)
 710                         ret = wret;
 711                 if (btrfs_header_nritems(right) == 0) {
 712                         u64 bytenr = right->start;
 713                         u32 blocksize = right->len;
 714
 715                         clean_tree_block(trans, root, right);
 716                         wait_on_tree_block_writeback(root, right);
 717                         free_extent_buffer(right);
 718                         right = NULL;
 719                         wret = btrfs_del_ptr(trans, root, path,
 720                                              level + 1, pslot + 1);
 721                         if (wret)
 722                                 ret = wret;
 723                         wret = btrfs_free_extent(trans, root, bytenr,
 724                                                  blocksize, 0,
 725                                                  root->root_key.objectid,
 726                                                  level, 0);
 727                         if (wret)
 728                                 ret = wret;
 729                 } else {
 730                         struct btrfs_disk_key right_key;
 731                         btrfs_node_key(right, &right_key, 0);
 732                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 733                         btrfs_mark_buffer_dirty(parent);
 734                 }
 735         }
 736         if (btrfs_header_nritems(mid) == 1) {
 737                 /*
 738                  * we're not allowed to leave a node with one item in the
 739                  * tree during a delete.  A deletion from lower in the tree
 740                  * could try to delete the only pointer in this node.
 741                  * So, pull some keys from the left.
 742                  * There has to be a left pointer at this point because
 743                  * otherwise we would have pulled some pointers from the
 744                  * right
 745                  */
 746                 BUG_ON(!left);
 747                 wret = balance_node_right(trans, root, mid, left);
 748                 if (wret < 0) {
 749                         ret = wret;
 750                         goto enospc;
 751                 }
 752                 if (wret == 1) {
 753                         wret = push_node_left(trans, root, left, mid, 1);
 754                         if (wret < 0)
 755                                 ret = wret;
 756                 }
 757                 BUG_ON(wret == 1);
 758         }
 759         if (btrfs_header_nritems(mid) == 0) {
 760                 /* we've managed to empty the middle node, drop it */
 761                 u64 bytenr = mid->start;
 762                 u32 blocksize = mid->len;
 763                 clean_tree_block(trans, root, mid);
 764                 wait_on_tree_block_writeback(root, mid);
 765                 free_extent_buffer(mid);
 766                 mid = NULL;
 767                 wret = btrfs_del_ptr(trans, root, path, level + 1, pslot);
 768                 if (wret)
 769                         ret = wret;
 770                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 771                                          0, root->root_key.objectid,
 772                                          level, 0);
 773                 if (wret)
 774                         ret = wret;
 775         } else {
 776                 /* update the parent key to reflect our changes */
 777                 struct btrfs_disk_key mid_key;
 778                 btrfs_node_key(mid, &mid_key, 0);
 779                 btrfs_set_node_key(parent, &mid_key, pslot);
 780                 btrfs_mark_buffer_dirty(parent);
 781         }
 782
 783         /* update the path */
 784         if (left) {
 785                 if (btrfs_header_nritems(left) > orig_slot) {
 786                         extent_buffer_get(left);
 787                         path->nodes[level] = left;
 788                         path->slots[level + 1] -= 1;
 789                         path->slots[level] = orig_slot;
 790                         if (mid)
 791                                 free_extent_buffer(mid);
 792                 } else {
 793                         orig_slot -= btrfs_header_nritems(left);
 794                         path->slots[level] = orig_slot;
 795                 }
 796         }
 797         /* double check we haven't messed things up */
 798         check_block(root, path, level);
 799         if (orig_ptr !=
 800             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 801                 BUG();
 802 enospc:
 803         if (right)
 804                 free_extent_buffer(right);
 805         if (left)
 806                 free_extent_buffer(left);
 807         return ret;
 808 }
 809
 810 /* returns zero if the push worked, non-zero otherwise */
 811 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
 812                                           struct btrfs_root *root,
 813                                           struct btrfs_path *path, int level)
 814 {
 815         struct extent_buffer *right = NULL;
 816         struct extent_buffer *mid;
 817         struct extent_buffer *left = NULL;
 818         struct extent_buffer *parent = NULL;
 819         int ret = 0;
 820         int wret;
 821         int pslot;
 822         int orig_slot = path->slots[level];
 823
 824         if (level == 0)
 825                 return 1;
 826
 827         mid = path->nodes[level];
 828         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 829
 830         if (level < BTRFS_MAX_LEVEL - 1) {
 831                 parent = path->nodes[level + 1];
 832                 pslot = path->slots[level + 1];
 833         }
 834
 835         if (!parent)
 836                 return 1;
 837
 838         left = read_node_slot(root, parent, pslot - 1);
 839
 840         /* first, try to make some room in the middle buffer */
 841         if (left) {
 842                 u32 left_nr;
 843                 left_nr = btrfs_header_nritems(left);
 844                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 845                         wret = 1;
 846                 } else {
 847                         ret = btrfs_cow_block(trans, root, left, parent,
 848                                               pslot - 1, &left);
 849                         if (ret)
 850                                 wret = 1;
 851                         else {
 852                                 wret = push_node_left(trans, root,
 853                                                       left, mid, 0);
 854                         }
 855                 }
 856                 if (wret < 0)
 857                         ret = wret;
 858                 if (wret == 0) {
 859                         struct btrfs_disk_key disk_key;
 860                         orig_slot += left_nr;
 861                         btrfs_node_key(mid, &disk_key, 0);
 862                         btrfs_set_node_key(parent, &disk_key, pslot);
 863                         btrfs_mark_buffer_dirty(parent);
 864                         if (btrfs_header_nritems(left) > orig_slot) {
 865                                 path->nodes[level] = left;
 866                                 path->slots[level + 1] -= 1;
 867                                 path->slots[level] = orig_slot;
 868                                 free_extent_buffer(mid);
 869                         } else {
 870                                 orig_slot -=
 871                                         btrfs_header_nritems(left);
 872                                 path->slots[level] = orig_slot;
 873                                 free_extent_buffer(left);
 874                         }
 875                         return 0;
 876                 }
 877                 free_extent_buffer(left);
 878         }
 879         right= read_node_slot(root, parent, pslot + 1);
 880
 881         /*
 882          * then try to empty the right most buffer into the middle
 883          */
 884         if (right) {
 885                 u32 right_nr;
 886                 right_nr = btrfs_header_nritems(right);
 887                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 888                         wret = 1;
 889                 } else {
 890                         ret = btrfs_cow_block(trans, root, right,
 891                                               parent, pslot + 1,
 892                                               &right);
 893                         if (ret)
 894                                 wret = 1;
 895                         else {
 896                                 wret = balance_node_right(trans, root,
 897                                                           right, mid);
 898                         }
 899                 }
 900                 if (wret < 0)
 901                         ret = wret;
 902                 if (wret == 0) {
 903                         struct btrfs_disk_key disk_key;
 904
 905                         btrfs_node_key(right, &disk_key, 0);
 906                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
 907                         btrfs_mark_buffer_dirty(parent);
 908
 909                         if (btrfs_header_nritems(mid) <= orig_slot) {
 910                                 path->nodes[level] = right;
 911                                 path->slots[level + 1] += 1;
 912                                 path->slots[level] = orig_slot -
 913                                         btrfs_header_nritems(mid);
 914                                 free_extent_buffer(mid);
 915                         } else {
 916                                 free_extent_buffer(right);
 917                         }
 918                         return 0;
 919                 }
 920                 free_extent_buffer(right);
 921         }
 922         return 1;
 923 }
 924
 925 /*
 926  * readahead one full node of leaves
 927  */
 928 void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
 929                              int level, int slot, u64 objectid)
 930 {
 931         struct extent_buffer *node;
 932         struct btrfs_disk_key disk_key;
 933         u32 nritems;
 934         u64 search;
 935         u64 lowest_read;
 936         u64 highest_read;
 937         u64 nread = 0;
 938         int direction = path->reada;
 939         struct extent_buffer *eb;
 940         u32 nr;
 941         u32 blocksize;
 942         u32 nscan = 0;
 943
 944         if (level != 1)
 945                 return;
 946
 947         if (!path->nodes[level])
 948                 return;
 949
 950         node = path->nodes[level];
 951         search = btrfs_node_blockptr(node, slot);
 952         blocksize = btrfs_level_size(root, level - 1);
 953         eb = btrfs_find_tree_block(root, search, blocksize);
 954         if (eb) {
 955                 free_extent_buffer(eb);
 956                 return;
 957         }
 958
 959         highest_read = search;
 960         lowest_read = search;
 961
 962         nritems = btrfs_header_nritems(node);
 963         nr = slot;
 964         while(1) {
 965                 if (direction < 0) {
 966                         if (nr == 0)
 967                                 break;
 968                         nr--;
 969                 } else if (direction > 0) {
 970                         nr++;
 971                         if (nr >= nritems)
 972                                 break;
 973                 }
 974                 if (path->reada < 0 && objectid) {
 975                         btrfs_node_key(node, &disk_key, nr);
 976                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
 977                                 break;
 978                 }
 979                 search = btrfs_node_blockptr(node, nr);
 980                 if ((search >= lowest_read && search <= highest_read) ||
 981                     (search < lowest_read && lowest_read - search <= 32768) ||
 982                     (search > highest_read && search - highest_read <= 32768)) {
 983                         readahead_tree_block(root, search, blocksize,
 984                                      btrfs_node_ptr_generation(node, nr));
 985                         nread += blocksize;
 986                 }
 987                 nscan++;
 988                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
 989                         break;
 990                 if(nread > (1024 * 1024) || nscan > 128)
 991                         break;
 992
 993                 if (search < lowest_read)
 994                         lowest_read = search;
 995                 if (search > highest_read)
 996                         highest_read = search;
 997         }
 998 }
 999
1000 /*
1001  * look for key in the tree.  path is filled in with nodes along the way
1002  * if key is found, we return zero and you can find the item in the leaf
1003  * level of the path (level 0)
1004  *
1005  * If the key isn't found, the path points to the slot where it should
1006  * be inserted, and 1 is returned.  If there are other errors during the
1007  * search a negative error number is returned.
1008  *
1009  * if ins_len > 0, nodes and leaves will be split as we walk down the
1010  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1011  * possible)
1012  */
1013 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1014                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1015                       ins_len, int cow)
1016 {
1017         struct extent_buffer *b;
1018         int slot;
1019         int ret;
1020         int level;
1021         int should_reada = p->reada;
1022         u8 lowest_level = 0;
1023
1024         lowest_level = p->lowest_level;
1025         WARN_ON(lowest_level && ins_len > 0);
1026         WARN_ON(p->nodes[0] != NULL);
1027         /*
1028         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
1029         */
1030 again:
1031         b = root->node;
1032         extent_buffer_get(b);
1033         while (b) {
1034                 level = btrfs_header_level(b);
1035                 if (cow) {
1036                         int wret;
1037                         wret = btrfs_cow_block(trans, root, b,
1038                                                p->nodes[level + 1],
1039                                                p->slots[level + 1],
1040                                                &b);
1041                         if (wret) {
1042                                 free_extent_buffer(b);
1043                                 return wret;
1044                         }
1045                 }
1046                 BUG_ON(!cow && ins_len);
1047                 if (level != btrfs_header_level(b))
1048                         WARN_ON(1);
1049                 level = btrfs_header_level(b);
1050                 p->nodes[level] = b;
1051                 ret = check_block(root, p, level);
1052                 if (ret)
1053                         return -1;
1054                 ret = bin_search(b, key, level, &slot);
1055                 if (level != 0) {
1056                         if (ret && slot > 0)
1057                                 slot -= 1;
1058                         p->slots[level] = slot;
1059                         if ((p->search_for_split || ins_len > 0) &&
1060                             btrfs_header_nritems(b) >=
1061                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1062                                 int sret = split_node(trans, root, p, level);
1063                                 BUG_ON(sret > 0);
1064                                 if (sret)
1065                                         return sret;
1066                                 b = p->nodes[level];
1067                                 slot = p->slots[level];
1068                         } else if (ins_len < 0) {
1069                                 int sret = balance_level(trans, root, p,
1070                                                          level);
1071                                 if (sret)
1072                                         return sret;
1073                                 b = p->nodes[level];
1074                                 if (!b) {
1075                                         btrfs_release_path(p);
1076                                         goto again;
1077                                 }
1078                                 slot = p->slots[level];
1079                                 BUG_ON(btrfs_header_nritems(b) == 1);
1080                         }
1081                         /* this is only true while dropping a snapshot */
1082                         if (level == lowest_level)
1083                                 break;
1084
1085                         if (should_reada)
1086                                 reada_for_search(root, p, level, slot,
1087                                                  key->objectid);
1088
1089                         b = read_node_slot(root, b, slot);
1090                         if (!extent_buffer_uptodate(b))
1091                                 return -EIO;
1092                 } else {
1093                         p->slots[level] = slot;
1094                         if (ins_len > 0 &&
1095                             ins_len > btrfs_leaf_free_space(root, b)) {
1096                                 int sret = split_leaf(trans, root, key,
1097                                                       p, ins_len, ret == 0);
1098                                 BUG_ON(sret > 0);
1099                                 if (sret)
1100                                         return sret;
1101                         }
1102                         return ret;
1103                 }
1104         }
1105         return 1;
1106 }
1107
1108 /*
1109  * adjust the pointers going up the tree, starting at level
1110  * making sure the right key of each node is points to 'key'.
1111  * This is used after shifting pointers to the left, so it stops
1112  * fixing up pointers when a given leaf/node is not in slot 0 of the
1113  * higher levels
1114  *
1115  * If this fails to write a tree block, it returns -1, but continues
1116  * fixing up the blocks in ram so the tree is consistent.
1117  */
1118 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1119                           struct btrfs_root *root, struct btrfs_path *path,
1120                           struct btrfs_disk_key *key, int level)
1121 {
1122         int i;
1123         int ret = 0;
1124         struct extent_buffer *t;
1125
1126         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1127                 int tslot = path->slots[i];
1128                 if (!path->nodes[i])
1129                         break;
1130                 t = path->nodes[i];
1131                 btrfs_set_node_key(t, key, tslot);
1132                 btrfs_mark_buffer_dirty(path->nodes[i]);
1133                 if (tslot != 0)
1134                         break;
1135         }
1136         return ret;
1137 }
1138
1139 /*
1140  * update item key.
1141  *
1142  * This function isn't completely safe. It's the caller's responsibility
1143  * that the new key won't break the order
1144  */
1145 int btrfs_set_item_key_safe(struct btrfs_trans_handle *trans,
1146                             struct btrfs_root *root, struct btrfs_path *path,
1147                             struct btrfs_key *new_key)
1148 {
1149         struct btrfs_disk_key disk_key;
1150         struct extent_buffer *eb;
1151         int slot;
1152
1153         eb = path->nodes[0];
1154         slot = path->slots[0];
1155         if (slot > 0) {
1156                 btrfs_item_key(eb, &disk_key, slot - 1);
1157                 if (btrfs_comp_keys(&disk_key, new_key) >= 0)
1158                         return -1;
1159         }
1160         if (slot < btrfs_header_nritems(eb) - 1) {
1161                 btrfs_item_key(eb, &disk_key, slot + 1);
1162                 if (btrfs_comp_keys(&disk_key, new_key) <= 0)
1163                         return -1;
1164         }
1165
1166         btrfs_cpu_key_to_disk(&disk_key, new_key);
1167         btrfs_set_item_key(eb, &disk_key, slot);
1168         btrfs_mark_buffer_dirty(eb);
1169         if (slot == 0)
1170                 fixup_low_keys(trans, root, path, &disk_key, 1);
1171         return 0;
1172 }
1173
1174 /*
1175  * try to push data from one node into the next node left in the
1176  * tree.
1177  *
1178  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1179  * error, and > 0 if there was no room in the left hand block.
1180  */
1181 static int push_node_left(struct btrfs_trans_handle *trans,
1182                           struct btrfs_root *root, struct extent_buffer *dst,
1183                           struct extent_buffer *src, int empty)
1184 {
1185         int push_items = 0;
1186         int src_nritems;
1187         int dst_nritems;
1188         int ret = 0;
1189
1190         src_nritems = btrfs_header_nritems(src);
1191         dst_nritems = btrfs_header_nritems(dst);
1192         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1193         WARN_ON(btrfs_header_generation(src) != trans->transid);
1194         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1195
1196         if (!empty && src_nritems <= 8)
1197                 return 1;
1198
1199         if (push_items <= 0) {
1200                 return 1;
1201         }
1202
1203         if (empty) {
1204                 push_items = min(src_nritems, push_items);
1205                 if (push_items < src_nritems) {
1206                         /* leave at least 8 pointers in the node if
1207                          * we aren't going to empty it
1208                          */
1209                         if (src_nritems - push_items < 8) {
1210                                 if (push_items <= 8)
1211                                         return 1;
1212                                 push_items -= 8;
1213                         }
1214                 }
1215         } else
1216                 push_items = min(src_nritems - 8, push_items);
1217
1218         copy_extent_buffer(dst, src,
1219                            btrfs_node_key_ptr_offset(dst_nritems),
1220                            btrfs_node_key_ptr_offset(0),
1221                            push_items * sizeof(struct btrfs_key_ptr));
1222
1223         if (push_items < src_nritems) {
1224                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1225                                       btrfs_node_key_ptr_offset(push_items),
1226                                       (src_nritems - push_items) *
1227                                       sizeof(struct btrfs_key_ptr));
1228         }
1229         btrfs_set_header_nritems(src, src_nritems - push_items);
1230         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1231         btrfs_mark_buffer_dirty(src);
1232         btrfs_mark_buffer_dirty(dst);
1233
1234         return ret;
1235 }
1236
1237 /*
1238  * try to push data from one node into the next node right in the
1239  * tree.
1240  *
1241  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1242  * error, and > 0 if there was no room in the right hand block.
1243  *
1244  * this will  only push up to 1/2 the contents of the left node over
1245  */
1246 static int balance_node_right(struct btrfs_trans_handle *trans,
1247                               struct btrfs_root *root,
1248                               struct extent_buffer *dst,
1249                               struct extent_buffer *src)
1250 {
1251         int push_items = 0;
1252         int max_push;
1253         int src_nritems;
1254         int dst_nritems;
1255         int ret = 0;
1256
1257         WARN_ON(btrfs_header_generation(src) != trans->transid);
1258         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1259
1260         src_nritems = btrfs_header_nritems(src);
1261         dst_nritems = btrfs_header_nritems(dst);
1262         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1263         if (push_items <= 0) {
1264                 return 1;
1265         }
1266
1267         if (src_nritems < 4) {
1268                 return 1;
1269         }
1270
1271         max_push = src_nritems / 2 + 1;
1272         /* don't try to empty the node */
1273         if (max_push >= src_nritems) {
1274                 return 1;
1275         }
1276
1277         if (max_push < push_items)
1278                 push_items = max_push;
1279
1280         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1281                                       btrfs_node_key_ptr_offset(0),
1282                                       (dst_nritems) *
1283                                       sizeof(struct btrfs_key_ptr));
1284
1285         copy_extent_buffer(dst, src,
1286                            btrfs_node_key_ptr_offset(0),
1287                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1288                            push_items * sizeof(struct btrfs_key_ptr));
1289
1290         btrfs_set_header_nritems(src, src_nritems - push_items);
1291         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1292
1293         btrfs_mark_buffer_dirty(src);
1294         btrfs_mark_buffer_dirty(dst);
1295
1296         return ret;
1297 }
1298
1299 /*
1300  * helper function to insert a new root level in the tree.
1301  * A new node is allocated, and a single item is inserted to
1302  * point to the existing root
1303  *
1304  * returns zero on success or < 0 on failure.
1305  */
1306 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1307                            struct btrfs_root *root,
1308                            struct btrfs_path *path, int level)
1309 {
1310         u64 lower_gen;
1311         struct extent_buffer *lower;
1312         struct extent_buffer *c;
1313         struct extent_buffer *old;
1314         struct btrfs_disk_key lower_key;
1315
1316         BUG_ON(path->nodes[level]);
1317         BUG_ON(path->nodes[level-1] != root->node);
1318
1319         lower = path->nodes[level-1];
1320         if (level == 1)
1321                 btrfs_item_key(lower, &lower_key, 0);
1322         else
1323                 btrfs_node_key(lower, &lower_key, 0);
1324
1325         c = btrfs_alloc_free_block(trans, root, root->nodesize,
1326                                    root->root_key.objectid, &lower_key,
1327                                    level, root->node->start, 0);
1328
1329         if (IS_ERR(c))
1330                 return PTR_ERR(c);
1331
1332         memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
1333         btrfs_set_header_nritems(c, 1);
1334         btrfs_set_header_level(c, level);
1335         btrfs_set_header_bytenr(c, c->start);
1336         btrfs_set_header_generation(c, trans->transid);
1337         btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
1338         btrfs_set_header_owner(c, root->root_key.objectid);
1339
1340         write_extent_buffer(c, root->fs_info->fsid,
1341                             (unsigned long)btrfs_header_fsid(c),
1342                             BTRFS_FSID_SIZE);
1343
1344         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1345                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
1346                             BTRFS_UUID_SIZE);
1347
1348         btrfs_set_node_key(c, &lower_key, 0);
1349         btrfs_set_node_blockptr(c, 0, lower->start);
1350         lower_gen = btrfs_header_generation(lower);
1351         WARN_ON(lower_gen != trans->transid);
1352
1353         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1354
1355         btrfs_mark_buffer_dirty(c);
1356
1357         old = root->node;
1358         root->node = c;
1359
1360         /* the super has an extra ref to root->node */
1361         free_extent_buffer(old);
1362
1363         add_root_to_dirty_list(root);
1364         extent_buffer_get(c);
1365         path->nodes[level] = c;
1366         path->slots[level] = 0;
1367         return 0;
1368 }
1369
1370 /*
1371  * worker function to insert a single pointer in a node.
1372  * the node should have enough room for the pointer already
1373  *
1374  * slot and level indicate where you want the key to go, and
1375  * blocknr is the block the key points to.
1376  *
1377  * returns zero on success and < 0 on any error
1378  */
1379 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1380                       *root, struct btrfs_path *path, struct btrfs_disk_key
1381                       *key, u64 bytenr, int slot, int level)
1382 {
1383         struct extent_buffer *lower;
1384         int nritems;
1385
1386         BUG_ON(!path->nodes[level]);
1387         lower = path->nodes[level];
1388         nritems = btrfs_header_nritems(lower);
1389         if (slot > nritems)
1390                 BUG();
1391         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1392                 BUG();
1393         if (slot != nritems) {
1394                 memmove_extent_buffer(lower,
1395                               btrfs_node_key_ptr_offset(slot + 1),
1396                               btrfs_node_key_ptr_offset(slot),
1397                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1398         }
1399         btrfs_set_node_key(lower, key, slot);
1400         btrfs_set_node_blockptr(lower, slot, bytenr);
1401         WARN_ON(trans->transid == 0);
1402         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1403         btrfs_set_header_nritems(lower, nritems + 1);
1404         btrfs_mark_buffer_dirty(lower);
1405         return 0;
1406 }
1407
1408 /*
1409  * split the node at the specified level in path in two.
1410  * The path is corrected to point to the appropriate node after the split
1411  *
1412  * Before splitting this tries to make some room in the node by pushing
1413  * left and right, if either one works, it returns right away.
1414  *
1415  * returns 0 on success and < 0 on failure
1416  */
1417 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1418                       *root, struct btrfs_path *path, int level)
1419 {
1420         struct extent_buffer *c;
1421         struct extent_buffer *split;
1422         struct btrfs_disk_key disk_key;
1423         int mid;
1424         int ret;
1425         int wret;
1426         u32 c_nritems;
1427
1428         c = path->nodes[level];
1429         WARN_ON(btrfs_header_generation(c) != trans->transid);
1430         if (c == root->node) {
1431                 /* trying to split the root, lets make a new one */
1432                 ret = insert_new_root(trans, root, path, level + 1);
1433                 if (ret)
1434                         return ret;
1435         } else {
1436                 ret = push_nodes_for_insert(trans, root, path, level);
1437                 c = path->nodes[level];
1438                 if (!ret && btrfs_header_nritems(c) <
1439                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
1440                         return 0;
1441                 if (ret < 0)
1442                         return ret;
1443         }
1444
1445         c_nritems = btrfs_header_nritems(c);
1446         mid = (c_nritems + 1) / 2;
1447         btrfs_node_key(c, &disk_key, mid);
1448
1449         split = btrfs_alloc_free_block(trans, root, root->nodesize,
1450                                         root->root_key.objectid,
1451                                         &disk_key, level, c->start, 0);
1452         if (IS_ERR(split))
1453                 return PTR_ERR(split);
1454
1455         memset_extent_buffer(split, 0, 0, sizeof(struct btrfs_header));
1456         btrfs_set_header_level(split, btrfs_header_level(c));
1457         btrfs_set_header_bytenr(split, split->start);
1458         btrfs_set_header_generation(split, trans->transid);
1459         btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
1460         btrfs_set_header_owner(split, root->root_key.objectid);
1461         write_extent_buffer(split, root->fs_info->fsid,
1462                             (unsigned long)btrfs_header_fsid(split),
1463                             BTRFS_FSID_SIZE);
1464         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1465                             (unsigned long)btrfs_header_chunk_tree_uuid(split),
1466                             BTRFS_UUID_SIZE);
1467
1468
1469         copy_extent_buffer(split, c,
1470                            btrfs_node_key_ptr_offset(0),
1471                            btrfs_node_key_ptr_offset(mid),
1472                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1473         btrfs_set_header_nritems(split, c_nritems - mid);
1474         btrfs_set_header_nritems(c, mid);
1475         ret = 0;
1476
1477         btrfs_mark_buffer_dirty(c);
1478         btrfs_mark_buffer_dirty(split);
1479
1480         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1481                           path->slots[level + 1] + 1,
1482                           level + 1);
1483         if (wret)
1484                 ret = wret;
1485
1486         if (path->slots[level] >= mid) {
1487                 path->slots[level] -= mid;
1488                 free_extent_buffer(c);
1489                 path->nodes[level] = split;
1490                 path->slots[level + 1] += 1;
1491         } else {
1492                 free_extent_buffer(split);
1493         }
1494         return ret;
1495 }
1496
1497 /*
1498  * how many bytes are required to store the items in a leaf.  start
1499  * and nr indicate which items in the leaf to check.  This totals up the
1500  * space used both by the item structs and the item data
1501  */
1502 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1503 {
1504         int data_len;
1505         int nritems = btrfs_header_nritems(l);
1506         int end = min(nritems, start + nr) - 1;
1507
1508         if (!nr)
1509                 return 0;
1510         data_len = btrfs_item_end_nr(l, start);
1511         data_len = data_len - btrfs_item_offset_nr(l, end);
1512         data_len += sizeof(struct btrfs_item) * nr;
1513         WARN_ON(data_len < 0);
1514         return data_len;
1515 }
1516
1517 /*
1518  * The space between the end of the leaf items and
1519  * the start of the leaf data.  IOW, how much room
1520  * the leaf has left for both items and data
1521  */
1522 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1523 {
1524         int nritems = btrfs_header_nritems(leaf);
1525         int ret;
1526         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1527         if (ret < 0) {
1528                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1529                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1530                        leaf_space_used(leaf, 0, nritems), nritems);
1531         }
1532         return ret;
1533 }
1534
1535 /*
1536  * push some data in the path leaf to the right, trying to free up at
1537  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1538  *
1539  * returns 1 if the push failed because the other node didn't have enough
1540  * room, 0 if everything worked out and < 0 if there were major errors.
1541  */
1542 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1543                            *root, struct btrfs_path *path, int data_size,
1544                            int empty)
1545 {
1546         struct extent_buffer *left = path->nodes[0];
1547         struct extent_buffer *right;
1548         struct extent_buffer *upper;
1549         struct btrfs_disk_key disk_key;
1550         int slot;
1551         u32 i;
1552         int free_space;
1553         int push_space = 0;
1554         int push_items = 0;
1555         struct btrfs_item *item;
1556         u32 left_nritems;
1557         u32 nr;
1558         u32 right_nritems;
1559         u32 data_end;
1560         u32 this_item_size;
1561         int ret;
1562
1563         slot = path->slots[1];
1564         if (!path->nodes[1]) {
1565                 return 1;
1566         }
1567         upper = path->nodes[1];
1568         if (slot >= btrfs_header_nritems(upper) - 1)
1569                 return 1;
1570
1571         right = read_node_slot(root, upper, slot + 1);
1572         free_space = btrfs_leaf_free_space(root, right);
1573         if (free_space < data_size) {
1574                 free_extent_buffer(right);
1575                 return 1;
1576         }
1577
1578         /* cow and double check */
1579         ret = btrfs_cow_block(trans, root, right, upper,
1580                               slot + 1, &right);
1581         if (ret) {
1582                 free_extent_buffer(right);
1583                 return 1;
1584         }
1585         free_space = btrfs_leaf_free_space(root, right);
1586         if (free_space < data_size) {
1587                 free_extent_buffer(right);
1588                 return 1;
1589         }
1590
1591         left_nritems = btrfs_header_nritems(left);
1592         if (left_nritems == 0) {
1593                 free_extent_buffer(right);
1594                 return 1;
1595         }
1596
1597         if (empty)
1598                 nr = 0;
1599         else
1600                 nr = 1;
1601
1602         i = left_nritems - 1;
1603         while (i >= nr) {
1604                 item = btrfs_item_nr(left, i);
1605
1606                 if (path->slots[0] == i)
1607                         push_space += data_size + sizeof(*item);
1608
1609                 this_item_size = btrfs_item_size(left, item);
1610                 if (this_item_size + sizeof(*item) + push_space > free_space)
1611                         break;
1612                 push_items++;
1613                 push_space += this_item_size + sizeof(*item);
1614                 if (i == 0)
1615                         break;
1616                 i--;
1617         }
1618
1619         if (push_items == 0) {
1620                 free_extent_buffer(right);
1621                 return 1;
1622         }
1623
1624         if (!empty && push_items == left_nritems)
1625                 WARN_ON(1);
1626
1627         /* push left to right */
1628         right_nritems = btrfs_header_nritems(right);
1629
1630         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1631         push_space -= leaf_data_end(root, left);
1632
1633         /* make room in the right data area */
1634         data_end = leaf_data_end(root, right);
1635         memmove_extent_buffer(right,
1636                               btrfs_leaf_data(right) + data_end - push_space,
1637                               btrfs_leaf_data(right) + data_end,
1638                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1639
1640         /* copy from the left data area */
1641         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1642                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1643                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1644                      push_space);
1645
1646         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1647                               btrfs_item_nr_offset(0),
1648                               right_nritems * sizeof(struct btrfs_item));
1649
1650         /* copy the items from left to right */
1651         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1652                    btrfs_item_nr_offset(left_nritems - push_items),
1653                    push_items * sizeof(struct btrfs_item));
1654
1655         /* update the item pointers */
1656         right_nritems += push_items;
1657         btrfs_set_header_nritems(right, right_nritems);
1658         push_space = BTRFS_LEAF_DATA_SIZE(root);
1659         for (i = 0; i < right_nritems; i++) {
1660                 item = btrfs_item_nr(right, i);
1661                 push_space -= btrfs_item_size(right, item);
1662                 btrfs_set_item_offset(right, item, push_space);
1663         }
1664
1665         left_nritems -= push_items;
1666         btrfs_set_header_nritems(left, left_nritems);
1667
1668         if (left_nritems)
1669                 btrfs_mark_buffer_dirty(left);
1670         btrfs_mark_buffer_dirty(right);
1671
1672         btrfs_item_key(right, &disk_key, 0);
1673         btrfs_set_node_key(upper, &disk_key, slot + 1);
1674         btrfs_mark_buffer_dirty(upper);
1675
1676         /* then fixup the leaf pointer in the path */
1677         if (path->slots[0] >= left_nritems) {
1678                 path->slots[0] -= left_nritems;
1679                 free_extent_buffer(path->nodes[0]);
1680                 path->nodes[0] = right;
1681                 path->slots[1] += 1;
1682         } else {
1683                 free_extent_buffer(right);
1684         }
1685         return 0;
1686 }
1687 /*
1688  * push some data in the path leaf to the left, trying to free up at
1689  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1690  */
1691 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1692                           *root, struct btrfs_path *path, int data_size,
1693                           int empty)
1694 {
1695         struct btrfs_disk_key disk_key;
1696         struct extent_buffer *right = path->nodes[0];
1697         struct extent_buffer *left;
1698         int slot;
1699         int i;
1700         int free_space;
1701         int push_space = 0;
1702         int push_items = 0;
1703         struct btrfs_item *item;
1704         u32 old_left_nritems;
1705         u32 right_nritems;
1706         u32 nr;
1707         int ret = 0;
1708         int wret;
1709         u32 this_item_size;
1710         u32 old_left_item_size;
1711
1712         slot = path->slots[1];
1713         if (slot == 0)
1714                 return 1;
1715         if (!path->nodes[1])
1716                 return 1;
1717
1718         right_nritems = btrfs_header_nritems(right);
1719         if (right_nritems == 0) {
1720                 return 1;
1721         }
1722
1723         left = read_node_slot(root, path->nodes[1], slot - 1);
1724         free_space = btrfs_leaf_free_space(root, left);
1725         if (free_space < data_size) {
1726                 free_extent_buffer(left);
1727                 return 1;
1728         }
1729
1730         /* cow and double check */
1731         ret = btrfs_cow_block(trans, root, left,
1732                               path->nodes[1], slot - 1, &left);
1733         if (ret) {
1734                 /* we hit -ENOSPC, but it isn't fatal here */
1735                 free_extent_buffer(left);
1736                 return 1;
1737         }
1738
1739         free_space = btrfs_leaf_free_space(root, left);
1740         if (free_space < data_size) {
1741                 free_extent_buffer(left);
1742                 return 1;
1743         }
1744
1745         if (empty)
1746                 nr = right_nritems;
1747         else
1748                 nr = right_nritems - 1;
1749
1750         for (i = 0; i < nr; i++) {
1751                 item = btrfs_item_nr(right, i);
1752
1753                 if (path->slots[0] == i)
1754                         push_space += data_size + sizeof(*item);
1755
1756                 this_item_size = btrfs_item_size(right, item);
1757                 if (this_item_size + sizeof(*item) + push_space > free_space)
1758                         break;
1759
1760                 push_items++;
1761                 push_space += this_item_size + sizeof(*item);
1762         }
1763
1764         if (push_items == 0) {
1765                 free_extent_buffer(left);
1766                 return 1;
1767         }
1768         if (!empty && push_items == btrfs_header_nritems(right))
1769                 WARN_ON(1);
1770
1771         /* push data from right to left */
1772         copy_extent_buffer(left, right,
1773                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1774                            btrfs_item_nr_offset(0),
1775                            push_items * sizeof(struct btrfs_item));
1776
1777         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1778                      btrfs_item_offset_nr(right, push_items -1);
1779
1780         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1781                      leaf_data_end(root, left) - push_space,
1782                      btrfs_leaf_data(right) +
1783                      btrfs_item_offset_nr(right, push_items - 1),
1784                      push_space);
1785         old_left_nritems = btrfs_header_nritems(left);
1786         BUG_ON(old_left_nritems == 0);
1787
1788         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1789         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1790                 u32 ioff;
1791
1792                 item = btrfs_item_nr(left, i);
1793                 ioff = btrfs_item_offset(left, item);
1794                 btrfs_set_item_offset(left, item,
1795                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1796         }
1797         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1798
1799         /* fixup right node */
1800         if (push_items > right_nritems) {
1801                 printk("push items %d nr %u\n", push_items, right_nritems);
1802                 WARN_ON(1);
1803         }
1804
1805         if (push_items < right_nritems) {
1806                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1807                                                   leaf_data_end(root, right);
1808                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
1809                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1810                                       btrfs_leaf_data(right) +
1811                                       leaf_data_end(root, right), push_space);
1812
1813                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1814                               btrfs_item_nr_offset(push_items),
1815                              (btrfs_header_nritems(right) - push_items) *
1816                              sizeof(struct btrfs_item));
1817         }
1818         right_nritems -= push_items;
1819         btrfs_set_header_nritems(right, right_nritems);
1820         push_space = BTRFS_LEAF_DATA_SIZE(root);
1821         for (i = 0; i < right_nritems; i++) {
1822                 item = btrfs_item_nr(right, i);
1823                 push_space = push_space - btrfs_item_size(right, item);
1824                 btrfs_set_item_offset(right, item, push_space);
1825         }
1826
1827         btrfs_mark_buffer_dirty(left);
1828         if (right_nritems)
1829                 btrfs_mark_buffer_dirty(right);
1830
1831         btrfs_item_key(right, &disk_key, 0);
1832         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1833         if (wret)
1834                 ret = wret;
1835
1836         /* then fixup the leaf pointer in the path */
1837         if (path->slots[0] < push_items) {
1838                 path->slots[0] += old_left_nritems;
1839                 free_extent_buffer(path->nodes[0]);
1840                 path->nodes[0] = left;
1841                 path->slots[1] -= 1;
1842         } else {
1843                 free_extent_buffer(left);
1844                 path->slots[0] -= push_items;
1845         }
1846         BUG_ON(path->slots[0] < 0);
1847         return ret;
1848 }
1849
1850 /*
1851  * split the path's leaf in two, making sure there is at least data_size
1852  * available for the resulting leaf level of the path.
1853  *
1854  * returns 0 if all went well and < 0 on failure.
1855  */
1856 static noinline int copy_for_split(struct btrfs_trans_handle *trans,
1857                                struct btrfs_root *root,
1858                                struct btrfs_path *path,
1859                                struct extent_buffer *l,
1860                                struct extent_buffer *right,
1861                                int slot, int mid, int nritems)
1862 {
1863         int data_copy_size;
1864         int rt_data_off;
1865         int i;
1866         int ret = 0;
1867         int wret;
1868         struct btrfs_disk_key disk_key;
1869
1870         nritems = nritems - mid;
1871         btrfs_set_header_nritems(right, nritems);
1872         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
1873
1874         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
1875                            btrfs_item_nr_offset(mid),
1876                            nritems * sizeof(struct btrfs_item));
1877
1878         copy_extent_buffer(right, l,
1879                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
1880                      data_copy_size, btrfs_leaf_data(l) +
1881                      leaf_data_end(root, l), data_copy_size);
1882
1883         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
1884                       btrfs_item_end_nr(l, mid);
1885
1886         for (i = 0; i < nritems; i++) {
1887                 struct btrfs_item *item = btrfs_item_nr(right, i);
1888                 u32 ioff = btrfs_item_offset(right, item);
1889                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
1890         }
1891
1892         btrfs_set_header_nritems(l, mid);
1893         ret = 0;
1894         btrfs_item_key(right, &disk_key, 0);
1895         wret = insert_ptr(trans, root, path, &disk_key, right->start,
1896                           path->slots[1] + 1, 1);
1897         if (wret)
1898                 ret = wret;
1899
1900         btrfs_mark_buffer_dirty(right);
1901         btrfs_mark_buffer_dirty(l);
1902         BUG_ON(path->slots[0] != slot);
1903
1904         if (mid <= slot) {
1905                 free_extent_buffer(path->nodes[0]);
1906                 path->nodes[0] = right;
1907                 path->slots[0] -= mid;
1908                 path->slots[1] += 1;
1909         } else {
1910                 free_extent_buffer(right);
1911         }
1912
1913         BUG_ON(path->slots[0] < 0);
1914
1915         return ret;
1916 }
1917
1918 /*
1919  * split the path's leaf in two, making sure there is at least data_size
1920  * available for the resulting leaf level of the path.
1921  *
1922  * returns 0 if all went well and < 0 on failure.
1923  */
1924 static noinline int split_leaf(struct btrfs_trans_handle *trans,
1925                                struct btrfs_root *root,
1926                                struct btrfs_key *ins_key,
1927                                struct btrfs_path *path, int data_size,
1928                                int extend)
1929 {
1930         struct btrfs_disk_key disk_key;
1931         struct extent_buffer *l;
1932         u32 nritems;
1933         int mid;
1934         int slot;
1935         struct extent_buffer *right;
1936         int ret = 0;
1937         int wret;
1938         int split;
1939         int num_doubles = 0;
1940
1941         /* first try to make some room by pushing left and right */
1942         if (data_size && ins_key->type != BTRFS_DIR_ITEM_KEY) {
1943                 wret = push_leaf_right(trans, root, path, data_size, 0);
1944                 if (wret < 0)
1945                         return wret;
1946                 if (wret) {
1947                         wret = push_leaf_left(trans, root, path, data_size, 0);
1948                         if (wret < 0)
1949                                 return wret;
1950                 }
1951                 l = path->nodes[0];
1952
1953                 /* did the pushes work? */
1954                 if (btrfs_leaf_free_space(root, l) >= data_size)
1955                         return 0;
1956         }
1957
1958         if (!path->nodes[1]) {
1959                 ret = insert_new_root(trans, root, path, 1);
1960                 if (ret)
1961                         return ret;
1962         }
1963 again:
1964         split = 1;
1965         l = path->nodes[0];
1966         slot = path->slots[0];
1967         nritems = btrfs_header_nritems(l);
1968         mid = (nritems + 1) / 2;
1969
1970         if (mid <= slot) {
1971                 if (nritems == 1 ||
1972                     leaf_space_used(l, mid, nritems - mid) + data_size >
1973                         BTRFS_LEAF_DATA_SIZE(root)) {
1974                         if (slot >= nritems) {
1975                                 split = 0;
1976                         } else {
1977                                 mid = slot;
1978                                 if (mid != nritems &&
1979                                     leaf_space_used(l, mid, nritems - mid) +
1980                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
1981                                         split = 2;
1982                                 }
1983                         }
1984                 }
1985         } else {
1986                 if (leaf_space_used(l, 0, mid) + data_size >
1987                         BTRFS_LEAF_DATA_SIZE(root)) {
1988                         if (!extend && data_size && slot == 0) {
1989                                 split = 0;
1990                         } else if ((extend || !data_size) && slot == 0) {
1991                                 mid = 1;
1992                         } else {
1993                                 mid = slot;
1994                                 if (mid != nritems &&
1995                                     leaf_space_used(l, mid, nritems - mid) +
1996                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
1997                                         split = 2 ;
1998                                 }
1999                         }
2000                 }
2001         }
2002
2003         if (split == 0)
2004                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2005         else
2006                 btrfs_item_key(l, &disk_key, mid);
2007
2008         right = btrfs_alloc_free_block(trans, root, root->leafsize,
2009                                         root->root_key.objectid,
2010                                         &disk_key, 0, l->start, 0);
2011         if (IS_ERR(right)) {
2012                 BUG_ON(1);
2013                 return PTR_ERR(right);
2014         }
2015
2016         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2017         btrfs_set_header_bytenr(right, right->start);
2018         btrfs_set_header_generation(right, trans->transid);
2019         btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
2020         btrfs_set_header_owner(right, root->root_key.objectid);
2021         btrfs_set_header_level(right, 0);
2022         write_extent_buffer(right, root->fs_info->fsid,
2023                             (unsigned long)btrfs_header_fsid(right),
2024                             BTRFS_FSID_SIZE);
2025
2026         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2027                             (unsigned long)btrfs_header_chunk_tree_uuid(right),
2028                             BTRFS_UUID_SIZE);
2029
2030         if (split == 0) {
2031                 if (mid <= slot) {
2032                         btrfs_set_header_nritems(right, 0);
2033                         wret = insert_ptr(trans, root, path,
2034                                           &disk_key, right->start,
2035                                           path->slots[1] + 1, 1);
2036                         if (wret)
2037                                 ret = wret;
2038
2039                         free_extent_buffer(path->nodes[0]);
2040                         path->nodes[0] = right;
2041                         path->slots[0] = 0;
2042                         path->slots[1] += 1;
2043                 } else {
2044                         btrfs_set_header_nritems(right, 0);
2045                         wret = insert_ptr(trans, root, path,
2046                                           &disk_key,
2047                                           right->start,
2048                                           path->slots[1], 1);
2049                         if (wret)
2050                                 ret = wret;
2051                         free_extent_buffer(path->nodes[0]);
2052                         path->nodes[0] = right;
2053                         path->slots[0] = 0;
2054                         if (path->slots[1] == 0) {
2055                                 wret = fixup_low_keys(trans, root,
2056                                                 path, &disk_key, 1);
2057                                 if (wret)
2058                                         ret = wret;
2059                         }
2060                 }
2061                 btrfs_mark_buffer_dirty(right);
2062                 return ret;
2063         }
2064
2065         ret = copy_for_split(trans, root, path, l, right, slot, mid, nritems);
2066         BUG_ON(ret);
2067
2068         if (split == 2) {
2069                 BUG_ON(num_doubles != 0);
2070                 num_doubles++;
2071                 goto again;
2072         }
2073
2074         return ret;
2075 }
2076
2077 /*
2078  * This function splits a single item into two items,
2079  * giving 'new_key' to the new item and splitting the
2080  * old one at split_offset (from the start of the item).
2081  *
2082  * The path may be released by this operation.  After
2083  * the split, the path is pointing to the old item.  The
2084  * new item is going to be in the same node as the old one.
2085  *
2086  * Note, the item being split must be smaller enough to live alone on
2087  * a tree block with room for one extra struct btrfs_item
2088  *
2089  * This allows us to split the item in place, keeping a lock on the
2090  * leaf the entire time.
2091  */
2092 int btrfs_split_item(struct btrfs_trans_handle *trans,
2093                      struct btrfs_root *root,
2094                      struct btrfs_path *path,
2095                      struct btrfs_key *new_key,
2096                      unsigned long split_offset)
2097 {
2098         u32 item_size;
2099         struct extent_buffer *leaf;
2100         struct btrfs_key orig_key;
2101         struct btrfs_item *item;
2102         struct btrfs_item *new_item;
2103         int ret = 0;
2104         int slot;
2105         u32 nritems;
2106         u32 orig_offset;
2107         struct btrfs_disk_key disk_key;
2108         char *buf;
2109
2110         leaf = path->nodes[0];
2111         btrfs_item_key_to_cpu(leaf, &orig_key, path->slots[0]);
2112         if (btrfs_leaf_free_space(root, leaf) >= sizeof(struct btrfs_item))
2113                 goto split;
2114
2115         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2116         btrfs_release_path(path);
2117
2118         path->search_for_split = 1;
2119
2120         ret = btrfs_search_slot(trans, root, &orig_key, path, 0, 1);
2121         path->search_for_split = 0;
2122
2123         /* if our item isn't there or got smaller, return now */
2124         if (ret != 0 || item_size != btrfs_item_size_nr(path->nodes[0],
2125                                                         path->slots[0])) {
2126                 return -EAGAIN;
2127         }
2128
2129         ret = split_leaf(trans, root, &orig_key, path, 0, 0);
2130         BUG_ON(ret);
2131
2132         BUG_ON(btrfs_leaf_free_space(root, leaf) < sizeof(struct btrfs_item));
2133         leaf = path->nodes[0];
2134
2135 split:
2136         item = btrfs_item_nr(leaf, path->slots[0]);
2137         orig_offset = btrfs_item_offset(leaf, item);
2138         item_size = btrfs_item_size(leaf, item);
2139
2140
2141         buf = kmalloc(item_size, GFP_NOFS);
2142         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
2143                             path->slots[0]), item_size);
2144         slot = path->slots[0] + 1;
2145         leaf = path->nodes[0];
2146
2147         nritems = btrfs_header_nritems(leaf);
2148
2149         if (slot != nritems) {
2150                 /* shift the items */
2151                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
2152                               btrfs_item_nr_offset(slot),
2153                               (nritems - slot) * sizeof(struct btrfs_item));
2154
2155         }
2156
2157         btrfs_cpu_key_to_disk(&disk_key, new_key);
2158         btrfs_set_item_key(leaf, &disk_key, slot);
2159
2160         new_item = btrfs_item_nr(leaf, slot);
2161
2162         btrfs_set_item_offset(leaf, new_item, orig_offset);
2163         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
2164
2165         btrfs_set_item_offset(leaf, item,
2166                               orig_offset + item_size - split_offset);
2167         btrfs_set_item_size(leaf, item, split_offset);
2168
2169         btrfs_set_header_nritems(leaf, nritems + 1);
2170
2171         /* write the data for the start of the original item */
2172         write_extent_buffer(leaf, buf,
2173                             btrfs_item_ptr_offset(leaf, path->slots[0]),
2174                             split_offset);
2175
2176         /* write the data for the new item */
2177         write_extent_buffer(leaf, buf + split_offset,
2178                             btrfs_item_ptr_offset(leaf, slot),
2179                             item_size - split_offset);
2180         btrfs_mark_buffer_dirty(leaf);
2181
2182         ret = 0;
2183         if (btrfs_leaf_free_space(root, leaf) < 0) {
2184                 btrfs_print_leaf(root, leaf);
2185                 BUG();
2186         }
2187         kfree(buf);
2188         return ret;
2189 }
2190
2191 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2192                         struct btrfs_root *root,
2193                         struct btrfs_path *path,
2194                         u32 new_size, int from_end)
2195 {
2196         int ret = 0;
2197         int slot;
2198         struct extent_buffer *leaf;
2199         struct btrfs_item *item;
2200         u32 nritems;
2201         unsigned int data_end;
2202         unsigned int old_data_start;
2203         unsigned int old_size;
2204         unsigned int size_diff;
2205         int i;
2206
2207         leaf = path->nodes[0];
2208         slot = path->slots[0];
2209
2210         old_size = btrfs_item_size_nr(leaf, slot);
2211         if (old_size == new_size)
2212                 return 0;
2213
2214         nritems = btrfs_header_nritems(leaf);
2215         data_end = leaf_data_end(root, leaf);
2216
2217         old_data_start = btrfs_item_offset_nr(leaf, slot);
2218
2219         size_diff = old_size - new_size;
2220
2221         BUG_ON(slot < 0);
2222         BUG_ON(slot >= nritems);
2223
2224         /*
2225          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2226          */
2227         /* first correct the data pointers */
2228         for (i = slot; i < nritems; i++) {
2229                 u32 ioff;
2230                 item = btrfs_item_nr(leaf, i);
2231                 ioff = btrfs_item_offset(leaf, item);
2232                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2233         }
2234
2235         /* shift the data */
2236         if (from_end) {
2237                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2238                               data_end + size_diff, btrfs_leaf_data(leaf) +
2239                               data_end, old_data_start + new_size - data_end);
2240         } else {
2241                 struct btrfs_disk_key disk_key;
2242                 u64 offset;
2243
2244                 btrfs_item_key(leaf, &disk_key, slot);
2245
2246                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2247                         unsigned long ptr;
2248                         struct btrfs_file_extent_item *fi;
2249
2250                         fi = btrfs_item_ptr(leaf, slot,
2251                                             struct btrfs_file_extent_item);
2252                         fi = (struct btrfs_file_extent_item *)(
2253                              (unsigned long)fi - size_diff);
2254
2255                         if (btrfs_file_extent_type(leaf, fi) ==
2256                             BTRFS_FILE_EXTENT_INLINE) {
2257                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2258                                 memmove_extent_buffer(leaf, ptr,
2259                                         (unsigned long)fi,
2260                                         offsetof(struct btrfs_file_extent_item,
2261                                                  disk_bytenr));
2262                         }
2263                 }
2264
2265                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2266                               data_end + size_diff, btrfs_leaf_data(leaf) +
2267                               data_end, old_data_start - data_end);
2268
2269                 offset = btrfs_disk_key_offset(&disk_key);
2270                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2271                 btrfs_set_item_key(leaf, &disk_key, slot);
2272                 if (slot == 0)
2273                         fixup_low_keys(trans, root, path, &disk_key, 1);
2274         }
2275
2276         item = btrfs_item_nr(leaf, slot);
2277         btrfs_set_item_size(leaf, item, new_size);
2278         btrfs_mark_buffer_dirty(leaf);
2279
2280         ret = 0;
2281         if (btrfs_leaf_free_space(root, leaf) < 0) {
2282                 btrfs_print_leaf(root, leaf);
2283                 BUG();
2284         }
2285         return ret;
2286 }
2287
2288 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2289                       struct btrfs_root *root, struct btrfs_path *path,
2290                       u32 data_size)
2291 {
2292         int ret = 0;
2293         int slot;
2294         struct extent_buffer *leaf;
2295         struct btrfs_item *item;
2296         u32 nritems;
2297         unsigned int data_end;
2298         unsigned int old_data;
2299         unsigned int old_size;
2300         int i;
2301
2302         leaf = path->nodes[0];
2303
2304         nritems = btrfs_header_nritems(leaf);
2305         data_end = leaf_data_end(root, leaf);
2306
2307         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2308                 btrfs_print_leaf(root, leaf);
2309                 BUG();
2310         }
2311         slot = path->slots[0];
2312         old_data = btrfs_item_end_nr(leaf, slot);
2313
2314         BUG_ON(slot < 0);
2315         if (slot >= nritems) {
2316                 btrfs_print_leaf(root, leaf);
2317                 printk("slot %d too large, nritems %d\n", slot, nritems);
2318                 BUG_ON(1);
2319         }
2320
2321         /*
2322          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2323          */
2324         /* first correct the data pointers */
2325         for (i = slot; i < nritems; i++) {
2326                 u32 ioff;
2327                 item = btrfs_item_nr(leaf, i);
2328                 ioff = btrfs_item_offset(leaf, item);
2329                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2330         }
2331
2332         /* shift the data */
2333         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2334                       data_end - data_size, btrfs_leaf_data(leaf) +
2335                       data_end, old_data - data_end);
2336
2337         data_end = old_data;
2338         old_size = btrfs_item_size_nr(leaf, slot);
2339         item = btrfs_item_nr(leaf, slot);
2340         btrfs_set_item_size(leaf, item, old_size + data_size);
2341         btrfs_mark_buffer_dirty(leaf);
2342
2343         ret = 0;
2344         if (btrfs_leaf_free_space(root, leaf) < 0) {
2345                 btrfs_print_leaf(root, leaf);
2346                 BUG();
2347         }
2348         return ret;
2349 }
2350
2351 /*
2352  * Given a key and some data, insert an item into the tree.
2353  * This does all the path init required, making room in the tree if needed.
2354  */
2355 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2356                             struct btrfs_root *root,
2357                             struct btrfs_path *path,
2358                             struct btrfs_key *cpu_key, u32 *data_size,
2359                             int nr)
2360 {
2361         struct extent_buffer *leaf;
2362         struct btrfs_item *item;
2363         int ret = 0;
2364         int slot;
2365         int i;
2366         u32 nritems;
2367         u32 total_size = 0;
2368         u32 total_data = 0;
2369         unsigned int data_end;
2370         struct btrfs_disk_key disk_key;
2371
2372         for (i = 0; i < nr; i++) {
2373                 total_data += data_size[i];
2374         }
2375
2376         /* create a root if there isn't one */
2377         if (!root->node)
2378                 BUG();
2379
2380         total_size = total_data + nr * sizeof(struct btrfs_item);
2381         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2382         if (ret == 0) {
2383                 return -EEXIST;
2384         }
2385         if (ret < 0)
2386                 goto out;
2387
2388         leaf = path->nodes[0];
2389
2390         nritems = btrfs_header_nritems(leaf);
2391         data_end = leaf_data_end(root, leaf);
2392
2393         if (btrfs_leaf_free_space(root, leaf) < total_size) {
2394                 btrfs_print_leaf(root, leaf);
2395                 printk("not enough freespace need %u have %d\n",
2396                        total_size, btrfs_leaf_free_space(root, leaf));
2397                 BUG();
2398         }
2399
2400         slot = path->slots[0];
2401         BUG_ON(slot < 0);
2402
2403         if (slot != nritems) {
2404                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2405
2406                 if (old_data < data_end) {
2407                         btrfs_print_leaf(root, leaf);
2408                         printk("slot %d old_data %d data_end %d\n",
2409                                slot, old_data, data_end);
2410                         BUG_ON(1);
2411                 }
2412                 /*
2413                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2414                  */
2415                 /* first correct the data pointers */
2416                 for (i = slot; i < nritems; i++) {
2417                         u32 ioff;
2418
2419                         item = btrfs_item_nr(leaf, i);
2420                         ioff = btrfs_item_offset(leaf, item);
2421                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2422                 }
2423
2424                 /* shift the items */
2425                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2426                               btrfs_item_nr_offset(slot),
2427                               (nritems - slot) * sizeof(struct btrfs_item));
2428
2429                 /* shift the data */
2430                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2431                               data_end - total_data, btrfs_leaf_data(leaf) +
2432                               data_end, old_data - data_end);
2433                 data_end = old_data;
2434         }
2435
2436         /* setup the item for the new data */
2437         for (i = 0; i < nr; i++) {
2438                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2439                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2440                 item = btrfs_item_nr(leaf, slot + i);
2441                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2442                 data_end -= data_size[i];
2443                 btrfs_set_item_size(leaf, item, data_size[i]);
2444         }
2445         btrfs_set_header_nritems(leaf, nritems + nr);
2446         btrfs_mark_buffer_dirty(leaf);
2447
2448         ret = 0;
2449         if (slot == 0) {
2450                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2451                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2452         }
2453
2454         if (btrfs_leaf_free_space(root, leaf) < 0) {
2455                 btrfs_print_leaf(root, leaf);
2456                 BUG();
2457         }
2458
2459 out:
2460         return ret;
2461 }
2462
2463 /*
2464  * Given a key and some data, insert an item into the tree.
2465  * This does all the path init required, making room in the tree if needed.
2466  */
2467 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2468                       *root, struct btrfs_key *cpu_key, void *data, u32
2469                       data_size)
2470 {
2471         int ret = 0;
2472         struct btrfs_path *path;
2473         struct extent_buffer *leaf;
2474         unsigned long ptr;
2475
2476         path = btrfs_alloc_path();
2477         BUG_ON(!path);
2478         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2479         if (!ret) {
2480                 leaf = path->nodes[0];
2481                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2482                 write_extent_buffer(leaf, data, ptr, data_size);
2483                 btrfs_mark_buffer_dirty(leaf);
2484         }
2485         btrfs_free_path(path);
2486         return ret;
2487 }
2488
2489 /*
2490  * delete the pointer from a given node.
2491  *
2492  * If the delete empties a node, the node is removed from the tree,
2493  * continuing all the way the root if required.  The root is converted into
2494  * a leaf if all the nodes are emptied.
2495  */
2496 int btrfs_del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2497                    struct btrfs_path *path, int level, int slot)
2498 {
2499         struct extent_buffer *parent = path->nodes[level];
2500         u32 nritems;
2501         int ret = 0;
2502         int wret;
2503
2504         nritems = btrfs_header_nritems(parent);
2505         if (slot != nritems -1) {
2506                 memmove_extent_buffer(parent,
2507                               btrfs_node_key_ptr_offset(slot),
2508                               btrfs_node_key_ptr_offset(slot + 1),
2509                               sizeof(struct btrfs_key_ptr) *
2510                               (nritems - slot - 1));
2511         }
2512         nritems--;
2513         btrfs_set_header_nritems(parent, nritems);
2514         if (nritems == 0 && parent == root->node) {
2515                 BUG_ON(btrfs_header_level(root->node) != 1);
2516                 /* just turn the root into a leaf and break */
2517                 btrfs_set_header_level(root->node, 0);
2518         } else if (slot == 0) {
2519                 struct btrfs_disk_key disk_key;
2520
2521                 btrfs_node_key(parent, &disk_key, 0);
2522                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2523                 if (wret)
2524                         ret = wret;
2525         }
2526         btrfs_mark_buffer_dirty(parent);
2527         return ret;
2528 }
2529
2530 /*
2531  * a helper function to delete the leaf pointed to by path->slots[1] and
2532  * path->nodes[1].
2533  *
2534  * This deletes the pointer in path->nodes[1] and frees the leaf
2535  * block extent.  zero is returned if it all worked out, < 0 otherwise.
2536  *
2537  * The path must have already been setup for deleting the leaf, including
2538  * all the proper balancing.  path->nodes[1] must be locked.
2539  */
2540 static noinline int btrfs_del_leaf(struct btrfs_trans_handle *trans,
2541                                    struct btrfs_root *root,
2542                                    struct btrfs_path *path,
2543                                    struct extent_buffer *leaf)
2544 {
2545         int ret;
2546
2547         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
2548         ret = btrfs_del_ptr(trans, root, path, 1, path->slots[1]);
2549         if (ret)
2550                 return ret;
2551
2552         ret = btrfs_free_extent(trans, root, leaf->start, leaf->len,
2553                                 0, root->root_key.objectid, 0, 0);
2554         return ret;
2555 }
2556
2557 /*
2558  * delete the item at the leaf level in path.  If that empties
2559  * the leaf, remove it from the tree
2560  */
2561 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2562                     struct btrfs_path *path, int slot, int nr)
2563 {
2564         struct extent_buffer *leaf;
2565         struct btrfs_item *item;
2566         int last_off;
2567         int dsize = 0;
2568         int ret = 0;
2569         int wret;
2570         int i;
2571         u32 nritems;
2572
2573         leaf = path->nodes[0];
2574         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2575
2576         for (i = 0; i < nr; i++)
2577                 dsize += btrfs_item_size_nr(leaf, slot + i);
2578
2579         nritems = btrfs_header_nritems(leaf);
2580
2581         if (slot + nr != nritems) {
2582                 int data_end = leaf_data_end(root, leaf);
2583
2584                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2585                               data_end + dsize,
2586                               btrfs_leaf_data(leaf) + data_end,
2587                               last_off - data_end);
2588
2589                 for (i = slot + nr; i < nritems; i++) {
2590                         u32 ioff;
2591
2592                         item = btrfs_item_nr(leaf, i);
2593                         ioff = btrfs_item_offset(leaf, item);
2594                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2595                 }
2596
2597                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2598                               btrfs_item_nr_offset(slot + nr),
2599                               sizeof(struct btrfs_item) *
2600                               (nritems - slot - nr));
2601         }
2602         btrfs_set_header_nritems(leaf, nritems - nr);
2603         nritems -= nr;
2604
2605         /* delete the leaf if we've emptied it */
2606         if (nritems == 0) {
2607                 if (leaf == root->node) {
2608                         btrfs_set_header_level(leaf, 0);
2609                 } else {
2610                         clean_tree_block(trans, root, leaf);
2611                         wait_on_tree_block_writeback(root, leaf);
2612
2613                         wret = btrfs_del_leaf(trans, root, path, leaf);
2614                         BUG_ON(ret);
2615                         if (wret)
2616                                 ret = wret;
2617                 }
2618         } else {
2619                 int used = leaf_space_used(leaf, 0, nritems);
2620                 if (slot == 0) {
2621                         struct btrfs_disk_key disk_key;
2622
2623                         btrfs_item_key(leaf, &disk_key, 0);
2624                         wret = fixup_low_keys(trans, root, path,
2625                                               &disk_key, 1);
2626                         if (wret)
2627                                 ret = wret;
2628                 }
2629
2630                 /* delete the leaf if it is mostly empty */
2631                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2632                         /* push_leaf_left fixes the path.
2633                          * make sure the path still points to our leaf
2634                          * for possible call to del_ptr below
2635                          */
2636                         slot = path->slots[1];
2637                         extent_buffer_get(leaf);
2638
2639                         wret = push_leaf_left(trans, root, path, 1, 1);
2640                         if (wret < 0 && wret != -ENOSPC)
2641                                 ret = wret;
2642
2643                         if (path->nodes[0] == leaf &&
2644                             btrfs_header_nritems(leaf)) {
2645                                 wret = push_leaf_right(trans, root, path, 1, 1);
2646                                 if (wret < 0 && wret != -ENOSPC)
2647                                         ret = wret;
2648                         }
2649
2650                         if (btrfs_header_nritems(leaf) == 0) {
2651                                 clean_tree_block(trans, root, leaf);
2652                                 wait_on_tree_block_writeback(root, leaf);
2653
2654                                 path->slots[1] = slot;
2655                                 ret = btrfs_del_leaf(trans, root, path, leaf);
2656                                 BUG_ON(ret);
2657                                 free_extent_buffer(leaf);
2658
2659                         } else {
2660                                 btrfs_mark_buffer_dirty(leaf);
2661                                 free_extent_buffer(leaf);
2662                         }
2663                 } else {
2664                         btrfs_mark_buffer_dirty(leaf);
2665                 }
2666         }
2667         return ret;
2668 }
2669
2670 /*
2671  * walk up the tree as far as required to find the previous leaf.
2672  * returns 0 if it found something or 1 if there are no lesser leaves.
2673  * returns < 0 on io errors.
2674  */
2675 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2676 {
2677         int slot;
2678         int level = 1;
2679         struct extent_buffer *c;
2680         struct extent_buffer *next = NULL;
2681
2682         while(level < BTRFS_MAX_LEVEL) {
2683                 if (!path->nodes[level])
2684                         return 1;
2685
2686                 slot = path->slots[level];
2687                 c = path->nodes[level];
2688                 if (slot == 0) {
2689                         level++;
2690                         if (level == BTRFS_MAX_LEVEL)
2691                                 return 1;
2692                         continue;
2693                 }
2694                 slot--;
2695
2696                 next = read_node_slot(root, c, slot);
2697                 break;
2698         }
2699         path->slots[level] = slot;
2700         while(1) {
2701                 level--;
2702                 c = path->nodes[level];
2703                 free_extent_buffer(c);
2704                 slot = btrfs_header_nritems(next);
2705                 if (slot != 0)
2706                         slot--;
2707                 path->nodes[level] = next;
2708                 path->slots[level] = slot;
2709                 if (!level)
2710                         break;
2711                 next = read_node_slot(root, next, slot);
2712         }
2713         return 0;
2714 }
2715
2716 /*
2717  * walk up the tree as far as required to find the next leaf.
2718  * returns 0 if it found something or 1 if there are no greater leaves.
2719  * returns < 0 on io errors.
2720  */
2721 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2722 {
2723         int slot;
2724         int level = 1;
2725         struct extent_buffer *c;
2726         struct extent_buffer *next = NULL;
2727
2728         while(level < BTRFS_MAX_LEVEL) {
2729                 if (!path->nodes[level])
2730                         return 1;
2731
2732                 slot = path->slots[level] + 1;
2733                 c = path->nodes[level];
2734                 if (slot >= btrfs_header_nritems(c)) {
2735                         level++;
2736                         if (level == BTRFS_MAX_LEVEL)
2737                                 return 1;
2738                         continue;
2739                 }
2740
2741                 if (path->reada)
2742                         reada_for_search(root, path, level, slot, 0);
2743
2744                 next = read_node_slot(root, c, slot);
2745                 if (!next)
2746                         return -EIO;
2747                 break;
2748         }
2749         path->slots[level] = slot;
2750         while(1) {
2751                 level--;
2752                 c = path->nodes[level];
2753                 free_extent_buffer(c);
2754                 path->nodes[level] = next;
2755                 path->slots[level] = 0;
2756                 if (!level)
2757                         break;
2758                 if (path->reada)
2759                         reada_for_search(root, path, level, 0, 0);
2760                 next = read_node_slot(root, next, 0);
2761                 if (!next)
2762                         return -EIO;
2763         }
2764         return 0;
2765 }
2766
2767 int btrfs_previous_item(struct btrfs_root *root,
2768                         struct btrfs_path *path, u64 min_objectid,
2769                         int type)
2770 {
2771         struct btrfs_key found_key;
2772         struct extent_buffer *leaf;
2773         int ret;
2774
2775         while(1) {
2776                 if (path->slots[0] == 0) {
2777                         ret = btrfs_prev_leaf(root, path);
2778                         if (ret != 0)
2779                                 return ret;
2780                 } else {
2781                         path->slots[0]--;
2782                 }
2783                 leaf = path->nodes[0];
2784                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2785                 if (found_key.type == type)
2786                         return 0;
2787         }
2788         return 1;
2789 }
2790