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