SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/btrfs/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
  19 #include <linux/highmem.h>
  20 #include "ctree.h"
  21 #include "disk-io.h"
  22 #include "transaction.h"
  23 #include "print-tree.h"
  24
  25 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  26                       *root, struct btrfs_path *path, int level);
  27 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  28                       *root, struct btrfs_key *ins_key,
  29                       struct btrfs_path *path, int data_size);
  30 static int push_node_left(struct btrfs_trans_handle *trans,
  31                           struct btrfs_root *root, struct extent_buffer *dst,
  32                           struct extent_buffer *src);
  33 static int balance_node_right(struct btrfs_trans_handle *trans,
  34                               struct btrfs_root *root,
  35                               struct extent_buffer *dst_buf,
  36                               struct extent_buffer *src_buf);
  37 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  38                    struct btrfs_path *path, int level, int slot);
  39
  40 inline void btrfs_init_path(struct btrfs_path *p)
  41 {
  42         memset(p, 0, sizeof(*p));
  43 }
  44
  45 struct btrfs_path *btrfs_alloc_path(void)
  46 {
  47         struct btrfs_path *path;
  48         path = kmem_cache_alloc(btrfs_path_cachep, GFP_NOFS);
  49         if (path) {
  50                 btrfs_init_path(path);
  51                 path->reada = 1;
  52         }
  53         return path;
  54 }
  55
  56 void btrfs_free_path(struct btrfs_path *p)
  57 {
  58         btrfs_release_path(NULL, p);
  59         kmem_cache_free(btrfs_path_cachep, p);
  60 }
  61
  62 void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
  63 {
  64         int i;
  65         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  66                 if (!p->nodes[i])
  67                         break;
  68                 free_extent_buffer(p->nodes[i]);
  69         }
  70         memset(p, 0, sizeof(*p));
  71 }
  72
  73 static int __btrfs_cow_block(struct btrfs_trans_handle *trans,
  74                              struct btrfs_root *root,
  75                              struct extent_buffer *buf,
  76                              struct extent_buffer *parent, int parent_slot,
  77                              struct extent_buffer **cow_ret,
  78                              u64 search_start, u64 empty_size)
  79 {
  80         struct extent_buffer *cow;
  81         int ret = 0;
  82         int different_trans = 0;
  83
  84         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
  85
  86         cow = btrfs_alloc_free_block(trans, root, buf->len,
  87                                      search_start, empty_size);
  88         if (IS_ERR(cow))
  89                 return PTR_ERR(cow);
  90
  91         cow->alloc_addr = (unsigned long)__builtin_return_address(0);
  92
  93         copy_extent_buffer(cow, buf, 0, 0, cow->len);
  94         btrfs_set_header_bytenr(cow, cow->start);
  95         btrfs_set_header_generation(cow, trans->transid);
  96         btrfs_set_header_owner(cow, root->root_key.objectid);
  97
  98         WARN_ON(btrfs_header_generation(buf) > trans->transid);
  99         if (btrfs_header_generation(buf) != trans->transid) {
 100                 different_trans = 1;
 101                 ret = btrfs_inc_ref(trans, root, buf);
 102                 if (ret)
 103                         return ret;
 104         } else {
 105                 clean_tree_block(trans, root, buf);
 106         }
 107
 108         if (buf == root->node) {
 109                 root->node = cow;
 110                 extent_buffer_get(cow);
 111                 if (buf != root->commit_root) {
 112                         btrfs_free_extent(trans, root, buf->start,
 113                                           buf->len, 1);
 114                 }
 115                 free_extent_buffer(buf);
 116         } else {
 117                 btrfs_set_node_blockptr(parent, parent_slot,
 118                                         cow->start);
 119                 btrfs_mark_buffer_dirty(parent);
 120                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 121                 btrfs_free_extent(trans, root, buf->start, buf->len, 1);
 122         }
 123         free_extent_buffer(buf);
 124         btrfs_mark_buffer_dirty(cow);
 125         *cow_ret = cow;
 126         return 0;
 127 }
 128
 129 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 130                     struct btrfs_root *root, struct extent_buffer *buf,
 131                     struct extent_buffer *parent, int parent_slot,
 132                     struct extent_buffer **cow_ret)
 133 {
 134         u64 search_start;
 135         int ret;
 136         if (trans->transaction != root->fs_info->running_transaction) {
 137                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 138                        root->fs_info->running_transaction->transid);
 139                 WARN_ON(1);
 140         }
 141         if (trans->transid != root->fs_info->generation) {
 142                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 143                        root->fs_info->generation);
 144                 WARN_ON(1);
 145         }
 146         if (btrfs_header_generation(buf) == trans->transid) {
 147                 *cow_ret = buf;
 148                 return 0;
 149         }
 150
 151         search_start = buf->start & ~((u64)BTRFS_BLOCK_GROUP_SIZE - 1);
 152         ret = __btrfs_cow_block(trans, root, buf, parent,
 153                                  parent_slot, cow_ret, search_start, 0);
 154         (*cow_ret)->alloc_addr = (unsigned long)__builtin_return_address(0);
 155         return ret;
 156 }
 157
 158 #if 0
 159 static int close_blocks(u64 blocknr, u64 other)
 160 {
 161         if (blocknr < other && other - blocknr < 8)
 162                 return 1;
 163         if (blocknr > other && blocknr - other < 8)
 164                 return 1;
 165         return 0;
 166 }
 167
 168 static int should_defrag_leaf(struct extent_buffer *eb)
 169 {
 170         return 0;
 171         struct btrfs_leaf *leaf = btrfs_buffer_leaf(eb);
 172         struct btrfs_disk_key *key;
 173         u32 nritems;
 174
 175         if (buffer_defrag(bh))
 176                 return 1;
 177
 178         nritems = btrfs_header_nritems(&leaf->header);
 179         if (nritems == 0)
 180                 return 0;
 181
 182         key = &leaf->items[0].key;
 183         if (btrfs_disk_key_type(key) == BTRFS_DIR_ITEM_KEY)
 184                 return 1;
 185
 186         key = &leaf->items[nritems-1].key;
 187         if (btrfs_disk_key_type(key) == BTRFS_DIR_ITEM_KEY)
 188                 return 1;
 189         if (nritems > 4) {
 190                 key = &leaf->items[nritems/2].key;
 191                 if (btrfs_disk_key_type(key) == BTRFS_DIR_ITEM_KEY)
 192                         return 1;
 193         }
 194         return 0;
 195 }
 196 #endif
 197
 198 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 199                        struct btrfs_root *root, struct extent_buffer *parent,
 200                        int cache_only, u64 *last_ret)
 201 {
 202         return 0;
 203 #if 0
 204         struct btrfs_node *parent_node;
 205         struct extent_buffer *cur_eb;
 206         struct extent_buffer *tmp_eb;
 207         u64 blocknr;
 208         u64 search_start = *last_ret;
 209         u64 last_block = 0;
 210         u64 other;
 211         u32 parent_nritems;
 212         int start_slot;
 213         int end_slot;
 214         int i;
 215         int err = 0;
 216         int parent_level;
 217
 218         if (trans->transaction != root->fs_info->running_transaction) {
 219                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 220                        root->fs_info->running_transaction->transid);
 221                 WARN_ON(1);
 222         }
 223         if (trans->transid != root->fs_info->generation) {
 224                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 225                        root->fs_info->generation);
 226                 WARN_ON(1);
 227         }
 228         if (buffer_defrag_done(parent))
 229                 return 0;
 230
 231         parent_node = btrfs_buffer_node(parent);
 232         parent_nritems = btrfs_header_nritems(&parent_node->header);
 233         parent_level = btrfs_header_level(&parent_node->header);
 234
 235         start_slot = 0;
 236         end_slot = parent_nritems;
 237
 238         if (parent_nritems == 1)
 239                 return 0;
 240
 241         for (i = start_slot; i < end_slot; i++) {
 242                 int close = 1;
 243                 blocknr = btrfs_node_blockptr(parent_node, i);
 244                 if (last_block == 0)
 245                         last_block = blocknr;
 246                 if (i > 0) {
 247                         other = btrfs_node_blockptr(parent_node, i - 1);
 248                         close = close_blocks(blocknr, other);
 249                 }
 250                 if (close && i < end_slot - 1) {
 251                         other = btrfs_node_blockptr(parent_node, i + 1);
 252                         close = close_blocks(blocknr, other);
 253                 }
 254                 if (close) {
 255                         last_block = blocknr;
 256                         continue;
 257                 }
 258
 259                 cur_bh = btrfs_find_tree_block(root, blocknr);
 260                 if (!cur_bh || !buffer_uptodate(cur_bh) ||
 261                     buffer_locked(cur_bh) ||
 262                     (parent_level != 1 && !buffer_defrag(cur_bh)) ||
 263                     (parent_level == 1 && !should_defrag_leaf(cur_bh))) {
 264                         if (cache_only) {
 265                                 brelse(cur_bh);
 266                                 continue;
 267                         }
 268                         if (!cur_bh || !buffer_uptodate(cur_bh) ||
 269                             buffer_locked(cur_bh)) {
 270                                 brelse(cur_bh);
 271                                 cur_bh = read_tree_block(root, blocknr);
 272                         }
 273                 }
 274                 if (search_start == 0)
 275                         search_start = last_block & ~((u64)65535);
 276
 277                 err = __btrfs_cow_block(trans, root, cur_bh, parent, i,
 278                                         &tmp_bh, search_start,
 279                                         min(8, end_slot - i));
 280                 if (err) {
 281                         brelse(cur_bh);
 282                         break;
 283                 }
 284                 search_start = bh_blocknr(tmp_bh);
 285                 *last_ret = search_start;
 286                 if (parent_level == 1)
 287                         clear_buffer_defrag(tmp_bh);
 288                 set_buffer_defrag_done(tmp_bh);
 289                 brelse(tmp_bh);
 290         }
 291         return err;
 292 #endif
 293 }
 294
 295 /*
 296  * The leaf data grows from end-to-front in the node.
 297  * this returns the address of the start of the last item,
 298  * which is the stop of the leaf data stack
 299  */
 300 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 301                                          struct extent_buffer *leaf)
 302 {
 303         u32 nr = btrfs_header_nritems(leaf);
 304         if (nr == 0)
 305                 return BTRFS_LEAF_DATA_SIZE(root);
 306         return btrfs_item_offset_nr(leaf, nr - 1);
 307 }
 308
 309 /*
 310  * compare two keys in a memcmp fashion
 311  */
 312 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 313 {
 314         struct btrfs_key k1;
 315
 316         btrfs_disk_key_to_cpu(&k1, disk);
 317
 318         if (k1.objectid > k2->objectid)
 319                 return 1;
 320         if (k1.objectid < k2->objectid)
 321                 return -1;
 322         if (k1.type > k2->type)
 323                 return 1;
 324         if (k1.type < k2->type)
 325                 return -1;
 326         if (k1.offset > k2->offset)
 327                 return 1;
 328         if (k1.offset < k2->offset)
 329                 return -1;
 330         return 0;
 331 }
 332
 333 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
 334                       int level)
 335 {
 336         struct extent_buffer *parent = NULL;
 337         struct extent_buffer *node = path->nodes[level];
 338         struct btrfs_disk_key parent_key;
 339         struct btrfs_disk_key node_key;
 340         int parent_slot;
 341         int slot;
 342         struct btrfs_key cpukey;
 343         u32 nritems = btrfs_header_nritems(node);
 344
 345         if (path->nodes[level + 1])
 346                 parent = path->nodes[level + 1];
 347
 348         slot = path->slots[level];
 349         BUG_ON(nritems == 0);
 350         if (parent) {
 351                 parent_slot = path->slots[level + 1];
 352                 btrfs_node_key(parent, &parent_key, parent_slot);
 353                 btrfs_node_key(node, &node_key, 0);
 354                 BUG_ON(memcmp(&parent_key, &node_key,
 355                               sizeof(struct btrfs_disk_key)));
 356                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 357                        btrfs_header_bytenr(node));
 358         }
 359         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 360         if (slot != 0) {
 361                 btrfs_node_key_to_cpu(node, &cpukey, slot - 1);
 362                 btrfs_node_key(node, &node_key, slot);
 363                 BUG_ON(comp_keys(&node_key, &cpukey) <= 0);
 364         }
 365         if (slot < nritems - 1) {
 366                 btrfs_node_key_to_cpu(node, &cpukey, slot + 1);
 367                 btrfs_node_key(node, &node_key, slot);
 368                 BUG_ON(comp_keys(&node_key, &cpukey) >= 0);
 369         }
 370         return 0;
 371 }
 372
 373 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 374                       int level)
 375 {
 376         struct extent_buffer *leaf = path->nodes[level];
 377         struct extent_buffer *parent = NULL;
 378         int parent_slot;
 379         struct btrfs_key cpukey;
 380         struct btrfs_disk_key parent_key;
 381         struct btrfs_disk_key leaf_key;
 382         int slot = path->slots[0];
 383
 384         u32 nritems = btrfs_header_nritems(leaf);
 385
 386         if (path->nodes[level + 1])
 387                 parent = path->nodes[level + 1];
 388
 389         if (nritems == 0)
 390                 return 0;
 391
 392         if (parent) {
 393                 parent_slot = path->slots[level + 1];
 394                 btrfs_node_key(parent, &parent_key, parent_slot);
 395                 btrfs_item_key(leaf, &leaf_key, 0);
 396
 397                 BUG_ON(memcmp(&parent_key, &leaf_key,
 398                        sizeof(struct btrfs_disk_key)));
 399                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 400                        btrfs_header_bytenr(leaf));
 401         }
 402 #if 0
 403         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 404                 btrfs_item_key_to_cpu(leaf, &cpukey, i + 1);
 405                 btrfs_item_key(leaf, &leaf_key, i);
 406                 if (comp_keys(&leaf_key, &cpukey) >= 0) {
 407                         btrfs_print_leaf(root, leaf);
 408                         printk("slot %d offset bad key\n", i);
 409                         BUG_ON(1);
 410                 }
 411                 if (btrfs_item_offset_nr(leaf, i) !=
 412                         btrfs_item_end_nr(leaf, i + 1)) {
 413                         btrfs_print_leaf(root, leaf);
 414                         printk("slot %d offset bad\n", i);
 415                         BUG_ON(1);
 416                 }
 417                 if (i == 0) {
 418                         if (btrfs_item_offset_nr(leaf, i) +
 419                                btrfs_item_size_nr(leaf, i) !=
 420                                BTRFS_LEAF_DATA_SIZE(root)) {
 421                                 btrfs_print_leaf(root, leaf);
 422                                 printk("slot %d first offset bad\n", i);
 423                                 BUG_ON(1);
 424                         }
 425                 }
 426         }
 427         if (nritems > 0) {
 428                 if (btrfs_item_size_nr(leaf, nritems - 1) > 4096) {
 429                                 btrfs_print_leaf(root, leaf);
 430                                 printk("slot %d bad size \n", nritems - 1);
 431                                 BUG_ON(1);
 432                 }
 433         }
 434 #endif
 435         if (slot != 0 && slot < nritems - 1) {
 436                 btrfs_item_key(leaf, &leaf_key, slot);
 437                 btrfs_item_key_to_cpu(leaf, &cpukey, slot - 1);
 438                 if (comp_keys(&leaf_key, &cpukey) <= 0) {
 439                         btrfs_print_leaf(root, leaf);
 440                         printk("slot %d offset bad key\n", slot);
 441                         BUG_ON(1);
 442                 }
 443                 if (btrfs_item_offset_nr(leaf, slot - 1) !=
 444                        btrfs_item_end_nr(leaf, slot)) {
 445                         btrfs_print_leaf(root, leaf);
 446                         printk("slot %d offset bad\n", slot);
 447                         BUG_ON(1);
 448                 }
 449         }
 450         if (slot < nritems - 1) {
 451                 btrfs_item_key(leaf, &leaf_key, slot);
 452                 btrfs_item_key_to_cpu(leaf, &cpukey, slot + 1);
 453                 BUG_ON(comp_keys(&leaf_key, &cpukey) >= 0);
 454                 if (btrfs_item_offset_nr(leaf, slot) !=
 455                         btrfs_item_end_nr(leaf, slot + 1)) {
 456                         btrfs_print_leaf(root, leaf);
 457                         printk("slot %d offset bad\n", slot);
 458                         BUG_ON(1);
 459                 }
 460         }
 461         BUG_ON(btrfs_item_offset_nr(leaf, 0) +
 462                btrfs_item_size_nr(leaf, 0) != BTRFS_LEAF_DATA_SIZE(root));
 463         return 0;
 464 }
 465
 466 static int check_block(struct btrfs_root *root, struct btrfs_path *path,
 467                         int level)
 468 {
 469 #if 0
 470         struct extent_buffer *buf = path->nodes[level];
 471
 472         if (memcmp_extent_buffer(buf, root->fs_info->fsid,
 473                                  (unsigned long)btrfs_header_fsid(buf),
 474                                  BTRFS_FSID_SIZE)) {
 475                 printk("warning bad block %Lu\n", buf->start);
 476                 return 1;
 477         }
 478 #endif
 479         if (level == 0)
 480                 return check_leaf(root, path, level);
 481         return check_node(root, path, level);
 482 }
 483
 484 /*
 485  * search for key in the extent_buffer.  The items start at offset p,
 486  * and they are item_size apart.  There are 'max' items in p.
 487  *
 488  * the slot in the array is returned via slot, and it points to
 489  * the place where you would insert key if it is not found in
 490  * the array.
 491  *
 492  * slot may point to max if the key is bigger than all of the keys
 493  */
 494 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 495                               int item_size, struct btrfs_key *key,
 496                               int max, int *slot)
 497 {
 498         int low = 0;
 499         int high = max;
 500         int mid;
 501         int ret;
 502         struct btrfs_disk_key *tmp = NULL;
 503         struct btrfs_disk_key unaligned;
 504         unsigned long offset;
 505         char *map_token = NULL;
 506         char *kaddr = NULL;
 507         unsigned long map_start = 0;
 508         unsigned long map_len = 0;
 509         int err;
 510
 511         while(low < high) {
 512                 mid = (low + high) / 2;
 513                 offset = p + mid * item_size;
 514
 515                 if (!map_token || offset < map_start ||
 516                     (offset + sizeof(struct btrfs_disk_key)) >
 517                     map_start + map_len) {
 518                         if (map_token) {
 519                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 520                                 map_token = NULL;
 521                         }
 522                         err = map_extent_buffer(eb, offset,
 523                                                 sizeof(struct btrfs_disk_key),
 524                                                 &map_token, &kaddr,
 525                                                 &map_start, &map_len, KM_USER0);
 526
 527                         if (!err) {
 528                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
 529                                                         map_start);
 530                         } else {
 531                                 read_extent_buffer(eb, &unaligned,
 532                                                    offset, sizeof(unaligned));
 533                                 tmp = &unaligned;
 534                         }
 535
 536                 } else {
 537                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
 538                                                         map_start);
 539                 }
 540                 ret = comp_keys(tmp, key);
 541
 542                 if (ret < 0)
 543                         low = mid + 1;
 544                 else if (ret > 0)
 545                         high = mid;
 546                 else {
 547                         *slot = mid;
 548                         if (map_token)
 549                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 550                         return 0;
 551                 }
 552         }
 553         *slot = low;
 554         if (map_token)
 555                 unmap_extent_buffer(eb, map_token, KM_USER0);
 556         return 1;
 557 }
 558
 559 /*
 560  * simple bin_search frontend that does the right thing for
 561  * leaves vs nodes
 562  */
 563 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 564                       int level, int *slot)
 565 {
 566         if (level == 0) {
 567                 return generic_bin_search(eb,
 568                                           offsetof(struct btrfs_leaf, items),
 569                                           sizeof(struct btrfs_item),
 570                                           key, btrfs_header_nritems(eb),
 571                                           slot);
 572         } else {
 573                 return generic_bin_search(eb,
 574                                           offsetof(struct btrfs_node, ptrs),
 575                                           sizeof(struct btrfs_key_ptr),
 576                                           key, btrfs_header_nritems(eb),
 577                                           slot);
 578         }
 579         return -1;
 580 }
 581
 582 static struct extent_buffer *read_node_slot(struct btrfs_root *root,
 583                                    struct extent_buffer *parent, int slot)
 584 {
 585         if (slot < 0)
 586                 return NULL;
 587         if (slot >= btrfs_header_nritems(parent))
 588                 return NULL;
 589         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 590                        btrfs_level_size(root, btrfs_header_level(parent) - 1));
 591 }
 592
 593 static int balance_level(struct btrfs_trans_handle *trans, struct btrfs_root
 594                          *root, struct btrfs_path *path, int level)
 595 {
 596         struct extent_buffer *right = NULL;
 597         struct extent_buffer *mid;
 598         struct extent_buffer *left = NULL;
 599         struct extent_buffer *parent = NULL;
 600         int ret = 0;
 601         int wret;
 602         int pslot;
 603         int orig_slot = path->slots[level];
 604         int err_on_enospc = 0;
 605         u64 orig_ptr;
 606
 607         if (level == 0)
 608                 return 0;
 609
 610         mid = path->nodes[level];
 611         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 612
 613         if (level < BTRFS_MAX_LEVEL - 1)
 614                 parent = path->nodes[level + 1];
 615         pslot = path->slots[level + 1];
 616
 617         /*
 618          * deal with the case where there is only one pointer in the root
 619          * by promoting the node below to a root
 620          */
 621         if (!parent) {
 622                 struct extent_buffer *child;
 623
 624                 if (btrfs_header_nritems(mid) != 1)
 625                         return 0;
 626
 627                 /* promote the child to a root */
 628                 child = read_node_slot(root, mid, 0);
 629                 BUG_ON(!child);
 630                 root->node = child;
 631                 path->nodes[level] = NULL;
 632                 clean_tree_block(trans, root, mid);
 633                 wait_on_tree_block_writeback(root, mid);
 634                 /* once for the path */
 635                 free_extent_buffer(mid);
 636                 ret = btrfs_free_extent(trans, root, mid->start, mid->len, 1);
 637                 /* once for the root ptr */
 638                 free_extent_buffer(mid);
 639                 return ret;
 640         }
 641         if (btrfs_header_nritems(mid) >
 642             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 643                 return 0;
 644
 645         if (btrfs_header_nritems(mid) < 2)
 646                 err_on_enospc = 1;
 647
 648         left = read_node_slot(root, parent, pslot - 1);
 649         if (left) {
 650                 wret = btrfs_cow_block(trans, root, left,
 651                                        parent, pslot - 1, &left);
 652                 if (wret) {
 653                         ret = wret;
 654                         goto enospc;
 655                 }
 656         }
 657         right = read_node_slot(root, parent, pslot + 1);
 658         if (right) {
 659                 wret = btrfs_cow_block(trans, root, right,
 660                                        parent, pslot + 1, &right);
 661                 if (wret) {
 662                         ret = wret;
 663                         goto enospc;
 664                 }
 665         }
 666
 667         /* first, try to make some room in the middle buffer */
 668         if (left) {
 669                 orig_slot += btrfs_header_nritems(left);
 670                 wret = push_node_left(trans, root, left, mid);
 671                 if (wret < 0)
 672                         ret = wret;
 673                 if (btrfs_header_nritems(mid) < 2)
 674                         err_on_enospc = 1;
 675         }
 676
 677         /*
 678          * then try to empty the right most buffer into the middle
 679          */
 680         if (right) {
 681                 wret = push_node_left(trans, root, mid, right);
 682                 if (wret < 0 && wret != -ENOSPC)
 683                         ret = wret;
 684                 if (btrfs_header_nritems(right) == 0) {
 685                         u64 bytenr = right->start;
 686                         u32 blocksize = right->len;
 687
 688                         clean_tree_block(trans, root, right);
 689                         wait_on_tree_block_writeback(root, right);
 690                         free_extent_buffer(right);
 691                         right = NULL;
 692                         wret = del_ptr(trans, root, path, level + 1, pslot +
 693                                        1);
 694                         if (wret)
 695                                 ret = wret;
 696                         wret = btrfs_free_extent(trans, root, bytenr,
 697                                                  blocksize, 1);
 698                         if (wret)
 699                                 ret = wret;
 700                 } else {
 701                         struct btrfs_disk_key right_key;
 702                         btrfs_node_key(right, &right_key, 0);
 703                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 704                         btrfs_mark_buffer_dirty(parent);
 705                 }
 706         }
 707         if (btrfs_header_nritems(mid) == 1) {
 708                 /*
 709                  * we're not allowed to leave a node with one item in the
 710                  * tree during a delete.  A deletion from lower in the tree
 711                  * could try to delete the only pointer in this node.
 712                  * So, pull some keys from the left.
 713                  * There has to be a left pointer at this point because
 714                  * otherwise we would have pulled some pointers from the
 715                  * right
 716                  */
 717                 BUG_ON(!left);
 718                 wret = balance_node_right(trans, root, mid, left);
 719                 if (wret < 0) {
 720                         ret = wret;
 721                         goto enospc;
 722                 }
 723                 BUG_ON(wret == 1);
 724         }
 725         if (btrfs_header_nritems(mid) == 0) {
 726                 /* we've managed to empty the middle node, drop it */
 727                 u64 bytenr = mid->start;
 728                 u32 blocksize = mid->len;
 729                 clean_tree_block(trans, root, mid);
 730                 wait_on_tree_block_writeback(root, mid);
 731                 free_extent_buffer(mid);
 732                 mid = NULL;
 733                 wret = del_ptr(trans, root, path, level + 1, pslot);
 734                 if (wret)
 735                         ret = wret;
 736                 wret = btrfs_free_extent(trans, root, bytenr, blocksize, 1);
 737                 if (wret)
 738                         ret = wret;
 739         } else {
 740                 /* update the parent key to reflect our changes */
 741                 struct btrfs_disk_key mid_key;
 742                 btrfs_node_key(mid, &mid_key, 0);
 743                 btrfs_set_node_key(parent, &mid_key, pslot);
 744                 btrfs_mark_buffer_dirty(parent);
 745         }
 746
 747         /* update the path */
 748         if (left) {
 749                 if (btrfs_header_nritems(left) > orig_slot) {
 750                         extent_buffer_get(left);
 751                         path->nodes[level] = left;
 752                         path->slots[level + 1] -= 1;
 753                         path->slots[level] = orig_slot;
 754                         if (mid)
 755                                 free_extent_buffer(mid);
 756                 } else {
 757                         orig_slot -= btrfs_header_nritems(left);
 758                         path->slots[level] = orig_slot;
 759                 }
 760         }
 761         /* double check we haven't messed things up */
 762         check_block(root, path, level);
 763         if (orig_ptr !=
 764             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 765                 BUG();
 766 enospc:
 767         if (right)
 768                 free_extent_buffer(right);
 769         if (left)
 770                 free_extent_buffer(left);
 771         return ret;
 772 }
 773
 774 /* returns zero if the push worked, non-zero otherwise */
 775 static int push_nodes_for_insert(struct btrfs_trans_handle *trans,
 776                                 struct btrfs_root *root,
 777                                 struct btrfs_path *path, int level)
 778 {
 779         struct extent_buffer *right = NULL;
 780         struct extent_buffer *mid;
 781         struct extent_buffer *left = NULL;
 782         struct extent_buffer *parent = NULL;
 783         int ret = 0;
 784         int wret;
 785         int pslot;
 786         int orig_slot = path->slots[level];
 787         u64 orig_ptr;
 788
 789         if (level == 0)
 790                 return 1;
 791
 792         mid = path->nodes[level];
 793         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 794
 795         if (level < BTRFS_MAX_LEVEL - 1)
 796                 parent = path->nodes[level + 1];
 797         pslot = path->slots[level + 1];
 798
 799         if (!parent)
 800                 return 1;
 801
 802         left = read_node_slot(root, parent, pslot - 1);
 803
 804         /* first, try to make some room in the middle buffer */
 805         if (left) {
 806                 u32 left_nr;
 807                 left_nr = btrfs_header_nritems(left);
 808                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 809                         wret = 1;
 810                 } else {
 811                         ret = btrfs_cow_block(trans, root, left, parent,
 812                                               pslot - 1, &left);
 813                         if (ret)
 814                                 wret = 1;
 815                         else {
 816                                 wret = push_node_left(trans, root,
 817                                                       left, mid);
 818                         }
 819                 }
 820                 if (wret < 0)
 821                         ret = wret;
 822                 if (wret == 0) {
 823                         struct btrfs_disk_key disk_key;
 824                         orig_slot += left_nr;
 825                         btrfs_node_key(mid, &disk_key, 0);
 826                         btrfs_set_node_key(parent, &disk_key, pslot);
 827                         btrfs_mark_buffer_dirty(parent);
 828                         if (btrfs_header_nritems(left) > orig_slot) {
 829                                 path->nodes[level] = left;
 830                                 path->slots[level + 1] -= 1;
 831                                 path->slots[level] = orig_slot;
 832                                 free_extent_buffer(mid);
 833                         } else {
 834                                 orig_slot -=
 835                                         btrfs_header_nritems(left);
 836                                 path->slots[level] = orig_slot;
 837                                 free_extent_buffer(left);
 838                         }
 839                         return 0;
 840                 }
 841                 free_extent_buffer(left);
 842         }
 843         right= read_node_slot(root, parent, pslot + 1);
 844
 845         /*
 846          * then try to empty the right most buffer into the middle
 847          */
 848         if (right) {
 849                 u32 right_nr;
 850                 right_nr = btrfs_header_nritems(right);
 851                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 852                         wret = 1;
 853                 } else {
 854                         ret = btrfs_cow_block(trans, root, right,
 855                                               parent, pslot + 1,
 856                                               &right);
 857                         if (ret)
 858                                 wret = 1;
 859                         else {
 860                                 wret = balance_node_right(trans, root,
 861                                                           right, mid);
 862                         }
 863                 }
 864                 if (wret < 0)
 865                         ret = wret;
 866                 if (wret == 0) {
 867                         struct btrfs_disk_key disk_key;
 868
 869                         btrfs_node_key(right, &disk_key, 0);
 870                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
 871                         btrfs_mark_buffer_dirty(parent);
 872
 873                         if (btrfs_header_nritems(mid) <= orig_slot) {
 874                                 path->nodes[level] = right;
 875                                 path->slots[level + 1] += 1;
 876                                 path->slots[level] = orig_slot -
 877                                         btrfs_header_nritems(mid);
 878                                 free_extent_buffer(mid);
 879                         } else {
 880                                 free_extent_buffer(right);
 881                         }
 882                         return 0;
 883                 }
 884                 free_extent_buffer(right);
 885         }
 886         return 1;
 887 }
 888
 889 /*
 890  * readahead one full node of leaves
 891  */
 892 static void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
 893                              int level, int slot)
 894 {
 895         return;
 896 #if 0
 897         struct extent_buffer *node;
 898         int i;
 899         u32 nritems;
 900         u64 bytenr;
 901         u64 search;
 902         u64 cluster_start;
 903         int ret;
 904         int nread = 0;
 905         int direction = path->reada;
 906         int level;
 907         struct radix_tree_root found;
 908         unsigned long gang[8];
 909         struct extent_buffer *eb;
 910
 911
 912         if (level == 0)
 913                 return;
 914
 915         if (!path->nodes[level])
 916                 return;
 917
 918         node = path->nodes[level];
 919         search = btrfs_node_blockptr(node, slot);
 920         eb = btrfs_find_tree_block(root, search);
 921         if (eb) {
 922                 free_extent_buffer(eb);
 923                 return;
 924         }
 925
 926         init_bit_radix(&found);
 927         nritems = btrfs_header_nritems(node);
 928         level = btrfs_header_level(node) - 1;
 929         for (i = slot; i < nritems; i++) {
 930                 bytenr = btrfs_node_blockptr(node, i);
 931                 set_radix_bit(&found, blocknr);
 932         }
 933         if (direction > 0) {
 934                 cluster_start = search - 4;
 935                 if (cluster_start > search)
 936                         cluster_start = 0;
 937         } else
 938                 cluster_start = search + 4;
 939         while(1) {
 940                 ret = find_first_radix_bit(&found, gang, 0, ARRAY_SIZE(gang));
 941                 if (!ret)
 942                         break;
 943                 for (i = 0; i < ret; i++) {
 944                         blocknr = gang[i];
 945                         clear_radix_bit(&found, blocknr);
 946                         if (path->reada == 1 && nread > 16)
 947                                 continue;
 948                         if (close_blocks(cluster_start, blocknr)) {
 949                                 readahead_tree_block(root, blocknr);
 950                                 nread++;
 951                                 cluster_start = blocknr;
 952                         }
 953                 }
 954         }
 955 #endif
 956 }
 957 /*
 958  * look for key in the tree.  path is filled in with nodes along the way
 959  * if key is found, we return zero and you can find the item in the leaf
 960  * level of the path (level 0)
 961  *
 962  * If the key isn't found, the path points to the slot where it should
 963  * be inserted, and 1 is returned.  If there are other errors during the
 964  * search a negative error number is returned.
 965  *
 966  * if ins_len > 0, nodes and leaves will be split as we walk down the
 967  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
 968  * possible)
 969  */
 970 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
 971                       *root, struct btrfs_key *key, struct btrfs_path *p, int
 972                       ins_len, int cow)
 973 {
 974         struct extent_buffer *b;
 975         u64 bytenr;
 976         int slot;
 977         int ret;
 978         int level;
 979         int should_reada = p->reada;
 980         u8 lowest_level = 0;
 981
 982         lowest_level = p->lowest_level;
 983         WARN_ON(lowest_level && ins_len);
 984         WARN_ON(p->nodes[0] != NULL);
 985         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
 986 again:
 987         b = root->node;
 988         extent_buffer_get(b);
 989         while (b) {
 990                 level = btrfs_header_level(b);
 991                 if (cow) {
 992                         int wret;
 993                         wret = btrfs_cow_block(trans, root, b,
 994                                                p->nodes[level + 1],
 995                                                p->slots[level + 1],
 996                                                &b);
 997                         if (wret) {
 998                                 free_extent_buffer(b);
 999                                 return wret;
1000                         }
1001                 }
1002                 BUG_ON(!cow && ins_len);
1003                 if (level != btrfs_header_level(b))
1004                         WARN_ON(1);
1005                 level = btrfs_header_level(b);
1006                 p->nodes[level] = b;
1007                 ret = check_block(root, p, level);
1008                 if (ret)
1009                         return -1;
1010                 ret = bin_search(b, key, level, &slot);
1011                 if (level != 0) {
1012                         if (ret && slot > 0)
1013                                 slot -= 1;
1014                         p->slots[level] = slot;
1015                         if (ins_len > 0 && btrfs_header_nritems(b) >=
1016                             BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1017                                 int sret = split_node(trans, root, p, level);
1018                                 BUG_ON(sret > 0);
1019                                 if (sret)
1020                                         return sret;
1021                                 b = p->nodes[level];
1022                                 slot = p->slots[level];
1023                         } else if (ins_len < 0) {
1024                                 int sret = balance_level(trans, root, p,
1025                                                          level);
1026                                 if (sret)
1027                                         return sret;
1028                                 b = p->nodes[level];
1029                                 if (!b) {
1030                                         btrfs_release_path(NULL, p);
1031                                         goto again;
1032                                 }
1033                                 slot = p->slots[level];
1034                                 BUG_ON(btrfs_header_nritems(b) == 1);
1035                         }
1036                         /* this is only true while dropping a snapshot */
1037                         if (level == lowest_level)
1038                                 break;
1039                         bytenr = btrfs_node_blockptr(b, slot);
1040                         if (should_reada)
1041                                 reada_for_search(root, p, level, slot);
1042                         b = read_tree_block(root, bytenr,
1043                                             btrfs_level_size(root, level - 1));
1044                 } else {
1045                         p->slots[level] = slot;
1046                         if (ins_len > 0 && btrfs_leaf_free_space(root, b) <
1047                             sizeof(struct btrfs_item) + ins_len) {
1048                                 int sret = split_leaf(trans, root, key,
1049                                                       p, ins_len);
1050                                 BUG_ON(sret > 0);
1051                                 if (sret)
1052                                         return sret;
1053                         }
1054                         return ret;
1055                 }
1056         }
1057         return 1;
1058 }
1059
1060 /*
1061  * adjust the pointers going up the tree, starting at level
1062  * making sure the right key of each node is points to 'key'.
1063  * This is used after shifting pointers to the left, so it stops
1064  * fixing up pointers when a given leaf/node is not in slot 0 of the
1065  * higher levels
1066  *
1067  * If this fails to write a tree block, it returns -1, but continues
1068  * fixing up the blocks in ram so the tree is consistent.
1069  */
1070 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1071                           struct btrfs_root *root, struct btrfs_path *path,
1072                           struct btrfs_disk_key *key, int level)
1073 {
1074         int i;
1075         int ret = 0;
1076         struct extent_buffer *t;
1077
1078         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1079                 int tslot = path->slots[i];
1080                 if (!path->nodes[i])
1081                         break;
1082                 t = path->nodes[i];
1083                 btrfs_set_node_key(t, key, tslot);
1084                 btrfs_mark_buffer_dirty(path->nodes[i]);
1085                 if (tslot != 0)
1086                         break;
1087         }
1088         return ret;
1089 }
1090
1091 /*
1092  * try to push data from one node into the next node left in the
1093  * tree.
1094  *
1095  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1096  * error, and > 0 if there was no room in the left hand block.
1097  */
1098 static int push_node_left(struct btrfs_trans_handle *trans, struct btrfs_root
1099                           *root, struct extent_buffer *dst,
1100                           struct extent_buffer *src)
1101 {
1102         int push_items = 0;
1103         int src_nritems;
1104         int dst_nritems;
1105         int ret = 0;
1106
1107         src_nritems = btrfs_header_nritems(src);
1108         dst_nritems = btrfs_header_nritems(dst);
1109         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1110
1111         if (push_items <= 0) {
1112                 return 1;
1113         }
1114
1115         if (src_nritems < push_items)
1116                 push_items = src_nritems;
1117
1118         copy_extent_buffer(dst, src,
1119                            btrfs_node_key_ptr_offset(dst_nritems),
1120                            btrfs_node_key_ptr_offset(0),
1121                            push_items * sizeof(struct btrfs_key_ptr));
1122
1123         if (push_items < src_nritems) {
1124                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1125                                       btrfs_node_key_ptr_offset(push_items),
1126                                       (src_nritems - push_items) *
1127                                       sizeof(struct btrfs_key_ptr));
1128         }
1129         btrfs_set_header_nritems(src, src_nritems - push_items);
1130         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1131         btrfs_mark_buffer_dirty(src);
1132         btrfs_mark_buffer_dirty(dst);
1133         return ret;
1134 }
1135
1136 /*
1137  * try to push data from one node into the next node right in the
1138  * tree.
1139  *
1140  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1141  * error, and > 0 if there was no room in the right hand block.
1142  *
1143  * this will  only push up to 1/2 the contents of the left node over
1144  */
1145 static int balance_node_right(struct btrfs_trans_handle *trans,
1146                               struct btrfs_root *root,
1147                               struct extent_buffer *dst,
1148                               struct extent_buffer *src)
1149 {
1150         int push_items = 0;
1151         int max_push;
1152         int src_nritems;
1153         int dst_nritems;
1154         int ret = 0;
1155
1156         src_nritems = btrfs_header_nritems(src);
1157         dst_nritems = btrfs_header_nritems(dst);
1158         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1159         if (push_items <= 0)
1160                 return 1;
1161
1162         max_push = src_nritems / 2 + 1;
1163         /* don't try to empty the node */
1164         if (max_push >= src_nritems)
1165                 return 1;
1166
1167         if (max_push < push_items)
1168                 push_items = max_push;
1169
1170         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1171                                       btrfs_node_key_ptr_offset(0),
1172                                       (dst_nritems) *
1173                                       sizeof(struct btrfs_key_ptr));
1174
1175         copy_extent_buffer(dst, src,
1176                            btrfs_node_key_ptr_offset(0),
1177                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1178                            push_items * sizeof(struct btrfs_key_ptr));
1179
1180         btrfs_set_header_nritems(src, src_nritems - push_items);
1181         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1182
1183         btrfs_mark_buffer_dirty(src);
1184         btrfs_mark_buffer_dirty(dst);
1185         return ret;
1186 }
1187
1188 /*
1189  * helper function to insert a new root level in the tree.
1190  * A new node is allocated, and a single item is inserted to
1191  * point to the existing root
1192  *
1193  * returns zero on success or < 0 on failure.
1194  */
1195 static int insert_new_root(struct btrfs_trans_handle *trans,
1196                            struct btrfs_root *root,
1197                            struct btrfs_path *path, int level)
1198 {
1199         struct extent_buffer *lower;
1200         struct extent_buffer *c;
1201         struct btrfs_disk_key lower_key;
1202
1203         BUG_ON(path->nodes[level]);
1204         BUG_ON(path->nodes[level-1] != root->node);
1205
1206         c = btrfs_alloc_free_block(trans, root, root->nodesize,
1207                                    root->node->start, 0);
1208         if (IS_ERR(c))
1209                 return PTR_ERR(c);
1210         memset_extent_buffer(c, 0, 0, root->nodesize);
1211         btrfs_set_header_nritems(c, 1);
1212         btrfs_set_header_level(c, level);
1213         btrfs_set_header_bytenr(c, c->start);
1214         btrfs_set_header_generation(c, trans->transid);
1215         btrfs_set_header_owner(c, root->root_key.objectid);
1216         lower = path->nodes[level-1];
1217
1218         write_extent_buffer(c, root->fs_info->fsid,
1219                             (unsigned long)btrfs_header_fsid(c),
1220                             BTRFS_FSID_SIZE);
1221         if (level == 1)
1222                 btrfs_item_key(lower, &lower_key, 0);
1223         else
1224                 btrfs_node_key(lower, &lower_key, 0);
1225         btrfs_set_node_key(c, &lower_key, 0);
1226         btrfs_set_node_blockptr(c, 0, lower->start);
1227
1228         btrfs_mark_buffer_dirty(c);
1229
1230         /* the super has an extra ref to root->node */
1231         free_extent_buffer(root->node);
1232         root->node = c;
1233         extent_buffer_get(c);
1234         path->nodes[level] = c;
1235         path->slots[level] = 0;
1236         return 0;
1237 }
1238
1239 /*
1240  * worker function to insert a single pointer in a node.
1241  * the node should have enough room for the pointer already
1242  *
1243  * slot and level indicate where you want the key to go, and
1244  * blocknr is the block the key points to.
1245  *
1246  * returns zero on success and < 0 on any error
1247  */
1248 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1249                       *root, struct btrfs_path *path, struct btrfs_disk_key
1250                       *key, u64 bytenr, int slot, int level)
1251 {
1252         struct extent_buffer *lower;
1253         int nritems;
1254
1255         BUG_ON(!path->nodes[level]);
1256         lower = path->nodes[level];
1257         nritems = btrfs_header_nritems(lower);
1258         if (slot > nritems)
1259                 BUG();
1260         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1261                 BUG();
1262         if (slot != nritems) {
1263                 memmove_extent_buffer(lower,
1264                               btrfs_node_key_ptr_offset(slot + 1),
1265                               btrfs_node_key_ptr_offset(slot),
1266                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1267         }
1268         btrfs_set_node_key(lower, key, slot);
1269         btrfs_set_node_blockptr(lower, slot, bytenr);
1270         btrfs_set_header_nritems(lower, nritems + 1);
1271         btrfs_mark_buffer_dirty(lower);
1272         return 0;
1273 }
1274
1275 /*
1276  * split the node at the specified level in path in two.
1277  * The path is corrected to point to the appropriate node after the split
1278  *
1279  * Before splitting this tries to make some room in the node by pushing
1280  * left and right, if either one works, it returns right away.
1281  *
1282  * returns 0 on success and < 0 on failure
1283  */
1284 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1285                       *root, struct btrfs_path *path, int level)
1286 {
1287         struct extent_buffer *c;
1288         struct extent_buffer *split;
1289         struct btrfs_disk_key disk_key;
1290         int mid;
1291         int ret;
1292         int wret;
1293         u32 c_nritems;
1294
1295         c = path->nodes[level];
1296         if (c == root->node) {
1297                 /* trying to split the root, lets make a new one */
1298                 ret = insert_new_root(trans, root, path, level + 1);
1299                 if (ret)
1300                         return ret;
1301         } else {
1302                 ret = push_nodes_for_insert(trans, root, path, level);
1303                 c = path->nodes[level];
1304                 if (!ret && btrfs_header_nritems(c) <
1305                     BTRFS_NODEPTRS_PER_BLOCK(root) - 1)
1306                         return 0;
1307                 if (ret < 0)
1308                         return ret;
1309         }
1310
1311         c_nritems = btrfs_header_nritems(c);
1312         split = btrfs_alloc_free_block(trans, root, root->nodesize,
1313                                        c->start, 0);
1314         if (IS_ERR(split))
1315                 return PTR_ERR(split);
1316
1317         btrfs_set_header_flags(split, btrfs_header_flags(c));
1318         btrfs_set_header_level(split, btrfs_header_level(c));
1319         btrfs_set_header_bytenr(split, split->start);
1320         btrfs_set_header_generation(split, trans->transid);
1321         btrfs_set_header_owner(split, root->root_key.objectid);
1322         write_extent_buffer(split, root->fs_info->fsid,
1323                             (unsigned long)btrfs_header_fsid(split),
1324                             BTRFS_FSID_SIZE);
1325
1326         mid = (c_nritems + 1) / 2;
1327
1328         copy_extent_buffer(split, c,
1329                            btrfs_node_key_ptr_offset(0),
1330                            btrfs_node_key_ptr_offset(mid),
1331                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1332         btrfs_set_header_nritems(split, c_nritems - mid);
1333         btrfs_set_header_nritems(c, mid);
1334         ret = 0;
1335
1336         btrfs_mark_buffer_dirty(c);
1337         btrfs_mark_buffer_dirty(split);
1338
1339         btrfs_node_key(split, &disk_key, 0);
1340         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1341                           path->slots[level + 1] + 1,
1342                           level + 1);
1343         if (wret)
1344                 ret = wret;
1345
1346         if (path->slots[level] >= mid) {
1347                 path->slots[level] -= mid;
1348                 free_extent_buffer(c);
1349                 path->nodes[level] = split;
1350                 path->slots[level + 1] += 1;
1351         } else {
1352                 free_extent_buffer(split);
1353         }
1354         return ret;
1355 }
1356
1357 /*
1358  * how many bytes are required to store the items in a leaf.  start
1359  * and nr indicate which items in the leaf to check.  This totals up the
1360  * space used both by the item structs and the item data
1361  */
1362 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1363 {
1364         int data_len;
1365         int nritems = btrfs_header_nritems(l);
1366         int end = min(nritems, start + nr) - 1;
1367
1368         if (!nr)
1369                 return 0;
1370         data_len = btrfs_item_end_nr(l, start);
1371         data_len = data_len - btrfs_item_offset_nr(l, end);
1372         data_len += sizeof(struct btrfs_item) * nr;
1373         WARN_ON(data_len < 0);
1374         return data_len;
1375 }
1376
1377 /*
1378  * The space between the end of the leaf items and
1379  * the start of the leaf data.  IOW, how much room
1380  * the leaf has left for both items and data
1381  */
1382 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1383 {
1384         int nritems = btrfs_header_nritems(leaf);
1385         int ret;
1386         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1387         if (ret < 0) {
1388                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1389                        ret, BTRFS_LEAF_DATA_SIZE(root),
1390                        leaf_space_used(leaf, 0, nritems), nritems);
1391         }
1392         return ret;
1393 }
1394
1395 /*
1396  * push some data in the path leaf to the right, trying to free up at
1397  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1398  *
1399  * returns 1 if the push failed because the other node didn't have enough
1400  * room, 0 if everything worked out and < 0 if there were major errors.
1401  */
1402 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1403                            *root, struct btrfs_path *path, int data_size)
1404 {
1405         struct extent_buffer *left = path->nodes[0];
1406         struct extent_buffer *right;
1407         struct extent_buffer *upper;
1408         struct btrfs_disk_key disk_key;
1409         int slot;
1410         int i;
1411         int free_space;
1412         int push_space = 0;
1413         int push_items = 0;
1414         struct btrfs_item *item;
1415         u32 left_nritems;
1416         u32 right_nritems;
1417         u32 data_end;
1418         u32 this_item_size;
1419         int ret;
1420
1421         slot = path->slots[1];
1422         if (!path->nodes[1]) {
1423                 return 1;
1424         }
1425         upper = path->nodes[1];
1426         if (slot >= btrfs_header_nritems(upper) - 1)
1427                 return 1;
1428
1429         right = read_tree_block(root, btrfs_node_blockptr(upper, slot + 1),
1430                                 root->leafsize);
1431         free_space = btrfs_leaf_free_space(root, right);
1432         if (free_space < data_size + sizeof(struct btrfs_item)) {
1433                 free_extent_buffer(right);
1434                 return 1;
1435         }
1436
1437         /* cow and double check */
1438         ret = btrfs_cow_block(trans, root, right, upper,
1439                               slot + 1, &right);
1440         if (ret) {
1441                 free_extent_buffer(right);
1442                 return 1;
1443         }
1444         free_space = btrfs_leaf_free_space(root, right);
1445         if (free_space < data_size + sizeof(struct btrfs_item)) {
1446                 free_extent_buffer(right);
1447                 return 1;
1448         }
1449
1450         left_nritems = btrfs_header_nritems(left);
1451         if (left_nritems == 0) {
1452                 free_extent_buffer(right);
1453                 return 1;
1454         }
1455
1456         for (i = left_nritems - 1; i >= 1; i--) {
1457                 item = btrfs_item_nr(left, i);
1458
1459                 if (path->slots[0] == i)
1460                         push_space += data_size + sizeof(*item);
1461
1462                 if (!left->map_token) {
1463                         map_extent_buffer(left, (unsigned long)item,
1464                                         sizeof(struct btrfs_item),
1465                                         &left->map_token, &left->kaddr,
1466                                         &left->map_start, &left->map_len,
1467                                         KM_USER1);
1468                 }
1469
1470                 this_item_size = btrfs_item_size(left, item);
1471                 if (this_item_size + sizeof(*item) + push_space > free_space)
1472                         break;
1473                 push_items++;
1474                 push_space += this_item_size + sizeof(*item);
1475         }
1476         if (left->map_token) {
1477                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1478                 left->map_token = NULL;
1479         }
1480
1481         if (push_items == 0) {
1482                 free_extent_buffer(right);
1483                 return 1;
1484         }
1485
1486         if (push_items == left_nritems)
1487                 WARN_ON(1);
1488
1489         /* push left to right */
1490         right_nritems = btrfs_header_nritems(right);
1491         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1492         push_space -= leaf_data_end(root, left);
1493
1494         /* make room in the right data area */
1495         data_end = leaf_data_end(root, right);
1496         memmove_extent_buffer(right,
1497                               btrfs_leaf_data(right) + data_end - push_space,
1498                               btrfs_leaf_data(right) + data_end,
1499                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1500
1501         /* copy from the left data area */
1502         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1503                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1504                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1505                      push_space);
1506
1507         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1508                               btrfs_item_nr_offset(0),
1509                               right_nritems * sizeof(struct btrfs_item));
1510
1511         /* copy the items from left to right */
1512         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1513                    btrfs_item_nr_offset(left_nritems - push_items),
1514                    push_items * sizeof(struct btrfs_item));
1515
1516         /* update the item pointers */
1517         right_nritems += push_items;
1518         btrfs_set_header_nritems(right, right_nritems);
1519         push_space = BTRFS_LEAF_DATA_SIZE(root);
1520
1521         for (i = 0; i < right_nritems; i++) {
1522                 item = btrfs_item_nr(right, i);
1523                 if (!right->map_token) {
1524                         map_extent_buffer(right, (unsigned long)item,
1525                                         sizeof(struct btrfs_item),
1526                                         &right->map_token, &right->kaddr,
1527                                         &right->map_start, &right->map_len,
1528                                         KM_USER1);
1529                 }
1530                 push_space -= btrfs_item_size(right, item);
1531                 btrfs_set_item_offset(right, item, push_space);
1532         }
1533
1534         if (right->map_token) {
1535                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1536                 right->map_token = NULL;
1537         }
1538         left_nritems -= push_items;
1539         btrfs_set_header_nritems(left, left_nritems);
1540
1541         btrfs_mark_buffer_dirty(left);
1542         btrfs_mark_buffer_dirty(right);
1543
1544         btrfs_item_key(right, &disk_key, 0);
1545         btrfs_set_node_key(upper, &disk_key, slot + 1);
1546         btrfs_mark_buffer_dirty(upper);
1547
1548         /* then fixup the leaf pointer in the path */
1549         if (path->slots[0] >= left_nritems) {
1550                 path->slots[0] -= left_nritems;
1551                 free_extent_buffer(path->nodes[0]);
1552                 path->nodes[0] = right;
1553                 path->slots[1] += 1;
1554         } else {
1555                 free_extent_buffer(right);
1556         }
1557         return 0;
1558 }
1559 /*
1560  * push some data in the path leaf to the left, trying to free up at
1561  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1562  */
1563 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1564                           *root, struct btrfs_path *path, int data_size)
1565 {
1566         struct btrfs_disk_key disk_key;
1567         struct extent_buffer *right = path->nodes[0];
1568         struct extent_buffer *left;
1569         int slot;
1570         int i;
1571         int free_space;
1572         int push_space = 0;
1573         int push_items = 0;
1574         struct btrfs_item *item;
1575         u32 old_left_nritems;
1576         u32 right_nritems;
1577         int ret = 0;
1578         int wret;
1579         u32 this_item_size;
1580         u32 old_left_item_size;
1581
1582         slot = path->slots[1];
1583         if (slot == 0)
1584                 return 1;
1585         if (!path->nodes[1])
1586                 return 1;
1587
1588         left = read_tree_block(root, btrfs_node_blockptr(path->nodes[1],
1589                                slot - 1), root->leafsize);
1590         free_space = btrfs_leaf_free_space(root, left);
1591         if (free_space < data_size + sizeof(struct btrfs_item)) {
1592                 free_extent_buffer(left);
1593                 return 1;
1594         }
1595
1596         /* cow and double check */
1597         ret = btrfs_cow_block(trans, root, left,
1598                               path->nodes[1], slot - 1, &left);
1599         if (ret) {
1600                 /* we hit -ENOSPC, but it isn't fatal here */
1601                 free_extent_buffer(left);
1602                 return 1;
1603         }
1604         free_space = btrfs_leaf_free_space(root, left);
1605         if (free_space < data_size + sizeof(struct btrfs_item)) {
1606                 free_extent_buffer(left);
1607                 return 1;
1608         }
1609
1610         right_nritems = btrfs_header_nritems(right);
1611         if (right_nritems == 0) {
1612                 free_extent_buffer(left);
1613                 return 1;
1614         }
1615
1616         for (i = 0; i < right_nritems - 1; i++) {
1617                 item = btrfs_item_nr(right, i);
1618                 if (!right->map_token) {
1619                         map_extent_buffer(right, (unsigned long)item,
1620                                         sizeof(struct btrfs_item),
1621                                         &right->map_token, &right->kaddr,
1622                                         &right->map_start, &right->map_len,
1623                                         KM_USER1);
1624                 }
1625
1626                 if (path->slots[0] == i)
1627                         push_space += data_size + sizeof(*item);
1628
1629                 this_item_size = btrfs_item_size(right, item);
1630                 if (this_item_size + sizeof(*item) + push_space > free_space)
1631                         break;
1632
1633                 push_items++;
1634                 push_space += this_item_size + sizeof(*item);
1635         }
1636
1637         if (right->map_token) {
1638                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1639                 right->map_token = NULL;
1640         }
1641
1642         if (push_items == 0) {
1643                 free_extent_buffer(left);
1644                 return 1;
1645         }
1646         if (push_items == btrfs_header_nritems(right))
1647                 WARN_ON(1);
1648
1649         /* push data from right to left */
1650         copy_extent_buffer(left, right,
1651                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1652                            btrfs_item_nr_offset(0),
1653                            push_items * sizeof(struct btrfs_item));
1654
1655         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1656                      btrfs_item_offset_nr(right, push_items -1);
1657
1658         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1659                      leaf_data_end(root, left) - push_space,
1660                      btrfs_leaf_data(right) +
1661                      btrfs_item_offset_nr(right, push_items - 1),
1662                      push_space);
1663         old_left_nritems = btrfs_header_nritems(left);
1664         BUG_ON(old_left_nritems < 0);
1665
1666         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1667         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1668                 u32 ioff;
1669
1670                 item = btrfs_item_nr(left, i);
1671                 if (!left->map_token) {
1672                         map_extent_buffer(left, (unsigned long)item,
1673                                         sizeof(struct btrfs_item),
1674                                         &left->map_token, &left->kaddr,
1675                                         &left->map_start, &left->map_len,
1676                                         KM_USER1);
1677                 }
1678
1679                 ioff = btrfs_item_offset(left, item);
1680                 btrfs_set_item_offset(left, item,
1681                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1682         }
1683         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1684         if (left->map_token) {
1685                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1686                 left->map_token = NULL;
1687         }
1688
1689         /* fixup right node */
1690         push_space = btrfs_item_offset_nr(right, push_items - 1) -
1691                                           leaf_data_end(root, right);
1692         memmove_extent_buffer(right, btrfs_leaf_data(right) +
1693                               BTRFS_LEAF_DATA_SIZE(root) - push_space,
1694                               btrfs_leaf_data(right) +
1695                               leaf_data_end(root, right), push_space);
1696
1697         memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1698                               btrfs_item_nr_offset(push_items),
1699                              (btrfs_header_nritems(right) - push_items) *
1700                              sizeof(struct btrfs_item));
1701
1702         right_nritems = btrfs_header_nritems(right) - push_items;
1703         btrfs_set_header_nritems(right, right_nritems);
1704         push_space = BTRFS_LEAF_DATA_SIZE(root);
1705
1706         for (i = 0; i < right_nritems; i++) {
1707                 item = btrfs_item_nr(right, i);
1708
1709                 if (!right->map_token) {
1710                         map_extent_buffer(right, (unsigned long)item,
1711                                         sizeof(struct btrfs_item),
1712                                         &right->map_token, &right->kaddr,
1713                                         &right->map_start, &right->map_len,
1714                                         KM_USER1);
1715                 }
1716
1717                 push_space = push_space - btrfs_item_size(right, item);
1718                 btrfs_set_item_offset(right, item, push_space);
1719         }
1720         if (right->map_token) {
1721                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1722                 right->map_token = NULL;
1723         }
1724
1725         btrfs_mark_buffer_dirty(left);
1726         btrfs_mark_buffer_dirty(right);
1727
1728         btrfs_item_key(right, &disk_key, 0);
1729         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1730         if (wret)
1731                 ret = wret;
1732
1733         /* then fixup the leaf pointer in the path */
1734         if (path->slots[0] < push_items) {
1735                 path->slots[0] += old_left_nritems;
1736                 free_extent_buffer(path->nodes[0]);
1737                 path->nodes[0] = left;
1738                 path->slots[1] -= 1;
1739         } else {
1740                 free_extent_buffer(left);
1741                 path->slots[0] -= push_items;
1742         }
1743         BUG_ON(path->slots[0] < 0);
1744         return ret;
1745 }
1746
1747 /*
1748  * split the path's leaf in two, making sure there is at least data_size
1749  * available for the resulting leaf level of the path.
1750  *
1751  * returns 0 if all went well and < 0 on failure.
1752  */
1753 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
1754                       *root, struct btrfs_key *ins_key,
1755                       struct btrfs_path *path, int data_size)
1756 {
1757         struct extent_buffer *l;
1758         u32 nritems;
1759         int mid;
1760         int slot;
1761         struct extent_buffer *right;
1762         int space_needed = data_size + sizeof(struct btrfs_item);
1763         int data_copy_size;
1764         int rt_data_off;
1765         int i;
1766         int ret = 0;
1767         int wret;
1768         int double_split = 0;
1769         struct btrfs_disk_key disk_key;
1770
1771         /* first try to make some room by pushing left and right */
1772         wret = push_leaf_left(trans, root, path, data_size);
1773         if (wret < 0)
1774                 return wret;
1775         if (wret) {
1776                 wret = push_leaf_right(trans, root, path, data_size);
1777                 if (wret < 0)
1778                         return wret;
1779         }
1780         l = path->nodes[0];
1781
1782         /* did the pushes work? */
1783         if (btrfs_leaf_free_space(root, l) >=
1784             sizeof(struct btrfs_item) + data_size)
1785                 return 0;
1786
1787         if (!path->nodes[1]) {
1788                 ret = insert_new_root(trans, root, path, 1);
1789                 if (ret)
1790                         return ret;
1791         }
1792         slot = path->slots[0];
1793         nritems = btrfs_header_nritems(l);
1794         mid = (nritems + 1)/ 2;
1795
1796         right = btrfs_alloc_free_block(trans, root, root->leafsize,
1797                                        l->start, 0);
1798         if (IS_ERR(right))
1799                 return PTR_ERR(right);
1800
1801         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
1802         btrfs_set_header_bytenr(right, right->start);
1803         btrfs_set_header_generation(right, trans->transid);
1804         btrfs_set_header_owner(right, root->root_key.objectid);
1805         btrfs_set_header_level(right, 0);
1806         write_extent_buffer(right, root->fs_info->fsid,
1807                             (unsigned long)btrfs_header_fsid(right),
1808                             BTRFS_FSID_SIZE);
1809
1810         if (mid <= slot) {
1811                 if (nritems == 1 ||
1812                     leaf_space_used(l, mid, nritems - mid) + space_needed >
1813                         BTRFS_LEAF_DATA_SIZE(root)) {
1814                         if (slot >= nritems) {
1815                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1816                                 btrfs_set_header_nritems(right, 0);
1817                                 wret = insert_ptr(trans, root, path,
1818                                                   &disk_key, right->start,
1819                                                   path->slots[1] + 1, 1);
1820                                 if (wret)
1821                                         ret = wret;
1822                                 free_extent_buffer(path->nodes[0]);
1823                                 path->nodes[0] = right;
1824                                 path->slots[0] = 0;
1825                                 path->slots[1] += 1;
1826                                 return ret;
1827                         }
1828                         mid = slot;
1829                         double_split = 1;
1830                 }
1831         } else {
1832                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
1833                         BTRFS_LEAF_DATA_SIZE(root)) {
1834                         if (slot == 0) {
1835                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1836                                 btrfs_set_header_nritems(right, 0);
1837                                 wret = insert_ptr(trans, root, path,
1838                                                   &disk_key,
1839                                                   right->start,
1840                                                   path->slots[1], 1);
1841                                 if (wret)
1842                                         ret = wret;
1843                                 free_extent_buffer(path->nodes[0]);
1844                                 path->nodes[0] = right;
1845                                 path->slots[0] = 0;
1846                                 if (path->slots[1] == 0) {
1847                                         wret = fixup_low_keys(trans, root,
1848                                                    path, &disk_key, 1);
1849                                         if (wret)
1850                                                 ret = wret;
1851                                 }
1852                                 return ret;
1853                         }
1854                         mid = slot;
1855                         double_split = 1;
1856                 }
1857         }
1858         nritems = nritems - mid;
1859         btrfs_set_header_nritems(right, nritems);
1860         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
1861
1862         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
1863                            btrfs_item_nr_offset(mid),
1864                            nritems * sizeof(struct btrfs_item));
1865
1866         copy_extent_buffer(right, l,
1867                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
1868                      data_copy_size, btrfs_leaf_data(l) +
1869                      leaf_data_end(root, l), data_copy_size);
1870
1871         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
1872                       btrfs_item_end_nr(l, mid);
1873
1874         for (i = 0; i < nritems; i++) {
1875                 struct btrfs_item *item = btrfs_item_nr(right, i);
1876                 u32 ioff;
1877
1878                 if (!right->map_token) {
1879                         map_extent_buffer(right, (unsigned long)item,
1880                                         sizeof(struct btrfs_item),
1881                                         &right->map_token, &right->kaddr,
1882                                         &right->map_start, &right->map_len,
1883                                         KM_USER1);
1884                 }
1885
1886                 ioff = btrfs_item_offset(right, item);
1887                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
1888         }
1889
1890         if (right->map_token) {
1891                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1892                 right->map_token = NULL;
1893         }
1894
1895         btrfs_set_header_nritems(l, mid);
1896         ret = 0;
1897         btrfs_item_key(right, &disk_key, 0);
1898         wret = insert_ptr(trans, root, path, &disk_key, right->start,
1899                           path->slots[1] + 1, 1);
1900         if (wret)
1901                 ret = wret;
1902
1903         btrfs_mark_buffer_dirty(right);
1904         btrfs_mark_buffer_dirty(l);
1905         BUG_ON(path->slots[0] != slot);
1906
1907         if (mid <= slot) {
1908                 free_extent_buffer(path->nodes[0]);
1909                 path->nodes[0] = right;
1910                 path->slots[0] -= mid;
1911                 path->slots[1] += 1;
1912         } else
1913                 free_extent_buffer(right);
1914
1915         BUG_ON(path->slots[0] < 0);
1916
1917         if (!double_split)
1918                 return ret;
1919
1920         right = btrfs_alloc_free_block(trans, root, root->leafsize,
1921                                        l->start, 0);
1922         if (IS_ERR(right))
1923                 return PTR_ERR(right);
1924
1925         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
1926         btrfs_set_header_bytenr(right, right->start);
1927         btrfs_set_header_generation(right, trans->transid);
1928         btrfs_set_header_owner(right, root->root_key.objectid);
1929         btrfs_set_header_level(right, 0);
1930         write_extent_buffer(right, root->fs_info->fsid,
1931                             (unsigned long)btrfs_header_fsid(right),
1932                             BTRFS_FSID_SIZE);
1933
1934         btrfs_cpu_key_to_disk(&disk_key, ins_key);
1935         btrfs_set_header_nritems(right, 0);
1936         wret = insert_ptr(trans, root, path,
1937                           &disk_key, right->start,
1938                           path->slots[1], 1);
1939         if (wret)
1940                 ret = wret;
1941         if (path->slots[1] == 0) {
1942                 wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1943                 if (wret)
1944                         ret = wret;
1945         }
1946         free_extent_buffer(path->nodes[0]);
1947         path->nodes[0] = right;
1948         path->slots[0] = 0;
1949         return ret;
1950 }
1951
1952 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
1953                         struct btrfs_root *root,
1954                         struct btrfs_path *path,
1955                         u32 new_size)
1956 {
1957         int ret = 0;
1958         int slot;
1959         int slot_orig;
1960         struct extent_buffer *leaf;
1961         struct btrfs_item *item;
1962         u32 nritems;
1963         unsigned int data_end;
1964         unsigned int old_data_start;
1965         unsigned int old_size;
1966         unsigned int size_diff;
1967         int i;
1968
1969         slot_orig = path->slots[0];
1970         leaf = path->nodes[0];
1971
1972         nritems = btrfs_header_nritems(leaf);
1973         data_end = leaf_data_end(root, leaf);
1974
1975         slot = path->slots[0];
1976         old_data_start = btrfs_item_offset_nr(leaf, slot);
1977         old_size = btrfs_item_size_nr(leaf, slot);
1978         BUG_ON(old_size <= new_size);
1979         size_diff = old_size - new_size;
1980
1981         BUG_ON(slot < 0);
1982         BUG_ON(slot >= nritems);
1983
1984         /*
1985          * item0..itemN ... dataN.offset..dataN.size .. data0.size
1986          */
1987         /* first correct the data pointers */
1988         for (i = slot; i < nritems; i++) {
1989                 u32 ioff;
1990                 item = btrfs_item_nr(leaf, i);
1991
1992                 if (!leaf->map_token) {
1993                         map_extent_buffer(leaf, (unsigned long)item,
1994                                         sizeof(struct btrfs_item),
1995                                         &leaf->map_token, &leaf->kaddr,
1996                                         &leaf->map_start, &leaf->map_len,
1997                                         KM_USER1);
1998                 }
1999
2000                 ioff = btrfs_item_offset(leaf, item);
2001                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2002         }
2003
2004         if (leaf->map_token) {
2005                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2006                 leaf->map_token = NULL;
2007         }
2008
2009         /* shift the data */
2010         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2011                       data_end + size_diff, btrfs_leaf_data(leaf) +
2012                       data_end, old_data_start + new_size - data_end);
2013
2014         item = btrfs_item_nr(leaf, slot);
2015         btrfs_set_item_size(leaf, item, new_size);
2016         btrfs_mark_buffer_dirty(leaf);
2017
2018         ret = 0;
2019         if (btrfs_leaf_free_space(root, leaf) < 0) {
2020                 btrfs_print_leaf(root, leaf);
2021                 BUG();
2022         }
2023         return ret;
2024 }
2025
2026 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2027                       struct btrfs_root *root, struct btrfs_path *path,
2028                       u32 data_size)
2029 {
2030         int ret = 0;
2031         int slot;
2032         int slot_orig;
2033         struct extent_buffer *leaf;
2034         struct btrfs_item *item;
2035         u32 nritems;
2036         unsigned int data_end;
2037         unsigned int old_data;
2038         unsigned int old_size;
2039         int i;
2040
2041         slot_orig = path->slots[0];
2042         leaf = path->nodes[0];
2043
2044         nritems = btrfs_header_nritems(leaf);
2045         data_end = leaf_data_end(root, leaf);
2046
2047         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2048                 btrfs_print_leaf(root, leaf);
2049                 BUG();
2050         }
2051         slot = path->slots[0];
2052         old_data = btrfs_item_end_nr(leaf, slot);
2053
2054         BUG_ON(slot < 0);
2055         BUG_ON(slot >= nritems);
2056
2057         /*
2058          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2059          */
2060         /* first correct the data pointers */
2061         for (i = slot; i < nritems; i++) {
2062                 u32 ioff;
2063                 item = btrfs_item_nr(leaf, i);
2064
2065                 if (!leaf->map_token) {
2066                         map_extent_buffer(leaf, (unsigned long)item,
2067                                         sizeof(struct btrfs_item),
2068                                         &leaf->map_token, &leaf->kaddr,
2069                                         &leaf->map_start, &leaf->map_len,
2070                                         KM_USER1);
2071                 }
2072                 ioff = btrfs_item_offset(leaf, item);
2073                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2074         }
2075
2076         if (leaf->map_token) {
2077                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2078                 leaf->map_token = NULL;
2079         }
2080
2081         /* shift the data */
2082         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2083                       data_end - data_size, btrfs_leaf_data(leaf) +
2084                       data_end, old_data - data_end);
2085
2086         data_end = old_data;
2087         old_size = btrfs_item_size_nr(leaf, slot);
2088         item = btrfs_item_nr(leaf, slot);
2089         btrfs_set_item_size(leaf, item, old_size + data_size);
2090         btrfs_mark_buffer_dirty(leaf);
2091
2092         ret = 0;
2093         if (btrfs_leaf_free_space(root, leaf) < 0) {
2094                 btrfs_print_leaf(root, leaf);
2095                 BUG();
2096         }
2097         return ret;
2098 }
2099
2100 /*
2101  * Given a key and some data, insert an item into the tree.
2102  * This does all the path init required, making room in the tree if needed.
2103  */
2104 int btrfs_insert_empty_item(struct btrfs_trans_handle *trans,
2105                             struct btrfs_root *root,
2106                             struct btrfs_path *path,
2107                             struct btrfs_key *cpu_key, u32 data_size)
2108 {
2109         struct extent_buffer *leaf;
2110         struct btrfs_item *item;
2111         int ret = 0;
2112         int slot;
2113         int slot_orig;
2114         u32 nritems;
2115         unsigned int data_end;
2116         struct btrfs_disk_key disk_key;
2117
2118         btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2119
2120         /* create a root if there isn't one */
2121         if (!root->node)
2122                 BUG();
2123
2124         ret = btrfs_search_slot(trans, root, cpu_key, path, data_size, 1);
2125         if (ret == 0) {
2126                 return -EEXIST;
2127         }
2128         if (ret < 0)
2129                 goto out;
2130
2131         slot_orig = path->slots[0];
2132         leaf = path->nodes[0];
2133
2134         nritems = btrfs_header_nritems(leaf);
2135         data_end = leaf_data_end(root, leaf);
2136
2137         if (btrfs_leaf_free_space(root, leaf) <
2138             sizeof(struct btrfs_item) + data_size) {
2139                 BUG();
2140         }
2141
2142         slot = path->slots[0];
2143         BUG_ON(slot < 0);
2144
2145         if (slot != nritems) {
2146                 int i;
2147                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2148
2149                 if (old_data < data_end) {
2150                         btrfs_print_leaf(root, leaf);
2151                         printk("slot %d old_data %d data_end %d\n",
2152                                slot, old_data, data_end);
2153                         BUG_ON(1);
2154                 }
2155                 /*
2156                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2157                  */
2158                 /* first correct the data pointers */
2159                 WARN_ON(leaf->map_token);
2160                 for (i = slot; i < nritems; i++) {
2161                         u32 ioff;
2162
2163                         item = btrfs_item_nr(leaf, i);
2164                         if (!leaf->map_token) {
2165                                 map_extent_buffer(leaf, (unsigned long)item,
2166                                         sizeof(struct btrfs_item),
2167                                         &leaf->map_token, &leaf->kaddr,
2168                                         &leaf->map_start, &leaf->map_len,
2169                                         KM_USER1);
2170                         }
2171
2172                         ioff = btrfs_item_offset(leaf, item);
2173                         btrfs_set_item_offset(leaf, item, ioff - data_size);
2174                 }
2175                 if (leaf->map_token) {
2176                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2177                         leaf->map_token = NULL;
2178                 }
2179
2180                 /* shift the items */
2181                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
2182                               btrfs_item_nr_offset(slot),
2183                               (nritems - slot) * sizeof(struct btrfs_item));
2184
2185                 /* shift the data */
2186                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2187                               data_end - data_size, btrfs_leaf_data(leaf) +
2188                               data_end, old_data - data_end);
2189                 data_end = old_data;
2190         }
2191
2192         /* setup the item for the new data */
2193         btrfs_set_item_key(leaf, &disk_key, slot);
2194         item = btrfs_item_nr(leaf, slot);
2195         btrfs_set_item_offset(leaf, item, data_end - data_size);
2196         btrfs_set_item_size(leaf, item, data_size);
2197         btrfs_set_header_nritems(leaf, nritems + 1);
2198         btrfs_mark_buffer_dirty(leaf);
2199
2200         ret = 0;
2201         if (slot == 0)
2202                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2203
2204         if (btrfs_leaf_free_space(root, leaf) < 0) {
2205                 btrfs_print_leaf(root, leaf);
2206                 BUG();
2207         }
2208 out:
2209         return ret;
2210 }
2211
2212 /*
2213  * Given a key and some data, insert an item into the tree.
2214  * This does all the path init required, making room in the tree if needed.
2215  */
2216 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2217                       *root, struct btrfs_key *cpu_key, void *data, u32
2218                       data_size)
2219 {
2220         int ret = 0;
2221         struct btrfs_path *path;
2222         struct extent_buffer *leaf;
2223         unsigned long ptr;
2224
2225         path = btrfs_alloc_path();
2226         BUG_ON(!path);
2227         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2228         if (!ret) {
2229                 leaf = path->nodes[0];
2230                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2231                 write_extent_buffer(leaf, data, ptr, data_size);
2232                 btrfs_mark_buffer_dirty(leaf);
2233         }
2234         btrfs_free_path(path);
2235         return ret;
2236 }
2237
2238 /*
2239  * delete the pointer from a given node.
2240  *
2241  * If the delete empties a node, the node is removed from the tree,
2242  * continuing all the way the root if required.  The root is converted into
2243  * a leaf if all the nodes are emptied.
2244  */
2245 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2246                    struct btrfs_path *path, int level, int slot)
2247 {
2248         struct extent_buffer *parent = path->nodes[level];
2249         u32 nritems;
2250         int ret = 0;
2251         int wret;
2252
2253         nritems = btrfs_header_nritems(parent);
2254         if (slot != nritems -1) {
2255                 memmove_extent_buffer(parent,
2256                               btrfs_node_key_ptr_offset(slot),
2257                               btrfs_node_key_ptr_offset(slot + 1),
2258                               sizeof(struct btrfs_key_ptr) *
2259                               (nritems - slot - 1));
2260         }
2261         nritems--;
2262         btrfs_set_header_nritems(parent, nritems);
2263         if (nritems == 0 && parent == root->node) {
2264                 BUG_ON(btrfs_header_level(root->node) != 1);
2265                 /* just turn the root into a leaf and break */
2266                 btrfs_set_header_level(root->node, 0);
2267         } else if (slot == 0) {
2268                 struct btrfs_disk_key disk_key;
2269
2270                 btrfs_node_key(parent, &disk_key, 0);
2271                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2272                 if (wret)
2273                         ret = wret;
2274         }
2275         btrfs_mark_buffer_dirty(parent);
2276         return ret;
2277 }
2278
2279 /*
2280  * delete the item at the leaf level in path.  If that empties
2281  * the leaf, remove it from the tree
2282  */
2283 int btrfs_del_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2284                    struct btrfs_path *path)
2285 {
2286         int slot;
2287         struct extent_buffer *leaf;
2288         struct btrfs_item *item;
2289         int doff;
2290         int dsize;
2291         int ret = 0;
2292         int wret;
2293         u32 nritems;
2294
2295         leaf = path->nodes[0];
2296         slot = path->slots[0];
2297         doff = btrfs_item_offset_nr(leaf, slot);
2298         dsize = btrfs_item_size_nr(leaf, slot);
2299         nritems = btrfs_header_nritems(leaf);
2300
2301         if (slot != nritems - 1) {
2302                 int i;
2303                 int data_end = leaf_data_end(root, leaf);
2304
2305                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2306                               data_end + dsize,
2307                               btrfs_leaf_data(leaf) + data_end,
2308                               doff - data_end);
2309
2310                 for (i = slot + 1; i < nritems; i++) {
2311                         u32 ioff;
2312
2313                         item = btrfs_item_nr(leaf, i);
2314                         if (!leaf->map_token) {
2315                                 map_extent_buffer(leaf, (unsigned long)item,
2316                                         sizeof(struct btrfs_item),
2317                                         &leaf->map_token, &leaf->kaddr,
2318                                         &leaf->map_start, &leaf->map_len,
2319                                         KM_USER1);
2320                         }
2321                         ioff = btrfs_item_offset(leaf, item);
2322                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2323                 }
2324
2325                 if (leaf->map_token) {
2326                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2327                         leaf->map_token = NULL;
2328                 }
2329
2330                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2331                               btrfs_item_nr_offset(slot + 1),
2332                               sizeof(struct btrfs_item) *
2333                               (nritems - slot - 1));
2334         }
2335         btrfs_set_header_nritems(leaf, nritems - 1);
2336         nritems--;
2337
2338         /* delete the leaf if we've emptied it */
2339         if (nritems == 0) {
2340                 if (leaf == root->node) {
2341                         btrfs_set_header_level(leaf, 0);
2342                 } else {
2343                         clean_tree_block(trans, root, leaf);
2344                         wait_on_tree_block_writeback(root, leaf);
2345                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
2346                         if (wret)
2347                                 ret = wret;
2348                         wret = btrfs_free_extent(trans, root,
2349                                                  leaf->start, leaf->len, 1);
2350                         if (wret)
2351                                 ret = wret;
2352                 }
2353         } else {
2354                 int used = leaf_space_used(leaf, 0, nritems);
2355                 if (slot == 0) {
2356                         struct btrfs_disk_key disk_key;
2357
2358                         btrfs_item_key(leaf, &disk_key, 0);
2359                         wret = fixup_low_keys(trans, root, path,
2360                                               &disk_key, 1);
2361                         if (wret)
2362                                 ret = wret;
2363                 }
2364
2365                 /* delete the leaf if it is mostly empty */
2366                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 3) {
2367                         /* push_leaf_left fixes the path.
2368                          * make sure the path still points to our leaf
2369                          * for possible call to del_ptr below
2370                          */
2371                         slot = path->slots[1];
2372                         extent_buffer_get(leaf);
2373
2374                         wret = push_leaf_left(trans, root, path, 1);
2375                         if (wret < 0 && wret != -ENOSPC)
2376                                 ret = wret;
2377
2378                         if (path->nodes[0] == leaf &&
2379                             btrfs_header_nritems(leaf)) {
2380                                 wret = push_leaf_right(trans, root, path, 1);
2381                                 if (wret < 0 && wret != -ENOSPC)
2382                                         ret = wret;
2383                         }
2384
2385                         if (btrfs_header_nritems(leaf) == 0) {
2386                                 u64 bytenr = leaf->start;
2387                                 u32 blocksize = leaf->len;
2388
2389                                 clean_tree_block(trans, root, leaf);
2390                                 wait_on_tree_block_writeback(root, leaf);
2391
2392                                 wret = del_ptr(trans, root, path, 1, slot);
2393                                 if (wret)
2394                                         ret = wret;
2395
2396                                 free_extent_buffer(leaf);
2397                                 wret = btrfs_free_extent(trans, root, bytenr,
2398                                                          blocksize, 1);
2399                                 if (wret)
2400                                         ret = wret;
2401                         } else {
2402                                 btrfs_mark_buffer_dirty(leaf);
2403                                 free_extent_buffer(leaf);
2404                         }
2405                 } else {
2406                         btrfs_mark_buffer_dirty(leaf);
2407                 }
2408         }
2409         return ret;
2410 }
2411
2412 /*
2413  * walk up the tree as far as required to find the next leaf.
2414  * returns 0 if it found something or 1 if there are no greater leaves.
2415  * returns < 0 on io errors.
2416  */
2417 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2418 {
2419         int slot;
2420         int level = 1;
2421         u64 bytenr;
2422         struct extent_buffer *c;
2423         struct extent_buffer *next = NULL;
2424
2425         while(level < BTRFS_MAX_LEVEL) {
2426                 if (!path->nodes[level])
2427                         return 1;
2428
2429                 slot = path->slots[level] + 1;
2430                 c = path->nodes[level];
2431                 if (slot >= btrfs_header_nritems(c)) {
2432                         level++;
2433                         continue;
2434                 }
2435
2436                 bytenr = btrfs_node_blockptr(c, slot);
2437                 if (next)
2438                         free_extent_buffer(next);
2439
2440                 if (path->reada)
2441                         reada_for_search(root, path, level, slot);
2442
2443                 next = read_tree_block(root, bytenr,
2444                                        btrfs_level_size(root, level -1));
2445                 break;
2446         }
2447         path->slots[level] = slot;
2448         while(1) {
2449                 level--;
2450                 c = path->nodes[level];
2451                 free_extent_buffer(c);
2452                 path->nodes[level] = next;
2453                 path->slots[level] = 0;
2454                 if (!level)
2455                         break;
2456                 if (path->reada)
2457                         reada_for_search(root, path, level, 0);
2458                 next = read_tree_block(root, btrfs_node_blockptr(next, 0),
2459                                        btrfs_level_size(root, level - 1));
2460         }
2461         return 0;
2462 }