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 "ctree.h"
  20 #include "disk-io.h"
  21 #include "transaction.h"
  22 #include "print-tree.h"
  23
  24 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  25                       *root, struct btrfs_path *path, int level);
  26 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  27                       *root, struct btrfs_key *ins_key,
  28                       struct btrfs_path *path, int data_size);
  29 static int push_node_left(struct btrfs_trans_handle *trans,
  30                           struct btrfs_root *root, struct extent_buffer *dst,
  31                           struct extent_buffer *src);
  32 static int balance_node_right(struct btrfs_trans_handle *trans,
  33                               struct btrfs_root *root,
  34                               struct extent_buffer *dst_buf,
  35                               struct extent_buffer *src_buf);
  36 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  37                    struct btrfs_path *path, int level, int slot);
  38
  39 inline void btrfs_init_path(struct btrfs_path *p)
  40 {
  41         memset(p, 0, sizeof(*p));
  42 }
  43
  44 struct btrfs_path *btrfs_alloc_path(void)
  45 {
  46         struct btrfs_path *path;
  47         path = kmem_cache_alloc(btrfs_path_cachep, GFP_NOFS);
  48         if (path) {
  49                 btrfs_init_path(path);
  50                 path->reada = 1;
  51         }
  52         return path;
  53 }
  54
  55 void btrfs_free_path(struct btrfs_path *p)
  56 {
  57         btrfs_release_path(NULL, p);
  58         kmem_cache_free(btrfs_path_cachep, p);
  59 }
  60
  61 void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
  62 {
  63         int i;
  64         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  65                 if (!p->nodes[i])
  66                         break;
  67                 free_extent_buffer(p->nodes[i]);
  68         }
  69         memset(p, 0, sizeof(*p));
  70 }
  71
  72 static int __btrfs_cow_block(struct btrfs_trans_handle *trans,
  73                              struct btrfs_root *root,
  74                              struct extent_buffer *buf,
  75                              struct extent_buffer *parent, int parent_slot,
  76                              struct extent_buffer **cow_ret,
  77                              u64 search_start, u64 empty_size)
  78 {
  79         struct extent_buffer *cow;
  80         int ret = 0;
  81         int different_trans = 0;
  82
  83         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
  84
  85         cow = btrfs_alloc_free_block(trans, root, buf->len,
  86                                      search_start, empty_size);
  87         if (IS_ERR(cow))
  88                 return PTR_ERR(cow);
  89
  90         copy_extent_buffer(cow, buf, 0, 0, cow->len);
  91         btrfs_set_header_bytenr(cow, cow->start);
  92         btrfs_set_header_generation(cow, trans->transid);
  93         btrfs_set_header_owner(cow, root->root_key.objectid);
  94
  95         WARN_ON(btrfs_header_generation(buf) > trans->transid);
  96         if (btrfs_header_generation(buf) != trans->transid) {
  97                 different_trans = 1;
  98                 ret = btrfs_inc_ref(trans, root, buf);
  99                 if (ret)
 100                         return ret;
 101         } else {
 102                 clean_tree_block(trans, root, buf);
 103         }
 104
 105         if (buf == root->node) {
 106                 root->node = cow;
 107                 extent_buffer_get(cow);
 108                 if (buf != root->commit_root) {
 109                         btrfs_free_extent(trans, root, buf->start,
 110                                           buf->len, 1);
 111                 }
 112                 free_extent_buffer(buf);
 113         } else {
 114                 btrfs_set_node_blockptr(parent, parent_slot,
 115                                         cow->start);
 116                 btrfs_mark_buffer_dirty(parent);
 117                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 118                 btrfs_free_extent(trans, root, buf->start, buf->len, 1);
 119         }
 120         free_extent_buffer(buf);
 121         btrfs_mark_buffer_dirty(cow);
 122         *cow_ret = cow;
 123         return 0;
 124 }
 125
 126 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 127                     struct btrfs_root *root, struct extent_buffer *buf,
 128                     struct extent_buffer *parent, int parent_slot,
 129                     struct extent_buffer **cow_ret)
 130 {
 131         u64 search_start;
 132         int ret;
 133         if (trans->transaction != root->fs_info->running_transaction) {
 134                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 135                        root->fs_info->running_transaction->transid);
 136                 WARN_ON(1);
 137         }
 138         if (trans->transid != root->fs_info->generation) {
 139                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 140                        root->fs_info->generation);
 141                 WARN_ON(1);
 142         }
 143         if (btrfs_header_generation(buf) == trans->transid) {
 144                 *cow_ret = buf;
 145                 return 0;
 146         }
 147
 148         search_start = buf->start & ~((u64)BTRFS_BLOCK_GROUP_SIZE - 1);
 149         ret = __btrfs_cow_block(trans, root, buf, parent,
 150                                  parent_slot, cow_ret, search_start, 0);
 151         return ret;
 152 }
 153
 154 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 155 {
 156         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 157                 return 1;
 158         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 159                 return 1;
 160         return 0;
 161 }
 162
 163 static int should_defrag_leaf(struct extent_buffer *leaf)
 164 {
 165         struct btrfs_key key;
 166         u32 nritems;
 167
 168         if (btrfs_buffer_defrag(leaf))
 169                 return 1;
 170
 171         nritems = btrfs_header_nritems(leaf);
 172         if (nritems == 0)
 173                 return 0;
 174
 175         btrfs_item_key_to_cpu(leaf, &key, 0);
 176         if (key.type == BTRFS_DIR_ITEM_KEY)
 177                 return 1;
 178
 179
 180         btrfs_item_key_to_cpu(leaf, &key, nritems - 1);
 181         if (key.type == BTRFS_DIR_ITEM_KEY)
 182                 return 1;
 183         if (nritems > 4) {
 184                 btrfs_item_key_to_cpu(leaf, &key, nritems / 2);
 185                 if (key.type == BTRFS_DIR_ITEM_KEY)
 186                         return 1;
 187         }
 188         return 0;
 189 }
 190
 191 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 192                        struct btrfs_root *root, struct extent_buffer *parent,
 193                        int cache_only, u64 *last_ret)
 194 {
 195         struct extent_buffer *cur;
 196         struct extent_buffer *tmp;
 197         u64 blocknr;
 198         u64 search_start = *last_ret;
 199         u64 last_block = 0;
 200         u64 other;
 201         u32 parent_nritems;
 202         int start_slot;
 203         int end_slot;
 204         int i;
 205         int err = 0;
 206         int parent_level;
 207         int uptodate;
 208         u32 blocksize;
 209
 210         if (trans->transaction != root->fs_info->running_transaction) {
 211                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 212                        root->fs_info->running_transaction->transid);
 213                 WARN_ON(1);
 214         }
 215         if (trans->transid != root->fs_info->generation) {
 216                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 217                        root->fs_info->generation);
 218                 WARN_ON(1);
 219         }
 220         parent_level = btrfs_header_level(parent);
 221
 222         parent_nritems = btrfs_header_nritems(parent);
 223         blocksize = btrfs_level_size(root, parent_level - 1);
 224
 225         start_slot = 0;
 226         end_slot = parent_nritems;
 227
 228         if (parent_nritems == 1)
 229                 return 0;
 230
 231         for (i = start_slot; i < end_slot; i++) {
 232                 int close = 1;
 233                 blocknr = btrfs_node_blockptr(parent, i);
 234                 if (last_block == 0)
 235                         last_block = blocknr;
 236                 if (i > 0) {
 237                         other = btrfs_node_blockptr(parent, i - 1);
 238                         close = close_blocks(blocknr, other, blocksize);
 239                 }
 240                 if (close && i < end_slot - 1) {
 241                         other = btrfs_node_blockptr(parent, i + 1);
 242                         close = close_blocks(blocknr, other, blocksize);
 243                 }
 244                 if (close) {
 245                         last_block = blocknr;
 246                         continue;
 247                 }
 248
 249                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
 250                 if (cur)
 251                         uptodate = btrfs_buffer_uptodate(cur);
 252                 else
 253                         uptodate = 0;
 254                 if (!cur || !uptodate ||
 255                     (parent_level != 1 && !btrfs_buffer_defrag(cur)) ||
 256                     (parent_level == 1 && !should_defrag_leaf(cur))) {
 257                         if (cache_only) {
 258                                 free_extent_buffer(cur);
 259                                 continue;
 260                         }
 261                         if (!cur) {
 262                                 cur = read_tree_block(root, blocknr,
 263                                                          blocksize);
 264                         } else if (!uptodate) {
 265                                 btrfs_read_buffer(cur);
 266                         }
 267                 }
 268                 if (search_start == 0)
 269                         search_start = last_block;
 270
 271                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 272                                         &tmp, search_start,
 273                                         min(16 * blocksize,
 274                                             (end_slot - i) * blocksize));
 275                 if (err) {
 276                         free_extent_buffer(cur);
 277                         break;
 278                 }
 279                 search_start = tmp->start;
 280                 *last_ret = search_start;
 281                 if (parent_level == 1)
 282                         btrfs_clear_buffer_defrag(tmp);
 283                 free_extent_buffer(tmp);
 284         }
 285         return err;
 286 }
 287
 288 /*
 289  * The leaf data grows from end-to-front in the node.
 290  * this returns the address of the start of the last item,
 291  * which is the stop of the leaf data stack
 292  */
 293 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 294                                          struct extent_buffer *leaf)
 295 {
 296         u32 nr = btrfs_header_nritems(leaf);
 297         if (nr == 0)
 298                 return BTRFS_LEAF_DATA_SIZE(root);
 299         return btrfs_item_offset_nr(leaf, nr - 1);
 300 }
 301
 302 /*
 303  * compare two keys in a memcmp fashion
 304  */
 305 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 306 {
 307         struct btrfs_key k1;
 308
 309         btrfs_disk_key_to_cpu(&k1, disk);
 310
 311         if (k1.objectid > k2->objectid)
 312                 return 1;
 313         if (k1.objectid < k2->objectid)
 314                 return -1;
 315         if (k1.type > k2->type)
 316                 return 1;
 317         if (k1.type < k2->type)
 318                 return -1;
 319         if (k1.offset > k2->offset)
 320                 return 1;
 321         if (k1.offset < k2->offset)
 322                 return -1;
 323         return 0;
 324 }
 325
 326 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
 327                       int level)
 328 {
 329         struct extent_buffer *parent = NULL;
 330         struct extent_buffer *node = path->nodes[level];
 331         struct btrfs_disk_key parent_key;
 332         struct btrfs_disk_key node_key;
 333         int parent_slot;
 334         int slot;
 335         struct btrfs_key cpukey;
 336         u32 nritems = btrfs_header_nritems(node);
 337
 338         if (path->nodes[level + 1])
 339                 parent = path->nodes[level + 1];
 340
 341         slot = path->slots[level];
 342         BUG_ON(nritems == 0);
 343         if (parent) {
 344                 parent_slot = path->slots[level + 1];
 345                 btrfs_node_key(parent, &parent_key, parent_slot);
 346                 btrfs_node_key(node, &node_key, 0);
 347                 BUG_ON(memcmp(&parent_key, &node_key,
 348                               sizeof(struct btrfs_disk_key)));
 349                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 350                        btrfs_header_bytenr(node));
 351         }
 352         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 353         if (slot != 0) {
 354                 btrfs_node_key_to_cpu(node, &cpukey, slot - 1);
 355                 btrfs_node_key(node, &node_key, slot);
 356                 BUG_ON(comp_keys(&node_key, &cpukey) <= 0);
 357         }
 358         if (slot < nritems - 1) {
 359                 btrfs_node_key_to_cpu(node, &cpukey, slot + 1);
 360                 btrfs_node_key(node, &node_key, slot);
 361                 BUG_ON(comp_keys(&node_key, &cpukey) >= 0);
 362         }
 363         return 0;
 364 }
 365
 366 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 367                       int level)
 368 {
 369         struct extent_buffer *leaf = path->nodes[level];
 370         struct extent_buffer *parent = NULL;
 371         int parent_slot;
 372         struct btrfs_key cpukey;
 373         struct btrfs_disk_key parent_key;
 374         struct btrfs_disk_key leaf_key;
 375         int slot = path->slots[0];
 376
 377         u32 nritems = btrfs_header_nritems(leaf);
 378
 379         if (path->nodes[level + 1])
 380                 parent = path->nodes[level + 1];
 381
 382         if (nritems == 0)
 383                 return 0;
 384
 385         if (parent) {
 386                 parent_slot = path->slots[level + 1];
 387                 btrfs_node_key(parent, &parent_key, parent_slot);
 388                 btrfs_item_key(leaf, &leaf_key, 0);
 389
 390                 BUG_ON(memcmp(&parent_key, &leaf_key,
 391                        sizeof(struct btrfs_disk_key)));
 392                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 393                        btrfs_header_bytenr(leaf));
 394         }
 395 #if 0
 396         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 397                 btrfs_item_key_to_cpu(leaf, &cpukey, i + 1);
 398                 btrfs_item_key(leaf, &leaf_key, i);
 399                 if (comp_keys(&leaf_key, &cpukey) >= 0) {
 400                         btrfs_print_leaf(root, leaf);
 401                         printk("slot %d offset bad key\n", i);
 402                         BUG_ON(1);
 403                 }
 404                 if (btrfs_item_offset_nr(leaf, i) !=
 405                         btrfs_item_end_nr(leaf, i + 1)) {
 406                         btrfs_print_leaf(root, leaf);
 407                         printk("slot %d offset bad\n", i);
 408                         BUG_ON(1);
 409                 }
 410                 if (i == 0) {
 411                         if (btrfs_item_offset_nr(leaf, i) +
 412                                btrfs_item_size_nr(leaf, i) !=
 413                                BTRFS_LEAF_DATA_SIZE(root)) {
 414                                 btrfs_print_leaf(root, leaf);
 415                                 printk("slot %d first offset bad\n", i);
 416                                 BUG_ON(1);
 417                         }
 418                 }
 419         }
 420         if (nritems > 0) {
 421                 if (btrfs_item_size_nr(leaf, nritems - 1) > 4096) {
 422                                 btrfs_print_leaf(root, leaf);
 423                                 printk("slot %d bad size \n", nritems - 1);
 424                                 BUG_ON(1);
 425                 }
 426         }
 427 #endif
 428         if (slot != 0 && slot < nritems - 1) {
 429                 btrfs_item_key(leaf, &leaf_key, slot);
 430                 btrfs_item_key_to_cpu(leaf, &cpukey, slot - 1);
 431                 if (comp_keys(&leaf_key, &cpukey) <= 0) {
 432                         btrfs_print_leaf(root, leaf);
 433                         printk("slot %d offset bad key\n", slot);
 434                         BUG_ON(1);
 435                 }
 436                 if (btrfs_item_offset_nr(leaf, slot - 1) !=
 437                        btrfs_item_end_nr(leaf, slot)) {
 438                         btrfs_print_leaf(root, leaf);
 439                         printk("slot %d offset bad\n", slot);
 440                         BUG_ON(1);
 441                 }
 442         }
 443         if (slot < nritems - 1) {
 444                 btrfs_item_key(leaf, &leaf_key, slot);
 445                 btrfs_item_key_to_cpu(leaf, &cpukey, slot + 1);
 446                 BUG_ON(comp_keys(&leaf_key, &cpukey) >= 0);
 447                 if (btrfs_item_offset_nr(leaf, slot) !=
 448                         btrfs_item_end_nr(leaf, slot + 1)) {
 449                         btrfs_print_leaf(root, leaf);
 450                         printk("slot %d offset bad\n", slot);
 451                         BUG_ON(1);
 452                 }
 453         }
 454         BUG_ON(btrfs_item_offset_nr(leaf, 0) +
 455                btrfs_item_size_nr(leaf, 0) != BTRFS_LEAF_DATA_SIZE(root));
 456         return 0;
 457 }
 458
 459 static int check_block(struct btrfs_root *root, struct btrfs_path *path,
 460                         int level)
 461 {
 462         return 0;
 463 #if 0
 464         struct extent_buffer *buf = path->nodes[level];
 465
 466         if (memcmp_extent_buffer(buf, root->fs_info->fsid,
 467                                  (unsigned long)btrfs_header_fsid(buf),
 468                                  BTRFS_FSID_SIZE)) {
 469                 printk("warning bad block %Lu\n", buf->start);
 470                 return 1;
 471         }
 472 #endif
 473         if (level == 0)
 474                 return check_leaf(root, path, level);
 475         return check_node(root, path, level);
 476 }
 477
 478 /*
 479  * search for key in the extent_buffer.  The items start at offset p,
 480  * and they are item_size apart.  There are 'max' items in p.
 481  *
 482  * the slot in the array is returned via slot, and it points to
 483  * the place where you would insert key if it is not found in
 484  * the array.
 485  *
 486  * slot may point to max if the key is bigger than all of the keys
 487  */
 488 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 489                               int item_size, struct btrfs_key *key,
 490                               int max, int *slot)
 491 {
 492         int low = 0;
 493         int high = max;
 494         int mid;
 495         int ret;
 496         struct btrfs_disk_key *tmp = NULL;
 497         struct btrfs_disk_key unaligned;
 498         unsigned long offset;
 499         char *map_token = NULL;
 500         char *kaddr = NULL;
 501         unsigned long map_start = 0;
 502         unsigned long map_len = 0;
 503         int err;
 504
 505         while(low < high) {
 506                 mid = (low + high) / 2;
 507                 offset = p + mid * item_size;
 508
 509                 if (!map_token || offset < map_start ||
 510                     (offset + sizeof(struct btrfs_disk_key)) >
 511                     map_start + map_len) {
 512                         if (map_token) {
 513                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 514                                 map_token = NULL;
 515                         }
 516                         err = map_extent_buffer(eb, offset,
 517                                                 sizeof(struct btrfs_disk_key),
 518                                                 &map_token, &kaddr,
 519                                                 &map_start, &map_len, KM_USER0);
 520
 521                         if (!err) {
 522                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
 523                                                         map_start);
 524                         } else {
 525                                 read_extent_buffer(eb, &unaligned,
 526                                                    offset, sizeof(unaligned));
 527                                 tmp = &unaligned;
 528                         }
 529
 530                 } else {
 531                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
 532                                                         map_start);
 533                 }
 534                 ret = comp_keys(tmp, key);
 535
 536                 if (ret < 0)
 537                         low = mid + 1;
 538                 else if (ret > 0)
 539                         high = mid;
 540                 else {
 541                         *slot = mid;
 542                         if (map_token)
 543                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 544                         return 0;
 545                 }
 546         }
 547         *slot = low;
 548         if (map_token)
 549                 unmap_extent_buffer(eb, map_token, KM_USER0);
 550         return 1;
 551 }
 552
 553 /*
 554  * simple bin_search frontend that does the right thing for
 555  * leaves vs nodes
 556  */
 557 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 558                       int level, int *slot)
 559 {
 560         if (level == 0) {
 561                 return generic_bin_search(eb,
 562                                           offsetof(struct btrfs_leaf, items),
 563                                           sizeof(struct btrfs_item),
 564                                           key, btrfs_header_nritems(eb),
 565                                           slot);
 566         } else {
 567                 return generic_bin_search(eb,
 568                                           offsetof(struct btrfs_node, ptrs),
 569                                           sizeof(struct btrfs_key_ptr),
 570                                           key, btrfs_header_nritems(eb),
 571                                           slot);
 572         }
 573         return -1;
 574 }
 575
 576 static struct extent_buffer *read_node_slot(struct btrfs_root *root,
 577                                    struct extent_buffer *parent, int slot)
 578 {
 579         if (slot < 0)
 580                 return NULL;
 581         if (slot >= btrfs_header_nritems(parent))
 582                 return NULL;
 583         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 584                        btrfs_level_size(root, btrfs_header_level(parent) - 1));
 585 }
 586
 587 static int balance_level(struct btrfs_trans_handle *trans, struct btrfs_root
 588                          *root, struct btrfs_path *path, int level)
 589 {
 590         struct extent_buffer *right = NULL;
 591         struct extent_buffer *mid;
 592         struct extent_buffer *left = NULL;
 593         struct extent_buffer *parent = NULL;
 594         int ret = 0;
 595         int wret;
 596         int pslot;
 597         int orig_slot = path->slots[level];
 598         int err_on_enospc = 0;
 599         u64 orig_ptr;
 600
 601         if (level == 0)
 602                 return 0;
 603
 604         mid = path->nodes[level];
 605         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 606
 607         if (level < BTRFS_MAX_LEVEL - 1)
 608                 parent = path->nodes[level + 1];
 609         pslot = path->slots[level + 1];
 610
 611         /*
 612          * deal with the case where there is only one pointer in the root
 613          * by promoting the node below to a root
 614          */
 615         if (!parent) {
 616                 struct extent_buffer *child;
 617
 618                 if (btrfs_header_nritems(mid) != 1)
 619                         return 0;
 620
 621                 /* promote the child to a root */
 622                 child = read_node_slot(root, mid, 0);
 623                 BUG_ON(!child);
 624                 root->node = child;
 625                 path->nodes[level] = NULL;
 626                 clean_tree_block(trans, root, mid);
 627                 wait_on_tree_block_writeback(root, mid);
 628                 /* once for the path */
 629                 free_extent_buffer(mid);
 630                 ret = btrfs_free_extent(trans, root, mid->start, mid->len, 1);
 631                 /* once for the root ptr */
 632                 free_extent_buffer(mid);
 633                 return ret;
 634         }
 635         if (btrfs_header_nritems(mid) >
 636             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 637                 return 0;
 638
 639         if (btrfs_header_nritems(mid) < 2)
 640                 err_on_enospc = 1;
 641
 642         left = read_node_slot(root, parent, pslot - 1);
 643         if (left) {
 644                 wret = btrfs_cow_block(trans, root, left,
 645                                        parent, pslot - 1, &left);
 646                 if (wret) {
 647                         ret = wret;
 648                         goto enospc;
 649                 }
 650         }
 651         right = read_node_slot(root, parent, pslot + 1);
 652         if (right) {
 653                 wret = btrfs_cow_block(trans, root, right,
 654                                        parent, pslot + 1, &right);
 655                 if (wret) {
 656                         ret = wret;
 657                         goto enospc;
 658                 }
 659         }
 660
 661         /* first, try to make some room in the middle buffer */
 662         if (left) {
 663                 orig_slot += btrfs_header_nritems(left);
 664                 wret = push_node_left(trans, root, left, mid);
 665                 if (wret < 0)
 666                         ret = wret;
 667                 if (btrfs_header_nritems(mid) < 2)
 668                         err_on_enospc = 1;
 669         }
 670
 671         /*
 672          * then try to empty the right most buffer into the middle
 673          */
 674         if (right) {
 675                 wret = push_node_left(trans, root, mid, right);
 676                 if (wret < 0 && wret != -ENOSPC)
 677                         ret = wret;
 678                 if (btrfs_header_nritems(right) == 0) {
 679                         u64 bytenr = right->start;
 680                         u32 blocksize = right->len;
 681
 682                         clean_tree_block(trans, root, right);
 683                         wait_on_tree_block_writeback(root, right);
 684                         free_extent_buffer(right);
 685                         right = NULL;
 686                         wret = del_ptr(trans, root, path, level + 1, pslot +
 687                                        1);
 688                         if (wret)
 689                                 ret = wret;
 690                         wret = btrfs_free_extent(trans, root, bytenr,
 691                                                  blocksize, 1);
 692                         if (wret)
 693                                 ret = wret;
 694                 } else {
 695                         struct btrfs_disk_key right_key;
 696                         btrfs_node_key(right, &right_key, 0);
 697                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 698                         btrfs_mark_buffer_dirty(parent);
 699                 }
 700         }
 701         if (btrfs_header_nritems(mid) == 1) {
 702                 /*
 703                  * we're not allowed to leave a node with one item in the
 704                  * tree during a delete.  A deletion from lower in the tree
 705                  * could try to delete the only pointer in this node.
 706                  * So, pull some keys from the left.
 707                  * There has to be a left pointer at this point because
 708                  * otherwise we would have pulled some pointers from the
 709                  * right
 710                  */
 711                 BUG_ON(!left);
 712                 wret = balance_node_right(trans, root, mid, left);
 713                 if (wret < 0) {
 714                         ret = wret;
 715                         goto enospc;
 716                 }
 717                 BUG_ON(wret == 1);
 718         }
 719         if (btrfs_header_nritems(mid) == 0) {
 720                 /* we've managed to empty the middle node, drop it */
 721                 u64 bytenr = mid->start;
 722                 u32 blocksize = mid->len;
 723                 clean_tree_block(trans, root, mid);
 724                 wait_on_tree_block_writeback(root, mid);
 725                 free_extent_buffer(mid);
 726                 mid = NULL;
 727                 wret = del_ptr(trans, root, path, level + 1, pslot);
 728                 if (wret)
 729                         ret = wret;
 730                 wret = btrfs_free_extent(trans, root, bytenr, blocksize, 1);
 731                 if (wret)
 732                         ret = wret;
 733         } else {
 734                 /* update the parent key to reflect our changes */
 735                 struct btrfs_disk_key mid_key;
 736                 btrfs_node_key(mid, &mid_key, 0);
 737                 btrfs_set_node_key(parent, &mid_key, pslot);
 738                 btrfs_mark_buffer_dirty(parent);
 739         }
 740
 741         /* update the path */
 742         if (left) {
 743                 if (btrfs_header_nritems(left) > orig_slot) {
 744                         extent_buffer_get(left);
 745                         path->nodes[level] = left;
 746                         path->slots[level + 1] -= 1;
 747                         path->slots[level] = orig_slot;
 748                         if (mid)
 749                                 free_extent_buffer(mid);
 750                 } else {
 751                         orig_slot -= btrfs_header_nritems(left);
 752                         path->slots[level] = orig_slot;
 753                 }
 754         }
 755         /* double check we haven't messed things up */
 756         check_block(root, path, level);
 757         if (orig_ptr !=
 758             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 759                 BUG();
 760 enospc:
 761         if (right)
 762                 free_extent_buffer(right);
 763         if (left)
 764                 free_extent_buffer(left);
 765         return ret;
 766 }
 767
 768 /* returns zero if the push worked, non-zero otherwise */
 769 static int push_nodes_for_insert(struct btrfs_trans_handle *trans,
 770                                 struct btrfs_root *root,
 771                                 struct btrfs_path *path, int level)
 772 {
 773         struct extent_buffer *right = NULL;
 774         struct extent_buffer *mid;
 775         struct extent_buffer *left = NULL;
 776         struct extent_buffer *parent = NULL;
 777         int ret = 0;
 778         int wret;
 779         int pslot;
 780         int orig_slot = path->slots[level];
 781         u64 orig_ptr;
 782
 783         if (level == 0)
 784                 return 1;
 785
 786         mid = path->nodes[level];
 787         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 788
 789         if (level < BTRFS_MAX_LEVEL - 1)
 790                 parent = path->nodes[level + 1];
 791         pslot = path->slots[level + 1];
 792
 793         if (!parent)
 794                 return 1;
 795
 796         left = read_node_slot(root, parent, pslot - 1);
 797
 798         /* first, try to make some room in the middle buffer */
 799         if (left) {
 800                 u32 left_nr;
 801                 left_nr = btrfs_header_nritems(left);
 802                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 803                         wret = 1;
 804                 } else {
 805                         ret = btrfs_cow_block(trans, root, left, parent,
 806                                               pslot - 1, &left);
 807                         if (ret)
 808                                 wret = 1;
 809                         else {
 810                                 wret = push_node_left(trans, root,
 811                                                       left, mid);
 812                         }
 813                 }
 814                 if (wret < 0)
 815                         ret = wret;
 816                 if (wret == 0) {
 817                         struct btrfs_disk_key disk_key;
 818                         orig_slot += left_nr;
 819                         btrfs_node_key(mid, &disk_key, 0);
 820                         btrfs_set_node_key(parent, &disk_key, pslot);
 821                         btrfs_mark_buffer_dirty(parent);
 822                         if (btrfs_header_nritems(left) > orig_slot) {
 823                                 path->nodes[level] = left;
 824                                 path->slots[level + 1] -= 1;
 825                                 path->slots[level] = orig_slot;
 826                                 free_extent_buffer(mid);
 827                         } else {
 828                                 orig_slot -=
 829                                         btrfs_header_nritems(left);
 830                                 path->slots[level] = orig_slot;
 831                                 free_extent_buffer(left);
 832                         }
 833                         return 0;
 834                 }
 835                 free_extent_buffer(left);
 836         }
 837         right= read_node_slot(root, parent, pslot + 1);
 838
 839         /*
 840          * then try to empty the right most buffer into the middle
 841          */
 842         if (right) {
 843                 u32 right_nr;
 844                 right_nr = btrfs_header_nritems(right);
 845                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 846                         wret = 1;
 847                 } else {
 848                         ret = btrfs_cow_block(trans, root, right,
 849                                               parent, pslot + 1,
 850                                               &right);
 851                         if (ret)
 852                                 wret = 1;
 853                         else {
 854                                 wret = balance_node_right(trans, root,
 855                                                           right, mid);
 856                         }
 857                 }
 858                 if (wret < 0)
 859                         ret = wret;
 860                 if (wret == 0) {
 861                         struct btrfs_disk_key disk_key;
 862
 863                         btrfs_node_key(right, &disk_key, 0);
 864                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
 865                         btrfs_mark_buffer_dirty(parent);
 866
 867                         if (btrfs_header_nritems(mid) <= orig_slot) {
 868                                 path->nodes[level] = right;
 869                                 path->slots[level + 1] += 1;
 870                                 path->slots[level] = orig_slot -
 871                                         btrfs_header_nritems(mid);
 872                                 free_extent_buffer(mid);
 873                         } else {
 874                                 free_extent_buffer(right);
 875                         }
 876                         return 0;
 877                 }
 878                 free_extent_buffer(right);
 879         }
 880         return 1;
 881 }
 882
 883 /*
 884  * readahead one full node of leaves
 885  */
 886 static void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
 887                              int level, int slot)
 888 {
 889         struct extent_buffer *node;
 890         u32 nritems;
 891         u64 search;
 892         u64 lowest_read;
 893         u64 highest_read;
 894         u64 nread = 0;
 895         int direction = path->reada;
 896         struct extent_buffer *eb;
 897         u32 nr;
 898         u32 blocksize;
 899         u32 nscan = 0;
 900
 901         if (level == 0)
 902                 return;
 903
 904         if (!path->nodes[level])
 905                 return;
 906
 907         node = path->nodes[level];
 908         search = btrfs_node_blockptr(node, slot);
 909         blocksize = btrfs_level_size(root, level - 1);
 910         eb = btrfs_find_tree_block(root, search, blocksize);
 911         if (eb) {
 912                 free_extent_buffer(eb);
 913                 return;
 914         }
 915
 916         highest_read = search;
 917         lowest_read = search;
 918
 919         nritems = btrfs_header_nritems(node);
 920         nr = slot;
 921         while(1) {
 922                 if (direction < 0) {
 923                         if (nr == 0)
 924                                 break;
 925                         nr--;
 926                 } else if (direction > 0) {
 927                         nr++;
 928                         if (nr >= nritems)
 929                                 break;
 930                 }
 931                 search = btrfs_node_blockptr(node, nr);
 932                 if ((search >= lowest_read && search <= highest_read) ||
 933                     (search < lowest_read && lowest_read - search <= 32768) ||
 934                     (search > highest_read && search - highest_read <= 32768)) {
 935                         readahead_tree_block(root, search, blocksize);
 936                         nread += blocksize;
 937                 }
 938                 nscan++;
 939                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
 940                         break;
 941                 if(nread > (1024 * 1024) || nscan > 128)
 942                         break;
 943
 944                 if (search < lowest_read)
 945                         lowest_read = search;
 946                 if (search > highest_read)
 947                         highest_read = search;
 948         }
 949 }
 950 /*
 951  * look for key in the tree.  path is filled in with nodes along the way
 952  * if key is found, we return zero and you can find the item in the leaf
 953  * level of the path (level 0)
 954  *
 955  * If the key isn't found, the path points to the slot where it should
 956  * be inserted, and 1 is returned.  If there are other errors during the
 957  * search a negative error number is returned.
 958  *
 959  * if ins_len > 0, nodes and leaves will be split as we walk down the
 960  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
 961  * possible)
 962  */
 963 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
 964                       *root, struct btrfs_key *key, struct btrfs_path *p, int
 965                       ins_len, int cow)
 966 {
 967         struct extent_buffer *b;
 968         u64 bytenr;
 969         int slot;
 970         int ret;
 971         int level;
 972         int should_reada = p->reada;
 973         u8 lowest_level = 0;
 974
 975         lowest_level = p->lowest_level;
 976         WARN_ON(lowest_level && ins_len);
 977         WARN_ON(p->nodes[0] != NULL);
 978         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
 979 again:
 980         b = root->node;
 981         extent_buffer_get(b);
 982         while (b) {
 983                 level = btrfs_header_level(b);
 984                 if (cow) {
 985                         int wret;
 986                         wret = btrfs_cow_block(trans, root, b,
 987                                                p->nodes[level + 1],
 988                                                p->slots[level + 1],
 989                                                &b);
 990                         if (wret) {
 991                                 free_extent_buffer(b);
 992                                 return wret;
 993                         }
 994                 }
 995                 BUG_ON(!cow && ins_len);
 996                 if (level != btrfs_header_level(b))
 997                         WARN_ON(1);
 998                 level = btrfs_header_level(b);
 999                 p->nodes[level] = b;
1000                 ret = check_block(root, p, level);
1001                 if (ret)
1002                         return -1;
1003                 ret = bin_search(b, key, level, &slot);
1004                 if (level != 0) {
1005                         if (ret && slot > 0)
1006                                 slot -= 1;
1007                         p->slots[level] = slot;
1008                         if (ins_len > 0 && btrfs_header_nritems(b) >=
1009                             BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1010                                 int sret = split_node(trans, root, p, level);
1011                                 BUG_ON(sret > 0);
1012                                 if (sret)
1013                                         return sret;
1014                                 b = p->nodes[level];
1015                                 slot = p->slots[level];
1016                         } else if (ins_len < 0) {
1017                                 int sret = balance_level(trans, root, p,
1018                                                          level);
1019                                 if (sret)
1020                                         return sret;
1021                                 b = p->nodes[level];
1022                                 if (!b) {
1023                                         btrfs_release_path(NULL, p);
1024                                         goto again;
1025                                 }
1026                                 slot = p->slots[level];
1027                                 BUG_ON(btrfs_header_nritems(b) == 1);
1028                         }
1029                         /* this is only true while dropping a snapshot */
1030                         if (level == lowest_level)
1031                                 break;
1032                         bytenr = btrfs_node_blockptr(b, slot);
1033                         if (should_reada)
1034                                 reada_for_search(root, p, level, slot);
1035                         b = read_tree_block(root, bytenr,
1036                                             btrfs_level_size(root, level - 1));
1037                 } else {
1038                         p->slots[level] = slot;
1039                         if (ins_len > 0 && btrfs_leaf_free_space(root, b) <
1040                             sizeof(struct btrfs_item) + ins_len) {
1041                                 int sret = split_leaf(trans, root, key,
1042                                                       p, ins_len);
1043                                 BUG_ON(sret > 0);
1044                                 if (sret)
1045                                         return sret;
1046                         }
1047                         return ret;
1048                 }
1049         }
1050         return 1;
1051 }
1052
1053 /*
1054  * adjust the pointers going up the tree, starting at level
1055  * making sure the right key of each node is points to 'key'.
1056  * This is used after shifting pointers to the left, so it stops
1057  * fixing up pointers when a given leaf/node is not in slot 0 of the
1058  * higher levels
1059  *
1060  * If this fails to write a tree block, it returns -1, but continues
1061  * fixing up the blocks in ram so the tree is consistent.
1062  */
1063 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1064                           struct btrfs_root *root, struct btrfs_path *path,
1065                           struct btrfs_disk_key *key, int level)
1066 {
1067         int i;
1068         int ret = 0;
1069         struct extent_buffer *t;
1070
1071         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1072                 int tslot = path->slots[i];
1073                 if (!path->nodes[i])
1074                         break;
1075                 t = path->nodes[i];
1076                 btrfs_set_node_key(t, key, tslot);
1077                 btrfs_mark_buffer_dirty(path->nodes[i]);
1078                 if (tslot != 0)
1079                         break;
1080         }
1081         return ret;
1082 }
1083
1084 /*
1085  * try to push data from one node into the next node left in the
1086  * tree.
1087  *
1088  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1089  * error, and > 0 if there was no room in the left hand block.
1090  */
1091 static int push_node_left(struct btrfs_trans_handle *trans, struct btrfs_root
1092                           *root, struct extent_buffer *dst,
1093                           struct extent_buffer *src)
1094 {
1095         int push_items = 0;
1096         int src_nritems;
1097         int dst_nritems;
1098         int ret = 0;
1099
1100         src_nritems = btrfs_header_nritems(src);
1101         dst_nritems = btrfs_header_nritems(dst);
1102         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1103
1104         if (push_items <= 0) {
1105                 return 1;
1106         }
1107
1108         if (src_nritems < push_items)
1109                 push_items = src_nritems;
1110
1111         copy_extent_buffer(dst, src,
1112                            btrfs_node_key_ptr_offset(dst_nritems),
1113                            btrfs_node_key_ptr_offset(0),
1114                            push_items * sizeof(struct btrfs_key_ptr));
1115
1116         if (push_items < src_nritems) {
1117                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1118                                       btrfs_node_key_ptr_offset(push_items),
1119                                       (src_nritems - push_items) *
1120                                       sizeof(struct btrfs_key_ptr));
1121         }
1122         btrfs_set_header_nritems(src, src_nritems - push_items);
1123         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1124         btrfs_mark_buffer_dirty(src);
1125         btrfs_mark_buffer_dirty(dst);
1126         return ret;
1127 }
1128
1129 /*
1130  * try to push data from one node into the next node right in the
1131  * tree.
1132  *
1133  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1134  * error, and > 0 if there was no room in the right hand block.
1135  *
1136  * this will  only push up to 1/2 the contents of the left node over
1137  */
1138 static int balance_node_right(struct btrfs_trans_handle *trans,
1139                               struct btrfs_root *root,
1140                               struct extent_buffer *dst,
1141                               struct extent_buffer *src)
1142 {
1143         int push_items = 0;
1144         int max_push;
1145         int src_nritems;
1146         int dst_nritems;
1147         int ret = 0;
1148
1149         src_nritems = btrfs_header_nritems(src);
1150         dst_nritems = btrfs_header_nritems(dst);
1151         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1152         if (push_items <= 0)
1153                 return 1;
1154
1155         max_push = src_nritems / 2 + 1;
1156         /* don't try to empty the node */
1157         if (max_push >= src_nritems)
1158                 return 1;
1159
1160         if (max_push < push_items)
1161                 push_items = max_push;
1162
1163         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1164                                       btrfs_node_key_ptr_offset(0),
1165                                       (dst_nritems) *
1166                                       sizeof(struct btrfs_key_ptr));
1167
1168         copy_extent_buffer(dst, src,
1169                            btrfs_node_key_ptr_offset(0),
1170                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1171                            push_items * sizeof(struct btrfs_key_ptr));
1172
1173         btrfs_set_header_nritems(src, src_nritems - push_items);
1174         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1175
1176         btrfs_mark_buffer_dirty(src);
1177         btrfs_mark_buffer_dirty(dst);
1178         return ret;
1179 }
1180
1181 /*
1182  * helper function to insert a new root level in the tree.
1183  * A new node is allocated, and a single item is inserted to
1184  * point to the existing root
1185  *
1186  * returns zero on success or < 0 on failure.
1187  */
1188 static int insert_new_root(struct btrfs_trans_handle *trans,
1189                            struct btrfs_root *root,
1190                            struct btrfs_path *path, int level)
1191 {
1192         struct extent_buffer *lower;
1193         struct extent_buffer *c;
1194         struct btrfs_disk_key lower_key;
1195
1196         BUG_ON(path->nodes[level]);
1197         BUG_ON(path->nodes[level-1] != root->node);
1198
1199         c = btrfs_alloc_free_block(trans, root, root->nodesize,
1200                                    root->node->start, 0);
1201         if (IS_ERR(c))
1202                 return PTR_ERR(c);
1203         memset_extent_buffer(c, 0, 0, root->nodesize);
1204         btrfs_set_header_nritems(c, 1);
1205         btrfs_set_header_level(c, level);
1206         btrfs_set_header_bytenr(c, c->start);
1207         btrfs_set_header_generation(c, trans->transid);
1208         btrfs_set_header_owner(c, root->root_key.objectid);
1209         lower = path->nodes[level-1];
1210
1211         write_extent_buffer(c, root->fs_info->fsid,
1212                             (unsigned long)btrfs_header_fsid(c),
1213                             BTRFS_FSID_SIZE);
1214         if (level == 1)
1215                 btrfs_item_key(lower, &lower_key, 0);
1216         else
1217                 btrfs_node_key(lower, &lower_key, 0);
1218         btrfs_set_node_key(c, &lower_key, 0);
1219         btrfs_set_node_blockptr(c, 0, lower->start);
1220
1221         btrfs_mark_buffer_dirty(c);
1222
1223         /* the super has an extra ref to root->node */
1224         free_extent_buffer(root->node);
1225         root->node = c;
1226         extent_buffer_get(c);
1227         path->nodes[level] = c;
1228         path->slots[level] = 0;
1229         return 0;
1230 }
1231
1232 /*
1233  * worker function to insert a single pointer in a node.
1234  * the node should have enough room for the pointer already
1235  *
1236  * slot and level indicate where you want the key to go, and
1237  * blocknr is the block the key points to.
1238  *
1239  * returns zero on success and < 0 on any error
1240  */
1241 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1242                       *root, struct btrfs_path *path, struct btrfs_disk_key
1243                       *key, u64 bytenr, int slot, int level)
1244 {
1245         struct extent_buffer *lower;
1246         int nritems;
1247
1248         BUG_ON(!path->nodes[level]);
1249         lower = path->nodes[level];
1250         nritems = btrfs_header_nritems(lower);
1251         if (slot > nritems)
1252                 BUG();
1253         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1254                 BUG();
1255         if (slot != nritems) {
1256                 memmove_extent_buffer(lower,
1257                               btrfs_node_key_ptr_offset(slot + 1),
1258                               btrfs_node_key_ptr_offset(slot),
1259                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1260         }
1261         btrfs_set_node_key(lower, key, slot);
1262         btrfs_set_node_blockptr(lower, slot, bytenr);
1263         btrfs_set_header_nritems(lower, nritems + 1);
1264         btrfs_mark_buffer_dirty(lower);
1265         return 0;
1266 }
1267
1268 /*
1269  * split the node at the specified level in path in two.
1270  * The path is corrected to point to the appropriate node after the split
1271  *
1272  * Before splitting this tries to make some room in the node by pushing
1273  * left and right, if either one works, it returns right away.
1274  *
1275  * returns 0 on success and < 0 on failure
1276  */
1277 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1278                       *root, struct btrfs_path *path, int level)
1279 {
1280         struct extent_buffer *c;
1281         struct extent_buffer *split;
1282         struct btrfs_disk_key disk_key;
1283         int mid;
1284         int ret;
1285         int wret;
1286         u32 c_nritems;
1287
1288         c = path->nodes[level];
1289         if (c == root->node) {
1290                 /* trying to split the root, lets make a new one */
1291                 ret = insert_new_root(trans, root, path, level + 1);
1292                 if (ret)
1293                         return ret;
1294         } else {
1295                 ret = push_nodes_for_insert(trans, root, path, level);
1296                 c = path->nodes[level];
1297                 if (!ret && btrfs_header_nritems(c) <
1298                     BTRFS_NODEPTRS_PER_BLOCK(root) - 1)
1299                         return 0;
1300                 if (ret < 0)
1301                         return ret;
1302         }
1303
1304         c_nritems = btrfs_header_nritems(c);
1305         split = btrfs_alloc_free_block(trans, root, root->nodesize,
1306                                        c->start, 0);
1307         if (IS_ERR(split))
1308                 return PTR_ERR(split);
1309
1310         btrfs_set_header_flags(split, btrfs_header_flags(c));
1311         btrfs_set_header_level(split, btrfs_header_level(c));
1312         btrfs_set_header_bytenr(split, split->start);
1313         btrfs_set_header_generation(split, trans->transid);
1314         btrfs_set_header_owner(split, root->root_key.objectid);
1315         write_extent_buffer(split, root->fs_info->fsid,
1316                             (unsigned long)btrfs_header_fsid(split),
1317                             BTRFS_FSID_SIZE);
1318
1319         mid = (c_nritems + 1) / 2;
1320
1321         copy_extent_buffer(split, c,
1322                            btrfs_node_key_ptr_offset(0),
1323                            btrfs_node_key_ptr_offset(mid),
1324                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1325         btrfs_set_header_nritems(split, c_nritems - mid);
1326         btrfs_set_header_nritems(c, mid);
1327         ret = 0;
1328
1329         btrfs_mark_buffer_dirty(c);
1330         btrfs_mark_buffer_dirty(split);
1331
1332         btrfs_node_key(split, &disk_key, 0);
1333         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1334                           path->slots[level + 1] + 1,
1335                           level + 1);
1336         if (wret)
1337                 ret = wret;
1338
1339         if (path->slots[level] >= mid) {
1340                 path->slots[level] -= mid;
1341                 free_extent_buffer(c);
1342                 path->nodes[level] = split;
1343                 path->slots[level + 1] += 1;
1344         } else {
1345                 free_extent_buffer(split);
1346         }
1347         return ret;
1348 }
1349
1350 /*
1351  * how many bytes are required to store the items in a leaf.  start
1352  * and nr indicate which items in the leaf to check.  This totals up the
1353  * space used both by the item structs and the item data
1354  */
1355 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1356 {
1357         int data_len;
1358         int nritems = btrfs_header_nritems(l);
1359         int end = min(nritems, start + nr) - 1;
1360
1361         if (!nr)
1362                 return 0;
1363         data_len = btrfs_item_end_nr(l, start);
1364         data_len = data_len - btrfs_item_offset_nr(l, end);
1365         data_len += sizeof(struct btrfs_item) * nr;
1366         WARN_ON(data_len < 0);
1367         return data_len;
1368 }
1369
1370 /*
1371  * The space between the end of the leaf items and
1372  * the start of the leaf data.  IOW, how much room
1373  * the leaf has left for both items and data
1374  */
1375 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1376 {
1377         int nritems = btrfs_header_nritems(leaf);
1378         int ret;
1379         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1380         if (ret < 0) {
1381                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1382                        ret, BTRFS_LEAF_DATA_SIZE(root),
1383                        leaf_space_used(leaf, 0, nritems), nritems);
1384         }
1385         return ret;
1386 }
1387
1388 /*
1389  * push some data in the path leaf to the right, trying to free up at
1390  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1391  *
1392  * returns 1 if the push failed because the other node didn't have enough
1393  * room, 0 if everything worked out and < 0 if there were major errors.
1394  */
1395 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1396                            *root, struct btrfs_path *path, int data_size)
1397 {
1398         struct extent_buffer *left = path->nodes[0];
1399         struct extent_buffer *right;
1400         struct extent_buffer *upper;
1401         struct btrfs_disk_key disk_key;
1402         int slot;
1403         int i;
1404         int free_space;
1405         int push_space = 0;
1406         int push_items = 0;
1407         struct btrfs_item *item;
1408         u32 left_nritems;
1409         u32 right_nritems;
1410         u32 data_end;
1411         u32 this_item_size;
1412         int ret;
1413
1414         slot = path->slots[1];
1415         if (!path->nodes[1]) {
1416                 return 1;
1417         }
1418         upper = path->nodes[1];
1419         if (slot >= btrfs_header_nritems(upper) - 1)
1420                 return 1;
1421
1422         right = read_tree_block(root, btrfs_node_blockptr(upper, slot + 1),
1423                                 root->leafsize);
1424         free_space = btrfs_leaf_free_space(root, right);
1425         if (free_space < data_size + sizeof(struct btrfs_item)) {
1426                 free_extent_buffer(right);
1427                 return 1;
1428         }
1429
1430         /* cow and double check */
1431         ret = btrfs_cow_block(trans, root, right, upper,
1432                               slot + 1, &right);
1433         if (ret) {
1434                 free_extent_buffer(right);
1435                 return 1;
1436         }
1437         free_space = btrfs_leaf_free_space(root, right);
1438         if (free_space < data_size + sizeof(struct btrfs_item)) {
1439                 free_extent_buffer(right);
1440                 return 1;
1441         }
1442
1443         left_nritems = btrfs_header_nritems(left);
1444         if (left_nritems == 0) {
1445                 free_extent_buffer(right);
1446                 return 1;
1447         }
1448
1449         for (i = left_nritems - 1; i >= 1; i--) {
1450                 item = btrfs_item_nr(left, i);
1451
1452                 if (path->slots[0] == i)
1453                         push_space += data_size + sizeof(*item);
1454
1455                 if (!left->map_token) {
1456                         map_extent_buffer(left, (unsigned long)item,
1457                                         sizeof(struct btrfs_item),
1458                                         &left->map_token, &left->kaddr,
1459                                         &left->map_start, &left->map_len,
1460                                         KM_USER1);
1461                 }
1462
1463                 this_item_size = btrfs_item_size(left, item);
1464                 if (this_item_size + sizeof(*item) + push_space > free_space)
1465                         break;
1466                 push_items++;
1467                 push_space += this_item_size + sizeof(*item);
1468         }
1469         if (left->map_token) {
1470                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1471                 left->map_token = NULL;
1472         }
1473
1474         if (push_items == 0) {
1475                 free_extent_buffer(right);
1476                 return 1;
1477         }
1478
1479         if (push_items == left_nritems)
1480                 WARN_ON(1);
1481
1482         /* push left to right */
1483         right_nritems = btrfs_header_nritems(right);
1484         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1485         push_space -= leaf_data_end(root, left);
1486
1487         /* make room in the right data area */
1488         data_end = leaf_data_end(root, right);
1489         memmove_extent_buffer(right,
1490                               btrfs_leaf_data(right) + data_end - push_space,
1491                               btrfs_leaf_data(right) + data_end,
1492                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1493
1494         /* copy from the left data area */
1495         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1496                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1497                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1498                      push_space);
1499
1500         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1501                               btrfs_item_nr_offset(0),
1502                               right_nritems * sizeof(struct btrfs_item));
1503
1504         /* copy the items from left to right */
1505         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1506                    btrfs_item_nr_offset(left_nritems - push_items),
1507                    push_items * sizeof(struct btrfs_item));
1508
1509         /* update the item pointers */
1510         right_nritems += push_items;
1511         btrfs_set_header_nritems(right, right_nritems);
1512         push_space = BTRFS_LEAF_DATA_SIZE(root);
1513
1514         for (i = 0; i < right_nritems; i++) {
1515                 item = btrfs_item_nr(right, i);
1516                 if (!right->map_token) {
1517                         map_extent_buffer(right, (unsigned long)item,
1518                                         sizeof(struct btrfs_item),
1519                                         &right->map_token, &right->kaddr,
1520                                         &right->map_start, &right->map_len,
1521                                         KM_USER1);
1522                 }
1523                 push_space -= btrfs_item_size(right, item);
1524                 btrfs_set_item_offset(right, item, push_space);
1525         }
1526
1527         if (right->map_token) {
1528                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1529                 right->map_token = NULL;
1530         }
1531         left_nritems -= push_items;
1532         btrfs_set_header_nritems(left, left_nritems);
1533
1534         btrfs_mark_buffer_dirty(left);
1535         btrfs_mark_buffer_dirty(right);
1536
1537         btrfs_item_key(right, &disk_key, 0);
1538         btrfs_set_node_key(upper, &disk_key, slot + 1);
1539         btrfs_mark_buffer_dirty(upper);
1540
1541         /* then fixup the leaf pointer in the path */
1542         if (path->slots[0] >= left_nritems) {
1543                 path->slots[0] -= left_nritems;
1544                 free_extent_buffer(path->nodes[0]);
1545                 path->nodes[0] = right;
1546                 path->slots[1] += 1;
1547         } else {
1548                 free_extent_buffer(right);
1549         }
1550         return 0;
1551 }
1552 /*
1553  * push some data in the path leaf to the left, trying to free up at
1554  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1555  */
1556 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1557                           *root, struct btrfs_path *path, int data_size)
1558 {
1559         struct btrfs_disk_key disk_key;
1560         struct extent_buffer *right = path->nodes[0];
1561         struct extent_buffer *left;
1562         int slot;
1563         int i;
1564         int free_space;
1565         int push_space = 0;
1566         int push_items = 0;
1567         struct btrfs_item *item;
1568         u32 old_left_nritems;
1569         u32 right_nritems;
1570         int ret = 0;
1571         int wret;
1572         u32 this_item_size;
1573         u32 old_left_item_size;
1574
1575         slot = path->slots[1];
1576         if (slot == 0)
1577                 return 1;
1578         if (!path->nodes[1])
1579                 return 1;
1580
1581         left = read_tree_block(root, btrfs_node_blockptr(path->nodes[1],
1582                                slot - 1), root->leafsize);
1583         free_space = btrfs_leaf_free_space(root, left);
1584         if (free_space < data_size + sizeof(struct btrfs_item)) {
1585                 free_extent_buffer(left);
1586                 return 1;
1587         }
1588
1589         /* cow and double check */
1590         ret = btrfs_cow_block(trans, root, left,
1591                               path->nodes[1], slot - 1, &left);
1592         if (ret) {
1593                 /* we hit -ENOSPC, but it isn't fatal here */
1594                 free_extent_buffer(left);
1595                 return 1;
1596         }
1597         free_space = btrfs_leaf_free_space(root, left);
1598         if (free_space < data_size + sizeof(struct btrfs_item)) {
1599                 free_extent_buffer(left);
1600                 return 1;
1601         }
1602
1603         right_nritems = btrfs_header_nritems(right);
1604         if (right_nritems == 0) {
1605                 free_extent_buffer(left);
1606                 return 1;
1607         }
1608
1609         for (i = 0; i < right_nritems - 1; i++) {
1610                 item = btrfs_item_nr(right, i);
1611                 if (!right->map_token) {
1612                         map_extent_buffer(right, (unsigned long)item,
1613                                         sizeof(struct btrfs_item),
1614                                         &right->map_token, &right->kaddr,
1615                                         &right->map_start, &right->map_len,
1616                                         KM_USER1);
1617                 }
1618
1619                 if (path->slots[0] == i)
1620                         push_space += data_size + sizeof(*item);
1621
1622                 this_item_size = btrfs_item_size(right, item);
1623                 if (this_item_size + sizeof(*item) + push_space > free_space)
1624                         break;
1625
1626                 push_items++;
1627                 push_space += this_item_size + sizeof(*item);
1628         }
1629
1630         if (right->map_token) {
1631                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1632                 right->map_token = NULL;
1633         }
1634
1635         if (push_items == 0) {
1636                 free_extent_buffer(left);
1637                 return 1;
1638         }
1639         if (push_items == btrfs_header_nritems(right))
1640                 WARN_ON(1);
1641
1642         /* push data from right to left */
1643         copy_extent_buffer(left, right,
1644                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1645                            btrfs_item_nr_offset(0),
1646                            push_items * sizeof(struct btrfs_item));
1647
1648         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1649                      btrfs_item_offset_nr(right, push_items -1);
1650
1651         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1652                      leaf_data_end(root, left) - push_space,
1653                      btrfs_leaf_data(right) +
1654                      btrfs_item_offset_nr(right, push_items - 1),
1655                      push_space);
1656         old_left_nritems = btrfs_header_nritems(left);
1657         BUG_ON(old_left_nritems < 0);
1658
1659         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1660         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1661                 u32 ioff;
1662
1663                 item = btrfs_item_nr(left, i);
1664                 if (!left->map_token) {
1665                         map_extent_buffer(left, (unsigned long)item,
1666                                         sizeof(struct btrfs_item),
1667                                         &left->map_token, &left->kaddr,
1668                                         &left->map_start, &left->map_len,
1669                                         KM_USER1);
1670                 }
1671
1672                 ioff = btrfs_item_offset(left, item);
1673                 btrfs_set_item_offset(left, item,
1674                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1675         }
1676         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1677         if (left->map_token) {
1678                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1679                 left->map_token = NULL;
1680         }
1681
1682         /* fixup right node */
1683         push_space = btrfs_item_offset_nr(right, push_items - 1) -
1684                                           leaf_data_end(root, right);
1685         memmove_extent_buffer(right, btrfs_leaf_data(right) +
1686                               BTRFS_LEAF_DATA_SIZE(root) - push_space,
1687                               btrfs_leaf_data(right) +
1688                               leaf_data_end(root, right), push_space);
1689
1690         memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1691                               btrfs_item_nr_offset(push_items),
1692                              (btrfs_header_nritems(right) - push_items) *
1693                              sizeof(struct btrfs_item));
1694
1695         right_nritems = btrfs_header_nritems(right) - push_items;
1696         btrfs_set_header_nritems(right, right_nritems);
1697         push_space = BTRFS_LEAF_DATA_SIZE(root);
1698
1699         for (i = 0; i < right_nritems; i++) {
1700                 item = btrfs_item_nr(right, i);
1701
1702                 if (!right->map_token) {
1703                         map_extent_buffer(right, (unsigned long)item,
1704                                         sizeof(struct btrfs_item),
1705                                         &right->map_token, &right->kaddr,
1706                                         &right->map_start, &right->map_len,
1707                                         KM_USER1);
1708                 }
1709
1710                 push_space = push_space - btrfs_item_size(right, item);
1711                 btrfs_set_item_offset(right, item, push_space);
1712         }
1713         if (right->map_token) {
1714                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1715                 right->map_token = NULL;
1716         }
1717
1718         btrfs_mark_buffer_dirty(left);
1719         btrfs_mark_buffer_dirty(right);
1720
1721         btrfs_item_key(right, &disk_key, 0);
1722         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1723         if (wret)
1724                 ret = wret;
1725
1726         /* then fixup the leaf pointer in the path */
1727         if (path->slots[0] < push_items) {
1728                 path->slots[0] += old_left_nritems;
1729                 free_extent_buffer(path->nodes[0]);
1730                 path->nodes[0] = left;
1731                 path->slots[1] -= 1;
1732         } else {
1733                 free_extent_buffer(left);
1734                 path->slots[0] -= push_items;
1735         }
1736         BUG_ON(path->slots[0] < 0);
1737         return ret;
1738 }
1739
1740 /*
1741  * split the path's leaf in two, making sure there is at least data_size
1742  * available for the resulting leaf level of the path.
1743  *
1744  * returns 0 if all went well and < 0 on failure.
1745  */
1746 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
1747                       *root, struct btrfs_key *ins_key,
1748                       struct btrfs_path *path, int data_size)
1749 {
1750         struct extent_buffer *l;
1751         u32 nritems;
1752         int mid;
1753         int slot;
1754         struct extent_buffer *right;
1755         int space_needed = data_size + sizeof(struct btrfs_item);
1756         int data_copy_size;
1757         int rt_data_off;
1758         int i;
1759         int ret = 0;
1760         int wret;
1761         int double_split = 0;
1762         struct btrfs_disk_key disk_key;
1763
1764         /* first try to make some room by pushing left and right */
1765         wret = push_leaf_left(trans, root, path, data_size);
1766         if (wret < 0) {
1767                 return wret;
1768         }
1769         if (wret) {
1770                 wret = push_leaf_right(trans, root, path, data_size);
1771                 if (wret < 0)
1772                         return wret;
1773         }
1774         l = path->nodes[0];
1775
1776         /* did the pushes work? */
1777         if (btrfs_leaf_free_space(root, l) >=
1778             sizeof(struct btrfs_item) + data_size) {
1779                 return 0;
1780         }
1781
1782         if (!path->nodes[1]) {
1783                 ret = insert_new_root(trans, root, path, 1);
1784                 if (ret)
1785                         return ret;
1786         }
1787         slot = path->slots[0];
1788         nritems = btrfs_header_nritems(l);
1789         mid = (nritems + 1)/ 2;
1790
1791         right = btrfs_alloc_free_block(trans, root, root->leafsize,
1792                                        l->start, 0);
1793         if (IS_ERR(right))
1794                 return PTR_ERR(right);
1795
1796         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
1797         btrfs_set_header_bytenr(right, right->start);
1798         btrfs_set_header_generation(right, trans->transid);
1799         btrfs_set_header_owner(right, root->root_key.objectid);
1800         btrfs_set_header_level(right, 0);
1801         write_extent_buffer(right, root->fs_info->fsid,
1802                             (unsigned long)btrfs_header_fsid(right),
1803                             BTRFS_FSID_SIZE);
1804
1805         if (mid <= slot) {
1806                 if (nritems == 1 ||
1807                     leaf_space_used(l, mid, nritems - mid) + space_needed >
1808                         BTRFS_LEAF_DATA_SIZE(root)) {
1809                         if (slot >= nritems) {
1810                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1811                                 btrfs_set_header_nritems(right, 0);
1812                                 wret = insert_ptr(trans, root, path,
1813                                                   &disk_key, right->start,
1814                                                   path->slots[1] + 1, 1);
1815                                 if (wret)
1816                                         ret = wret;
1817                                 free_extent_buffer(path->nodes[0]);
1818                                 path->nodes[0] = right;
1819                                 path->slots[0] = 0;
1820                                 path->slots[1] += 1;
1821                                 return ret;
1822                         }
1823                         mid = slot;
1824                         if (mid != nritems &&
1825                             leaf_space_used(l, mid, nritems - mid) +
1826                             space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
1827                                 double_split = 1;
1828                         }
1829                 }
1830         } else {
1831                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
1832                         BTRFS_LEAF_DATA_SIZE(root)) {
1833                         if (slot == 0) {
1834                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1835                                 btrfs_set_header_nritems(right, 0);
1836                                 wret = insert_ptr(trans, root, path,
1837                                                   &disk_key,
1838                                                   right->start,
1839                                                   path->slots[1], 1);
1840                                 if (wret)
1841                                         ret = wret;
1842                                 free_extent_buffer(path->nodes[0]);
1843                                 path->nodes[0] = right;
1844                                 path->slots[0] = 0;
1845                                 if (path->slots[1] == 0) {
1846                                         wret = fixup_low_keys(trans, root,
1847                                                    path, &disk_key, 1);
1848                                         if (wret)
1849                                                 ret = wret;
1850                                 }
1851                                 return ret;
1852                         }
1853                         mid = slot;
1854                         double_split = 1;
1855                 }
1856         }
1857         nritems = nritems - mid;
1858         btrfs_set_header_nritems(right, nritems);
1859         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
1860
1861         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
1862                            btrfs_item_nr_offset(mid),
1863                            nritems * sizeof(struct btrfs_item));
1864
1865         copy_extent_buffer(right, l,
1866                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
1867                      data_copy_size, btrfs_leaf_data(l) +
1868                      leaf_data_end(root, l), data_copy_size);
1869
1870         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
1871                       btrfs_item_end_nr(l, mid);
1872
1873         for (i = 0; i < nritems; i++) {
1874                 struct btrfs_item *item = btrfs_item_nr(right, i);
1875                 u32 ioff;
1876
1877                 if (!right->map_token) {
1878                         map_extent_buffer(right, (unsigned long)item,
1879                                         sizeof(struct btrfs_item),
1880                                         &right->map_token, &right->kaddr,
1881                                         &right->map_start, &right->map_len,
1882                                         KM_USER1);
1883                 }
1884
1885                 ioff = btrfs_item_offset(right, item);
1886                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
1887         }
1888
1889         if (right->map_token) {
1890                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1891                 right->map_token = NULL;
1892         }
1893
1894         btrfs_set_header_nritems(l, mid);
1895         ret = 0;
1896         btrfs_item_key(right, &disk_key, 0);
1897         wret = insert_ptr(trans, root, path, &disk_key, right->start,
1898                           path->slots[1] + 1, 1);
1899         if (wret)
1900                 ret = wret;
1901
1902         btrfs_mark_buffer_dirty(right);
1903         btrfs_mark_buffer_dirty(l);
1904         BUG_ON(path->slots[0] != slot);
1905
1906         if (mid <= slot) {
1907                 free_extent_buffer(path->nodes[0]);
1908                 path->nodes[0] = right;
1909                 path->slots[0] -= mid;
1910                 path->slots[1] += 1;
1911         } else
1912                 free_extent_buffer(right);
1913
1914         BUG_ON(path->slots[0] < 0);
1915
1916         if (!double_split) {
1917                 return ret;
1918         }
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         if (slot >= nritems) {
2056                 btrfs_print_leaf(root, leaf);
2057                 printk("slot %d too large, nritems %d\n", slot, nritems);
2058                 BUG_ON(1);
2059         }
2060
2061         /*
2062          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2063          */
2064         /* first correct the data pointers */
2065         for (i = slot; i < nritems; i++) {
2066                 u32 ioff;
2067                 item = btrfs_item_nr(leaf, i);
2068
2069                 if (!leaf->map_token) {
2070                         map_extent_buffer(leaf, (unsigned long)item,
2071                                         sizeof(struct btrfs_item),
2072                                         &leaf->map_token, &leaf->kaddr,
2073                                         &leaf->map_start, &leaf->map_len,
2074                                         KM_USER1);
2075                 }
2076                 ioff = btrfs_item_offset(leaf, item);
2077                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2078         }
2079
2080         if (leaf->map_token) {
2081                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2082                 leaf->map_token = NULL;
2083         }
2084
2085         /* shift the data */
2086         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2087                       data_end - data_size, btrfs_leaf_data(leaf) +
2088                       data_end, old_data - data_end);
2089
2090         data_end = old_data;
2091         old_size = btrfs_item_size_nr(leaf, slot);
2092         item = btrfs_item_nr(leaf, slot);
2093         btrfs_set_item_size(leaf, item, old_size + data_size);
2094         btrfs_mark_buffer_dirty(leaf);
2095
2096         ret = 0;
2097         if (btrfs_leaf_free_space(root, leaf) < 0) {
2098                 btrfs_print_leaf(root, leaf);
2099                 BUG();
2100         }
2101         return ret;
2102 }
2103
2104 /*
2105  * Given a key and some data, insert an item into the tree.
2106  * This does all the path init required, making room in the tree if needed.
2107  */
2108 int btrfs_insert_empty_item(struct btrfs_trans_handle *trans,
2109                             struct btrfs_root *root,
2110                             struct btrfs_path *path,
2111                             struct btrfs_key *cpu_key, u32 data_size)
2112 {
2113         struct extent_buffer *leaf;
2114         struct btrfs_item *item;
2115         int ret = 0;
2116         int slot;
2117         int slot_orig;
2118         u32 nritems;
2119         unsigned int data_end;
2120         struct btrfs_disk_key disk_key;
2121
2122         btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2123
2124         /* create a root if there isn't one */
2125         if (!root->node)
2126                 BUG();
2127
2128         ret = btrfs_search_slot(trans, root, cpu_key, path, data_size, 1);
2129         if (ret == 0) {
2130                 return -EEXIST;
2131         }
2132         if (ret < 0)
2133                 goto out;
2134
2135         slot_orig = path->slots[0];
2136         leaf = path->nodes[0];
2137
2138         nritems = btrfs_header_nritems(leaf);
2139         data_end = leaf_data_end(root, leaf);
2140
2141         if (btrfs_leaf_free_space(root, leaf) <
2142             sizeof(struct btrfs_item) + data_size) {
2143                 btrfs_print_leaf(root, leaf);
2144                 printk("not enough freespace need %u have %d\n",
2145                        data_size, btrfs_leaf_free_space(root, leaf));
2146                 BUG();
2147         }
2148
2149         slot = path->slots[0];
2150         BUG_ON(slot < 0);
2151
2152         if (slot != nritems) {
2153                 int i;
2154                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2155
2156                 if (old_data < data_end) {
2157                         btrfs_print_leaf(root, leaf);
2158                         printk("slot %d old_data %d data_end %d\n",
2159                                slot, old_data, data_end);
2160                         BUG_ON(1);
2161                 }
2162                 /*
2163                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2164                  */
2165                 /* first correct the data pointers */
2166                 WARN_ON(leaf->map_token);
2167                 for (i = slot; i < nritems; i++) {
2168                         u32 ioff;
2169
2170                         item = btrfs_item_nr(leaf, i);
2171                         if (!leaf->map_token) {
2172                                 map_extent_buffer(leaf, (unsigned long)item,
2173                                         sizeof(struct btrfs_item),
2174                                         &leaf->map_token, &leaf->kaddr,
2175                                         &leaf->map_start, &leaf->map_len,
2176                                         KM_USER1);
2177                         }
2178
2179                         ioff = btrfs_item_offset(leaf, item);
2180                         btrfs_set_item_offset(leaf, item, ioff - data_size);
2181                 }
2182                 if (leaf->map_token) {
2183                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2184                         leaf->map_token = NULL;
2185                 }
2186
2187                 /* shift the items */
2188                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
2189                               btrfs_item_nr_offset(slot),
2190                               (nritems - slot) * sizeof(struct btrfs_item));
2191
2192                 /* shift the data */
2193                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2194                               data_end - data_size, btrfs_leaf_data(leaf) +
2195                               data_end, old_data - data_end);
2196                 data_end = old_data;
2197         }
2198
2199         /* setup the item for the new data */
2200         btrfs_set_item_key(leaf, &disk_key, slot);
2201         item = btrfs_item_nr(leaf, slot);
2202         btrfs_set_item_offset(leaf, item, data_end - data_size);
2203         btrfs_set_item_size(leaf, item, data_size);
2204         btrfs_set_header_nritems(leaf, nritems + 1);
2205         btrfs_mark_buffer_dirty(leaf);
2206
2207         ret = 0;
2208         if (slot == 0)
2209                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2210
2211         if (btrfs_leaf_free_space(root, leaf) < 0) {
2212                 btrfs_print_leaf(root, leaf);
2213                 BUG();
2214         }
2215 out:
2216         return ret;
2217 }
2218
2219 /*
2220  * Given a key and some data, insert an item into the tree.
2221  * This does all the path init required, making room in the tree if needed.
2222  */
2223 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2224                       *root, struct btrfs_key *cpu_key, void *data, u32
2225                       data_size)
2226 {
2227         int ret = 0;
2228         struct btrfs_path *path;
2229         struct extent_buffer *leaf;
2230         unsigned long ptr;
2231
2232         path = btrfs_alloc_path();
2233         BUG_ON(!path);
2234         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2235         if (!ret) {
2236                 leaf = path->nodes[0];
2237                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2238                 write_extent_buffer(leaf, data, ptr, data_size);
2239                 btrfs_mark_buffer_dirty(leaf);
2240         }
2241         btrfs_free_path(path);
2242         return ret;
2243 }
2244
2245 /*
2246  * delete the pointer from a given node.
2247  *
2248  * If the delete empties a node, the node is removed from the tree,
2249  * continuing all the way the root if required.  The root is converted into
2250  * a leaf if all the nodes are emptied.
2251  */
2252 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2253                    struct btrfs_path *path, int level, int slot)
2254 {
2255         struct extent_buffer *parent = path->nodes[level];
2256         u32 nritems;
2257         int ret = 0;
2258         int wret;
2259
2260         nritems = btrfs_header_nritems(parent);
2261         if (slot != nritems -1) {
2262                 memmove_extent_buffer(parent,
2263                               btrfs_node_key_ptr_offset(slot),
2264                               btrfs_node_key_ptr_offset(slot + 1),
2265                               sizeof(struct btrfs_key_ptr) *
2266                               (nritems - slot - 1));
2267         }
2268         nritems--;
2269         btrfs_set_header_nritems(parent, nritems);
2270         if (nritems == 0 && parent == root->node) {
2271                 BUG_ON(btrfs_header_level(root->node) != 1);
2272                 /* just turn the root into a leaf and break */
2273                 btrfs_set_header_level(root->node, 0);
2274         } else if (slot == 0) {
2275                 struct btrfs_disk_key disk_key;
2276
2277                 btrfs_node_key(parent, &disk_key, 0);
2278                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2279                 if (wret)
2280                         ret = wret;
2281         }
2282         btrfs_mark_buffer_dirty(parent);
2283         return ret;
2284 }
2285
2286 /*
2287  * delete the item at the leaf level in path.  If that empties
2288  * the leaf, remove it from the tree
2289  */
2290 int btrfs_del_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2291                    struct btrfs_path *path)
2292 {
2293         int slot;
2294         struct extent_buffer *leaf;
2295         struct btrfs_item *item;
2296         int doff;
2297         int dsize;
2298         int ret = 0;
2299         int wret;
2300         u32 nritems;
2301
2302         leaf = path->nodes[0];
2303         slot = path->slots[0];
2304         doff = btrfs_item_offset_nr(leaf, slot);
2305         dsize = btrfs_item_size_nr(leaf, slot);
2306         nritems = btrfs_header_nritems(leaf);
2307
2308         if (slot != nritems - 1) {
2309                 int i;
2310                 int data_end = leaf_data_end(root, leaf);
2311
2312                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2313                               data_end + dsize,
2314                               btrfs_leaf_data(leaf) + data_end,
2315                               doff - data_end);
2316
2317                 for (i = slot + 1; i < nritems; i++) {
2318                         u32 ioff;
2319
2320                         item = btrfs_item_nr(leaf, i);
2321                         if (!leaf->map_token) {
2322                                 map_extent_buffer(leaf, (unsigned long)item,
2323                                         sizeof(struct btrfs_item),
2324                                         &leaf->map_token, &leaf->kaddr,
2325                                         &leaf->map_start, &leaf->map_len,
2326                                         KM_USER1);
2327                         }
2328                         ioff = btrfs_item_offset(leaf, item);
2329                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2330                 }
2331
2332                 if (leaf->map_token) {
2333                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2334                         leaf->map_token = NULL;
2335                 }
2336
2337                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2338                               btrfs_item_nr_offset(slot + 1),
2339                               sizeof(struct btrfs_item) *
2340                               (nritems - slot - 1));
2341         }
2342         btrfs_set_header_nritems(leaf, nritems - 1);
2343         nritems--;
2344
2345         /* delete the leaf if we've emptied it */
2346         if (nritems == 0) {
2347                 if (leaf == root->node) {
2348                         btrfs_set_header_level(leaf, 0);
2349                 } else {
2350                         clean_tree_block(trans, root, leaf);
2351                         wait_on_tree_block_writeback(root, leaf);
2352                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
2353                         if (wret)
2354                                 ret = wret;
2355                         wret = btrfs_free_extent(trans, root,
2356                                                  leaf->start, leaf->len, 1);
2357                         if (wret)
2358                                 ret = wret;
2359                 }
2360         } else {
2361                 int used = leaf_space_used(leaf, 0, nritems);
2362                 if (slot == 0) {
2363                         struct btrfs_disk_key disk_key;
2364
2365                         btrfs_item_key(leaf, &disk_key, 0);
2366                         wret = fixup_low_keys(trans, root, path,
2367                                               &disk_key, 1);
2368                         if (wret)
2369                                 ret = wret;
2370                 }
2371
2372                 /* delete the leaf if it is mostly empty */
2373                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 3) {
2374                         /* push_leaf_left fixes the path.
2375                          * make sure the path still points to our leaf
2376                          * for possible call to del_ptr below
2377                          */
2378                         slot = path->slots[1];
2379                         extent_buffer_get(leaf);
2380
2381                         wret = push_leaf_left(trans, root, path, 1);
2382                         if (wret < 0 && wret != -ENOSPC)
2383                                 ret = wret;
2384
2385                         if (path->nodes[0] == leaf &&
2386                             btrfs_header_nritems(leaf)) {
2387                                 wret = push_leaf_right(trans, root, path, 1);
2388                                 if (wret < 0 && wret != -ENOSPC)
2389                                         ret = wret;
2390                         }
2391
2392                         if (btrfs_header_nritems(leaf) == 0) {
2393                                 u64 bytenr = leaf->start;
2394                                 u32 blocksize = leaf->len;
2395
2396                                 clean_tree_block(trans, root, leaf);
2397                                 wait_on_tree_block_writeback(root, leaf);
2398
2399                                 wret = del_ptr(trans, root, path, 1, slot);
2400                                 if (wret)
2401                                         ret = wret;
2402
2403                                 free_extent_buffer(leaf);
2404                                 wret = btrfs_free_extent(trans, root, bytenr,
2405                                                          blocksize, 1);
2406                                 if (wret)
2407                                         ret = wret;
2408                         } else {
2409                                 btrfs_mark_buffer_dirty(leaf);
2410                                 free_extent_buffer(leaf);
2411                         }
2412                 } else {
2413                         btrfs_mark_buffer_dirty(leaf);
2414                 }
2415         }
2416         return ret;
2417 }
2418
2419 /*
2420  * walk up the tree as far as required to find the next leaf.
2421  * returns 0 if it found something or 1 if there are no greater leaves.
2422  * returns < 0 on io errors.
2423  */
2424 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2425 {
2426         int slot;
2427         int level = 1;
2428         u64 bytenr;
2429         struct extent_buffer *c;
2430         struct extent_buffer *next = NULL;
2431
2432         while(level < BTRFS_MAX_LEVEL) {
2433                 if (!path->nodes[level])
2434                         return 1;
2435
2436                 slot = path->slots[level] + 1;
2437                 c = path->nodes[level];
2438                 if (slot >= btrfs_header_nritems(c)) {
2439                         level++;
2440                         continue;
2441                 }
2442
2443                 bytenr = btrfs_node_blockptr(c, slot);
2444                 if (next)
2445                         free_extent_buffer(next);
2446
2447                 if (path->reada)
2448                         reada_for_search(root, path, level, slot);
2449
2450                 next = read_tree_block(root, bytenr,
2451                                        btrfs_level_size(root, level -1));
2452                 break;
2453         }
2454         path->slots[level] = slot;
2455         while(1) {
2456                 level--;
2457                 c = path->nodes[level];
2458                 free_extent_buffer(c);
2459                 path->nodes[level] = next;
2460                 path->slots[level] = 0;
2461                 if (!level)
2462                         break;
2463                 if (path->reada)
2464                         reada_for_search(root, path, level, 0);
2465                 next = read_tree_block(root, btrfs_node_blockptr(next, 0),
2466                                        btrfs_level_size(root, level - 1));
2467         }
2468         return 0;
2469 }