SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/btrfs/ctree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/highmem.h>
  20 #include "ctree.h"
  21 #include "disk-io.h"
  22 #include "transaction.h"
  23 #include "print-tree.h"
  24
  25 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  26                       *root, struct btrfs_path *path, int level);
  27 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  28                       *root, struct btrfs_key *ins_key,
  29                       struct btrfs_path *path, int data_size);
  30 static int push_node_left(struct btrfs_trans_handle *trans,
  31                           struct btrfs_root *root, struct extent_buffer *dst,
  32                           struct extent_buffer *src);
  33 static int balance_node_right(struct btrfs_trans_handle *trans,
  34                               struct btrfs_root *root,
  35                               struct extent_buffer *dst_buf,
  36                               struct extent_buffer *src_buf);
  37 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  38                    struct btrfs_path *path, int level, int slot);
  39
  40 inline void btrfs_init_path(struct btrfs_path *p)
  41 {
  42         memset(p, 0, sizeof(*p));
  43 }
  44
  45 struct btrfs_path *btrfs_alloc_path(void)
  46 {
  47         struct btrfs_path *path;
  48         path = kmem_cache_alloc(btrfs_path_cachep, GFP_NOFS);
  49         if (path) {
  50                 btrfs_init_path(path);
  51                 path->reada = 1;
  52         }
  53         return path;
  54 }
  55
  56 void btrfs_free_path(struct btrfs_path *p)
  57 {
  58         btrfs_release_path(NULL, p);
  59         kmem_cache_free(btrfs_path_cachep, p);
  60 }
  61
  62 void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
  63 {
  64         int i;
  65         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  66                 if (!p->nodes[i])
  67                         break;
  68                 free_extent_buffer(p->nodes[i]);
  69         }
  70         memset(p, 0, sizeof(*p));
  71 }
  72
  73 static int __btrfs_cow_block(struct btrfs_trans_handle *trans,
  74                              struct btrfs_root *root,
  75                              struct extent_buffer *buf,
  76                              struct extent_buffer *parent, int parent_slot,
  77                              struct extent_buffer **cow_ret,
  78                              u64 search_start, u64 empty_size)
  79 {
  80         struct extent_buffer *cow;
  81         int ret = 0;
  82         int different_trans = 0;
  83
  84         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
  85
  86         cow = btrfs_alloc_free_block(trans, root, buf->len,
  87                                      search_start, empty_size);
  88         if (IS_ERR(cow))
  89                 return PTR_ERR(cow);
  90
  91         copy_extent_buffer(cow, buf, 0, 0, cow->len);
  92         btrfs_set_header_bytenr(cow, cow->start);
  93         btrfs_set_header_generation(cow, trans->transid);
  94         btrfs_set_header_owner(cow, root->root_key.objectid);
  95
  96         WARN_ON(btrfs_header_generation(buf) > trans->transid);
  97         if (btrfs_header_generation(buf) != trans->transid) {
  98                 different_trans = 1;
  99                 ret = btrfs_inc_ref(trans, root, buf);
 100                 if (ret)
 101                         return ret;
 102         } else {
 103                 clean_tree_block(trans, root, buf);
 104         }
 105
 106         if (buf == root->node) {
 107                 root->node = cow;
 108                 extent_buffer_get(cow);
 109                 if (buf != root->commit_root) {
 110                         btrfs_free_extent(trans, root, buf->start,
 111                                           buf->len, 1);
 112                 }
 113                 free_extent_buffer(buf);
 114         } else {
 115                 btrfs_set_node_blockptr(parent, parent_slot,
 116                                         cow->start);
 117                 btrfs_mark_buffer_dirty(parent);
 118                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 119                 btrfs_free_extent(trans, root, buf->start, buf->len, 1);
 120         }
 121         free_extent_buffer(buf);
 122         btrfs_mark_buffer_dirty(cow);
 123         *cow_ret = cow;
 124         return 0;
 125 }
 126
 127 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 128                     struct btrfs_root *root, struct extent_buffer *buf,
 129                     struct extent_buffer *parent, int parent_slot,
 130                     struct extent_buffer **cow_ret)
 131 {
 132         u64 search_start;
 133         int ret;
 134         if (trans->transaction != root->fs_info->running_transaction) {
 135                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 136                        root->fs_info->running_transaction->transid);
 137                 WARN_ON(1);
 138         }
 139         if (trans->transid != root->fs_info->generation) {
 140                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 141                        root->fs_info->generation);
 142                 WARN_ON(1);
 143         }
 144         if (btrfs_header_generation(buf) == trans->transid) {
 145                 *cow_ret = buf;
 146                 return 0;
 147         }
 148
 149         search_start = buf->start & ~((u64)BTRFS_BLOCK_GROUP_SIZE - 1);
 150         ret = __btrfs_cow_block(trans, root, buf, parent,
 151                                  parent_slot, cow_ret, search_start, 0);
 152         return ret;
 153 }
 154
 155 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 156 {
 157         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 158                 return 1;
 159         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 160                 return 1;
 161         return 0;
 162 }
 163
 164 static int should_defrag_leaf(struct extent_buffer *leaf)
 165 {
 166         struct btrfs_key key;
 167         u32 nritems;
 168
 169         if (btrfs_buffer_defrag(leaf))
 170                 return 1;
 171
 172         nritems = btrfs_header_nritems(leaf);
 173         if (nritems == 0)
 174                 return 0;
 175
 176         btrfs_item_key_to_cpu(leaf, &key, 0);
 177         if (key.type == BTRFS_DIR_ITEM_KEY)
 178                 return 1;
 179
 180
 181         btrfs_item_key_to_cpu(leaf, &key, nritems - 1);
 182         if (key.type == BTRFS_DIR_ITEM_KEY)
 183                 return 1;
 184         if (nritems > 4) {
 185                 btrfs_item_key_to_cpu(leaf, &key, nritems / 2);
 186                 if (key.type == BTRFS_DIR_ITEM_KEY)
 187                         return 1;
 188         }
 189         return 0;
 190 }
 191
 192 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 193                        struct btrfs_root *root, struct extent_buffer *parent,
 194                        int cache_only, u64 *last_ret)
 195 {
 196         struct extent_buffer *cur;
 197         struct extent_buffer *tmp;
 198         u64 blocknr;
 199         u64 search_start = *last_ret;
 200         u64 last_block = 0;
 201         u64 other;
 202         u32 parent_nritems;
 203         int start_slot;
 204         int end_slot;
 205         int i;
 206         int err = 0;
 207         int parent_level;
 208         int uptodate;
 209         u32 blocksize;
 210
 211         if (trans->transaction != root->fs_info->running_transaction) {
 212                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 213                        root->fs_info->running_transaction->transid);
 214                 WARN_ON(1);
 215         }
 216         if (trans->transid != root->fs_info->generation) {
 217                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 218                        root->fs_info->generation);
 219                 WARN_ON(1);
 220         }
 221         if (btrfs_buffer_defrag_done(parent))
 222                 return 0;
 223
 224         parent_nritems = btrfs_header_nritems(parent);
 225         parent_level = btrfs_header_level(parent);
 226         blocksize = btrfs_level_size(root, parent_level - 1);
 227
 228         start_slot = 0;
 229         end_slot = parent_nritems;
 230
 231         if (parent_nritems == 1)
 232                 return 0;
 233
 234         for (i = start_slot; i < end_slot; i++) {
 235                 int close = 1;
 236                 blocknr = btrfs_node_blockptr(parent, i);
 237                 if (last_block == 0)
 238                         last_block = blocknr;
 239                 if (i > 0) {
 240                         other = btrfs_node_blockptr(parent, i - 1);
 241                         close = close_blocks(blocknr, other, blocksize);
 242                 }
 243                 if (close && i < end_slot - 1) {
 244                         other = btrfs_node_blockptr(parent, i + 1);
 245                         close = close_blocks(blocknr, other, blocksize);
 246                 }
 247                 if (close) {
 248                         last_block = blocknr;
 249                         continue;
 250                 }
 251
 252                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
 253                 if (cur)
 254                         uptodate = btrfs_buffer_uptodate(cur);
 255                 else
 256                         uptodate = 0;
 257                 if (!cur || !uptodate ||
 258                     (parent_level != 1 && !btrfs_buffer_defrag(cur)) ||
 259                     (parent_level == 1 && !should_defrag_leaf(cur))) {
 260                         if (cache_only) {
 261                                 free_extent_buffer(cur);
 262                                 continue;
 263                         }
 264                         if (!cur) {
 265                                 cur = read_tree_block(root, blocknr,
 266                                                          blocksize);
 267                         } else if (!uptodate) {
 268                                 btrfs_read_buffer(cur);
 269                         }
 270                 }
 271                 if (search_start == 0)
 272                         search_start = last_block;
 273
 274                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 275                                         &tmp, search_start,
 276                                         min(16 * blocksize,
 277                                             (end_slot - i) * blocksize));
 278                 if (err) {
 279                         free_extent_buffer(cur);
 280                         break;
 281                 }
 282                 search_start = tmp->start;
 283                 *last_ret = search_start;
 284                 if (parent_level == 1)
 285                         btrfs_clear_buffer_defrag(tmp);
 286                 btrfs_set_buffer_defrag_done(tmp);
 287                 free_extent_buffer(tmp);
 288         }
 289         return err;
 290 }
 291
 292 /*
 293  * The leaf data grows from end-to-front in the node.
 294  * this returns the address of the start of the last item,
 295  * which is the stop of the leaf data stack
 296  */
 297 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 298                                          struct extent_buffer *leaf)
 299 {
 300         u32 nr = btrfs_header_nritems(leaf);
 301         if (nr == 0)
 302                 return BTRFS_LEAF_DATA_SIZE(root);
 303         return btrfs_item_offset_nr(leaf, nr - 1);
 304 }
 305
 306 /*
 307  * compare two keys in a memcmp fashion
 308  */
 309 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 310 {
 311         struct btrfs_key k1;
 312
 313         btrfs_disk_key_to_cpu(&k1, disk);
 314
 315         if (k1.objectid > k2->objectid)
 316                 return 1;
 317         if (k1.objectid < k2->objectid)
 318                 return -1;
 319         if (k1.type > k2->type)
 320                 return 1;
 321         if (k1.type < k2->type)
 322                 return -1;
 323         if (k1.offset > k2->offset)
 324                 return 1;
 325         if (k1.offset < k2->offset)
 326                 return -1;
 327         return 0;
 328 }
 329
 330 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
 331                       int level)
 332 {
 333         struct extent_buffer *parent = NULL;
 334         struct extent_buffer *node = path->nodes[level];
 335         struct btrfs_disk_key parent_key;
 336         struct btrfs_disk_key node_key;
 337         int parent_slot;
 338         int slot;
 339         struct btrfs_key cpukey;
 340         u32 nritems = btrfs_header_nritems(node);
 341
 342         if (path->nodes[level + 1])
 343                 parent = path->nodes[level + 1];
 344
 345         slot = path->slots[level];
 346         BUG_ON(nritems == 0);
 347         if (parent) {
 348                 parent_slot = path->slots[level + 1];
 349                 btrfs_node_key(parent, &parent_key, parent_slot);
 350                 btrfs_node_key(node, &node_key, 0);
 351                 BUG_ON(memcmp(&parent_key, &node_key,
 352                               sizeof(struct btrfs_disk_key)));
 353                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 354                        btrfs_header_bytenr(node));
 355         }
 356         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 357         if (slot != 0) {
 358                 btrfs_node_key_to_cpu(node, &cpukey, slot - 1);
 359                 btrfs_node_key(node, &node_key, slot);
 360                 BUG_ON(comp_keys(&node_key, &cpukey) <= 0);
 361         }
 362         if (slot < nritems - 1) {
 363                 btrfs_node_key_to_cpu(node, &cpukey, slot + 1);
 364                 btrfs_node_key(node, &node_key, slot);
 365                 BUG_ON(comp_keys(&node_key, &cpukey) >= 0);
 366         }
 367         return 0;
 368 }
 369
 370 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 371                       int level)
 372 {
 373         struct extent_buffer *leaf = path->nodes[level];
 374         struct extent_buffer *parent = NULL;
 375         int parent_slot;
 376         struct btrfs_key cpukey;
 377         struct btrfs_disk_key parent_key;
 378         struct btrfs_disk_key leaf_key;
 379         int slot = path->slots[0];
 380
 381         u32 nritems = btrfs_header_nritems(leaf);
 382
 383         if (path->nodes[level + 1])
 384                 parent = path->nodes[level + 1];
 385
 386         if (nritems == 0)
 387                 return 0;
 388
 389         if (parent) {
 390                 parent_slot = path->slots[level + 1];
 391                 btrfs_node_key(parent, &parent_key, parent_slot);
 392                 btrfs_item_key(leaf, &leaf_key, 0);
 393
 394                 BUG_ON(memcmp(&parent_key, &leaf_key,
 395                        sizeof(struct btrfs_disk_key)));
 396                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 397                        btrfs_header_bytenr(leaf));
 398         }
 399 #if 0
 400         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 401                 btrfs_item_key_to_cpu(leaf, &cpukey, i + 1);
 402                 btrfs_item_key(leaf, &leaf_key, i);
 403                 if (comp_keys(&leaf_key, &cpukey) >= 0) {
 404                         btrfs_print_leaf(root, leaf);
 405                         printk("slot %d offset bad key\n", i);
 406                         BUG_ON(1);
 407                 }
 408                 if (btrfs_item_offset_nr(leaf, i) !=
 409                         btrfs_item_end_nr(leaf, i + 1)) {
 410                         btrfs_print_leaf(root, leaf);
 411                         printk("slot %d offset bad\n", i);
 412                         BUG_ON(1);
 413                 }
 414                 if (i == 0) {
 415                         if (btrfs_item_offset_nr(leaf, i) +
 416                                btrfs_item_size_nr(leaf, i) !=
 417                                BTRFS_LEAF_DATA_SIZE(root)) {
 418                                 btrfs_print_leaf(root, leaf);
 419                                 printk("slot %d first offset bad\n", i);
 420                                 BUG_ON(1);
 421                         }
 422                 }
 423         }
 424         if (nritems > 0) {
 425                 if (btrfs_item_size_nr(leaf, nritems - 1) > 4096) {
 426                                 btrfs_print_leaf(root, leaf);
 427                                 printk("slot %d bad size \n", nritems - 1);
 428                                 BUG_ON(1);
 429                 }
 430         }
 431 #endif
 432         if (slot != 0 && slot < nritems - 1) {
 433                 btrfs_item_key(leaf, &leaf_key, slot);
 434                 btrfs_item_key_to_cpu(leaf, &cpukey, slot - 1);
 435                 if (comp_keys(&leaf_key, &cpukey) <= 0) {
 436                         btrfs_print_leaf(root, leaf);
 437                         printk("slot %d offset bad key\n", slot);
 438                         BUG_ON(1);
 439                 }
 440                 if (btrfs_item_offset_nr(leaf, slot - 1) !=
 441                        btrfs_item_end_nr(leaf, slot)) {
 442                         btrfs_print_leaf(root, leaf);
 443                         printk("slot %d offset bad\n", slot);
 444                         BUG_ON(1);
 445                 }
 446         }
 447         if (slot < nritems - 1) {
 448                 btrfs_item_key(leaf, &leaf_key, slot);
 449                 btrfs_item_key_to_cpu(leaf, &cpukey, slot + 1);
 450                 BUG_ON(comp_keys(&leaf_key, &cpukey) >= 0);
 451                 if (btrfs_item_offset_nr(leaf, slot) !=
 452                         btrfs_item_end_nr(leaf, slot + 1)) {
 453                         btrfs_print_leaf(root, leaf);
 454                         printk("slot %d offset bad\n", slot);
 455                         BUG_ON(1);
 456                 }
 457         }
 458         BUG_ON(btrfs_item_offset_nr(leaf, 0) +
 459                btrfs_item_size_nr(leaf, 0) != BTRFS_LEAF_DATA_SIZE(root));
 460         return 0;
 461 }
 462
 463 static int check_block(struct btrfs_root *root, struct btrfs_path *path,
 464                         int level)
 465 {
 466         return 0;
 467 #if 0
 468         struct extent_buffer *buf = path->nodes[level];
 469
 470         if (memcmp_extent_buffer(buf, root->fs_info->fsid,
 471                                  (unsigned long)btrfs_header_fsid(buf),
 472                                  BTRFS_FSID_SIZE)) {
 473                 printk("warning bad block %Lu\n", buf->start);
 474                 return 1;
 475         }
 476 #endif
 477         if (level == 0)
 478                 return check_leaf(root, path, level);
 479         return check_node(root, path, level);
 480 }
 481
 482 /*
 483  * search for key in the extent_buffer.  The items start at offset p,
 484  * and they are item_size apart.  There are 'max' items in p.
 485  *
 486  * the slot in the array is returned via slot, and it points to
 487  * the place where you would insert key if it is not found in
 488  * the array.
 489  *
 490  * slot may point to max if the key is bigger than all of the keys
 491  */
 492 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 493                               int item_size, struct btrfs_key *key,
 494                               int max, int *slot)
 495 {
 496         int low = 0;
 497         int high = max;
 498         int mid;
 499         int ret;
 500         struct btrfs_disk_key *tmp = NULL;
 501         struct btrfs_disk_key unaligned;
 502         unsigned long offset;
 503         char *map_token = NULL;
 504         char *kaddr = NULL;
 505         unsigned long map_start = 0;
 506         unsigned long map_len = 0;
 507         int err;
 508
 509         while(low < high) {
 510                 mid = (low + high) / 2;
 511                 offset = p + mid * item_size;
 512
 513                 if (!map_token || offset < map_start ||
 514                     (offset + sizeof(struct btrfs_disk_key)) >
 515                     map_start + map_len) {
 516                         if (map_token) {
 517                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 518                                 map_token = NULL;
 519                         }
 520                         err = map_extent_buffer(eb, offset,
 521                                                 sizeof(struct btrfs_disk_key),
 522                                                 &map_token, &kaddr,
 523                                                 &map_start, &map_len, KM_USER0);
 524
 525                         if (!err) {
 526                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
 527                                                         map_start);
 528                         } else {
 529                                 read_extent_buffer(eb, &unaligned,
 530                                                    offset, sizeof(unaligned));
 531                                 tmp = &unaligned;
 532                         }
 533
 534                 } else {
 535                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
 536                                                         map_start);
 537                 }
 538                 ret = comp_keys(tmp, key);
 539
 540                 if (ret < 0)
 541                         low = mid + 1;
 542                 else if (ret > 0)
 543                         high = mid;
 544                 else {
 545                         *slot = mid;
 546                         if (map_token)
 547                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 548                         return 0;
 549                 }
 550         }
 551         *slot = low;
 552         if (map_token)
 553                 unmap_extent_buffer(eb, map_token, KM_USER0);
 554         return 1;
 555 }
 556
 557 /*
 558  * simple bin_search frontend that does the right thing for
 559  * leaves vs nodes
 560  */
 561 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 562                       int level, int *slot)
 563 {
 564         if (level == 0) {
 565                 return generic_bin_search(eb,
 566                                           offsetof(struct btrfs_leaf, items),
 567                                           sizeof(struct btrfs_item),
 568                                           key, btrfs_header_nritems(eb),
 569                                           slot);
 570         } else {
 571                 return generic_bin_search(eb,
 572                                           offsetof(struct btrfs_node, ptrs),
 573                                           sizeof(struct btrfs_key_ptr),
 574                                           key, btrfs_header_nritems(eb),
 575                                           slot);
 576         }
 577         return -1;
 578 }
 579
 580 static struct extent_buffer *read_node_slot(struct btrfs_root *root,
 581                                    struct extent_buffer *parent, int slot)
 582 {
 583         if (slot < 0)
 584                 return NULL;
 585         if (slot >= btrfs_header_nritems(parent))
 586                 return NULL;
 587         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 588                        btrfs_level_size(root, btrfs_header_level(parent) - 1));
 589 }
 590
 591 static int balance_level(struct btrfs_trans_handle *trans, struct btrfs_root
 592                          *root, struct btrfs_path *path, int level)
 593 {
 594         struct extent_buffer *right = NULL;
 595         struct extent_buffer *mid;
 596         struct extent_buffer *left = NULL;
 597         struct extent_buffer *parent = NULL;
 598         int ret = 0;
 599         int wret;
 600         int pslot;
 601         int orig_slot = path->slots[level];
 602         int err_on_enospc = 0;
 603         u64 orig_ptr;
 604
 605         if (level == 0)
 606                 return 0;
 607
 608         mid = path->nodes[level];
 609         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 610
 611         if (level < BTRFS_MAX_LEVEL - 1)
 612                 parent = path->nodes[level + 1];
 613         pslot = path->slots[level + 1];
 614
 615         /*
 616          * deal with the case where there is only one pointer in the root
 617          * by promoting the node below to a root
 618          */
 619         if (!parent) {
 620                 struct extent_buffer *child;
 621
 622                 if (btrfs_header_nritems(mid) != 1)
 623                         return 0;
 624
 625                 /* promote the child to a root */
 626                 child = read_node_slot(root, mid, 0);
 627                 BUG_ON(!child);
 628                 root->node = child;
 629                 path->nodes[level] = NULL;
 630                 clean_tree_block(trans, root, mid);
 631                 wait_on_tree_block_writeback(root, mid);
 632                 /* once for the path */
 633                 free_extent_buffer(mid);
 634                 ret = btrfs_free_extent(trans, root, mid->start, mid->len, 1);
 635                 /* once for the root ptr */
 636                 free_extent_buffer(mid);
 637                 return ret;
 638         }
 639         if (btrfs_header_nritems(mid) >
 640             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 641                 return 0;
 642
 643         if (btrfs_header_nritems(mid) < 2)
 644                 err_on_enospc = 1;
 645
 646         left = read_node_slot(root, parent, pslot - 1);
 647         if (left) {
 648                 wret = btrfs_cow_block(trans, root, left,
 649                                        parent, pslot - 1, &left);
 650                 if (wret) {
 651                         ret = wret;
 652                         goto enospc;
 653                 }
 654         }
 655         right = read_node_slot(root, parent, pslot + 1);
 656         if (right) {
 657                 wret = btrfs_cow_block(trans, root, right,
 658                                        parent, pslot + 1, &right);
 659                 if (wret) {
 660                         ret = wret;
 661                         goto enospc;
 662                 }
 663         }
 664
 665         /* first, try to make some room in the middle buffer */
 666         if (left) {
 667                 orig_slot += btrfs_header_nritems(left);
 668                 wret = push_node_left(trans, root, left, mid);
 669                 if (wret < 0)
 670                         ret = wret;
 671                 if (btrfs_header_nritems(mid) < 2)
 672                         err_on_enospc = 1;
 673         }
 674
 675         /*
 676          * then try to empty the right most buffer into the middle
 677          */
 678         if (right) {
 679                 wret = push_node_left(trans, root, mid, right);
 680                 if (wret < 0 && wret != -ENOSPC)
 681                         ret = wret;
 682                 if (btrfs_header_nritems(right) == 0) {
 683                         u64 bytenr = right->start;
 684                         u32 blocksize = right->len;
 685
 686                         clean_tree_block(trans, root, right);
 687                         wait_on_tree_block_writeback(root, right);
 688                         free_extent_buffer(right);
 689                         right = NULL;
 690                         wret = del_ptr(trans, root, path, level + 1, pslot +
 691                                        1);
 692                         if (wret)
 693                                 ret = wret;
 694                         wret = btrfs_free_extent(trans, root, bytenr,
 695                                                  blocksize, 1);
 696                         if (wret)
 697                                 ret = wret;
 698                 } else {
 699                         struct btrfs_disk_key right_key;
 700                         btrfs_node_key(right, &right_key, 0);
 701                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 702                         btrfs_mark_buffer_dirty(parent);
 703                 }
 704         }
 705         if (btrfs_header_nritems(mid) == 1) {
 706                 /*
 707                  * we're not allowed to leave a node with one item in the
 708                  * tree during a delete.  A deletion from lower in the tree
 709                  * could try to delete the only pointer in this node.
 710                  * So, pull some keys from the left.
 711                  * There has to be a left pointer at this point because
 712                  * otherwise we would have pulled some pointers from the
 713                  * right
 714                  */
 715                 BUG_ON(!left);
 716                 wret = balance_node_right(trans, root, mid, left);
 717                 if (wret < 0) {
 718                         ret = wret;
 719                         goto enospc;
 720                 }
 721                 BUG_ON(wret == 1);
 722         }
 723         if (btrfs_header_nritems(mid) == 0) {
 724                 /* we've managed to empty the middle node, drop it */
 725                 u64 bytenr = mid->start;
 726                 u32 blocksize = mid->len;
 727                 clean_tree_block(trans, root, mid);
 728                 wait_on_tree_block_writeback(root, mid);
 729                 free_extent_buffer(mid);
 730                 mid = NULL;
 731                 wret = del_ptr(trans, root, path, level + 1, pslot);
 732                 if (wret)
 733                         ret = wret;
 734                 wret = btrfs_free_extent(trans, root, bytenr, blocksize, 1);
 735                 if (wret)
 736                         ret = wret;
 737         } else {
 738                 /* update the parent key to reflect our changes */
 739                 struct btrfs_disk_key mid_key;
 740                 btrfs_node_key(mid, &mid_key, 0);
 741                 btrfs_set_node_key(parent, &mid_key, pslot);
 742                 btrfs_mark_buffer_dirty(parent);
 743         }
 744
 745         /* update the path */
 746         if (left) {
 747                 if (btrfs_header_nritems(left) > orig_slot) {
 748                         extent_buffer_get(left);
 749                         path->nodes[level] = left;
 750                         path->slots[level + 1] -= 1;
 751                         path->slots[level] = orig_slot;
 752                         if (mid)
 753                                 free_extent_buffer(mid);
 754                 } else {
 755                         orig_slot -= btrfs_header_nritems(left);
 756                         path->slots[level] = orig_slot;
 757                 }
 758         }
 759         /* double check we haven't messed things up */
 760         check_block(root, path, level);
 761         if (orig_ptr !=
 762             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 763                 BUG();
 764 enospc:
 765         if (right)
 766                 free_extent_buffer(right);
 767         if (left)
 768                 free_extent_buffer(left);
 769         return ret;
 770 }
 771
 772 /* returns zero if the push worked, non-zero otherwise */
 773 static int push_nodes_for_insert(struct btrfs_trans_handle *trans,
 774                                 struct btrfs_root *root,
 775                                 struct btrfs_path *path, int level)
 776 {
 777         struct extent_buffer *right = NULL;
 778         struct extent_buffer *mid;
 779         struct extent_buffer *left = NULL;
 780         struct extent_buffer *parent = NULL;
 781         int ret = 0;
 782         int wret;
 783         int pslot;
 784         int orig_slot = path->slots[level];
 785         u64 orig_ptr;
 786
 787         if (level == 0)
 788                 return 1;
 789
 790         mid = path->nodes[level];
 791         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 792
 793         if (level < BTRFS_MAX_LEVEL - 1)
 794                 parent = path->nodes[level + 1];
 795         pslot = path->slots[level + 1];
 796
 797         if (!parent)
 798                 return 1;
 799
 800         left = read_node_slot(root, parent, pslot - 1);
 801
 802         /* first, try to make some room in the middle buffer */
 803         if (left) {
 804                 u32 left_nr;
 805                 left_nr = btrfs_header_nritems(left);
 806                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 807                         wret = 1;
 808                 } else {
 809                         ret = btrfs_cow_block(trans, root, left, parent,
 810                                               pslot - 1, &left);
 811                         if (ret)
 812                                 wret = 1;
 813                         else {
 814                                 wret = push_node_left(trans, root,
 815                                                       left, mid);
 816                         }
 817                 }
 818                 if (wret < 0)
 819                         ret = wret;
 820                 if (wret == 0) {
 821                         struct btrfs_disk_key disk_key;
 822                         orig_slot += left_nr;
 823                         btrfs_node_key(mid, &disk_key, 0);
 824                         btrfs_set_node_key(parent, &disk_key, pslot);
 825                         btrfs_mark_buffer_dirty(parent);
 826                         if (btrfs_header_nritems(left) > orig_slot) {
 827                                 path->nodes[level] = left;
 828                                 path->slots[level + 1] -= 1;
 829                                 path->slots[level] = orig_slot;
 830                                 free_extent_buffer(mid);
 831                         } else {
 832                                 orig_slot -=
 833                                         btrfs_header_nritems(left);
 834                                 path->slots[level] = orig_slot;
 835                                 free_extent_buffer(left);
 836                         }
 837                         return 0;
 838                 }
 839                 free_extent_buffer(left);
 840         }
 841         right= read_node_slot(root, parent, pslot + 1);
 842
 843         /*
 844          * then try to empty the right most buffer into the middle
 845          */
 846         if (right) {
 847                 u32 right_nr;
 848                 right_nr = btrfs_header_nritems(right);
 849                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 850                         wret = 1;
 851                 } else {
 852                         ret = btrfs_cow_block(trans, root, right,
 853                                               parent, pslot + 1,
 854                                               &right);
 855                         if (ret)
 856                                 wret = 1;
 857                         else {
 858                                 wret = balance_node_right(trans, root,
 859                                                           right, mid);
 860                         }
 861                 }
 862                 if (wret < 0)
 863                         ret = wret;
 864                 if (wret == 0) {
 865                         struct btrfs_disk_key disk_key;
 866
 867                         btrfs_node_key(right, &disk_key, 0);
 868                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
 869                         btrfs_mark_buffer_dirty(parent);
 870
 871                         if (btrfs_header_nritems(mid) <= orig_slot) {
 872                                 path->nodes[level] = right;
 873                                 path->slots[level + 1] += 1;
 874                                 path->slots[level] = orig_slot -
 875                                         btrfs_header_nritems(mid);
 876                                 free_extent_buffer(mid);
 877                         } else {
 878                                 free_extent_buffer(right);
 879                         }
 880                         return 0;
 881                 }
 882                 free_extent_buffer(right);
 883         }
 884         return 1;
 885 }
 886
 887 /*
 888  * readahead one full node of leaves
 889  */
 890 static void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
 891                              int level, int slot)
 892 {
 893         struct extent_buffer *node;
 894         u32 nritems;
 895         u64 search;
 896         u64 lowest_read;
 897         u64 highest_read;
 898         u64 nread = 0;
 899         int direction = path->reada;
 900         struct extent_buffer *eb;
 901         u32 nr;
 902         u32 blocksize;
 903         u32 nscan = 0;
 904
 905         if (level == 0)
 906                 return;
 907
 908         if (!path->nodes[level])
 909                 return;
 910
 911         node = path->nodes[level];
 912         search = btrfs_node_blockptr(node, slot);
 913         blocksize = btrfs_level_size(root, level - 1);
 914         eb = btrfs_find_tree_block(root, search, blocksize);
 915         if (eb) {
 916                 free_extent_buffer(eb);
 917                 return;
 918         }
 919
 920         highest_read = search;
 921         lowest_read = search;
 922
 923         nritems = btrfs_header_nritems(node);
 924         nr = slot;
 925         while(1) {
 926                 if (direction < 0) {
 927                         if (nr == 0)
 928                                 break;
 929                         nr--;
 930                 } else if (direction > 0) {
 931                         nr++;
 932                         if (nr >= nritems)
 933                                 break;
 934                 }
 935                 search = btrfs_node_blockptr(node, nr);
 936                 if ((search >= lowest_read && search <= highest_read) ||
 937                     (search < lowest_read && lowest_read - search <= 32768) ||
 938                     (search > highest_read && search - highest_read <= 32768)) {
 939                         readahead_tree_block(root, search, blocksize);
 940                         nread += blocksize;
 941                 }
 942                 nscan++;
 943                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
 944                         break;
 945                 if(nread > (1024 * 1024) || nscan > 128)
 946                         break;
 947
 948                 if (search < lowest_read)
 949                         lowest_read = search;
 950                 if (search > highest_read)
 951                         highest_read = search;
 952         }
 953 }
 954 /*
 955  * look for key in the tree.  path is filled in with nodes along the way
 956  * if key is found, we return zero and you can find the item in the leaf
 957  * level of the path (level 0)
 958  *
 959  * If the key isn't found, the path points to the slot where it should
 960  * be inserted, and 1 is returned.  If there are other errors during the
 961  * search a negative error number is returned.
 962  *
 963  * if ins_len > 0, nodes and leaves will be split as we walk down the
 964  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
 965  * possible)
 966  */
 967 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
 968                       *root, struct btrfs_key *key, struct btrfs_path *p, int
 969                       ins_len, int cow)
 970 {
 971         struct extent_buffer *b;
 972         u64 bytenr;
 973         int slot;
 974         int ret;
 975         int level;
 976         int should_reada = p->reada;
 977         u8 lowest_level = 0;
 978
 979         lowest_level = p->lowest_level;
 980         WARN_ON(lowest_level && ins_len);
 981         WARN_ON(p->nodes[0] != NULL);
 982         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
 983 again:
 984         b = root->node;
 985         extent_buffer_get(b);
 986         while (b) {
 987                 level = btrfs_header_level(b);
 988                 if (cow) {
 989                         int wret;
 990                         wret = btrfs_cow_block(trans, root, b,
 991                                                p->nodes[level + 1],
 992                                                p->slots[level + 1],
 993                                                &b);
 994                         if (wret) {
 995                                 free_extent_buffer(b);
 996                                 return wret;
 997                         }
 998                 }
 999                 BUG_ON(!cow && ins_len);
1000                 if (level != btrfs_header_level(b))
1001                         WARN_ON(1);
1002                 level = btrfs_header_level(b);
1003                 p->nodes[level] = b;
1004                 ret = check_block(root, p, level);
1005                 if (ret)
1006                         return -1;
1007                 ret = bin_search(b, key, level, &slot);
1008                 if (level != 0) {
1009                         if (ret && slot > 0)
1010                                 slot -= 1;
1011                         p->slots[level] = slot;
1012                         if (ins_len > 0 && btrfs_header_nritems(b) >=
1013                             BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1014                                 int sret = split_node(trans, root, p, level);
1015                                 BUG_ON(sret > 0);
1016                                 if (sret)
1017                                         return sret;
1018                                 b = p->nodes[level];
1019                                 slot = p->slots[level];
1020                         } else if (ins_len < 0) {
1021                                 int sret = balance_level(trans, root, p,
1022                                                          level);
1023                                 if (sret)
1024                                         return sret;
1025                                 b = p->nodes[level];
1026                                 if (!b) {
1027                                         btrfs_release_path(NULL, p);
1028                                         goto again;
1029                                 }
1030                                 slot = p->slots[level];
1031                                 BUG_ON(btrfs_header_nritems(b) == 1);
1032                         }
1033                         /* this is only true while dropping a snapshot */
1034                         if (level == lowest_level)
1035                                 break;
1036                         bytenr = btrfs_node_blockptr(b, slot);
1037                         if (should_reada)
1038                                 reada_for_search(root, p, level, slot);
1039                         b = read_tree_block(root, bytenr,
1040                                             btrfs_level_size(root, level - 1));
1041                 } else {
1042                         p->slots[level] = slot;
1043                         if (ins_len > 0 && btrfs_leaf_free_space(root, b) <
1044                             sizeof(struct btrfs_item) + ins_len) {
1045                                 int sret = split_leaf(trans, root, key,
1046                                                       p, ins_len);
1047                                 BUG_ON(sret > 0);
1048                                 if (sret)
1049                                         return sret;
1050                         }
1051                         return ret;
1052                 }
1053         }
1054         return 1;
1055 }
1056
1057 /*
1058  * adjust the pointers going up the tree, starting at level
1059  * making sure the right key of each node is points to 'key'.
1060  * This is used after shifting pointers to the left, so it stops
1061  * fixing up pointers when a given leaf/node is not in slot 0 of the
1062  * higher levels
1063  *
1064  * If this fails to write a tree block, it returns -1, but continues
1065  * fixing up the blocks in ram so the tree is consistent.
1066  */
1067 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1068                           struct btrfs_root *root, struct btrfs_path *path,
1069                           struct btrfs_disk_key *key, int level)
1070 {
1071         int i;
1072         int ret = 0;
1073         struct extent_buffer *t;
1074
1075         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1076                 int tslot = path->slots[i];
1077                 if (!path->nodes[i])
1078                         break;
1079                 t = path->nodes[i];
1080                 btrfs_set_node_key(t, key, tslot);
1081                 btrfs_mark_buffer_dirty(path->nodes[i]);
1082                 if (tslot != 0)
1083                         break;
1084         }
1085         return ret;
1086 }
1087
1088 /*
1089  * try to push data from one node into the next node left in the
1090  * tree.
1091  *
1092  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1093  * error, and > 0 if there was no room in the left hand block.
1094  */
1095 static int push_node_left(struct btrfs_trans_handle *trans, struct btrfs_root
1096                           *root, struct extent_buffer *dst,
1097                           struct extent_buffer *src)
1098 {
1099         int push_items = 0;
1100         int src_nritems;
1101         int dst_nritems;
1102         int ret = 0;
1103
1104         src_nritems = btrfs_header_nritems(src);
1105         dst_nritems = btrfs_header_nritems(dst);
1106         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1107
1108         if (push_items <= 0) {
1109                 return 1;
1110         }
1111
1112         if (src_nritems < push_items)
1113                 push_items = src_nritems;
1114
1115         copy_extent_buffer(dst, src,
1116                            btrfs_node_key_ptr_offset(dst_nritems),
1117                            btrfs_node_key_ptr_offset(0),
1118                            push_items * sizeof(struct btrfs_key_ptr));
1119
1120         if (push_items < src_nritems) {
1121                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1122                                       btrfs_node_key_ptr_offset(push_items),
1123                                       (src_nritems - push_items) *
1124                                       sizeof(struct btrfs_key_ptr));
1125         }
1126         btrfs_set_header_nritems(src, src_nritems - push_items);
1127         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1128         btrfs_mark_buffer_dirty(src);
1129         btrfs_mark_buffer_dirty(dst);
1130         return ret;
1131 }
1132
1133 /*
1134  * try to push data from one node into the next node right in the
1135  * tree.
1136  *
1137  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1138  * error, and > 0 if there was no room in the right hand block.
1139  *
1140  * this will  only push up to 1/2 the contents of the left node over
1141  */
1142 static int balance_node_right(struct btrfs_trans_handle *trans,
1143                               struct btrfs_root *root,
1144                               struct extent_buffer *dst,
1145                               struct extent_buffer *src)
1146 {
1147         int push_items = 0;
1148         int max_push;
1149         int src_nritems;
1150         int dst_nritems;
1151         int ret = 0;
1152
1153         src_nritems = btrfs_header_nritems(src);
1154         dst_nritems = btrfs_header_nritems(dst);
1155         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1156         if (push_items <= 0)
1157                 return 1;
1158
1159         max_push = src_nritems / 2 + 1;
1160         /* don't try to empty the node */
1161         if (max_push >= src_nritems)
1162                 return 1;
1163
1164         if (max_push < push_items)
1165                 push_items = max_push;
1166
1167         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1168                                       btrfs_node_key_ptr_offset(0),
1169                                       (dst_nritems) *
1170                                       sizeof(struct btrfs_key_ptr));
1171
1172         copy_extent_buffer(dst, src,
1173                            btrfs_node_key_ptr_offset(0),
1174                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1175                            push_items * sizeof(struct btrfs_key_ptr));
1176
1177         btrfs_set_header_nritems(src, src_nritems - push_items);
1178         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1179
1180         btrfs_mark_buffer_dirty(src);
1181         btrfs_mark_buffer_dirty(dst);
1182         return ret;
1183 }
1184
1185 /*
1186  * helper function to insert a new root level in the tree.
1187  * A new node is allocated, and a single item is inserted to
1188  * point to the existing root
1189  *
1190  * returns zero on success or < 0 on failure.
1191  */
1192 static int insert_new_root(struct btrfs_trans_handle *trans,
1193                            struct btrfs_root *root,
1194                            struct btrfs_path *path, int level)
1195 {
1196         struct extent_buffer *lower;
1197         struct extent_buffer *c;
1198         struct btrfs_disk_key lower_key;
1199
1200         BUG_ON(path->nodes[level]);
1201         BUG_ON(path->nodes[level-1] != root->node);
1202
1203         c = btrfs_alloc_free_block(trans, root, root->nodesize,
1204                                    root->node->start, 0);
1205         if (IS_ERR(c))
1206                 return PTR_ERR(c);
1207         memset_extent_buffer(c, 0, 0, root->nodesize);
1208         btrfs_set_header_nritems(c, 1);
1209         btrfs_set_header_level(c, level);
1210         btrfs_set_header_bytenr(c, c->start);
1211         btrfs_set_header_generation(c, trans->transid);
1212         btrfs_set_header_owner(c, root->root_key.objectid);
1213         lower = path->nodes[level-1];
1214
1215         write_extent_buffer(c, root->fs_info->fsid,
1216                             (unsigned long)btrfs_header_fsid(c),
1217                             BTRFS_FSID_SIZE);
1218         if (level == 1)
1219                 btrfs_item_key(lower, &lower_key, 0);
1220         else
1221                 btrfs_node_key(lower, &lower_key, 0);
1222         btrfs_set_node_key(c, &lower_key, 0);
1223         btrfs_set_node_blockptr(c, 0, lower->start);
1224
1225         btrfs_mark_buffer_dirty(c);
1226
1227         /* the super has an extra ref to root->node */
1228         free_extent_buffer(root->node);
1229         root->node = c;
1230         extent_buffer_get(c);
1231         path->nodes[level] = c;
1232         path->slots[level] = 0;
1233         return 0;
1234 }
1235
1236 /*
1237  * worker function to insert a single pointer in a node.
1238  * the node should have enough room for the pointer already
1239  *
1240  * slot and level indicate where you want the key to go, and
1241  * blocknr is the block the key points to.
1242  *
1243  * returns zero on success and < 0 on any error
1244  */
1245 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1246                       *root, struct btrfs_path *path, struct btrfs_disk_key
1247                       *key, u64 bytenr, int slot, int level)
1248 {
1249         struct extent_buffer *lower;
1250         int nritems;
1251
1252         BUG_ON(!path->nodes[level]);
1253         lower = path->nodes[level];
1254         nritems = btrfs_header_nritems(lower);
1255         if (slot > nritems)
1256                 BUG();
1257         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1258                 BUG();
1259         if (slot != nritems) {
1260                 memmove_extent_buffer(lower,
1261                               btrfs_node_key_ptr_offset(slot + 1),
1262                               btrfs_node_key_ptr_offset(slot),
1263                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1264         }
1265         btrfs_set_node_key(lower, key, slot);
1266         btrfs_set_node_blockptr(lower, slot, bytenr);
1267         btrfs_set_header_nritems(lower, nritems + 1);
1268         btrfs_mark_buffer_dirty(lower);
1269         return 0;
1270 }
1271
1272 /*
1273  * split the node at the specified level in path in two.
1274  * The path is corrected to point to the appropriate node after the split
1275  *
1276  * Before splitting this tries to make some room in the node by pushing
1277  * left and right, if either one works, it returns right away.
1278  *
1279  * returns 0 on success and < 0 on failure
1280  */
1281 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1282                       *root, struct btrfs_path *path, int level)
1283 {
1284         struct extent_buffer *c;
1285         struct extent_buffer *split;
1286         struct btrfs_disk_key disk_key;
1287         int mid;
1288         int ret;
1289         int wret;
1290         u32 c_nritems;
1291
1292         c = path->nodes[level];
1293         if (c == root->node) {
1294                 /* trying to split the root, lets make a new one */
1295                 ret = insert_new_root(trans, root, path, level + 1);
1296                 if (ret)
1297                         return ret;
1298         } else {
1299                 ret = push_nodes_for_insert(trans, root, path, level);
1300                 c = path->nodes[level];
1301                 if (!ret && btrfs_header_nritems(c) <
1302                     BTRFS_NODEPTRS_PER_BLOCK(root) - 1)
1303                         return 0;
1304                 if (ret < 0)
1305                         return ret;
1306         }
1307
1308         c_nritems = btrfs_header_nritems(c);
1309         split = btrfs_alloc_free_block(trans, root, root->nodesize,
1310                                        c->start, 0);
1311         if (IS_ERR(split))
1312                 return PTR_ERR(split);
1313
1314         btrfs_set_header_flags(split, btrfs_header_flags(c));
1315         btrfs_set_header_level(split, btrfs_header_level(c));
1316         btrfs_set_header_bytenr(split, split->start);
1317         btrfs_set_header_generation(split, trans->transid);
1318         btrfs_set_header_owner(split, root->root_key.objectid);
1319         write_extent_buffer(split, root->fs_info->fsid,
1320                             (unsigned long)btrfs_header_fsid(split),
1321                             BTRFS_FSID_SIZE);
1322
1323         mid = (c_nritems + 1) / 2;
1324
1325         copy_extent_buffer(split, c,
1326                            btrfs_node_key_ptr_offset(0),
1327                            btrfs_node_key_ptr_offset(mid),
1328                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1329         btrfs_set_header_nritems(split, c_nritems - mid);
1330         btrfs_set_header_nritems(c, mid);
1331         ret = 0;
1332
1333         btrfs_mark_buffer_dirty(c);
1334         btrfs_mark_buffer_dirty(split);
1335
1336         btrfs_node_key(split, &disk_key, 0);
1337         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1338                           path->slots[level + 1] + 1,
1339                           level + 1);
1340         if (wret)
1341                 ret = wret;
1342
1343         if (path->slots[level] >= mid) {
1344                 path->slots[level] -= mid;
1345                 free_extent_buffer(c);
1346                 path->nodes[level] = split;
1347                 path->slots[level + 1] += 1;
1348         } else {
1349                 free_extent_buffer(split);
1350         }
1351         return ret;
1352 }
1353
1354 /*
1355  * how many bytes are required to store the items in a leaf.  start
1356  * and nr indicate which items in the leaf to check.  This totals up the
1357  * space used both by the item structs and the item data
1358  */
1359 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1360 {
1361         int data_len;
1362         int nritems = btrfs_header_nritems(l);
1363         int end = min(nritems, start + nr) - 1;
1364
1365         if (!nr)
1366                 return 0;
1367         data_len = btrfs_item_end_nr(l, start);
1368         data_len = data_len - btrfs_item_offset_nr(l, end);
1369         data_len += sizeof(struct btrfs_item) * nr;
1370         WARN_ON(data_len < 0);
1371         return data_len;
1372 }
1373
1374 /*
1375  * The space between the end of the leaf items and
1376  * the start of the leaf data.  IOW, how much room
1377  * the leaf has left for both items and data
1378  */
1379 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1380 {
1381         int nritems = btrfs_header_nritems(leaf);
1382         int ret;
1383         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1384         if (ret < 0) {
1385                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1386                        ret, BTRFS_LEAF_DATA_SIZE(root),
1387                        leaf_space_used(leaf, 0, nritems), nritems);
1388         }
1389         return ret;
1390 }
1391
1392 /*
1393  * push some data in the path leaf to the right, trying to free up at
1394  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1395  *
1396  * returns 1 if the push failed because the other node didn't have enough
1397  * room, 0 if everything worked out and < 0 if there were major errors.
1398  */
1399 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1400                            *root, struct btrfs_path *path, int data_size)
1401 {
1402         struct extent_buffer *left = path->nodes[0];
1403         struct extent_buffer *right;
1404         struct extent_buffer *upper;
1405         struct btrfs_disk_key disk_key;
1406         int slot;
1407         int i;
1408         int free_space;
1409         int push_space = 0;
1410         int push_items = 0;
1411         struct btrfs_item *item;
1412         u32 left_nritems;
1413         u32 right_nritems;
1414         u32 data_end;
1415         u32 this_item_size;
1416         int ret;
1417
1418         slot = path->slots[1];
1419         if (!path->nodes[1]) {
1420                 return 1;
1421         }
1422         upper = path->nodes[1];
1423         if (slot >= btrfs_header_nritems(upper) - 1)
1424                 return 1;
1425
1426         right = read_tree_block(root, btrfs_node_blockptr(upper, slot + 1),
1427                                 root->leafsize);
1428         free_space = btrfs_leaf_free_space(root, right);
1429         if (free_space < data_size + sizeof(struct btrfs_item)) {
1430                 free_extent_buffer(right);
1431                 return 1;
1432         }
1433
1434         /* cow and double check */
1435         ret = btrfs_cow_block(trans, root, right, upper,
1436                               slot + 1, &right);
1437         if (ret) {
1438                 free_extent_buffer(right);
1439                 return 1;
1440         }
1441         free_space = btrfs_leaf_free_space(root, right);
1442         if (free_space < data_size + sizeof(struct btrfs_item)) {
1443                 free_extent_buffer(right);
1444                 return 1;
1445         }
1446
1447         left_nritems = btrfs_header_nritems(left);
1448         if (left_nritems == 0) {
1449                 free_extent_buffer(right);
1450                 return 1;
1451         }
1452
1453         for (i = left_nritems - 1; i >= 1; i--) {
1454                 item = btrfs_item_nr(left, i);
1455
1456                 if (path->slots[0] == i)
1457                         push_space += data_size + sizeof(*item);
1458
1459                 if (!left->map_token) {
1460                         map_extent_buffer(left, (unsigned long)item,
1461                                         sizeof(struct btrfs_item),
1462                                         &left->map_token, &left->kaddr,
1463                                         &left->map_start, &left->map_len,
1464                                         KM_USER1);
1465                 }
1466
1467                 this_item_size = btrfs_item_size(left, item);
1468                 if (this_item_size + sizeof(*item) + push_space > free_space)
1469                         break;
1470                 push_items++;
1471                 push_space += this_item_size + sizeof(*item);
1472         }
1473         if (left->map_token) {
1474                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1475                 left->map_token = NULL;
1476         }
1477
1478         if (push_items == 0) {
1479                 free_extent_buffer(right);
1480                 return 1;
1481         }
1482
1483         if (push_items == left_nritems)
1484                 WARN_ON(1);
1485
1486         /* push left to right */
1487         right_nritems = btrfs_header_nritems(right);
1488         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1489         push_space -= leaf_data_end(root, left);
1490
1491         /* make room in the right data area */
1492         data_end = leaf_data_end(root, right);
1493         memmove_extent_buffer(right,
1494                               btrfs_leaf_data(right) + data_end - push_space,
1495                               btrfs_leaf_data(right) + data_end,
1496                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1497
1498         /* copy from the left data area */
1499         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1500                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1501                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1502                      push_space);
1503
1504         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1505                               btrfs_item_nr_offset(0),
1506                               right_nritems * sizeof(struct btrfs_item));
1507
1508         /* copy the items from left to right */
1509         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1510                    btrfs_item_nr_offset(left_nritems - push_items),
1511                    push_items * sizeof(struct btrfs_item));
1512
1513         /* update the item pointers */
1514         right_nritems += push_items;
1515         btrfs_set_header_nritems(right, right_nritems);
1516         push_space = BTRFS_LEAF_DATA_SIZE(root);
1517
1518         for (i = 0; i < right_nritems; i++) {
1519                 item = btrfs_item_nr(right, i);
1520                 if (!right->map_token) {
1521                         map_extent_buffer(right, (unsigned long)item,
1522                                         sizeof(struct btrfs_item),
1523                                         &right->map_token, &right->kaddr,
1524                                         &right->map_start, &right->map_len,
1525                                         KM_USER1);
1526                 }
1527                 push_space -= btrfs_item_size(right, item);
1528                 btrfs_set_item_offset(right, item, push_space);
1529         }
1530
1531         if (right->map_token) {
1532                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1533                 right->map_token = NULL;
1534         }
1535         left_nritems -= push_items;
1536         btrfs_set_header_nritems(left, left_nritems);
1537
1538         btrfs_mark_buffer_dirty(left);
1539         btrfs_mark_buffer_dirty(right);
1540
1541         btrfs_item_key(right, &disk_key, 0);
1542         btrfs_set_node_key(upper, &disk_key, slot + 1);
1543         btrfs_mark_buffer_dirty(upper);
1544
1545         /* then fixup the leaf pointer in the path */
1546         if (path->slots[0] >= left_nritems) {
1547                 path->slots[0] -= left_nritems;
1548                 free_extent_buffer(path->nodes[0]);
1549                 path->nodes[0] = right;
1550                 path->slots[1] += 1;
1551         } else {
1552                 free_extent_buffer(right);
1553         }
1554         return 0;
1555 }
1556 /*
1557  * push some data in the path leaf to the left, trying to free up at
1558  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1559  */
1560 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1561                           *root, struct btrfs_path *path, int data_size)
1562 {
1563         struct btrfs_disk_key disk_key;
1564         struct extent_buffer *right = path->nodes[0];
1565         struct extent_buffer *left;
1566         int slot;
1567         int i;
1568         int free_space;
1569         int push_space = 0;
1570         int push_items = 0;
1571         struct btrfs_item *item;
1572         u32 old_left_nritems;
1573         u32 right_nritems;
1574         int ret = 0;
1575         int wret;
1576         u32 this_item_size;
1577         u32 old_left_item_size;
1578
1579         slot = path->slots[1];
1580         if (slot == 0)
1581                 return 1;
1582         if (!path->nodes[1])
1583                 return 1;
1584
1585         left = read_tree_block(root, btrfs_node_blockptr(path->nodes[1],
1586                                slot - 1), root->leafsize);
1587         free_space = btrfs_leaf_free_space(root, left);
1588         if (free_space < data_size + sizeof(struct btrfs_item)) {
1589                 free_extent_buffer(left);
1590                 return 1;
1591         }
1592
1593         /* cow and double check */
1594         ret = btrfs_cow_block(trans, root, left,
1595                               path->nodes[1], slot - 1, &left);
1596         if (ret) {
1597                 /* we hit -ENOSPC, but it isn't fatal here */
1598                 free_extent_buffer(left);
1599                 return 1;
1600         }
1601         free_space = btrfs_leaf_free_space(root, left);
1602         if (free_space < data_size + sizeof(struct btrfs_item)) {
1603                 free_extent_buffer(left);
1604                 return 1;
1605         }
1606
1607         right_nritems = btrfs_header_nritems(right);
1608         if (right_nritems == 0) {
1609                 free_extent_buffer(left);
1610                 return 1;
1611         }
1612
1613         for (i = 0; i < right_nritems - 1; i++) {
1614                 item = btrfs_item_nr(right, i);
1615                 if (!right->map_token) {
1616                         map_extent_buffer(right, (unsigned long)item,
1617                                         sizeof(struct btrfs_item),
1618                                         &right->map_token, &right->kaddr,
1619                                         &right->map_start, &right->map_len,
1620                                         KM_USER1);
1621                 }
1622
1623                 if (path->slots[0] == i)
1624                         push_space += data_size + sizeof(*item);
1625
1626                 this_item_size = btrfs_item_size(right, item);
1627                 if (this_item_size + sizeof(*item) + push_space > free_space)
1628                         break;
1629
1630                 push_items++;
1631                 push_space += this_item_size + sizeof(*item);
1632         }
1633
1634         if (right->map_token) {
1635                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1636                 right->map_token = NULL;
1637         }
1638
1639         if (push_items == 0) {
1640                 free_extent_buffer(left);
1641                 return 1;
1642         }
1643         if (push_items == btrfs_header_nritems(right))
1644                 WARN_ON(1);
1645
1646         /* push data from right to left */
1647         copy_extent_buffer(left, right,
1648                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1649                            btrfs_item_nr_offset(0),
1650                            push_items * sizeof(struct btrfs_item));
1651
1652         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1653                      btrfs_item_offset_nr(right, push_items -1);
1654
1655         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1656                      leaf_data_end(root, left) - push_space,
1657                      btrfs_leaf_data(right) +
1658                      btrfs_item_offset_nr(right, push_items - 1),
1659                      push_space);
1660         old_left_nritems = btrfs_header_nritems(left);
1661         BUG_ON(old_left_nritems < 0);
1662
1663         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1664         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1665                 u32 ioff;
1666
1667                 item = btrfs_item_nr(left, i);
1668                 if (!left->map_token) {
1669                         map_extent_buffer(left, (unsigned long)item,
1670                                         sizeof(struct btrfs_item),
1671                                         &left->map_token, &left->kaddr,
1672                                         &left->map_start, &left->map_len,
1673                                         KM_USER1);
1674                 }
1675
1676                 ioff = btrfs_item_offset(left, item);
1677                 btrfs_set_item_offset(left, item,
1678                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1679         }
1680         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1681         if (left->map_token) {
1682                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1683                 left->map_token = NULL;
1684         }
1685
1686         /* fixup right node */
1687         push_space = btrfs_item_offset_nr(right, push_items - 1) -
1688                                           leaf_data_end(root, right);
1689         memmove_extent_buffer(right, btrfs_leaf_data(right) +
1690                               BTRFS_LEAF_DATA_SIZE(root) - push_space,
1691                               btrfs_leaf_data(right) +
1692                               leaf_data_end(root, right), push_space);
1693
1694         memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1695                               btrfs_item_nr_offset(push_items),
1696                              (btrfs_header_nritems(right) - push_items) *
1697                              sizeof(struct btrfs_item));
1698
1699         right_nritems = btrfs_header_nritems(right) - push_items;
1700         btrfs_set_header_nritems(right, right_nritems);
1701         push_space = BTRFS_LEAF_DATA_SIZE(root);
1702
1703         for (i = 0; i < right_nritems; i++) {
1704                 item = btrfs_item_nr(right, i);
1705
1706                 if (!right->map_token) {
1707                         map_extent_buffer(right, (unsigned long)item,
1708                                         sizeof(struct btrfs_item),
1709                                         &right->map_token, &right->kaddr,
1710                                         &right->map_start, &right->map_len,
1711                                         KM_USER1);
1712                 }
1713
1714                 push_space = push_space - btrfs_item_size(right, item);
1715                 btrfs_set_item_offset(right, item, push_space);
1716         }
1717         if (right->map_token) {
1718                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1719                 right->map_token = NULL;
1720         }
1721
1722         btrfs_mark_buffer_dirty(left);
1723         btrfs_mark_buffer_dirty(right);
1724
1725         btrfs_item_key(right, &disk_key, 0);
1726         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1727         if (wret)
1728                 ret = wret;
1729
1730         /* then fixup the leaf pointer in the path */
1731         if (path->slots[0] < push_items) {
1732                 path->slots[0] += old_left_nritems;
1733                 free_extent_buffer(path->nodes[0]);
1734                 path->nodes[0] = left;
1735                 path->slots[1] -= 1;
1736         } else {
1737                 free_extent_buffer(left);
1738                 path->slots[0] -= push_items;
1739         }
1740         BUG_ON(path->slots[0] < 0);
1741         return ret;
1742 }
1743
1744 /*
1745  * split the path's leaf in two, making sure there is at least data_size
1746  * available for the resulting leaf level of the path.
1747  *
1748  * returns 0 if all went well and < 0 on failure.
1749  */
1750 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
1751                       *root, struct btrfs_key *ins_key,
1752                       struct btrfs_path *path, int data_size)
1753 {
1754         struct extent_buffer *l;
1755         u32 nritems;
1756         int mid;
1757         int slot;
1758         struct extent_buffer *right;
1759         int space_needed = data_size + sizeof(struct btrfs_item);
1760         int data_copy_size;
1761         int rt_data_off;
1762         int i;
1763         int ret = 0;
1764         int wret;
1765         int double_split = 0;
1766         struct btrfs_disk_key disk_key;
1767
1768         /* first try to make some room by pushing left and right */
1769         wret = push_leaf_left(trans, root, path, data_size);
1770         if (wret < 0) {
1771                 return wret;
1772         }
1773         if (wret) {
1774                 wret = push_leaf_right(trans, root, path, data_size);
1775                 if (wret < 0)
1776                         return wret;
1777         }
1778         l = path->nodes[0];
1779
1780         /* did the pushes work? */
1781         if (btrfs_leaf_free_space(root, l) >=
1782             sizeof(struct btrfs_item) + data_size) {
1783                 return 0;
1784         }
1785
1786         if (!path->nodes[1]) {
1787                 ret = insert_new_root(trans, root, path, 1);
1788                 if (ret)
1789                         return ret;
1790         }
1791         slot = path->slots[0];
1792         nritems = btrfs_header_nritems(l);
1793         mid = (nritems + 1)/ 2;
1794
1795         right = btrfs_alloc_free_block(trans, root, root->leafsize,
1796                                        l->start, 0);
1797         if (IS_ERR(right))
1798                 return PTR_ERR(right);
1799
1800         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
1801         btrfs_set_header_bytenr(right, right->start);
1802         btrfs_set_header_generation(right, trans->transid);
1803         btrfs_set_header_owner(right, root->root_key.objectid);
1804         btrfs_set_header_level(right, 0);
1805         write_extent_buffer(right, root->fs_info->fsid,
1806                             (unsigned long)btrfs_header_fsid(right),
1807                             BTRFS_FSID_SIZE);
1808
1809         if (mid <= slot) {
1810                 if (nritems == 1 ||
1811                     leaf_space_used(l, mid, nritems - mid) + space_needed >
1812                         BTRFS_LEAF_DATA_SIZE(root)) {
1813                         if (slot >= nritems) {
1814                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1815                                 btrfs_set_header_nritems(right, 0);
1816                                 wret = insert_ptr(trans, root, path,
1817                                                   &disk_key, right->start,
1818                                                   path->slots[1] + 1, 1);
1819                                 if (wret)
1820                                         ret = wret;
1821                                 free_extent_buffer(path->nodes[0]);
1822                                 path->nodes[0] = right;
1823                                 path->slots[0] = 0;
1824                                 path->slots[1] += 1;
1825                                 return ret;
1826                         }
1827                         mid = slot;
1828                         if (mid != nritems &&
1829                             leaf_space_used(l, mid, nritems - mid) +
1830                             space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
1831                                 double_split = 1;
1832                         }
1833                 }
1834         } else {
1835                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
1836                         BTRFS_LEAF_DATA_SIZE(root)) {
1837                         if (slot == 0) {
1838                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1839                                 btrfs_set_header_nritems(right, 0);
1840                                 wret = insert_ptr(trans, root, path,
1841                                                   &disk_key,
1842                                                   right->start,
1843                                                   path->slots[1], 1);
1844                                 if (wret)
1845                                         ret = wret;
1846                                 free_extent_buffer(path->nodes[0]);
1847                                 path->nodes[0] = right;
1848                                 path->slots[0] = 0;
1849                                 if (path->slots[1] == 0) {
1850                                         wret = fixup_low_keys(trans, root,
1851                                                    path, &disk_key, 1);
1852                                         if (wret)
1853                                                 ret = wret;
1854                                 }
1855                                 return ret;
1856                         }
1857                         mid = slot;
1858                         double_split = 1;
1859                 }
1860         }
1861         nritems = nritems - mid;
1862         btrfs_set_header_nritems(right, nritems);
1863         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
1864
1865         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
1866                            btrfs_item_nr_offset(mid),
1867                            nritems * sizeof(struct btrfs_item));
1868
1869         copy_extent_buffer(right, l,
1870                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
1871                      data_copy_size, btrfs_leaf_data(l) +
1872                      leaf_data_end(root, l), data_copy_size);
1873
1874         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
1875                       btrfs_item_end_nr(l, mid);
1876
1877         for (i = 0; i < nritems; i++) {
1878                 struct btrfs_item *item = btrfs_item_nr(right, i);
1879                 u32 ioff;
1880
1881                 if (!right->map_token) {
1882                         map_extent_buffer(right, (unsigned long)item,
1883                                         sizeof(struct btrfs_item),
1884                                         &right->map_token, &right->kaddr,
1885                                         &right->map_start, &right->map_len,
1886                                         KM_USER1);
1887                 }
1888
1889                 ioff = btrfs_item_offset(right, item);
1890                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
1891         }
1892
1893         if (right->map_token) {
1894                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1895                 right->map_token = NULL;
1896         }
1897
1898         btrfs_set_header_nritems(l, mid);
1899         ret = 0;
1900         btrfs_item_key(right, &disk_key, 0);
1901         wret = insert_ptr(trans, root, path, &disk_key, right->start,
1902                           path->slots[1] + 1, 1);
1903         if (wret)
1904                 ret = wret;
1905
1906         btrfs_mark_buffer_dirty(right);
1907         btrfs_mark_buffer_dirty(l);
1908         BUG_ON(path->slots[0] != slot);
1909
1910         if (mid <= slot) {
1911                 free_extent_buffer(path->nodes[0]);
1912                 path->nodes[0] = right;
1913                 path->slots[0] -= mid;
1914                 path->slots[1] += 1;
1915         } else
1916                 free_extent_buffer(right);
1917
1918         BUG_ON(path->slots[0] < 0);
1919
1920         if (!double_split) {
1921                 return ret;
1922         }
1923
1924         right = btrfs_alloc_free_block(trans, root, root->leafsize,
1925                                        l->start, 0);
1926         if (IS_ERR(right))
1927                 return PTR_ERR(right);
1928
1929         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
1930         btrfs_set_header_bytenr(right, right->start);
1931         btrfs_set_header_generation(right, trans->transid);
1932         btrfs_set_header_owner(right, root->root_key.objectid);
1933         btrfs_set_header_level(right, 0);
1934         write_extent_buffer(right, root->fs_info->fsid,
1935                             (unsigned long)btrfs_header_fsid(right),
1936                             BTRFS_FSID_SIZE);
1937
1938         btrfs_cpu_key_to_disk(&disk_key, ins_key);
1939         btrfs_set_header_nritems(right, 0);
1940         wret = insert_ptr(trans, root, path,
1941                           &disk_key, right->start,
1942                           path->slots[1], 1);
1943         if (wret)
1944                 ret = wret;
1945         if (path->slots[1] == 0) {
1946                 wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1947                 if (wret)
1948                         ret = wret;
1949         }
1950         free_extent_buffer(path->nodes[0]);
1951         path->nodes[0] = right;
1952         path->slots[0] = 0;
1953         return ret;
1954 }
1955
1956 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
1957                         struct btrfs_root *root,
1958                         struct btrfs_path *path,
1959                         u32 new_size)
1960 {
1961         int ret = 0;
1962         int slot;
1963         int slot_orig;
1964         struct extent_buffer *leaf;
1965         struct btrfs_item *item;
1966         u32 nritems;
1967         unsigned int data_end;
1968         unsigned int old_data_start;
1969         unsigned int old_size;
1970         unsigned int size_diff;
1971         int i;
1972
1973         slot_orig = path->slots[0];
1974         leaf = path->nodes[0];
1975
1976         nritems = btrfs_header_nritems(leaf);
1977         data_end = leaf_data_end(root, leaf);
1978
1979         slot = path->slots[0];
1980         old_data_start = btrfs_item_offset_nr(leaf, slot);
1981         old_size = btrfs_item_size_nr(leaf, slot);
1982         BUG_ON(old_size <= new_size);
1983         size_diff = old_size - new_size;
1984
1985         BUG_ON(slot < 0);
1986         BUG_ON(slot >= nritems);
1987
1988         /*
1989          * item0..itemN ... dataN.offset..dataN.size .. data0.size
1990          */
1991         /* first correct the data pointers */
1992         for (i = slot; i < nritems; i++) {
1993                 u32 ioff;
1994                 item = btrfs_item_nr(leaf, i);
1995
1996                 if (!leaf->map_token) {
1997                         map_extent_buffer(leaf, (unsigned long)item,
1998                                         sizeof(struct btrfs_item),
1999                                         &leaf->map_token, &leaf->kaddr,
2000                                         &leaf->map_start, &leaf->map_len,
2001                                         KM_USER1);
2002                 }
2003
2004                 ioff = btrfs_item_offset(leaf, item);
2005                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2006         }
2007
2008         if (leaf->map_token) {
2009                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2010                 leaf->map_token = NULL;
2011         }
2012
2013         /* shift the data */
2014         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2015                       data_end + size_diff, btrfs_leaf_data(leaf) +
2016                       data_end, old_data_start + new_size - data_end);
2017
2018         item = btrfs_item_nr(leaf, slot);
2019         btrfs_set_item_size(leaf, item, new_size);
2020         btrfs_mark_buffer_dirty(leaf);
2021
2022         ret = 0;
2023         if (btrfs_leaf_free_space(root, leaf) < 0) {
2024                 btrfs_print_leaf(root, leaf);
2025                 BUG();
2026         }
2027         return ret;
2028 }
2029
2030 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2031                       struct btrfs_root *root, struct btrfs_path *path,
2032                       u32 data_size)
2033 {
2034         int ret = 0;
2035         int slot;
2036         int slot_orig;
2037         struct extent_buffer *leaf;
2038         struct btrfs_item *item;
2039         u32 nritems;
2040         unsigned int data_end;
2041         unsigned int old_data;
2042         unsigned int old_size;
2043         int i;
2044
2045         slot_orig = path->slots[0];
2046         leaf = path->nodes[0];
2047
2048         nritems = btrfs_header_nritems(leaf);
2049         data_end = leaf_data_end(root, leaf);
2050
2051         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2052                 btrfs_print_leaf(root, leaf);
2053                 BUG();
2054         }
2055         slot = path->slots[0];
2056         old_data = btrfs_item_end_nr(leaf, slot);
2057
2058         BUG_ON(slot < 0);
2059         if (slot >= nritems) {
2060                 btrfs_print_leaf(root, leaf);
2061                 printk("slot %d too large, nritems %d\n", slot, nritems);
2062                 BUG_ON(1);
2063         }
2064
2065         /*
2066          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2067          */
2068         /* first correct the data pointers */
2069         for (i = slot; i < nritems; i++) {
2070                 u32 ioff;
2071                 item = btrfs_item_nr(leaf, i);
2072
2073                 if (!leaf->map_token) {
2074                         map_extent_buffer(leaf, (unsigned long)item,
2075                                         sizeof(struct btrfs_item),
2076                                         &leaf->map_token, &leaf->kaddr,
2077                                         &leaf->map_start, &leaf->map_len,
2078                                         KM_USER1);
2079                 }
2080                 ioff = btrfs_item_offset(leaf, item);
2081                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2082         }
2083
2084         if (leaf->map_token) {
2085                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2086                 leaf->map_token = NULL;
2087         }
2088
2089         /* shift the data */
2090         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2091                       data_end - data_size, btrfs_leaf_data(leaf) +
2092                       data_end, old_data - data_end);
2093
2094         data_end = old_data;
2095         old_size = btrfs_item_size_nr(leaf, slot);
2096         item = btrfs_item_nr(leaf, slot);
2097         btrfs_set_item_size(leaf, item, old_size + data_size);
2098         btrfs_mark_buffer_dirty(leaf);
2099
2100         ret = 0;
2101         if (btrfs_leaf_free_space(root, leaf) < 0) {
2102                 btrfs_print_leaf(root, leaf);
2103                 BUG();
2104         }
2105         return ret;
2106 }
2107
2108 /*
2109  * Given a key and some data, insert an item into the tree.
2110  * This does all the path init required, making room in the tree if needed.
2111  */
2112 int btrfs_insert_empty_item(struct btrfs_trans_handle *trans,
2113                             struct btrfs_root *root,
2114                             struct btrfs_path *path,
2115                             struct btrfs_key *cpu_key, u32 data_size)
2116 {
2117         struct extent_buffer *leaf;
2118         struct btrfs_item *item;
2119         int ret = 0;
2120         int slot;
2121         int slot_orig;
2122         u32 nritems;
2123         unsigned int data_end;
2124         struct btrfs_disk_key disk_key;
2125
2126         btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2127
2128         /* create a root if there isn't one */
2129         if (!root->node)
2130                 BUG();
2131
2132         ret = btrfs_search_slot(trans, root, cpu_key, path, data_size, 1);
2133         if (ret == 0) {
2134                 return -EEXIST;
2135         }
2136         if (ret < 0)
2137                 goto out;
2138
2139         slot_orig = path->slots[0];
2140         leaf = path->nodes[0];
2141
2142         nritems = btrfs_header_nritems(leaf);
2143         data_end = leaf_data_end(root, leaf);
2144
2145         if (btrfs_leaf_free_space(root, leaf) <
2146             sizeof(struct btrfs_item) + data_size) {
2147                 btrfs_print_leaf(root, leaf);
2148                 printk("not enough freespace need %u have %d\n",
2149                        data_size, btrfs_leaf_free_space(root, leaf));
2150                 BUG();
2151         }
2152
2153         slot = path->slots[0];
2154         BUG_ON(slot < 0);
2155
2156         if (slot != nritems) {
2157                 int i;
2158                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2159
2160                 if (old_data < data_end) {
2161                         btrfs_print_leaf(root, leaf);
2162                         printk("slot %d old_data %d data_end %d\n",
2163                                slot, old_data, data_end);
2164                         BUG_ON(1);
2165                 }
2166                 /*
2167                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2168                  */
2169                 /* first correct the data pointers */
2170                 WARN_ON(leaf->map_token);
2171                 for (i = slot; i < nritems; i++) {
2172                         u32 ioff;
2173
2174                         item = btrfs_item_nr(leaf, i);
2175                         if (!leaf->map_token) {
2176                                 map_extent_buffer(leaf, (unsigned long)item,
2177                                         sizeof(struct btrfs_item),
2178                                         &leaf->map_token, &leaf->kaddr,
2179                                         &leaf->map_start, &leaf->map_len,
2180                                         KM_USER1);
2181                         }
2182
2183                         ioff = btrfs_item_offset(leaf, item);
2184                         btrfs_set_item_offset(leaf, item, ioff - data_size);
2185                 }
2186                 if (leaf->map_token) {
2187                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2188                         leaf->map_token = NULL;
2189                 }
2190
2191                 /* shift the items */
2192                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
2193                               btrfs_item_nr_offset(slot),
2194                               (nritems - slot) * sizeof(struct btrfs_item));
2195
2196                 /* shift the data */
2197                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2198                               data_end - data_size, btrfs_leaf_data(leaf) +
2199                               data_end, old_data - data_end);
2200                 data_end = old_data;
2201         }
2202
2203         /* setup the item for the new data */
2204         btrfs_set_item_key(leaf, &disk_key, slot);
2205         item = btrfs_item_nr(leaf, slot);
2206         btrfs_set_item_offset(leaf, item, data_end - data_size);
2207         btrfs_set_item_size(leaf, item, data_size);
2208         btrfs_set_header_nritems(leaf, nritems + 1);
2209         btrfs_mark_buffer_dirty(leaf);
2210
2211         ret = 0;
2212         if (slot == 0)
2213                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2214
2215         if (btrfs_leaf_free_space(root, leaf) < 0) {
2216                 btrfs_print_leaf(root, leaf);
2217                 BUG();
2218         }
2219 out:
2220         return ret;
2221 }
2222
2223 /*
2224  * Given a key and some data, insert an item into the tree.
2225  * This does all the path init required, making room in the tree if needed.
2226  */
2227 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2228                       *root, struct btrfs_key *cpu_key, void *data, u32
2229                       data_size)
2230 {
2231         int ret = 0;
2232         struct btrfs_path *path;
2233         struct extent_buffer *leaf;
2234         unsigned long ptr;
2235
2236         path = btrfs_alloc_path();
2237         BUG_ON(!path);
2238         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2239         if (!ret) {
2240                 leaf = path->nodes[0];
2241                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2242                 write_extent_buffer(leaf, data, ptr, data_size);
2243                 btrfs_mark_buffer_dirty(leaf);
2244         }
2245         btrfs_free_path(path);
2246         return ret;
2247 }
2248
2249 /*
2250  * delete the pointer from a given node.
2251  *
2252  * If the delete empties a node, the node is removed from the tree,
2253  * continuing all the way the root if required.  The root is converted into
2254  * a leaf if all the nodes are emptied.
2255  */
2256 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2257                    struct btrfs_path *path, int level, int slot)
2258 {
2259         struct extent_buffer *parent = path->nodes[level];
2260         u32 nritems;
2261         int ret = 0;
2262         int wret;
2263
2264         nritems = btrfs_header_nritems(parent);
2265         if (slot != nritems -1) {
2266                 memmove_extent_buffer(parent,
2267                               btrfs_node_key_ptr_offset(slot),
2268                               btrfs_node_key_ptr_offset(slot + 1),
2269                               sizeof(struct btrfs_key_ptr) *
2270                               (nritems - slot - 1));
2271         }
2272         nritems--;
2273         btrfs_set_header_nritems(parent, nritems);
2274         if (nritems == 0 && parent == root->node) {
2275                 BUG_ON(btrfs_header_level(root->node) != 1);
2276                 /* just turn the root into a leaf and break */
2277                 btrfs_set_header_level(root->node, 0);
2278         } else if (slot == 0) {
2279                 struct btrfs_disk_key disk_key;
2280
2281                 btrfs_node_key(parent, &disk_key, 0);
2282                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2283                 if (wret)
2284                         ret = wret;
2285         }
2286         btrfs_mark_buffer_dirty(parent);
2287         return ret;
2288 }
2289
2290 /*
2291  * delete the item at the leaf level in path.  If that empties
2292  * the leaf, remove it from the tree
2293  */
2294 int btrfs_del_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2295                    struct btrfs_path *path)
2296 {
2297         int slot;
2298         struct extent_buffer *leaf;
2299         struct btrfs_item *item;
2300         int doff;
2301         int dsize;
2302         int ret = 0;
2303         int wret;
2304         u32 nritems;
2305
2306         leaf = path->nodes[0];
2307         slot = path->slots[0];
2308         doff = btrfs_item_offset_nr(leaf, slot);
2309         dsize = btrfs_item_size_nr(leaf, slot);
2310         nritems = btrfs_header_nritems(leaf);
2311
2312         if (slot != nritems - 1) {
2313                 int i;
2314                 int data_end = leaf_data_end(root, leaf);
2315
2316                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2317                               data_end + dsize,
2318                               btrfs_leaf_data(leaf) + data_end,
2319                               doff - data_end);
2320
2321                 for (i = slot + 1; i < nritems; i++) {
2322                         u32 ioff;
2323
2324                         item = btrfs_item_nr(leaf, i);
2325                         if (!leaf->map_token) {
2326                                 map_extent_buffer(leaf, (unsigned long)item,
2327                                         sizeof(struct btrfs_item),
2328                                         &leaf->map_token, &leaf->kaddr,
2329                                         &leaf->map_start, &leaf->map_len,
2330                                         KM_USER1);
2331                         }
2332                         ioff = btrfs_item_offset(leaf, item);
2333                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2334                 }
2335
2336                 if (leaf->map_token) {
2337                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2338                         leaf->map_token = NULL;
2339                 }
2340
2341                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2342                               btrfs_item_nr_offset(slot + 1),
2343                               sizeof(struct btrfs_item) *
2344                               (nritems - slot - 1));
2345         }
2346         btrfs_set_header_nritems(leaf, nritems - 1);
2347         nritems--;
2348
2349         /* delete the leaf if we've emptied it */
2350         if (nritems == 0) {
2351                 if (leaf == root->node) {
2352                         btrfs_set_header_level(leaf, 0);
2353                 } else {
2354                         clean_tree_block(trans, root, leaf);
2355                         wait_on_tree_block_writeback(root, leaf);
2356                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
2357                         if (wret)
2358                                 ret = wret;
2359                         wret = btrfs_free_extent(trans, root,
2360                                                  leaf->start, leaf->len, 1);
2361                         if (wret)
2362                                 ret = wret;
2363                 }
2364         } else {
2365                 int used = leaf_space_used(leaf, 0, nritems);
2366                 if (slot == 0) {
2367                         struct btrfs_disk_key disk_key;
2368
2369                         btrfs_item_key(leaf, &disk_key, 0);
2370                         wret = fixup_low_keys(trans, root, path,
2371                                               &disk_key, 1);
2372                         if (wret)
2373                                 ret = wret;
2374                 }
2375
2376                 /* delete the leaf if it is mostly empty */
2377                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 3) {
2378                         /* push_leaf_left fixes the path.
2379                          * make sure the path still points to our leaf
2380                          * for possible call to del_ptr below
2381                          */
2382                         slot = path->slots[1];
2383                         extent_buffer_get(leaf);
2384
2385                         wret = push_leaf_left(trans, root, path, 1);
2386                         if (wret < 0 && wret != -ENOSPC)
2387                                 ret = wret;
2388
2389                         if (path->nodes[0] == leaf &&
2390                             btrfs_header_nritems(leaf)) {
2391                                 wret = push_leaf_right(trans, root, path, 1);
2392                                 if (wret < 0 && wret != -ENOSPC)
2393                                         ret = wret;
2394                         }
2395
2396                         if (btrfs_header_nritems(leaf) == 0) {
2397                                 u64 bytenr = leaf->start;
2398                                 u32 blocksize = leaf->len;
2399
2400                                 clean_tree_block(trans, root, leaf);
2401                                 wait_on_tree_block_writeback(root, leaf);
2402
2403                                 wret = del_ptr(trans, root, path, 1, slot);
2404                                 if (wret)
2405                                         ret = wret;
2406
2407                                 free_extent_buffer(leaf);
2408                                 wret = btrfs_free_extent(trans, root, bytenr,
2409                                                          blocksize, 1);
2410                                 if (wret)
2411                                         ret = wret;
2412                         } else {
2413                                 btrfs_mark_buffer_dirty(leaf);
2414                                 free_extent_buffer(leaf);
2415                         }
2416                 } else {
2417                         btrfs_mark_buffer_dirty(leaf);
2418                 }
2419         }
2420         return ret;
2421 }
2422
2423 /*
2424  * walk up the tree as far as required to find the next leaf.
2425  * returns 0 if it found something or 1 if there are no greater leaves.
2426  * returns < 0 on io errors.
2427  */
2428 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2429 {
2430         int slot;
2431         int level = 1;
2432         u64 bytenr;
2433         struct extent_buffer *c;
2434         struct extent_buffer *next = NULL;
2435
2436         while(level < BTRFS_MAX_LEVEL) {
2437                 if (!path->nodes[level])
2438                         return 1;
2439
2440                 slot = path->slots[level] + 1;
2441                 c = path->nodes[level];
2442                 if (slot >= btrfs_header_nritems(c)) {
2443                         level++;
2444                         continue;
2445                 }
2446
2447                 bytenr = btrfs_node_blockptr(c, slot);
2448                 if (next)
2449                         free_extent_buffer(next);
2450
2451                 if (path->reada)
2452                         reada_for_search(root, path, level, slot);
2453
2454                 next = read_tree_block(root, bytenr,
2455                                        btrfs_level_size(root, level -1));
2456                 break;
2457         }
2458         path->slots[level] = slot;
2459         while(1) {
2460                 level--;
2461                 c = path->nodes[level];
2462                 free_extent_buffer(c);
2463                 path->nodes[level] = next;
2464                 path->slots[level] = 0;
2465                 if (!level)
2466                         break;
2467                 if (path->reada)
2468                         reada_for_search(root, path, level, 0);
2469                 next = read_tree_block(root, btrfs_node_blockptr(next, 0),
2470                                        btrfs_level_size(root, level - 1));
2471         }
2472         return 0;
2473 }