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