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/module.h>
  20 #include "ctree.h"
  21 #include "disk-io.h"
  22 #include "transaction.h"
  23
  24 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  25                       *root, struct btrfs_path *path, int level);
  26 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  27                       *root, struct btrfs_key *ins_key,
  28                       struct btrfs_path *path, int data_size);
  29 static int push_node_left(struct btrfs_trans_handle *trans, struct btrfs_root
  30                           *root, struct buffer_head *dst, struct buffer_head
  31                           *src);
  32 static int balance_node_right(struct btrfs_trans_handle *trans, struct
  33                               btrfs_root *root, struct buffer_head *dst_buf,
  34                               struct buffer_head *src_buf);
  35 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  36                    struct btrfs_path *path, int level, int slot);
  37
  38 inline void btrfs_init_path(struct btrfs_path *p)
  39 {
  40         memset(p, 0, sizeof(*p));
  41 }
  42
  43 struct btrfs_path *btrfs_alloc_path(void)
  44 {
  45         struct btrfs_path *path;
  46         path = kmem_cache_alloc(btrfs_path_cachep, GFP_NOFS);
  47         if (path)
  48                 btrfs_init_path(path);
  49         return path;
  50 }
  51
  52 void btrfs_free_path(struct btrfs_path *p)
  53 {
  54         btrfs_release_path(NULL, p);
  55         kmem_cache_free(btrfs_path_cachep, p);
  56 }
  57
  58 void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
  59 {
  60         int i;
  61         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  62                 if (!p->nodes[i])
  63                         break;
  64                 btrfs_block_release(root, p->nodes[i]);
  65         }
  66         memset(p, 0, sizeof(*p));
  67 }
  68
  69 static int btrfs_cow_block(struct btrfs_trans_handle *trans, struct btrfs_root
  70                            *root, struct buffer_head *buf, struct buffer_head
  71                            *parent, int parent_slot, struct buffer_head
  72                            **cow_ret)
  73 {
  74         struct buffer_head *cow;
  75         struct btrfs_node *cow_node;
  76         int ret;
  77
  78         WARN_ON(!buffer_uptodate(buf));
  79         if (trans->transaction != root->fs_info->running_transaction) {
  80                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
  81                        root->fs_info->running_transaction->transid);
  82                 WARN_ON(1);
  83         }
  84         if (trans->transid != root->fs_info->generation) {
  85                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
  86                        root->fs_info->generation);
  87                 WARN_ON(1);
  88         }
  89         if (btrfs_header_generation(btrfs_buffer_header(buf)) ==
  90                                     trans->transid) {
  91                 *cow_ret = buf;
  92                 return 0;
  93         }
  94         cow = btrfs_alloc_free_block(trans, root, buf->b_blocknr);
  95         if (IS_ERR(cow))
  96                 return PTR_ERR(cow);
  97         cow_node = btrfs_buffer_node(cow);
  98         if (buf->b_size != root->blocksize || cow->b_size != root->blocksize)
  99                 WARN_ON(1);
 100         memcpy(cow_node, btrfs_buffer_node(buf), root->blocksize);
 101         btrfs_set_header_blocknr(&cow_node->header, bh_blocknr(cow));
 102         btrfs_set_header_generation(&cow_node->header, trans->transid);
 103         btrfs_set_header_owner(&cow_node->header, root->root_key.objectid);
 104         ret = btrfs_inc_ref(trans, root, buf);
 105         if (ret)
 106                 return ret;
 107         if (buf == root->node) {
 108                 root->node = cow;
 109                 get_bh(cow);
 110                 if (buf != root->commit_root) {
 111                         btrfs_free_extent(trans, root, bh_blocknr(buf), 1, 1);
 112                 }
 113                 btrfs_block_release(root, buf);
 114         } else {
 115                 btrfs_set_node_blockptr(btrfs_buffer_node(parent), parent_slot,
 116                                         bh_blocknr(cow));
 117                 btrfs_mark_buffer_dirty(parent);
 118                 btrfs_free_extent(trans, root, bh_blocknr(buf), 1, 1);
 119         }
 120         btrfs_block_release(root, buf);
 121         btrfs_mark_buffer_dirty(cow);
 122         *cow_ret = cow;
 123         return 0;
 124 }
 125
 126 /*
 127  * The leaf data grows from end-to-front in the node.
 128  * this returns the address of the start of the last item,
 129  * which is the stop of the leaf data stack
 130  */
 131 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 132                                          struct btrfs_leaf *leaf)
 133 {
 134         u32 nr = btrfs_header_nritems(&leaf->header);
 135         if (nr == 0)
 136                 return BTRFS_LEAF_DATA_SIZE(root);
 137         return btrfs_item_offset(leaf->items + nr - 1);
 138 }
 139
 140 /*
 141  * compare two keys in a memcmp fashion
 142  */
 143 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 144 {
 145         struct btrfs_key k1;
 146
 147         btrfs_disk_key_to_cpu(&k1, disk);
 148
 149         if (k1.objectid > k2->objectid)
 150                 return 1;
 151         if (k1.objectid < k2->objectid)
 152                 return -1;
 153         if (k1.flags > k2->flags)
 154                 return 1;
 155         if (k1.flags < k2->flags)
 156                 return -1;
 157         if (k1.offset > k2->offset)
 158                 return 1;
 159         if (k1.offset < k2->offset)
 160                 return -1;
 161         return 0;
 162 }
 163
 164 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
 165                       int level)
 166 {
 167         struct btrfs_node *parent = NULL;
 168         struct btrfs_node *node = btrfs_buffer_node(path->nodes[level]);
 169         int parent_slot;
 170         int slot;
 171         struct btrfs_key cpukey;
 172         u32 nritems = btrfs_header_nritems(&node->header);
 173
 174         if (path->nodes[level + 1])
 175                 parent = btrfs_buffer_node(path->nodes[level + 1]);
 176         parent_slot = path->slots[level + 1];
 177         slot = path->slots[level];
 178         BUG_ON(nritems == 0);
 179         if (parent) {
 180                 struct btrfs_disk_key *parent_key;
 181                 parent_key = &parent->ptrs[parent_slot].key;
 182                 BUG_ON(memcmp(parent_key, &node->ptrs[0].key,
 183                               sizeof(struct btrfs_disk_key)));
 184                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 185                        btrfs_header_blocknr(&node->header));
 186         }
 187         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 188         if (slot != 0) {
 189                 btrfs_disk_key_to_cpu(&cpukey, &node->ptrs[slot - 1].key);
 190                 BUG_ON(comp_keys(&node->ptrs[slot].key, &cpukey) <= 0);
 191         }
 192         if (slot < nritems - 1) {
 193                 btrfs_disk_key_to_cpu(&cpukey, &node->ptrs[slot + 1].key);
 194                 BUG_ON(comp_keys(&node->ptrs[slot].key, &cpukey) >= 0);
 195         }
 196         return 0;
 197 }
 198
 199 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 200                       int level)
 201 {
 202         struct btrfs_leaf *leaf = btrfs_buffer_leaf(path->nodes[level]);
 203         struct btrfs_node *parent = NULL;
 204         int parent_slot;
 205         int slot = path->slots[0];
 206         struct btrfs_key cpukey;
 207
 208         u32 nritems = btrfs_header_nritems(&leaf->header);
 209
 210         if (path->nodes[level + 1])
 211                 parent = btrfs_buffer_node(path->nodes[level + 1]);
 212         parent_slot = path->slots[level + 1];
 213         BUG_ON(btrfs_leaf_free_space(root, leaf) < 0);
 214
 215         if (nritems == 0)
 216                 return 0;
 217
 218         if (parent) {
 219                 struct btrfs_disk_key *parent_key;
 220                 parent_key = &parent->ptrs[parent_slot].key;
 221                 BUG_ON(memcmp(parent_key, &leaf->items[0].key,
 222                        sizeof(struct btrfs_disk_key)));
 223                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 224                        btrfs_header_blocknr(&leaf->header));
 225         }
 226         if (slot != 0) {
 227                 btrfs_disk_key_to_cpu(&cpukey, &leaf->items[slot - 1].key);
 228                 BUG_ON(comp_keys(&leaf->items[slot].key, &cpukey) <= 0);
 229                 BUG_ON(btrfs_item_offset(leaf->items + slot - 1) !=
 230                         btrfs_item_end(leaf->items + slot));
 231         }
 232         if (slot < nritems - 1) {
 233                 btrfs_disk_key_to_cpu(&cpukey, &leaf->items[slot + 1].key);
 234                 BUG_ON(comp_keys(&leaf->items[slot].key, &cpukey) >= 0);
 235                 BUG_ON(btrfs_item_offset(leaf->items + slot) !=
 236                         btrfs_item_end(leaf->items + slot + 1));
 237         }
 238         BUG_ON(btrfs_item_offset(leaf->items) +
 239                btrfs_item_size(leaf->items) != BTRFS_LEAF_DATA_SIZE(root));
 240         return 0;
 241 }
 242
 243 static int check_block(struct btrfs_root *root, struct btrfs_path *path,
 244                         int level)
 245 {
 246         struct btrfs_node *node = btrfs_buffer_node(path->nodes[level]);
 247         if (memcmp(node->header.fsid, root->fs_info->disk_super->fsid,
 248                    sizeof(node->header.fsid)))
 249                 BUG();
 250         if (level == 0)
 251                 return check_leaf(root, path, level);
 252         return check_node(root, path, level);
 253 }
 254
 255 /*
 256  * search for key in the array p.  items p are item_size apart
 257  * and there are 'max' items in p
 258  * the slot in the array is returned via slot, and it points to
 259  * the place where you would insert key if it is not found in
 260  * the array.
 261  *
 262  * slot may point to max if the key is bigger than all of the keys
 263  */
 264 static int generic_bin_search(char *p, int item_size, struct btrfs_key *key,
 265                        int max, int *slot)
 266 {
 267         int low = 0;
 268         int high = max;
 269         int mid;
 270         int ret;
 271         struct btrfs_disk_key *tmp;
 272
 273         while(low < high) {
 274                 mid = (low + high) / 2;
 275                 tmp = (struct btrfs_disk_key *)(p + mid * item_size);
 276                 ret = comp_keys(tmp, key);
 277
 278                 if (ret < 0)
 279                         low = mid + 1;
 280                 else if (ret > 0)
 281                         high = mid;
 282                 else {
 283                         *slot = mid;
 284                         return 0;
 285                 }
 286         }
 287         *slot = low;
 288         return 1;
 289 }
 290
 291 /*
 292  * simple bin_search frontend that does the right thing for
 293  * leaves vs nodes
 294  */
 295 static int bin_search(struct btrfs_node *c, struct btrfs_key *key, int *slot)
 296 {
 297         if (btrfs_is_leaf(c)) {
 298                 struct btrfs_leaf *l = (struct btrfs_leaf *)c;
 299                 return generic_bin_search((void *)l->items,
 300                                           sizeof(struct btrfs_item),
 301                                           key, btrfs_header_nritems(&c->header),
 302                                           slot);
 303         } else {
 304                 return generic_bin_search((void *)c->ptrs,
 305                                           sizeof(struct btrfs_key_ptr),
 306                                           key, btrfs_header_nritems(&c->header),
 307                                           slot);
 308         }
 309         return -1;
 310 }
 311
 312 static struct buffer_head *read_node_slot(struct btrfs_root *root,
 313                                    struct buffer_head *parent_buf,
 314                                    int slot)
 315 {
 316         struct btrfs_node *node = btrfs_buffer_node(parent_buf);
 317         if (slot < 0)
 318                 return NULL;
 319         if (slot >= btrfs_header_nritems(&node->header))
 320                 return NULL;
 321         return read_tree_block(root, btrfs_node_blockptr(node, slot));
 322 }
 323
 324 static int balance_level(struct btrfs_trans_handle *trans, struct btrfs_root
 325                          *root, struct btrfs_path *path, int level)
 326 {
 327         struct buffer_head *right_buf;
 328         struct buffer_head *mid_buf;
 329         struct buffer_head *left_buf;
 330         struct buffer_head *parent_buf = NULL;
 331         struct btrfs_node *right = NULL;
 332         struct btrfs_node *mid;
 333         struct btrfs_node *left = NULL;
 334         struct btrfs_node *parent = NULL;
 335         int ret = 0;
 336         int wret;
 337         int pslot;
 338         int orig_slot = path->slots[level];
 339         int err_on_enospc = 0;
 340         u64 orig_ptr;
 341
 342         if (level == 0)
 343                 return 0;
 344
 345         mid_buf = path->nodes[level];
 346         mid = btrfs_buffer_node(mid_buf);
 347         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 348
 349         if (level < BTRFS_MAX_LEVEL - 1)
 350                 parent_buf = path->nodes[level + 1];
 351         pslot = path->slots[level + 1];
 352
 353         /*
 354          * deal with the case where there is only one pointer in the root
 355          * by promoting the node below to a root
 356          */
 357         if (!parent_buf) {
 358                 struct buffer_head *child;
 359                 u64 blocknr = bh_blocknr(mid_buf);
 360
 361                 if (btrfs_header_nritems(&mid->header) != 1)
 362                         return 0;
 363
 364                 /* promote the child to a root */
 365                 child = read_node_slot(root, mid_buf, 0);
 366                 BUG_ON(!child);
 367                 root->node = child;
 368                 path->nodes[level] = NULL;
 369                 clean_tree_block(trans, root, mid_buf);
 370                 wait_on_buffer(mid_buf);
 371                 /* once for the path */
 372                 btrfs_block_release(root, mid_buf);
 373                 /* once for the root ptr */
 374                 btrfs_block_release(root, mid_buf);
 375                 return btrfs_free_extent(trans, root, blocknr, 1, 1);
 376         }
 377         parent = btrfs_buffer_node(parent_buf);
 378
 379         if (btrfs_header_nritems(&mid->header) >
 380             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 381                 return 0;
 382
 383         if (btrfs_header_nritems(&mid->header) < 2)
 384                 err_on_enospc = 1;
 385
 386         left_buf = read_node_slot(root, parent_buf, pslot - 1);
 387         right_buf = read_node_slot(root, parent_buf, pslot + 1);
 388
 389         /* first, try to make some room in the middle buffer */
 390         if (left_buf) {
 391                 wret = btrfs_cow_block(trans, root, left_buf,
 392                                        parent_buf, pslot - 1, &left_buf);
 393                 if (wret) {
 394                         ret = wret;
 395                         goto enospc;
 396                 }
 397                 left = btrfs_buffer_node(left_buf);
 398                 orig_slot += btrfs_header_nritems(&left->header);
 399                 wret = push_node_left(trans, root, left_buf, mid_buf);
 400                 if (wret < 0)
 401                         ret = wret;
 402                 if (btrfs_header_nritems(&mid->header) < 2)
 403                         err_on_enospc = 1;
 404         }
 405
 406         /*
 407          * then try to empty the right most buffer into the middle
 408          */
 409         if (right_buf) {
 410                 wret = btrfs_cow_block(trans, root, right_buf,
 411                                        parent_buf, pslot + 1, &right_buf);
 412                 if (wret) {
 413                         ret = wret;
 414                         goto enospc;
 415                 }
 416
 417                 right = btrfs_buffer_node(right_buf);
 418                 wret = push_node_left(trans, root, mid_buf, right_buf);
 419                 if (wret < 0 && wret != -ENOSPC)
 420                         ret = wret;
 421                 if (btrfs_header_nritems(&right->header) == 0) {
 422                         u64 blocknr = bh_blocknr(right_buf);
 423                         clean_tree_block(trans, root, right_buf);
 424                         wait_on_buffer(right_buf);
 425                         btrfs_block_release(root, right_buf);
 426                         right_buf = NULL;
 427                         right = NULL;
 428                         wret = del_ptr(trans, root, path, level + 1, pslot +
 429                                        1);
 430                         if (wret)
 431                                 ret = wret;
 432                         wret = btrfs_free_extent(trans, root, blocknr, 1, 1);
 433                         if (wret)
 434                                 ret = wret;
 435                 } else {
 436                         btrfs_memcpy(root, parent,
 437                                      &parent->ptrs[pslot + 1].key,
 438                                      &right->ptrs[0].key,
 439                                      sizeof(struct btrfs_disk_key));
 440                         btrfs_mark_buffer_dirty(parent_buf);
 441                 }
 442         }
 443         if (btrfs_header_nritems(&mid->header) == 1) {
 444                 /*
 445                  * we're not allowed to leave a node with one item in the
 446                  * tree during a delete.  A deletion from lower in the tree
 447                  * could try to delete the only pointer in this node.
 448                  * So, pull some keys from the left.
 449                  * There has to be a left pointer at this point because
 450                  * otherwise we would have pulled some pointers from the
 451                  * right
 452                  */
 453                 BUG_ON(!left_buf);
 454                 wret = balance_node_right(trans, root, mid_buf, left_buf);
 455                 if (wret < 0) {
 456                         ret = wret;
 457                         goto enospc;
 458                 }
 459                 BUG_ON(wret == 1);
 460         }
 461         if (btrfs_header_nritems(&mid->header) == 0) {
 462                 /* we've managed to empty the middle node, drop it */
 463                 u64 blocknr = bh_blocknr(mid_buf);
 464                 clean_tree_block(trans, root, mid_buf);
 465                 wait_on_buffer(mid_buf);
 466                 btrfs_block_release(root, mid_buf);
 467                 mid_buf = NULL;
 468                 mid = NULL;
 469                 wret = del_ptr(trans, root, path, level + 1, pslot);
 470                 if (wret)
 471                         ret = wret;
 472                 wret = btrfs_free_extent(trans, root, blocknr, 1, 1);
 473                 if (wret)
 474                         ret = wret;
 475         } else {
 476                 /* update the parent key to reflect our changes */
 477                 btrfs_memcpy(root, parent,
 478                              &parent->ptrs[pslot].key, &mid->ptrs[0].key,
 479                              sizeof(struct btrfs_disk_key));
 480                 btrfs_mark_buffer_dirty(parent_buf);
 481         }
 482
 483         /* update the path */
 484         if (left_buf) {
 485                 if (btrfs_header_nritems(&left->header) > orig_slot) {
 486                         get_bh(left_buf);
 487                         path->nodes[level] = left_buf;
 488                         path->slots[level + 1] -= 1;
 489                         path->slots[level] = orig_slot;
 490                         if (mid_buf)
 491                                 btrfs_block_release(root, mid_buf);
 492                 } else {
 493                         orig_slot -= btrfs_header_nritems(&left->header);
 494                         path->slots[level] = orig_slot;
 495                 }
 496         }
 497         /* double check we haven't messed things up */
 498         check_block(root, path, level);
 499         if (orig_ptr !=
 500             btrfs_node_blockptr(btrfs_buffer_node(path->nodes[level]),
 501                                 path->slots[level]))
 502                 BUG();
 503 enospc:
 504         if (right_buf)
 505                 btrfs_block_release(root, right_buf);
 506         if (left_buf)
 507                 btrfs_block_release(root, left_buf);
 508         return ret;
 509 }
 510
 511 /* returns zero if the push worked, non-zero otherwise */
 512 static int push_nodes_for_insert(struct btrfs_trans_handle *trans,
 513                                 struct btrfs_root *root,
 514                                 struct btrfs_path *path, int level)
 515 {
 516         struct buffer_head *right_buf;
 517         struct buffer_head *mid_buf;
 518         struct buffer_head *left_buf;
 519         struct buffer_head *parent_buf = NULL;
 520         struct btrfs_node *right = NULL;
 521         struct btrfs_node *mid;
 522         struct btrfs_node *left = NULL;
 523         struct btrfs_node *parent = NULL;
 524         int ret = 0;
 525         int wret;
 526         int pslot;
 527         int orig_slot = path->slots[level];
 528         u64 orig_ptr;
 529
 530         if (level == 0)
 531                 return 1;
 532
 533         mid_buf = path->nodes[level];
 534         mid = btrfs_buffer_node(mid_buf);
 535         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 536
 537         if (level < BTRFS_MAX_LEVEL - 1)
 538                 parent_buf = path->nodes[level + 1];
 539         pslot = path->slots[level + 1];
 540
 541         if (!parent_buf)
 542                 return 1;
 543         parent = btrfs_buffer_node(parent_buf);
 544
 545         left_buf = read_node_slot(root, parent_buf, pslot - 1);
 546
 547         /* first, try to make some room in the middle buffer */
 548         if (left_buf) {
 549                 u32 left_nr;
 550                 left = btrfs_buffer_node(left_buf);
 551                 left_nr = btrfs_header_nritems(&left->header);
 552                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 553                         wret = 1;
 554                 } else {
 555                         ret = btrfs_cow_block(trans, root, left_buf, parent_buf,
 556                                               pslot - 1, &left_buf);
 557                         if (ret)
 558                                 wret = 1;
 559                         else {
 560                                 left = btrfs_buffer_node(left_buf);
 561                                 wret = push_node_left(trans, root,
 562                                                       left_buf, mid_buf);
 563                         }
 564                 }
 565                 if (wret < 0)
 566                         ret = wret;
 567                 if (wret == 0) {
 568                         orig_slot += left_nr;
 569                         btrfs_memcpy(root, parent,
 570                                      &parent->ptrs[pslot].key,
 571                                      &mid->ptrs[0].key,
 572                                      sizeof(struct btrfs_disk_key));
 573                         btrfs_mark_buffer_dirty(parent_buf);
 574                         if (btrfs_header_nritems(&left->header) > orig_slot) {
 575                                 path->nodes[level] = left_buf;
 576                                 path->slots[level + 1] -= 1;
 577                                 path->slots[level] = orig_slot;
 578                                 btrfs_block_release(root, mid_buf);
 579                         } else {
 580                                 orig_slot -=
 581                                         btrfs_header_nritems(&left->header);
 582                                 path->slots[level] = orig_slot;
 583                                 btrfs_block_release(root, left_buf);
 584                         }
 585                         check_node(root, path, level);
 586                         return 0;
 587                 }
 588                 btrfs_block_release(root, left_buf);
 589         }
 590         right_buf = read_node_slot(root, parent_buf, pslot + 1);
 591
 592         /*
 593          * then try to empty the right most buffer into the middle
 594          */
 595         if (right_buf) {
 596                 u32 right_nr;
 597                 right = btrfs_buffer_node(right_buf);
 598                 right_nr = btrfs_header_nritems(&right->header);
 599                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 600                         wret = 1;
 601                 } else {
 602                         ret = btrfs_cow_block(trans, root, right_buf,
 603                                               parent_buf, pslot + 1,
 604                                               &right_buf);
 605                         if (ret)
 606                                 wret = 1;
 607                         else {
 608                                 right = btrfs_buffer_node(right_buf);
 609                                 wret = balance_node_right(trans, root,
 610                                                           right_buf, mid_buf);
 611                         }
 612                 }
 613                 if (wret < 0)
 614                         ret = wret;
 615                 if (wret == 0) {
 616                         btrfs_memcpy(root, parent,
 617                                      &parent->ptrs[pslot + 1].key,
 618                                      &right->ptrs[0].key,
 619                                      sizeof(struct btrfs_disk_key));
 620                         btrfs_mark_buffer_dirty(parent_buf);
 621                         if (btrfs_header_nritems(&mid->header) <= orig_slot) {
 622                                 path->nodes[level] = right_buf;
 623                                 path->slots[level + 1] += 1;
 624                                 path->slots[level] = orig_slot -
 625                                         btrfs_header_nritems(&mid->header);
 626                                 btrfs_block_release(root, mid_buf);
 627                         } else {
 628                                 btrfs_block_release(root, right_buf);
 629                         }
 630                         check_node(root, path, level);
 631                         return 0;
 632                 }
 633                 btrfs_block_release(root, right_buf);
 634         }
 635         check_node(root, path, level);
 636         return 1;
 637 }
 638
 639 /*
 640  * look for key in the tree.  path is filled in with nodes along the way
 641  * if key is found, we return zero and you can find the item in the leaf
 642  * level of the path (level 0)
 643  *
 644  * If the key isn't found, the path points to the slot where it should
 645  * be inserted, and 1 is returned.  If there are other errors during the
 646  * search a negative error number is returned.
 647  *
 648  * if ins_len > 0, nodes and leaves will be split as we walk down the
 649  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
 650  * possible)
 651  */
 652 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
 653                       *root, struct btrfs_key *key, struct btrfs_path *p, int
 654                       ins_len, int cow)
 655 {
 656         struct buffer_head *b;
 657         struct buffer_head *cow_buf;
 658         struct btrfs_node *c;
 659         int slot;
 660         int ret;
 661         int level;
 662
 663         WARN_ON(p->nodes[0] != NULL);
 664         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
 665 again:
 666         b = root->node;
 667         get_bh(b);
 668         while (b) {
 669                 c = btrfs_buffer_node(b);
 670                 level = btrfs_header_level(&c->header);
 671                 if (cow) {
 672                         int wret;
 673                         wret = btrfs_cow_block(trans, root, b,
 674                                                p->nodes[level + 1],
 675                                                p->slots[level + 1],
 676                                                &cow_buf);
 677                         if (wret) {
 678                                 btrfs_block_release(root, cow_buf);
 679                                 return wret;
 680                         }
 681                         b = cow_buf;
 682                         c = btrfs_buffer_node(b);
 683                 }
 684                 BUG_ON(!cow && ins_len);
 685                 if (level != btrfs_header_level(&c->header))
 686                         WARN_ON(1);
 687                 level = btrfs_header_level(&c->header);
 688                 p->nodes[level] = b;
 689                 ret = check_block(root, p, level);
 690                 if (ret)
 691                         return -1;
 692                 ret = bin_search(c, key, &slot);
 693                 if (!btrfs_is_leaf(c)) {
 694                         if (ret && slot > 0)
 695                                 slot -= 1;
 696                         p->slots[level] = slot;
 697                         if (ins_len > 0 && btrfs_header_nritems(&c->header) >=
 698                             BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 699                                 int sret = split_node(trans, root, p, level);
 700                                 BUG_ON(sret > 0);
 701                                 if (sret)
 702                                         return sret;
 703                                 b = p->nodes[level];
 704                                 c = btrfs_buffer_node(b);
 705                                 slot = p->slots[level];
 706                         } else if (ins_len < 0) {
 707                                 int sret = balance_level(trans, root, p,
 708                                                          level);
 709                                 if (sret)
 710                                         return sret;
 711                                 b = p->nodes[level];
 712                                 if (!b)
 713                                         goto again;
 714                                 c = btrfs_buffer_node(b);
 715                                 slot = p->slots[level];
 716                                 BUG_ON(btrfs_header_nritems(&c->header) == 1);
 717                         }
 718                         b = read_tree_block(root, btrfs_node_blockptr(c, slot));
 719                 } else {
 720                         struct btrfs_leaf *l = (struct btrfs_leaf *)c;
 721                         p->slots[level] = slot;
 722                         if (ins_len > 0 && btrfs_leaf_free_space(root, l) <
 723                             sizeof(struct btrfs_item) + ins_len) {
 724                                 int sret = split_leaf(trans, root, key,
 725                                                       p, ins_len);
 726                                 BUG_ON(sret > 0);
 727                                 if (sret)
 728                                         return sret;
 729                         }
 730                         return ret;
 731                 }
 732         }
 733         return 1;
 734 }
 735
 736 /*
 737  * adjust the pointers going up the tree, starting at level
 738  * making sure the right key of each node is points to 'key'.
 739  * This is used after shifting pointers to the left, so it stops
 740  * fixing up pointers when a given leaf/node is not in slot 0 of the
 741  * higher levels
 742  *
 743  * If this fails to write a tree block, it returns -1, but continues
 744  * fixing up the blocks in ram so the tree is consistent.
 745  */
 746 static int fixup_low_keys(struct btrfs_trans_handle *trans, struct btrfs_root
 747                           *root, struct btrfs_path *path, struct btrfs_disk_key
 748                           *key, int level)
 749 {
 750         int i;
 751         int ret = 0;
 752         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
 753                 struct btrfs_node *t;
 754                 int tslot = path->slots[i];
 755                 if (!path->nodes[i])
 756                         break;
 757                 t = btrfs_buffer_node(path->nodes[i]);
 758                 btrfs_memcpy(root, t, &t->ptrs[tslot].key, key, sizeof(*key));
 759                 btrfs_mark_buffer_dirty(path->nodes[i]);
 760                 if (tslot != 0)
 761                         break;
 762         }
 763         return ret;
 764 }
 765
 766 /*
 767  * try to push data from one node into the next node left in the
 768  * tree.
 769  *
 770  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
 771  * error, and > 0 if there was no room in the left hand block.
 772  */
 773 static int push_node_left(struct btrfs_trans_handle *trans, struct btrfs_root
 774                           *root, struct buffer_head *dst_buf, struct
 775                           buffer_head *src_buf)
 776 {
 777         struct btrfs_node *src = btrfs_buffer_node(src_buf);
 778         struct btrfs_node *dst = btrfs_buffer_node(dst_buf);
 779         int push_items = 0;
 780         int src_nritems;
 781         int dst_nritems;
 782         int ret = 0;
 783
 784         src_nritems = btrfs_header_nritems(&src->header);
 785         dst_nritems = btrfs_header_nritems(&dst->header);
 786         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
 787
 788         if (push_items <= 0) {
 789                 return 1;
 790         }
 791
 792         if (src_nritems < push_items)
 793                 push_items = src_nritems;
 794
 795         btrfs_memcpy(root, dst, dst->ptrs + dst_nritems, src->ptrs,
 796                      push_items * sizeof(struct btrfs_key_ptr));
 797         if (push_items < src_nritems) {
 798                 btrfs_memmove(root, src, src->ptrs, src->ptrs + push_items,
 799                         (src_nritems - push_items) *
 800                         sizeof(struct btrfs_key_ptr));
 801         }
 802         btrfs_set_header_nritems(&src->header, src_nritems - push_items);
 803         btrfs_set_header_nritems(&dst->header, dst_nritems + push_items);
 804         btrfs_mark_buffer_dirty(src_buf);
 805         btrfs_mark_buffer_dirty(dst_buf);
 806         return ret;
 807 }
 808
 809 /*
 810  * try to push data from one node into the next node right in the
 811  * tree.
 812  *
 813  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
 814  * error, and > 0 if there was no room in the right hand block.
 815  *
 816  * this will  only push up to 1/2 the contents of the left node over
 817  */
 818 static int balance_node_right(struct btrfs_trans_handle *trans, struct
 819                               btrfs_root *root, struct buffer_head *dst_buf,
 820                               struct buffer_head *src_buf)
 821 {
 822         struct btrfs_node *src = btrfs_buffer_node(src_buf);
 823         struct btrfs_node *dst = btrfs_buffer_node(dst_buf);
 824         int push_items = 0;
 825         int max_push;
 826         int src_nritems;
 827         int dst_nritems;
 828         int ret = 0;
 829
 830         src_nritems = btrfs_header_nritems(&src->header);
 831         dst_nritems = btrfs_header_nritems(&dst->header);
 832         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
 833         if (push_items <= 0) {
 834                 return 1;
 835         }
 836
 837         max_push = src_nritems / 2 + 1;
 838         /* don't try to empty the node */
 839         if (max_push > src_nritems)
 840                 return 1;
 841         if (max_push < push_items)
 842                 push_items = max_push;
 843
 844         btrfs_memmove(root, dst, dst->ptrs + push_items, dst->ptrs,
 845                       dst_nritems * sizeof(struct btrfs_key_ptr));
 846
 847         btrfs_memcpy(root, dst, dst->ptrs,
 848                      src->ptrs + src_nritems - push_items,
 849                      push_items * sizeof(struct btrfs_key_ptr));
 850
 851         btrfs_set_header_nritems(&src->header, src_nritems - push_items);
 852         btrfs_set_header_nritems(&dst->header, dst_nritems + push_items);
 853
 854         btrfs_mark_buffer_dirty(src_buf);
 855         btrfs_mark_buffer_dirty(dst_buf);
 856         return ret;
 857 }
 858
 859 /*
 860  * helper function to insert a new root level in the tree.
 861  * A new node is allocated, and a single item is inserted to
 862  * point to the existing root
 863  *
 864  * returns zero on success or < 0 on failure.
 865  */
 866 static int insert_new_root(struct btrfs_trans_handle *trans, struct btrfs_root
 867                            *root, struct btrfs_path *path, int level)
 868 {
 869         struct buffer_head *t;
 870         struct btrfs_node *lower;
 871         struct btrfs_node *c;
 872         struct btrfs_disk_key *lower_key;
 873
 874         BUG_ON(path->nodes[level]);
 875         BUG_ON(path->nodes[level-1] != root->node);
 876
 877         t = btrfs_alloc_free_block(trans, root, root->node->b_blocknr);
 878         if (IS_ERR(t))
 879                 return PTR_ERR(t);
 880         c = btrfs_buffer_node(t);
 881         memset(c, 0, root->blocksize);
 882         btrfs_set_header_nritems(&c->header, 1);
 883         btrfs_set_header_level(&c->header, level);
 884         btrfs_set_header_blocknr(&c->header, bh_blocknr(t));
 885         btrfs_set_header_generation(&c->header, trans->transid);
 886         btrfs_set_header_owner(&c->header, root->root_key.objectid);
 887         lower = btrfs_buffer_node(path->nodes[level-1]);
 888         memcpy(c->header.fsid, root->fs_info->disk_super->fsid,
 889                sizeof(c->header.fsid));
 890         if (btrfs_is_leaf(lower))
 891                 lower_key = &((struct btrfs_leaf *)lower)->items[0].key;
 892         else
 893                 lower_key = &lower->ptrs[0].key;
 894         btrfs_memcpy(root, c, &c->ptrs[0].key, lower_key,
 895                      sizeof(struct btrfs_disk_key));
 896         btrfs_set_node_blockptr(c, 0, bh_blocknr(path->nodes[level - 1]));
 897
 898         btrfs_mark_buffer_dirty(t);
 899
 900         /* the super has an extra ref to root->node */
 901         btrfs_block_release(root, root->node);
 902         root->node = t;
 903         get_bh(t);
 904         path->nodes[level] = t;
 905         path->slots[level] = 0;
 906         return 0;
 907 }
 908
 909 /*
 910  * worker function to insert a single pointer in a node.
 911  * the node should have enough room for the pointer already
 912  *
 913  * slot and level indicate where you want the key to go, and
 914  * blocknr is the block the key points to.
 915  *
 916  * returns zero on success and < 0 on any error
 917  */
 918 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
 919                       *root, struct btrfs_path *path, struct btrfs_disk_key
 920                       *key, u64 blocknr, int slot, int level)
 921 {
 922         struct btrfs_node *lower;
 923         int nritems;
 924
 925         BUG_ON(!path->nodes[level]);
 926         lower = btrfs_buffer_node(path->nodes[level]);
 927         nritems = btrfs_header_nritems(&lower->header);
 928         if (slot > nritems)
 929                 BUG();
 930         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
 931                 BUG();
 932         if (slot != nritems) {
 933                 btrfs_memmove(root, lower, lower->ptrs + slot + 1,
 934                               lower->ptrs + slot,
 935                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
 936         }
 937         btrfs_memcpy(root, lower, &lower->ptrs[slot].key,
 938                      key, sizeof(struct btrfs_disk_key));
 939         btrfs_set_node_blockptr(lower, slot, blocknr);
 940         btrfs_set_header_nritems(&lower->header, nritems + 1);
 941         btrfs_mark_buffer_dirty(path->nodes[level]);
 942         check_node(root, path, level);
 943         return 0;
 944 }
 945
 946 /*
 947  * split the node at the specified level in path in two.
 948  * The path is corrected to point to the appropriate node after the split
 949  *
 950  * Before splitting this tries to make some room in the node by pushing
 951  * left and right, if either one works, it returns right away.
 952  *
 953  * returns 0 on success and < 0 on failure
 954  */
 955 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
 956                       *root, struct btrfs_path *path, int level)
 957 {
 958         struct buffer_head *t;
 959         struct btrfs_node *c;
 960         struct buffer_head *split_buffer;
 961         struct btrfs_node *split;
 962         int mid;
 963         int ret;
 964         int wret;
 965         u32 c_nritems;
 966
 967         t = path->nodes[level];
 968         c = btrfs_buffer_node(t);
 969         if (t == root->node) {
 970                 /* trying to split the root, lets make a new one */
 971                 ret = insert_new_root(trans, root, path, level + 1);
 972                 if (ret)
 973                         return ret;
 974         } else {
 975                 ret = push_nodes_for_insert(trans, root, path, level);
 976                 t = path->nodes[level];
 977                 c = btrfs_buffer_node(t);
 978                 if (!ret &&
 979                     btrfs_header_nritems(&c->header) <
 980                     BTRFS_NODEPTRS_PER_BLOCK(root) - 1)
 981                         return 0;
 982                 if (ret < 0)
 983                         return ret;
 984         }
 985
 986         c_nritems = btrfs_header_nritems(&c->header);
 987         split_buffer = btrfs_alloc_free_block(trans, root, t->b_blocknr);
 988         if (IS_ERR(split_buffer))
 989                 return PTR_ERR(split_buffer);
 990
 991         split = btrfs_buffer_node(split_buffer);
 992         btrfs_set_header_flags(&split->header, btrfs_header_flags(&c->header));
 993         btrfs_set_header_level(&split->header, btrfs_header_level(&c->header));
 994         btrfs_set_header_blocknr(&split->header, bh_blocknr(split_buffer));
 995         btrfs_set_header_generation(&split->header, trans->transid);
 996         btrfs_set_header_owner(&split->header, root->root_key.objectid);
 997         memcpy(split->header.fsid, root->fs_info->disk_super->fsid,
 998                sizeof(split->header.fsid));
 999         mid = (c_nritems + 1) / 2;
1000         btrfs_memcpy(root, split, split->ptrs, c->ptrs + mid,
1001                      (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1002         btrfs_set_header_nritems(&split->header, c_nritems - mid);
1003         btrfs_set_header_nritems(&c->header, mid);
1004         ret = 0;
1005
1006         btrfs_mark_buffer_dirty(t);
1007         btrfs_mark_buffer_dirty(split_buffer);
1008         wret = insert_ptr(trans, root, path, &split->ptrs[0].key,
1009                           bh_blocknr(split_buffer), path->slots[level + 1] + 1,
1010                           level + 1);
1011         if (wret)
1012                 ret = wret;
1013
1014         if (path->slots[level] >= mid) {
1015                 path->slots[level] -= mid;
1016                 btrfs_block_release(root, t);
1017                 path->nodes[level] = split_buffer;
1018                 path->slots[level + 1] += 1;
1019         } else {
1020                 btrfs_block_release(root, split_buffer);
1021         }
1022         return ret;
1023 }
1024
1025 /*
1026  * how many bytes are required to store the items in a leaf.  start
1027  * and nr indicate which items in the leaf to check.  This totals up the
1028  * space used both by the item structs and the item data
1029  */
1030 static int leaf_space_used(struct btrfs_leaf *l, int start, int nr)
1031 {
1032         int data_len;
1033         int nritems = btrfs_header_nritems(&l->header);
1034         int end = min(nritems, start + nr) - 1;
1035
1036         if (!nr)
1037                 return 0;
1038         data_len = btrfs_item_end(l->items + start);
1039         data_len = data_len - btrfs_item_offset(l->items + end);
1040         data_len += sizeof(struct btrfs_item) * nr;
1041         WARN_ON(data_len < 0);
1042         return data_len;
1043 }
1044
1045 /*
1046  * The space between the end of the leaf items and
1047  * the start of the leaf data.  IOW, how much room
1048  * the leaf has left for both items and data
1049  */
1050 int btrfs_leaf_free_space(struct btrfs_root *root, struct btrfs_leaf *leaf)
1051 {
1052         int nritems = btrfs_header_nritems(&leaf->header);
1053         return BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1054 }
1055
1056 /*
1057  * push some data in the path leaf to the right, trying to free up at
1058  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1059  *
1060  * returns 1 if the push failed because the other node didn't have enough
1061  * room, 0 if everything worked out and < 0 if there were major errors.
1062  */
1063 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1064                            *root, struct btrfs_path *path, int data_size)
1065 {
1066         struct buffer_head *left_buf = path->nodes[0];
1067         struct btrfs_leaf *left = btrfs_buffer_leaf(left_buf);
1068         struct btrfs_leaf *right;
1069         struct buffer_head *right_buf;
1070         struct buffer_head *upper;
1071         struct btrfs_node *upper_node;
1072         int slot;
1073         int i;
1074         int free_space;
1075         int push_space = 0;
1076         int push_items = 0;
1077         struct btrfs_item *item;
1078         u32 left_nritems;
1079         u32 right_nritems;
1080         int ret;
1081
1082         slot = path->slots[1];
1083         if (!path->nodes[1]) {
1084                 return 1;
1085         }
1086         upper = path->nodes[1];
1087         upper_node = btrfs_buffer_node(upper);
1088         if (slot >= btrfs_header_nritems(&upper_node->header) - 1) {
1089                 return 1;
1090         }
1091         right_buf = read_tree_block(root,
1092                     btrfs_node_blockptr(btrfs_buffer_node(upper), slot + 1));
1093         right = btrfs_buffer_leaf(right_buf);
1094         free_space = btrfs_leaf_free_space(root, right);
1095         if (free_space < data_size + sizeof(struct btrfs_item)) {
1096                 btrfs_block_release(root, right_buf);
1097                 return 1;
1098         }
1099         /* cow and double check */
1100         ret = btrfs_cow_block(trans, root, right_buf, upper,
1101                               slot + 1, &right_buf);
1102         if (ret) {
1103                 btrfs_block_release(root, right_buf);
1104                 return 1;
1105         }
1106         right = btrfs_buffer_leaf(right_buf);
1107         free_space = btrfs_leaf_free_space(root, right);
1108         if (free_space < data_size + sizeof(struct btrfs_item)) {
1109                 btrfs_block_release(root, right_buf);
1110                 return 1;
1111         }
1112
1113         left_nritems = btrfs_header_nritems(&left->header);
1114         if (left_nritems == 0) {
1115                 btrfs_block_release(root, right_buf);
1116                 return 1;
1117         }
1118         for (i = left_nritems - 1; i >= 1; i--) {
1119                 item = left->items + i;
1120                 if (path->slots[0] == i)
1121                         push_space += data_size + sizeof(*item);
1122                 if (btrfs_item_size(item) + sizeof(*item) + push_space >
1123                     free_space)
1124                         break;
1125                 push_items++;
1126                 push_space += btrfs_item_size(item) + sizeof(*item);
1127         }
1128         if (push_items == 0) {
1129                 btrfs_block_release(root, right_buf);
1130                 return 1;
1131         }
1132         if (push_items == left_nritems)
1133                 WARN_ON(1);
1134         right_nritems = btrfs_header_nritems(&right->header);
1135         /* push left to right */
1136         push_space = btrfs_item_end(left->items + left_nritems - push_items);
1137         push_space -= leaf_data_end(root, left);
1138         /* make room in the right data area */
1139         btrfs_memmove(root, right, btrfs_leaf_data(right) +
1140                       leaf_data_end(root, right) - push_space,
1141                       btrfs_leaf_data(right) +
1142                       leaf_data_end(root, right), BTRFS_LEAF_DATA_SIZE(root) -
1143                       leaf_data_end(root, right));
1144         /* copy from the left data area */
1145         btrfs_memcpy(root, right, btrfs_leaf_data(right) +
1146                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1147                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1148                      push_space);
1149         btrfs_memmove(root, right, right->items + push_items, right->items,
1150                 right_nritems * sizeof(struct btrfs_item));
1151         /* copy the items from left to right */
1152         btrfs_memcpy(root, right, right->items, left->items +
1153                      left_nritems - push_items,
1154                      push_items * sizeof(struct btrfs_item));
1155
1156         /* update the item pointers */
1157         right_nritems += push_items;
1158         btrfs_set_header_nritems(&right->header, right_nritems);
1159         push_space = BTRFS_LEAF_DATA_SIZE(root);
1160         for (i = 0; i < right_nritems; i++) {
1161                 btrfs_set_item_offset(right->items + i, push_space -
1162                                       btrfs_item_size(right->items + i));
1163                 push_space = btrfs_item_offset(right->items + i);
1164         }
1165         left_nritems -= push_items;
1166         btrfs_set_header_nritems(&left->header, left_nritems);
1167
1168         btrfs_mark_buffer_dirty(left_buf);
1169         btrfs_mark_buffer_dirty(right_buf);
1170
1171         btrfs_memcpy(root, upper_node, &upper_node->ptrs[slot + 1].key,
1172                 &right->items[0].key, sizeof(struct btrfs_disk_key));
1173         btrfs_mark_buffer_dirty(upper);
1174
1175         /* then fixup the leaf pointer in the path */
1176         if (path->slots[0] >= left_nritems) {
1177                 path->slots[0] -= left_nritems;
1178                 btrfs_block_release(root, path->nodes[0]);
1179                 path->nodes[0] = right_buf;
1180                 path->slots[1] += 1;
1181         } else {
1182                 btrfs_block_release(root, right_buf);
1183         }
1184         if (path->nodes[1])
1185                 check_node(root, path, 1);
1186         return 0;
1187 }
1188 /*
1189  * push some data in the path leaf to the left, trying to free up at
1190  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1191  */
1192 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1193                           *root, struct btrfs_path *path, int data_size)
1194 {
1195         struct buffer_head *right_buf = path->nodes[0];
1196         struct btrfs_leaf *right = btrfs_buffer_leaf(right_buf);
1197         struct buffer_head *t;
1198         struct btrfs_leaf *left;
1199         int slot;
1200         int i;
1201         int free_space;
1202         int push_space = 0;
1203         int push_items = 0;
1204         struct btrfs_item *item;
1205         u32 old_left_nritems;
1206         int ret = 0;
1207         int wret;
1208
1209         slot = path->slots[1];
1210         if (slot == 0) {
1211                 return 1;
1212         }
1213         if (!path->nodes[1]) {
1214                 return 1;
1215         }
1216         t = read_tree_block(root,
1217             btrfs_node_blockptr(btrfs_buffer_node(path->nodes[1]), slot - 1));
1218         left = btrfs_buffer_leaf(t);
1219         free_space = btrfs_leaf_free_space(root, left);
1220         if (free_space < data_size + sizeof(struct btrfs_item)) {
1221                 btrfs_block_release(root, t);
1222                 return 1;
1223         }
1224
1225         /* cow and double check */
1226         ret = btrfs_cow_block(trans, root, t, path->nodes[1], slot - 1, &t);
1227         if (ret) {
1228                 /* we hit -ENOSPC, but it isn't fatal here */
1229                 return 1;
1230         }
1231         left = btrfs_buffer_leaf(t);
1232         free_space = btrfs_leaf_free_space(root, left);
1233         if (free_space < data_size + sizeof(struct btrfs_item)) {
1234                 btrfs_block_release(root, t);
1235                 return 1;
1236         }
1237
1238         if (btrfs_header_nritems(&right->header) == 0) {
1239                 btrfs_block_release(root, t);
1240                 return 1;
1241         }
1242
1243         for (i = 0; i < btrfs_header_nritems(&right->header) - 1; i++) {
1244                 item = right->items + i;
1245                 if (path->slots[0] == i)
1246                         push_space += data_size + sizeof(*item);
1247                 if (btrfs_item_size(item) + sizeof(*item) + push_space >
1248                     free_space)
1249                         break;
1250                 push_items++;
1251                 push_space += btrfs_item_size(item) + sizeof(*item);
1252         }
1253         if (push_items == 0) {
1254                 btrfs_block_release(root, t);
1255                 return 1;
1256         }
1257         if (push_items == btrfs_header_nritems(&right->header))
1258                 WARN_ON(1);
1259         /* push data from right to left */
1260         btrfs_memcpy(root, left, left->items +
1261                      btrfs_header_nritems(&left->header),
1262                      right->items, push_items * sizeof(struct btrfs_item));
1263         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1264                      btrfs_item_offset(right->items + push_items -1);
1265         btrfs_memcpy(root, left, btrfs_leaf_data(left) +
1266                      leaf_data_end(root, left) - push_space,
1267                      btrfs_leaf_data(right) +
1268                      btrfs_item_offset(right->items + push_items - 1),
1269                      push_space);
1270         old_left_nritems = btrfs_header_nritems(&left->header);
1271         BUG_ON(old_left_nritems < 0);
1272
1273         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1274                 u32 ioff = btrfs_item_offset(left->items + i);
1275                 btrfs_set_item_offset(left->items + i, ioff -
1276                                      (BTRFS_LEAF_DATA_SIZE(root) -
1277                                       btrfs_item_offset(left->items +
1278                                                         old_left_nritems - 1)));
1279         }
1280         btrfs_set_header_nritems(&left->header, old_left_nritems + push_items);
1281
1282         /* fixup right node */
1283         push_space = btrfs_item_offset(right->items + push_items - 1) -
1284                      leaf_data_end(root, right);
1285         btrfs_memmove(root, right, btrfs_leaf_data(right) +
1286                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1287                       btrfs_leaf_data(right) +
1288                       leaf_data_end(root, right), push_space);
1289         btrfs_memmove(root, right, right->items, right->items + push_items,
1290                 (btrfs_header_nritems(&right->header) - push_items) *
1291                 sizeof(struct btrfs_item));
1292         btrfs_set_header_nritems(&right->header,
1293                                  btrfs_header_nritems(&right->header) -
1294                                  push_items);
1295         push_space = BTRFS_LEAF_DATA_SIZE(root);
1296
1297         for (i = 0; i < btrfs_header_nritems(&right->header); i++) {
1298                 btrfs_set_item_offset(right->items + i, push_space -
1299                                       btrfs_item_size(right->items + i));
1300                 push_space = btrfs_item_offset(right->items + i);
1301         }
1302
1303         btrfs_mark_buffer_dirty(t);
1304         btrfs_mark_buffer_dirty(right_buf);
1305
1306         wret = fixup_low_keys(trans, root, path, &right->items[0].key, 1);
1307         if (wret)
1308                 ret = wret;
1309
1310         /* then fixup the leaf pointer in the path */
1311         if (path->slots[0] < push_items) {
1312                 path->slots[0] += old_left_nritems;
1313                 btrfs_block_release(root, path->nodes[0]);
1314                 path->nodes[0] = t;
1315                 path->slots[1] -= 1;
1316         } else {
1317                 btrfs_block_release(root, t);
1318                 path->slots[0] -= push_items;
1319         }
1320         BUG_ON(path->slots[0] < 0);
1321         if (path->nodes[1])
1322                 check_node(root, path, 1);
1323         return ret;
1324 }
1325
1326 /*
1327  * split the path's leaf in two, making sure there is at least data_size
1328  * available for the resulting leaf level of the path.
1329  *
1330  * returns 0 if all went well and < 0 on failure.
1331  */
1332 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
1333                       *root, struct btrfs_key *ins_key,
1334                       struct btrfs_path *path, int data_size)
1335 {
1336         struct buffer_head *l_buf;
1337         struct btrfs_leaf *l;
1338         u32 nritems;
1339         int mid;
1340         int slot;
1341         struct btrfs_leaf *right;
1342         struct buffer_head *right_buffer;
1343         int space_needed = data_size + sizeof(struct btrfs_item);
1344         int data_copy_size;
1345         int rt_data_off;
1346         int i;
1347         int ret = 0;
1348         int wret;
1349         int double_split = 0;
1350         struct btrfs_disk_key disk_key;
1351
1352         /* first try to make some room by pushing left and right */
1353         wret = push_leaf_left(trans, root, path, data_size);
1354         if (wret < 0)
1355                 return wret;
1356         if (wret) {
1357                 wret = push_leaf_right(trans, root, path, data_size);
1358                 if (wret < 0)
1359                         return wret;
1360         }
1361         l_buf = path->nodes[0];
1362         l = btrfs_buffer_leaf(l_buf);
1363
1364         /* did the pushes work? */
1365         if (btrfs_leaf_free_space(root, l) >=
1366             sizeof(struct btrfs_item) + data_size)
1367                 return 0;
1368
1369         if (!path->nodes[1]) {
1370                 ret = insert_new_root(trans, root, path, 1);
1371                 if (ret)
1372                         return ret;
1373         }
1374         slot = path->slots[0];
1375         nritems = btrfs_header_nritems(&l->header);
1376         mid = (nritems + 1)/ 2;
1377
1378         right_buffer = btrfs_alloc_free_block(trans, root, l_buf->b_blocknr);
1379         if (IS_ERR(right_buffer))
1380                 return PTR_ERR(right_buffer);
1381
1382         right = btrfs_buffer_leaf(right_buffer);
1383         memset(&right->header, 0, sizeof(right->header));
1384         btrfs_set_header_blocknr(&right->header, bh_blocknr(right_buffer));
1385         btrfs_set_header_generation(&right->header, trans->transid);
1386         btrfs_set_header_owner(&right->header, root->root_key.objectid);
1387         btrfs_set_header_level(&right->header, 0);
1388         memcpy(right->header.fsid, root->fs_info->disk_super->fsid,
1389                sizeof(right->header.fsid));
1390         if (mid <= slot) {
1391                 if (nritems == 1 ||
1392                     leaf_space_used(l, mid, nritems - mid) + space_needed >
1393                         BTRFS_LEAF_DATA_SIZE(root)) {
1394                         if (slot >= nritems) {
1395                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1396                                 btrfs_set_header_nritems(&right->header, 0);
1397                                 wret = insert_ptr(trans, root, path,
1398                                                   &disk_key,
1399                                                   bh_blocknr(right_buffer),
1400                                                   path->slots[1] + 1, 1);
1401                                 if (wret)
1402                                         ret = wret;
1403                                 btrfs_block_release(root, path->nodes[0]);
1404                                 path->nodes[0] = right_buffer;
1405                                 path->slots[0] = 0;
1406                                 path->slots[1] += 1;
1407                                 return ret;
1408                         }
1409                         mid = slot;
1410                         double_split = 1;
1411                 }
1412         } else {
1413                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
1414                         BTRFS_LEAF_DATA_SIZE(root)) {
1415                         if (slot == 0) {
1416                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1417                                 btrfs_set_header_nritems(&right->header, 0);
1418                                 wret = insert_ptr(trans, root, path,
1419                                                   &disk_key,
1420                                                   bh_blocknr(right_buffer),
1421                                                   path->slots[1], 1);
1422                                 if (wret)
1423                                         ret = wret;
1424                                 btrfs_block_release(root, path->nodes[0]);
1425                                 path->nodes[0] = right_buffer;
1426                                 path->slots[0] = 0;
1427                                 if (path->slots[1] == 0) {
1428                                         wret = fixup_low_keys(trans, root,
1429                                                    path, &disk_key, 1);
1430                                         if (wret)
1431                                                 ret = wret;
1432                                 }
1433                                 return ret;
1434                         }
1435                         mid = slot;
1436                         double_split = 1;
1437                 }
1438         }
1439         btrfs_set_header_nritems(&right->header, nritems - mid);
1440         data_copy_size = btrfs_item_end(l->items + mid) -
1441                          leaf_data_end(root, l);
1442         btrfs_memcpy(root, right, right->items, l->items + mid,
1443                      (nritems - mid) * sizeof(struct btrfs_item));
1444         btrfs_memcpy(root, right,
1445                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
1446                      data_copy_size, btrfs_leaf_data(l) +
1447                      leaf_data_end(root, l), data_copy_size);
1448         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
1449                       btrfs_item_end(l->items + mid);
1450
1451         for (i = 0; i < btrfs_header_nritems(&right->header); i++) {
1452                 u32 ioff = btrfs_item_offset(right->items + i);
1453                 btrfs_set_item_offset(right->items + i, ioff + rt_data_off);
1454         }
1455
1456         btrfs_set_header_nritems(&l->header, mid);
1457         ret = 0;
1458         wret = insert_ptr(trans, root, path, &right->items[0].key,
1459                           bh_blocknr(right_buffer), path->slots[1] + 1, 1);
1460         if (wret)
1461                 ret = wret;
1462         btrfs_mark_buffer_dirty(right_buffer);
1463         btrfs_mark_buffer_dirty(l_buf);
1464         BUG_ON(path->slots[0] != slot);
1465         if (mid <= slot) {
1466                 btrfs_block_release(root, path->nodes[0]);
1467                 path->nodes[0] = right_buffer;
1468                 path->slots[0] -= mid;
1469                 path->slots[1] += 1;
1470         } else
1471                 btrfs_block_release(root, right_buffer);
1472         BUG_ON(path->slots[0] < 0);
1473         check_node(root, path, 1);
1474
1475         if (!double_split)
1476                 return ret;
1477         right_buffer = btrfs_alloc_free_block(trans, root, l_buf->b_blocknr);
1478         if (IS_ERR(right_buffer))
1479                 return PTR_ERR(right_buffer);
1480
1481         right = btrfs_buffer_leaf(right_buffer);
1482         memset(&right->header, 0, sizeof(right->header));
1483         btrfs_set_header_blocknr(&right->header, bh_blocknr(right_buffer));
1484         btrfs_set_header_generation(&right->header, trans->transid);
1485         btrfs_set_header_owner(&right->header, root->root_key.objectid);
1486         btrfs_set_header_level(&right->header, 0);
1487         memcpy(right->header.fsid, root->fs_info->disk_super->fsid,
1488                sizeof(right->header.fsid));
1489         btrfs_cpu_key_to_disk(&disk_key, ins_key);
1490         btrfs_set_header_nritems(&right->header, 0);
1491         wret = insert_ptr(trans, root, path,
1492                           &disk_key,
1493                           bh_blocknr(right_buffer),
1494                           path->slots[1], 1);
1495         if (wret)
1496                 ret = wret;
1497         if (path->slots[1] == 0) {
1498                 wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1499                 if (wret)
1500                         ret = wret;
1501         }
1502         btrfs_block_release(root, path->nodes[0]);
1503         path->nodes[0] = right_buffer;
1504         path->slots[0] = 0;
1505         check_node(root, path, 1);
1506         check_leaf(root, path, 0);
1507         return ret;
1508 }
1509
1510 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
1511                         struct btrfs_root *root,
1512                         struct btrfs_path *path,
1513                         u32 new_size)
1514 {
1515         int ret = 0;
1516         int slot;
1517         int slot_orig;
1518         struct btrfs_leaf *leaf;
1519         struct buffer_head *leaf_buf;
1520         u32 nritems;
1521         unsigned int data_end;
1522         unsigned int old_data_start;
1523         unsigned int old_size;
1524         unsigned int size_diff;
1525         int i;
1526
1527         slot_orig = path->slots[0];
1528         leaf_buf = path->nodes[0];
1529         leaf = btrfs_buffer_leaf(leaf_buf);
1530
1531         nritems = btrfs_header_nritems(&leaf->header);
1532         data_end = leaf_data_end(root, leaf);
1533
1534         slot = path->slots[0];
1535         old_data_start = btrfs_item_offset(leaf->items + slot);
1536         old_size = btrfs_item_size(leaf->items + slot);
1537         BUG_ON(old_size <= new_size);
1538         size_diff = old_size - new_size;
1539
1540         BUG_ON(slot < 0);
1541         BUG_ON(slot >= nritems);
1542
1543         /*
1544          * item0..itemN ... dataN.offset..dataN.size .. data0.size
1545          */
1546         /* first correct the data pointers */
1547         for (i = slot; i < nritems; i++) {
1548                 u32 ioff = btrfs_item_offset(leaf->items + i);
1549                 btrfs_set_item_offset(leaf->items + i,
1550                                       ioff + size_diff);
1551         }
1552         /* shift the data */
1553         btrfs_memmove(root, leaf, btrfs_leaf_data(leaf) +
1554                       data_end + size_diff, btrfs_leaf_data(leaf) +
1555                       data_end, old_data_start + new_size - data_end);
1556         btrfs_set_item_size(leaf->items + slot, new_size);
1557         btrfs_mark_buffer_dirty(leaf_buf);
1558
1559         ret = 0;
1560         if (btrfs_leaf_free_space(root, leaf) < 0)
1561                 BUG();
1562         check_leaf(root, path, 0);
1563         return ret;
1564 }
1565
1566 int btrfs_extend_item(struct btrfs_trans_handle *trans, struct btrfs_root
1567                       *root, struct btrfs_path *path, u32 data_size)
1568 {
1569         int ret = 0;
1570         int slot;
1571         int slot_orig;
1572         struct btrfs_leaf *leaf;
1573         struct buffer_head *leaf_buf;
1574         u32 nritems;
1575         unsigned int data_end;
1576         unsigned int old_data;
1577         unsigned int old_size;
1578         int i;
1579
1580         slot_orig = path->slots[0];
1581         leaf_buf = path->nodes[0];
1582         leaf = btrfs_buffer_leaf(leaf_buf);
1583
1584         nritems = btrfs_header_nritems(&leaf->header);
1585         data_end = leaf_data_end(root, leaf);
1586
1587         if (btrfs_leaf_free_space(root, leaf) < data_size)
1588                 BUG();
1589         slot = path->slots[0];
1590         old_data = btrfs_item_end(leaf->items + slot);
1591
1592         BUG_ON(slot < 0);
1593         BUG_ON(slot >= nritems);
1594
1595         /*
1596          * item0..itemN ... dataN.offset..dataN.size .. data0.size
1597          */
1598         /* first correct the data pointers */
1599         for (i = slot; i < nritems; i++) {
1600                 u32 ioff = btrfs_item_offset(leaf->items + i);
1601                 btrfs_set_item_offset(leaf->items + i,
1602                                       ioff - data_size);
1603         }
1604         /* shift the data */
1605         btrfs_memmove(root, leaf, btrfs_leaf_data(leaf) +
1606                       data_end - data_size, btrfs_leaf_data(leaf) +
1607                       data_end, old_data - data_end);
1608         data_end = old_data;
1609         old_size = btrfs_item_size(leaf->items + slot);
1610         btrfs_set_item_size(leaf->items + slot, old_size + data_size);
1611         btrfs_mark_buffer_dirty(leaf_buf);
1612
1613         ret = 0;
1614         if (btrfs_leaf_free_space(root, leaf) < 0)
1615                 BUG();
1616         check_leaf(root, path, 0);
1617         return ret;
1618 }
1619
1620 /*
1621  * Given a key and some data, insert an item into the tree.
1622  * This does all the path init required, making room in the tree if needed.
1623  */
1624 int btrfs_insert_empty_item(struct btrfs_trans_handle *trans, struct btrfs_root
1625                             *root, struct btrfs_path *path, struct btrfs_key
1626                             *cpu_key, u32 data_size)
1627 {
1628         int ret = 0;
1629         int slot;
1630         int slot_orig;
1631         struct btrfs_leaf *leaf;
1632         struct buffer_head *leaf_buf;
1633         u32 nritems;
1634         unsigned int data_end;
1635         struct btrfs_disk_key disk_key;
1636
1637         btrfs_cpu_key_to_disk(&disk_key, cpu_key);
1638
1639         /* create a root if there isn't one */
1640         if (!root->node)
1641                 BUG();
1642         ret = btrfs_search_slot(trans, root, cpu_key, path, data_size, 1);
1643         if (ret == 0) {
1644                 return -EEXIST;
1645         }
1646         if (ret < 0)
1647                 goto out;
1648
1649         slot_orig = path->slots[0];
1650         leaf_buf = path->nodes[0];
1651         leaf = btrfs_buffer_leaf(leaf_buf);
1652
1653         nritems = btrfs_header_nritems(&leaf->header);
1654         data_end = leaf_data_end(root, leaf);
1655
1656         if (btrfs_leaf_free_space(root, leaf) <
1657             sizeof(struct btrfs_item) + data_size) {
1658                 BUG();
1659         }
1660         slot = path->slots[0];
1661         BUG_ON(slot < 0);
1662         if (slot != nritems) {
1663                 int i;
1664                 unsigned int old_data = btrfs_item_end(leaf->items + slot);
1665
1666                 /*
1667                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
1668                  */
1669                 /* first correct the data pointers */
1670                 for (i = slot; i < nritems; i++) {
1671                         u32 ioff = btrfs_item_offset(leaf->items + i);
1672                         btrfs_set_item_offset(leaf->items + i,
1673                                               ioff - data_size);
1674                 }
1675
1676                 /* shift the items */
1677                 btrfs_memmove(root, leaf, leaf->items + slot + 1,
1678                               leaf->items + slot,
1679                               (nritems - slot) * sizeof(struct btrfs_item));
1680
1681                 /* shift the data */
1682                 btrfs_memmove(root, leaf, btrfs_leaf_data(leaf) +
1683                               data_end - data_size, btrfs_leaf_data(leaf) +
1684                               data_end, old_data - data_end);
1685                 data_end = old_data;
1686         }
1687         /* setup the item for the new data */
1688         btrfs_memcpy(root, leaf, &leaf->items[slot].key, &disk_key,
1689                      sizeof(struct btrfs_disk_key));
1690         btrfs_set_item_offset(leaf->items + slot, data_end - data_size);
1691         btrfs_set_item_size(leaf->items + slot, data_size);
1692         btrfs_set_header_nritems(&leaf->header, nritems + 1);
1693         btrfs_mark_buffer_dirty(leaf_buf);
1694
1695         ret = 0;
1696         if (slot == 0)
1697                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
1698
1699         if (btrfs_leaf_free_space(root, leaf) < 0)
1700                 BUG();
1701         check_leaf(root, path, 0);
1702 out:
1703         return ret;
1704 }
1705
1706 /*
1707  * Given a key and some data, insert an item into the tree.
1708  * This does all the path init required, making room in the tree if needed.
1709  */
1710 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
1711                       *root, struct btrfs_key *cpu_key, void *data, u32
1712                       data_size)
1713 {
1714         int ret = 0;
1715         struct btrfs_path *path;
1716         u8 *ptr;
1717
1718         path = btrfs_alloc_path();
1719         BUG_ON(!path);
1720         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
1721         if (!ret) {
1722                 ptr = btrfs_item_ptr(btrfs_buffer_leaf(path->nodes[0]),
1723                                      path->slots[0], u8);
1724                 btrfs_memcpy(root, path->nodes[0]->b_data,
1725                              ptr, data, data_size);
1726                 btrfs_mark_buffer_dirty(path->nodes[0]);
1727         }
1728         btrfs_free_path(path);
1729         return ret;
1730 }
1731
1732 /*
1733  * delete the pointer from a given node.
1734  *
1735  * If the delete empties a node, the node is removed from the tree,
1736  * continuing all the way the root if required.  The root is converted into
1737  * a leaf if all the nodes are emptied.
1738  */
1739 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1740                    struct btrfs_path *path, int level, int slot)
1741 {
1742         struct btrfs_node *node;
1743         struct buffer_head *parent = path->nodes[level];
1744         u32 nritems;
1745         int ret = 0;
1746         int wret;
1747
1748         node = btrfs_buffer_node(parent);
1749         nritems = btrfs_header_nritems(&node->header);
1750         if (slot != nritems -1) {
1751                 btrfs_memmove(root, node, node->ptrs + slot,
1752                               node->ptrs + slot + 1,
1753                               sizeof(struct btrfs_key_ptr) *
1754                               (nritems - slot - 1));
1755         }
1756         nritems--;
1757         btrfs_set_header_nritems(&node->header, nritems);
1758         if (nritems == 0 && parent == root->node) {
1759                 struct btrfs_header *header = btrfs_buffer_header(root->node);
1760                 BUG_ON(btrfs_header_level(header) != 1);
1761                 /* just turn the root into a leaf and break */
1762                 btrfs_set_header_level(header, 0);
1763         } else if (slot == 0) {
1764                 wret = fixup_low_keys(trans, root, path, &node->ptrs[0].key,
1765                                       level + 1);
1766                 if (wret)
1767                         ret = wret;
1768         }
1769         btrfs_mark_buffer_dirty(parent);
1770         return ret;
1771 }
1772
1773 /*
1774  * delete the item at the leaf level in path.  If that empties
1775  * the leaf, remove it from the tree
1776  */
1777 int btrfs_del_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1778                    struct btrfs_path *path)
1779 {
1780         int slot;
1781         struct btrfs_leaf *leaf;
1782         struct buffer_head *leaf_buf;
1783         int doff;
1784         int dsize;
1785         int ret = 0;
1786         int wret;
1787         u32 nritems;
1788
1789         leaf_buf = path->nodes[0];
1790         leaf = btrfs_buffer_leaf(leaf_buf);
1791         slot = path->slots[0];
1792         doff = btrfs_item_offset(leaf->items + slot);
1793         dsize = btrfs_item_size(leaf->items + slot);
1794         nritems = btrfs_header_nritems(&leaf->header);
1795
1796         if (slot != nritems - 1) {
1797                 int i;
1798                 int data_end = leaf_data_end(root, leaf);
1799                 btrfs_memmove(root, leaf, btrfs_leaf_data(leaf) +
1800                               data_end + dsize,
1801                               btrfs_leaf_data(leaf) + data_end,
1802                               doff - data_end);
1803                 for (i = slot + 1; i < nritems; i++) {
1804                         u32 ioff = btrfs_item_offset(leaf->items + i);
1805                         btrfs_set_item_offset(leaf->items + i, ioff + dsize);
1806                 }
1807                 btrfs_memmove(root, leaf, leaf->items + slot,
1808                               leaf->items + slot + 1,
1809                               sizeof(struct btrfs_item) *
1810                               (nritems - slot - 1));
1811         }
1812         btrfs_set_header_nritems(&leaf->header, nritems - 1);
1813         nritems--;
1814         /* delete the leaf if we've emptied it */
1815         if (nritems == 0) {
1816                 if (leaf_buf == root->node) {
1817                         btrfs_set_header_level(&leaf->header, 0);
1818                 } else {
1819                         clean_tree_block(trans, root, leaf_buf);
1820                         wait_on_buffer(leaf_buf);
1821                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
1822                         if (wret)
1823                                 ret = wret;
1824                         wret = btrfs_free_extent(trans, root,
1825                                                  bh_blocknr(leaf_buf), 1, 1);
1826                         if (wret)
1827                                 ret = wret;
1828                 }
1829         } else {
1830                 int used = leaf_space_used(leaf, 0, nritems);
1831                 if (slot == 0) {
1832                         wret = fixup_low_keys(trans, root, path,
1833                                               &leaf->items[0].key, 1);
1834                         if (wret)
1835                                 ret = wret;
1836                 }
1837
1838                 /* delete the leaf if it is mostly empty */
1839                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 3) {
1840                         /* push_leaf_left fixes the path.
1841                          * make sure the path still points to our leaf
1842                          * for possible call to del_ptr below
1843                          */
1844                         slot = path->slots[1];
1845                         get_bh(leaf_buf);
1846                         wret = push_leaf_left(trans, root, path, 1);
1847                         if (wret < 0 && wret != -ENOSPC)
1848                                 ret = wret;
1849                         if (path->nodes[0] == leaf_buf &&
1850                             btrfs_header_nritems(&leaf->header)) {
1851                                 wret = push_leaf_right(trans, root, path, 1);
1852                                 if (wret < 0 && wret != -ENOSPC)
1853                                         ret = wret;
1854                         }
1855                         if (btrfs_header_nritems(&leaf->header) == 0) {
1856                                 u64 blocknr = bh_blocknr(leaf_buf);
1857                                 clean_tree_block(trans, root, leaf_buf);
1858                                 wait_on_buffer(leaf_buf);
1859                                 wret = del_ptr(trans, root, path, 1, slot);
1860                                 if (wret)
1861                                         ret = wret;
1862                                 btrfs_block_release(root, leaf_buf);
1863                                 wret = btrfs_free_extent(trans, root, blocknr,
1864                                                          1, 1);
1865                                 if (wret)
1866                                         ret = wret;
1867                         } else {
1868                                 btrfs_mark_buffer_dirty(leaf_buf);
1869                                 btrfs_block_release(root, leaf_buf);
1870                         }
1871                 } else {
1872                         btrfs_mark_buffer_dirty(leaf_buf);
1873                 }
1874         }
1875         return ret;
1876 }
1877
1878 /*
1879  * walk up the tree as far as required to find the next leaf.
1880  * returns 0 if it found something or 1 if there are no greater leaves.
1881  * returns < 0 on io errors.
1882  */
1883 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
1884 {
1885         int slot;
1886         int level = 1;
1887         u64 blocknr;
1888         struct buffer_head *c;
1889         struct btrfs_node *c_node;
1890         struct buffer_head *next = NULL;
1891
1892         while(level < BTRFS_MAX_LEVEL) {
1893                 if (!path->nodes[level])
1894                         return 1;
1895                 slot = path->slots[level] + 1;
1896                 c = path->nodes[level];
1897                 c_node = btrfs_buffer_node(c);
1898                 if (slot >= btrfs_header_nritems(&c_node->header)) {
1899                         level++;
1900                         continue;
1901                 }
1902                 blocknr = btrfs_node_blockptr(c_node, slot);
1903                 if (next)
1904                         btrfs_block_release(root, next);
1905                 next = read_tree_block(root, blocknr);
1906                 break;
1907         }
1908         path->slots[level] = slot;
1909         while(1) {
1910                 level--;
1911                 c = path->nodes[level];
1912                 btrfs_block_release(root, c);
1913                 path->nodes[level] = next;
1914                 path->slots[level] = 0;
1915                 if (!level)
1916                         break;
1917                 next = read_tree_block(root,
1918                        btrfs_node_blockptr(btrfs_buffer_node(next), 0));
1919         }
1920         return 0;
1921 }