fs/btrfs/ctree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include "ctree.h"
  20 #include "disk-io.h"
  21 #include "transaction.h"
  22
  23 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  24                       *root, struct btrfs_path *path, int level);
  25 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  26                       *root, struct btrfs_key *ins_key,
  27                       struct btrfs_path *path, int data_size);
  28 static int push_node_left(struct btrfs_trans_handle *trans, struct btrfs_root
  29                           *root, struct buffer_head *dst, struct buffer_head
  30                           *src);
  31 static int balance_node_right(struct btrfs_trans_handle *trans, struct
  32                               btrfs_root *root, struct buffer_head *dst_buf,
  33                               struct buffer_head *src_buf);
  34 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  35                    struct btrfs_path *path, int level, int slot);
  36
  37 inline void btrfs_init_path(struct btrfs_path *p)
  38 {
  39         memset(p, 0, sizeof(*p));
  40 }
  41
  42 struct btrfs_path *btrfs_alloc_path(void)
  43 {
  44         struct btrfs_path *path;
  45         path = kmem_cache_alloc(btrfs_path_cachep, GFP_NOFS);
  46         if (path)
  47                 btrfs_init_path(path);
  48         return path;
  49 }
  50
  51 void btrfs_free_path(struct btrfs_path *p)
  52 {
  53         btrfs_release_path(NULL, p);
  54         kmem_cache_free(btrfs_path_cachep, p);
  55 }
  56
  57 void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
  58 {
  59         int i;
  60         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  61                 if (!p->nodes[i])
  62                         break;
  63                 btrfs_block_release(root, p->nodes[i]);
  64         }
  65         memset(p, 0, sizeof(*p));
  66 }
  67
  68 static int btrfs_cow_block(struct btrfs_trans_handle *trans, struct btrfs_root
  69                            *root, struct buffer_head *buf, struct buffer_head
  70                            *parent, int parent_slot, struct buffer_head
  71                            **cow_ret)
  72 {
  73         struct buffer_head *cow;
  74         struct btrfs_node *cow_node;
  75         int ret;
  76
  77         WARN_ON(!buffer_uptodate(buf));
  78         if (trans->transaction != root->fs_info->running_transaction) {
  79                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
  80                        root->fs_info->running_transaction->transid);
  81                 WARN_ON(1);
  82         }
  83         if (trans->transid != root->fs_info->generation) {
  84                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
  85                        root->fs_info->generation);
  86                 WARN_ON(1);
  87         }
  88         if (btrfs_header_generation(btrfs_buffer_header(buf)) ==
  89                                     trans->transid) {
  90                 *cow_ret = buf;
  91                 return 0;
  92         }
  93         cow = btrfs_alloc_free_block(trans, root, buf->b_blocknr);
  94         if (IS_ERR(cow))
  95                 return PTR_ERR(cow);
  96         cow_node = btrfs_buffer_node(cow);
  97         if (buf->b_size != root->blocksize || cow->b_size != root->blocksize)
  98                 WARN_ON(1);
  99         memcpy(cow_node, btrfs_buffer_node(buf), root->blocksize);
 100         btrfs_set_header_blocknr(&cow_node->header, bh_blocknr(cow));
 101         btrfs_set_header_generation(&cow_node->header, trans->transid);
 102         btrfs_set_header_owner(&cow_node->header, root->root_key.objectid);
 103         ret = btrfs_inc_ref(trans, root, buf);
 104         if (ret)
 105                 return ret;
 106         if (buf == root->node) {
 107                 root->node = cow;
 108                 get_bh(cow);
 109                 if (buf != root->commit_root) {
 110                         btrfs_free_extent(trans, root, bh_blocknr(buf), 1, 1);
 111                 }
 112                 btrfs_block_release(root, buf);
 113         } else {
 114                 btrfs_set_node_blockptr(btrfs_buffer_node(parent), parent_slot,
 115                                         bh_blocknr(cow));
 116                 btrfs_mark_buffer_dirty(parent);
 117                 btrfs_free_extent(trans, root, bh_blocknr(buf), 1, 1);
 118         }
 119         btrfs_block_release(root, buf);
 120         btrfs_mark_buffer_dirty(cow);
 121         *cow_ret = cow;
 122         return 0;
 123 }
 124
 125 /*
 126  * The leaf data grows from end-to-front in the node.
 127  * this returns the address of the start of the last item,
 128  * which is the stop of the leaf data stack
 129  */
 130 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 131                                          struct btrfs_leaf *leaf)
 132 {
 133         u32 nr = btrfs_header_nritems(&leaf->header);
 134         if (nr == 0)
 135                 return BTRFS_LEAF_DATA_SIZE(root);
 136         return btrfs_item_offset(leaf->items + nr - 1);
 137 }
 138
 139 /*
 140  * compare two keys in a memcmp fashion
 141  */
 142 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 143 {
 144         struct btrfs_key k1;
 145
 146         btrfs_disk_key_to_cpu(&k1, disk);
 147
 148         if (k1.objectid > k2->objectid)
 149                 return 1;
 150         if (k1.objectid < k2->objectid)
 151                 return -1;
 152         if (k1.flags > k2->flags)
 153                 return 1;
 154         if (k1.flags < k2->flags)
 155                 return -1;
 156         if (k1.offset > k2->offset)
 157                 return 1;
 158         if (k1.offset < k2->offset)
 159                 return -1;
 160         return 0;
 161 }
 162
 163 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
 164                       int level)
 165 {
 166         struct btrfs_node *parent = NULL;
 167         struct btrfs_node *node = btrfs_buffer_node(path->nodes[level]);
 168         int parent_slot;
 169         int slot;
 170         struct btrfs_key cpukey;
 171         u32 nritems = btrfs_header_nritems(&node->header);
 172
 173         if (path->nodes[level + 1])
 174                 parent = btrfs_buffer_node(path->nodes[level + 1]);
 175         parent_slot = path->slots[level + 1];
 176         slot = path->slots[level];
 177         BUG_ON(nritems == 0);
 178         if (parent) {
 179                 struct btrfs_disk_key *parent_key;
 180                 parent_key = &parent->ptrs[parent_slot].key;
 181                 BUG_ON(memcmp(parent_key, &node->ptrs[0].key,
 182                               sizeof(struct btrfs_disk_key)));
 183                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 184                        btrfs_header_blocknr(&node->header));
 185         }
 186         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 187         if (slot != 0) {
 188                 btrfs_disk_key_to_cpu(&cpukey, &node->ptrs[slot - 1].key);
 189                 BUG_ON(comp_keys(&node->ptrs[slot].key, &cpukey) <= 0);
 190         }
 191         if (slot < nritems - 1) {
 192                 btrfs_disk_key_to_cpu(&cpukey, &node->ptrs[slot + 1].key);
 193                 BUG_ON(comp_keys(&node->ptrs[slot].key, &cpukey) >= 0);
 194         }
 195         return 0;
 196 }
 197
 198 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 199                       int level)
 200 {
 201         struct btrfs_leaf *leaf = btrfs_buffer_leaf(path->nodes[level]);
 202         struct btrfs_node *parent = NULL;
 203         int parent_slot;
 204         int slot = path->slots[0];
 205         struct btrfs_key cpukey;
 206
 207         u32 nritems = btrfs_header_nritems(&leaf->header);
 208
 209         if (path->nodes[level + 1])
 210                 parent = btrfs_buffer_node(path->nodes[level + 1]);
 211         parent_slot = path->slots[level + 1];
 212         BUG_ON(btrfs_leaf_free_space(root, leaf) < 0);
 213
 214         if (nritems == 0)
 215                 return 0;
 216
 217         if (parent) {
 218                 struct btrfs_disk_key *parent_key;
 219                 parent_key = &parent->ptrs[parent_slot].key;
 220                 BUG_ON(memcmp(parent_key, &leaf->items[0].key,
 221                        sizeof(struct btrfs_disk_key)));
 222                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 223                        btrfs_header_blocknr(&leaf->header));
 224         }
 225         if (slot != 0) {
 226                 btrfs_disk_key_to_cpu(&cpukey, &leaf->items[slot - 1].key);
 227                 BUG_ON(comp_keys(&leaf->items[slot].key, &cpukey) <= 0);
 228                 BUG_ON(btrfs_item_offset(leaf->items + slot - 1) !=
 229                         btrfs_item_end(leaf->items + slot));
 230         }
 231         if (slot < nritems - 1) {
 232                 btrfs_disk_key_to_cpu(&cpukey, &leaf->items[slot + 1].key);
 233                 BUG_ON(comp_keys(&leaf->items[slot].key, &cpukey) >= 0);
 234                 BUG_ON(btrfs_item_offset(leaf->items + slot) !=
 235                         btrfs_item_end(leaf->items + slot + 1));
 236         }
 237         BUG_ON(btrfs_item_offset(leaf->items) +
 238                btrfs_item_size(leaf->items) != BTRFS_LEAF_DATA_SIZE(root));
 239         return 0;
 240 }
 241
 242 static int check_block(struct btrfs_root *root, struct btrfs_path *path,
 243                         int level)
 244 {
 245         struct btrfs_node *node = btrfs_buffer_node(path->nodes[level]);
 246         if (memcmp(node->header.fsid, root->fs_info->disk_super->fsid,
 247                    sizeof(node->header.fsid)))
 248                 BUG();
 249         if (level == 0)
 250                 return check_leaf(root, path, level);
 251         return check_node(root, path, level);
 252 }
 253
 254 /*
 255  * search for key in the array p.  items p are item_size apart
 256  * and there are 'max' items in p
 257  * the slot in the array is returned via slot, and it points to
 258  * the place where you would insert key if it is not found in
 259  * the array.
 260  *
 261  * slot may point to max if the key is bigger than all of the keys
 262  */
 263 static int generic_bin_search(char *p, int item_size, struct btrfs_key *key,
 264                        int max, int *slot)
 265 {
 266         int low = 0;
 267         int high = max;
 268         int mid;
 269         int ret;
 270         struct btrfs_disk_key *tmp;
 271
 272         while(low < high) {
 273                 mid = (low + high) / 2;
 274                 tmp = (struct btrfs_disk_key *)(p + mid * item_size);
 275                 ret = comp_keys(tmp, key);
 276
 277                 if (ret < 0)
 278                         low = mid + 1;
 279                 else if (ret > 0)
 280                         high = mid;
 281                 else {
 282                         *slot = mid;
 283                         return 0;
 284                 }
 285         }
 286         *slot = low;
 287         return 1;
 288 }
 289
 290 /*
 291  * simple bin_search frontend that does the right thing for
 292  * leaves vs nodes
 293  */
 294 static int bin_search(struct btrfs_node *c, struct btrfs_key *key, int *slot)
 295 {
 296         if (btrfs_is_leaf(c)) {
 297                 struct btrfs_leaf *l = (struct btrfs_leaf *)c;
 298                 return generic_bin_search((void *)l->items,
 299                                           sizeof(struct btrfs_item),
 300                                           key, btrfs_header_nritems(&c->header),
 301                                           slot);
 302         } else {
 303                 return generic_bin_search((void *)c->ptrs,
 304                                           sizeof(struct btrfs_key_ptr),
 305                                           key, btrfs_header_nritems(&c->header),
 306                                           slot);
 307         }
 308         return -1;
 309 }
 310
 311 static struct buffer_head *read_node_slot(struct btrfs_root *root,
 312                                    struct buffer_head *parent_buf,
 313                                    int slot)
 314 {
 315         struct btrfs_node *node = btrfs_buffer_node(parent_buf);
 316         if (slot < 0)
 317                 return NULL;
 318         if (slot >= btrfs_header_nritems(&node->header))
 319                 return NULL;
 320         return read_tree_block(root, btrfs_node_blockptr(node, slot));
 321 }
 322
 323 static int balance_level(struct btrfs_trans_handle *trans, struct btrfs_root
 324                          *root, struct btrfs_path *path, int level)
 325 {
 326         struct buffer_head *right_buf;
 327         struct buffer_head *mid_buf;
 328         struct buffer_head *left_buf;
 329         struct buffer_head *parent_buf = NULL;
 330         struct btrfs_node *right = NULL;
 331         struct btrfs_node *mid;
 332         struct btrfs_node *left = NULL;
 333         struct btrfs_node *parent = NULL;
 334         int ret = 0;
 335         int wret;
 336         int pslot;
 337         int orig_slot = path->slots[level];
 338         int err_on_enospc = 0;
 339         u64 orig_ptr;
 340
 341         if (level == 0)
 342                 return 0;
 343
 344         mid_buf = path->nodes[level];
 345         mid = btrfs_buffer_node(mid_buf);
 346         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 347
 348         if (level < BTRFS_MAX_LEVEL - 1)
 349                 parent_buf = path->nodes[level + 1];
 350         pslot = path->slots[level + 1];
 351
 352         /*
 353          * deal with the case where there is only one pointer in the root
 354          * by promoting the node below to a root
 355          */
 356         if (!parent_buf) {
 357                 struct buffer_head *child;
 358                 u64 blocknr = bh_blocknr(mid_buf);
 359
 360                 if (btrfs_header_nritems(&mid->header) != 1)
 361                         return 0;
 362
 363                 /* promote the child to a root */
 364                 child = read_node_slot(root, mid_buf, 0);
 365                 BUG_ON(!child);
 366                 root->node = child;
 367                 path->nodes[level] = NULL;
 368                 clean_tree_block(trans, root, mid_buf);
 369                 wait_on_buffer(mid_buf);
 370                 /* once for the path */
 371                 btrfs_block_release(root, mid_buf);
 372                 /* once for the root ptr */
 373                 btrfs_block_release(root, mid_buf);
 374                 return btrfs_free_extent(trans, root, blocknr, 1, 1);
 375         }
 376         parent = btrfs_buffer_node(parent_buf);
 377
 378         if (btrfs_header_nritems(&mid->header) >
 379             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 380                 return 0;
 381
 382         if (btrfs_header_nritems(&mid->header) < 2)
 383                 err_on_enospc = 1;
 384
 385         left_buf = read_node_slot(root, parent_buf, pslot - 1);
 386         right_buf = read_node_slot(root, parent_buf, pslot + 1);
 387
 388         /* first, try to make some room in the middle buffer */
 389         if (left_buf) {
 390                 wret = btrfs_cow_block(trans, root, left_buf,
 391                                        parent_buf, pslot - 1, &left_buf);
 392                 if (wret) {
 393                         ret = wret;
 394                         goto enospc;
 395                 }
 396                 left = btrfs_buffer_node(left_buf);
 397                 orig_slot += btrfs_header_nritems(&left->header);
 398                 wret = push_node_left(trans, root, left_buf, mid_buf);
 399                 if (wret < 0)
 400                         ret = wret;
 401                 if (btrfs_header_nritems(&mid->header) < 2)
 402                         err_on_enospc = 1;
 403         }
 404
 405         /*
 406          * then try to empty the right most buffer into the middle
 407          */
 408         if (right_buf) {
 409                 wret = btrfs_cow_block(trans, root, right_buf,
 410                                        parent_buf, pslot + 1, &right_buf);
 411                 if (wret) {
 412                         ret = wret;
 413                         goto enospc;
 414                 }
 415
 416                 right = btrfs_buffer_node(right_buf);
 417                 wret = push_node_left(trans, root, mid_buf, right_buf);
 418                 if (wret < 0 && wret != -ENOSPC)
 419                         ret = wret;
 420                 if (btrfs_header_nritems(&right->header) == 0) {
 421                         u64 blocknr = bh_blocknr(right_buf);
 422                         clean_tree_block(trans, root, right_buf);
 423                         wait_on_buffer(right_buf);
 424                         btrfs_block_release(root, right_buf);
 425                         right_buf = NULL;
 426                         right = NULL;
 427                         wret = del_ptr(trans, root, path, level + 1, pslot +
 428                                        1);
 429                         if (wret)
 430                                 ret = wret;
 431                         wret = btrfs_free_extent(trans, root, blocknr, 1, 1);
 432                         if (wret)
 433                                 ret = wret;
 434                 } else {
 435                         btrfs_memcpy(root, parent,
 436                                      &parent->ptrs[pslot + 1].key,
 437                                      &right->ptrs[0].key,
 438                                      sizeof(struct btrfs_disk_key));
 439                         btrfs_mark_buffer_dirty(parent_buf);
 440                 }
 441         }
 442         if (btrfs_header_nritems(&mid->header) == 1) {
 443                 /*
 444                  * we're not allowed to leave a node with one item in the
 445                  * tree during a delete.  A deletion from lower in the tree
 446                  * could try to delete the only pointer in this node.
 447                  * So, pull some keys from the left.
 448                  * There has to be a left pointer at this point because
 449                  * otherwise we would have pulled some pointers from the
 450                  * right
 451                  */
 452                 BUG_ON(!left_buf);
 453                 wret = balance_node_right(trans, root, mid_buf, left_buf);
 454                 if (wret < 0) {
 455                         ret = wret;
 456                         goto enospc;
 457                 }
 458                 BUG_ON(wret == 1);
 459         }
 460         if (btrfs_header_nritems(&mid->header) == 0) {
 461                 /* we've managed to empty the middle node, drop it */
 462                 u64 blocknr = bh_blocknr(mid_buf);
 463                 clean_tree_block(trans, root, mid_buf);
 464                 wait_on_buffer(mid_buf);
 465                 btrfs_block_release(root, mid_buf);
 466                 mid_buf = NULL;
 467                 mid = NULL;
 468                 wret = del_ptr(trans, root, path, level + 1, pslot);
 469                 if (wret)
 470                         ret = wret;
 471                 wret = btrfs_free_extent(trans, root, blocknr, 1, 1);
 472                 if (wret)
 473                         ret = wret;
 474         } else {
 475                 /* update the parent key to reflect our changes */
 476                 btrfs_memcpy(root, parent,
 477                              &parent->ptrs[pslot].key, &mid->ptrs[0].key,
 478                              sizeof(struct btrfs_disk_key));
 479                 btrfs_mark_buffer_dirty(parent_buf);
 480         }
 481
 482         /* update the path */
 483         if (left_buf) {
 484                 if (btrfs_header_nritems(&left->header) > orig_slot) {
 485                         get_bh(left_buf);
 486                         path->nodes[level] = left_buf;
 487                         path->slots[level + 1] -= 1;
 488                         path->slots[level] = orig_slot;
 489                         if (mid_buf)
 490                                 btrfs_block_release(root, mid_buf);
 491                 } else {
 492                         orig_slot -= btrfs_header_nritems(&left->header);
 493                         path->slots[level] = orig_slot;
 494                 }
 495         }
 496         /* double check we haven't messed things up */
 497         check_block(root, path, level);
 498         if (orig_ptr !=
 499             btrfs_node_blockptr(btrfs_buffer_node(path->nodes[level]),
 500                                 path->slots[level]))
 501                 BUG();
 502 enospc:
 503         if (right_buf)
 504                 btrfs_block_release(root, right_buf);
 505         if (left_buf)
 506                 btrfs_block_release(root, left_buf);
 507         return ret;
 508 }
 509
 510 /* returns zero if the push worked, non-zero otherwise */
 511 static int push_nodes_for_insert(struct btrfs_trans_handle *trans,
 512                                 struct btrfs_root *root,
 513                                 struct btrfs_path *path, int level)
 514 {
 515         struct buffer_head *right_buf;
 516         struct buffer_head *mid_buf;
 517         struct buffer_head *left_buf;
 518         struct buffer_head *parent_buf = NULL;
 519         struct btrfs_node *right = NULL;
 520         struct btrfs_node *mid;
 521         struct btrfs_node *left = NULL;
 522         struct btrfs_node *parent = NULL;
 523         int ret = 0;
 524         int wret;
 525         int pslot;
 526         int orig_slot = path->slots[level];
 527         u64 orig_ptr;
 528
 529         if (level == 0)
 530                 return 1;
 531
 532         mid_buf = path->nodes[level];
 533         mid = btrfs_buffer_node(mid_buf);
 534         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 535
 536         if (level < BTRFS_MAX_LEVEL - 1)
 537                 parent_buf = path->nodes[level + 1];
 538         pslot = path->slots[level + 1];
 539
 540         if (!parent_buf)
 541                 return 1;
 542         parent = btrfs_buffer_node(parent_buf);
 543
 544         left_buf = read_node_slot(root, parent_buf, pslot - 1);
 545
 546         /* first, try to make some room in the middle buffer */
 547         if (left_buf) {
 548                 u32 left_nr;
 549                 left = btrfs_buffer_node(left_buf);
 550                 left_nr = btrfs_header_nritems(&left->header);
 551                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 552                         wret = 1;
 553                 } else {
 554                         ret = btrfs_cow_block(trans, root, left_buf, parent_buf,
 555                                               pslot - 1, &left_buf);
 556                         if (ret)
 557                                 wret = 1;
 558                         else {
 559                                 left = btrfs_buffer_node(left_buf);
 560                                 wret = push_node_left(trans, root,
 561                                                       left_buf, mid_buf);
 562                         }
 563                 }
 564                 if (wret < 0)
 565                         ret = wret;
 566                 if (wret == 0) {
 567                         orig_slot += left_nr;
 568                         btrfs_memcpy(root, parent,
 569                                      &parent->ptrs[pslot].key,
 570                                      &mid->ptrs[0].key,
 571                                      sizeof(struct btrfs_disk_key));
 572                         btrfs_mark_buffer_dirty(parent_buf);
 573                         if (btrfs_header_nritems(&left->header) > orig_slot) {
 574                                 path->nodes[level] = left_buf;
 575                                 path->slots[level + 1] -= 1;
 576                                 path->slots[level] = orig_slot;
 577                                 btrfs_block_release(root, mid_buf);
 578                         } else {
 579                                 orig_slot -=
 580                                         btrfs_header_nritems(&left->header);
 581                                 path->slots[level] = orig_slot;
 582                                 btrfs_block_release(root, left_buf);
 583                         }
 584                         check_node(root, path, level);
 585                         return 0;
 586                 }
 587                 btrfs_block_release(root, left_buf);
 588         }
 589         right_buf = read_node_slot(root, parent_buf, pslot + 1);
 590
 591         /*
 592          * then try to empty the right most buffer into the middle
 593          */
 594         if (right_buf) {
 595                 u32 right_nr;
 596                 right = btrfs_buffer_node(right_buf);
 597                 right_nr = btrfs_header_nritems(&right->header);
 598                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 599                         wret = 1;
 600                 } else {
 601                         ret = btrfs_cow_block(trans, root, right_buf,
 602                                               parent_buf, pslot + 1,
 603                                               &right_buf);
 604                         if (ret)
 605                                 wret = 1;
 606                         else {
 607                                 right = btrfs_buffer_node(right_buf);
 608                                 wret = balance_node_right(trans, root,
 609                                                           right_buf, mid_buf);
 610                         }
 611                 }
 612                 if (wret < 0)
 613                         ret = wret;
 614                 if (wret == 0) {
 615                         btrfs_memcpy(root, parent,
 616                                      &parent->ptrs[pslot + 1].key,
 617                                      &right->ptrs[0].key,
 618                                      sizeof(struct btrfs_disk_key));
 619                         btrfs_mark_buffer_dirty(parent_buf);
 620                         if (btrfs_header_nritems(&mid->header) <= orig_slot) {
 621                                 path->nodes[level] = right_buf;
 622                                 path->slots[level + 1] += 1;
 623                                 path->slots[level] = orig_slot -
 624                                         btrfs_header_nritems(&mid->header);
 625                                 btrfs_block_release(root, mid_buf);
 626                         } else {
 627                                 btrfs_block_release(root, right_buf);
 628                         }
 629                         check_node(root, path, level);
 630                         return 0;
 631                 }
 632                 btrfs_block_release(root, right_buf);
 633         }
 634         check_node(root, path, level);
 635         return 1;
 636 }
 637
 638 /*
 639  * look for key in the tree.  path is filled in with nodes along the way
 640  * if key is found, we return zero and you can find the item in the leaf
 641  * level of the path (level 0)
 642  *
 643  * If the key isn't found, the path points to the slot where it should
 644  * be inserted, and 1 is returned.  If there are other errors during the
 645  * search a negative error number is returned.
 646  *
 647  * if ins_len > 0, nodes and leaves will be split as we walk down the
 648  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
 649  * possible)
 650  */
 651 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
 652                       *root, struct btrfs_key *key, struct btrfs_path *p, int
 653                       ins_len, int cow)
 654 {
 655         struct buffer_head *b;
 656         struct buffer_head *cow_buf;
 657         struct btrfs_node *c;
 658         int slot;
 659         int ret;
 660         int level;
 661
 662         WARN_ON(p->nodes[0] != NULL);
 663         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
 664 again:
 665         b = root->node;
 666         get_bh(b);
 667         while (b) {
 668                 c = btrfs_buffer_node(b);
 669                 level = btrfs_header_level(&c->header);
 670                 if (cow) {
 671                         int wret;
 672                         wret = btrfs_cow_block(trans, root, b,
 673                                                p->nodes[level + 1],
 674                                                p->slots[level + 1],
 675                                                &cow_buf);
 676                         if (wret) {
 677                                 btrfs_block_release(root, cow_buf);
 678                                 return wret;
 679                         }
 680                         b = cow_buf;
 681                         c = btrfs_buffer_node(b);
 682                 }
 683                 BUG_ON(!cow && ins_len);
 684                 if (level != btrfs_header_level(&c->header))
 685                         WARN_ON(1);
 686                 level = btrfs_header_level(&c->header);
 687                 p->nodes[level] = b;
 688                 ret = check_block(root, p, level);
 689                 if (ret)
 690                         return -1;
 691                 ret = bin_search(c, key, &slot);
 692                 if (!btrfs_is_leaf(c)) {
 693                         if (ret && slot > 0)
 694                                 slot -= 1;
 695                         p->slots[level] = slot;
 696                         if (ins_len > 0 && btrfs_header_nritems(&c->header) >=
 697                             BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 698                                 int sret = split_node(trans, root, p, level);
 699                                 BUG_ON(sret > 0);
 700                                 if (sret)
 701                                         return sret;
 702                                 b = p->nodes[level];
 703                                 c = btrfs_buffer_node(b);
 704                                 slot = p->slots[level];
 705                         } else if (ins_len < 0) {
 706                                 int sret = balance_level(trans, root, p,
 707                                                          level);
 708                                 if (sret)
 709                                         return sret;
 710                                 b = p->nodes[level];
 711                                 if (!b)
 712                                         goto again;
 713                                 c = btrfs_buffer_node(b);
 714                                 slot = p->slots[level];
 715                                 BUG_ON(btrfs_header_nritems(&c->header) == 1);
 716                         }
 717                         b = read_tree_block(root, btrfs_node_blockptr(c, slot));
 718                 } else {
 719                         struct btrfs_leaf *l = (struct btrfs_leaf *)c;
 720                         p->slots[level] = slot;
 721                         if (ins_len > 0 && btrfs_leaf_free_space(root, l) <
 722                             sizeof(struct btrfs_item) + ins_len) {
 723                                 int sret = split_leaf(trans, root, key,
 724                                                       p, ins_len);
 725                                 BUG_ON(sret > 0);
 726                                 if (sret)
 727                                         return sret;
 728                         }
 729                         return ret;
 730                 }
 731         }
 732         return 1;
 733 }
 734
 735 /*
 736  * adjust the pointers going up the tree, starting at level
 737  * making sure the right key of each node is points to 'key'.
 738  * This is used after shifting pointers to the left, so it stops
 739  * fixing up pointers when a given leaf/node is not in slot 0 of the
 740  * higher levels
 741  *
 742  * If this fails to write a tree block, it returns -1, but continues
 743  * fixing up the blocks in ram so the tree is consistent.
 744  */
 745 static int fixup_low_keys(struct btrfs_trans_handle *trans, struct btrfs_root
 746                           *root, struct btrfs_path *path, struct btrfs_disk_key
 747                           *key, int level)
 748 {
 749         int i;
 750         int ret = 0;
 751         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
 752                 struct btrfs_node *t;
 753                 int tslot = path->slots[i];
 754                 if (!path->nodes[i])
 755                         break;
 756                 t = btrfs_buffer_node(path->nodes[i]);
 757                 btrfs_memcpy(root, t, &t->ptrs[tslot].key, key, sizeof(*key));
 758                 btrfs_mark_buffer_dirty(path->nodes[i]);
 759                 if (tslot != 0)
 760                         break;
 761         }
 762         return ret;
 763 }
 764
 765 /*
 766  * try to push data from one node into the next node left in the
 767  * tree.
 768  *
 769  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
 770  * error, and > 0 if there was no room in the left hand block.
 771  */
 772 static int push_node_left(struct btrfs_trans_handle *trans, struct btrfs_root
 773                           *root, struct buffer_head *dst_buf, struct
 774                           buffer_head *src_buf)
 775 {
 776         struct btrfs_node *src = btrfs_buffer_node(src_buf);
 777         struct btrfs_node *dst = btrfs_buffer_node(dst_buf);
 778         int push_items = 0;
 779         int src_nritems;
 780         int dst_nritems;
 781         int ret = 0;
 782
 783         src_nritems = btrfs_header_nritems(&src->header);
 784         dst_nritems = btrfs_header_nritems(&dst->header);
 785         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
 786
 787         if (push_items <= 0) {
 788                 return 1;
 789         }
 790
 791         if (src_nritems < push_items)
 792                 push_items = src_nritems;
 793
 794         btrfs_memcpy(root, dst, dst->ptrs + dst_nritems, src->ptrs,
 795                      push_items * sizeof(struct btrfs_key_ptr));
 796         if (push_items < src_nritems) {
 797                 btrfs_memmove(root, src, src->ptrs, src->ptrs + push_items,
 798                         (src_nritems - push_items) *
 799                         sizeof(struct btrfs_key_ptr));
 800         }
 801         btrfs_set_header_nritems(&src->header, src_nritems - push_items);
 802         btrfs_set_header_nritems(&dst->header, dst_nritems + push_items);
 803         btrfs_mark_buffer_dirty(src_buf);
 804         btrfs_mark_buffer_dirty(dst_buf);
 805         return ret;
 806 }
 807
 808 /*
 809  * try to push data from one node into the next node right in the
 810  * tree.
 811  *
 812  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
 813  * error, and > 0 if there was no room in the right hand block.
 814  *
 815  * this will  only push up to 1/2 the contents of the left node over
 816  */
 817 static int balance_node_right(struct btrfs_trans_handle *trans, struct
 818                               btrfs_root *root, struct buffer_head *dst_buf,
 819                               struct buffer_head *src_buf)
 820 {
 821         struct btrfs_node *src = btrfs_buffer_node(src_buf);
 822         struct btrfs_node *dst = btrfs_buffer_node(dst_buf);
 823         int push_items = 0;
 824         int max_push;
 825         int src_nritems;
 826         int dst_nritems;
 827         int ret = 0;
 828
 829         src_nritems = btrfs_header_nritems(&src->header);
 830         dst_nritems = btrfs_header_nritems(&dst->header);
 831         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
 832         if (push_items <= 0) {
 833                 return 1;
 834         }
 835
 836         max_push = src_nritems / 2 + 1;
 837         /* don't try to empty the node */
 838         if (max_push > src_nritems)
 839                 return 1;
 840         if (max_push < push_items)
 841                 push_items = max_push;
 842
 843         btrfs_memmove(root, dst, dst->ptrs + push_items, dst->ptrs,
 844                       dst_nritems * sizeof(struct btrfs_key_ptr));
 845
 846         btrfs_memcpy(root, dst, dst->ptrs,
 847                      src->ptrs + src_nritems - push_items,
 848                      push_items * sizeof(struct btrfs_key_ptr));
 849
 850         btrfs_set_header_nritems(&src->header, src_nritems - push_items);
 851         btrfs_set_header_nritems(&dst->header, dst_nritems + push_items);
 852
 853         btrfs_mark_buffer_dirty(src_buf);
 854         btrfs_mark_buffer_dirty(dst_buf);
 855         return ret;
 856 }
 857
 858 /*
 859  * helper function to insert a new root level in the tree.
 860  * A new node is allocated, and a single item is inserted to
 861  * point to the existing root
 862  *
 863  * returns zero on success or < 0 on failure.
 864  */
 865 static int insert_new_root(struct btrfs_trans_handle *trans, struct btrfs_root
 866                            *root, struct btrfs_path *path, int level)
 867 {
 868         struct buffer_head *t;
 869         struct btrfs_node *lower;
 870         struct btrfs_node *c;
 871         struct btrfs_disk_key *lower_key;
 872
 873         BUG_ON(path->nodes[level]);
 874         BUG_ON(path->nodes[level-1] != root->node);
 875
 876         t = btrfs_alloc_free_block(trans, root, root->node->b_blocknr);
 877         if (IS_ERR(t))
 878                 return PTR_ERR(t);
 879         c = btrfs_buffer_node(t);
 880         memset(c, 0, root->blocksize);
 881         btrfs_set_header_nritems(&c->header, 1);
 882         btrfs_set_header_level(&c->header, level);
 883         btrfs_set_header_blocknr(&c->header, bh_blocknr(t));
 884         btrfs_set_header_generation(&c->header, trans->transid);
 885         btrfs_set_header_owner(&c->header, root->root_key.objectid);
 886         lower = btrfs_buffer_node(path->nodes[level-1]);
 887         memcpy(c->header.fsid, root->fs_info->disk_super->fsid,
 888                sizeof(c->header.fsid));
 889         if (btrfs_is_leaf(lower))
 890                 lower_key = &((struct btrfs_leaf *)lower)->items[0].key;
 891         else
 892                 lower_key = &lower->ptrs[0].key;
 893         btrfs_memcpy(root, c, &c->ptrs[0].key, lower_key,
 894                      sizeof(struct btrfs_disk_key));
 895         btrfs_set_node_blockptr(c, 0, bh_blocknr(path->nodes[level - 1]));
 896
 897         btrfs_mark_buffer_dirty(t);
 898
 899         /* the super has an extra ref to root->node */
 900         btrfs_block_release(root, root->node);
 901         root->node = t;
 902         get_bh(t);
 903         path->nodes[level] = t;
 904         path->slots[level] = 0;
 905         return 0;
 906 }
 907
 908 /*
 909  * worker function to insert a single pointer in a node.
 910  * the node should have enough room for the pointer already
 911  *
 912  * slot and level indicate where you want the key to go, and
 913  * blocknr is the block the key points to.
 914  *
 915  * returns zero on success and < 0 on any error
 916  */
 917 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
 918                       *root, struct btrfs_path *path, struct btrfs_disk_key
 919                       *key, u64 blocknr, int slot, int level)
 920 {
 921         struct btrfs_node *lower;
 922         int nritems;
 923
 924         BUG_ON(!path->nodes[level]);
 925         lower = btrfs_buffer_node(path->nodes[level]);
 926         nritems = btrfs_header_nritems(&lower->header);
 927         if (slot > nritems)
 928                 BUG();
 929         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
 930                 BUG();
 931         if (slot != nritems) {
 932                 btrfs_memmove(root, lower, lower->ptrs + slot + 1,
 933                               lower->ptrs + slot,
 934                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
 935         }
 936         btrfs_memcpy(root, lower, &lower->ptrs[slot].key,
 937                      key, sizeof(struct btrfs_disk_key));
 938         btrfs_set_node_blockptr(lower, slot, blocknr);
 939         btrfs_set_header_nritems(&lower->header, nritems + 1);
 940         btrfs_mark_buffer_dirty(path->nodes[level]);
 941         check_node(root, path, level);
 942         return 0;
 943 }
 944
 945 /*
 946  * split the node at the specified level in path in two.
 947  * The path is corrected to point to the appropriate node after the split
 948  *
 949  * Before splitting this tries to make some room in the node by pushing
 950  * left and right, if either one works, it returns right away.
 951  *
 952  * returns 0 on success and < 0 on failure
 953  */
 954 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
 955                       *root, struct btrfs_path *path, int level)
 956 {
 957         struct buffer_head *t;
 958         struct btrfs_node *c;
 959         struct buffer_head *split_buffer;
 960         struct btrfs_node *split;
 961         int mid;
 962         int ret;
 963         int wret;
 964         u32 c_nritems;
 965
 966         t = path->nodes[level];
 967         c = btrfs_buffer_node(t);
 968         if (t == root->node) {
 969                 /* trying to split the root, lets make a new one */
 970                 ret = insert_new_root(trans, root, path, level + 1);
 971                 if (ret)
 972                         return ret;
 973         } else {
 974                 ret = push_nodes_for_insert(trans, root, path, level);
 975                 t = path->nodes[level];
 976                 c = btrfs_buffer_node(t);
 977                 if (!ret &&
 978                     btrfs_header_nritems(&c->header) <
 979                     BTRFS_NODEPTRS_PER_BLOCK(root) - 1)
 980                         return 0;
 981                 if (ret < 0)
 982                         return ret;
 983         }
 984
 985         c_nritems = btrfs_header_nritems(&c->header);
 986         split_buffer = btrfs_alloc_free_block(trans, root, t->b_blocknr);
 987         if (IS_ERR(split_buffer))
 988                 return PTR_ERR(split_buffer);
 989
 990         split = btrfs_buffer_node(split_buffer);
 991         btrfs_set_header_flags(&split->header, btrfs_header_flags(&c->header));
 992         btrfs_set_header_level(&split->header, btrfs_header_level(&c->header));
 993         btrfs_set_header_blocknr(&split->header, bh_blocknr(split_buffer));
 994         btrfs_set_header_generation(&split->header, trans->transid);
 995         btrfs_set_header_owner(&split->header, root->root_key.objectid);
 996         memcpy(split->header.fsid, root->fs_info->disk_super->fsid,
 997                sizeof(split->header.fsid));
 998         mid = (c_nritems + 1) / 2;
 999         btrfs_memcpy(root, split, split->ptrs, c->ptrs + mid,
1000                      (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1001         btrfs_set_header_nritems(&split->header, c_nritems - mid);
1002         btrfs_set_header_nritems(&c->header, mid);
1003         ret = 0;
1004
1005         btrfs_mark_buffer_dirty(t);
1006         btrfs_mark_buffer_dirty(split_buffer);
1007         wret = insert_ptr(trans, root, path, &split->ptrs[0].key,
1008                           bh_blocknr(split_buffer), path->slots[level + 1] + 1,
1009                           level + 1);
1010         if (wret)
1011                 ret = wret;
1012
1013         if (path->slots[level] >= mid) {
1014                 path->slots[level] -= mid;
1015                 btrfs_block_release(root, t);
1016                 path->nodes[level] = split_buffer;
1017                 path->slots[level + 1] += 1;
1018         } else {
1019                 btrfs_block_release(root, split_buffer);
1020         }
1021         return ret;
1022 }
1023
1024 /*
1025  * how many bytes are required to store the items in a leaf.  start
1026  * and nr indicate which items in the leaf to check.  This totals up the
1027  * space used both by the item structs and the item data
1028  */
1029 static int leaf_space_used(struct btrfs_leaf *l, int start, int nr)
1030 {
1031         int data_len;
1032         int nritems = btrfs_header_nritems(&l->header);
1033         int end = min(nritems, start + nr) - 1;
1034
1035         if (!nr)
1036                 return 0;
1037         data_len = btrfs_item_end(l->items + start);
1038         data_len = data_len - btrfs_item_offset(l->items + end);
1039         data_len += sizeof(struct btrfs_item) * nr;
1040         WARN_ON(data_len < 0);
1041         return data_len;
1042 }
1043
1044 /*
1045  * The space between the end of the leaf items and
1046  * the start of the leaf data.  IOW, how much room
1047  * the leaf has left for both items and data
1048  */
1049 int btrfs_leaf_free_space(struct btrfs_root *root, struct btrfs_leaf *leaf)
1050 {
1051         int nritems = btrfs_header_nritems(&leaf->header);
1052         return BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1053 }
1054
1055 /*
1056  * push some data in the path leaf to the right, trying to free up at
1057  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1058  *
1059  * returns 1 if the push failed because the other node didn't have enough
1060  * room, 0 if everything worked out and < 0 if there were major errors.
1061  */
1062 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1063                            *root, struct btrfs_path *path, int data_size)
1064 {
1065         struct buffer_head *left_buf = path->nodes[0];
1066         struct btrfs_leaf *left = btrfs_buffer_leaf(left_buf);
1067         struct btrfs_leaf *right;
1068         struct buffer_head *right_buf;
1069         struct buffer_head *upper;
1070         struct btrfs_node *upper_node;
1071         int slot;
1072         int i;
1073         int free_space;
1074         int push_space = 0;
1075         int push_items = 0;
1076         struct btrfs_item *item;
1077         u32 left_nritems;
1078         u32 right_nritems;
1079         int ret;
1080
1081         slot = path->slots[1];
1082         if (!path->nodes[1]) {
1083                 return 1;
1084         }
1085         upper = path->nodes[1];
1086         upper_node = btrfs_buffer_node(upper);
1087         if (slot >= btrfs_header_nritems(&upper_node->header) - 1) {
1088                 return 1;
1089         }
1090         right_buf = read_tree_block(root,
1091                     btrfs_node_blockptr(btrfs_buffer_node(upper), slot + 1));
1092         right = btrfs_buffer_leaf(right_buf);
1093         free_space = btrfs_leaf_free_space(root, right);
1094         if (free_space < data_size + sizeof(struct btrfs_item)) {
1095                 btrfs_block_release(root, right_buf);
1096                 return 1;
1097         }
1098         /* cow and double check */
1099         ret = btrfs_cow_block(trans, root, right_buf, upper,
1100                               slot + 1, &right_buf);
1101         if (ret) {
1102                 btrfs_block_release(root, right_buf);
1103                 return 1;
1104         }
1105         right = btrfs_buffer_leaf(right_buf);
1106         free_space = btrfs_leaf_free_space(root, right);
1107         if (free_space < data_size + sizeof(struct btrfs_item)) {
1108                 btrfs_block_release(root, right_buf);
1109                 return 1;
1110         }
1111
1112         left_nritems = btrfs_header_nritems(&left->header);
1113         if (left_nritems == 0) {
1114                 btrfs_block_release(root, right_buf);
1115                 return 1;
1116         }
1117         for (i = left_nritems - 1; i >= 1; i--) {
1118                 item = left->items + i;
1119                 if (path->slots[0] == i)
1120                         push_space += data_size + sizeof(*item);
1121                 if (btrfs_item_size(item) + sizeof(*item) + push_space >
1122                     free_space)
1123                         break;
1124                 push_items++;
1125                 push_space += btrfs_item_size(item) + sizeof(*item);
1126         }
1127         if (push_items == 0) {
1128                 btrfs_block_release(root, right_buf);
1129                 return 1;
1130         }
1131         if (push_items == left_nritems)
1132                 WARN_ON(1);
1133         right_nritems = btrfs_header_nritems(&right->header);
1134         /* push left to right */
1135         push_space = btrfs_item_end(left->items + left_nritems - push_items);
1136         push_space -= leaf_data_end(root, left);
1137         /* make room in the right data area */
1138         btrfs_memmove(root, right, btrfs_leaf_data(right) +
1139                       leaf_data_end(root, right) - push_space,
1140                       btrfs_leaf_data(right) +
1141                       leaf_data_end(root, right), BTRFS_LEAF_DATA_SIZE(root) -
1142                       leaf_data_end(root, right));
1143         /* copy from the left data area */
1144         btrfs_memcpy(root, right, btrfs_leaf_data(right) +
1145                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1146                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1147                      push_space);
1148         btrfs_memmove(root, right, right->items + push_items, right->items,
1149                 right_nritems * sizeof(struct btrfs_item));
1150         /* copy the items from left to right */
1151         btrfs_memcpy(root, right, right->items, left->items +
1152                      left_nritems - push_items,
1153                      push_items * sizeof(struct btrfs_item));
1154
1155         /* update the item pointers */
1156         right_nritems += push_items;
1157         btrfs_set_header_nritems(&right->header, right_nritems);
1158         push_space = BTRFS_LEAF_DATA_SIZE(root);
1159         for (i = 0; i < right_nritems; i++) {
1160                 btrfs_set_item_offset(right->items + i, push_space -
1161                                       btrfs_item_size(right->items + i));
1162                 push_space = btrfs_item_offset(right->items + i);
1163         }
1164         left_nritems -= push_items;
1165         btrfs_set_header_nritems(&left->header, left_nritems);
1166
1167         btrfs_mark_buffer_dirty(left_buf);
1168         btrfs_mark_buffer_dirty(right_buf);
1169
1170         btrfs_memcpy(root, upper_node, &upper_node->ptrs[slot + 1].key,
1171                 &right->items[0].key, sizeof(struct btrfs_disk_key));
1172         btrfs_mark_buffer_dirty(upper);
1173
1174         /* then fixup the leaf pointer in the path */
1175         if (path->slots[0] >= left_nritems) {
1176                 path->slots[0] -= left_nritems;
1177                 btrfs_block_release(root, path->nodes[0]);
1178                 path->nodes[0] = right_buf;
1179                 path->slots[1] += 1;
1180         } else {
1181                 btrfs_block_release(root, right_buf);
1182         }
1183         if (path->nodes[1])
1184                 check_node(root, path, 1);
1185         return 0;
1186 }
1187 /*
1188  * push some data in the path leaf to the left, trying to free up at
1189  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1190  */
1191 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1192                           *root, struct btrfs_path *path, int data_size)
1193 {
1194         struct buffer_head *right_buf = path->nodes[0];
1195         struct btrfs_leaf *right = btrfs_buffer_leaf(right_buf);
1196         struct buffer_head *t;
1197         struct btrfs_leaf *left;
1198         int slot;
1199         int i;
1200         int free_space;
1201         int push_space = 0;
1202         int push_items = 0;
1203         struct btrfs_item *item;
1204         u32 old_left_nritems;
1205         int ret = 0;
1206         int wret;
1207
1208         slot = path->slots[1];
1209         if (slot == 0) {
1210                 return 1;
1211         }
1212         if (!path->nodes[1]) {
1213                 return 1;
1214         }
1215         t = read_tree_block(root,
1216             btrfs_node_blockptr(btrfs_buffer_node(path->nodes[1]), slot - 1));
1217         left = btrfs_buffer_leaf(t);
1218         free_space = btrfs_leaf_free_space(root, left);
1219         if (free_space < data_size + sizeof(struct btrfs_item)) {
1220                 btrfs_block_release(root, t);
1221                 return 1;
1222         }
1223
1224         /* cow and double check */
1225         ret = btrfs_cow_block(trans, root, t, path->nodes[1], slot - 1, &t);
1226         if (ret) {
1227                 /* we hit -ENOSPC, but it isn't fatal here */
1228                 return 1;
1229         }
1230         left = btrfs_buffer_leaf(t);
1231         free_space = btrfs_leaf_free_space(root, left);
1232         if (free_space < data_size + sizeof(struct btrfs_item)) {
1233                 btrfs_block_release(root, t);
1234                 return 1;
1235         }
1236
1237         if (btrfs_header_nritems(&right->header) == 0) {
1238                 btrfs_block_release(root, t);
1239                 return 1;
1240         }
1241
1242         for (i = 0; i < btrfs_header_nritems(&right->header) - 1; i++) {
1243                 item = right->items + i;
1244                 if (path->slots[0] == i)
1245                         push_space += data_size + sizeof(*item);
1246                 if (btrfs_item_size(item) + sizeof(*item) + push_space >
1247                     free_space)
1248                         break;
1249                 push_items++;
1250                 push_space += btrfs_item_size(item) + sizeof(*item);
1251         }
1252         if (push_items == 0) {
1253                 btrfs_block_release(root, t);
1254                 return 1;
1255         }
1256         if (push_items == btrfs_header_nritems(&right->header))
1257                 WARN_ON(1);
1258         /* push data from right to left */
1259         btrfs_memcpy(root, left, left->items +
1260                      btrfs_header_nritems(&left->header),
1261                      right->items, push_items * sizeof(struct btrfs_item));
1262         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1263                      btrfs_item_offset(right->items + push_items -1);
1264         btrfs_memcpy(root, left, btrfs_leaf_data(left) +
1265                      leaf_data_end(root, left) - push_space,
1266                      btrfs_leaf_data(right) +
1267                      btrfs_item_offset(right->items + push_items - 1),
1268                      push_space);
1269         old_left_nritems = btrfs_header_nritems(&left->header);
1270         BUG_ON(old_left_nritems < 0);
1271
1272         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1273                 u32 ioff = btrfs_item_offset(left->items + i);
1274                 btrfs_set_item_offset(left->items + i, ioff -
1275                                      (BTRFS_LEAF_DATA_SIZE(root) -
1276                                       btrfs_item_offset(left->items +
1277                                                         old_left_nritems - 1)));
1278         }
1279         btrfs_set_header_nritems(&left->header, old_left_nritems + push_items);
1280
1281         /* fixup right node */
1282         push_space = btrfs_item_offset(right->items + push_items - 1) -
1283                      leaf_data_end(root, right);
1284         btrfs_memmove(root, right, btrfs_leaf_data(right) +
1285                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1286                       btrfs_leaf_data(right) +
1287                       leaf_data_end(root, right), push_space);
1288         btrfs_memmove(root, right, right->items, right->items + push_items,
1289                 (btrfs_header_nritems(&right->header) - push_items) *
1290                 sizeof(struct btrfs_item));
1291         btrfs_set_header_nritems(&right->header,
1292                                  btrfs_header_nritems(&right->header) -
1293                                  push_items);
1294         push_space = BTRFS_LEAF_DATA_SIZE(root);
1295
1296         for (i = 0; i < btrfs_header_nritems(&right->header); i++) {
1297                 btrfs_set_item_offset(right->items + i, push_space -
1298                                       btrfs_item_size(right->items + i));
1299                 push_space = btrfs_item_offset(right->items + i);
1300         }
1301
1302         btrfs_mark_buffer_dirty(t);
1303         btrfs_mark_buffer_dirty(right_buf);
1304
1305         wret = fixup_low_keys(trans, root, path, &right->items[0].key, 1);
1306         if (wret)
1307                 ret = wret;
1308
1309         /* then fixup the leaf pointer in the path */
1310         if (path->slots[0] < push_items) {
1311                 path->slots[0] += old_left_nritems;
1312                 btrfs_block_release(root, path->nodes[0]);
1313                 path->nodes[0] = t;
1314                 path->slots[1] -= 1;
1315         } else {
1316                 btrfs_block_release(root, t);
1317                 path->slots[0] -= push_items;
1318         }
1319         BUG_ON(path->slots[0] < 0);
1320         if (path->nodes[1])
1321                 check_node(root, path, 1);
1322         return ret;
1323 }
1324
1325 /*
1326  * split the path's leaf in two, making sure there is at least data_size
1327  * available for the resulting leaf level of the path.
1328  *
1329  * returns 0 if all went well and < 0 on failure.
1330  */
1331 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
1332                       *root, struct btrfs_key *ins_key,
1333                       struct btrfs_path *path, int data_size)
1334 {
1335         struct buffer_head *l_buf;
1336         struct btrfs_leaf *l;
1337         u32 nritems;
1338         int mid;
1339         int slot;
1340         struct btrfs_leaf *right;
1341         struct buffer_head *right_buffer;
1342         int space_needed = data_size + sizeof(struct btrfs_item);
1343         int data_copy_size;
1344         int rt_data_off;
1345         int i;
1346         int ret = 0;
1347         int wret;
1348         int double_split = 0;
1349         struct btrfs_disk_key disk_key;
1350
1351         /* first try to make some room by pushing left and right */
1352         wret = push_leaf_left(trans, root, path, data_size);
1353         if (wret < 0)
1354                 return wret;
1355         if (wret) {
1356                 wret = push_leaf_right(trans, root, path, data_size);
1357                 if (wret < 0)
1358                         return wret;
1359         }
1360         l_buf = path->nodes[0];
1361         l = btrfs_buffer_leaf(l_buf);
1362
1363         /* did the pushes work? */
1364         if (btrfs_leaf_free_space(root, l) >=
1365             sizeof(struct btrfs_item) + data_size)
1366                 return 0;
1367
1368         if (!path->nodes[1]) {
1369                 ret = insert_new_root(trans, root, path, 1);
1370                 if (ret)
1371                         return ret;
1372         }
1373         slot = path->slots[0];
1374         nritems = btrfs_header_nritems(&l->header);
1375         mid = (nritems + 1)/ 2;
1376
1377         right_buffer = btrfs_alloc_free_block(trans, root, l_buf->b_blocknr);
1378         if (IS_ERR(right_buffer))
1379                 return PTR_ERR(right_buffer);
1380
1381         right = btrfs_buffer_leaf(right_buffer);
1382         memset(&right->header, 0, sizeof(right->header));
1383         btrfs_set_header_blocknr(&right->header, bh_blocknr(right_buffer));
1384         btrfs_set_header_generation(&right->header, trans->transid);
1385         btrfs_set_header_owner(&right->header, root->root_key.objectid);
1386         btrfs_set_header_level(&right->header, 0);
1387         memcpy(right->header.fsid, root->fs_info->disk_super->fsid,
1388                sizeof(right->header.fsid));
1389         if (mid <= slot) {
1390                 if (nritems == 1 ||
1391                     leaf_space_used(l, mid, nritems - mid) + space_needed >
1392                         BTRFS_LEAF_DATA_SIZE(root)) {
1393                         if (slot >= nritems) {
1394                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1395                                 btrfs_set_header_nritems(&right->header, 0);
1396                                 wret = insert_ptr(trans, root, path,
1397                                                   &disk_key,
1398                                                   bh_blocknr(right_buffer),
1399                                                   path->slots[1] + 1, 1);
1400                                 if (wret)
1401                                         ret = wret;
1402                                 btrfs_block_release(root, path->nodes[0]);
1403                                 path->nodes[0] = right_buffer;
1404                                 path->slots[0] = 0;
1405                                 path->slots[1] += 1;
1406                                 return ret;
1407                         }
1408                         mid = slot;
1409                         double_split = 1;
1410                 }
1411         } else {
1412                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
1413                         BTRFS_LEAF_DATA_SIZE(root)) {
1414                         if (slot == 0) {
1415                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1416                                 btrfs_set_header_nritems(&right->header, 0);
1417                                 wret = insert_ptr(trans, root, path,
1418                                                   &disk_key,
1419                                                   bh_blocknr(right_buffer),
1420                                                   path->slots[1], 1);
1421                                 if (wret)
1422                                         ret = wret;
1423                                 btrfs_block_release(root, path->nodes[0]);
1424                                 path->nodes[0] = right_buffer;
1425                                 path->slots[0] = 0;
1426                                 if (path->slots[1] == 0) {
1427                                         wret = fixup_low_keys(trans, root,
1428                                                    path, &disk_key, 1);
1429                                         if (wret)
1430                                                 ret = wret;
1431                                 }
1432                                 return ret;
1433                         }
1434                         mid = slot;
1435                         double_split = 1;
1436                 }
1437         }
1438         btrfs_set_header_nritems(&right->header, nritems - mid);
1439         data_copy_size = btrfs_item_end(l->items + mid) -
1440                          leaf_data_end(root, l);
1441         btrfs_memcpy(root, right, right->items, l->items + mid,
1442                      (nritems - mid) * sizeof(struct btrfs_item));
1443         btrfs_memcpy(root, right,
1444                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
1445                      data_copy_size, btrfs_leaf_data(l) +
1446                      leaf_data_end(root, l), data_copy_size);
1447         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
1448                       btrfs_item_end(l->items + mid);
1449
1450         for (i = 0; i < btrfs_header_nritems(&right->header); i++) {
1451                 u32 ioff = btrfs_item_offset(right->items + i);
1452                 btrfs_set_item_offset(right->items + i, ioff + rt_data_off);
1453         }
1454
1455         btrfs_set_header_nritems(&l->header, mid);
1456         ret = 0;
1457         wret = insert_ptr(trans, root, path, &right->items[0].key,
1458                           bh_blocknr(right_buffer), path->slots[1] + 1, 1);
1459         if (wret)
1460                 ret = wret;
1461         btrfs_mark_buffer_dirty(right_buffer);
1462         btrfs_mark_buffer_dirty(l_buf);
1463         BUG_ON(path->slots[0] != slot);
1464         if (mid <= slot) {
1465                 btrfs_block_release(root, path->nodes[0]);
1466                 path->nodes[0] = right_buffer;
1467                 path->slots[0] -= mid;
1468                 path->slots[1] += 1;
1469         } else
1470                 btrfs_block_release(root, right_buffer);
1471         BUG_ON(path->slots[0] < 0);
1472         check_node(root, path, 1);
1473
1474         if (!double_split)
1475                 return ret;
1476         right_buffer = btrfs_alloc_free_block(trans, root, l_buf->b_blocknr);
1477         if (IS_ERR(right_buffer))
1478                 return PTR_ERR(right_buffer);
1479
1480         right = btrfs_buffer_leaf(right_buffer);
1481         memset(&right->header, 0, sizeof(right->header));
1482         btrfs_set_header_blocknr(&right->header, bh_blocknr(right_buffer));
1483         btrfs_set_header_generation(&right->header, trans->transid);
1484         btrfs_set_header_owner(&right->header, root->root_key.objectid);
1485         btrfs_set_header_level(&right->header, 0);
1486         memcpy(right->header.fsid, root->fs_info->disk_super->fsid,
1487                sizeof(right->header.fsid));
1488         btrfs_cpu_key_to_disk(&disk_key, ins_key);
1489         btrfs_set_header_nritems(&right->header, 0);
1490         wret = insert_ptr(trans, root, path,
1491                           &disk_key,
1492                           bh_blocknr(right_buffer),
1493                           path->slots[1], 1);
1494         if (wret)
1495                 ret = wret;
1496         if (path->slots[1] == 0) {
1497                 wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1498                 if (wret)
1499                         ret = wret;
1500         }
1501         btrfs_block_release(root, path->nodes[0]);
1502         path->nodes[0] = right_buffer;
1503         path->slots[0] = 0;
1504         check_node(root, path, 1);
1505         check_leaf(root, path, 0);
1506         return ret;
1507 }
1508
1509 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
1510                         struct btrfs_root *root,
1511                         struct btrfs_path *path,
1512                         u32 new_size)
1513 {
1514         int ret = 0;
1515         int slot;
1516         int slot_orig;
1517         struct btrfs_leaf *leaf;
1518         struct buffer_head *leaf_buf;
1519         u32 nritems;
1520         unsigned int data_end;
1521         unsigned int old_data_start;
1522         unsigned int old_size;
1523         unsigned int size_diff;
1524         int i;
1525
1526         slot_orig = path->slots[0];
1527         leaf_buf = path->nodes[0];
1528         leaf = btrfs_buffer_leaf(leaf_buf);
1529
1530         nritems = btrfs_header_nritems(&leaf->header);
1531         data_end = leaf_data_end(root, leaf);
1532
1533         slot = path->slots[0];
1534         old_data_start = btrfs_item_offset(leaf->items + slot);
1535         old_size = btrfs_item_size(leaf->items + slot);
1536         BUG_ON(old_size <= new_size);
1537         size_diff = old_size - new_size;
1538
1539         BUG_ON(slot < 0);
1540         BUG_ON(slot >= nritems);
1541
1542         /*
1543          * item0..itemN ... dataN.offset..dataN.size .. data0.size
1544          */
1545         /* first correct the data pointers */
1546         for (i = slot; i < nritems; i++) {
1547                 u32 ioff = btrfs_item_offset(leaf->items + i);
1548                 btrfs_set_item_offset(leaf->items + i,
1549                                       ioff + size_diff);
1550         }
1551         /* shift the data */
1552         btrfs_memmove(root, leaf, btrfs_leaf_data(leaf) +
1553                       data_end + size_diff, btrfs_leaf_data(leaf) +
1554                       data_end, old_data_start + new_size - data_end);
1555         btrfs_set_item_size(leaf->items + slot, new_size);
1556         btrfs_mark_buffer_dirty(leaf_buf);
1557
1558         ret = 0;
1559         if (btrfs_leaf_free_space(root, leaf) < 0)
1560                 BUG();
1561         check_leaf(root, path, 0);
1562         return ret;
1563 }
1564
1565 int btrfs_extend_item(struct btrfs_trans_handle *trans, struct btrfs_root
1566                       *root, struct btrfs_path *path, u32 data_size)
1567 {
1568         int ret = 0;
1569         int slot;
1570         int slot_orig;
1571         struct btrfs_leaf *leaf;
1572         struct buffer_head *leaf_buf;
1573         u32 nritems;
1574         unsigned int data_end;
1575         unsigned int old_data;
1576         unsigned int old_size;
1577         int i;
1578
1579         slot_orig = path->slots[0];
1580         leaf_buf = path->nodes[0];
1581         leaf = btrfs_buffer_leaf(leaf_buf);
1582
1583         nritems = btrfs_header_nritems(&leaf->header);
1584         data_end = leaf_data_end(root, leaf);
1585
1586         if (btrfs_leaf_free_space(root, leaf) < data_size)
1587                 BUG();
1588         slot = path->slots[0];
1589         old_data = btrfs_item_end(leaf->items + slot);
1590
1591         BUG_ON(slot < 0);
1592         BUG_ON(slot >= nritems);
1593
1594         /*
1595          * item0..itemN ... dataN.offset..dataN.size .. data0.size
1596          */
1597         /* first correct the data pointers */
1598         for (i = slot; i < nritems; i++) {
1599                 u32 ioff = btrfs_item_offset(leaf->items + i);
1600                 btrfs_set_item_offset(leaf->items + i,
1601                                       ioff - data_size);
1602         }
1603         /* shift the data */
1604         btrfs_memmove(root, leaf, btrfs_leaf_data(leaf) +
1605                       data_end - data_size, btrfs_leaf_data(leaf) +
1606                       data_end, old_data - data_end);
1607         data_end = old_data;
1608         old_size = btrfs_item_size(leaf->items + slot);
1609         btrfs_set_item_size(leaf->items + slot, old_size + data_size);
1610         btrfs_mark_buffer_dirty(leaf_buf);
1611
1612         ret = 0;
1613         if (btrfs_leaf_free_space(root, leaf) < 0)
1614                 BUG();
1615         check_leaf(root, path, 0);
1616         return ret;
1617 }
1618
1619 /*
1620  * Given a key and some data, insert an item into the tree.
1621  * This does all the path init required, making room in the tree if needed.
1622  */
1623 int btrfs_insert_empty_item(struct btrfs_trans_handle *trans, struct btrfs_root
1624                             *root, struct btrfs_path *path, struct btrfs_key
1625                             *cpu_key, u32 data_size)
1626 {
1627         int ret = 0;
1628         int slot;
1629         int slot_orig;
1630         struct btrfs_leaf *leaf;
1631         struct buffer_head *leaf_buf;
1632         u32 nritems;
1633         unsigned int data_end;
1634         struct btrfs_disk_key disk_key;
1635
1636         btrfs_cpu_key_to_disk(&disk_key, cpu_key);
1637
1638         /* create a root if there isn't one */
1639         if (!root->node)
1640                 BUG();
1641         ret = btrfs_search_slot(trans, root, cpu_key, path, data_size, 1);
1642         if (ret == 0) {
1643                 return -EEXIST;
1644         }
1645         if (ret < 0)
1646                 goto out;
1647
1648         slot_orig = path->slots[0];
1649         leaf_buf = path->nodes[0];
1650         leaf = btrfs_buffer_leaf(leaf_buf);
1651
1652         nritems = btrfs_header_nritems(&leaf->header);
1653         data_end = leaf_data_end(root, leaf);
1654
1655         if (btrfs_leaf_free_space(root, leaf) <
1656             sizeof(struct btrfs_item) + data_size) {
1657                 BUG();
1658         }
1659         slot = path->slots[0];
1660         BUG_ON(slot < 0);
1661         if (slot != nritems) {
1662                 int i;
1663                 unsigned int old_data = btrfs_item_end(leaf->items + slot);
1664
1665                 /*
1666                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
1667                  */
1668                 /* first correct the data pointers */
1669                 for (i = slot; i < nritems; i++) {
1670                         u32 ioff = btrfs_item_offset(leaf->items + i);
1671                         btrfs_set_item_offset(leaf->items + i,
1672                                               ioff - data_size);
1673                 }
1674
1675                 /* shift the items */
1676                 btrfs_memmove(root, leaf, leaf->items + slot + 1,
1677                               leaf->items + slot,
1678                               (nritems - slot) * sizeof(struct btrfs_item));
1679
1680                 /* shift the data */
1681                 btrfs_memmove(root, leaf, btrfs_leaf_data(leaf) +
1682                               data_end - data_size, btrfs_leaf_data(leaf) +
1683                               data_end, old_data - data_end);
1684                 data_end = old_data;
1685         }
1686         /* setup the item for the new data */
1687         btrfs_memcpy(root, leaf, &leaf->items[slot].key, &disk_key,
1688                      sizeof(struct btrfs_disk_key));
1689         btrfs_set_item_offset(leaf->items + slot, data_end - data_size);
1690         btrfs_set_item_size(leaf->items + slot, data_size);
1691         btrfs_set_header_nritems(&leaf->header, nritems + 1);
1692         btrfs_mark_buffer_dirty(leaf_buf);
1693
1694         ret = 0;
1695         if (slot == 0)
1696                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
1697
1698         if (btrfs_leaf_free_space(root, leaf) < 0)
1699                 BUG();
1700         check_leaf(root, path, 0);
1701 out:
1702         return ret;
1703 }
1704
1705 /*
1706  * Given a key and some data, insert an item into the tree.
1707  * This does all the path init required, making room in the tree if needed.
1708  */
1709 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
1710                       *root, struct btrfs_key *cpu_key, void *data, u32
1711                       data_size)
1712 {
1713         int ret = 0;
1714         struct btrfs_path *path;
1715         u8 *ptr;
1716
1717         path = btrfs_alloc_path();
1718         BUG_ON(!path);
1719         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
1720         if (!ret) {
1721                 ptr = btrfs_item_ptr(btrfs_buffer_leaf(path->nodes[0]),
1722                                      path->slots[0], u8);
1723                 btrfs_memcpy(root, path->nodes[0]->b_data,
1724                              ptr, data, data_size);
1725                 btrfs_mark_buffer_dirty(path->nodes[0]);
1726         }
1727         btrfs_free_path(path);
1728         return ret;
1729 }
1730
1731 /*
1732  * delete the pointer from a given node.
1733  *
1734  * If the delete empties a node, the node is removed from the tree,
1735  * continuing all the way the root if required.  The root is converted into
1736  * a leaf if all the nodes are emptied.
1737  */
1738 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1739                    struct btrfs_path *path, int level, int slot)
1740 {
1741         struct btrfs_node *node;
1742         struct buffer_head *parent = path->nodes[level];
1743         u32 nritems;
1744         int ret = 0;
1745         int wret;
1746
1747         node = btrfs_buffer_node(parent);
1748         nritems = btrfs_header_nritems(&node->header);
1749         if (slot != nritems -1) {
1750                 btrfs_memmove(root, node, node->ptrs + slot,
1751                               node->ptrs + slot + 1,
1752                               sizeof(struct btrfs_key_ptr) *
1753                               (nritems - slot - 1));
1754         }
1755         nritems--;
1756         btrfs_set_header_nritems(&node->header, nritems);
1757         if (nritems == 0 && parent == root->node) {
1758                 struct btrfs_header *header = btrfs_buffer_header(root->node);
1759                 BUG_ON(btrfs_header_level(header) != 1);
1760                 /* just turn the root into a leaf and break */
1761                 btrfs_set_header_level(header, 0);
1762         } else if (slot == 0) {
1763                 wret = fixup_low_keys(trans, root, path, &node->ptrs[0].key,
1764                                       level + 1);
1765                 if (wret)
1766                         ret = wret;
1767         }
1768         btrfs_mark_buffer_dirty(parent);
1769         return ret;
1770 }
1771
1772 /*
1773  * delete the item at the leaf level in path.  If that empties
1774  * the leaf, remove it from the tree
1775  */
1776 int btrfs_del_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1777                    struct btrfs_path *path)
1778 {
1779         int slot;
1780         struct btrfs_leaf *leaf;
1781         struct buffer_head *leaf_buf;
1782         int doff;
1783         int dsize;
1784         int ret = 0;
1785         int wret;
1786         u32 nritems;
1787
1788         leaf_buf = path->nodes[0];
1789         leaf = btrfs_buffer_leaf(leaf_buf);
1790         slot = path->slots[0];
1791         doff = btrfs_item_offset(leaf->items + slot);
1792         dsize = btrfs_item_size(leaf->items + slot);
1793         nritems = btrfs_header_nritems(&leaf->header);
1794
1795         if (slot != nritems - 1) {
1796                 int i;
1797                 int data_end = leaf_data_end(root, leaf);
1798                 btrfs_memmove(root, leaf, btrfs_leaf_data(leaf) +
1799                               data_end + dsize,
1800                               btrfs_leaf_data(leaf) + data_end,
1801                               doff - data_end);
1802                 for (i = slot + 1; i < nritems; i++) {
1803                         u32 ioff = btrfs_item_offset(leaf->items + i);
1804                         btrfs_set_item_offset(leaf->items + i, ioff + dsize);
1805                 }
1806                 btrfs_memmove(root, leaf, leaf->items + slot,
1807                               leaf->items + slot + 1,
1808                               sizeof(struct btrfs_item) *
1809                               (nritems - slot - 1));
1810         }
1811         btrfs_set_header_nritems(&leaf->header, nritems - 1);
1812         nritems--;
1813         /* delete the leaf if we've emptied it */
1814         if (nritems == 0) {
1815                 if (leaf_buf == root->node) {
1816                         btrfs_set_header_level(&leaf->header, 0);
1817                 } else {
1818                         clean_tree_block(trans, root, leaf_buf);
1819                         wait_on_buffer(leaf_buf);
1820                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
1821                         if (wret)
1822                                 ret = wret;
1823                         wret = btrfs_free_extent(trans, root,
1824                                                  bh_blocknr(leaf_buf), 1, 1);
1825                         if (wret)
1826                                 ret = wret;
1827                 }
1828         } else {
1829                 int used = leaf_space_used(leaf, 0, nritems);
1830                 if (slot == 0) {
1831                         wret = fixup_low_keys(trans, root, path,
1832                                               &leaf->items[0].key, 1);
1833                         if (wret)
1834                                 ret = wret;
1835                 }
1836
1837                 /* delete the leaf if it is mostly empty */
1838                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 3) {
1839                         /* push_leaf_left fixes the path.
1840                          * make sure the path still points to our leaf
1841                          * for possible call to del_ptr below
1842                          */
1843                         slot = path->slots[1];
1844                         get_bh(leaf_buf);
1845                         wret = push_leaf_left(trans, root, path, 1);
1846                         if (wret < 0 && wret != -ENOSPC)
1847                                 ret = wret;
1848                         if (path->nodes[0] == leaf_buf &&
1849                             btrfs_header_nritems(&leaf->header)) {
1850                                 wret = push_leaf_right(trans, root, path, 1);
1851                                 if (wret < 0 && wret != -ENOSPC)
1852                                         ret = wret;
1853                         }
1854                         if (btrfs_header_nritems(&leaf->header) == 0) {
1855                                 u64 blocknr = bh_blocknr(leaf_buf);
1856                                 clean_tree_block(trans, root, leaf_buf);
1857                                 wait_on_buffer(leaf_buf);
1858                                 wret = del_ptr(trans, root, path, 1, slot);
1859                                 if (wret)
1860                                         ret = wret;
1861                                 btrfs_block_release(root, leaf_buf);
1862                                 wret = btrfs_free_extent(trans, root, blocknr,
1863                                                          1, 1);
1864                                 if (wret)
1865                                         ret = wret;
1866                         } else {
1867                                 btrfs_mark_buffer_dirty(leaf_buf);
1868                                 btrfs_block_release(root, leaf_buf);
1869                         }
1870                 } else {
1871                         btrfs_mark_buffer_dirty(leaf_buf);
1872                 }
1873         }
1874         return ret;
1875 }
1876
1877 /*
1878  * walk up the tree as far as required to find the next leaf.
1879  * returns 0 if it found something or 1 if there are no greater leaves.
1880  * returns < 0 on io errors.
1881  */
1882 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
1883 {
1884         int slot;
1885         int level = 1;
1886         u64 blocknr;
1887         struct buffer_head *c;
1888         struct btrfs_node *c_node;
1889         struct buffer_head *next = NULL;
1890
1891         while(level < BTRFS_MAX_LEVEL) {
1892                 if (!path->nodes[level])
1893                         return 1;
1894                 slot = path->slots[level] + 1;
1895                 c = path->nodes[level];
1896                 c_node = btrfs_buffer_node(c);
1897                 if (slot >= btrfs_header_nritems(&c_node->header)) {
1898                         level++;
1899                         continue;
1900                 }
1901                 blocknr = btrfs_node_blockptr(c_node, slot);
1902                 if (next)
1903                         btrfs_block_release(root, next);
1904                 next = read_tree_block(root, blocknr);
1905                 break;
1906         }
1907         path->slots[level] = slot;
1908         while(1) {
1909                 level--;
1910                 c = path->nodes[level];
1911                 btrfs_block_release(root, c);
1912                 path->nodes[level] = next;
1913                 path->slots[level] = 0;
1914                 if (!level)
1915                         break;
1916                 next = read_tree_block(root,
1917                        btrfs_node_blockptr(btrfs_buffer_node(next), 0));
1918         }
1919         return 0;
1920 }