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
 176         slot = path->slots[level];
 177         BUG_ON(nritems == 0);
 178         if (parent) {
 179                 struct btrfs_disk_key *parent_key;
 180
 181                 parent_slot = path->slots[level + 1];
 182                 parent_key = &parent->ptrs[parent_slot].key;
 183                 BUG_ON(memcmp(parent_key, &node->ptrs[0].key,
 184                               sizeof(struct btrfs_disk_key)));
 185                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 186                        btrfs_header_blocknr(&node->header));
 187         }
 188         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 189         if (slot != 0) {
 190                 btrfs_disk_key_to_cpu(&cpukey, &node->ptrs[slot - 1].key);
 191                 BUG_ON(comp_keys(&node->ptrs[slot].key, &cpukey) <= 0);
 192         }
 193         if (slot < nritems - 1) {
 194                 btrfs_disk_key_to_cpu(&cpukey, &node->ptrs[slot + 1].key);
 195                 BUG_ON(comp_keys(&node->ptrs[slot].key, &cpukey) >= 0);
 196         }
 197         return 0;
 198 }
 199
 200 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 201                       int level)
 202 {
 203         struct btrfs_leaf *leaf = btrfs_buffer_leaf(path->nodes[level]);
 204         struct btrfs_node *parent = NULL;
 205         int parent_slot;
 206         int slot = path->slots[0];
 207         struct btrfs_key cpukey;
 208
 209         u32 nritems = btrfs_header_nritems(&leaf->header);
 210
 211         if (path->nodes[level + 1])
 212                 parent = btrfs_buffer_node(path->nodes[level + 1]);
 213
 214         BUG_ON(btrfs_leaf_free_space(root, leaf) < 0);
 215
 216         if (nritems == 0)
 217                 return 0;
 218
 219         if (parent) {
 220                 struct btrfs_disk_key *parent_key;
 221
 222                 parent_slot = path->slots[level + 1];
 223                 parent_key = &parent->ptrs[parent_slot].key;
 224                 BUG_ON(memcmp(parent_key, &leaf->items[0].key,
 225                        sizeof(struct btrfs_disk_key)));
 226                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 227                        btrfs_header_blocknr(&leaf->header));
 228         }
 229         if (slot != 0) {
 230                 btrfs_disk_key_to_cpu(&cpukey, &leaf->items[slot - 1].key);
 231                 BUG_ON(comp_keys(&leaf->items[slot].key, &cpukey) <= 0);
 232                 BUG_ON(btrfs_item_offset(leaf->items + slot - 1) !=
 233                         btrfs_item_end(leaf->items + slot));
 234         }
 235         if (slot < nritems - 1) {
 236                 btrfs_disk_key_to_cpu(&cpukey, &leaf->items[slot + 1].key);
 237                 BUG_ON(comp_keys(&leaf->items[slot].key, &cpukey) >= 0);
 238                 BUG_ON(btrfs_item_offset(leaf->items + slot) !=
 239                         btrfs_item_end(leaf->items + slot + 1));
 240         }
 241         BUG_ON(btrfs_item_offset(leaf->items) +
 242                btrfs_item_size(leaf->items) != BTRFS_LEAF_DATA_SIZE(root));
 243         return 0;
 244 }
 245
 246 static int check_block(struct btrfs_root *root, struct btrfs_path *path,
 247                         int level)
 248 {
 249         struct btrfs_node *node = btrfs_buffer_node(path->nodes[level]);
 250         if (memcmp(node->header.fsid, root->fs_info->disk_super->fsid,
 251                    sizeof(node->header.fsid)))
 252                 BUG();
 253         if (level == 0)
 254                 return check_leaf(root, path, level);
 255         return check_node(root, path, level);
 256 }
 257
 258 /*
 259  * search for key in the array p.  items p are item_size apart
 260  * and there are 'max' items in p
 261  * the slot in the array is returned via slot, and it points to
 262  * the place where you would insert key if it is not found in
 263  * the array.
 264  *
 265  * slot may point to max if the key is bigger than all of the keys
 266  */
 267 static int generic_bin_search(char *p, int item_size, struct btrfs_key *key,
 268                        int max, int *slot)
 269 {
 270         int low = 0;
 271         int high = max;
 272         int mid;
 273         int ret;
 274         struct btrfs_disk_key *tmp;
 275
 276         while(low < high) {
 277                 mid = (low + high) / 2;
 278                 tmp = (struct btrfs_disk_key *)(p + mid * item_size);
 279                 ret = comp_keys(tmp, key);
 280
 281                 if (ret < 0)
 282                         low = mid + 1;
 283                 else if (ret > 0)
 284                         high = mid;
 285                 else {
 286                         *slot = mid;
 287                         return 0;
 288                 }
 289         }
 290         *slot = low;
 291         return 1;
 292 }
 293
 294 /*
 295  * simple bin_search frontend that does the right thing for
 296  * leaves vs nodes
 297  */
 298 static int bin_search(struct btrfs_node *c, struct btrfs_key *key, int *slot)
 299 {
 300         if (btrfs_is_leaf(c)) {
 301                 struct btrfs_leaf *l = (struct btrfs_leaf *)c;
 302                 return generic_bin_search((void *)l->items,
 303                                           sizeof(struct btrfs_item),
 304                                           key, btrfs_header_nritems(&c->header),
 305                                           slot);
 306         } else {
 307                 return generic_bin_search((void *)c->ptrs,
 308                                           sizeof(struct btrfs_key_ptr),
 309                                           key, btrfs_header_nritems(&c->header),
 310                                           slot);
 311         }
 312         return -1;
 313 }
 314
 315 static struct buffer_head *read_node_slot(struct btrfs_root *root,
 316                                    struct buffer_head *parent_buf,
 317                                    int slot)
 318 {
 319         struct btrfs_node *node = btrfs_buffer_node(parent_buf);
 320         if (slot < 0)
 321                 return NULL;
 322         if (slot >= btrfs_header_nritems(&node->header))
 323                 return NULL;
 324         return read_tree_block(root, btrfs_node_blockptr(node, slot));
 325 }
 326
 327 static int balance_level(struct btrfs_trans_handle *trans, struct btrfs_root
 328                          *root, struct btrfs_path *path, int level)
 329 {
 330         struct buffer_head *right_buf;
 331         struct buffer_head *mid_buf;
 332         struct buffer_head *left_buf;
 333         struct buffer_head *parent_buf = NULL;
 334         struct btrfs_node *right = NULL;
 335         struct btrfs_node *mid;
 336         struct btrfs_node *left = NULL;
 337         struct btrfs_node *parent = NULL;
 338         int ret = 0;
 339         int wret;
 340         int pslot;
 341         int orig_slot = path->slots[level];
 342         int err_on_enospc = 0;
 343         u64 orig_ptr;
 344
 345         if (level == 0)
 346                 return 0;
 347
 348         mid_buf = path->nodes[level];
 349         mid = btrfs_buffer_node(mid_buf);
 350         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 351
 352         if (level < BTRFS_MAX_LEVEL - 1)
 353                 parent_buf = path->nodes[level + 1];
 354         pslot = path->slots[level + 1];
 355
 356         /*
 357          * deal with the case where there is only one pointer in the root
 358          * by promoting the node below to a root
 359          */
 360         if (!parent_buf) {
 361                 struct buffer_head *child;
 362                 u64 blocknr = bh_blocknr(mid_buf);
 363
 364                 if (btrfs_header_nritems(&mid->header) != 1)
 365                         return 0;
 366
 367                 /* promote the child to a root */
 368                 child = read_node_slot(root, mid_buf, 0);
 369                 BUG_ON(!child);
 370                 root->node = child;
 371                 path->nodes[level] = NULL;
 372                 clean_tree_block(trans, root, mid_buf);
 373                 wait_on_buffer(mid_buf);
 374                 /* once for the path */
 375                 btrfs_block_release(root, mid_buf);
 376                 /* once for the root ptr */
 377                 btrfs_block_release(root, mid_buf);
 378                 return btrfs_free_extent(trans, root, blocknr, 1, 1);
 379         }
 380         parent = btrfs_buffer_node(parent_buf);
 381
 382         if (btrfs_header_nritems(&mid->header) >
 383             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 384                 return 0;
 385
 386         if (btrfs_header_nritems(&mid->header) < 2)
 387                 err_on_enospc = 1;
 388
 389         left_buf = read_node_slot(root, parent_buf, pslot - 1);
 390         right_buf = read_node_slot(root, parent_buf, pslot + 1);
 391
 392         /* first, try to make some room in the middle buffer */
 393         if (left_buf) {
 394                 wret = btrfs_cow_block(trans, root, left_buf,
 395                                        parent_buf, pslot - 1, &left_buf);
 396                 if (wret) {
 397                         ret = wret;
 398                         goto enospc;
 399                 }
 400                 left = btrfs_buffer_node(left_buf);
 401                 orig_slot += btrfs_header_nritems(&left->header);
 402                 wret = push_node_left(trans, root, left_buf, mid_buf);
 403                 if (wret < 0)
 404                         ret = wret;
 405                 if (btrfs_header_nritems(&mid->header) < 2)
 406                         err_on_enospc = 1;
 407         }
 408
 409         /*
 410          * then try to empty the right most buffer into the middle
 411          */
 412         if (right_buf) {
 413                 wret = btrfs_cow_block(trans, root, right_buf,
 414                                        parent_buf, pslot + 1, &right_buf);
 415                 if (wret) {
 416                         ret = wret;
 417                         goto enospc;
 418                 }
 419
 420                 right = btrfs_buffer_node(right_buf);
 421                 wret = push_node_left(trans, root, mid_buf, right_buf);
 422                 if (wret < 0 && wret != -ENOSPC)
 423                         ret = wret;
 424                 if (btrfs_header_nritems(&right->header) == 0) {
 425                         u64 blocknr = bh_blocknr(right_buf);
 426                         clean_tree_block(trans, root, right_buf);
 427                         wait_on_buffer(right_buf);
 428                         btrfs_block_release(root, right_buf);
 429                         right_buf = NULL;
 430                         right = NULL;
 431                         wret = del_ptr(trans, root, path, level + 1, pslot +
 432                                        1);
 433                         if (wret)
 434                                 ret = wret;
 435                         wret = btrfs_free_extent(trans, root, blocknr, 1, 1);
 436                         if (wret)
 437                                 ret = wret;
 438                 } else {
 439                         btrfs_memcpy(root, parent,
 440                                      &parent->ptrs[pslot + 1].key,
 441                                      &right->ptrs[0].key,
 442                                      sizeof(struct btrfs_disk_key));
 443                         btrfs_mark_buffer_dirty(parent_buf);
 444                 }
 445         }
 446         if (btrfs_header_nritems(&mid->header) == 1) {
 447                 /*
 448                  * we're not allowed to leave a node with one item in the
 449                  * tree during a delete.  A deletion from lower in the tree
 450                  * could try to delete the only pointer in this node.
 451                  * So, pull some keys from the left.
 452                  * There has to be a left pointer at this point because
 453                  * otherwise we would have pulled some pointers from the
 454                  * right
 455                  */
 456                 BUG_ON(!left_buf);
 457                 wret = balance_node_right(trans, root, mid_buf, left_buf);
 458                 if (wret < 0) {
 459                         ret = wret;
 460                         goto enospc;
 461                 }
 462                 BUG_ON(wret == 1);
 463         }
 464         if (btrfs_header_nritems(&mid->header) == 0) {
 465                 /* we've managed to empty the middle node, drop it */
 466                 u64 blocknr = bh_blocknr(mid_buf);
 467                 clean_tree_block(trans, root, mid_buf);
 468                 wait_on_buffer(mid_buf);
 469                 btrfs_block_release(root, mid_buf);
 470                 mid_buf = NULL;
 471                 mid = NULL;
 472                 wret = del_ptr(trans, root, path, level + 1, pslot);
 473                 if (wret)
 474                         ret = wret;
 475                 wret = btrfs_free_extent(trans, root, blocknr, 1, 1);
 476                 if (wret)
 477                         ret = wret;
 478         } else {
 479                 /* update the parent key to reflect our changes */
 480                 btrfs_memcpy(root, parent,
 481                              &parent->ptrs[pslot].key, &mid->ptrs[0].key,
 482                              sizeof(struct btrfs_disk_key));
 483                 btrfs_mark_buffer_dirty(parent_buf);
 484         }
 485
 486         /* update the path */
 487         if (left_buf) {
 488                 if (btrfs_header_nritems(&left->header) > orig_slot) {
 489                         get_bh(left_buf);
 490                         path->nodes[level] = left_buf;
 491                         path->slots[level + 1] -= 1;
 492                         path->slots[level] = orig_slot;
 493                         if (mid_buf)
 494                                 btrfs_block_release(root, mid_buf);
 495                 } else {
 496                         orig_slot -= btrfs_header_nritems(&left->header);
 497                         path->slots[level] = orig_slot;
 498                 }
 499         }
 500         /* double check we haven't messed things up */
 501         check_block(root, path, level);
 502         if (orig_ptr !=
 503             btrfs_node_blockptr(btrfs_buffer_node(path->nodes[level]),
 504                                 path->slots[level]))
 505                 BUG();
 506 enospc:
 507         if (right_buf)
 508                 btrfs_block_release(root, right_buf);
 509         if (left_buf)
 510                 btrfs_block_release(root, left_buf);
 511         return ret;
 512 }
 513
 514 /* returns zero if the push worked, non-zero otherwise */
 515 static int push_nodes_for_insert(struct btrfs_trans_handle *trans,
 516                                 struct btrfs_root *root,
 517                                 struct btrfs_path *path, int level)
 518 {
 519         struct buffer_head *right_buf;
 520         struct buffer_head *mid_buf;
 521         struct buffer_head *left_buf;
 522         struct buffer_head *parent_buf = NULL;
 523         struct btrfs_node *right = NULL;
 524         struct btrfs_node *mid;
 525         struct btrfs_node *left = NULL;
 526         struct btrfs_node *parent = NULL;
 527         int ret = 0;
 528         int wret;
 529         int pslot;
 530         int orig_slot = path->slots[level];
 531         u64 orig_ptr;
 532
 533         if (level == 0)
 534                 return 1;
 535
 536         mid_buf = path->nodes[level];
 537         mid = btrfs_buffer_node(mid_buf);
 538         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 539
 540         if (level < BTRFS_MAX_LEVEL - 1)
 541                 parent_buf = path->nodes[level + 1];
 542         pslot = path->slots[level + 1];
 543
 544         if (!parent_buf)
 545                 return 1;
 546         parent = btrfs_buffer_node(parent_buf);
 547
 548         left_buf = read_node_slot(root, parent_buf, pslot - 1);
 549
 550         /* first, try to make some room in the middle buffer */
 551         if (left_buf) {
 552                 u32 left_nr;
 553                 left = btrfs_buffer_node(left_buf);
 554                 left_nr = btrfs_header_nritems(&left->header);
 555                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 556                         wret = 1;
 557                 } else {
 558                         ret = btrfs_cow_block(trans, root, left_buf, parent_buf,
 559                                               pslot - 1, &left_buf);
 560                         if (ret)
 561                                 wret = 1;
 562                         else {
 563                                 left = btrfs_buffer_node(left_buf);
 564                                 wret = push_node_left(trans, root,
 565                                                       left_buf, mid_buf);
 566                         }
 567                 }
 568                 if (wret < 0)
 569                         ret = wret;
 570                 if (wret == 0) {
 571                         orig_slot += left_nr;
 572                         btrfs_memcpy(root, parent,
 573                                      &parent->ptrs[pslot].key,
 574                                      &mid->ptrs[0].key,
 575                                      sizeof(struct btrfs_disk_key));
 576                         btrfs_mark_buffer_dirty(parent_buf);
 577                         if (btrfs_header_nritems(&left->header) > orig_slot) {
 578                                 path->nodes[level] = left_buf;
 579                                 path->slots[level + 1] -= 1;
 580                                 path->slots[level] = orig_slot;
 581                                 btrfs_block_release(root, mid_buf);
 582                         } else {
 583                                 orig_slot -=
 584                                         btrfs_header_nritems(&left->header);
 585                                 path->slots[level] = orig_slot;
 586                                 btrfs_block_release(root, left_buf);
 587                         }
 588                         check_node(root, path, level);
 589                         return 0;
 590                 }
 591                 btrfs_block_release(root, left_buf);
 592         }
 593         right_buf = read_node_slot(root, parent_buf, pslot + 1);
 594
 595         /*
 596          * then try to empty the right most buffer into the middle
 597          */
 598         if (right_buf) {
 599                 u32 right_nr;
 600                 right = btrfs_buffer_node(right_buf);
 601                 right_nr = btrfs_header_nritems(&right->header);
 602                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 603                         wret = 1;
 604                 } else {
 605                         ret = btrfs_cow_block(trans, root, right_buf,
 606                                               parent_buf, pslot + 1,
 607                                               &right_buf);
 608                         if (ret)
 609                                 wret = 1;
 610                         else {
 611                                 right = btrfs_buffer_node(right_buf);
 612                                 wret = balance_node_right(trans, root,
 613                                                           right_buf, mid_buf);
 614                         }
 615                 }
 616                 if (wret < 0)
 617                         ret = wret;
 618                 if (wret == 0) {
 619                         btrfs_memcpy(root, parent,
 620                                      &parent->ptrs[pslot + 1].key,
 621                                      &right->ptrs[0].key,
 622                                      sizeof(struct btrfs_disk_key));
 623                         btrfs_mark_buffer_dirty(parent_buf);
 624                         if (btrfs_header_nritems(&mid->header) <= orig_slot) {
 625                                 path->nodes[level] = right_buf;
 626                                 path->slots[level + 1] += 1;
 627                                 path->slots[level] = orig_slot -
 628                                         btrfs_header_nritems(&mid->header);
 629                                 btrfs_block_release(root, mid_buf);
 630                         } else {
 631                                 btrfs_block_release(root, right_buf);
 632                         }
 633                         check_node(root, path, level);
 634                         return 0;
 635                 }
 636                 btrfs_block_release(root, right_buf);
 637         }
 638         check_node(root, path, level);
 639         return 1;
 640 }
 641
 642 /*
 643  * readahead one full node of leaves
 644  */
 645 static void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
 646                              int slot)
 647 {
 648         struct btrfs_node *node;
 649         int i;
 650         u32 nritems;
 651         u64 item_objectid;
 652         u64 blocknr;
 653         u64 search;
 654         u64 cluster_start;
 655         int ret;
 656         int nread = 0;
 657         int direction = path->reada;
 658         struct radix_tree_root found;
 659         unsigned long gang[8];
 660         struct buffer_head *bh;
 661
 662         if (!path->nodes[1])
 663                 return;
 664
 665         node = btrfs_buffer_node(path->nodes[1]);
 666         search = btrfs_node_blockptr(node, slot);
 667         bh = btrfs_find_tree_block(root, search);
 668         if (bh) {
 669                 brelse(bh);
 670                 return;
 671         }
 672
 673         init_bit_radix(&found);
 674         nritems = btrfs_header_nritems(&node->header);
 675         for (i = slot; i < nritems; i++) {
 676                 item_objectid = btrfs_disk_key_objectid(&node->ptrs[i].key);
 677                 blocknr = btrfs_node_blockptr(node, i);
 678                 set_radix_bit(&found, blocknr);
 679         }
 680         if (direction > 0) {
 681                 cluster_start = search - 4;
 682                 if (cluster_start > search)
 683                         cluster_start = 0;
 684         } else
 685                 cluster_start = search + 4;
 686         while(1) {
 687                 ret = find_first_radix_bit(&found, gang, 0, ARRAY_SIZE(gang));
 688                 if (!ret)
 689                         break;
 690                 for (i = 0; i < ret; i++) {
 691                         blocknr = gang[i];
 692                         clear_radix_bit(&found, blocknr);
 693                         if (nread > 64)
 694                                 continue;
 695                         if (direction > 0 && cluster_start <= blocknr &&
 696                             cluster_start + 8 > blocknr) {
 697                                 cluster_start = blocknr;
 698                                 readahead_tree_block(root, blocknr);
 699                                 nread++;
 700                         } else if (direction < 0 && cluster_start >= blocknr &&
 701                                    blocknr + 8 > cluster_start) {
 702                                 cluster_start = blocknr;
 703                                 readahead_tree_block(root, blocknr);
 704                                 nread++;
 705                         }
 706                 }
 707         }
 708 }
 709 /*
 710  * look for key in the tree.  path is filled in with nodes along the way
 711  * if key is found, we return zero and you can find the item in the leaf
 712  * level of the path (level 0)
 713  *
 714  * If the key isn't found, the path points to the slot where it should
 715  * be inserted, and 1 is returned.  If there are other errors during the
 716  * search a negative error number is returned.
 717  *
 718  * if ins_len > 0, nodes and leaves will be split as we walk down the
 719  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
 720  * possible)
 721  */
 722 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
 723                       *root, struct btrfs_key *key, struct btrfs_path *p, int
 724                       ins_len, int cow)
 725 {
 726         struct buffer_head *b;
 727         struct buffer_head *cow_buf;
 728         struct btrfs_node *c;
 729         struct btrfs_root_item *root_item = &root->root_item;
 730         u64 blocknr;
 731         int slot;
 732         int ret;
 733         int level;
 734         int should_reada = p->reada;
 735         u8 lowest_level = 0;
 736
 737         if (btrfs_root_refs(root_item) == 0 && root->ref_cows) {
 738                 lowest_level = root_item->drop_level;
 739                 WARN_ON(ins_len || cow);
 740         }
 741
 742         WARN_ON(p->nodes[0] != NULL);
 743         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
 744 again:
 745         b = root->node;
 746         get_bh(b);
 747         while (b) {
 748                 c = btrfs_buffer_node(b);
 749                 level = btrfs_header_level(&c->header);
 750                 if (cow) {
 751                         int wret;
 752                         wret = btrfs_cow_block(trans, root, b,
 753                                                p->nodes[level + 1],
 754                                                p->slots[level + 1],
 755                                                &cow_buf);
 756                         if (wret) {
 757                                 btrfs_block_release(root, cow_buf);
 758                                 return wret;
 759                         }
 760                         b = cow_buf;
 761                         c = btrfs_buffer_node(b);
 762                 }
 763                 BUG_ON(!cow && ins_len);
 764                 if (level != btrfs_header_level(&c->header))
 765                         WARN_ON(1);
 766                 level = btrfs_header_level(&c->header);
 767                 p->nodes[level] = b;
 768                 ret = check_block(root, p, level);
 769                 if (ret)
 770                         return -1;
 771                 ret = bin_search(c, key, &slot);
 772                 if (!btrfs_is_leaf(c)) {
 773                         if (ret && slot > 0)
 774                                 slot -= 1;
 775                         p->slots[level] = slot;
 776                         if (ins_len > 0 && btrfs_header_nritems(&c->header) >=
 777                             BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 778                                 int sret = split_node(trans, root, p, level);
 779                                 BUG_ON(sret > 0);
 780                                 if (sret)
 781                                         return sret;
 782                                 b = p->nodes[level];
 783                                 c = btrfs_buffer_node(b);
 784                                 slot = p->slots[level];
 785                         } else if (ins_len < 0) {
 786                                 int sret = balance_level(trans, root, p,
 787                                                          level);
 788                                 if (sret)
 789                                         return sret;
 790                                 b = p->nodes[level];
 791                                 if (!b)
 792                                         goto again;
 793                                 c = btrfs_buffer_node(b);
 794                                 slot = p->slots[level];
 795                                 BUG_ON(btrfs_header_nritems(&c->header) == 1);
 796                         }
 797                         /* this is only true while dropping a snapshot */
 798                         if (level == lowest_level)
 799                                 break;
 800                         blocknr = btrfs_node_blockptr(c, slot);
 801                         if (level == 1 && should_reada)
 802                                 reada_for_search(root, p, slot);
 803                         b = read_tree_block(root, btrfs_node_blockptr(c, slot));
 804
 805                 } else {
 806                         struct btrfs_leaf *l = (struct btrfs_leaf *)c;
 807                         p->slots[level] = slot;
 808                         if (ins_len > 0 && btrfs_leaf_free_space(root, l) <
 809                             sizeof(struct btrfs_item) + ins_len) {
 810                                 int sret = split_leaf(trans, root, key,
 811                                                       p, ins_len);
 812                                 BUG_ON(sret > 0);
 813                                 if (sret)
 814                                         return sret;
 815                         }
 816                         return ret;
 817                 }
 818         }
 819         return 1;
 820 }
 821
 822 /*
 823  * adjust the pointers going up the tree, starting at level
 824  * making sure the right key of each node is points to 'key'.
 825  * This is used after shifting pointers to the left, so it stops
 826  * fixing up pointers when a given leaf/node is not in slot 0 of the
 827  * higher levels
 828  *
 829  * If this fails to write a tree block, it returns -1, but continues
 830  * fixing up the blocks in ram so the tree is consistent.
 831  */
 832 static int fixup_low_keys(struct btrfs_trans_handle *trans, struct btrfs_root
 833                           *root, struct btrfs_path *path, struct btrfs_disk_key
 834                           *key, int level)
 835 {
 836         int i;
 837         int ret = 0;
 838         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
 839                 struct btrfs_node *t;
 840                 int tslot = path->slots[i];
 841                 if (!path->nodes[i])
 842                         break;
 843                 t = btrfs_buffer_node(path->nodes[i]);
 844                 btrfs_memcpy(root, t, &t->ptrs[tslot].key, key, sizeof(*key));
 845                 btrfs_mark_buffer_dirty(path->nodes[i]);
 846                 if (tslot != 0)
 847                         break;
 848         }
 849         return ret;
 850 }
 851
 852 /*
 853  * try to push data from one node into the next node left in the
 854  * tree.
 855  *
 856  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
 857  * error, and > 0 if there was no room in the left hand block.
 858  */
 859 static int push_node_left(struct btrfs_trans_handle *trans, struct btrfs_root
 860                           *root, struct buffer_head *dst_buf, struct
 861                           buffer_head *src_buf)
 862 {
 863         struct btrfs_node *src = btrfs_buffer_node(src_buf);
 864         struct btrfs_node *dst = btrfs_buffer_node(dst_buf);
 865         int push_items = 0;
 866         int src_nritems;
 867         int dst_nritems;
 868         int ret = 0;
 869
 870         src_nritems = btrfs_header_nritems(&src->header);
 871         dst_nritems = btrfs_header_nritems(&dst->header);
 872         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
 873
 874         if (push_items <= 0) {
 875                 return 1;
 876         }
 877
 878         if (src_nritems < push_items)
 879                 push_items = src_nritems;
 880
 881         btrfs_memcpy(root, dst, dst->ptrs + dst_nritems, src->ptrs,
 882                      push_items * sizeof(struct btrfs_key_ptr));
 883         if (push_items < src_nritems) {
 884                 btrfs_memmove(root, src, src->ptrs, src->ptrs + push_items,
 885                         (src_nritems - push_items) *
 886                         sizeof(struct btrfs_key_ptr));
 887         }
 888         btrfs_set_header_nritems(&src->header, src_nritems - push_items);
 889         btrfs_set_header_nritems(&dst->header, dst_nritems + push_items);
 890         btrfs_mark_buffer_dirty(src_buf);
 891         btrfs_mark_buffer_dirty(dst_buf);
 892         return ret;
 893 }
 894
 895 /*
 896  * try to push data from one node into the next node right in the
 897  * tree.
 898  *
 899  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
 900  * error, and > 0 if there was no room in the right hand block.
 901  *
 902  * this will  only push up to 1/2 the contents of the left node over
 903  */
 904 static int balance_node_right(struct btrfs_trans_handle *trans, struct
 905                               btrfs_root *root, struct buffer_head *dst_buf,
 906                               struct buffer_head *src_buf)
 907 {
 908         struct btrfs_node *src = btrfs_buffer_node(src_buf);
 909         struct btrfs_node *dst = btrfs_buffer_node(dst_buf);
 910         int push_items = 0;
 911         int max_push;
 912         int src_nritems;
 913         int dst_nritems;
 914         int ret = 0;
 915
 916         src_nritems = btrfs_header_nritems(&src->header);
 917         dst_nritems = btrfs_header_nritems(&dst->header);
 918         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
 919         if (push_items <= 0) {
 920                 return 1;
 921         }
 922
 923         max_push = src_nritems / 2 + 1;
 924         /* don't try to empty the node */
 925         if (max_push > src_nritems)
 926                 return 1;
 927         if (max_push < push_items)
 928                 push_items = max_push;
 929
 930         btrfs_memmove(root, dst, dst->ptrs + push_items, dst->ptrs,
 931                       dst_nritems * sizeof(struct btrfs_key_ptr));
 932
 933         btrfs_memcpy(root, dst, dst->ptrs,
 934                      src->ptrs + src_nritems - push_items,
 935                      push_items * sizeof(struct btrfs_key_ptr));
 936
 937         btrfs_set_header_nritems(&src->header, src_nritems - push_items);
 938         btrfs_set_header_nritems(&dst->header, dst_nritems + push_items);
 939
 940         btrfs_mark_buffer_dirty(src_buf);
 941         btrfs_mark_buffer_dirty(dst_buf);
 942         return ret;
 943 }
 944
 945 /*
 946  * helper function to insert a new root level in the tree.
 947  * A new node is allocated, and a single item is inserted to
 948  * point to the existing root
 949  *
 950  * returns zero on success or < 0 on failure.
 951  */
 952 static int insert_new_root(struct btrfs_trans_handle *trans, struct btrfs_root
 953                            *root, struct btrfs_path *path, int level)
 954 {
 955         struct buffer_head *t;
 956         struct btrfs_node *lower;
 957         struct btrfs_node *c;
 958         struct btrfs_disk_key *lower_key;
 959
 960         BUG_ON(path->nodes[level]);
 961         BUG_ON(path->nodes[level-1] != root->node);
 962
 963         t = btrfs_alloc_free_block(trans, root, root->node->b_blocknr);
 964         if (IS_ERR(t))
 965                 return PTR_ERR(t);
 966         c = btrfs_buffer_node(t);
 967         memset(c, 0, root->blocksize);
 968         btrfs_set_header_nritems(&c->header, 1);
 969         btrfs_set_header_level(&c->header, level);
 970         btrfs_set_header_blocknr(&c->header, bh_blocknr(t));
 971         btrfs_set_header_generation(&c->header, trans->transid);
 972         btrfs_set_header_owner(&c->header, root->root_key.objectid);
 973         lower = btrfs_buffer_node(path->nodes[level-1]);
 974         memcpy(c->header.fsid, root->fs_info->disk_super->fsid,
 975                sizeof(c->header.fsid));
 976         if (btrfs_is_leaf(lower))
 977                 lower_key = &((struct btrfs_leaf *)lower)->items[0].key;
 978         else
 979                 lower_key = &lower->ptrs[0].key;
 980         btrfs_memcpy(root, c, &c->ptrs[0].key, lower_key,
 981                      sizeof(struct btrfs_disk_key));
 982         btrfs_set_node_blockptr(c, 0, bh_blocknr(path->nodes[level - 1]));
 983
 984         btrfs_mark_buffer_dirty(t);
 985
 986         /* the super has an extra ref to root->node */
 987         btrfs_block_release(root, root->node);
 988         root->node = t;
 989         get_bh(t);
 990         path->nodes[level] = t;
 991         path->slots[level] = 0;
 992         return 0;
 993 }
 994
 995 /*
 996  * worker function to insert a single pointer in a node.
 997  * the node should have enough room for the pointer already
 998  *
 999  * slot and level indicate where you want the key to go, and
1000  * blocknr is the block the key points to.
1001  *
1002  * returns zero on success and < 0 on any error
1003  */
1004 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1005                       *root, struct btrfs_path *path, struct btrfs_disk_key
1006                       *key, u64 blocknr, int slot, int level)
1007 {
1008         struct btrfs_node *lower;
1009         int nritems;
1010
1011         BUG_ON(!path->nodes[level]);
1012         lower = btrfs_buffer_node(path->nodes[level]);
1013         nritems = btrfs_header_nritems(&lower->header);
1014         if (slot > nritems)
1015                 BUG();
1016         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1017                 BUG();
1018         if (slot != nritems) {
1019                 btrfs_memmove(root, lower, lower->ptrs + slot + 1,
1020                               lower->ptrs + slot,
1021                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1022         }
1023         btrfs_memcpy(root, lower, &lower->ptrs[slot].key,
1024                      key, sizeof(struct btrfs_disk_key));
1025         btrfs_set_node_blockptr(lower, slot, blocknr);
1026         btrfs_set_header_nritems(&lower->header, nritems + 1);
1027         btrfs_mark_buffer_dirty(path->nodes[level]);
1028         check_node(root, path, level);
1029         return 0;
1030 }
1031
1032 /*
1033  * split the node at the specified level in path in two.
1034  * The path is corrected to point to the appropriate node after the split
1035  *
1036  * Before splitting this tries to make some room in the node by pushing
1037  * left and right, if either one works, it returns right away.
1038  *
1039  * returns 0 on success and < 0 on failure
1040  */
1041 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1042                       *root, struct btrfs_path *path, int level)
1043 {
1044         struct buffer_head *t;
1045         struct btrfs_node *c;
1046         struct buffer_head *split_buffer;
1047         struct btrfs_node *split;
1048         int mid;
1049         int ret;
1050         int wret;
1051         u32 c_nritems;
1052
1053         t = path->nodes[level];
1054         c = btrfs_buffer_node(t);
1055         if (t == root->node) {
1056                 /* trying to split the root, lets make a new one */
1057                 ret = insert_new_root(trans, root, path, level + 1);
1058                 if (ret)
1059                         return ret;
1060         } else {
1061                 ret = push_nodes_for_insert(trans, root, path, level);
1062                 t = path->nodes[level];
1063                 c = btrfs_buffer_node(t);
1064                 if (!ret &&
1065                     btrfs_header_nritems(&c->header) <
1066                     BTRFS_NODEPTRS_PER_BLOCK(root) - 1)
1067                         return 0;
1068                 if (ret < 0)
1069                         return ret;
1070         }
1071
1072         c_nritems = btrfs_header_nritems(&c->header);
1073         split_buffer = btrfs_alloc_free_block(trans, root, t->b_blocknr);
1074         if (IS_ERR(split_buffer))
1075                 return PTR_ERR(split_buffer);
1076
1077         split = btrfs_buffer_node(split_buffer);
1078         btrfs_set_header_flags(&split->header, btrfs_header_flags(&c->header));
1079         btrfs_set_header_level(&split->header, btrfs_header_level(&c->header));
1080         btrfs_set_header_blocknr(&split->header, bh_blocknr(split_buffer));
1081         btrfs_set_header_generation(&split->header, trans->transid);
1082         btrfs_set_header_owner(&split->header, root->root_key.objectid);
1083         memcpy(split->header.fsid, root->fs_info->disk_super->fsid,
1084                sizeof(split->header.fsid));
1085         mid = (c_nritems + 1) / 2;
1086         btrfs_memcpy(root, split, split->ptrs, c->ptrs + mid,
1087                      (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1088         btrfs_set_header_nritems(&split->header, c_nritems - mid);
1089         btrfs_set_header_nritems(&c->header, mid);
1090         ret = 0;
1091
1092         btrfs_mark_buffer_dirty(t);
1093         btrfs_mark_buffer_dirty(split_buffer);
1094         wret = insert_ptr(trans, root, path, &split->ptrs[0].key,
1095                           bh_blocknr(split_buffer), path->slots[level + 1] + 1,
1096                           level + 1);
1097         if (wret)
1098                 ret = wret;
1099
1100         if (path->slots[level] >= mid) {
1101                 path->slots[level] -= mid;
1102                 btrfs_block_release(root, t);
1103                 path->nodes[level] = split_buffer;
1104                 path->slots[level + 1] += 1;
1105         } else {
1106                 btrfs_block_release(root, split_buffer);
1107         }
1108         return ret;
1109 }
1110
1111 /*
1112  * how many bytes are required to store the items in a leaf.  start
1113  * and nr indicate which items in the leaf to check.  This totals up the
1114  * space used both by the item structs and the item data
1115  */
1116 static int leaf_space_used(struct btrfs_leaf *l, int start, int nr)
1117 {
1118         int data_len;
1119         int nritems = btrfs_header_nritems(&l->header);
1120         int end = min(nritems, start + nr) - 1;
1121
1122         if (!nr)
1123                 return 0;
1124         data_len = btrfs_item_end(l->items + start);
1125         data_len = data_len - btrfs_item_offset(l->items + end);
1126         data_len += sizeof(struct btrfs_item) * nr;
1127         WARN_ON(data_len < 0);
1128         return data_len;
1129 }
1130
1131 /*
1132  * The space between the end of the leaf items and
1133  * the start of the leaf data.  IOW, how much room
1134  * the leaf has left for both items and data
1135  */
1136 int btrfs_leaf_free_space(struct btrfs_root *root, struct btrfs_leaf *leaf)
1137 {
1138         int nritems = btrfs_header_nritems(&leaf->header);
1139         return BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1140 }
1141
1142 /*
1143  * push some data in the path leaf to the right, trying to free up at
1144  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1145  *
1146  * returns 1 if the push failed because the other node didn't have enough
1147  * room, 0 if everything worked out and < 0 if there were major errors.
1148  */
1149 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1150                            *root, struct btrfs_path *path, int data_size)
1151 {
1152         struct buffer_head *left_buf = path->nodes[0];
1153         struct btrfs_leaf *left = btrfs_buffer_leaf(left_buf);
1154         struct btrfs_leaf *right;
1155         struct buffer_head *right_buf;
1156         struct buffer_head *upper;
1157         struct btrfs_node *upper_node;
1158         int slot;
1159         int i;
1160         int free_space;
1161         int push_space = 0;
1162         int push_items = 0;
1163         struct btrfs_item *item;
1164         u32 left_nritems;
1165         u32 right_nritems;
1166         int ret;
1167
1168         slot = path->slots[1];
1169         if (!path->nodes[1]) {
1170                 return 1;
1171         }
1172         upper = path->nodes[1];
1173         upper_node = btrfs_buffer_node(upper);
1174         if (slot >= btrfs_header_nritems(&upper_node->header) - 1) {
1175                 return 1;
1176         }
1177         right_buf = read_tree_block(root,
1178                     btrfs_node_blockptr(btrfs_buffer_node(upper), slot + 1));
1179         right = btrfs_buffer_leaf(right_buf);
1180         free_space = btrfs_leaf_free_space(root, right);
1181         if (free_space < data_size + sizeof(struct btrfs_item)) {
1182                 btrfs_block_release(root, right_buf);
1183                 return 1;
1184         }
1185         /* cow and double check */
1186         ret = btrfs_cow_block(trans, root, right_buf, upper,
1187                               slot + 1, &right_buf);
1188         if (ret) {
1189                 btrfs_block_release(root, right_buf);
1190                 return 1;
1191         }
1192         right = btrfs_buffer_leaf(right_buf);
1193         free_space = btrfs_leaf_free_space(root, right);
1194         if (free_space < data_size + sizeof(struct btrfs_item)) {
1195                 btrfs_block_release(root, right_buf);
1196                 return 1;
1197         }
1198
1199         left_nritems = btrfs_header_nritems(&left->header);
1200         if (left_nritems == 0) {
1201                 btrfs_block_release(root, right_buf);
1202                 return 1;
1203         }
1204         for (i = left_nritems - 1; i >= 1; i--) {
1205                 item = left->items + i;
1206                 if (path->slots[0] == i)
1207                         push_space += data_size + sizeof(*item);
1208                 if (btrfs_item_size(item) + sizeof(*item) + push_space >
1209                     free_space)
1210                         break;
1211                 push_items++;
1212                 push_space += btrfs_item_size(item) + sizeof(*item);
1213         }
1214         if (push_items == 0) {
1215                 btrfs_block_release(root, right_buf);
1216                 return 1;
1217         }
1218         if (push_items == left_nritems)
1219                 WARN_ON(1);
1220         right_nritems = btrfs_header_nritems(&right->header);
1221         /* push left to right */
1222         push_space = btrfs_item_end(left->items + left_nritems - push_items);
1223         push_space -= leaf_data_end(root, left);
1224         /* make room in the right data area */
1225         btrfs_memmove(root, right, btrfs_leaf_data(right) +
1226                       leaf_data_end(root, right) - push_space,
1227                       btrfs_leaf_data(right) +
1228                       leaf_data_end(root, right), BTRFS_LEAF_DATA_SIZE(root) -
1229                       leaf_data_end(root, right));
1230         /* copy from the left data area */
1231         btrfs_memcpy(root, right, btrfs_leaf_data(right) +
1232                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1233                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1234                      push_space);
1235         btrfs_memmove(root, right, right->items + push_items, right->items,
1236                 right_nritems * sizeof(struct btrfs_item));
1237         /* copy the items from left to right */
1238         btrfs_memcpy(root, right, right->items, left->items +
1239                      left_nritems - push_items,
1240                      push_items * sizeof(struct btrfs_item));
1241
1242         /* update the item pointers */
1243         right_nritems += push_items;
1244         btrfs_set_header_nritems(&right->header, right_nritems);
1245         push_space = BTRFS_LEAF_DATA_SIZE(root);
1246         for (i = 0; i < right_nritems; i++) {
1247                 btrfs_set_item_offset(right->items + i, push_space -
1248                                       btrfs_item_size(right->items + i));
1249                 push_space = btrfs_item_offset(right->items + i);
1250         }
1251         left_nritems -= push_items;
1252         btrfs_set_header_nritems(&left->header, left_nritems);
1253
1254         btrfs_mark_buffer_dirty(left_buf);
1255         btrfs_mark_buffer_dirty(right_buf);
1256
1257         btrfs_memcpy(root, upper_node, &upper_node->ptrs[slot + 1].key,
1258                 &right->items[0].key, sizeof(struct btrfs_disk_key));
1259         btrfs_mark_buffer_dirty(upper);
1260
1261         /* then fixup the leaf pointer in the path */
1262         if (path->slots[0] >= left_nritems) {
1263                 path->slots[0] -= left_nritems;
1264                 btrfs_block_release(root, path->nodes[0]);
1265                 path->nodes[0] = right_buf;
1266                 path->slots[1] += 1;
1267         } else {
1268                 btrfs_block_release(root, right_buf);
1269         }
1270         if (path->nodes[1])
1271                 check_node(root, path, 1);
1272         return 0;
1273 }
1274 /*
1275  * push some data in the path leaf to the left, trying to free up at
1276  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1277  */
1278 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1279                           *root, struct btrfs_path *path, int data_size)
1280 {
1281         struct buffer_head *right_buf = path->nodes[0];
1282         struct btrfs_leaf *right = btrfs_buffer_leaf(right_buf);
1283         struct buffer_head *t;
1284         struct btrfs_leaf *left;
1285         int slot;
1286         int i;
1287         int free_space;
1288         int push_space = 0;
1289         int push_items = 0;
1290         struct btrfs_item *item;
1291         u32 old_left_nritems;
1292         int ret = 0;
1293         int wret;
1294
1295         slot = path->slots[1];
1296         if (slot == 0) {
1297                 return 1;
1298         }
1299         if (!path->nodes[1]) {
1300                 return 1;
1301         }
1302         t = read_tree_block(root,
1303             btrfs_node_blockptr(btrfs_buffer_node(path->nodes[1]), slot - 1));
1304         left = btrfs_buffer_leaf(t);
1305         free_space = btrfs_leaf_free_space(root, left);
1306         if (free_space < data_size + sizeof(struct btrfs_item)) {
1307                 btrfs_block_release(root, t);
1308                 return 1;
1309         }
1310
1311         /* cow and double check */
1312         ret = btrfs_cow_block(trans, root, t, path->nodes[1], slot - 1, &t);
1313         if (ret) {
1314                 /* we hit -ENOSPC, but it isn't fatal here */
1315                 return 1;
1316         }
1317         left = btrfs_buffer_leaf(t);
1318         free_space = btrfs_leaf_free_space(root, left);
1319         if (free_space < data_size + sizeof(struct btrfs_item)) {
1320                 btrfs_block_release(root, t);
1321                 return 1;
1322         }
1323
1324         if (btrfs_header_nritems(&right->header) == 0) {
1325                 btrfs_block_release(root, t);
1326                 return 1;
1327         }
1328
1329         for (i = 0; i < btrfs_header_nritems(&right->header) - 1; i++) {
1330                 item = right->items + i;
1331                 if (path->slots[0] == i)
1332                         push_space += data_size + sizeof(*item);
1333                 if (btrfs_item_size(item) + sizeof(*item) + push_space >
1334                     free_space)
1335                         break;
1336                 push_items++;
1337                 push_space += btrfs_item_size(item) + sizeof(*item);
1338         }
1339         if (push_items == 0) {
1340                 btrfs_block_release(root, t);
1341                 return 1;
1342         }
1343         if (push_items == btrfs_header_nritems(&right->header))
1344                 WARN_ON(1);
1345         /* push data from right to left */
1346         btrfs_memcpy(root, left, left->items +
1347                      btrfs_header_nritems(&left->header),
1348                      right->items, push_items * sizeof(struct btrfs_item));
1349         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1350                      btrfs_item_offset(right->items + push_items -1);
1351         btrfs_memcpy(root, left, btrfs_leaf_data(left) +
1352                      leaf_data_end(root, left) - push_space,
1353                      btrfs_leaf_data(right) +
1354                      btrfs_item_offset(right->items + push_items - 1),
1355                      push_space);
1356         old_left_nritems = btrfs_header_nritems(&left->header);
1357         BUG_ON(old_left_nritems < 0);
1358
1359         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1360                 u32 ioff = btrfs_item_offset(left->items + i);
1361                 btrfs_set_item_offset(left->items + i, ioff -
1362                                      (BTRFS_LEAF_DATA_SIZE(root) -
1363                                       btrfs_item_offset(left->items +
1364                                                         old_left_nritems - 1)));
1365         }
1366         btrfs_set_header_nritems(&left->header, old_left_nritems + push_items);
1367
1368         /* fixup right node */
1369         push_space = btrfs_item_offset(right->items + push_items - 1) -
1370                      leaf_data_end(root, right);
1371         btrfs_memmove(root, right, btrfs_leaf_data(right) +
1372                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1373                       btrfs_leaf_data(right) +
1374                       leaf_data_end(root, right), push_space);
1375         btrfs_memmove(root, right, right->items, right->items + push_items,
1376                 (btrfs_header_nritems(&right->header) - push_items) *
1377                 sizeof(struct btrfs_item));
1378         btrfs_set_header_nritems(&right->header,
1379                                  btrfs_header_nritems(&right->header) -
1380                                  push_items);
1381         push_space = BTRFS_LEAF_DATA_SIZE(root);
1382
1383         for (i = 0; i < btrfs_header_nritems(&right->header); i++) {
1384                 btrfs_set_item_offset(right->items + i, push_space -
1385                                       btrfs_item_size(right->items + i));
1386                 push_space = btrfs_item_offset(right->items + i);
1387         }
1388
1389         btrfs_mark_buffer_dirty(t);
1390         btrfs_mark_buffer_dirty(right_buf);
1391
1392         wret = fixup_low_keys(trans, root, path, &right->items[0].key, 1);
1393         if (wret)
1394                 ret = wret;
1395
1396         /* then fixup the leaf pointer in the path */
1397         if (path->slots[0] < push_items) {
1398                 path->slots[0] += old_left_nritems;
1399                 btrfs_block_release(root, path->nodes[0]);
1400                 path->nodes[0] = t;
1401                 path->slots[1] -= 1;
1402         } else {
1403                 btrfs_block_release(root, t);
1404                 path->slots[0] -= push_items;
1405         }
1406         BUG_ON(path->slots[0] < 0);
1407         if (path->nodes[1])
1408                 check_node(root, path, 1);
1409         return ret;
1410 }
1411
1412 /*
1413  * split the path's leaf in two, making sure there is at least data_size
1414  * available for the resulting leaf level of the path.
1415  *
1416  * returns 0 if all went well and < 0 on failure.
1417  */
1418 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
1419                       *root, struct btrfs_key *ins_key,
1420                       struct btrfs_path *path, int data_size)
1421 {
1422         struct buffer_head *l_buf;
1423         struct btrfs_leaf *l;
1424         u32 nritems;
1425         int mid;
1426         int slot;
1427         struct btrfs_leaf *right;
1428         struct buffer_head *right_buffer;
1429         int space_needed = data_size + sizeof(struct btrfs_item);
1430         int data_copy_size;
1431         int rt_data_off;
1432         int i;
1433         int ret = 0;
1434         int wret;
1435         int double_split = 0;
1436         struct btrfs_disk_key disk_key;
1437
1438         /* first try to make some room by pushing left and right */
1439         wret = push_leaf_left(trans, root, path, data_size);
1440         if (wret < 0)
1441                 return wret;
1442         if (wret) {
1443                 wret = push_leaf_right(trans, root, path, data_size);
1444                 if (wret < 0)
1445                         return wret;
1446         }
1447         l_buf = path->nodes[0];
1448         l = btrfs_buffer_leaf(l_buf);
1449
1450         /* did the pushes work? */
1451         if (btrfs_leaf_free_space(root, l) >=
1452             sizeof(struct btrfs_item) + data_size)
1453                 return 0;
1454
1455         if (!path->nodes[1]) {
1456                 ret = insert_new_root(trans, root, path, 1);
1457                 if (ret)
1458                         return ret;
1459         }
1460         slot = path->slots[0];
1461         nritems = btrfs_header_nritems(&l->header);
1462         mid = (nritems + 1)/ 2;
1463
1464         right_buffer = btrfs_alloc_free_block(trans, root, l_buf->b_blocknr);
1465         if (IS_ERR(right_buffer))
1466                 return PTR_ERR(right_buffer);
1467
1468         right = btrfs_buffer_leaf(right_buffer);
1469         memset(&right->header, 0, sizeof(right->header));
1470         btrfs_set_header_blocknr(&right->header, bh_blocknr(right_buffer));
1471         btrfs_set_header_generation(&right->header, trans->transid);
1472         btrfs_set_header_owner(&right->header, root->root_key.objectid);
1473         btrfs_set_header_level(&right->header, 0);
1474         memcpy(right->header.fsid, root->fs_info->disk_super->fsid,
1475                sizeof(right->header.fsid));
1476         if (mid <= slot) {
1477                 if (nritems == 1 ||
1478                     leaf_space_used(l, mid, nritems - mid) + space_needed >
1479                         BTRFS_LEAF_DATA_SIZE(root)) {
1480                         if (slot >= nritems) {
1481                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1482                                 btrfs_set_header_nritems(&right->header, 0);
1483                                 wret = insert_ptr(trans, root, path,
1484                                                   &disk_key,
1485                                                   bh_blocknr(right_buffer),
1486                                                   path->slots[1] + 1, 1);
1487                                 if (wret)
1488                                         ret = wret;
1489                                 btrfs_block_release(root, path->nodes[0]);
1490                                 path->nodes[0] = right_buffer;
1491                                 path->slots[0] = 0;
1492                                 path->slots[1] += 1;
1493                                 return ret;
1494                         }
1495                         mid = slot;
1496                         double_split = 1;
1497                 }
1498         } else {
1499                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
1500                         BTRFS_LEAF_DATA_SIZE(root)) {
1501                         if (slot == 0) {
1502                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1503                                 btrfs_set_header_nritems(&right->header, 0);
1504                                 wret = insert_ptr(trans, root, path,
1505                                                   &disk_key,
1506                                                   bh_blocknr(right_buffer),
1507                                                   path->slots[1], 1);
1508                                 if (wret)
1509                                         ret = wret;
1510                                 btrfs_block_release(root, path->nodes[0]);
1511                                 path->nodes[0] = right_buffer;
1512                                 path->slots[0] = 0;
1513                                 if (path->slots[1] == 0) {
1514                                         wret = fixup_low_keys(trans, root,
1515                                                    path, &disk_key, 1);
1516                                         if (wret)
1517                                                 ret = wret;
1518                                 }
1519                                 return ret;
1520                         }
1521                         mid = slot;
1522                         double_split = 1;
1523                 }
1524         }
1525         btrfs_set_header_nritems(&right->header, nritems - mid);
1526         data_copy_size = btrfs_item_end(l->items + mid) -
1527                          leaf_data_end(root, l);
1528         btrfs_memcpy(root, right, right->items, l->items + mid,
1529                      (nritems - mid) * sizeof(struct btrfs_item));
1530         btrfs_memcpy(root, right,
1531                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
1532                      data_copy_size, btrfs_leaf_data(l) +
1533                      leaf_data_end(root, l), data_copy_size);
1534         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
1535                       btrfs_item_end(l->items + mid);
1536
1537         for (i = 0; i < btrfs_header_nritems(&right->header); i++) {
1538                 u32 ioff = btrfs_item_offset(right->items + i);
1539                 btrfs_set_item_offset(right->items + i, ioff + rt_data_off);
1540         }
1541
1542         btrfs_set_header_nritems(&l->header, mid);
1543         ret = 0;
1544         wret = insert_ptr(trans, root, path, &right->items[0].key,
1545                           bh_blocknr(right_buffer), path->slots[1] + 1, 1);
1546         if (wret)
1547                 ret = wret;
1548         btrfs_mark_buffer_dirty(right_buffer);
1549         btrfs_mark_buffer_dirty(l_buf);
1550         BUG_ON(path->slots[0] != slot);
1551         if (mid <= slot) {
1552                 btrfs_block_release(root, path->nodes[0]);
1553                 path->nodes[0] = right_buffer;
1554                 path->slots[0] -= mid;
1555                 path->slots[1] += 1;
1556         } else
1557                 btrfs_block_release(root, right_buffer);
1558         BUG_ON(path->slots[0] < 0);
1559         check_node(root, path, 1);
1560
1561         if (!double_split)
1562                 return ret;
1563         right_buffer = btrfs_alloc_free_block(trans, root, l_buf->b_blocknr);
1564         if (IS_ERR(right_buffer))
1565                 return PTR_ERR(right_buffer);
1566
1567         right = btrfs_buffer_leaf(right_buffer);
1568         memset(&right->header, 0, sizeof(right->header));
1569         btrfs_set_header_blocknr(&right->header, bh_blocknr(right_buffer));
1570         btrfs_set_header_generation(&right->header, trans->transid);
1571         btrfs_set_header_owner(&right->header, root->root_key.objectid);
1572         btrfs_set_header_level(&right->header, 0);
1573         memcpy(right->header.fsid, root->fs_info->disk_super->fsid,
1574                sizeof(right->header.fsid));
1575         btrfs_cpu_key_to_disk(&disk_key, ins_key);
1576         btrfs_set_header_nritems(&right->header, 0);
1577         wret = insert_ptr(trans, root, path,
1578                           &disk_key,
1579                           bh_blocknr(right_buffer),
1580                           path->slots[1], 1);
1581         if (wret)
1582                 ret = wret;
1583         if (path->slots[1] == 0) {
1584                 wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1585                 if (wret)
1586                         ret = wret;
1587         }
1588         btrfs_block_release(root, path->nodes[0]);
1589         path->nodes[0] = right_buffer;
1590         path->slots[0] = 0;
1591         check_node(root, path, 1);
1592         check_leaf(root, path, 0);
1593         return ret;
1594 }
1595
1596 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
1597                         struct btrfs_root *root,
1598                         struct btrfs_path *path,
1599                         u32 new_size)
1600 {
1601         int ret = 0;
1602         int slot;
1603         int slot_orig;
1604         struct btrfs_leaf *leaf;
1605         struct buffer_head *leaf_buf;
1606         u32 nritems;
1607         unsigned int data_end;
1608         unsigned int old_data_start;
1609         unsigned int old_size;
1610         unsigned int size_diff;
1611         int i;
1612
1613         slot_orig = path->slots[0];
1614         leaf_buf = path->nodes[0];
1615         leaf = btrfs_buffer_leaf(leaf_buf);
1616
1617         nritems = btrfs_header_nritems(&leaf->header);
1618         data_end = leaf_data_end(root, leaf);
1619
1620         slot = path->slots[0];
1621         old_data_start = btrfs_item_offset(leaf->items + slot);
1622         old_size = btrfs_item_size(leaf->items + slot);
1623         BUG_ON(old_size <= new_size);
1624         size_diff = old_size - new_size;
1625
1626         BUG_ON(slot < 0);
1627         BUG_ON(slot >= nritems);
1628
1629         /*
1630          * item0..itemN ... dataN.offset..dataN.size .. data0.size
1631          */
1632         /* first correct the data pointers */
1633         for (i = slot; i < nritems; i++) {
1634                 u32 ioff = btrfs_item_offset(leaf->items + i);
1635                 btrfs_set_item_offset(leaf->items + i,
1636                                       ioff + size_diff);
1637         }
1638         /* shift the data */
1639         btrfs_memmove(root, leaf, btrfs_leaf_data(leaf) +
1640                       data_end + size_diff, btrfs_leaf_data(leaf) +
1641                       data_end, old_data_start + new_size - data_end);
1642         btrfs_set_item_size(leaf->items + slot, new_size);
1643         btrfs_mark_buffer_dirty(leaf_buf);
1644
1645         ret = 0;
1646         if (btrfs_leaf_free_space(root, leaf) < 0)
1647                 BUG();
1648         check_leaf(root, path, 0);
1649         return ret;
1650 }
1651
1652 int btrfs_extend_item(struct btrfs_trans_handle *trans, struct btrfs_root
1653                       *root, struct btrfs_path *path, u32 data_size)
1654 {
1655         int ret = 0;
1656         int slot;
1657         int slot_orig;
1658         struct btrfs_leaf *leaf;
1659         struct buffer_head *leaf_buf;
1660         u32 nritems;
1661         unsigned int data_end;
1662         unsigned int old_data;
1663         unsigned int old_size;
1664         int i;
1665
1666         slot_orig = path->slots[0];
1667         leaf_buf = path->nodes[0];
1668         leaf = btrfs_buffer_leaf(leaf_buf);
1669
1670         nritems = btrfs_header_nritems(&leaf->header);
1671         data_end = leaf_data_end(root, leaf);
1672
1673         if (btrfs_leaf_free_space(root, leaf) < data_size)
1674                 BUG();
1675         slot = path->slots[0];
1676         old_data = btrfs_item_end(leaf->items + slot);
1677
1678         BUG_ON(slot < 0);
1679         BUG_ON(slot >= nritems);
1680
1681         /*
1682          * item0..itemN ... dataN.offset..dataN.size .. data0.size
1683          */
1684         /* first correct the data pointers */
1685         for (i = slot; i < nritems; i++) {
1686                 u32 ioff = btrfs_item_offset(leaf->items + i);
1687                 btrfs_set_item_offset(leaf->items + i,
1688                                       ioff - data_size);
1689         }
1690         /* shift the data */
1691         btrfs_memmove(root, leaf, btrfs_leaf_data(leaf) +
1692                       data_end - data_size, btrfs_leaf_data(leaf) +
1693                       data_end, old_data - data_end);
1694         data_end = old_data;
1695         old_size = btrfs_item_size(leaf->items + slot);
1696         btrfs_set_item_size(leaf->items + slot, old_size + data_size);
1697         btrfs_mark_buffer_dirty(leaf_buf);
1698
1699         ret = 0;
1700         if (btrfs_leaf_free_space(root, leaf) < 0)
1701                 BUG();
1702         check_leaf(root, path, 0);
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_empty_item(struct btrfs_trans_handle *trans, struct btrfs_root
1711                             *root, struct btrfs_path *path, struct btrfs_key
1712                             *cpu_key, u32 data_size)
1713 {
1714         int ret = 0;
1715         int slot;
1716         int slot_orig;
1717         struct btrfs_leaf *leaf;
1718         struct buffer_head *leaf_buf;
1719         u32 nritems;
1720         unsigned int data_end;
1721         struct btrfs_disk_key disk_key;
1722
1723         btrfs_cpu_key_to_disk(&disk_key, cpu_key);
1724
1725         /* create a root if there isn't one */
1726         if (!root->node)
1727                 BUG();
1728         ret = btrfs_search_slot(trans, root, cpu_key, path, data_size, 1);
1729         if (ret == 0) {
1730                 return -EEXIST;
1731         }
1732         if (ret < 0)
1733                 goto out;
1734
1735         slot_orig = path->slots[0];
1736         leaf_buf = path->nodes[0];
1737         leaf = btrfs_buffer_leaf(leaf_buf);
1738
1739         nritems = btrfs_header_nritems(&leaf->header);
1740         data_end = leaf_data_end(root, leaf);
1741
1742         if (btrfs_leaf_free_space(root, leaf) <
1743             sizeof(struct btrfs_item) + data_size) {
1744                 BUG();
1745         }
1746         slot = path->slots[0];
1747         BUG_ON(slot < 0);
1748         if (slot != nritems) {
1749                 int i;
1750                 unsigned int old_data = btrfs_item_end(leaf->items + slot);
1751
1752                 /*
1753                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
1754                  */
1755                 /* first correct the data pointers */
1756                 for (i = slot; i < nritems; i++) {
1757                         u32 ioff = btrfs_item_offset(leaf->items + i);
1758                         btrfs_set_item_offset(leaf->items + i,
1759                                               ioff - data_size);
1760                 }
1761
1762                 /* shift the items */
1763                 btrfs_memmove(root, leaf, leaf->items + slot + 1,
1764                               leaf->items + slot,
1765                               (nritems - slot) * sizeof(struct btrfs_item));
1766
1767                 /* shift the data */
1768                 btrfs_memmove(root, leaf, btrfs_leaf_data(leaf) +
1769                               data_end - data_size, btrfs_leaf_data(leaf) +
1770                               data_end, old_data - data_end);
1771                 data_end = old_data;
1772         }
1773         /* setup the item for the new data */
1774         btrfs_memcpy(root, leaf, &leaf->items[slot].key, &disk_key,
1775                      sizeof(struct btrfs_disk_key));
1776         btrfs_set_item_offset(leaf->items + slot, data_end - data_size);
1777         btrfs_set_item_size(leaf->items + slot, data_size);
1778         btrfs_set_header_nritems(&leaf->header, nritems + 1);
1779         btrfs_mark_buffer_dirty(leaf_buf);
1780
1781         ret = 0;
1782         if (slot == 0)
1783                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
1784
1785         if (btrfs_leaf_free_space(root, leaf) < 0)
1786                 BUG();
1787         check_leaf(root, path, 0);
1788 out:
1789         return ret;
1790 }
1791
1792 /*
1793  * Given a key and some data, insert an item into the tree.
1794  * This does all the path init required, making room in the tree if needed.
1795  */
1796 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
1797                       *root, struct btrfs_key *cpu_key, void *data, u32
1798                       data_size)
1799 {
1800         int ret = 0;
1801         struct btrfs_path *path;
1802         u8 *ptr;
1803
1804         path = btrfs_alloc_path();
1805         BUG_ON(!path);
1806         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
1807         if (!ret) {
1808                 ptr = btrfs_item_ptr(btrfs_buffer_leaf(path->nodes[0]),
1809                                      path->slots[0], u8);
1810                 btrfs_memcpy(root, path->nodes[0]->b_data,
1811                              ptr, data, data_size);
1812                 btrfs_mark_buffer_dirty(path->nodes[0]);
1813         }
1814         btrfs_free_path(path);
1815         return ret;
1816 }
1817
1818 /*
1819  * delete the pointer from a given node.
1820  *
1821  * If the delete empties a node, the node is removed from the tree,
1822  * continuing all the way the root if required.  The root is converted into
1823  * a leaf if all the nodes are emptied.
1824  */
1825 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1826                    struct btrfs_path *path, int level, int slot)
1827 {
1828         struct btrfs_node *node;
1829         struct buffer_head *parent = path->nodes[level];
1830         u32 nritems;
1831         int ret = 0;
1832         int wret;
1833
1834         node = btrfs_buffer_node(parent);
1835         nritems = btrfs_header_nritems(&node->header);
1836         if (slot != nritems -1) {
1837                 btrfs_memmove(root, node, node->ptrs + slot,
1838                               node->ptrs + slot + 1,
1839                               sizeof(struct btrfs_key_ptr) *
1840                               (nritems - slot - 1));
1841         }
1842         nritems--;
1843         btrfs_set_header_nritems(&node->header, nritems);
1844         if (nritems == 0 && parent == root->node) {
1845                 struct btrfs_header *header = btrfs_buffer_header(root->node);
1846                 BUG_ON(btrfs_header_level(header) != 1);
1847                 /* just turn the root into a leaf and break */
1848                 btrfs_set_header_level(header, 0);
1849         } else if (slot == 0) {
1850                 wret = fixup_low_keys(trans, root, path, &node->ptrs[0].key,
1851                                       level + 1);
1852                 if (wret)
1853                         ret = wret;
1854         }
1855         btrfs_mark_buffer_dirty(parent);
1856         return ret;
1857 }
1858
1859 /*
1860  * delete the item at the leaf level in path.  If that empties
1861  * the leaf, remove it from the tree
1862  */
1863 int btrfs_del_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1864                    struct btrfs_path *path)
1865 {
1866         int slot;
1867         struct btrfs_leaf *leaf;
1868         struct buffer_head *leaf_buf;
1869         int doff;
1870         int dsize;
1871         int ret = 0;
1872         int wret;
1873         u32 nritems;
1874
1875         leaf_buf = path->nodes[0];
1876         leaf = btrfs_buffer_leaf(leaf_buf);
1877         slot = path->slots[0];
1878         doff = btrfs_item_offset(leaf->items + slot);
1879         dsize = btrfs_item_size(leaf->items + slot);
1880         nritems = btrfs_header_nritems(&leaf->header);
1881
1882         if (slot != nritems - 1) {
1883                 int i;
1884                 int data_end = leaf_data_end(root, leaf);
1885                 btrfs_memmove(root, leaf, btrfs_leaf_data(leaf) +
1886                               data_end + dsize,
1887                               btrfs_leaf_data(leaf) + data_end,
1888                               doff - data_end);
1889                 for (i = slot + 1; i < nritems; i++) {
1890                         u32 ioff = btrfs_item_offset(leaf->items + i);
1891                         btrfs_set_item_offset(leaf->items + i, ioff + dsize);
1892                 }
1893                 btrfs_memmove(root, leaf, leaf->items + slot,
1894                               leaf->items + slot + 1,
1895                               sizeof(struct btrfs_item) *
1896                               (nritems - slot - 1));
1897         }
1898         btrfs_set_header_nritems(&leaf->header, nritems - 1);
1899         nritems--;
1900         /* delete the leaf if we've emptied it */
1901         if (nritems == 0) {
1902                 if (leaf_buf == root->node) {
1903                         btrfs_set_header_level(&leaf->header, 0);
1904                 } else {
1905                         clean_tree_block(trans, root, leaf_buf);
1906                         wait_on_buffer(leaf_buf);
1907                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
1908                         if (wret)
1909                                 ret = wret;
1910                         wret = btrfs_free_extent(trans, root,
1911                                                  bh_blocknr(leaf_buf), 1, 1);
1912                         if (wret)
1913                                 ret = wret;
1914                 }
1915         } else {
1916                 int used = leaf_space_used(leaf, 0, nritems);
1917                 if (slot == 0) {
1918                         wret = fixup_low_keys(trans, root, path,
1919                                               &leaf->items[0].key, 1);
1920                         if (wret)
1921                                 ret = wret;
1922                 }
1923
1924                 /* delete the leaf if it is mostly empty */
1925                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 3) {
1926                         /* push_leaf_left fixes the path.
1927                          * make sure the path still points to our leaf
1928                          * for possible call to del_ptr below
1929                          */
1930                         slot = path->slots[1];
1931                         get_bh(leaf_buf);
1932                         wret = push_leaf_left(trans, root, path, 1);
1933                         if (wret < 0 && wret != -ENOSPC)
1934                                 ret = wret;
1935                         if (path->nodes[0] == leaf_buf &&
1936                             btrfs_header_nritems(&leaf->header)) {
1937                                 wret = push_leaf_right(trans, root, path, 1);
1938                                 if (wret < 0 && wret != -ENOSPC)
1939                                         ret = wret;
1940                         }
1941                         if (btrfs_header_nritems(&leaf->header) == 0) {
1942                                 u64 blocknr = bh_blocknr(leaf_buf);
1943                                 clean_tree_block(trans, root, leaf_buf);
1944                                 wait_on_buffer(leaf_buf);
1945                                 wret = del_ptr(trans, root, path, 1, slot);
1946                                 if (wret)
1947                                         ret = wret;
1948                                 btrfs_block_release(root, leaf_buf);
1949                                 wret = btrfs_free_extent(trans, root, blocknr,
1950                                                          1, 1);
1951                                 if (wret)
1952                                         ret = wret;
1953                         } else {
1954                                 btrfs_mark_buffer_dirty(leaf_buf);
1955                                 btrfs_block_release(root, leaf_buf);
1956                         }
1957                 } else {
1958                         btrfs_mark_buffer_dirty(leaf_buf);
1959                 }
1960         }
1961         return ret;
1962 }
1963
1964 /*
1965  * walk up the tree as far as required to find the next leaf.
1966  * returns 0 if it found something or 1 if there are no greater leaves.
1967  * returns < 0 on io errors.
1968  */
1969 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
1970 {
1971         int slot;
1972         int level = 1;
1973         u64 blocknr;
1974         struct buffer_head *c;
1975         struct btrfs_node *c_node;
1976         struct buffer_head *next = NULL;
1977
1978         while(level < BTRFS_MAX_LEVEL) {
1979                 if (!path->nodes[level])
1980                         return 1;
1981                 slot = path->slots[level] + 1;
1982                 c = path->nodes[level];
1983                 c_node = btrfs_buffer_node(c);
1984                 if (slot >= btrfs_header_nritems(&c_node->header)) {
1985                         level++;
1986                         continue;
1987                 }
1988                 blocknr = btrfs_node_blockptr(c_node, slot);
1989                 if (next)
1990                         btrfs_block_release(root, next);
1991                 if (level == 1 && path->reada)
1992                         reada_for_search(root, path, slot);
1993                 next = read_tree_block(root, blocknr);
1994                 break;
1995         }
1996         path->slots[level] = slot;
1997         while(1) {
1998                 level--;
1999                 c = path->nodes[level];
2000                 btrfs_block_release(root, c);
2001                 path->nodes[level] = next;
2002                 path->slots[level] = 0;
2003                 if (!level)
2004                         break;
2005                 if (level == 1 && path->reada)
2006                         reada_for_search(root, path, slot);
2007                 next = read_tree_block(root,
2008                        btrfs_node_blockptr(btrfs_buffer_node(next), 0));
2009         }
2010         return 0;
2011 }