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

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/sched.h>
  20 #include "ctree.h"
  21 #include "disk-io.h"
  22 #include "transaction.h"
  23 #include "print-tree.h"
  24
  25 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  26                       *root, struct btrfs_path *path, int level);
  27 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  28                       *root, struct btrfs_key *ins_key,
  29                       struct btrfs_path *path, int data_size, int extend);
  30 static int push_node_left(struct btrfs_trans_handle *trans,
  31                           struct btrfs_root *root, struct extent_buffer *dst,
  32                           struct extent_buffer *src);
  33 static int balance_node_right(struct btrfs_trans_handle *trans,
  34                               struct btrfs_root *root,
  35                               struct extent_buffer *dst_buf,
  36                               struct extent_buffer *src_buf);
  37 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  38                    struct btrfs_path *path, int level, int slot);
  39
  40 inline void btrfs_init_path(struct btrfs_path *p)
  41 {
  42         memset(p, 0, sizeof(*p));
  43 }
  44
  45 struct btrfs_path *btrfs_alloc_path(void)
  46 {
  47         struct btrfs_path *path;
  48         path = kmem_cache_alloc(btrfs_path_cachep, GFP_NOFS);
  49         if (path) {
  50                 btrfs_init_path(path);
  51                 path->reada = 1;
  52         }
  53         return path;
  54 }
  55
  56 void btrfs_free_path(struct btrfs_path *p)
  57 {
  58         btrfs_release_path(NULL, p);
  59         kmem_cache_free(btrfs_path_cachep, p);
  60 }
  61
  62 void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
  63 {
  64         int i;
  65         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  66                 if (!p->nodes[i])
  67                         break;
  68                 free_extent_buffer(p->nodes[i]);
  69         }
  70         memset(p, 0, sizeof(*p));
  71 }
  72
  73 static int __btrfs_cow_block(struct btrfs_trans_handle *trans,
  74                              struct btrfs_root *root,
  75                              struct extent_buffer *buf,
  76                              struct extent_buffer *parent, int parent_slot,
  77                              struct extent_buffer **cow_ret,
  78                              u64 search_start, u64 empty_size)
  79 {
  80         struct extent_buffer *cow;
  81         int ret = 0;
  82         int different_trans = 0;
  83
  84         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
  85
  86         cow = btrfs_alloc_free_block(trans, root, buf->len,
  87                                      search_start, empty_size);
  88         if (IS_ERR(cow))
  89                 return PTR_ERR(cow);
  90
  91         copy_extent_buffer(cow, buf, 0, 0, cow->len);
  92         btrfs_set_header_bytenr(cow, cow->start);
  93         btrfs_set_header_generation(cow, trans->transid);
  94         btrfs_set_header_owner(cow, root->root_key.objectid);
  95
  96         WARN_ON(btrfs_header_generation(buf) > trans->transid);
  97         if (btrfs_header_generation(buf) != trans->transid) {
  98                 different_trans = 1;
  99                 ret = btrfs_inc_ref(trans, root, buf);
 100                 if (ret)
 101                         return ret;
 102         } else {
 103                 clean_tree_block(trans, root, buf);
 104         }
 105
 106         if (buf == root->node) {
 107                 root->node = cow;
 108                 extent_buffer_get(cow);
 109                 if (buf != root->commit_root) {
 110                         btrfs_free_extent(trans, root, buf->start,
 111                                           buf->len, 1);
 112                 }
 113                 free_extent_buffer(buf);
 114         } else {
 115                 btrfs_set_node_blockptr(parent, parent_slot,
 116                                         cow->start);
 117                 btrfs_mark_buffer_dirty(parent);
 118                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 119                 btrfs_free_extent(trans, root, buf->start, buf->len, 1);
 120         }
 121         free_extent_buffer(buf);
 122         btrfs_mark_buffer_dirty(cow);
 123         *cow_ret = cow;
 124         return 0;
 125 }
 126
 127 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 128                     struct btrfs_root *root, struct extent_buffer *buf,
 129                     struct extent_buffer *parent, int parent_slot,
 130                     struct extent_buffer **cow_ret)
 131 {
 132         u64 search_start;
 133         int ret;
 134         if (trans->transaction != root->fs_info->running_transaction) {
 135                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 136                        root->fs_info->running_transaction->transid);
 137                 WARN_ON(1);
 138         }
 139         if (trans->transid != root->fs_info->generation) {
 140                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 141                        root->fs_info->generation);
 142                 WARN_ON(1);
 143         }
 144         if (btrfs_header_generation(buf) == trans->transid) {
 145                 *cow_ret = buf;
 146                 return 0;
 147         }
 148
 149         search_start = buf->start & ~((u64)BTRFS_BLOCK_GROUP_SIZE - 1);
 150         ret = __btrfs_cow_block(trans, root, buf, parent,
 151                                  parent_slot, cow_ret, search_start, 0);
 152         return ret;
 153 }
 154
 155 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 156 {
 157         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 158                 return 1;
 159         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 160                 return 1;
 161         return 0;
 162 }
 163
 164 /*
 165  * compare two keys in a memcmp fashion
 166  */
 167 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 168 {
 169         struct btrfs_key k1;
 170
 171         btrfs_disk_key_to_cpu(&k1, disk);
 172
 173         if (k1.objectid > k2->objectid)
 174                 return 1;
 175         if (k1.objectid < k2->objectid)
 176                 return -1;
 177         if (k1.type > k2->type)
 178                 return 1;
 179         if (k1.type < k2->type)
 180                 return -1;
 181         if (k1.offset > k2->offset)
 182                 return 1;
 183         if (k1.offset < k2->offset)
 184                 return -1;
 185         return 0;
 186 }
 187
 188
 189 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 190                        struct btrfs_root *root, struct extent_buffer *parent,
 191                        int start_slot, int cache_only, u64 *last_ret,
 192                        struct btrfs_key *progress)
 193 {
 194         struct extent_buffer *cur;
 195         struct extent_buffer *tmp;
 196         u64 blocknr;
 197         u64 search_start = *last_ret;
 198         u64 last_block = 0;
 199         u64 other;
 200         u32 parent_nritems;
 201         int end_slot;
 202         int i;
 203         int err = 0;
 204         int parent_level;
 205         int uptodate;
 206         u32 blocksize;
 207         int progress_passed = 0;
 208         struct btrfs_disk_key disk_key;
 209
 210         parent_level = btrfs_header_level(parent);
 211         if (cache_only && parent_level != 1)
 212                 return 0;
 213
 214         if (trans->transaction != root->fs_info->running_transaction) {
 215                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 216                        root->fs_info->running_transaction->transid);
 217                 WARN_ON(1);
 218         }
 219         if (trans->transid != root->fs_info->generation) {
 220                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 221                        root->fs_info->generation);
 222                 WARN_ON(1);
 223         }
 224
 225         parent_nritems = btrfs_header_nritems(parent);
 226         blocksize = btrfs_level_size(root, parent_level - 1);
 227         end_slot = parent_nritems;
 228
 229         if (parent_nritems == 1)
 230                 return 0;
 231
 232         for (i = start_slot; i < end_slot; i++) {
 233                 int close = 1;
 234
 235                 if (!parent->map_token) {
 236                         map_extent_buffer(parent,
 237                                         btrfs_node_key_ptr_offset(i),
 238                                         sizeof(struct btrfs_key_ptr),
 239                                         &parent->map_token, &parent->kaddr,
 240                                         &parent->map_start, &parent->map_len,
 241                                         KM_USER1);
 242                 }
 243                 btrfs_node_key(parent, &disk_key, i);
 244                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
 245                         continue;
 246
 247                 progress_passed = 1;
 248                 blocknr = btrfs_node_blockptr(parent, i);
 249                 if (last_block == 0)
 250                         last_block = blocknr;
 251
 252                 if (i > 0) {
 253                         other = btrfs_node_blockptr(parent, i - 1);
 254                         close = close_blocks(blocknr, other, blocksize);
 255                 }
 256                 if (close && i < end_slot - 2) {
 257                         other = btrfs_node_blockptr(parent, i + 1);
 258                         close = close_blocks(blocknr, other, blocksize);
 259                 }
 260                 if (close) {
 261                         last_block = blocknr;
 262                         continue;
 263                 }
 264                 if (parent->map_token) {
 265                         unmap_extent_buffer(parent, parent->map_token,
 266                                             KM_USER1);
 267                         parent->map_token = NULL;
 268                 }
 269
 270                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
 271                 if (cur)
 272                         uptodate = btrfs_buffer_uptodate(cur);
 273                 else
 274                         uptodate = 0;
 275                 if (!cur || !uptodate) {
 276                         if (cache_only) {
 277                                 free_extent_buffer(cur);
 278                                 continue;
 279                         }
 280                         if (!cur) {
 281                                 cur = read_tree_block(root, blocknr,
 282                                                          blocksize);
 283                         } else if (!uptodate) {
 284                                 btrfs_read_buffer(cur);
 285                         }
 286                 }
 287                 if (search_start == 0)
 288                         search_start = last_block;
 289
 290                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 291                                         &tmp, search_start,
 292                                         min(16 * blocksize,
 293                                             (end_slot - i) * blocksize));
 294                 if (err) {
 295                         free_extent_buffer(cur);
 296                         break;
 297                 }
 298                 search_start = tmp->start;
 299                 last_block = tmp->start;
 300                 *last_ret = search_start;
 301                 if (parent_level == 1)
 302                         btrfs_clear_buffer_defrag(tmp);
 303                 free_extent_buffer(tmp);
 304         }
 305         if (parent->map_token) {
 306                 unmap_extent_buffer(parent, parent->map_token,
 307                                     KM_USER1);
 308                 parent->map_token = NULL;
 309         }
 310         return err;
 311 }
 312
 313 /*
 314  * The leaf data grows from end-to-front in the node.
 315  * this returns the address of the start of the last item,
 316  * which is the stop of the leaf data stack
 317  */
 318 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 319                                          struct extent_buffer *leaf)
 320 {
 321         u32 nr = btrfs_header_nritems(leaf);
 322         if (nr == 0)
 323                 return BTRFS_LEAF_DATA_SIZE(root);
 324         return btrfs_item_offset_nr(leaf, nr - 1);
 325 }
 326
 327 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
 328                       int level)
 329 {
 330         struct extent_buffer *parent = NULL;
 331         struct extent_buffer *node = path->nodes[level];
 332         struct btrfs_disk_key parent_key;
 333         struct btrfs_disk_key node_key;
 334         int parent_slot;
 335         int slot;
 336         struct btrfs_key cpukey;
 337         u32 nritems = btrfs_header_nritems(node);
 338
 339         if (path->nodes[level + 1])
 340                 parent = path->nodes[level + 1];
 341
 342         slot = path->slots[level];
 343         BUG_ON(nritems == 0);
 344         if (parent) {
 345                 parent_slot = path->slots[level + 1];
 346                 btrfs_node_key(parent, &parent_key, parent_slot);
 347                 btrfs_node_key(node, &node_key, 0);
 348                 BUG_ON(memcmp(&parent_key, &node_key,
 349                               sizeof(struct btrfs_disk_key)));
 350                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 351                        btrfs_header_bytenr(node));
 352         }
 353         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 354         if (slot != 0) {
 355                 btrfs_node_key_to_cpu(node, &cpukey, slot - 1);
 356                 btrfs_node_key(node, &node_key, slot);
 357                 BUG_ON(comp_keys(&node_key, &cpukey) <= 0);
 358         }
 359         if (slot < nritems - 1) {
 360                 btrfs_node_key_to_cpu(node, &cpukey, slot + 1);
 361                 btrfs_node_key(node, &node_key, slot);
 362                 BUG_ON(comp_keys(&node_key, &cpukey) >= 0);
 363         }
 364         return 0;
 365 }
 366
 367 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 368                       int level)
 369 {
 370         struct extent_buffer *leaf = path->nodes[level];
 371         struct extent_buffer *parent = NULL;
 372         int parent_slot;
 373         struct btrfs_key cpukey;
 374         struct btrfs_disk_key parent_key;
 375         struct btrfs_disk_key leaf_key;
 376         int slot = path->slots[0];
 377
 378         u32 nritems = btrfs_header_nritems(leaf);
 379
 380         if (path->nodes[level + 1])
 381                 parent = path->nodes[level + 1];
 382
 383         if (nritems == 0)
 384                 return 0;
 385
 386         if (parent) {
 387                 parent_slot = path->slots[level + 1];
 388                 btrfs_node_key(parent, &parent_key, parent_slot);
 389                 btrfs_item_key(leaf, &leaf_key, 0);
 390
 391                 BUG_ON(memcmp(&parent_key, &leaf_key,
 392                        sizeof(struct btrfs_disk_key)));
 393                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 394                        btrfs_header_bytenr(leaf));
 395         }
 396 #if 0
 397         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 398                 btrfs_item_key_to_cpu(leaf, &cpukey, i + 1);
 399                 btrfs_item_key(leaf, &leaf_key, i);
 400                 if (comp_keys(&leaf_key, &cpukey) >= 0) {
 401                         btrfs_print_leaf(root, leaf);
 402                         printk("slot %d offset bad key\n", i);
 403                         BUG_ON(1);
 404                 }
 405                 if (btrfs_item_offset_nr(leaf, i) !=
 406                         btrfs_item_end_nr(leaf, i + 1)) {
 407                         btrfs_print_leaf(root, leaf);
 408                         printk("slot %d offset bad\n", i);
 409                         BUG_ON(1);
 410                 }
 411                 if (i == 0) {
 412                         if (btrfs_item_offset_nr(leaf, i) +
 413                                btrfs_item_size_nr(leaf, i) !=
 414                                BTRFS_LEAF_DATA_SIZE(root)) {
 415                                 btrfs_print_leaf(root, leaf);
 416                                 printk("slot %d first offset bad\n", i);
 417                                 BUG_ON(1);
 418                         }
 419                 }
 420         }
 421         if (nritems > 0) {
 422                 if (btrfs_item_size_nr(leaf, nritems - 1) > 4096) {
 423                                 btrfs_print_leaf(root, leaf);
 424                                 printk("slot %d bad size \n", nritems - 1);
 425                                 BUG_ON(1);
 426                 }
 427         }
 428 #endif
 429         if (slot != 0 && slot < nritems - 1) {
 430                 btrfs_item_key(leaf, &leaf_key, slot);
 431                 btrfs_item_key_to_cpu(leaf, &cpukey, slot - 1);
 432                 if (comp_keys(&leaf_key, &cpukey) <= 0) {
 433                         btrfs_print_leaf(root, leaf);
 434                         printk("slot %d offset bad key\n", slot);
 435                         BUG_ON(1);
 436                 }
 437                 if (btrfs_item_offset_nr(leaf, slot - 1) !=
 438                        btrfs_item_end_nr(leaf, slot)) {
 439                         btrfs_print_leaf(root, leaf);
 440                         printk("slot %d offset bad\n", slot);
 441                         BUG_ON(1);
 442                 }
 443         }
 444         if (slot < nritems - 1) {
 445                 btrfs_item_key(leaf, &leaf_key, slot);
 446                 btrfs_item_key_to_cpu(leaf, &cpukey, slot + 1);
 447                 BUG_ON(comp_keys(&leaf_key, &cpukey) >= 0);
 448                 if (btrfs_item_offset_nr(leaf, slot) !=
 449                         btrfs_item_end_nr(leaf, slot + 1)) {
 450                         btrfs_print_leaf(root, leaf);
 451                         printk("slot %d offset bad\n", slot);
 452                         BUG_ON(1);
 453                 }
 454         }
 455         BUG_ON(btrfs_item_offset_nr(leaf, 0) +
 456                btrfs_item_size_nr(leaf, 0) != BTRFS_LEAF_DATA_SIZE(root));
 457         return 0;
 458 }
 459
 460 static int check_block(struct btrfs_root *root, struct btrfs_path *path,
 461                         int level)
 462 {
 463         return 0;
 464 #if 0
 465         struct extent_buffer *buf = path->nodes[level];
 466
 467         if (memcmp_extent_buffer(buf, root->fs_info->fsid,
 468                                  (unsigned long)btrfs_header_fsid(buf),
 469                                  BTRFS_FSID_SIZE)) {
 470                 printk("warning bad block %Lu\n", buf->start);
 471                 return 1;
 472         }
 473 #endif
 474         if (level == 0)
 475                 return check_leaf(root, path, level);
 476         return check_node(root, path, level);
 477 }
 478
 479 /*
 480  * search for key in the extent_buffer.  The items start at offset p,
 481  * and they are item_size apart.  There are 'max' items in p.
 482  *
 483  * the slot in the array is returned via slot, and it points to
 484  * the place where you would insert key if it is not found in
 485  * the array.
 486  *
 487  * slot may point to max if the key is bigger than all of the keys
 488  */
 489 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 490                               int item_size, struct btrfs_key *key,
 491                               int max, int *slot)
 492 {
 493         int low = 0;
 494         int high = max;
 495         int mid;
 496         int ret;
 497         struct btrfs_disk_key *tmp = NULL;
 498         struct btrfs_disk_key unaligned;
 499         unsigned long offset;
 500         char *map_token = NULL;
 501         char *kaddr = NULL;
 502         unsigned long map_start = 0;
 503         unsigned long map_len = 0;
 504         int err;
 505
 506         while(low < high) {
 507                 mid = (low + high) / 2;
 508                 offset = p + mid * item_size;
 509
 510                 if (!map_token || offset < map_start ||
 511                     (offset + sizeof(struct btrfs_disk_key)) >
 512                     map_start + map_len) {
 513                         if (map_token) {
 514                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 515                                 map_token = NULL;
 516                         }
 517                         err = map_extent_buffer(eb, offset,
 518                                                 sizeof(struct btrfs_disk_key),
 519                                                 &map_token, &kaddr,
 520                                                 &map_start, &map_len, KM_USER0);
 521
 522                         if (!err) {
 523                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
 524                                                         map_start);
 525                         } else {
 526                                 read_extent_buffer(eb, &unaligned,
 527                                                    offset, sizeof(unaligned));
 528                                 tmp = &unaligned;
 529                         }
 530
 531                 } else {
 532                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
 533                                                         map_start);
 534                 }
 535                 ret = comp_keys(tmp, key);
 536
 537                 if (ret < 0)
 538                         low = mid + 1;
 539                 else if (ret > 0)
 540                         high = mid;
 541                 else {
 542                         *slot = mid;
 543                         if (map_token)
 544                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 545                         return 0;
 546                 }
 547         }
 548         *slot = low;
 549         if (map_token)
 550                 unmap_extent_buffer(eb, map_token, KM_USER0);
 551         return 1;
 552 }
 553
 554 /*
 555  * simple bin_search frontend that does the right thing for
 556  * leaves vs nodes
 557  */
 558 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 559                       int level, int *slot)
 560 {
 561         if (level == 0) {
 562                 return generic_bin_search(eb,
 563                                           offsetof(struct btrfs_leaf, items),
 564                                           sizeof(struct btrfs_item),
 565                                           key, btrfs_header_nritems(eb),
 566                                           slot);
 567         } else {
 568                 return generic_bin_search(eb,
 569                                           offsetof(struct btrfs_node, ptrs),
 570                                           sizeof(struct btrfs_key_ptr),
 571                                           key, btrfs_header_nritems(eb),
 572                                           slot);
 573         }
 574         return -1;
 575 }
 576
 577 static struct extent_buffer *read_node_slot(struct btrfs_root *root,
 578                                    struct extent_buffer *parent, int slot)
 579 {
 580         if (slot < 0)
 581                 return NULL;
 582         if (slot >= btrfs_header_nritems(parent))
 583                 return NULL;
 584         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 585                        btrfs_level_size(root, btrfs_header_level(parent) - 1));
 586 }
 587
 588 static int balance_level(struct btrfs_trans_handle *trans, struct btrfs_root
 589                          *root, struct btrfs_path *path, int level)
 590 {
 591         struct extent_buffer *right = NULL;
 592         struct extent_buffer *mid;
 593         struct extent_buffer *left = NULL;
 594         struct extent_buffer *parent = NULL;
 595         int ret = 0;
 596         int wret;
 597         int pslot;
 598         int orig_slot = path->slots[level];
 599         int err_on_enospc = 0;
 600         u64 orig_ptr;
 601
 602         if (level == 0)
 603                 return 0;
 604
 605         mid = path->nodes[level];
 606         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 607
 608         if (level < BTRFS_MAX_LEVEL - 1)
 609                 parent = path->nodes[level + 1];
 610         pslot = path->slots[level + 1];
 611
 612         /*
 613          * deal with the case where there is only one pointer in the root
 614          * by promoting the node below to a root
 615          */
 616         if (!parent) {
 617                 struct extent_buffer *child;
 618
 619                 if (btrfs_header_nritems(mid) != 1)
 620                         return 0;
 621
 622                 /* promote the child to a root */
 623                 child = read_node_slot(root, mid, 0);
 624                 BUG_ON(!child);
 625                 root->node = child;
 626                 path->nodes[level] = NULL;
 627                 clean_tree_block(trans, root, mid);
 628                 wait_on_tree_block_writeback(root, mid);
 629                 /* once for the path */
 630                 free_extent_buffer(mid);
 631                 ret = btrfs_free_extent(trans, root, mid->start, mid->len, 1);
 632                 /* once for the root ptr */
 633                 free_extent_buffer(mid);
 634                 return ret;
 635         }
 636         if (btrfs_header_nritems(mid) >
 637             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 638                 return 0;
 639
 640         if (btrfs_header_nritems(mid) < 2)
 641                 err_on_enospc = 1;
 642
 643         left = read_node_slot(root, parent, pslot - 1);
 644         if (left) {
 645                 wret = btrfs_cow_block(trans, root, left,
 646                                        parent, pslot - 1, &left);
 647                 if (wret) {
 648                         ret = wret;
 649                         goto enospc;
 650                 }
 651         }
 652         right = read_node_slot(root, parent, pslot + 1);
 653         if (right) {
 654                 wret = btrfs_cow_block(trans, root, right,
 655                                        parent, pslot + 1, &right);
 656                 if (wret) {
 657                         ret = wret;
 658                         goto enospc;
 659                 }
 660         }
 661
 662         /* first, try to make some room in the middle buffer */
 663         if (left) {
 664                 orig_slot += btrfs_header_nritems(left);
 665                 wret = push_node_left(trans, root, left, mid);
 666                 if (wret < 0)
 667                         ret = wret;
 668                 if (btrfs_header_nritems(mid) < 2)
 669                         err_on_enospc = 1;
 670         }
 671
 672         /*
 673          * then try to empty the right most buffer into the middle
 674          */
 675         if (right) {
 676                 wret = push_node_left(trans, root, mid, right);
 677                 if (wret < 0 && wret != -ENOSPC)
 678                         ret = wret;
 679                 if (btrfs_header_nritems(right) == 0) {
 680                         u64 bytenr = right->start;
 681                         u32 blocksize = right->len;
 682
 683                         clean_tree_block(trans, root, right);
 684                         wait_on_tree_block_writeback(root, right);
 685                         free_extent_buffer(right);
 686                         right = NULL;
 687                         wret = del_ptr(trans, root, path, level + 1, pslot +
 688                                        1);
 689                         if (wret)
 690                                 ret = wret;
 691                         wret = btrfs_free_extent(trans, root, bytenr,
 692                                                  blocksize, 1);
 693                         if (wret)
 694                                 ret = wret;
 695                 } else {
 696                         struct btrfs_disk_key right_key;
 697                         btrfs_node_key(right, &right_key, 0);
 698                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 699                         btrfs_mark_buffer_dirty(parent);
 700                 }
 701         }
 702         if (btrfs_header_nritems(mid) == 1) {
 703                 /*
 704                  * we're not allowed to leave a node with one item in the
 705                  * tree during a delete.  A deletion from lower in the tree
 706                  * could try to delete the only pointer in this node.
 707                  * So, pull some keys from the left.
 708                  * There has to be a left pointer at this point because
 709                  * otherwise we would have pulled some pointers from the
 710                  * right
 711                  */
 712                 BUG_ON(!left);
 713                 wret = balance_node_right(trans, root, mid, left);
 714                 if (wret < 0) {
 715                         ret = wret;
 716                         goto enospc;
 717                 }
 718                 BUG_ON(wret == 1);
 719         }
 720         if (btrfs_header_nritems(mid) == 0) {
 721                 /* we've managed to empty the middle node, drop it */
 722                 u64 bytenr = mid->start;
 723                 u32 blocksize = mid->len;
 724                 clean_tree_block(trans, root, mid);
 725                 wait_on_tree_block_writeback(root, mid);
 726                 free_extent_buffer(mid);
 727                 mid = NULL;
 728                 wret = del_ptr(trans, root, path, level + 1, pslot);
 729                 if (wret)
 730                         ret = wret;
 731                 wret = btrfs_free_extent(trans, root, bytenr, blocksize, 1);
 732                 if (wret)
 733                         ret = wret;
 734         } else {
 735                 /* update the parent key to reflect our changes */
 736                 struct btrfs_disk_key mid_key;
 737                 btrfs_node_key(mid, &mid_key, 0);
 738                 btrfs_set_node_key(parent, &mid_key, pslot);
 739                 btrfs_mark_buffer_dirty(parent);
 740         }
 741
 742         /* update the path */
 743         if (left) {
 744                 if (btrfs_header_nritems(left) > orig_slot) {
 745                         extent_buffer_get(left);
 746                         path->nodes[level] = left;
 747                         path->slots[level + 1] -= 1;
 748                         path->slots[level] = orig_slot;
 749                         if (mid)
 750                                 free_extent_buffer(mid);
 751                 } else {
 752                         orig_slot -= btrfs_header_nritems(left);
 753                         path->slots[level] = orig_slot;
 754                 }
 755         }
 756         /* double check we haven't messed things up */
 757         check_block(root, path, level);
 758         if (orig_ptr !=
 759             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 760                 BUG();
 761 enospc:
 762         if (right)
 763                 free_extent_buffer(right);
 764         if (left)
 765                 free_extent_buffer(left);
 766         return ret;
 767 }
 768
 769 /* returns zero if the push worked, non-zero otherwise */
 770 static int push_nodes_for_insert(struct btrfs_trans_handle *trans,
 771                                 struct btrfs_root *root,
 772                                 struct btrfs_path *path, int level)
 773 {
 774         struct extent_buffer *right = NULL;
 775         struct extent_buffer *mid;
 776         struct extent_buffer *left = NULL;
 777         struct extent_buffer *parent = NULL;
 778         int ret = 0;
 779         int wret;
 780         int pslot;
 781         int orig_slot = path->slots[level];
 782         u64 orig_ptr;
 783
 784         if (level == 0)
 785                 return 1;
 786
 787         mid = path->nodes[level];
 788         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 789
 790         if (level < BTRFS_MAX_LEVEL - 1)
 791                 parent = path->nodes[level + 1];
 792         pslot = path->slots[level + 1];
 793
 794         if (!parent)
 795                 return 1;
 796
 797         left = read_node_slot(root, parent, pslot - 1);
 798
 799         /* first, try to make some room in the middle buffer */
 800         if (left) {
 801                 u32 left_nr;
 802                 left_nr = btrfs_header_nritems(left);
 803                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 804                         wret = 1;
 805                 } else {
 806                         ret = btrfs_cow_block(trans, root, left, parent,
 807                                               pslot - 1, &left);
 808                         if (ret)
 809                                 wret = 1;
 810                         else {
 811                                 wret = push_node_left(trans, root,
 812                                                       left, mid);
 813                         }
 814                 }
 815                 if (wret < 0)
 816                         ret = wret;
 817                 if (wret == 0) {
 818                         struct btrfs_disk_key disk_key;
 819                         orig_slot += left_nr;
 820                         btrfs_node_key(mid, &disk_key, 0);
 821                         btrfs_set_node_key(parent, &disk_key, pslot);
 822                         btrfs_mark_buffer_dirty(parent);
 823                         if (btrfs_header_nritems(left) > orig_slot) {
 824                                 path->nodes[level] = left;
 825                                 path->slots[level + 1] -= 1;
 826                                 path->slots[level] = orig_slot;
 827                                 free_extent_buffer(mid);
 828                         } else {
 829                                 orig_slot -=
 830                                         btrfs_header_nritems(left);
 831                                 path->slots[level] = orig_slot;
 832                                 free_extent_buffer(left);
 833                         }
 834                         return 0;
 835                 }
 836                 free_extent_buffer(left);
 837         }
 838         right= read_node_slot(root, parent, pslot + 1);
 839
 840         /*
 841          * then try to empty the right most buffer into the middle
 842          */
 843         if (right) {
 844                 u32 right_nr;
 845                 right_nr = btrfs_header_nritems(right);
 846                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 847                         wret = 1;
 848                 } else {
 849                         ret = btrfs_cow_block(trans, root, right,
 850                                               parent, pslot + 1,
 851                                               &right);
 852                         if (ret)
 853                                 wret = 1;
 854                         else {
 855                                 wret = balance_node_right(trans, root,
 856                                                           right, mid);
 857                         }
 858                 }
 859                 if (wret < 0)
 860                         ret = wret;
 861                 if (wret == 0) {
 862                         struct btrfs_disk_key disk_key;
 863
 864                         btrfs_node_key(right, &disk_key, 0);
 865                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
 866                         btrfs_mark_buffer_dirty(parent);
 867
 868                         if (btrfs_header_nritems(mid) <= orig_slot) {
 869                                 path->nodes[level] = right;
 870                                 path->slots[level + 1] += 1;
 871                                 path->slots[level] = orig_slot -
 872                                         btrfs_header_nritems(mid);
 873                                 free_extent_buffer(mid);
 874                         } else {
 875                                 free_extent_buffer(right);
 876                         }
 877                         return 0;
 878                 }
 879                 free_extent_buffer(right);
 880         }
 881         return 1;
 882 }
 883
 884 /*
 885  * readahead one full node of leaves
 886  */
 887 static void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
 888                              int level, int slot)
 889 {
 890         struct extent_buffer *node;
 891         u32 nritems;
 892         u64 search;
 893         u64 lowest_read;
 894         u64 highest_read;
 895         u64 nread = 0;
 896         int direction = path->reada;
 897         struct extent_buffer *eb;
 898         u32 nr;
 899         u32 blocksize;
 900         u32 nscan = 0;
 901
 902         if (level != 1)
 903                 return;
 904
 905         if (!path->nodes[level])
 906                 return;
 907
 908         node = path->nodes[level];
 909         search = btrfs_node_blockptr(node, slot);
 910         blocksize = btrfs_level_size(root, level - 1);
 911         eb = btrfs_find_tree_block(root, search, blocksize);
 912         if (eb) {
 913                 free_extent_buffer(eb);
 914                 return;
 915         }
 916
 917         highest_read = search;
 918         lowest_read = search;
 919
 920         nritems = btrfs_header_nritems(node);
 921         nr = slot;
 922         while(1) {
 923                 if (direction < 0) {
 924                         if (nr == 0)
 925                                 break;
 926                         nr--;
 927                 } else if (direction > 0) {
 928                         nr++;
 929                         if (nr >= nritems)
 930                                 break;
 931                 }
 932                 search = btrfs_node_blockptr(node, nr);
 933                 if ((search >= lowest_read && search <= highest_read) ||
 934                     (search < lowest_read && lowest_read - search <= 32768) ||
 935                     (search > highest_read && search - highest_read <= 32768)) {
 936                         readahead_tree_block(root, search, blocksize);
 937                         nread += blocksize;
 938                 }
 939                 nscan++;
 940                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
 941                         break;
 942                 if(nread > (1024 * 1024) || nscan > 128)
 943                         break;
 944
 945                 if (search < lowest_read)
 946                         lowest_read = search;
 947                 if (search > highest_read)
 948                         highest_read = search;
 949         }
 950 }
 951 /*
 952  * look for key in the tree.  path is filled in with nodes along the way
 953  * if key is found, we return zero and you can find the item in the leaf
 954  * level of the path (level 0)
 955  *
 956  * If the key isn't found, the path points to the slot where it should
 957  * be inserted, and 1 is returned.  If there are other errors during the
 958  * search a negative error number is returned.
 959  *
 960  * if ins_len > 0, nodes and leaves will be split as we walk down the
 961  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
 962  * possible)
 963  */
 964 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
 965                       *root, struct btrfs_key *key, struct btrfs_path *p, int
 966                       ins_len, int cow)
 967 {
 968         struct extent_buffer *b;
 969         u64 bytenr;
 970         int slot;
 971         int ret;
 972         int level;
 973         int should_reada = p->reada;
 974         u8 lowest_level = 0;
 975
 976         lowest_level = p->lowest_level;
 977         WARN_ON(lowest_level && ins_len);
 978         WARN_ON(p->nodes[0] != NULL);
 979         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
 980 again:
 981         b = root->node;
 982         extent_buffer_get(b);
 983         while (b) {
 984                 level = btrfs_header_level(b);
 985                 if (cow) {
 986                         int wret;
 987                         wret = btrfs_cow_block(trans, root, b,
 988                                                p->nodes[level + 1],
 989                                                p->slots[level + 1],
 990                                                &b);
 991                         if (wret) {
 992                                 free_extent_buffer(b);
 993                                 return wret;
 994                         }
 995                 }
 996                 BUG_ON(!cow && ins_len);
 997                 if (level != btrfs_header_level(b))
 998                         WARN_ON(1);
 999                 level = btrfs_header_level(b);
1000                 p->nodes[level] = b;
1001                 ret = check_block(root, p, level);
1002                 if (ret)
1003                         return -1;
1004                 ret = bin_search(b, key, level, &slot);
1005                 if (level != 0) {
1006                         if (ret && slot > 0)
1007                                 slot -= 1;
1008                         p->slots[level] = slot;
1009                         if (ins_len > 0 && btrfs_header_nritems(b) >=
1010                             BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1011                                 int sret = split_node(trans, root, p, level);
1012                                 BUG_ON(sret > 0);
1013                                 if (sret)
1014                                         return sret;
1015                                 b = p->nodes[level];
1016                                 slot = p->slots[level];
1017                         } else if (ins_len < 0) {
1018                                 int sret = balance_level(trans, root, p,
1019                                                          level);
1020                                 if (sret)
1021                                         return sret;
1022                                 b = p->nodes[level];
1023                                 if (!b) {
1024                                         btrfs_release_path(NULL, p);
1025                                         goto again;
1026                                 }
1027                                 slot = p->slots[level];
1028                                 BUG_ON(btrfs_header_nritems(b) == 1);
1029                         }
1030                         /* this is only true while dropping a snapshot */
1031                         if (level == lowest_level)
1032                                 break;
1033                         bytenr = btrfs_node_blockptr(b, slot);
1034                         if (should_reada)
1035                                 reada_for_search(root, p, level, slot);
1036                         b = read_tree_block(root, bytenr,
1037                                             btrfs_level_size(root, level - 1));
1038                 } else {
1039                         p->slots[level] = slot;
1040                         if (ins_len > 0 && btrfs_leaf_free_space(root, b) <
1041                             sizeof(struct btrfs_item) + ins_len) {
1042                                 int sret = split_leaf(trans, root, key,
1043                                                       p, ins_len, ret == 0);
1044                                 BUG_ON(sret > 0);
1045                                 if (sret)
1046                                         return sret;
1047                         }
1048                         return ret;
1049                 }
1050         }
1051         return 1;
1052 }
1053
1054 /*
1055  * adjust the pointers going up the tree, starting at level
1056  * making sure the right key of each node is points to 'key'.
1057  * This is used after shifting pointers to the left, so it stops
1058  * fixing up pointers when a given leaf/node is not in slot 0 of the
1059  * higher levels
1060  *
1061  * If this fails to write a tree block, it returns -1, but continues
1062  * fixing up the blocks in ram so the tree is consistent.
1063  */
1064 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1065                           struct btrfs_root *root, struct btrfs_path *path,
1066                           struct btrfs_disk_key *key, int level)
1067 {
1068         int i;
1069         int ret = 0;
1070         struct extent_buffer *t;
1071
1072         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1073                 int tslot = path->slots[i];
1074                 if (!path->nodes[i])
1075                         break;
1076                 t = path->nodes[i];
1077                 btrfs_set_node_key(t, key, tslot);
1078                 btrfs_mark_buffer_dirty(path->nodes[i]);
1079                 if (tslot != 0)
1080                         break;
1081         }
1082         return ret;
1083 }
1084
1085 /*
1086  * try to push data from one node into the next node left in the
1087  * tree.
1088  *
1089  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1090  * error, and > 0 if there was no room in the left hand block.
1091  */
1092 static int push_node_left(struct btrfs_trans_handle *trans, struct btrfs_root
1093                           *root, struct extent_buffer *dst,
1094                           struct extent_buffer *src)
1095 {
1096         int push_items = 0;
1097         int src_nritems;
1098         int dst_nritems;
1099         int ret = 0;
1100
1101         src_nritems = btrfs_header_nritems(src);
1102         dst_nritems = btrfs_header_nritems(dst);
1103         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1104
1105         if (push_items <= 0) {
1106                 return 1;
1107         }
1108
1109         if (src_nritems < push_items)
1110                 push_items = src_nritems;
1111
1112         copy_extent_buffer(dst, src,
1113                            btrfs_node_key_ptr_offset(dst_nritems),
1114                            btrfs_node_key_ptr_offset(0),
1115                            push_items * sizeof(struct btrfs_key_ptr));
1116
1117         if (push_items < src_nritems) {
1118                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1119                                       btrfs_node_key_ptr_offset(push_items),
1120                                       (src_nritems - push_items) *
1121                                       sizeof(struct btrfs_key_ptr));
1122         }
1123         btrfs_set_header_nritems(src, src_nritems - push_items);
1124         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1125         btrfs_mark_buffer_dirty(src);
1126         btrfs_mark_buffer_dirty(dst);
1127         return ret;
1128 }
1129
1130 /*
1131  * try to push data from one node into the next node right in the
1132  * tree.
1133  *
1134  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1135  * error, and > 0 if there was no room in the right hand block.
1136  *
1137  * this will  only push up to 1/2 the contents of the left node over
1138  */
1139 static int balance_node_right(struct btrfs_trans_handle *trans,
1140                               struct btrfs_root *root,
1141                               struct extent_buffer *dst,
1142                               struct extent_buffer *src)
1143 {
1144         int push_items = 0;
1145         int max_push;
1146         int src_nritems;
1147         int dst_nritems;
1148         int ret = 0;
1149
1150         src_nritems = btrfs_header_nritems(src);
1151         dst_nritems = btrfs_header_nritems(dst);
1152         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1153         if (push_items <= 0)
1154                 return 1;
1155
1156         max_push = src_nritems / 2 + 1;
1157         /* don't try to empty the node */
1158         if (max_push >= src_nritems)
1159                 return 1;
1160
1161         if (max_push < push_items)
1162                 push_items = max_push;
1163
1164         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1165                                       btrfs_node_key_ptr_offset(0),
1166                                       (dst_nritems) *
1167                                       sizeof(struct btrfs_key_ptr));
1168
1169         copy_extent_buffer(dst, src,
1170                            btrfs_node_key_ptr_offset(0),
1171                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1172                            push_items * sizeof(struct btrfs_key_ptr));
1173
1174         btrfs_set_header_nritems(src, src_nritems - push_items);
1175         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1176
1177         btrfs_mark_buffer_dirty(src);
1178         btrfs_mark_buffer_dirty(dst);
1179         return ret;
1180 }
1181
1182 /*
1183  * helper function to insert a new root level in the tree.
1184  * A new node is allocated, and a single item is inserted to
1185  * point to the existing root
1186  *
1187  * returns zero on success or < 0 on failure.
1188  */
1189 static int insert_new_root(struct btrfs_trans_handle *trans,
1190                            struct btrfs_root *root,
1191                            struct btrfs_path *path, int level)
1192 {
1193         struct extent_buffer *lower;
1194         struct extent_buffer *c;
1195         struct btrfs_disk_key lower_key;
1196
1197         BUG_ON(path->nodes[level]);
1198         BUG_ON(path->nodes[level-1] != root->node);
1199
1200         c = btrfs_alloc_free_block(trans, root, root->nodesize,
1201                                    root->node->start, 0);
1202         if (IS_ERR(c))
1203                 return PTR_ERR(c);
1204         memset_extent_buffer(c, 0, 0, root->nodesize);
1205         btrfs_set_header_nritems(c, 1);
1206         btrfs_set_header_level(c, level);
1207         btrfs_set_header_bytenr(c, c->start);
1208         btrfs_set_header_generation(c, trans->transid);
1209         btrfs_set_header_owner(c, root->root_key.objectid);
1210         lower = path->nodes[level-1];
1211
1212         write_extent_buffer(c, root->fs_info->fsid,
1213                             (unsigned long)btrfs_header_fsid(c),
1214                             BTRFS_FSID_SIZE);
1215         if (level == 1)
1216                 btrfs_item_key(lower, &lower_key, 0);
1217         else
1218                 btrfs_node_key(lower, &lower_key, 0);
1219         btrfs_set_node_key(c, &lower_key, 0);
1220         btrfs_set_node_blockptr(c, 0, lower->start);
1221
1222         btrfs_mark_buffer_dirty(c);
1223
1224         /* the super has an extra ref to root->node */
1225         free_extent_buffer(root->node);
1226         root->node = c;
1227         extent_buffer_get(c);
1228         path->nodes[level] = c;
1229         path->slots[level] = 0;
1230         return 0;
1231 }
1232
1233 /*
1234  * worker function to insert a single pointer in a node.
1235  * the node should have enough room for the pointer already
1236  *
1237  * slot and level indicate where you want the key to go, and
1238  * blocknr is the block the key points to.
1239  *
1240  * returns zero on success and < 0 on any error
1241  */
1242 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1243                       *root, struct btrfs_path *path, struct btrfs_disk_key
1244                       *key, u64 bytenr, int slot, int level)
1245 {
1246         struct extent_buffer *lower;
1247         int nritems;
1248
1249         BUG_ON(!path->nodes[level]);
1250         lower = path->nodes[level];
1251         nritems = btrfs_header_nritems(lower);
1252         if (slot > nritems)
1253                 BUG();
1254         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1255                 BUG();
1256         if (slot != nritems) {
1257                 memmove_extent_buffer(lower,
1258                               btrfs_node_key_ptr_offset(slot + 1),
1259                               btrfs_node_key_ptr_offset(slot),
1260                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1261         }
1262         btrfs_set_node_key(lower, key, slot);
1263         btrfs_set_node_blockptr(lower, slot, bytenr);
1264         btrfs_set_header_nritems(lower, nritems + 1);
1265         btrfs_mark_buffer_dirty(lower);
1266         return 0;
1267 }
1268
1269 /*
1270  * split the node at the specified level in path in two.
1271  * The path is corrected to point to the appropriate node after the split
1272  *
1273  * Before splitting this tries to make some room in the node by pushing
1274  * left and right, if either one works, it returns right away.
1275  *
1276  * returns 0 on success and < 0 on failure
1277  */
1278 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1279                       *root, struct btrfs_path *path, int level)
1280 {
1281         struct extent_buffer *c;
1282         struct extent_buffer *split;
1283         struct btrfs_disk_key disk_key;
1284         int mid;
1285         int ret;
1286         int wret;
1287         u32 c_nritems;
1288
1289         c = path->nodes[level];
1290         if (c == root->node) {
1291                 /* trying to split the root, lets make a new one */
1292                 ret = insert_new_root(trans, root, path, level + 1);
1293                 if (ret)
1294                         return ret;
1295         } else {
1296                 ret = push_nodes_for_insert(trans, root, path, level);
1297                 c = path->nodes[level];
1298                 if (!ret && btrfs_header_nritems(c) <
1299                     BTRFS_NODEPTRS_PER_BLOCK(root) - 1)
1300                         return 0;
1301                 if (ret < 0)
1302                         return ret;
1303         }
1304
1305         c_nritems = btrfs_header_nritems(c);
1306         split = btrfs_alloc_free_block(trans, root, root->nodesize,
1307                                        c->start, 0);
1308         if (IS_ERR(split))
1309                 return PTR_ERR(split);
1310
1311         btrfs_set_header_flags(split, btrfs_header_flags(c));
1312         btrfs_set_header_level(split, btrfs_header_level(c));
1313         btrfs_set_header_bytenr(split, split->start);
1314         btrfs_set_header_generation(split, trans->transid);
1315         btrfs_set_header_owner(split, root->root_key.objectid);
1316         write_extent_buffer(split, root->fs_info->fsid,
1317                             (unsigned long)btrfs_header_fsid(split),
1318                             BTRFS_FSID_SIZE);
1319
1320         mid = (c_nritems + 1) / 2;
1321
1322         copy_extent_buffer(split, c,
1323                            btrfs_node_key_ptr_offset(0),
1324                            btrfs_node_key_ptr_offset(mid),
1325                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1326         btrfs_set_header_nritems(split, c_nritems - mid);
1327         btrfs_set_header_nritems(c, mid);
1328         ret = 0;
1329
1330         btrfs_mark_buffer_dirty(c);
1331         btrfs_mark_buffer_dirty(split);
1332
1333         btrfs_node_key(split, &disk_key, 0);
1334         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1335                           path->slots[level + 1] + 1,
1336                           level + 1);
1337         if (wret)
1338                 ret = wret;
1339
1340         if (path->slots[level] >= mid) {
1341                 path->slots[level] -= mid;
1342                 free_extent_buffer(c);
1343                 path->nodes[level] = split;
1344                 path->slots[level + 1] += 1;
1345         } else {
1346                 free_extent_buffer(split);
1347         }
1348         return ret;
1349 }
1350
1351 /*
1352  * how many bytes are required to store the items in a leaf.  start
1353  * and nr indicate which items in the leaf to check.  This totals up the
1354  * space used both by the item structs and the item data
1355  */
1356 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1357 {
1358         int data_len;
1359         int nritems = btrfs_header_nritems(l);
1360         int end = min(nritems, start + nr) - 1;
1361
1362         if (!nr)
1363                 return 0;
1364         data_len = btrfs_item_end_nr(l, start);
1365         data_len = data_len - btrfs_item_offset_nr(l, end);
1366         data_len += sizeof(struct btrfs_item) * nr;
1367         WARN_ON(data_len < 0);
1368         return data_len;
1369 }
1370
1371 /*
1372  * The space between the end of the leaf items and
1373  * the start of the leaf data.  IOW, how much room
1374  * the leaf has left for both items and data
1375  */
1376 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1377 {
1378         int nritems = btrfs_header_nritems(leaf);
1379         int ret;
1380         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1381         if (ret < 0) {
1382                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1383                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1384                        leaf_space_used(leaf, 0, nritems), nritems);
1385         }
1386         return ret;
1387 }
1388
1389 /*
1390  * push some data in the path leaf to the right, trying to free up at
1391  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1392  *
1393  * returns 1 if the push failed because the other node didn't have enough
1394  * room, 0 if everything worked out and < 0 if there were major errors.
1395  */
1396 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1397                            *root, struct btrfs_path *path, int data_size,
1398                            int empty)
1399 {
1400         struct extent_buffer *left = path->nodes[0];
1401         struct extent_buffer *right;
1402         struct extent_buffer *upper;
1403         struct btrfs_disk_key disk_key;
1404         int slot;
1405         u32 i;
1406         int free_space;
1407         int push_space = 0;
1408         int push_items = 0;
1409         struct btrfs_item *item;
1410         u32 left_nritems;
1411         u32 nr;
1412         u32 right_nritems;
1413         u32 data_end;
1414         u32 this_item_size;
1415         int ret;
1416
1417         slot = path->slots[1];
1418         if (!path->nodes[1]) {
1419                 return 1;
1420         }
1421         upper = path->nodes[1];
1422         if (slot >= btrfs_header_nritems(upper) - 1)
1423                 return 1;
1424
1425         right = read_tree_block(root, btrfs_node_blockptr(upper, slot + 1),
1426                                 root->leafsize);
1427         free_space = btrfs_leaf_free_space(root, right);
1428         if (free_space < data_size + sizeof(struct btrfs_item)) {
1429                 free_extent_buffer(right);
1430                 return 1;
1431         }
1432
1433         /* cow and double check */
1434         ret = btrfs_cow_block(trans, root, right, upper,
1435                               slot + 1, &right);
1436         if (ret) {
1437                 free_extent_buffer(right);
1438                 return 1;
1439         }
1440         free_space = btrfs_leaf_free_space(root, right);
1441         if (free_space < data_size + sizeof(struct btrfs_item)) {
1442                 free_extent_buffer(right);
1443                 return 1;
1444         }
1445
1446         left_nritems = btrfs_header_nritems(left);
1447         if (left_nritems == 0) {
1448                 free_extent_buffer(right);
1449                 return 1;
1450         }
1451
1452         if (empty)
1453                 nr = 0;
1454         else
1455                 nr = 1;
1456
1457         i = left_nritems - 1;
1458         while (i >= nr) {
1459                 item = btrfs_item_nr(left, i);
1460
1461                 if (path->slots[0] == i)
1462                         push_space += data_size + sizeof(*item);
1463
1464                 if (!left->map_token) {
1465                         map_extent_buffer(left, (unsigned long)item,
1466                                         sizeof(struct btrfs_item),
1467                                         &left->map_token, &left->kaddr,
1468                                         &left->map_start, &left->map_len,
1469                                         KM_USER1);
1470                 }
1471
1472                 this_item_size = btrfs_item_size(left, item);
1473                 if (this_item_size + sizeof(*item) + push_space > free_space)
1474                         break;
1475                 push_items++;
1476                 push_space += this_item_size + sizeof(*item);
1477                 if (i == 0)
1478                         break;
1479                 i--;
1480         }
1481         if (left->map_token) {
1482                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1483                 left->map_token = NULL;
1484         }
1485
1486         if (push_items == 0) {
1487                 free_extent_buffer(right);
1488                 return 1;
1489         }
1490
1491         if (!empty && push_items == left_nritems)
1492                 WARN_ON(1);
1493
1494         /* push left to right */
1495         right_nritems = btrfs_header_nritems(right);
1496
1497         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1498         push_space -= leaf_data_end(root, left);
1499
1500         /* make room in the right data area */
1501         data_end = leaf_data_end(root, right);
1502         memmove_extent_buffer(right,
1503                               btrfs_leaf_data(right) + data_end - push_space,
1504                               btrfs_leaf_data(right) + data_end,
1505                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1506
1507         /* copy from the left data area */
1508         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1509                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1510                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1511                      push_space);
1512
1513         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1514                               btrfs_item_nr_offset(0),
1515                               right_nritems * sizeof(struct btrfs_item));
1516
1517         /* copy the items from left to right */
1518         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1519                    btrfs_item_nr_offset(left_nritems - push_items),
1520                    push_items * sizeof(struct btrfs_item));
1521
1522         /* update the item pointers */
1523         right_nritems += push_items;
1524         btrfs_set_header_nritems(right, right_nritems);
1525         push_space = BTRFS_LEAF_DATA_SIZE(root);
1526         for (i = 0; i < right_nritems; i++) {
1527                 item = btrfs_item_nr(right, i);
1528                 if (!right->map_token) {
1529                         map_extent_buffer(right, (unsigned long)item,
1530                                         sizeof(struct btrfs_item),
1531                                         &right->map_token, &right->kaddr,
1532                                         &right->map_start, &right->map_len,
1533                                         KM_USER1);
1534                 }
1535                 push_space -= btrfs_item_size(right, item);
1536                 btrfs_set_item_offset(right, item, push_space);
1537         }
1538
1539         if (right->map_token) {
1540                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1541                 right->map_token = NULL;
1542         }
1543         left_nritems -= push_items;
1544         btrfs_set_header_nritems(left, left_nritems);
1545
1546         if (left_nritems)
1547                 btrfs_mark_buffer_dirty(left);
1548         btrfs_mark_buffer_dirty(right);
1549
1550         btrfs_item_key(right, &disk_key, 0);
1551         btrfs_set_node_key(upper, &disk_key, slot + 1);
1552         btrfs_mark_buffer_dirty(upper);
1553
1554         /* then fixup the leaf pointer in the path */
1555         if (path->slots[0] >= left_nritems) {
1556                 path->slots[0] -= left_nritems;
1557                 free_extent_buffer(path->nodes[0]);
1558                 path->nodes[0] = right;
1559                 path->slots[1] += 1;
1560         } else {
1561                 free_extent_buffer(right);
1562         }
1563         return 0;
1564 }
1565 /*
1566  * push some data in the path leaf to the left, trying to free up at
1567  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1568  */
1569 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1570                           *root, struct btrfs_path *path, int data_size,
1571                           int empty)
1572 {
1573         struct btrfs_disk_key disk_key;
1574         struct extent_buffer *right = path->nodes[0];
1575         struct extent_buffer *left;
1576         int slot;
1577         int i;
1578         int free_space;
1579         int push_space = 0;
1580         int push_items = 0;
1581         struct btrfs_item *item;
1582         u32 old_left_nritems;
1583         u32 right_nritems;
1584         u32 nr;
1585         int ret = 0;
1586         int wret;
1587         u32 this_item_size;
1588         u32 old_left_item_size;
1589
1590         slot = path->slots[1];
1591         if (slot == 0)
1592                 return 1;
1593         if (!path->nodes[1])
1594                 return 1;
1595
1596         right_nritems = btrfs_header_nritems(right);
1597         if (right_nritems == 0) {
1598                 return 1;
1599         }
1600
1601         left = read_tree_block(root, btrfs_node_blockptr(path->nodes[1],
1602                                slot - 1), root->leafsize);
1603         free_space = btrfs_leaf_free_space(root, left);
1604         if (free_space < data_size + sizeof(struct btrfs_item)) {
1605                 free_extent_buffer(left);
1606                 return 1;
1607         }
1608
1609         /* cow and double check */
1610         ret = btrfs_cow_block(trans, root, left,
1611                               path->nodes[1], slot - 1, &left);
1612         if (ret) {
1613                 /* we hit -ENOSPC, but it isn't fatal here */
1614                 free_extent_buffer(left);
1615                 return 1;
1616         }
1617
1618         free_space = btrfs_leaf_free_space(root, left);
1619         if (free_space < data_size + sizeof(struct btrfs_item)) {
1620                 free_extent_buffer(left);
1621                 return 1;
1622         }
1623
1624         if (empty)
1625                 nr = right_nritems;
1626         else
1627                 nr = right_nritems - 1;
1628
1629         for (i = 0; i < nr; i++) {
1630                 item = btrfs_item_nr(right, i);
1631                 if (!right->map_token) {
1632                         map_extent_buffer(right, (unsigned long)item,
1633                                         sizeof(struct btrfs_item),
1634                                         &right->map_token, &right->kaddr,
1635                                         &right->map_start, &right->map_len,
1636                                         KM_USER1);
1637                 }
1638
1639                 if (path->slots[0] == i)
1640                         push_space += data_size + sizeof(*item);
1641
1642                 this_item_size = btrfs_item_size(right, item);
1643                 if (this_item_size + sizeof(*item) + push_space > free_space)
1644                         break;
1645
1646                 push_items++;
1647                 push_space += this_item_size + sizeof(*item);
1648         }
1649
1650         if (right->map_token) {
1651                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1652                 right->map_token = NULL;
1653         }
1654
1655         if (push_items == 0) {
1656                 free_extent_buffer(left);
1657                 return 1;
1658         }
1659         if (!empty && push_items == btrfs_header_nritems(right))
1660                 WARN_ON(1);
1661
1662         /* push data from right to left */
1663         copy_extent_buffer(left, right,
1664                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1665                            btrfs_item_nr_offset(0),
1666                            push_items * sizeof(struct btrfs_item));
1667
1668         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1669                      btrfs_item_offset_nr(right, push_items -1);
1670
1671         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1672                      leaf_data_end(root, left) - push_space,
1673                      btrfs_leaf_data(right) +
1674                      btrfs_item_offset_nr(right, push_items - 1),
1675                      push_space);
1676         old_left_nritems = btrfs_header_nritems(left);
1677         BUG_ON(old_left_nritems < 0);
1678
1679         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1680         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1681                 u32 ioff;
1682
1683                 item = btrfs_item_nr(left, i);
1684                 if (!left->map_token) {
1685                         map_extent_buffer(left, (unsigned long)item,
1686                                         sizeof(struct btrfs_item),
1687                                         &left->map_token, &left->kaddr,
1688                                         &left->map_start, &left->map_len,
1689                                         KM_USER1);
1690                 }
1691
1692                 ioff = btrfs_item_offset(left, item);
1693                 btrfs_set_item_offset(left, item,
1694                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1695         }
1696         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1697         if (left->map_token) {
1698                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1699                 left->map_token = NULL;
1700         }
1701
1702         /* fixup right node */
1703         if (push_items > right_nritems) {
1704                 printk("push items %d nr %u\n", push_items, right_nritems);
1705                 WARN_ON(1);
1706         }
1707
1708         if (push_items < right_nritems) {
1709                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1710                                                   leaf_data_end(root, right);
1711                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
1712                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1713                                       btrfs_leaf_data(right) +
1714                                       leaf_data_end(root, right), push_space);
1715
1716                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1717                               btrfs_item_nr_offset(push_items),
1718                              (btrfs_header_nritems(right) - push_items) *
1719                              sizeof(struct btrfs_item));
1720
1721         }
1722         btrfs_set_header_nritems(right, right_nritems - push_items);
1723         push_space = BTRFS_LEAF_DATA_SIZE(root);
1724
1725         for (i = 0; i < right_nritems; i++) {
1726                 item = btrfs_item_nr(right, i);
1727
1728                 if (!right->map_token) {
1729                         map_extent_buffer(right, (unsigned long)item,
1730                                         sizeof(struct btrfs_item),
1731                                         &right->map_token, &right->kaddr,
1732                                         &right->map_start, &right->map_len,
1733                                         KM_USER1);
1734                 }
1735
1736                 push_space = push_space - btrfs_item_size(right, item);
1737                 btrfs_set_item_offset(right, item, push_space);
1738         }
1739         if (right->map_token) {
1740                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1741                 right->map_token = NULL;
1742         }
1743
1744         btrfs_mark_buffer_dirty(left);
1745         if (right_nritems)
1746                 btrfs_mark_buffer_dirty(right);
1747
1748         btrfs_item_key(right, &disk_key, 0);
1749         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1750         if (wret)
1751                 ret = wret;
1752
1753         /* then fixup the leaf pointer in the path */
1754         if (path->slots[0] < push_items) {
1755                 path->slots[0] += old_left_nritems;
1756                 free_extent_buffer(path->nodes[0]);
1757                 path->nodes[0] = left;
1758                 path->slots[1] -= 1;
1759         } else {
1760                 free_extent_buffer(left);
1761                 path->slots[0] -= push_items;
1762         }
1763         BUG_ON(path->slots[0] < 0);
1764         return ret;
1765 }
1766
1767 /*
1768  * split the path's leaf in two, making sure there is at least data_size
1769  * available for the resulting leaf level of the path.
1770  *
1771  * returns 0 if all went well and < 0 on failure.
1772  */
1773 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
1774                       *root, struct btrfs_key *ins_key,
1775                       struct btrfs_path *path, int data_size, int extend)
1776 {
1777         struct extent_buffer *l;
1778         u32 nritems;
1779         int mid;
1780         int slot;
1781         struct extent_buffer *right;
1782         int space_needed = data_size + sizeof(struct btrfs_item);
1783         int data_copy_size;
1784         int rt_data_off;
1785         int i;
1786         int ret = 0;
1787         int wret;
1788         int double_split;
1789         int num_doubles = 0;
1790         struct btrfs_disk_key disk_key;
1791
1792         if (extend)
1793                 space_needed = data_size;
1794
1795         /* first try to make some room by pushing left and right */
1796         if (ins_key->type != BTRFS_DIR_ITEM_KEY) {
1797                 wret = push_leaf_right(trans, root, path, data_size, 0);
1798                 if (wret < 0) {
1799                         return wret;
1800                 }
1801                 if (wret) {
1802                         wret = push_leaf_left(trans, root, path, data_size, 0);
1803                         if (wret < 0)
1804                                 return wret;
1805                 }
1806                 l = path->nodes[0];
1807
1808                 /* did the pushes work? */
1809                 if (btrfs_leaf_free_space(root, l) >= space_needed)
1810                         return 0;
1811         }
1812
1813         if (!path->nodes[1]) {
1814                 ret = insert_new_root(trans, root, path, 1);
1815                 if (ret)
1816                         return ret;
1817         }
1818 again:
1819         double_split = 0;
1820         l = path->nodes[0];
1821         slot = path->slots[0];
1822         nritems = btrfs_header_nritems(l);
1823         mid = (nritems + 1)/ 2;
1824
1825         right = btrfs_alloc_free_block(trans, root, root->leafsize,
1826                                        l->start, 0);
1827         if (IS_ERR(right))
1828                 return PTR_ERR(right);
1829
1830         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
1831         btrfs_set_header_bytenr(right, right->start);
1832         btrfs_set_header_generation(right, trans->transid);
1833         btrfs_set_header_owner(right, root->root_key.objectid);
1834         btrfs_set_header_level(right, 0);
1835         write_extent_buffer(right, root->fs_info->fsid,
1836                             (unsigned long)btrfs_header_fsid(right),
1837                             BTRFS_FSID_SIZE);
1838         if (mid <= slot) {
1839                 if (nritems == 1 ||
1840                     leaf_space_used(l, mid, nritems - mid) + space_needed >
1841                         BTRFS_LEAF_DATA_SIZE(root)) {
1842                         if (slot >= nritems) {
1843                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1844                                 btrfs_set_header_nritems(right, 0);
1845                                 wret = insert_ptr(trans, root, path,
1846                                                   &disk_key, right->start,
1847                                                   path->slots[1] + 1, 1);
1848                                 if (wret)
1849                                         ret = wret;
1850                                 free_extent_buffer(path->nodes[0]);
1851                                 path->nodes[0] = right;
1852                                 path->slots[0] = 0;
1853                                 path->slots[1] += 1;
1854                                 return ret;
1855                         }
1856                         mid = slot;
1857                         if (mid != nritems &&
1858                             leaf_space_used(l, mid, nritems - mid) +
1859                             space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
1860                                 double_split = 1;
1861                         }
1862                 }
1863         } else {
1864                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
1865                         BTRFS_LEAF_DATA_SIZE(root)) {
1866                         if (!extend && slot == 0) {
1867                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1868                                 btrfs_set_header_nritems(right, 0);
1869                                 wret = insert_ptr(trans, root, path,
1870                                                   &disk_key,
1871                                                   right->start,
1872                                                   path->slots[1], 1);
1873                                 if (wret)
1874                                         ret = wret;
1875                                 free_extent_buffer(path->nodes[0]);
1876                                 path->nodes[0] = right;
1877                                 path->slots[0] = 0;
1878                                 if (path->slots[1] == 0) {
1879                                         wret = fixup_low_keys(trans, root,
1880                                                    path, &disk_key, 1);
1881                                         if (wret)
1882                                                 ret = wret;
1883                                 }
1884                                 return ret;
1885                         } else if (extend && slot == 0) {
1886                                 mid = 1;
1887                         } else {
1888                                 mid = slot;
1889                                 if (mid != nritems &&
1890                                     leaf_space_used(l, mid, nritems - mid) +
1891                                     space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
1892                                         double_split = 1;
1893                                 }
1894                         }
1895                 }
1896         }
1897         nritems = nritems - mid;
1898         btrfs_set_header_nritems(right, nritems);
1899         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
1900
1901         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
1902                            btrfs_item_nr_offset(mid),
1903                            nritems * sizeof(struct btrfs_item));
1904
1905         copy_extent_buffer(right, l,
1906                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
1907                      data_copy_size, btrfs_leaf_data(l) +
1908                      leaf_data_end(root, l), data_copy_size);
1909
1910         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
1911                       btrfs_item_end_nr(l, mid);
1912
1913         for (i = 0; i < nritems; i++) {
1914                 struct btrfs_item *item = btrfs_item_nr(right, i);
1915                 u32 ioff;
1916
1917                 if (!right->map_token) {
1918                         map_extent_buffer(right, (unsigned long)item,
1919                                         sizeof(struct btrfs_item),
1920                                         &right->map_token, &right->kaddr,
1921                                         &right->map_start, &right->map_len,
1922                                         KM_USER1);
1923                 }
1924
1925                 ioff = btrfs_item_offset(right, item);
1926                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
1927         }
1928
1929         if (right->map_token) {
1930                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1931                 right->map_token = NULL;
1932         }
1933
1934         btrfs_set_header_nritems(l, mid);
1935         ret = 0;
1936         btrfs_item_key(right, &disk_key, 0);
1937         wret = insert_ptr(trans, root, path, &disk_key, right->start,
1938                           path->slots[1] + 1, 1);
1939         if (wret)
1940                 ret = wret;
1941
1942         btrfs_mark_buffer_dirty(right);
1943         btrfs_mark_buffer_dirty(l);
1944         BUG_ON(path->slots[0] != slot);
1945
1946         if (mid <= slot) {
1947                 free_extent_buffer(path->nodes[0]);
1948                 path->nodes[0] = right;
1949                 path->slots[0] -= mid;
1950                 path->slots[1] += 1;
1951         } else
1952                 free_extent_buffer(right);
1953
1954         BUG_ON(path->slots[0] < 0);
1955
1956         if (double_split) {
1957                 BUG_ON(num_doubles != 0);
1958                 num_doubles++;
1959                 goto again;
1960         }
1961         return ret;
1962 }
1963
1964 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
1965                         struct btrfs_root *root,
1966                         struct btrfs_path *path,
1967                         u32 new_size, int from_end)
1968 {
1969         int ret = 0;
1970         int slot;
1971         int slot_orig;
1972         struct extent_buffer *leaf;
1973         struct btrfs_item *item;
1974         u32 nritems;
1975         unsigned int data_end;
1976         unsigned int old_data_start;
1977         unsigned int old_size;
1978         unsigned int size_diff;
1979         int i;
1980
1981         slot_orig = path->slots[0];
1982         leaf = path->nodes[0];
1983         slot = path->slots[0];
1984
1985         old_size = btrfs_item_size_nr(leaf, slot);
1986         if (old_size == new_size)
1987                 return 0;
1988
1989         nritems = btrfs_header_nritems(leaf);
1990         data_end = leaf_data_end(root, leaf);
1991
1992         old_data_start = btrfs_item_offset_nr(leaf, slot);
1993
1994         size_diff = old_size - new_size;
1995
1996         BUG_ON(slot < 0);
1997         BUG_ON(slot >= nritems);
1998
1999         /*
2000          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2001          */
2002         /* first correct the data pointers */
2003         for (i = slot; i < nritems; i++) {
2004                 u32 ioff;
2005                 item = btrfs_item_nr(leaf, i);
2006
2007                 if (!leaf->map_token) {
2008                         map_extent_buffer(leaf, (unsigned long)item,
2009                                         sizeof(struct btrfs_item),
2010                                         &leaf->map_token, &leaf->kaddr,
2011                                         &leaf->map_start, &leaf->map_len,
2012                                         KM_USER1);
2013                 }
2014
2015                 ioff = btrfs_item_offset(leaf, item);
2016                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2017         }
2018
2019         if (leaf->map_token) {
2020                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2021                 leaf->map_token = NULL;
2022         }
2023
2024         /* shift the data */
2025         if (from_end) {
2026                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2027                               data_end + size_diff, btrfs_leaf_data(leaf) +
2028                               data_end, old_data_start + new_size - data_end);
2029         } else {
2030                 struct btrfs_disk_key disk_key;
2031                 u64 offset;
2032
2033                 btrfs_item_key(leaf, &disk_key, slot);
2034
2035                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2036                         unsigned long ptr;
2037                         struct btrfs_file_extent_item *fi;
2038
2039                         fi = btrfs_item_ptr(leaf, slot,
2040                                             struct btrfs_file_extent_item);
2041                         fi = (struct btrfs_file_extent_item *)(
2042                              (unsigned long)fi - size_diff);
2043
2044                         if (btrfs_file_extent_type(leaf, fi) ==
2045                             BTRFS_FILE_EXTENT_INLINE) {
2046                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2047                                 memmove_extent_buffer(leaf, ptr,
2048                                         (unsigned long)fi,
2049                                         offsetof(struct btrfs_file_extent_item,
2050                                                  disk_bytenr));
2051                         }
2052                 }
2053
2054                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2055                               data_end + size_diff, btrfs_leaf_data(leaf) +
2056                               data_end, old_data_start - data_end);
2057
2058                 offset = btrfs_disk_key_offset(&disk_key);
2059                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2060                 btrfs_set_item_key(leaf, &disk_key, slot);
2061                 if (slot == 0)
2062                         fixup_low_keys(trans, root, path, &disk_key, 1);
2063         }
2064
2065         item = btrfs_item_nr(leaf, slot);
2066         btrfs_set_item_size(leaf, item, new_size);
2067         btrfs_mark_buffer_dirty(leaf);
2068
2069         ret = 0;
2070         if (btrfs_leaf_free_space(root, leaf) < 0) {
2071                 btrfs_print_leaf(root, leaf);
2072                 BUG();
2073         }
2074         return ret;
2075 }
2076
2077 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2078                       struct btrfs_root *root, struct btrfs_path *path,
2079                       u32 data_size)
2080 {
2081         int ret = 0;
2082         int slot;
2083         int slot_orig;
2084         struct extent_buffer *leaf;
2085         struct btrfs_item *item;
2086         u32 nritems;
2087         unsigned int data_end;
2088         unsigned int old_data;
2089         unsigned int old_size;
2090         int i;
2091
2092         slot_orig = path->slots[0];
2093         leaf = path->nodes[0];
2094
2095         nritems = btrfs_header_nritems(leaf);
2096         data_end = leaf_data_end(root, leaf);
2097
2098         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2099                 btrfs_print_leaf(root, leaf);
2100                 BUG();
2101         }
2102         slot = path->slots[0];
2103         old_data = btrfs_item_end_nr(leaf, slot);
2104
2105         BUG_ON(slot < 0);
2106         if (slot >= nritems) {
2107                 btrfs_print_leaf(root, leaf);
2108                 printk("slot %d too large, nritems %d\n", slot, nritems);
2109                 BUG_ON(1);
2110         }
2111
2112         /*
2113          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2114          */
2115         /* first correct the data pointers */
2116         for (i = slot; i < nritems; i++) {
2117                 u32 ioff;
2118                 item = btrfs_item_nr(leaf, i);
2119
2120                 if (!leaf->map_token) {
2121                         map_extent_buffer(leaf, (unsigned long)item,
2122                                         sizeof(struct btrfs_item),
2123                                         &leaf->map_token, &leaf->kaddr,
2124                                         &leaf->map_start, &leaf->map_len,
2125                                         KM_USER1);
2126                 }
2127                 ioff = btrfs_item_offset(leaf, item);
2128                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2129         }
2130
2131         if (leaf->map_token) {
2132                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2133                 leaf->map_token = NULL;
2134         }
2135
2136         /* shift the data */
2137         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2138                       data_end - data_size, btrfs_leaf_data(leaf) +
2139                       data_end, old_data - data_end);
2140
2141         data_end = old_data;
2142         old_size = btrfs_item_size_nr(leaf, slot);
2143         item = btrfs_item_nr(leaf, slot);
2144         btrfs_set_item_size(leaf, item, old_size + data_size);
2145         btrfs_mark_buffer_dirty(leaf);
2146
2147         ret = 0;
2148         if (btrfs_leaf_free_space(root, leaf) < 0) {
2149                 btrfs_print_leaf(root, leaf);
2150                 BUG();
2151         }
2152         return ret;
2153 }
2154
2155 /*
2156  * Given a key and some data, insert an item into the tree.
2157  * This does all the path init required, making room in the tree if needed.
2158  */
2159 int btrfs_insert_empty_item(struct btrfs_trans_handle *trans,
2160                             struct btrfs_root *root,
2161                             struct btrfs_path *path,
2162                             struct btrfs_key *cpu_key, u32 data_size)
2163 {
2164         struct extent_buffer *leaf;
2165         struct btrfs_item *item;
2166         int ret = 0;
2167         int slot;
2168         int slot_orig;
2169         u32 nritems;
2170         unsigned int data_end;
2171         struct btrfs_disk_key disk_key;
2172
2173         btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2174
2175         /* create a root if there isn't one */
2176         if (!root->node)
2177                 BUG();
2178
2179         ret = btrfs_search_slot(trans, root, cpu_key, path, data_size, 1);
2180         if (ret == 0) {
2181                 return -EEXIST;
2182         }
2183         if (ret < 0)
2184                 goto out;
2185
2186         slot_orig = path->slots[0];
2187         leaf = path->nodes[0];
2188
2189         nritems = btrfs_header_nritems(leaf);
2190         data_end = leaf_data_end(root, leaf);
2191
2192         if (btrfs_leaf_free_space(root, leaf) <
2193             sizeof(struct btrfs_item) + data_size) {
2194                 btrfs_print_leaf(root, leaf);
2195                 printk("not enough freespace need %u have %d\n",
2196                        data_size, btrfs_leaf_free_space(root, leaf));
2197                 BUG();
2198         }
2199
2200         slot = path->slots[0];
2201         BUG_ON(slot < 0);
2202
2203         if (slot != nritems) {
2204                 int i;
2205                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2206
2207                 if (old_data < data_end) {
2208                         btrfs_print_leaf(root, leaf);
2209                         printk("slot %d old_data %d data_end %d\n",
2210                                slot, old_data, data_end);
2211                         BUG_ON(1);
2212                 }
2213                 /*
2214                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2215                  */
2216                 /* first correct the data pointers */
2217                 WARN_ON(leaf->map_token);
2218                 for (i = slot; i < nritems; i++) {
2219                         u32 ioff;
2220
2221                         item = btrfs_item_nr(leaf, i);
2222                         if (!leaf->map_token) {
2223                                 map_extent_buffer(leaf, (unsigned long)item,
2224                                         sizeof(struct btrfs_item),
2225                                         &leaf->map_token, &leaf->kaddr,
2226                                         &leaf->map_start, &leaf->map_len,
2227                                         KM_USER1);
2228                         }
2229
2230                         ioff = btrfs_item_offset(leaf, item);
2231                         btrfs_set_item_offset(leaf, item, ioff - data_size);
2232                 }
2233                 if (leaf->map_token) {
2234                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2235                         leaf->map_token = NULL;
2236                 }
2237
2238                 /* shift the items */
2239                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
2240                               btrfs_item_nr_offset(slot),
2241                               (nritems - slot) * sizeof(struct btrfs_item));
2242
2243                 /* shift the data */
2244                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2245                               data_end - data_size, btrfs_leaf_data(leaf) +
2246                               data_end, old_data - data_end);
2247                 data_end = old_data;
2248         }
2249
2250         /* setup the item for the new data */
2251         btrfs_set_item_key(leaf, &disk_key, slot);
2252         item = btrfs_item_nr(leaf, slot);
2253         btrfs_set_item_offset(leaf, item, data_end - data_size);
2254         btrfs_set_item_size(leaf, item, data_size);
2255         btrfs_set_header_nritems(leaf, nritems + 1);
2256         btrfs_mark_buffer_dirty(leaf);
2257
2258         ret = 0;
2259         if (slot == 0)
2260                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2261
2262         if (btrfs_leaf_free_space(root, leaf) < 0) {
2263                 btrfs_print_leaf(root, leaf);
2264                 BUG();
2265         }
2266 out:
2267         return ret;
2268 }
2269
2270 /*
2271  * Given a key and some data, insert an item into the tree.
2272  * This does all the path init required, making room in the tree if needed.
2273  */
2274 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2275                       *root, struct btrfs_key *cpu_key, void *data, u32
2276                       data_size)
2277 {
2278         int ret = 0;
2279         struct btrfs_path *path;
2280         struct extent_buffer *leaf;
2281         unsigned long ptr;
2282
2283         path = btrfs_alloc_path();
2284         BUG_ON(!path);
2285         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2286         if (!ret) {
2287                 leaf = path->nodes[0];
2288                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2289                 write_extent_buffer(leaf, data, ptr, data_size);
2290                 btrfs_mark_buffer_dirty(leaf);
2291         }
2292         btrfs_free_path(path);
2293         return ret;
2294 }
2295
2296 /*
2297  * delete the pointer from a given node.
2298  *
2299  * If the delete empties a node, the node is removed from the tree,
2300  * continuing all the way the root if required.  The root is converted into
2301  * a leaf if all the nodes are emptied.
2302  */
2303 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2304                    struct btrfs_path *path, int level, int slot)
2305 {
2306         struct extent_buffer *parent = path->nodes[level];
2307         u32 nritems;
2308         int ret = 0;
2309         int wret;
2310
2311         nritems = btrfs_header_nritems(parent);
2312         if (slot != nritems -1) {
2313                 memmove_extent_buffer(parent,
2314                               btrfs_node_key_ptr_offset(slot),
2315                               btrfs_node_key_ptr_offset(slot + 1),
2316                               sizeof(struct btrfs_key_ptr) *
2317                               (nritems - slot - 1));
2318         }
2319         nritems--;
2320         btrfs_set_header_nritems(parent, nritems);
2321         if (nritems == 0 && parent == root->node) {
2322                 BUG_ON(btrfs_header_level(root->node) != 1);
2323                 /* just turn the root into a leaf and break */
2324                 btrfs_set_header_level(root->node, 0);
2325         } else if (slot == 0) {
2326                 struct btrfs_disk_key disk_key;
2327
2328                 btrfs_node_key(parent, &disk_key, 0);
2329                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2330                 if (wret)
2331                         ret = wret;
2332         }
2333         btrfs_mark_buffer_dirty(parent);
2334         return ret;
2335 }
2336
2337 /*
2338  * delete the item at the leaf level in path.  If that empties
2339  * the leaf, remove it from the tree
2340  */
2341 int btrfs_del_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2342                    struct btrfs_path *path)
2343 {
2344         int slot;
2345         struct extent_buffer *leaf;
2346         struct btrfs_item *item;
2347         int doff;
2348         int dsize;
2349         int ret = 0;
2350         int wret;
2351         u32 nritems;
2352
2353         leaf = path->nodes[0];
2354         slot = path->slots[0];
2355         doff = btrfs_item_offset_nr(leaf, slot);
2356         dsize = btrfs_item_size_nr(leaf, slot);
2357         nritems = btrfs_header_nritems(leaf);
2358
2359         if (slot != nritems - 1) {
2360                 int i;
2361                 int data_end = leaf_data_end(root, leaf);
2362
2363                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2364                               data_end + dsize,
2365                               btrfs_leaf_data(leaf) + data_end,
2366                               doff - data_end);
2367
2368                 for (i = slot + 1; i < nritems; i++) {
2369                         u32 ioff;
2370
2371                         item = btrfs_item_nr(leaf, i);
2372                         if (!leaf->map_token) {
2373                                 map_extent_buffer(leaf, (unsigned long)item,
2374                                         sizeof(struct btrfs_item),
2375                                         &leaf->map_token, &leaf->kaddr,
2376                                         &leaf->map_start, &leaf->map_len,
2377                                         KM_USER1);
2378                         }
2379                         ioff = btrfs_item_offset(leaf, item);
2380                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2381                 }
2382
2383                 if (leaf->map_token) {
2384                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2385                         leaf->map_token = NULL;
2386                 }
2387
2388                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2389                               btrfs_item_nr_offset(slot + 1),
2390                               sizeof(struct btrfs_item) *
2391                               (nritems - slot - 1));
2392         }
2393         btrfs_set_header_nritems(leaf, nritems - 1);
2394         nritems--;
2395
2396         /* delete the leaf if we've emptied it */
2397         if (nritems == 0) {
2398                 if (leaf == root->node) {
2399                         btrfs_set_header_level(leaf, 0);
2400                 } else {
2401                         clean_tree_block(trans, root, leaf);
2402                         wait_on_tree_block_writeback(root, leaf);
2403                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
2404                         if (wret)
2405                                 ret = wret;
2406                         wret = btrfs_free_extent(trans, root,
2407                                                  leaf->start, leaf->len, 1);
2408                         if (wret)
2409                                 ret = wret;
2410                 }
2411         } else {
2412                 int used = leaf_space_used(leaf, 0, nritems);
2413                 if (slot == 0) {
2414                         struct btrfs_disk_key disk_key;
2415
2416                         btrfs_item_key(leaf, &disk_key, 0);
2417                         wret = fixup_low_keys(trans, root, path,
2418                                               &disk_key, 1);
2419                         if (wret)
2420                                 ret = wret;
2421                 }
2422
2423                 /* delete the leaf if it is mostly empty */
2424                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 3) {
2425                         /* push_leaf_left fixes the path.
2426                          * make sure the path still points to our leaf
2427                          * for possible call to del_ptr below
2428                          */
2429                         slot = path->slots[1];
2430                         extent_buffer_get(leaf);
2431
2432                         wret = push_leaf_right(trans, root, path, 1, 1);
2433                         if (wret < 0 && wret != -ENOSPC)
2434                                 ret = wret;
2435
2436                         if (path->nodes[0] == leaf &&
2437                             btrfs_header_nritems(leaf)) {
2438                                 wret = push_leaf_left(trans, root, path, 1, 1);
2439                                 if (wret < 0 && wret != -ENOSPC)
2440                                         ret = wret;
2441                         }
2442
2443                         if (btrfs_header_nritems(leaf) == 0) {
2444                                 u64 bytenr = leaf->start;
2445                                 u32 blocksize = leaf->len;
2446
2447                                 clean_tree_block(trans, root, leaf);
2448                                 wait_on_tree_block_writeback(root, leaf);
2449
2450                                 wret = del_ptr(trans, root, path, 1, slot);
2451                                 if (wret)
2452                                         ret = wret;
2453
2454                                 free_extent_buffer(leaf);
2455                                 wret = btrfs_free_extent(trans, root, bytenr,
2456                                                          blocksize, 1);
2457                                 if (wret)
2458                                         ret = wret;
2459                         } else {
2460                                 btrfs_mark_buffer_dirty(leaf);
2461                                 free_extent_buffer(leaf);
2462                         }
2463                 } else {
2464                         btrfs_mark_buffer_dirty(leaf);
2465                 }
2466         }
2467         return ret;
2468 }
2469
2470 /*
2471  * walk up the tree as far as required to find the next leaf.
2472  * returns 0 if it found something or 1 if there are no greater leaves.
2473  * returns < 0 on io errors.
2474  */
2475 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2476 {
2477         int slot;
2478         int level = 1;
2479         u64 bytenr;
2480         struct extent_buffer *c;
2481         struct extent_buffer *next = NULL;
2482
2483         while(level < BTRFS_MAX_LEVEL) {
2484                 if (!path->nodes[level])
2485                         return 1;
2486
2487                 slot = path->slots[level] + 1;
2488                 c = path->nodes[level];
2489                 if (slot >= btrfs_header_nritems(c)) {
2490                         level++;
2491                         continue;
2492                 }
2493
2494                 bytenr = btrfs_node_blockptr(c, slot);
2495                 if (next)
2496                         free_extent_buffer(next);
2497
2498                 if (path->reada)
2499                         reada_for_search(root, path, level, slot);
2500
2501                 next = read_tree_block(root, bytenr,
2502                                        btrfs_level_size(root, level -1));
2503                 break;
2504         }
2505         path->slots[level] = slot;
2506         while(1) {
2507                 level--;
2508                 c = path->nodes[level];
2509                 free_extent_buffer(c);
2510                 path->nodes[level] = next;
2511                 path->slots[level] = 0;
2512                 if (!level)
2513                         break;
2514                 if (path->reada)
2515                         reada_for_search(root, path, level, 0);
2516                 next = read_tree_block(root, btrfs_node_blockptr(next, 0),
2517                                        btrfs_level_size(root, level - 1));
2518         }
2519         return 0;
2520 }