fs/btrfs/ctree.c

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