fs/btrfs/ctree.c

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