fs/btrfs/ctree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/sched.h>
  20 #include "ctree.h"
  21 #include "disk-io.h"
  22 #include "transaction.h"
  23 #include "print-tree.h"
  24 #include "locking.h"
  25
  26 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  27                       *root, struct btrfs_path *path, int level);
  28 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  29                       *root, struct btrfs_key *ins_key,
  30                       struct btrfs_path *path, int data_size, int extend);
  31 static int push_node_left(struct btrfs_trans_handle *trans,
  32                           struct btrfs_root *root, struct extent_buffer *dst,
  33                           struct extent_buffer *src, int empty);
  34 static int balance_node_right(struct btrfs_trans_handle *trans,
  35                               struct btrfs_root *root,
  36                               struct extent_buffer *dst_buf,
  37                               struct extent_buffer *src_buf);
  38 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  39                    struct btrfs_path *path, int level, int slot);
  40
  41 inline void btrfs_init_path(struct btrfs_path *p)
  42 {
  43         memset(p, 0, sizeof(*p));
  44 }
  45
  46 struct btrfs_path *btrfs_alloc_path(void)
  47 {
  48         struct btrfs_path *path;
  49         path = kmem_cache_alloc(btrfs_path_cachep, GFP_NOFS);
  50         if (path) {
  51                 btrfs_init_path(path);
  52                 path->reada = 1;
  53         }
  54         return path;
  55 }
  56
  57 void btrfs_free_path(struct btrfs_path *p)
  58 {
  59         btrfs_release_path(NULL, p);
  60         kmem_cache_free(btrfs_path_cachep, p);
  61 }
  62
  63 void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
  64 {
  65         int i;
  66         int keep = p->keep_locks;
  67
  68         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  69                 if (!p->nodes[i])
  70                         continue;
  71                 if (p->locks[i]) {
  72                         btrfs_tree_unlock(p->nodes[i]);
  73                         p->locks[i] = 0;
  74                 }
  75                 free_extent_buffer(p->nodes[i]);
  76         }
  77         memset(p, 0, sizeof(*p));
  78         p->keep_locks = keep;
  79 }
  80
  81 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
  82 {
  83         struct extent_buffer *eb;
  84         spin_lock(&root->node_lock);
  85         eb = root->node;
  86         extent_buffer_get(eb);
  87         spin_unlock(&root->node_lock);
  88         return eb;
  89 }
  90
  91 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
  92 {
  93         struct extent_buffer *eb;
  94
  95         while(1) {
  96                 eb = btrfs_root_node(root);
  97                 btrfs_tree_lock(eb);
  98
  99                 spin_lock(&root->node_lock);
 100                 if (eb == root->node) {
 101                         spin_unlock(&root->node_lock);
 102                         break;
 103                 }
 104                 spin_unlock(&root->node_lock);
 105
 106                 btrfs_tree_unlock(eb);
 107                 free_extent_buffer(eb);
 108         }
 109         return eb;
 110 }
 111
 112 static void add_root_to_dirty_list(struct btrfs_root *root)
 113 {
 114         if (root->track_dirty && list_empty(&root->dirty_list)) {
 115                 list_add(&root->dirty_list,
 116                          &root->fs_info->dirty_cowonly_roots);
 117         }
 118 }
 119
 120 int btrfs_copy_root(struct btrfs_trans_handle *trans,
 121                       struct btrfs_root *root,
 122                       struct extent_buffer *buf,
 123                       struct extent_buffer **cow_ret, u64 new_root_objectid)
 124 {
 125         struct extent_buffer *cow;
 126         u32 nritems;
 127         int ret = 0;
 128         int level;
 129         struct btrfs_key first_key;
 130         struct btrfs_root *new_root;
 131
 132         new_root = kmalloc(sizeof(*new_root), GFP_NOFS);
 133         if (!new_root)
 134                 return -ENOMEM;
 135
 136         memcpy(new_root, root, sizeof(*new_root));
 137         new_root->root_key.objectid = new_root_objectid;
 138
 139         WARN_ON(root->ref_cows && trans->transid !=
 140                 root->fs_info->running_transaction->transid);
 141         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 142
 143         level = btrfs_header_level(buf);
 144         nritems = btrfs_header_nritems(buf);
 145         if (nritems) {
 146                 if (level == 0)
 147                         btrfs_item_key_to_cpu(buf, &first_key, 0);
 148                 else
 149                         btrfs_node_key_to_cpu(buf, &first_key, 0);
 150         } else {
 151                 first_key.objectid = 0;
 152         }
 153         cow = btrfs_alloc_free_block(trans, new_root, buf->len,
 154                                        new_root_objectid,
 155                                        trans->transid, first_key.objectid,
 156                                        level, buf->start, 0);
 157         if (IS_ERR(cow)) {
 158                 kfree(new_root);
 159                 return PTR_ERR(cow);
 160         }
 161
 162         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 163         btrfs_set_header_bytenr(cow, cow->start);
 164         btrfs_set_header_generation(cow, trans->transid);
 165         btrfs_set_header_owner(cow, new_root_objectid);
 166         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN);
 167
 168         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 169         ret = btrfs_inc_ref(trans, new_root, buf);
 170         kfree(new_root);
 171
 172         if (ret)
 173                 return ret;
 174
 175         btrfs_mark_buffer_dirty(cow);
 176         *cow_ret = cow;
 177         return 0;
 178 }
 179
 180 int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 181                              struct btrfs_root *root,
 182                              struct extent_buffer *buf,
 183                              struct extent_buffer *parent, int parent_slot,
 184                              struct extent_buffer **cow_ret,
 185                              u64 search_start, u64 empty_size)
 186 {
 187         u64 root_gen;
 188         struct extent_buffer *cow;
 189         u32 nritems;
 190         int ret = 0;
 191         int different_trans = 0;
 192         int level;
 193         int unlock_orig = 0;
 194         struct btrfs_key first_key;
 195
 196         if (*cow_ret == buf)
 197                 unlock_orig = 1;
 198
 199         WARN_ON(!btrfs_tree_locked(buf));
 200
 201         if (root->ref_cows) {
 202                 root_gen = trans->transid;
 203         } else {
 204                 root_gen = 0;
 205         }
 206         WARN_ON(root->ref_cows && trans->transid !=
 207                 root->fs_info->running_transaction->transid);
 208         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 209
 210         level = btrfs_header_level(buf);
 211         nritems = btrfs_header_nritems(buf);
 212         if (nritems) {
 213                 if (level == 0)
 214                         btrfs_item_key_to_cpu(buf, &first_key, 0);
 215                 else
 216                         btrfs_node_key_to_cpu(buf, &first_key, 0);
 217         } else {
 218                 first_key.objectid = 0;
 219         }
 220         cow = btrfs_alloc_free_block(trans, root, buf->len,
 221                                      root->root_key.objectid,
 222                                      root_gen, first_key.objectid, level,
 223                                      search_start, empty_size);
 224         if (IS_ERR(cow))
 225                 return PTR_ERR(cow);
 226
 227         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 228         btrfs_set_header_bytenr(cow, cow->start);
 229         btrfs_set_header_generation(cow, trans->transid);
 230         btrfs_set_header_owner(cow, root->root_key.objectid);
 231         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN);
 232
 233         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 234         if (btrfs_header_generation(buf) != trans->transid) {
 235                 different_trans = 1;
 236                 ret = btrfs_inc_ref(trans, root, buf);
 237                 if (ret)
 238                         return ret;
 239         } else {
 240                 clean_tree_block(trans, root, buf);
 241         }
 242
 243         if (buf == root->node) {
 244                 WARN_ON(parent && parent != buf);
 245                 root_gen = btrfs_header_generation(buf);
 246
 247                 spin_lock(&root->node_lock);
 248                 root->node = cow;
 249                 extent_buffer_get(cow);
 250                 spin_unlock(&root->node_lock);
 251
 252                 if (buf != root->commit_root) {
 253                         btrfs_free_extent(trans, root, buf->start,
 254                                           buf->len, root->root_key.objectid,
 255                                           root_gen, 0, 0, 1);
 256                 }
 257                 free_extent_buffer(buf);
 258                 add_root_to_dirty_list(root);
 259         } else {
 260                 root_gen = btrfs_header_generation(parent);
 261                 btrfs_set_node_blockptr(parent, parent_slot,
 262                                         cow->start);
 263                 WARN_ON(trans->transid == 0);
 264                 btrfs_set_node_ptr_generation(parent, parent_slot,
 265                                               trans->transid);
 266                 btrfs_mark_buffer_dirty(parent);
 267                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 268                 btrfs_free_extent(trans, root, buf->start, buf->len,
 269                                   btrfs_header_owner(parent), root_gen,
 270                                   0, 0, 1);
 271         }
 272         if (unlock_orig)
 273                 btrfs_tree_unlock(buf);
 274         free_extent_buffer(buf);
 275         btrfs_mark_buffer_dirty(cow);
 276         *cow_ret = cow;
 277         return 0;
 278 }
 279
 280 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 281                     struct btrfs_root *root, struct extent_buffer *buf,
 282                     struct extent_buffer *parent, int parent_slot,
 283                     struct extent_buffer **cow_ret)
 284 {
 285         u64 search_start;
 286         u64 header_trans;
 287         int ret;
 288
 289         if (trans->transaction != root->fs_info->running_transaction) {
 290                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 291                        root->fs_info->running_transaction->transid);
 292                 WARN_ON(1);
 293         }
 294         if (trans->transid != root->fs_info->generation) {
 295                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 296                        root->fs_info->generation);
 297                 WARN_ON(1);
 298         }
 299
 300         header_trans = btrfs_header_generation(buf);
 301         spin_lock(&root->fs_info->hash_lock);
 302         if (header_trans == trans->transid &&
 303             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
 304                 *cow_ret = buf;
 305                 spin_unlock(&root->fs_info->hash_lock);
 306                 return 0;
 307         }
 308         spin_unlock(&root->fs_info->hash_lock);
 309         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
 310         ret = __btrfs_cow_block(trans, root, buf, parent,
 311                                  parent_slot, cow_ret, search_start, 0);
 312         return ret;
 313 }
 314
 315 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 316 {
 317         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 318                 return 1;
 319         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 320                 return 1;
 321         return 0;
 322 }
 323
 324 /*
 325  * compare two keys in a memcmp fashion
 326  */
 327 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 328 {
 329         struct btrfs_key k1;
 330
 331         btrfs_disk_key_to_cpu(&k1, disk);
 332
 333         if (k1.objectid > k2->objectid)
 334                 return 1;
 335         if (k1.objectid < k2->objectid)
 336                 return -1;
 337         if (k1.type > k2->type)
 338                 return 1;
 339         if (k1.type < k2->type)
 340                 return -1;
 341         if (k1.offset > k2->offset)
 342                 return 1;
 343         if (k1.offset < k2->offset)
 344                 return -1;
 345         return 0;
 346 }
 347
 348
 349 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 350                        struct btrfs_root *root, struct extent_buffer *parent,
 351                        int start_slot, int cache_only, u64 *last_ret,
 352                        struct btrfs_key *progress)
 353 {
 354         struct extent_buffer *cur;
 355         struct extent_buffer *tmp;
 356         u64 blocknr;
 357         u64 gen;
 358         u64 search_start = *last_ret;
 359         u64 last_block = 0;
 360         u64 other;
 361         u32 parent_nritems;
 362         int end_slot;
 363         int i;
 364         int err = 0;
 365         int parent_level;
 366         int uptodate;
 367         u32 blocksize;
 368         int progress_passed = 0;
 369         struct btrfs_disk_key disk_key;
 370
 371         /* FIXME this code needs locking */
 372         return 0;
 373
 374         parent_level = btrfs_header_level(parent);
 375         if (cache_only && parent_level != 1)
 376                 return 0;
 377
 378         if (trans->transaction != root->fs_info->running_transaction) {
 379                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 380                        root->fs_info->running_transaction->transid);
 381                 WARN_ON(1);
 382         }
 383         if (trans->transid != root->fs_info->generation) {
 384                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 385                        root->fs_info->generation);
 386                 WARN_ON(1);
 387         }
 388
 389         parent_nritems = btrfs_header_nritems(parent);
 390         blocksize = btrfs_level_size(root, parent_level - 1);
 391         end_slot = parent_nritems;
 392
 393         if (parent_nritems == 1)
 394                 return 0;
 395
 396         for (i = start_slot; i < end_slot; i++) {
 397                 int close = 1;
 398
 399                 if (!parent->map_token) {
 400                         map_extent_buffer(parent,
 401                                         btrfs_node_key_ptr_offset(i),
 402                                         sizeof(struct btrfs_key_ptr),
 403                                         &parent->map_token, &parent->kaddr,
 404                                         &parent->map_start, &parent->map_len,
 405                                         KM_USER1);
 406                 }
 407                 btrfs_node_key(parent, &disk_key, i);
 408                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
 409                         continue;
 410
 411                 progress_passed = 1;
 412                 blocknr = btrfs_node_blockptr(parent, i);
 413                 gen = btrfs_node_ptr_generation(parent, i);
 414                 if (last_block == 0)
 415                         last_block = blocknr;
 416
 417                 if (i > 0) {
 418                         other = btrfs_node_blockptr(parent, i - 1);
 419                         close = close_blocks(blocknr, other, blocksize);
 420                 }
 421                 if (!close && i < end_slot - 2) {
 422                         other = btrfs_node_blockptr(parent, i + 1);
 423                         close = close_blocks(blocknr, other, blocksize);
 424                 }
 425                 if (close) {
 426                         last_block = blocknr;
 427                         continue;
 428                 }
 429                 if (parent->map_token) {
 430                         unmap_extent_buffer(parent, parent->map_token,
 431                                             KM_USER1);
 432                         parent->map_token = NULL;
 433                 }
 434
 435                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
 436                 if (cur)
 437                         uptodate = btrfs_buffer_uptodate(cur, gen);
 438                 else
 439                         uptodate = 0;
 440                 if (!cur || !uptodate) {
 441                         if (cache_only) {
 442                                 free_extent_buffer(cur);
 443                                 continue;
 444                         }
 445                         if (!cur) {
 446                                 cur = read_tree_block(root, blocknr,
 447                                                          blocksize, gen);
 448                         } else if (!uptodate) {
 449                                 btrfs_read_buffer(cur, gen);
 450                         }
 451                 }
 452                 if (search_start == 0)
 453                         search_start = last_block;
 454
 455                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 456                                         &tmp, search_start,
 457                                         min(16 * blocksize,
 458                                             (end_slot - i) * blocksize));
 459                 if (err) {
 460                         free_extent_buffer(cur);
 461                         break;
 462                 }
 463                 search_start = tmp->start;
 464                 last_block = tmp->start;
 465                 *last_ret = search_start;
 466                 if (parent_level == 1)
 467                         btrfs_clear_buffer_defrag(tmp);
 468                 free_extent_buffer(tmp);
 469         }
 470         if (parent->map_token) {
 471                 unmap_extent_buffer(parent, parent->map_token,
 472                                     KM_USER1);
 473                 parent->map_token = NULL;
 474         }
 475         return err;
 476 }
 477
 478 /*
 479  * The leaf data grows from end-to-front in the node.
 480  * this returns the address of the start of the last item,
 481  * which is the stop of the leaf data stack
 482  */
 483 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 484                                          struct extent_buffer *leaf)
 485 {
 486         u32 nr = btrfs_header_nritems(leaf);
 487         if (nr == 0)
 488                 return BTRFS_LEAF_DATA_SIZE(root);
 489         return btrfs_item_offset_nr(leaf, nr - 1);
 490 }
 491
 492 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
 493                       int level)
 494 {
 495         struct extent_buffer *parent = NULL;
 496         struct extent_buffer *node = path->nodes[level];
 497         struct btrfs_disk_key parent_key;
 498         struct btrfs_disk_key node_key;
 499         int parent_slot;
 500         int slot;
 501         struct btrfs_key cpukey;
 502         u32 nritems = btrfs_header_nritems(node);
 503
 504         if (path->nodes[level + 1])
 505                 parent = path->nodes[level + 1];
 506
 507         slot = path->slots[level];
 508         BUG_ON(nritems == 0);
 509         if (parent) {
 510                 parent_slot = path->slots[level + 1];
 511                 btrfs_node_key(parent, &parent_key, parent_slot);
 512                 btrfs_node_key(node, &node_key, 0);
 513                 BUG_ON(memcmp(&parent_key, &node_key,
 514                               sizeof(struct btrfs_disk_key)));
 515                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 516                        btrfs_header_bytenr(node));
 517         }
 518         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 519         if (slot != 0) {
 520                 btrfs_node_key_to_cpu(node, &cpukey, slot - 1);
 521                 btrfs_node_key(node, &node_key, slot);
 522                 BUG_ON(comp_keys(&node_key, &cpukey) <= 0);
 523         }
 524         if (slot < nritems - 1) {
 525                 btrfs_node_key_to_cpu(node, &cpukey, slot + 1);
 526                 btrfs_node_key(node, &node_key, slot);
 527                 BUG_ON(comp_keys(&node_key, &cpukey) >= 0);
 528         }
 529         return 0;
 530 }
 531
 532 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 533                       int level)
 534 {
 535         struct extent_buffer *leaf = path->nodes[level];
 536         struct extent_buffer *parent = NULL;
 537         int parent_slot;
 538         struct btrfs_key cpukey;
 539         struct btrfs_disk_key parent_key;
 540         struct btrfs_disk_key leaf_key;
 541         int slot = path->slots[0];
 542
 543         u32 nritems = btrfs_header_nritems(leaf);
 544
 545         if (path->nodes[level + 1])
 546                 parent = path->nodes[level + 1];
 547
 548         if (nritems == 0)
 549                 return 0;
 550
 551         if (parent) {
 552                 parent_slot = path->slots[level + 1];
 553                 btrfs_node_key(parent, &parent_key, parent_slot);
 554                 btrfs_item_key(leaf, &leaf_key, 0);
 555
 556                 BUG_ON(memcmp(&parent_key, &leaf_key,
 557                        sizeof(struct btrfs_disk_key)));
 558                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 559                        btrfs_header_bytenr(leaf));
 560         }
 561 #if 0
 562         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 563                 btrfs_item_key_to_cpu(leaf, &cpukey, i + 1);
 564                 btrfs_item_key(leaf, &leaf_key, i);
 565                 if (comp_keys(&leaf_key, &cpukey) >= 0) {
 566                         btrfs_print_leaf(root, leaf);
 567                         printk("slot %d offset bad key\n", i);
 568                         BUG_ON(1);
 569                 }
 570                 if (btrfs_item_offset_nr(leaf, i) !=
 571                         btrfs_item_end_nr(leaf, i + 1)) {
 572                         btrfs_print_leaf(root, leaf);
 573                         printk("slot %d offset bad\n", i);
 574                         BUG_ON(1);
 575                 }
 576                 if (i == 0) {
 577                         if (btrfs_item_offset_nr(leaf, i) +
 578                                btrfs_item_size_nr(leaf, i) !=
 579                                BTRFS_LEAF_DATA_SIZE(root)) {
 580                                 btrfs_print_leaf(root, leaf);
 581                                 printk("slot %d first offset bad\n", i);
 582                                 BUG_ON(1);
 583                         }
 584                 }
 585         }
 586         if (nritems > 0) {
 587                 if (btrfs_item_size_nr(leaf, nritems - 1) > 4096) {
 588                                 btrfs_print_leaf(root, leaf);
 589                                 printk("slot %d bad size \n", nritems - 1);
 590                                 BUG_ON(1);
 591                 }
 592         }
 593 #endif
 594         if (slot != 0 && slot < nritems - 1) {
 595                 btrfs_item_key(leaf, &leaf_key, slot);
 596                 btrfs_item_key_to_cpu(leaf, &cpukey, slot - 1);
 597                 if (comp_keys(&leaf_key, &cpukey) <= 0) {
 598                         btrfs_print_leaf(root, leaf);
 599                         printk("slot %d offset bad key\n", slot);
 600                         BUG_ON(1);
 601                 }
 602                 if (btrfs_item_offset_nr(leaf, slot - 1) !=
 603                        btrfs_item_end_nr(leaf, slot)) {
 604                         btrfs_print_leaf(root, leaf);
 605                         printk("slot %d offset bad\n", slot);
 606                         BUG_ON(1);
 607                 }
 608         }
 609         if (slot < nritems - 1) {
 610                 btrfs_item_key(leaf, &leaf_key, slot);
 611                 btrfs_item_key_to_cpu(leaf, &cpukey, slot + 1);
 612                 BUG_ON(comp_keys(&leaf_key, &cpukey) >= 0);
 613                 if (btrfs_item_offset_nr(leaf, slot) !=
 614                         btrfs_item_end_nr(leaf, slot + 1)) {
 615                         btrfs_print_leaf(root, leaf);
 616                         printk("slot %d offset bad\n", slot);
 617                         BUG_ON(1);
 618                 }
 619         }
 620         BUG_ON(btrfs_item_offset_nr(leaf, 0) +
 621                btrfs_item_size_nr(leaf, 0) != BTRFS_LEAF_DATA_SIZE(root));
 622         return 0;
 623 }
 624
 625 static int noinline check_block(struct btrfs_root *root,
 626                                 struct btrfs_path *path, int level)
 627 {
 628         u64 found_start;
 629         return 0;
 630         if (btrfs_header_level(path->nodes[level]) != level)
 631             printk("warning: bad level %Lu wanted %d found %d\n",
 632                    path->nodes[level]->start, level,
 633                    btrfs_header_level(path->nodes[level]));
 634         found_start = btrfs_header_bytenr(path->nodes[level]);
 635         if (found_start != path->nodes[level]->start) {
 636             printk("warning: bad bytentr %Lu found %Lu\n",
 637                    path->nodes[level]->start, found_start);
 638         }
 639 #if 0
 640         struct extent_buffer *buf = path->nodes[level];
 641
 642         if (memcmp_extent_buffer(buf, root->fs_info->fsid,
 643                                  (unsigned long)btrfs_header_fsid(buf),
 644                                  BTRFS_FSID_SIZE)) {
 645                 printk("warning bad block %Lu\n", buf->start);
 646                 return 1;
 647         }
 648 #endif
 649         if (level == 0)
 650                 return check_leaf(root, path, level);
 651         return check_node(root, path, level);
 652 }
 653
 654 /*
 655  * search for key in the extent_buffer.  The items start at offset p,
 656  * and they are item_size apart.  There are 'max' items in p.
 657  *
 658  * the slot in the array is returned via slot, and it points to
 659  * the place where you would insert key if it is not found in
 660  * the array.
 661  *
 662  * slot may point to max if the key is bigger than all of the keys
 663  */
 664 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 665                               int item_size, struct btrfs_key *key,
 666                               int max, int *slot)
 667 {
 668         int low = 0;
 669         int high = max;
 670         int mid;
 671         int ret;
 672         struct btrfs_disk_key *tmp = NULL;
 673         struct btrfs_disk_key unaligned;
 674         unsigned long offset;
 675         char *map_token = NULL;
 676         char *kaddr = NULL;
 677         unsigned long map_start = 0;
 678         unsigned long map_len = 0;
 679         int err;
 680
 681         while(low < high) {
 682                 mid = (low + high) / 2;
 683                 offset = p + mid * item_size;
 684
 685                 if (!map_token || offset < map_start ||
 686                     (offset + sizeof(struct btrfs_disk_key)) >
 687                     map_start + map_len) {
 688                         if (map_token) {
 689                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 690                                 map_token = NULL;
 691                         }
 692                         err = map_extent_buffer(eb, offset,
 693                                                 sizeof(struct btrfs_disk_key),
 694                                                 &map_token, &kaddr,
 695                                                 &map_start, &map_len, KM_USER0);
 696
 697                         if (!err) {
 698                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
 699                                                         map_start);
 700                         } else {
 701                                 read_extent_buffer(eb, &unaligned,
 702                                                    offset, sizeof(unaligned));
 703                                 tmp = &unaligned;
 704                         }
 705
 706                 } else {
 707                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
 708                                                         map_start);
 709                 }
 710                 ret = comp_keys(tmp, key);
 711
 712                 if (ret < 0)
 713                         low = mid + 1;
 714                 else if (ret > 0)
 715                         high = mid;
 716                 else {
 717                         *slot = mid;
 718                         if (map_token)
 719                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 720                         return 0;
 721                 }
 722         }
 723         *slot = low;
 724         if (map_token)
 725                 unmap_extent_buffer(eb, map_token, KM_USER0);
 726         return 1;
 727 }
 728
 729 /*
 730  * simple bin_search frontend that does the right thing for
 731  * leaves vs nodes
 732  */
 733 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 734                       int level, int *slot)
 735 {
 736         if (level == 0) {
 737                 return generic_bin_search(eb,
 738                                           offsetof(struct btrfs_leaf, items),
 739                                           sizeof(struct btrfs_item),
 740                                           key, btrfs_header_nritems(eb),
 741                                           slot);
 742         } else {
 743                 return generic_bin_search(eb,
 744                                           offsetof(struct btrfs_node, ptrs),
 745                                           sizeof(struct btrfs_key_ptr),
 746                                           key, btrfs_header_nritems(eb),
 747                                           slot);
 748         }
 749         return -1;
 750 }
 751
 752 static struct extent_buffer *read_node_slot(struct btrfs_root *root,
 753                                    struct extent_buffer *parent, int slot)
 754 {
 755         int level = btrfs_header_level(parent);
 756         if (slot < 0)
 757                 return NULL;
 758         if (slot >= btrfs_header_nritems(parent))
 759                 return NULL;
 760
 761         BUG_ON(level == 0);
 762
 763         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 764                        btrfs_level_size(root, level - 1),
 765                        btrfs_node_ptr_generation(parent, slot));
 766 }
 767
 768 static int balance_level(struct btrfs_trans_handle *trans,
 769                          struct btrfs_root *root,
 770                          struct btrfs_path *path, int level)
 771 {
 772         struct extent_buffer *right = NULL;
 773         struct extent_buffer *mid;
 774         struct extent_buffer *left = NULL;
 775         struct extent_buffer *parent = NULL;
 776         int ret = 0;
 777         int wret;
 778         int pslot;
 779         int orig_slot = path->slots[level];
 780         int err_on_enospc = 0;
 781         u64 orig_ptr;
 782
 783         if (level == 0)
 784                 return 0;
 785
 786         mid = path->nodes[level];
 787         WARN_ON(!path->locks[level]);
 788         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 789
 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         /*
 797          * deal with the case where there is only one pointer in the root
 798          * by promoting the node below to a root
 799          */
 800         if (!parent) {
 801                 struct extent_buffer *child;
 802
 803                 if (btrfs_header_nritems(mid) != 1)
 804                         return 0;
 805
 806                 /* promote the child to a root */
 807                 child = read_node_slot(root, mid, 0);
 808                 btrfs_tree_lock(child);
 809                 BUG_ON(!child);
 810                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
 811                 BUG_ON(ret);
 812
 813                 spin_lock(&root->node_lock);
 814                 root->node = child;
 815                 spin_unlock(&root->node_lock);
 816
 817                 add_root_to_dirty_list(root);
 818                 btrfs_tree_unlock(child);
 819                 path->locks[level] = 0;
 820                 path->nodes[level] = NULL;
 821                 clean_tree_block(trans, root, mid);
 822                 btrfs_tree_unlock(mid);
 823                 /* once for the path */
 824                 free_extent_buffer(mid);
 825                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 826                                         root->root_key.objectid,
 827                                         btrfs_header_generation(mid), 0, 0, 1);
 828                 /* once for the root ptr */
 829                 free_extent_buffer(mid);
 830                 return ret;
 831         }
 832         if (btrfs_header_nritems(mid) >
 833             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 834                 return 0;
 835
 836         if (btrfs_header_nritems(mid) < 2)
 837                 err_on_enospc = 1;
 838
 839         left = read_node_slot(root, parent, pslot - 1);
 840         if (left) {
 841                 btrfs_tree_lock(left);
 842                 wret = btrfs_cow_block(trans, root, left,
 843                                        parent, pslot - 1, &left);
 844                 if (wret) {
 845                         ret = wret;
 846                         goto enospc;
 847                 }
 848         }
 849         right = read_node_slot(root, parent, pslot + 1);
 850         if (right) {
 851                 btrfs_tree_lock(right);
 852                 wret = btrfs_cow_block(trans, root, right,
 853                                        parent, pslot + 1, &right);
 854                 if (wret) {
 855                         ret = wret;
 856                         goto enospc;
 857                 }
 858         }
 859
 860         /* first, try to make some room in the middle buffer */
 861         if (left) {
 862                 orig_slot += btrfs_header_nritems(left);
 863                 wret = push_node_left(trans, root, left, mid, 1);
 864                 if (wret < 0)
 865                         ret = wret;
 866                 if (btrfs_header_nritems(mid) < 2)
 867                         err_on_enospc = 1;
 868         }
 869
 870         /*
 871          * then try to empty the right most buffer into the middle
 872          */
 873         if (right) {
 874                 wret = push_node_left(trans, root, mid, right, 1);
 875                 if (wret < 0 && wret != -ENOSPC)
 876                         ret = wret;
 877                 if (btrfs_header_nritems(right) == 0) {
 878                         u64 bytenr = right->start;
 879                         u64 generation = btrfs_header_generation(parent);
 880                         u32 blocksize = right->len;
 881
 882                         clean_tree_block(trans, root, right);
 883                         btrfs_tree_unlock(right);
 884                         free_extent_buffer(right);
 885                         right = NULL;
 886                         wret = del_ptr(trans, root, path, level + 1, pslot +
 887                                        1);
 888                         if (wret)
 889                                 ret = wret;
 890                         wret = btrfs_free_extent(trans, root, bytenr,
 891                                                  blocksize,
 892                                                  btrfs_header_owner(parent),
 893                                                  generation, 0, 0, 1);
 894                         if (wret)
 895                                 ret = wret;
 896                 } else {
 897                         struct btrfs_disk_key right_key;
 898                         btrfs_node_key(right, &right_key, 0);
 899                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 900                         btrfs_mark_buffer_dirty(parent);
 901                 }
 902         }
 903         if (btrfs_header_nritems(mid) == 1) {
 904                 /*
 905                  * we're not allowed to leave a node with one item in the
 906                  * tree during a delete.  A deletion from lower in the tree
 907                  * could try to delete the only pointer in this node.
 908                  * So, pull some keys from the left.
 909                  * There has to be a left pointer at this point because
 910                  * otherwise we would have pulled some pointers from the
 911                  * right
 912                  */
 913                 BUG_ON(!left);
 914                 wret = balance_node_right(trans, root, mid, left);
 915                 if (wret < 0) {
 916                         ret = wret;
 917                         goto enospc;
 918                 }
 919                 if (wret == 1) {
 920                         wret = push_node_left(trans, root, left, mid, 1);
 921                         if (wret < 0)
 922                                 ret = wret;
 923                 }
 924                 BUG_ON(wret == 1);
 925         }
 926         if (btrfs_header_nritems(mid) == 0) {
 927                 /* we've managed to empty the middle node, drop it */
 928                 u64 root_gen = btrfs_header_generation(parent);
 929                 u64 bytenr = mid->start;
 930                 u32 blocksize = mid->len;
 931
 932                 clean_tree_block(trans, root, mid);
 933                 btrfs_tree_unlock(mid);
 934                 free_extent_buffer(mid);
 935                 mid = NULL;
 936                 wret = del_ptr(trans, root, path, level + 1, pslot);
 937                 if (wret)
 938                         ret = wret;
 939                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 940                                          btrfs_header_owner(parent),
 941                                          root_gen, 0, 0, 1);
 942                 if (wret)
 943                         ret = wret;
 944         } else {
 945                 /* update the parent key to reflect our changes */
 946                 struct btrfs_disk_key mid_key;
 947                 btrfs_node_key(mid, &mid_key, 0);
 948                 btrfs_set_node_key(parent, &mid_key, pslot);
 949                 btrfs_mark_buffer_dirty(parent);
 950         }
 951
 952         /* update the path */
 953         if (left) {
 954                 if (btrfs_header_nritems(left) > orig_slot) {
 955                         extent_buffer_get(left);
 956                         /* left was locked after cow */
 957                         path->nodes[level] = left;
 958                         path->slots[level + 1] -= 1;
 959                         path->slots[level] = orig_slot;
 960                         if (mid) {
 961                                 btrfs_tree_unlock(mid);
 962                                 free_extent_buffer(mid);
 963                         }
 964                 } else {
 965                         orig_slot -= btrfs_header_nritems(left);
 966                         path->slots[level] = orig_slot;
 967                 }
 968         }
 969         /* double check we haven't messed things up */
 970         check_block(root, path, level);
 971         if (orig_ptr !=
 972             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 973                 BUG();
 974 enospc:
 975         if (right) {
 976                 btrfs_tree_unlock(right);
 977                 free_extent_buffer(right);
 978         }
 979         if (left) {
 980                 if (path->nodes[level] != left)
 981                         btrfs_tree_unlock(left);
 982                 free_extent_buffer(left);
 983         }
 984         return ret;
 985 }
 986
 987 /* returns zero if the push worked, non-zero otherwise */
 988 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
 989                                           struct btrfs_root *root,
 990                                           struct btrfs_path *path, int level)
 991 {
 992         struct extent_buffer *right = NULL;
 993         struct extent_buffer *mid;
 994         struct extent_buffer *left = NULL;
 995         struct extent_buffer *parent = NULL;
 996         int ret = 0;
 997         int wret;
 998         int pslot;
 999         int orig_slot = path->slots[level];
1000         u64 orig_ptr;
1001
1002         if (level == 0)
1003                 return 1;
1004
1005         mid = path->nodes[level];
1006         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1007         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1008
1009         if (level < BTRFS_MAX_LEVEL - 1)
1010                 parent = path->nodes[level + 1];
1011         pslot = path->slots[level + 1];
1012
1013         if (!parent)
1014                 return 1;
1015
1016         left = read_node_slot(root, parent, pslot - 1);
1017
1018         /* first, try to make some room in the middle buffer */
1019         if (left) {
1020                 u32 left_nr;
1021
1022                 btrfs_tree_lock(left);
1023                 left_nr = btrfs_header_nritems(left);
1024                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1025                         wret = 1;
1026                 } else {
1027                         ret = btrfs_cow_block(trans, root, left, parent,
1028                                               pslot - 1, &left);
1029                         if (ret)
1030                                 wret = 1;
1031                         else {
1032                                 wret = push_node_left(trans, root,
1033                                                       left, mid, 0);
1034                         }
1035                 }
1036                 if (wret < 0)
1037                         ret = wret;
1038                 if (wret == 0) {
1039                         struct btrfs_disk_key disk_key;
1040                         orig_slot += left_nr;
1041                         btrfs_node_key(mid, &disk_key, 0);
1042                         btrfs_set_node_key(parent, &disk_key, pslot);
1043                         btrfs_mark_buffer_dirty(parent);
1044                         if (btrfs_header_nritems(left) > orig_slot) {
1045                                 path->nodes[level] = left;
1046                                 path->slots[level + 1] -= 1;
1047                                 path->slots[level] = orig_slot;
1048                                 btrfs_tree_unlock(mid);
1049                                 free_extent_buffer(mid);
1050                         } else {
1051                                 orig_slot -=
1052                                         btrfs_header_nritems(left);
1053                                 path->slots[level] = orig_slot;
1054                                 btrfs_tree_unlock(left);
1055                                 free_extent_buffer(left);
1056                         }
1057                         return 0;
1058                 }
1059                 btrfs_tree_unlock(left);
1060                 free_extent_buffer(left);
1061         }
1062         right = read_node_slot(root, parent, pslot + 1);
1063
1064         /*
1065          * then try to empty the right most buffer into the middle
1066          */
1067         if (right) {
1068                 u32 right_nr;
1069                 btrfs_tree_lock(right);
1070                 right_nr = btrfs_header_nritems(right);
1071                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1072                         wret = 1;
1073                 } else {
1074                         ret = btrfs_cow_block(trans, root, right,
1075                                               parent, pslot + 1,
1076                                               &right);
1077                         if (ret)
1078                                 wret = 1;
1079                         else {
1080                                 wret = balance_node_right(trans, root,
1081                                                           right, mid);
1082                         }
1083                 }
1084                 if (wret < 0)
1085                         ret = wret;
1086                 if (wret == 0) {
1087                         struct btrfs_disk_key disk_key;
1088
1089                         btrfs_node_key(right, &disk_key, 0);
1090                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
1091                         btrfs_mark_buffer_dirty(parent);
1092
1093                         if (btrfs_header_nritems(mid) <= orig_slot) {
1094                                 path->nodes[level] = right;
1095                                 path->slots[level + 1] += 1;
1096                                 path->slots[level] = orig_slot -
1097                                         btrfs_header_nritems(mid);
1098                                 btrfs_tree_unlock(mid);
1099                                 free_extent_buffer(mid);
1100                         } else {
1101                                 btrfs_tree_unlock(right);
1102                                 free_extent_buffer(right);
1103                         }
1104                         return 0;
1105                 }
1106                 btrfs_tree_unlock(right);
1107                 free_extent_buffer(right);
1108         }
1109         return 1;
1110 }
1111
1112 /*
1113  * readahead one full node of leaves
1114  */
1115 static void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
1116                              int level, int slot, u64 objectid)
1117 {
1118         struct extent_buffer *node;
1119         struct btrfs_disk_key disk_key;
1120         u32 nritems;
1121         u64 search;
1122         u64 lowest_read;
1123         u64 highest_read;
1124         u64 nread = 0;
1125         int direction = path->reada;
1126         struct extent_buffer *eb;
1127         u32 nr;
1128         u32 blocksize;
1129         u32 nscan = 0;
1130
1131         if (level != 1)
1132                 return;
1133
1134         if (!path->nodes[level])
1135                 return;
1136
1137         node = path->nodes[level];
1138
1139         search = btrfs_node_blockptr(node, slot);
1140         blocksize = btrfs_level_size(root, level - 1);
1141         eb = btrfs_find_tree_block(root, search, blocksize);
1142         if (eb) {
1143                 free_extent_buffer(eb);
1144                 return;
1145         }
1146
1147         highest_read = search;
1148         lowest_read = search;
1149
1150         nritems = btrfs_header_nritems(node);
1151         nr = slot;
1152         while(1) {
1153                 if (direction < 0) {
1154                         if (nr == 0)
1155                                 break;
1156                         nr--;
1157                 } else if (direction > 0) {
1158                         nr++;
1159                         if (nr >= nritems)
1160                                 break;
1161                 }
1162                 if (path->reada < 0 && objectid) {
1163                         btrfs_node_key(node, &disk_key, nr);
1164                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1165                                 break;
1166                 }
1167                 search = btrfs_node_blockptr(node, nr);
1168                 if ((search >= lowest_read && search <= highest_read) ||
1169                     (search < lowest_read && lowest_read - search <= 32768) ||
1170                     (search > highest_read && search - highest_read <= 32768)) {
1171                         readahead_tree_block(root, search, blocksize,
1172                                      btrfs_node_ptr_generation(node, nr));
1173                         nread += blocksize;
1174                 }
1175                 nscan++;
1176                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
1177                         break;
1178                 if(nread > (1024 * 1024) || nscan > 128)
1179                         break;
1180
1181                 if (search < lowest_read)
1182                         lowest_read = search;
1183                 if (search > highest_read)
1184                         highest_read = search;
1185         }
1186 }
1187
1188 static void unlock_up(struct btrfs_path *path, int level, int lowest_unlock)
1189 {
1190         int i;
1191         int skip_level = level;
1192         int no_skips = 0;
1193         struct extent_buffer *t;
1194
1195         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1196                 if (!path->nodes[i])
1197                         break;
1198                 if (!path->locks[i])
1199                         break;
1200                 if (!no_skips && path->slots[i] == 0) {
1201                         skip_level = i + 1;
1202                         continue;
1203                 }
1204                 if (!no_skips && path->keep_locks) {
1205                         u32 nritems;
1206                         t = path->nodes[i];
1207                         nritems = btrfs_header_nritems(t);
1208                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
1209                                 skip_level = i + 1;
1210                                 continue;
1211                         }
1212                 }
1213                 if (skip_level < i && i >= lowest_unlock)
1214                         no_skips = 1;
1215
1216                 t = path->nodes[i];
1217                 if (i >= lowest_unlock && i > skip_level && path->locks[i]) {
1218                         btrfs_tree_unlock(t);
1219                         path->locks[i] = 0;
1220                 }
1221         }
1222 }
1223
1224 /*
1225  * look for key in the tree.  path is filled in with nodes along the way
1226  * if key is found, we return zero and you can find the item in the leaf
1227  * level of the path (level 0)
1228  *
1229  * If the key isn't found, the path points to the slot where it should
1230  * be inserted, and 1 is returned.  If there are other errors during the
1231  * search a negative error number is returned.
1232  *
1233  * if ins_len > 0, nodes and leaves will be split as we walk down the
1234  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1235  * possible)
1236  */
1237 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1238                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1239                       ins_len, int cow)
1240 {
1241         struct extent_buffer *b;
1242         struct extent_buffer *tmp;
1243         int slot;
1244         int ret;
1245         int level;
1246         int should_reada = p->reada;
1247         int lowest_unlock = 1;
1248         u8 lowest_level = 0;
1249
1250         lowest_level = p->lowest_level;
1251         WARN_ON(lowest_level && ins_len);
1252         WARN_ON(p->nodes[0] != NULL);
1253         WARN_ON(root == root->fs_info->extent_root &&
1254                 !mutex_is_locked(&root->fs_info->alloc_mutex));
1255         WARN_ON(root == root->fs_info->chunk_root &&
1256                 !mutex_is_locked(&root->fs_info->chunk_mutex));
1257         WARN_ON(root == root->fs_info->dev_root &&
1258                 !mutex_is_locked(&root->fs_info->chunk_mutex));
1259         if (ins_len < 0)
1260                 lowest_unlock = 2;
1261 again:
1262         b = btrfs_lock_root_node(root);
1263
1264         while (b) {
1265                 level = btrfs_header_level(b);
1266                 if (cow) {
1267                         int wret;
1268                         wret = btrfs_cow_block(trans, root, b,
1269                                                p->nodes[level + 1],
1270                                                p->slots[level + 1],
1271                                                &b);
1272                         if (wret) {
1273                                 free_extent_buffer(b);
1274                                 return wret;
1275                         }
1276                 }
1277                 BUG_ON(!cow && ins_len);
1278                 if (level != btrfs_header_level(b))
1279                         WARN_ON(1);
1280                 level = btrfs_header_level(b);
1281                 p->nodes[level] = b;
1282                 p->locks[level] = 1;
1283                 ret = check_block(root, p, level);
1284                 if (ret)
1285                         return -1;
1286
1287                 ret = bin_search(b, key, level, &slot);
1288                 if (level != 0) {
1289                         if (ret && slot > 0)
1290                                 slot -= 1;
1291                         p->slots[level] = slot;
1292                         if (ins_len > 0 && btrfs_header_nritems(b) >=
1293                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1294                                 int sret = split_node(trans, root, p, level);
1295                                 BUG_ON(sret > 0);
1296                                 if (sret)
1297                                         return sret;
1298                                 b = p->nodes[level];
1299                                 slot = p->slots[level];
1300                         } else if (ins_len < 0) {
1301                                 int sret = balance_level(trans, root, p,
1302                                                          level);
1303                                 if (sret)
1304                                         return sret;
1305                                 b = p->nodes[level];
1306                                 if (!b) {
1307                                         btrfs_release_path(NULL, p);
1308                                         goto again;
1309                                 }
1310                                 slot = p->slots[level];
1311                                 BUG_ON(btrfs_header_nritems(b) == 1);
1312                         }
1313                         /* this is only true while dropping a snapshot */
1314                         if (level == lowest_level) {
1315                                 unlock_up(p, level, lowest_unlock);
1316                                 break;
1317                         }
1318
1319                         if (should_reada)
1320                                 reada_for_search(root, p, level, slot,
1321                                                  key->objectid);
1322
1323                         tmp = btrfs_find_tree_block(root,
1324                                           btrfs_node_blockptr(b, slot),
1325                                           btrfs_level_size(root, level - 1));
1326                         if (tmp && btrfs_buffer_uptodate(tmp,
1327                                    btrfs_node_ptr_generation(b, slot))) {
1328                                 b = tmp;
1329                         } else {
1330                                 /*
1331                                  * reduce lock contention at high levels
1332                                  * of the btree by dropping locks before
1333                                  * we read.
1334                                  */
1335                                 if (level > 1) {
1336                                         btrfs_release_path(NULL, p);
1337                                         if (tmp)
1338                                                 free_extent_buffer(tmp);
1339                                         goto again;
1340                                 } else {
1341                                         b = read_node_slot(root, b, slot);
1342                                 }
1343                         }
1344                         btrfs_tree_lock(b);
1345                         unlock_up(p, level, lowest_unlock);
1346                 } else {
1347                         p->slots[level] = slot;
1348                         if (ins_len > 0 && btrfs_leaf_free_space(root, b) <
1349                             sizeof(struct btrfs_item) + ins_len) {
1350                                 int sret = split_leaf(trans, root, key,
1351                                                       p, ins_len, ret == 0);
1352                                 BUG_ON(sret > 0);
1353                                 if (sret)
1354                                         return sret;
1355                         }
1356                         unlock_up(p, level, lowest_unlock);
1357                         return ret;
1358                 }
1359         }
1360         return 1;
1361 }
1362
1363 /*
1364  * adjust the pointers going up the tree, starting at level
1365  * making sure the right key of each node is points to 'key'.
1366  * This is used after shifting pointers to the left, so it stops
1367  * fixing up pointers when a given leaf/node is not in slot 0 of the
1368  * higher levels
1369  *
1370  * If this fails to write a tree block, it returns -1, but continues
1371  * fixing up the blocks in ram so the tree is consistent.
1372  */
1373 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1374                           struct btrfs_root *root, struct btrfs_path *path,
1375                           struct btrfs_disk_key *key, int level)
1376 {
1377         int i;
1378         int ret = 0;
1379         struct extent_buffer *t;
1380
1381         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1382                 int tslot = path->slots[i];
1383                 if (!path->nodes[i])
1384                         break;
1385                 t = path->nodes[i];
1386                 btrfs_set_node_key(t, key, tslot);
1387                 if (!btrfs_tree_locked(path->nodes[i])) {
1388                         int ii;
1389 printk("fixup without lock on level %d\n", btrfs_header_level(path->nodes[i]));
1390                         for (ii = 0; ii < BTRFS_MAX_LEVEL; ii++) {
1391 printk("level %d slot %d\n", ii, path->slots[ii]);
1392                         }
1393                 }
1394                 btrfs_mark_buffer_dirty(path->nodes[i]);
1395                 if (tslot != 0)
1396                         break;
1397         }
1398         return ret;
1399 }
1400
1401 /*
1402  * try to push data from one node into the next node left in the
1403  * tree.
1404  *
1405  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1406  * error, and > 0 if there was no room in the left hand block.
1407  */
1408 static int push_node_left(struct btrfs_trans_handle *trans,
1409                           struct btrfs_root *root, struct extent_buffer *dst,
1410                           struct extent_buffer *src, int empty)
1411 {
1412         int push_items = 0;
1413         int src_nritems;
1414         int dst_nritems;
1415         int ret = 0;
1416
1417         src_nritems = btrfs_header_nritems(src);
1418         dst_nritems = btrfs_header_nritems(dst);
1419         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1420         WARN_ON(btrfs_header_generation(src) != trans->transid);
1421         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1422
1423         if (!empty && src_nritems <= 8)
1424                 return 1;
1425
1426         if (push_items <= 0) {
1427                 return 1;
1428         }
1429
1430         if (empty) {
1431                 push_items = min(src_nritems, push_items);
1432                 if (push_items < src_nritems) {
1433                         /* leave at least 8 pointers in the node if
1434                          * we aren't going to empty it
1435                          */
1436                         if (src_nritems - push_items < 8) {
1437                                 if (push_items <= 8)
1438                                         return 1;
1439                                 push_items -= 8;
1440                         }
1441                 }
1442         } else
1443                 push_items = min(src_nritems - 8, push_items);
1444
1445         copy_extent_buffer(dst, src,
1446                            btrfs_node_key_ptr_offset(dst_nritems),
1447                            btrfs_node_key_ptr_offset(0),
1448                            push_items * sizeof(struct btrfs_key_ptr));
1449
1450         if (push_items < src_nritems) {
1451                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1452                                       btrfs_node_key_ptr_offset(push_items),
1453                                       (src_nritems - push_items) *
1454                                       sizeof(struct btrfs_key_ptr));
1455         }
1456         btrfs_set_header_nritems(src, src_nritems - push_items);
1457         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1458         btrfs_mark_buffer_dirty(src);
1459         btrfs_mark_buffer_dirty(dst);
1460         return ret;
1461 }
1462
1463 /*
1464  * try to push data from one node into the next node right in the
1465  * tree.
1466  *
1467  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1468  * error, and > 0 if there was no room in the right hand block.
1469  *
1470  * this will  only push up to 1/2 the contents of the left node over
1471  */
1472 static int balance_node_right(struct btrfs_trans_handle *trans,
1473                               struct btrfs_root *root,
1474                               struct extent_buffer *dst,
1475                               struct extent_buffer *src)
1476 {
1477         int push_items = 0;
1478         int max_push;
1479         int src_nritems;
1480         int dst_nritems;
1481         int ret = 0;
1482
1483         WARN_ON(btrfs_header_generation(src) != trans->transid);
1484         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1485
1486         src_nritems = btrfs_header_nritems(src);
1487         dst_nritems = btrfs_header_nritems(dst);
1488         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1489         if (push_items <= 0) {
1490                 return 1;
1491         }
1492
1493         if (src_nritems < 4) {
1494                 return 1;
1495         }
1496
1497         max_push = src_nritems / 2 + 1;
1498         /* don't try to empty the node */
1499         if (max_push >= src_nritems) {
1500                 return 1;
1501         }
1502
1503         if (max_push < push_items)
1504                 push_items = max_push;
1505
1506         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1507                                       btrfs_node_key_ptr_offset(0),
1508                                       (dst_nritems) *
1509                                       sizeof(struct btrfs_key_ptr));
1510
1511         copy_extent_buffer(dst, src,
1512                            btrfs_node_key_ptr_offset(0),
1513                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1514                            push_items * sizeof(struct btrfs_key_ptr));
1515
1516         btrfs_set_header_nritems(src, src_nritems - push_items);
1517         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1518
1519         btrfs_mark_buffer_dirty(src);
1520         btrfs_mark_buffer_dirty(dst);
1521         return ret;
1522 }
1523
1524 /*
1525  * helper function to insert a new root level in the tree.
1526  * A new node is allocated, and a single item is inserted to
1527  * point to the existing root
1528  *
1529  * returns zero on success or < 0 on failure.
1530  */
1531 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1532                            struct btrfs_root *root,
1533                            struct btrfs_path *path, int level)
1534 {
1535         u64 root_gen;
1536         u64 lower_gen;
1537         struct extent_buffer *lower;
1538         struct extent_buffer *c;
1539         struct extent_buffer *old;
1540         struct btrfs_disk_key lower_key;
1541
1542         BUG_ON(path->nodes[level]);
1543         BUG_ON(path->nodes[level-1] != root->node);
1544
1545         if (root->ref_cows)
1546                 root_gen = trans->transid;
1547         else
1548                 root_gen = 0;
1549
1550         lower = path->nodes[level-1];
1551         if (level == 1)
1552                 btrfs_item_key(lower, &lower_key, 0);
1553         else
1554                 btrfs_node_key(lower, &lower_key, 0);
1555
1556         c = btrfs_alloc_free_block(trans, root, root->nodesize,
1557                                    root->root_key.objectid,
1558                                    root_gen, lower_key.objectid, level,
1559                                    root->node->start, 0);
1560         if (IS_ERR(c))
1561                 return PTR_ERR(c);
1562
1563         memset_extent_buffer(c, 0, 0, root->nodesize);
1564         btrfs_set_header_nritems(c, 1);
1565         btrfs_set_header_level(c, level);
1566         btrfs_set_header_bytenr(c, c->start);
1567         btrfs_set_header_generation(c, trans->transid);
1568         btrfs_set_header_owner(c, root->root_key.objectid);
1569
1570         write_extent_buffer(c, root->fs_info->fsid,
1571                             (unsigned long)btrfs_header_fsid(c),
1572                             BTRFS_FSID_SIZE);
1573
1574         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1575                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
1576                             BTRFS_UUID_SIZE);
1577
1578         btrfs_set_node_key(c, &lower_key, 0);
1579         btrfs_set_node_blockptr(c, 0, lower->start);
1580         lower_gen = btrfs_header_generation(lower);
1581         WARN_ON(lower_gen == 0);
1582
1583         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1584
1585         btrfs_mark_buffer_dirty(c);
1586
1587         spin_lock(&root->node_lock);
1588         old = root->node;
1589         root->node = c;
1590         spin_unlock(&root->node_lock);
1591
1592         /* the super has an extra ref to root->node */
1593         free_extent_buffer(old);
1594
1595         add_root_to_dirty_list(root);
1596         extent_buffer_get(c);
1597         path->nodes[level] = c;
1598         path->locks[level] = 1;
1599         path->slots[level] = 0;
1600
1601         if (root->ref_cows && lower_gen != trans->transid) {
1602                 struct btrfs_path *back_path = btrfs_alloc_path();
1603                 int ret;
1604                 mutex_lock(&root->fs_info->alloc_mutex);
1605                 ret = btrfs_insert_extent_backref(trans,
1606                                                   root->fs_info->extent_root,
1607                                                   path, lower->start,
1608                                                   root->root_key.objectid,
1609                                                   trans->transid, 0, 0);
1610                 BUG_ON(ret);
1611                 mutex_unlock(&root->fs_info->alloc_mutex);
1612                 btrfs_free_path(back_path);
1613         }
1614         return 0;
1615 }
1616
1617 /*
1618  * worker function to insert a single pointer in a node.
1619  * the node should have enough room for the pointer already
1620  *
1621  * slot and level indicate where you want the key to go, and
1622  * blocknr is the block the key points to.
1623  *
1624  * returns zero on success and < 0 on any error
1625  */
1626 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1627                       *root, struct btrfs_path *path, struct btrfs_disk_key
1628                       *key, u64 bytenr, int slot, int level)
1629 {
1630         struct extent_buffer *lower;
1631         int nritems;
1632
1633         BUG_ON(!path->nodes[level]);
1634         lower = path->nodes[level];
1635         nritems = btrfs_header_nritems(lower);
1636         if (slot > nritems)
1637                 BUG();
1638         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1639                 BUG();
1640         if (slot != nritems) {
1641                 memmove_extent_buffer(lower,
1642                               btrfs_node_key_ptr_offset(slot + 1),
1643                               btrfs_node_key_ptr_offset(slot),
1644                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1645         }
1646         btrfs_set_node_key(lower, key, slot);
1647         btrfs_set_node_blockptr(lower, slot, bytenr);
1648         WARN_ON(trans->transid == 0);
1649         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1650         btrfs_set_header_nritems(lower, nritems + 1);
1651         btrfs_mark_buffer_dirty(lower);
1652         return 0;
1653 }
1654
1655 /*
1656  * split the node at the specified level in path in two.
1657  * The path is corrected to point to the appropriate node after the split
1658  *
1659  * Before splitting this tries to make some room in the node by pushing
1660  * left and right, if either one works, it returns right away.
1661  *
1662  * returns 0 on success and < 0 on failure
1663  */
1664 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1665                       *root, struct btrfs_path *path, int level)
1666 {
1667         u64 root_gen;
1668         struct extent_buffer *c;
1669         struct extent_buffer *split;
1670         struct btrfs_disk_key disk_key;
1671         int mid;
1672         int ret;
1673         int wret;
1674         u32 c_nritems;
1675
1676         c = path->nodes[level];
1677         WARN_ON(btrfs_header_generation(c) != trans->transid);
1678         if (c == root->node) {
1679                 /* trying to split the root, lets make a new one */
1680                 ret = insert_new_root(trans, root, path, level + 1);
1681                 if (ret)
1682                         return ret;
1683         } else {
1684                 ret = push_nodes_for_insert(trans, root, path, level);
1685                 c = path->nodes[level];
1686                 if (!ret && btrfs_header_nritems(c) <
1687                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
1688                         return 0;
1689                 if (ret < 0)
1690                         return ret;
1691         }
1692
1693         c_nritems = btrfs_header_nritems(c);
1694         if (root->ref_cows)
1695                 root_gen = trans->transid;
1696         else
1697                 root_gen = 0;
1698
1699         btrfs_node_key(c, &disk_key, 0);
1700         split = btrfs_alloc_free_block(trans, root, root->nodesize,
1701                                          root->root_key.objectid,
1702                                          root_gen,
1703                                          btrfs_disk_key_objectid(&disk_key),
1704                                          level, c->start, 0);
1705         if (IS_ERR(split))
1706                 return PTR_ERR(split);
1707
1708         btrfs_set_header_flags(split, btrfs_header_flags(c));
1709         btrfs_set_header_level(split, btrfs_header_level(c));
1710         btrfs_set_header_bytenr(split, split->start);
1711         btrfs_set_header_generation(split, trans->transid);
1712         btrfs_set_header_owner(split, root->root_key.objectid);
1713         btrfs_set_header_flags(split, 0);
1714         write_extent_buffer(split, root->fs_info->fsid,
1715                             (unsigned long)btrfs_header_fsid(split),
1716                             BTRFS_FSID_SIZE);
1717         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1718                             (unsigned long)btrfs_header_chunk_tree_uuid(split),
1719                             BTRFS_UUID_SIZE);
1720
1721         mid = (c_nritems + 1) / 2;
1722
1723         copy_extent_buffer(split, c,
1724                            btrfs_node_key_ptr_offset(0),
1725                            btrfs_node_key_ptr_offset(mid),
1726                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1727         btrfs_set_header_nritems(split, c_nritems - mid);
1728         btrfs_set_header_nritems(c, mid);
1729         ret = 0;
1730
1731         btrfs_mark_buffer_dirty(c);
1732         btrfs_mark_buffer_dirty(split);
1733
1734         btrfs_node_key(split, &disk_key, 0);
1735         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1736                           path->slots[level + 1] + 1,
1737                           level + 1);
1738         if (wret)
1739                 ret = wret;
1740
1741         if (path->slots[level] >= mid) {
1742                 path->slots[level] -= mid;
1743                 btrfs_tree_unlock(c);
1744                 free_extent_buffer(c);
1745                 path->nodes[level] = split;
1746                 path->slots[level + 1] += 1;
1747         } else {
1748                 btrfs_tree_unlock(split);
1749                 free_extent_buffer(split);
1750         }
1751         return ret;
1752 }
1753
1754 /*
1755  * how many bytes are required to store the items in a leaf.  start
1756  * and nr indicate which items in the leaf to check.  This totals up the
1757  * space used both by the item structs and the item data
1758  */
1759 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1760 {
1761         int data_len;
1762         int nritems = btrfs_header_nritems(l);
1763         int end = min(nritems, start + nr) - 1;
1764
1765         if (!nr)
1766                 return 0;
1767         data_len = btrfs_item_end_nr(l, start);
1768         data_len = data_len - btrfs_item_offset_nr(l, end);
1769         data_len += sizeof(struct btrfs_item) * nr;
1770         WARN_ON(data_len < 0);
1771         return data_len;
1772 }
1773
1774 /*
1775  * The space between the end of the leaf items and
1776  * the start of the leaf data.  IOW, how much room
1777  * the leaf has left for both items and data
1778  */
1779 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1780 {
1781         int nritems = btrfs_header_nritems(leaf);
1782         int ret;
1783         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1784         if (ret < 0) {
1785                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1786                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1787                        leaf_space_used(leaf, 0, nritems), nritems);
1788         }
1789         return ret;
1790 }
1791
1792 /*
1793  * push some data in the path leaf to the right, trying to free up at
1794  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1795  *
1796  * returns 1 if the push failed because the other node didn't have enough
1797  * room, 0 if everything worked out and < 0 if there were major errors.
1798  */
1799 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1800                            *root, struct btrfs_path *path, int data_size,
1801                            int empty)
1802 {
1803         struct extent_buffer *left = path->nodes[0];
1804         struct extent_buffer *right;
1805         struct extent_buffer *upper;
1806         struct btrfs_disk_key disk_key;
1807         int slot;
1808         u32 i;
1809         int free_space;
1810         int push_space = 0;
1811         int push_items = 0;
1812         struct btrfs_item *item;
1813         u32 left_nritems;
1814         u32 nr;
1815         u32 right_nritems;
1816         u32 data_end;
1817         u32 this_item_size;
1818         int ret;
1819
1820         slot = path->slots[1];
1821         if (!path->nodes[1]) {
1822                 return 1;
1823         }
1824         upper = path->nodes[1];
1825         if (slot >= btrfs_header_nritems(upper) - 1)
1826                 return 1;
1827
1828         WARN_ON(!btrfs_tree_locked(path->nodes[1]));
1829
1830         right = read_node_slot(root, upper, slot + 1);
1831         btrfs_tree_lock(right);
1832         free_space = btrfs_leaf_free_space(root, right);
1833         if (free_space < data_size + sizeof(struct btrfs_item))
1834                 goto out_unlock;
1835
1836         /* cow and double check */
1837         ret = btrfs_cow_block(trans, root, right, upper,
1838                               slot + 1, &right);
1839         if (ret)
1840                 goto out_unlock;
1841
1842         free_space = btrfs_leaf_free_space(root, right);
1843         if (free_space < data_size + sizeof(struct btrfs_item))
1844                 goto out_unlock;
1845
1846         left_nritems = btrfs_header_nritems(left);
1847         if (left_nritems == 0)
1848                 goto out_unlock;
1849
1850         if (empty)
1851                 nr = 0;
1852         else
1853                 nr = 1;
1854
1855         i = left_nritems - 1;
1856         while (i >= nr) {
1857                 item = btrfs_item_nr(left, i);
1858
1859                 if (path->slots[0] == i)
1860                         push_space += data_size + sizeof(*item);
1861
1862                 if (!left->map_token) {
1863                         map_extent_buffer(left, (unsigned long)item,
1864                                         sizeof(struct btrfs_item),
1865                                         &left->map_token, &left->kaddr,
1866                                         &left->map_start, &left->map_len,
1867                                         KM_USER1);
1868                 }
1869
1870                 this_item_size = btrfs_item_size(left, item);
1871                 if (this_item_size + sizeof(*item) + push_space > free_space)
1872                         break;
1873                 push_items++;
1874                 push_space += this_item_size + sizeof(*item);
1875                 if (i == 0)
1876                         break;
1877                 i--;
1878         }
1879         if (left->map_token) {
1880                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1881                 left->map_token = NULL;
1882         }
1883
1884         if (push_items == 0)
1885                 goto out_unlock;
1886
1887         if (!empty && push_items == left_nritems)
1888                 WARN_ON(1);
1889
1890         /* push left to right */
1891         right_nritems = btrfs_header_nritems(right);
1892
1893         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1894         push_space -= leaf_data_end(root, left);
1895
1896         /* make room in the right data area */
1897         data_end = leaf_data_end(root, right);
1898         memmove_extent_buffer(right,
1899                               btrfs_leaf_data(right) + data_end - push_space,
1900                               btrfs_leaf_data(right) + data_end,
1901                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1902
1903         /* copy from the left data area */
1904         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1905                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1906                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1907                      push_space);
1908
1909         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1910                               btrfs_item_nr_offset(0),
1911                               right_nritems * sizeof(struct btrfs_item));
1912
1913         /* copy the items from left to right */
1914         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1915                    btrfs_item_nr_offset(left_nritems - push_items),
1916                    push_items * sizeof(struct btrfs_item));
1917
1918         /* update the item pointers */
1919         right_nritems += push_items;
1920         btrfs_set_header_nritems(right, right_nritems);
1921         push_space = BTRFS_LEAF_DATA_SIZE(root);
1922         for (i = 0; i < right_nritems; i++) {
1923                 item = btrfs_item_nr(right, i);
1924                 if (!right->map_token) {
1925                         map_extent_buffer(right, (unsigned long)item,
1926                                         sizeof(struct btrfs_item),
1927                                         &right->map_token, &right->kaddr,
1928                                         &right->map_start, &right->map_len,
1929                                         KM_USER1);
1930                 }
1931                 push_space -= btrfs_item_size(right, item);
1932                 btrfs_set_item_offset(right, item, push_space);
1933         }
1934
1935         if (right->map_token) {
1936                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1937                 right->map_token = NULL;
1938         }
1939         left_nritems -= push_items;
1940         btrfs_set_header_nritems(left, left_nritems);
1941
1942         if (left_nritems)
1943                 btrfs_mark_buffer_dirty(left);
1944         btrfs_mark_buffer_dirty(right);
1945
1946         btrfs_item_key(right, &disk_key, 0);
1947         btrfs_set_node_key(upper, &disk_key, slot + 1);
1948         btrfs_mark_buffer_dirty(upper);
1949
1950         /* then fixup the leaf pointer in the path */
1951         if (path->slots[0] >= left_nritems) {
1952                 path->slots[0] -= left_nritems;
1953                 if (btrfs_header_nritems(path->nodes[0]) == 0)
1954                         clean_tree_block(trans, root, path->nodes[0]);
1955                 btrfs_tree_unlock(path->nodes[0]);
1956                 free_extent_buffer(path->nodes[0]);
1957                 path->nodes[0] = right;
1958                 path->slots[1] += 1;
1959         } else {
1960                 btrfs_tree_unlock(right);
1961                 free_extent_buffer(right);
1962         }
1963         return 0;
1964
1965 out_unlock:
1966         btrfs_tree_unlock(right);
1967         free_extent_buffer(right);
1968         return 1;
1969 }
1970
1971 /*
1972  * push some data in the path leaf to the left, trying to free up at
1973  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1974  */
1975 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1976                           *root, struct btrfs_path *path, int data_size,
1977                           int empty)
1978 {
1979         struct btrfs_disk_key disk_key;
1980         struct extent_buffer *right = path->nodes[0];
1981         struct extent_buffer *left;
1982         int slot;
1983         int i;
1984         int free_space;
1985         int push_space = 0;
1986         int push_items = 0;
1987         struct btrfs_item *item;
1988         u32 old_left_nritems;
1989         u32 right_nritems;
1990         u32 nr;
1991         int ret = 0;
1992         int wret;
1993         u32 this_item_size;
1994         u32 old_left_item_size;
1995
1996         slot = path->slots[1];
1997         if (slot == 0)
1998                 return 1;
1999         if (!path->nodes[1])
2000                 return 1;
2001
2002         right_nritems = btrfs_header_nritems(right);
2003         if (right_nritems == 0) {
2004                 return 1;
2005         }
2006
2007         WARN_ON(!btrfs_tree_locked(path->nodes[1]));
2008
2009         left = read_node_slot(root, path->nodes[1], slot - 1);
2010         btrfs_tree_lock(left);
2011         free_space = btrfs_leaf_free_space(root, left);
2012         if (free_space < data_size + sizeof(struct btrfs_item)) {
2013                 ret = 1;
2014                 goto out;
2015         }
2016
2017         /* cow and double check */
2018         ret = btrfs_cow_block(trans, root, left,
2019                               path->nodes[1], slot - 1, &left);
2020         if (ret) {
2021                 /* we hit -ENOSPC, but it isn't fatal here */
2022                 ret = 1;
2023                 goto out;
2024         }
2025
2026         free_space = btrfs_leaf_free_space(root, left);
2027         if (free_space < data_size + sizeof(struct btrfs_item)) {
2028                 ret = 1;
2029                 goto out;
2030         }
2031
2032         if (empty)
2033                 nr = right_nritems;
2034         else
2035                 nr = right_nritems - 1;
2036
2037         for (i = 0; i < nr; i++) {
2038                 item = btrfs_item_nr(right, i);
2039                 if (!right->map_token) {
2040                         map_extent_buffer(right, (unsigned long)item,
2041                                         sizeof(struct btrfs_item),
2042                                         &right->map_token, &right->kaddr,
2043                                         &right->map_start, &right->map_len,
2044                                         KM_USER1);
2045                 }
2046
2047                 if (path->slots[0] == i)
2048                         push_space += data_size + sizeof(*item);
2049
2050                 this_item_size = btrfs_item_size(right, item);
2051                 if (this_item_size + sizeof(*item) + push_space > free_space)
2052                         break;
2053
2054                 push_items++;
2055                 push_space += this_item_size + sizeof(*item);
2056         }
2057
2058         if (right->map_token) {
2059                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2060                 right->map_token = NULL;
2061         }
2062
2063         if (push_items == 0) {
2064                 ret = 1;
2065                 goto out;
2066         }
2067         if (!empty && push_items == btrfs_header_nritems(right))
2068                 WARN_ON(1);
2069
2070         /* push data from right to left */
2071         copy_extent_buffer(left, right,
2072                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
2073                            btrfs_item_nr_offset(0),
2074                            push_items * sizeof(struct btrfs_item));
2075
2076         push_space = BTRFS_LEAF_DATA_SIZE(root) -
2077                      btrfs_item_offset_nr(right, push_items -1);
2078
2079         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
2080                      leaf_data_end(root, left) - push_space,
2081                      btrfs_leaf_data(right) +
2082                      btrfs_item_offset_nr(right, push_items - 1),
2083                      push_space);
2084         old_left_nritems = btrfs_header_nritems(left);
2085         BUG_ON(old_left_nritems < 0);
2086
2087         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
2088         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
2089                 u32 ioff;
2090
2091                 item = btrfs_item_nr(left, i);
2092                 if (!left->map_token) {
2093                         map_extent_buffer(left, (unsigned long)item,
2094                                         sizeof(struct btrfs_item),
2095                                         &left->map_token, &left->kaddr,
2096                                         &left->map_start, &left->map_len,
2097                                         KM_USER1);
2098                 }
2099
2100                 ioff = btrfs_item_offset(left, item);
2101                 btrfs_set_item_offset(left, item,
2102                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
2103         }
2104         btrfs_set_header_nritems(left, old_left_nritems + push_items);
2105         if (left->map_token) {
2106                 unmap_extent_buffer(left, left->map_token, KM_USER1);
2107                 left->map_token = NULL;
2108         }
2109
2110         /* fixup right node */
2111         if (push_items > right_nritems) {
2112                 printk("push items %d nr %u\n", push_items, right_nritems);
2113                 WARN_ON(1);
2114         }
2115
2116         if (push_items < right_nritems) {
2117                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
2118                                                   leaf_data_end(root, right);
2119                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
2120                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
2121                                       btrfs_leaf_data(right) +
2122                                       leaf_data_end(root, right), push_space);
2123
2124                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
2125                               btrfs_item_nr_offset(push_items),
2126                              (btrfs_header_nritems(right) - push_items) *
2127                              sizeof(struct btrfs_item));
2128         }
2129         right_nritems -= push_items;
2130         btrfs_set_header_nritems(right, right_nritems);
2131         push_space = BTRFS_LEAF_DATA_SIZE(root);
2132         for (i = 0; i < right_nritems; i++) {
2133                 item = btrfs_item_nr(right, i);
2134
2135                 if (!right->map_token) {
2136                         map_extent_buffer(right, (unsigned long)item,
2137                                         sizeof(struct btrfs_item),
2138                                         &right->map_token, &right->kaddr,
2139                                         &right->map_start, &right->map_len,
2140                                         KM_USER1);
2141                 }
2142
2143                 push_space = push_space - btrfs_item_size(right, item);
2144                 btrfs_set_item_offset(right, item, push_space);
2145         }
2146         if (right->map_token) {
2147                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2148                 right->map_token = NULL;
2149         }
2150
2151         btrfs_mark_buffer_dirty(left);
2152         if (right_nritems)
2153                 btrfs_mark_buffer_dirty(right);
2154
2155         btrfs_item_key(right, &disk_key, 0);
2156         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
2157         if (wret)
2158                 ret = wret;
2159
2160         /* then fixup the leaf pointer in the path */
2161         if (path->slots[0] < push_items) {
2162                 path->slots[0] += old_left_nritems;
2163                 if (btrfs_header_nritems(path->nodes[0]) == 0)
2164                         clean_tree_block(trans, root, path->nodes[0]);
2165                 btrfs_tree_unlock(path->nodes[0]);
2166                 free_extent_buffer(path->nodes[0]);
2167                 path->nodes[0] = left;
2168                 path->slots[1] -= 1;
2169         } else {
2170                 btrfs_tree_unlock(left);
2171                 free_extent_buffer(left);
2172                 path->slots[0] -= push_items;
2173         }
2174         BUG_ON(path->slots[0] < 0);
2175         return ret;
2176 out:
2177         btrfs_tree_unlock(left);
2178         free_extent_buffer(left);
2179         return ret;
2180 }
2181
2182 /*
2183  * split the path's leaf in two, making sure there is at least data_size
2184  * available for the resulting leaf level of the path.
2185  *
2186  * returns 0 if all went well and < 0 on failure.
2187  */
2188 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
2189                       *root, struct btrfs_key *ins_key,
2190                       struct btrfs_path *path, int data_size, int extend)
2191 {
2192         u64 root_gen;
2193         struct extent_buffer *l;
2194         u32 nritems;
2195         int mid;
2196         int slot;
2197         struct extent_buffer *right;
2198         int space_needed = data_size + sizeof(struct btrfs_item);
2199         int data_copy_size;
2200         int rt_data_off;
2201         int i;
2202         int ret = 0;
2203         int wret;
2204         int double_split;
2205         int num_doubles = 0;
2206         struct btrfs_disk_key disk_key;
2207
2208         if (extend)
2209                 space_needed = data_size;
2210
2211         if (root->ref_cows)
2212                 root_gen = trans->transid;
2213         else
2214                 root_gen = 0;
2215
2216         /* first try to make some room by pushing left and right */
2217         if (ins_key->type != BTRFS_DIR_ITEM_KEY) {
2218                 wret = push_leaf_right(trans, root, path, data_size, 0);
2219                 if (wret < 0) {
2220                         return wret;
2221                 }
2222                 if (wret) {
2223                         wret = push_leaf_left(trans, root, path, data_size, 0);
2224                         if (wret < 0)
2225                                 return wret;
2226                 }
2227                 l = path->nodes[0];
2228
2229                 /* did the pushes work? */
2230                 if (btrfs_leaf_free_space(root, l) >= space_needed)
2231                         return 0;
2232         }
2233
2234         if (!path->nodes[1]) {
2235                 ret = insert_new_root(trans, root, path, 1);
2236                 if (ret)
2237                         return ret;
2238         }
2239 again:
2240         double_split = 0;
2241         l = path->nodes[0];
2242         slot = path->slots[0];
2243         nritems = btrfs_header_nritems(l);
2244         mid = (nritems + 1)/ 2;
2245
2246         btrfs_item_key(l, &disk_key, 0);
2247
2248         right = btrfs_alloc_free_block(trans, root, root->leafsize,
2249                                          root->root_key.objectid,
2250                                          root_gen, disk_key.objectid, 0,
2251                                          l->start, 0);
2252         if (IS_ERR(right)) {
2253                 BUG_ON(1);
2254                 return PTR_ERR(right);
2255         }
2256
2257         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2258         btrfs_set_header_bytenr(right, right->start);
2259         btrfs_set_header_generation(right, trans->transid);
2260         btrfs_set_header_owner(right, root->root_key.objectid);
2261         btrfs_set_header_level(right, 0);
2262         write_extent_buffer(right, root->fs_info->fsid,
2263                             (unsigned long)btrfs_header_fsid(right),
2264                             BTRFS_FSID_SIZE);
2265
2266         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2267                             (unsigned long)btrfs_header_chunk_tree_uuid(right),
2268                             BTRFS_UUID_SIZE);
2269         if (mid <= slot) {
2270                 if (nritems == 1 ||
2271                     leaf_space_used(l, mid, nritems - mid) + space_needed >
2272                         BTRFS_LEAF_DATA_SIZE(root)) {
2273                         if (slot >= nritems) {
2274                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2275                                 btrfs_set_header_nritems(right, 0);
2276                                 wret = insert_ptr(trans, root, path,
2277                                                   &disk_key, right->start,
2278                                                   path->slots[1] + 1, 1);
2279                                 if (wret)
2280                                         ret = wret;
2281
2282                                 btrfs_tree_unlock(path->nodes[0]);
2283                                 free_extent_buffer(path->nodes[0]);
2284                                 path->nodes[0] = right;
2285                                 path->slots[0] = 0;
2286                                 path->slots[1] += 1;
2287                                 btrfs_mark_buffer_dirty(right);
2288                                 return ret;
2289                         }
2290                         mid = slot;
2291                         if (mid != nritems &&
2292                             leaf_space_used(l, mid, nritems - mid) +
2293                             space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2294                                 double_split = 1;
2295                         }
2296                 }
2297         } else {
2298                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
2299                         BTRFS_LEAF_DATA_SIZE(root)) {
2300                         if (!extend && slot == 0) {
2301                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2302                                 btrfs_set_header_nritems(right, 0);
2303                                 wret = insert_ptr(trans, root, path,
2304                                                   &disk_key,
2305                                                   right->start,
2306                                                   path->slots[1], 1);
2307                                 if (wret)
2308                                         ret = wret;
2309                                 btrfs_tree_unlock(path->nodes[0]);
2310                                 free_extent_buffer(path->nodes[0]);
2311                                 path->nodes[0] = right;
2312                                 path->slots[0] = 0;
2313                                 if (path->slots[1] == 0) {
2314                                         wret = fixup_low_keys(trans, root,
2315                                                    path, &disk_key, 1);
2316                                         if (wret)
2317                                                 ret = wret;
2318                                 }
2319                                 btrfs_mark_buffer_dirty(right);
2320                                 return ret;
2321                         } else if (extend && slot == 0) {
2322                                 mid = 1;
2323                         } else {
2324                                 mid = slot;
2325                                 if (mid != nritems &&
2326                                     leaf_space_used(l, mid, nritems - mid) +
2327                                     space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2328                                         double_split = 1;
2329                                 }
2330                         }
2331                 }
2332         }
2333         nritems = nritems - mid;
2334         btrfs_set_header_nritems(right, nritems);
2335         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2336
2337         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2338                            btrfs_item_nr_offset(mid),
2339                            nritems * sizeof(struct btrfs_item));
2340
2341         copy_extent_buffer(right, l,
2342                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2343                      data_copy_size, btrfs_leaf_data(l) +
2344                      leaf_data_end(root, l), data_copy_size);
2345
2346         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2347                       btrfs_item_end_nr(l, mid);
2348
2349         for (i = 0; i < nritems; i++) {
2350                 struct btrfs_item *item = btrfs_item_nr(right, i);
2351                 u32 ioff;
2352
2353                 if (!right->map_token) {
2354                         map_extent_buffer(right, (unsigned long)item,
2355                                         sizeof(struct btrfs_item),
2356                                         &right->map_token, &right->kaddr,
2357                                         &right->map_start, &right->map_len,
2358                                         KM_USER1);
2359                 }
2360
2361                 ioff = btrfs_item_offset(right, item);
2362                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2363         }
2364
2365         if (right->map_token) {
2366                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2367                 right->map_token = NULL;
2368         }
2369
2370         btrfs_set_header_nritems(l, mid);
2371         ret = 0;
2372         btrfs_item_key(right, &disk_key, 0);
2373         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2374                           path->slots[1] + 1, 1);
2375         if (wret)
2376                 ret = wret;
2377
2378         btrfs_mark_buffer_dirty(right);
2379         btrfs_mark_buffer_dirty(l);
2380         BUG_ON(path->slots[0] != slot);
2381
2382         if (mid <= slot) {
2383                 btrfs_tree_unlock(path->nodes[0]);
2384                 free_extent_buffer(path->nodes[0]);
2385                 path->nodes[0] = right;
2386                 path->slots[0] -= mid;
2387                 path->slots[1] += 1;
2388         } else {
2389                 btrfs_tree_unlock(right);
2390                 free_extent_buffer(right);
2391         }
2392
2393         BUG_ON(path->slots[0] < 0);
2394
2395         if (double_split) {
2396                 BUG_ON(num_doubles != 0);
2397                 num_doubles++;
2398                 goto again;
2399         }
2400         return ret;
2401 }
2402
2403 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2404                         struct btrfs_root *root,
2405                         struct btrfs_path *path,
2406                         u32 new_size, int from_end)
2407 {
2408         int ret = 0;
2409         int slot;
2410         int slot_orig;
2411         struct extent_buffer *leaf;
2412         struct btrfs_item *item;
2413         u32 nritems;
2414         unsigned int data_end;
2415         unsigned int old_data_start;
2416         unsigned int old_size;
2417         unsigned int size_diff;
2418         int i;
2419
2420         slot_orig = path->slots[0];
2421         leaf = path->nodes[0];
2422         slot = path->slots[0];
2423
2424         old_size = btrfs_item_size_nr(leaf, slot);
2425         if (old_size == new_size)
2426                 return 0;
2427
2428         nritems = btrfs_header_nritems(leaf);
2429         data_end = leaf_data_end(root, leaf);
2430
2431         old_data_start = btrfs_item_offset_nr(leaf, slot);
2432
2433         size_diff = old_size - new_size;
2434
2435         BUG_ON(slot < 0);
2436         BUG_ON(slot >= nritems);
2437
2438         /*
2439          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2440          */
2441         /* first correct the data pointers */
2442         for (i = slot; i < nritems; i++) {
2443                 u32 ioff;
2444                 item = btrfs_item_nr(leaf, i);
2445
2446                 if (!leaf->map_token) {
2447                         map_extent_buffer(leaf, (unsigned long)item,
2448                                         sizeof(struct btrfs_item),
2449                                         &leaf->map_token, &leaf->kaddr,
2450                                         &leaf->map_start, &leaf->map_len,
2451                                         KM_USER1);
2452                 }
2453
2454                 ioff = btrfs_item_offset(leaf, item);
2455                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2456         }
2457
2458         if (leaf->map_token) {
2459                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2460                 leaf->map_token = NULL;
2461         }
2462
2463         /* shift the data */
2464         if (from_end) {
2465                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2466                               data_end + size_diff, btrfs_leaf_data(leaf) +
2467                               data_end, old_data_start + new_size - data_end);
2468         } else {
2469                 struct btrfs_disk_key disk_key;
2470                 u64 offset;
2471
2472                 btrfs_item_key(leaf, &disk_key, slot);
2473
2474                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2475                         unsigned long ptr;
2476                         struct btrfs_file_extent_item *fi;
2477
2478                         fi = btrfs_item_ptr(leaf, slot,
2479                                             struct btrfs_file_extent_item);
2480                         fi = (struct btrfs_file_extent_item *)(
2481                              (unsigned long)fi - size_diff);
2482
2483                         if (btrfs_file_extent_type(leaf, fi) ==
2484                             BTRFS_FILE_EXTENT_INLINE) {
2485                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2486                                 memmove_extent_buffer(leaf, ptr,
2487                                         (unsigned long)fi,
2488                                         offsetof(struct btrfs_file_extent_item,
2489                                                  disk_bytenr));
2490                         }
2491                 }
2492
2493                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2494                               data_end + size_diff, btrfs_leaf_data(leaf) +
2495                               data_end, old_data_start - data_end);
2496
2497                 offset = btrfs_disk_key_offset(&disk_key);
2498                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2499                 btrfs_set_item_key(leaf, &disk_key, slot);
2500                 if (slot == 0)
2501                         fixup_low_keys(trans, root, path, &disk_key, 1);
2502         }
2503
2504         item = btrfs_item_nr(leaf, slot);
2505         btrfs_set_item_size(leaf, item, new_size);
2506         btrfs_mark_buffer_dirty(leaf);
2507
2508         ret = 0;
2509         if (btrfs_leaf_free_space(root, leaf) < 0) {
2510                 btrfs_print_leaf(root, leaf);
2511                 BUG();
2512         }
2513         return ret;
2514 }
2515
2516 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2517                       struct btrfs_root *root, struct btrfs_path *path,
2518                       u32 data_size)
2519 {
2520         int ret = 0;
2521         int slot;
2522         int slot_orig;
2523         struct extent_buffer *leaf;
2524         struct btrfs_item *item;
2525         u32 nritems;
2526         unsigned int data_end;
2527         unsigned int old_data;
2528         unsigned int old_size;
2529         int i;
2530
2531         slot_orig = path->slots[0];
2532         leaf = path->nodes[0];
2533
2534         nritems = btrfs_header_nritems(leaf);
2535         data_end = leaf_data_end(root, leaf);
2536
2537         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2538                 btrfs_print_leaf(root, leaf);
2539                 BUG();
2540         }
2541         slot = path->slots[0];
2542         old_data = btrfs_item_end_nr(leaf, slot);
2543
2544         BUG_ON(slot < 0);
2545         if (slot >= nritems) {
2546                 btrfs_print_leaf(root, leaf);
2547                 printk("slot %d too large, nritems %d\n", slot, nritems);
2548                 BUG_ON(1);
2549         }
2550
2551         /*
2552          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2553          */
2554         /* first correct the data pointers */
2555         for (i = slot; i < nritems; i++) {
2556                 u32 ioff;
2557                 item = btrfs_item_nr(leaf, i);
2558
2559                 if (!leaf->map_token) {
2560                         map_extent_buffer(leaf, (unsigned long)item,
2561                                         sizeof(struct btrfs_item),
2562                                         &leaf->map_token, &leaf->kaddr,
2563                                         &leaf->map_start, &leaf->map_len,
2564                                         KM_USER1);
2565                 }
2566                 ioff = btrfs_item_offset(leaf, item);
2567                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2568         }
2569
2570         if (leaf->map_token) {
2571                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2572                 leaf->map_token = NULL;
2573         }
2574
2575         /* shift the data */
2576         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2577                       data_end - data_size, btrfs_leaf_data(leaf) +
2578                       data_end, old_data - data_end);
2579
2580         data_end = old_data;
2581         old_size = btrfs_item_size_nr(leaf, slot);
2582         item = btrfs_item_nr(leaf, slot);
2583         btrfs_set_item_size(leaf, item, old_size + data_size);
2584         btrfs_mark_buffer_dirty(leaf);
2585
2586         ret = 0;
2587         if (btrfs_leaf_free_space(root, leaf) < 0) {
2588                 btrfs_print_leaf(root, leaf);
2589                 BUG();
2590         }
2591         return ret;
2592 }
2593
2594 /*
2595  * Given a key and some data, insert an item into the tree.
2596  * This does all the path init required, making room in the tree if needed.
2597  */
2598 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2599                             struct btrfs_root *root,
2600                             struct btrfs_path *path,
2601                             struct btrfs_key *cpu_key, u32 *data_size,
2602                             int nr)
2603 {
2604         struct extent_buffer *leaf;
2605         struct btrfs_item *item;
2606         int ret = 0;
2607         int slot;
2608         int slot_orig;
2609         int i;
2610         u32 nritems;
2611         u32 total_size = 0;
2612         u32 total_data = 0;
2613         unsigned int data_end;
2614         struct btrfs_disk_key disk_key;
2615
2616         for (i = 0; i < nr; i++) {
2617                 total_data += data_size[i];
2618         }
2619
2620         total_size = total_data + (nr - 1) * sizeof(struct btrfs_item);
2621         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2622         if (ret == 0) {
2623                 return -EEXIST;
2624         }
2625         if (ret < 0)
2626                 goto out;
2627
2628         slot_orig = path->slots[0];
2629         leaf = path->nodes[0];
2630
2631         nritems = btrfs_header_nritems(leaf);
2632         data_end = leaf_data_end(root, leaf);
2633
2634         if (btrfs_leaf_free_space(root, leaf) <
2635             sizeof(struct btrfs_item) + total_size) {
2636                 btrfs_print_leaf(root, leaf);
2637                 printk("not enough freespace need %u have %d\n",
2638                        total_size, btrfs_leaf_free_space(root, leaf));
2639                 BUG();
2640         }
2641
2642         slot = path->slots[0];
2643         BUG_ON(slot < 0);
2644
2645         if (slot != nritems) {
2646                 int i;
2647                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2648
2649                 if (old_data < data_end) {
2650                         btrfs_print_leaf(root, leaf);
2651                         printk("slot %d old_data %d data_end %d\n",
2652                                slot, old_data, data_end);
2653                         BUG_ON(1);
2654                 }
2655                 /*
2656                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2657                  */
2658                 /* first correct the data pointers */
2659                 WARN_ON(leaf->map_token);
2660                 for (i = slot; i < nritems; i++) {
2661                         u32 ioff;
2662
2663                         item = btrfs_item_nr(leaf, i);
2664                         if (!leaf->map_token) {
2665                                 map_extent_buffer(leaf, (unsigned long)item,
2666                                         sizeof(struct btrfs_item),
2667                                         &leaf->map_token, &leaf->kaddr,
2668                                         &leaf->map_start, &leaf->map_len,
2669                                         KM_USER1);
2670                         }
2671
2672                         ioff = btrfs_item_offset(leaf, item);
2673                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2674                 }
2675                 if (leaf->map_token) {
2676                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2677                         leaf->map_token = NULL;
2678                 }
2679
2680                 /* shift the items */
2681                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2682                               btrfs_item_nr_offset(slot),
2683                               (nritems - slot) * sizeof(struct btrfs_item));
2684
2685                 /* shift the data */
2686                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2687                               data_end - total_data, btrfs_leaf_data(leaf) +
2688                               data_end, old_data - data_end);
2689                 data_end = old_data;
2690         }
2691
2692         /* setup the item for the new data */
2693         for (i = 0; i < nr; i++) {
2694                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2695                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2696                 item = btrfs_item_nr(leaf, slot + i);
2697                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2698                 data_end -= data_size[i];
2699                 btrfs_set_item_size(leaf, item, data_size[i]);
2700         }
2701         btrfs_set_header_nritems(leaf, nritems + nr);
2702         btrfs_mark_buffer_dirty(leaf);
2703
2704         ret = 0;
2705         if (slot == 0) {
2706                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2707                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2708         }
2709
2710         if (btrfs_leaf_free_space(root, leaf) < 0) {
2711                 btrfs_print_leaf(root, leaf);
2712                 BUG();
2713         }
2714 out:
2715         return ret;
2716 }
2717
2718 /*
2719  * Given a key and some data, insert an item into the tree.
2720  * This does all the path init required, making room in the tree if needed.
2721  */
2722 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2723                       *root, struct btrfs_key *cpu_key, void *data, u32
2724                       data_size)
2725 {
2726         int ret = 0;
2727         struct btrfs_path *path;
2728         struct extent_buffer *leaf;
2729         unsigned long ptr;
2730
2731         path = btrfs_alloc_path();
2732         BUG_ON(!path);
2733         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2734         if (!ret) {
2735                 leaf = path->nodes[0];
2736                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2737                 write_extent_buffer(leaf, data, ptr, data_size);
2738                 btrfs_mark_buffer_dirty(leaf);
2739         }
2740         btrfs_free_path(path);
2741         return ret;
2742 }
2743
2744 /*
2745  * delete the pointer from a given node.
2746  *
2747  * If the delete empties a node, the node is removed from the tree,
2748  * continuing all the way the root if required.  The root is converted into
2749  * a leaf if all the nodes are emptied.
2750  */
2751 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2752                    struct btrfs_path *path, int level, int slot)
2753 {
2754         struct extent_buffer *parent = path->nodes[level];
2755         u32 nritems;
2756         int ret = 0;
2757         int wret;
2758
2759         nritems = btrfs_header_nritems(parent);
2760         if (slot != nritems -1) {
2761                 memmove_extent_buffer(parent,
2762                               btrfs_node_key_ptr_offset(slot),
2763                               btrfs_node_key_ptr_offset(slot + 1),
2764                               sizeof(struct btrfs_key_ptr) *
2765                               (nritems - slot - 1));
2766         }
2767         nritems--;
2768         btrfs_set_header_nritems(parent, nritems);
2769         if (nritems == 0 && parent == root->node) {
2770                 BUG_ON(btrfs_header_level(root->node) != 1);
2771                 /* just turn the root into a leaf and break */
2772                 btrfs_set_header_level(root->node, 0);
2773         } else if (slot == 0) {
2774                 struct btrfs_disk_key disk_key;
2775
2776                 btrfs_node_key(parent, &disk_key, 0);
2777                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2778                 if (wret)
2779                         ret = wret;
2780         }
2781         btrfs_mark_buffer_dirty(parent);
2782         return ret;
2783 }
2784
2785 /*
2786  * delete the item at the leaf level in path.  If that empties
2787  * the leaf, remove it from the tree
2788  */
2789 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2790                     struct btrfs_path *path, int slot, int nr)
2791 {
2792         struct extent_buffer *leaf;
2793         struct btrfs_item *item;
2794         int last_off;
2795         int dsize = 0;
2796         int ret = 0;
2797         int wret;
2798         int i;
2799         u32 nritems;
2800
2801         leaf = path->nodes[0];
2802         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2803
2804         for (i = 0; i < nr; i++)
2805                 dsize += btrfs_item_size_nr(leaf, slot + i);
2806
2807         nritems = btrfs_header_nritems(leaf);
2808
2809         if (slot + nr != nritems) {
2810                 int i;
2811                 int data_end = leaf_data_end(root, leaf);
2812
2813                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2814                               data_end + dsize,
2815                               btrfs_leaf_data(leaf) + data_end,
2816                               last_off - data_end);
2817
2818                 for (i = slot + nr; i < nritems; i++) {
2819                         u32 ioff;
2820
2821                         item = btrfs_item_nr(leaf, i);
2822                         if (!leaf->map_token) {
2823                                 map_extent_buffer(leaf, (unsigned long)item,
2824                                         sizeof(struct btrfs_item),
2825                                         &leaf->map_token, &leaf->kaddr,
2826                                         &leaf->map_start, &leaf->map_len,
2827                                         KM_USER1);
2828                         }
2829                         ioff = btrfs_item_offset(leaf, item);
2830                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2831                 }
2832
2833                 if (leaf->map_token) {
2834                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2835                         leaf->map_token = NULL;
2836                 }
2837
2838                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2839                               btrfs_item_nr_offset(slot + nr),
2840                               sizeof(struct btrfs_item) *
2841                               (nritems - slot - nr));
2842         }
2843         btrfs_set_header_nritems(leaf, nritems - nr);
2844         nritems -= nr;
2845
2846         /* delete the leaf if we've emptied it */
2847         if (nritems == 0) {
2848                 if (leaf == root->node) {
2849                         btrfs_set_header_level(leaf, 0);
2850                 } else {
2851                         u64 root_gen = btrfs_header_generation(path->nodes[1]);
2852                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
2853                         if (wret)
2854                                 ret = wret;
2855                         wret = btrfs_free_extent(trans, root,
2856                                          leaf->start, leaf->len,
2857                                          btrfs_header_owner(path->nodes[1]),
2858                                          root_gen, 0, 0, 1);
2859                         if (wret)
2860                                 ret = wret;
2861                 }
2862         } else {
2863                 int used = leaf_space_used(leaf, 0, nritems);
2864                 if (slot == 0) {
2865                         struct btrfs_disk_key disk_key;
2866
2867                         btrfs_item_key(leaf, &disk_key, 0);
2868                         wret = fixup_low_keys(trans, root, path,
2869                                               &disk_key, 1);
2870                         if (wret)
2871                                 ret = wret;
2872                 }
2873
2874                 /* delete the leaf if it is mostly empty */
2875                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2876                         /* push_leaf_left fixes the path.
2877                          * make sure the path still points to our leaf
2878                          * for possible call to del_ptr below
2879                          */
2880                         slot = path->slots[1];
2881                         extent_buffer_get(leaf);
2882
2883                         wret = push_leaf_left(trans, root, path, 1, 1);
2884                         if (wret < 0 && wret != -ENOSPC)
2885                                 ret = wret;
2886
2887                         if (path->nodes[0] == leaf &&
2888                             btrfs_header_nritems(leaf)) {
2889                                 wret = push_leaf_right(trans, root, path, 1, 1);
2890                                 if (wret < 0 && wret != -ENOSPC)
2891                                         ret = wret;
2892                         }
2893
2894                         if (btrfs_header_nritems(leaf) == 0) {
2895                                 u64 root_gen;
2896                                 u64 bytenr = leaf->start;
2897                                 u32 blocksize = leaf->len;
2898
2899                                 root_gen = btrfs_header_generation(
2900                                                            path->nodes[1]);
2901
2902                                 wret = del_ptr(trans, root, path, 1, slot);
2903                                 if (wret)
2904                                         ret = wret;
2905
2906                                 free_extent_buffer(leaf);
2907                                 wret = btrfs_free_extent(trans, root, bytenr,
2908                                              blocksize,
2909                                              btrfs_header_owner(path->nodes[1]),
2910                                              root_gen, 0, 0, 1);
2911                                 if (wret)
2912                                         ret = wret;
2913                         } else {
2914                                 /* if we're still in the path, make sure
2915                                  * we're dirty.  Otherwise, one of the
2916                                  * push_leaf functions must have already
2917                                  * dirtied this buffer
2918                                  */
2919                                 if (path->nodes[0] == leaf)
2920                                         btrfs_mark_buffer_dirty(leaf);
2921                                 free_extent_buffer(leaf);
2922                         }
2923                 } else {
2924                         btrfs_mark_buffer_dirty(leaf);
2925                 }
2926         }
2927         return ret;
2928 }
2929
2930 /*
2931  * search the tree again to find a leaf with lesser keys
2932  * returns 0 if it found something or 1 if there are no lesser leaves.
2933  * returns < 0 on io errors.
2934  */
2935 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2936 {
2937         struct btrfs_key key;
2938         struct btrfs_disk_key found_key;
2939         int ret;
2940
2941         btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
2942
2943         if (key.offset > 0)
2944                 key.offset--;
2945         else if (key.type > 0)
2946                 key.type--;
2947         else if (key.objectid > 0)
2948                 key.objectid--;
2949         else
2950                 return 1;
2951
2952         btrfs_release_path(root, path);
2953         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2954         if (ret < 0)
2955                 return ret;
2956         btrfs_item_key(path->nodes[0], &found_key, 0);
2957         ret = comp_keys(&found_key, &key);
2958         if (ret < 0)
2959                 return 0;
2960         return 1;
2961 }
2962
2963 /*
2964  * search the tree again to find a leaf with greater keys
2965  * returns 0 if it found something or 1 if there are no greater leaves.
2966  * returns < 0 on io errors.
2967  */
2968 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2969 {
2970         int slot;
2971         int level = 1;
2972         struct extent_buffer *c;
2973         struct extent_buffer *next = NULL;
2974         struct btrfs_key key;
2975         u32 nritems;
2976         int ret;
2977
2978         nritems = btrfs_header_nritems(path->nodes[0]);
2979         if (nritems == 0) {
2980                 return 1;
2981         }
2982
2983         btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
2984
2985         btrfs_release_path(root, path);
2986         path->keep_locks = 1;
2987         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2988         path->keep_locks = 0;
2989
2990         if (ret < 0)
2991                 return ret;
2992
2993         nritems = btrfs_header_nritems(path->nodes[0]);
2994         if (nritems > 0 && path->slots[0] < nritems - 1) {
2995                 goto done;
2996         }
2997
2998         while(level < BTRFS_MAX_LEVEL) {
2999                 if (!path->nodes[level])
3000                         return 1;
3001
3002                 slot = path->slots[level] + 1;
3003                 c = path->nodes[level];
3004                 if (slot >= btrfs_header_nritems(c)) {
3005                         level++;
3006                         if (level == BTRFS_MAX_LEVEL) {
3007                                 return 1;
3008                         }
3009                         continue;
3010                 }
3011
3012                 if (next) {
3013                         btrfs_tree_unlock(next);
3014                         free_extent_buffer(next);
3015                 }
3016
3017                 if (level == 1 && path->locks[1] && path->reada)
3018                         reada_for_search(root, path, level, slot, 0);
3019
3020                 next = read_node_slot(root, c, slot);
3021                 WARN_ON(!btrfs_tree_locked(c));
3022                 btrfs_tree_lock(next);
3023                 break;
3024         }
3025         path->slots[level] = slot;
3026         while(1) {
3027                 level--;
3028                 c = path->nodes[level];
3029                 if (path->locks[level])
3030                         btrfs_tree_unlock(c);
3031                 free_extent_buffer(c);
3032                 path->nodes[level] = next;
3033                 path->slots[level] = 0;
3034                 path->locks[level] = 1;
3035                 if (!level)
3036                         break;
3037                 if (level == 1 && path->locks[1] && path->reada)
3038                         reada_for_search(root, path, level, slot, 0);
3039                 next = read_node_slot(root, next, 0);
3040                 WARN_ON(!btrfs_tree_locked(path->nodes[level]));
3041                 btrfs_tree_lock(next);
3042         }
3043 done:
3044         unlock_up(path, 0, 1);
3045         return 0;
3046 }
3047
3048 int btrfs_previous_item(struct btrfs_root *root,
3049                         struct btrfs_path *path, u64 min_objectid,
3050                         int type)
3051 {
3052         struct btrfs_key found_key;
3053         struct extent_buffer *leaf;
3054         int ret;
3055
3056         while(1) {
3057                 if (path->slots[0] == 0) {
3058                         ret = btrfs_prev_leaf(root, path);
3059                         if (ret != 0)
3060                                 return ret;
3061                 } else {
3062                         path->slots[0]--;
3063                 }
3064                 leaf = path->nodes[0];
3065                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3066                 if (found_key.type == type)
3067                         return 0;
3068         }
3069         return 1;
3070 }
3071