Pileus Git - ~andy/linux/blob - fs/btrfs/ctree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/sched.h>
  20 #include "ctree.h"
  21 #include "disk-io.h"
  22 #include "transaction.h"
  23 #include "print-tree.h"
  24 #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 skip = p->skip_locking;
  67         int keep = p->keep_locks;
  68
  69         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  70                 if (!p->nodes[i])
  71                         continue;
  72                 if (p->locks[i]) {
  73                         btrfs_tree_unlock(p->nodes[i]);
  74                         p->locks[i] = 0;
  75                 }
  76                 free_extent_buffer(p->nodes[i]);
  77         }
  78         memset(p, 0, sizeof(*p));
  79         p->skip_locking = skip;
  80         p->keep_locks = keep;
  81 }
  82
  83 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
  84 {
  85         struct extent_buffer *eb;
  86         spin_lock(&root->node_lock);
  87         eb = root->node;
  88         extent_buffer_get(eb);
  89         spin_unlock(&root->node_lock);
  90         return eb;
  91 }
  92
  93 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
  94 {
  95         struct extent_buffer *eb;
  96
  97         while(1) {
  98                 eb = btrfs_root_node(root);
  99                 btrfs_tree_lock(eb);
 100
 101                 spin_lock(&root->node_lock);
 102                 if (eb == root->node) {
 103                         spin_unlock(&root->node_lock);
 104                         break;
 105                 }
 106                 spin_unlock(&root->node_lock);
 107
 108                 btrfs_tree_unlock(eb);
 109                 free_extent_buffer(eb);
 110         }
 111         return eb;
 112 }
 113
 114 static void add_root_to_dirty_list(struct btrfs_root *root)
 115 {
 116         if (root->track_dirty && list_empty(&root->dirty_list)) {
 117                 list_add(&root->dirty_list,
 118                          &root->fs_info->dirty_cowonly_roots);
 119         }
 120 }
 121
 122 int btrfs_copy_root(struct btrfs_trans_handle *trans,
 123                       struct btrfs_root *root,
 124                       struct extent_buffer *buf,
 125                       struct extent_buffer **cow_ret, u64 new_root_objectid)
 126 {
 127         struct extent_buffer *cow;
 128         u32 nritems;
 129         int ret = 0;
 130         int level;
 131         struct btrfs_key first_key;
 132         struct btrfs_root *new_root;
 133
 134         new_root = kmalloc(sizeof(*new_root), GFP_NOFS);
 135         if (!new_root)
 136                 return -ENOMEM;
 137
 138         memcpy(new_root, root, sizeof(*new_root));
 139         new_root->root_key.objectid = new_root_objectid;
 140
 141         WARN_ON(root->ref_cows && trans->transid !=
 142                 root->fs_info->running_transaction->transid);
 143         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 144
 145         level = btrfs_header_level(buf);
 146         nritems = btrfs_header_nritems(buf);
 147         if (nritems) {
 148                 if (level == 0)
 149                         btrfs_item_key_to_cpu(buf, &first_key, 0);
 150                 else
 151                         btrfs_node_key_to_cpu(buf, &first_key, 0);
 152         } else {
 153                 first_key.objectid = 0;
 154         }
 155         cow = btrfs_alloc_free_block(trans, new_root, buf->len,
 156                                        new_root_objectid,
 157                                        trans->transid, first_key.objectid,
 158                                        level, buf->start, 0);
 159         if (IS_ERR(cow)) {
 160                 kfree(new_root);
 161                 return PTR_ERR(cow);
 162         }
 163
 164         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 165         btrfs_set_header_bytenr(cow, cow->start);
 166         btrfs_set_header_generation(cow, trans->transid);
 167         btrfs_set_header_owner(cow, new_root_objectid);
 168         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN);
 169
 170         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 171         ret = btrfs_inc_ref(trans, new_root, buf);
 172         kfree(new_root);
 173
 174         if (ret)
 175                 return ret;
 176
 177         btrfs_mark_buffer_dirty(cow);
 178         *cow_ret = cow;
 179         return 0;
 180 }
 181
 182 int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 183                              struct btrfs_root *root,
 184                              struct extent_buffer *buf,
 185                              struct extent_buffer *parent, int parent_slot,
 186                              struct extent_buffer **cow_ret,
 187                              u64 search_start, u64 empty_size)
 188 {
 189         u64 root_gen;
 190         struct extent_buffer *cow;
 191         u32 nritems;
 192         int ret = 0;
 193         int different_trans = 0;
 194         int level;
 195         int unlock_orig = 0;
 196         struct btrfs_key first_key;
 197
 198         if (*cow_ret == buf)
 199                 unlock_orig = 1;
 200
 201         WARN_ON(!btrfs_tree_locked(buf));
 202
 203         if (root->ref_cows) {
 204                 root_gen = trans->transid;
 205         } else {
 206                 root_gen = 0;
 207         }
 208         WARN_ON(root->ref_cows && trans->transid !=
 209                 root->fs_info->running_transaction->transid);
 210         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 211
 212         level = btrfs_header_level(buf);
 213         nritems = btrfs_header_nritems(buf);
 214         if (nritems) {
 215                 if (level == 0)
 216                         btrfs_item_key_to_cpu(buf, &first_key, 0);
 217                 else
 218                         btrfs_node_key_to_cpu(buf, &first_key, 0);
 219         } else {
 220                 first_key.objectid = 0;
 221         }
 222         cow = btrfs_alloc_free_block(trans, root, buf->len,
 223                                      root->root_key.objectid,
 224                                      root_gen, first_key.objectid, level,
 225                                      search_start, empty_size);
 226         if (IS_ERR(cow))
 227                 return PTR_ERR(cow);
 228
 229         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 230         btrfs_set_header_bytenr(cow, cow->start);
 231         btrfs_set_header_generation(cow, trans->transid);
 232         btrfs_set_header_owner(cow, root->root_key.objectid);
 233         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN);
 234
 235         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 236         if (btrfs_header_generation(buf) != trans->transid) {
 237                 different_trans = 1;
 238                 ret = btrfs_inc_ref(trans, root, buf);
 239                 if (ret)
 240                         return ret;
 241         } else {
 242                 clean_tree_block(trans, root, buf);
 243         }
 244
 245         if (buf == root->node) {
 246                 WARN_ON(parent && parent != buf);
 247                 root_gen = btrfs_header_generation(buf);
 248
 249                 spin_lock(&root->node_lock);
 250                 root->node = cow;
 251                 extent_buffer_get(cow);
 252                 spin_unlock(&root->node_lock);
 253
 254                 if (buf != root->commit_root) {
 255                         btrfs_free_extent(trans, root, buf->start,
 256                                           buf->len, root->root_key.objectid,
 257                                           root_gen, 0, 0, 1);
 258                 }
 259                 free_extent_buffer(buf);
 260                 add_root_to_dirty_list(root);
 261         } else {
 262                 root_gen = btrfs_header_generation(parent);
 263                 btrfs_set_node_blockptr(parent, parent_slot,
 264                                         cow->start);
 265                 WARN_ON(trans->transid == 0);
 266                 btrfs_set_node_ptr_generation(parent, parent_slot,
 267                                               trans->transid);
 268                 btrfs_mark_buffer_dirty(parent);
 269                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 270                 btrfs_free_extent(trans, root, buf->start, buf->len,
 271                                   btrfs_header_owner(parent), root_gen,
 272                                   0, 0, 1);
 273         }
 274         if (unlock_orig)
 275                 btrfs_tree_unlock(buf);
 276         free_extent_buffer(buf);
 277         btrfs_mark_buffer_dirty(cow);
 278         *cow_ret = cow;
 279         return 0;
 280 }
 281
 282 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 283                     struct btrfs_root *root, struct extent_buffer *buf,
 284                     struct extent_buffer *parent, int parent_slot,
 285                     struct extent_buffer **cow_ret)
 286 {
 287         u64 search_start;
 288         u64 header_trans;
 289         int ret;
 290
 291         if (trans->transaction != root->fs_info->running_transaction) {
 292                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 293                        root->fs_info->running_transaction->transid);
 294                 WARN_ON(1);
 295         }
 296         if (trans->transid != root->fs_info->generation) {
 297                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 298                        root->fs_info->generation);
 299                 WARN_ON(1);
 300         }
 301
 302         header_trans = btrfs_header_generation(buf);
 303         spin_lock(&root->fs_info->hash_lock);
 304         if (header_trans == trans->transid &&
 305             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
 306                 *cow_ret = buf;
 307                 spin_unlock(&root->fs_info->hash_lock);
 308                 return 0;
 309         }
 310         spin_unlock(&root->fs_info->hash_lock);
 311         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
 312         ret = __btrfs_cow_block(trans, root, buf, parent,
 313                                  parent_slot, cow_ret, search_start, 0);
 314         return ret;
 315 }
 316
 317 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 318 {
 319         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 320                 return 1;
 321         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 322                 return 1;
 323         return 0;
 324 }
 325
 326 /*
 327  * compare two keys in a memcmp fashion
 328  */
 329 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 330 {
 331         struct btrfs_key k1;
 332
 333         btrfs_disk_key_to_cpu(&k1, disk);
 334
 335         if (k1.objectid > k2->objectid)
 336                 return 1;
 337         if (k1.objectid < k2->objectid)
 338                 return -1;
 339         if (k1.type > k2->type)
 340                 return 1;
 341         if (k1.type < k2->type)
 342                 return -1;
 343         if (k1.offset > k2->offset)
 344                 return 1;
 345         if (k1.offset < k2->offset)
 346                 return -1;
 347         return 0;
 348 }
 349
 350
 351 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 352                        struct btrfs_root *root, struct extent_buffer *parent,
 353                        int start_slot, int cache_only, u64 *last_ret,
 354                        struct btrfs_key *progress)
 355 {
 356         struct extent_buffer *cur;
 357         struct extent_buffer *tmp;
 358         u64 blocknr;
 359         u64 gen;
 360         u64 search_start = *last_ret;
 361         u64 last_block = 0;
 362         u64 other;
 363         u32 parent_nritems;
 364         int end_slot;
 365         int i;
 366         int err = 0;
 367         int parent_level;
 368         int uptodate;
 369         u32 blocksize;
 370         int progress_passed = 0;
 371         struct btrfs_disk_key disk_key;
 372
 373         /* FIXME this code needs locking */
 374         return 0;
 375
 376         parent_level = btrfs_header_level(parent);
 377         if (cache_only && parent_level != 1)
 378                 return 0;
 379
 380         if (trans->transaction != root->fs_info->running_transaction) {
 381                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 382                        root->fs_info->running_transaction->transid);
 383                 WARN_ON(1);
 384         }
 385         if (trans->transid != root->fs_info->generation) {
 386                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 387                        root->fs_info->generation);
 388                 WARN_ON(1);
 389         }
 390
 391         parent_nritems = btrfs_header_nritems(parent);
 392         blocksize = btrfs_level_size(root, parent_level - 1);
 393         end_slot = parent_nritems;
 394
 395         if (parent_nritems == 1)
 396                 return 0;
 397
 398         for (i = start_slot; i < end_slot; i++) {
 399                 int close = 1;
 400
 401                 if (!parent->map_token) {
 402                         map_extent_buffer(parent,
 403                                         btrfs_node_key_ptr_offset(i),
 404                                         sizeof(struct btrfs_key_ptr),
 405                                         &parent->map_token, &parent->kaddr,
 406                                         &parent->map_start, &parent->map_len,
 407                                         KM_USER1);
 408                 }
 409                 btrfs_node_key(parent, &disk_key, i);
 410                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
 411                         continue;
 412
 413                 progress_passed = 1;
 414                 blocknr = btrfs_node_blockptr(parent, i);
 415                 gen = btrfs_node_ptr_generation(parent, i);
 416                 if (last_block == 0)
 417                         last_block = blocknr;
 418
 419                 if (i > 0) {
 420                         other = btrfs_node_blockptr(parent, i - 1);
 421                         close = close_blocks(blocknr, other, blocksize);
 422                 }
 423                 if (!close && i < end_slot - 2) {
 424                         other = btrfs_node_blockptr(parent, i + 1);
 425                         close = close_blocks(blocknr, other, blocksize);
 426                 }
 427                 if (close) {
 428                         last_block = blocknr;
 429                         continue;
 430                 }
 431                 if (parent->map_token) {
 432                         unmap_extent_buffer(parent, parent->map_token,
 433                                             KM_USER1);
 434                         parent->map_token = NULL;
 435                 }
 436
 437                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
 438                 if (cur)
 439                         uptodate = btrfs_buffer_uptodate(cur, gen);
 440                 else
 441                         uptodate = 0;
 442                 if (!cur || !uptodate) {
 443                         if (cache_only) {
 444                                 free_extent_buffer(cur);
 445                                 continue;
 446                         }
 447                         if (!cur) {
 448                                 cur = read_tree_block(root, blocknr,
 449                                                          blocksize, gen);
 450                         } else if (!uptodate) {
 451                                 btrfs_read_buffer(cur, gen);
 452                         }
 453                 }
 454                 if (search_start == 0)
 455                         search_start = last_block;
 456
 457                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 458                                         &tmp, search_start,
 459                                         min(16 * blocksize,
 460                                             (end_slot - i) * blocksize));
 461                 if (err) {
 462                         free_extent_buffer(cur);
 463                         break;
 464                 }
 465                 search_start = tmp->start;
 466                 last_block = tmp->start;
 467                 *last_ret = search_start;
 468                 if (parent_level == 1)
 469                         btrfs_clear_buffer_defrag(tmp);
 470                 free_extent_buffer(tmp);
 471         }
 472         if (parent->map_token) {
 473                 unmap_extent_buffer(parent, parent->map_token,
 474                                     KM_USER1);
 475                 parent->map_token = NULL;
 476         }
 477         return err;
 478 }
 479
 480 /*
 481  * The leaf data grows from end-to-front in the node.
 482  * this returns the address of the start of the last item,
 483  * which is the stop of the leaf data stack
 484  */
 485 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 486                                          struct extent_buffer *leaf)
 487 {
 488         u32 nr = btrfs_header_nritems(leaf);
 489         if (nr == 0)
 490                 return BTRFS_LEAF_DATA_SIZE(root);
 491         return btrfs_item_offset_nr(leaf, nr - 1);
 492 }
 493
 494 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
 495                       int level)
 496 {
 497         struct extent_buffer *parent = NULL;
 498         struct extent_buffer *node = path->nodes[level];
 499         struct btrfs_disk_key parent_key;
 500         struct btrfs_disk_key node_key;
 501         int parent_slot;
 502         int slot;
 503         struct btrfs_key cpukey;
 504         u32 nritems = btrfs_header_nritems(node);
 505
 506         if (path->nodes[level + 1])
 507                 parent = path->nodes[level + 1];
 508
 509         slot = path->slots[level];
 510         BUG_ON(nritems == 0);
 511         if (parent) {
 512                 parent_slot = path->slots[level + 1];
 513                 btrfs_node_key(parent, &parent_key, parent_slot);
 514                 btrfs_node_key(node, &node_key, 0);
 515                 BUG_ON(memcmp(&parent_key, &node_key,
 516                               sizeof(struct btrfs_disk_key)));
 517                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 518                        btrfs_header_bytenr(node));
 519         }
 520         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 521         if (slot != 0) {
 522                 btrfs_node_key_to_cpu(node, &cpukey, slot - 1);
 523                 btrfs_node_key(node, &node_key, slot);
 524                 BUG_ON(comp_keys(&node_key, &cpukey) <= 0);
 525         }
 526         if (slot < nritems - 1) {
 527                 btrfs_node_key_to_cpu(node, &cpukey, slot + 1);
 528                 btrfs_node_key(node, &node_key, slot);
 529                 BUG_ON(comp_keys(&node_key, &cpukey) >= 0);
 530         }
 531         return 0;
 532 }
 533
 534 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 535                       int level)
 536 {
 537         struct extent_buffer *leaf = path->nodes[level];
 538         struct extent_buffer *parent = NULL;
 539         int parent_slot;
 540         struct btrfs_key cpukey;
 541         struct btrfs_disk_key parent_key;
 542         struct btrfs_disk_key leaf_key;
 543         int slot = path->slots[0];
 544
 545         u32 nritems = btrfs_header_nritems(leaf);
 546
 547         if (path->nodes[level + 1])
 548                 parent = path->nodes[level + 1];
 549
 550         if (nritems == 0)
 551                 return 0;
 552
 553         if (parent) {
 554                 parent_slot = path->slots[level + 1];
 555                 btrfs_node_key(parent, &parent_key, parent_slot);
 556                 btrfs_item_key(leaf, &leaf_key, 0);
 557
 558                 BUG_ON(memcmp(&parent_key, &leaf_key,
 559                        sizeof(struct btrfs_disk_key)));
 560                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 561                        btrfs_header_bytenr(leaf));
 562         }
 563 #if 0
 564         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 565                 btrfs_item_key_to_cpu(leaf, &cpukey, i + 1);
 566                 btrfs_item_key(leaf, &leaf_key, i);
 567                 if (comp_keys(&leaf_key, &cpukey) >= 0) {
 568                         btrfs_print_leaf(root, leaf);
 569                         printk("slot %d offset bad key\n", i);
 570                         BUG_ON(1);
 571                 }
 572                 if (btrfs_item_offset_nr(leaf, i) !=
 573                         btrfs_item_end_nr(leaf, i + 1)) {
 574                         btrfs_print_leaf(root, leaf);
 575                         printk("slot %d offset bad\n", i);
 576                         BUG_ON(1);
 577                 }
 578                 if (i == 0) {
 579                         if (btrfs_item_offset_nr(leaf, i) +
 580                                btrfs_item_size_nr(leaf, i) !=
 581                                BTRFS_LEAF_DATA_SIZE(root)) {
 582                                 btrfs_print_leaf(root, leaf);
 583                                 printk("slot %d first offset bad\n", i);
 584                                 BUG_ON(1);
 585                         }
 586                 }
 587         }
 588         if (nritems > 0) {
 589                 if (btrfs_item_size_nr(leaf, nritems - 1) > 4096) {
 590                                 btrfs_print_leaf(root, leaf);
 591                                 printk("slot %d bad size \n", nritems - 1);
 592                                 BUG_ON(1);
 593                 }
 594         }
 595 #endif
 596         if (slot != 0 && slot < nritems - 1) {
 597                 btrfs_item_key(leaf, &leaf_key, slot);
 598                 btrfs_item_key_to_cpu(leaf, &cpukey, slot - 1);
 599                 if (comp_keys(&leaf_key, &cpukey) <= 0) {
 600                         btrfs_print_leaf(root, leaf);
 601                         printk("slot %d offset bad key\n", slot);
 602                         BUG_ON(1);
 603                 }
 604                 if (btrfs_item_offset_nr(leaf, slot - 1) !=
 605                        btrfs_item_end_nr(leaf, slot)) {
 606                         btrfs_print_leaf(root, leaf);
 607                         printk("slot %d offset bad\n", slot);
 608                         BUG_ON(1);
 609                 }
 610         }
 611         if (slot < nritems - 1) {
 612                 btrfs_item_key(leaf, &leaf_key, slot);
 613                 btrfs_item_key_to_cpu(leaf, &cpukey, slot + 1);
 614                 BUG_ON(comp_keys(&leaf_key, &cpukey) >= 0);
 615                 if (btrfs_item_offset_nr(leaf, slot) !=
 616                         btrfs_item_end_nr(leaf, slot + 1)) {
 617                         btrfs_print_leaf(root, leaf);
 618                         printk("slot %d offset bad\n", slot);
 619                         BUG_ON(1);
 620                 }
 621         }
 622         BUG_ON(btrfs_item_offset_nr(leaf, 0) +
 623                btrfs_item_size_nr(leaf, 0) != BTRFS_LEAF_DATA_SIZE(root));
 624         return 0;
 625 }
 626
 627 static int noinline check_block(struct btrfs_root *root,
 628                                 struct btrfs_path *path, int level)
 629 {
 630         u64 found_start;
 631         return 0;
 632         if (btrfs_header_level(path->nodes[level]) != level)
 633             printk("warning: bad level %Lu wanted %d found %d\n",
 634                    path->nodes[level]->start, level,
 635                    btrfs_header_level(path->nodes[level]));
 636         found_start = btrfs_header_bytenr(path->nodes[level]);
 637         if (found_start != path->nodes[level]->start) {
 638             printk("warning: bad bytentr %Lu found %Lu\n",
 639                    path->nodes[level]->start, found_start);
 640         }
 641 #if 0
 642         struct extent_buffer *buf = path->nodes[level];
 643
 644         if (memcmp_extent_buffer(buf, root->fs_info->fsid,
 645                                  (unsigned long)btrfs_header_fsid(buf),
 646                                  BTRFS_FSID_SIZE)) {
 647                 printk("warning bad block %Lu\n", buf->start);
 648                 return 1;
 649         }
 650 #endif
 651         if (level == 0)
 652                 return check_leaf(root, path, level);
 653         return check_node(root, path, level);
 654 }
 655
 656 /*
 657  * search for key in the extent_buffer.  The items start at offset p,
 658  * and they are item_size apart.  There are 'max' items in p.
 659  *
 660  * the slot in the array is returned via slot, and it points to
 661  * the place where you would insert key if it is not found in
 662  * the array.
 663  *
 664  * slot may point to max if the key is bigger than all of the keys
 665  */
 666 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 667                               int item_size, struct btrfs_key *key,
 668                               int max, int *slot)
 669 {
 670         int low = 0;
 671         int high = max;
 672         int mid;
 673         int ret;
 674         struct btrfs_disk_key *tmp = NULL;
 675         struct btrfs_disk_key unaligned;
 676         unsigned long offset;
 677         char *map_token = NULL;
 678         char *kaddr = NULL;
 679         unsigned long map_start = 0;
 680         unsigned long map_len = 0;
 681         int err;
 682
 683         while(low < high) {
 684                 mid = (low + high) / 2;
 685                 offset = p + mid * item_size;
 686
 687                 if (!map_token || offset < map_start ||
 688                     (offset + sizeof(struct btrfs_disk_key)) >
 689                     map_start + map_len) {
 690                         if (map_token) {
 691                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 692                                 map_token = NULL;
 693                         }
 694                         err = map_extent_buffer(eb, offset,
 695                                                 sizeof(struct btrfs_disk_key),
 696                                                 &map_token, &kaddr,
 697                                                 &map_start, &map_len, KM_USER0);
 698
 699                         if (!err) {
 700                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
 701                                                         map_start);
 702                         } else {
 703                                 read_extent_buffer(eb, &unaligned,
 704                                                    offset, sizeof(unaligned));
 705                                 tmp = &unaligned;
 706                         }
 707
 708                 } else {
 709                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
 710                                                         map_start);
 711                 }
 712                 ret = comp_keys(tmp, key);
 713
 714                 if (ret < 0)
 715                         low = mid + 1;
 716                 else if (ret > 0)
 717                         high = mid;
 718                 else {
 719                         *slot = mid;
 720                         if (map_token)
 721                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 722                         return 0;
 723                 }
 724         }
 725         *slot = low;
 726         if (map_token)
 727                 unmap_extent_buffer(eb, map_token, KM_USER0);
 728         return 1;
 729 }
 730
 731 /*
 732  * simple bin_search frontend that does the right thing for
 733  * leaves vs nodes
 734  */
 735 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 736                       int level, int *slot)
 737 {
 738         if (level == 0) {
 739                 return generic_bin_search(eb,
 740                                           offsetof(struct btrfs_leaf, items),
 741                                           sizeof(struct btrfs_item),
 742                                           key, btrfs_header_nritems(eb),
 743                                           slot);
 744         } else {
 745                 return generic_bin_search(eb,
 746                                           offsetof(struct btrfs_node, ptrs),
 747                                           sizeof(struct btrfs_key_ptr),
 748                                           key, btrfs_header_nritems(eb),
 749                                           slot);
 750         }
 751         return -1;
 752 }
 753
 754 static struct extent_buffer *read_node_slot(struct btrfs_root *root,
 755                                    struct extent_buffer *parent, int slot)
 756 {
 757         int level = btrfs_header_level(parent);
 758         if (slot < 0)
 759                 return NULL;
 760         if (slot >= btrfs_header_nritems(parent))
 761                 return NULL;
 762
 763         BUG_ON(level == 0);
 764
 765         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 766                        btrfs_level_size(root, level - 1),
 767                        btrfs_node_ptr_generation(parent, slot));
 768 }
 769
 770 static int balance_level(struct btrfs_trans_handle *trans,
 771                          struct btrfs_root *root,
 772                          struct btrfs_path *path, int level)
 773 {
 774         struct extent_buffer *right = NULL;
 775         struct extent_buffer *mid;
 776         struct extent_buffer *left = NULL;
 777         struct extent_buffer *parent = NULL;
 778         int ret = 0;
 779         int wret;
 780         int pslot;
 781         int orig_slot = path->slots[level];
 782         int err_on_enospc = 0;
 783         u64 orig_ptr;
 784
 785         if (level == 0)
 786                 return 0;
 787
 788         mid = path->nodes[level];
 789         WARN_ON(!path->locks[level]);
 790         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 791
 792         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 793
 794         if (level < BTRFS_MAX_LEVEL - 1)
 795                 parent = path->nodes[level + 1];
 796         pslot = path->slots[level + 1];
 797
 798         /*
 799          * deal with the case where there is only one pointer in the root
 800          * by promoting the node below to a root
 801          */
 802         if (!parent) {
 803                 struct extent_buffer *child;
 804
 805                 if (btrfs_header_nritems(mid) != 1)
 806                         return 0;
 807
 808                 /* promote the child to a root */
 809                 child = read_node_slot(root, mid, 0);
 810                 btrfs_tree_lock(child);
 811                 BUG_ON(!child);
 812                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
 813                 BUG_ON(ret);
 814
 815                 spin_lock(&root->node_lock);
 816                 root->node = child;
 817                 spin_unlock(&root->node_lock);
 818
 819                 add_root_to_dirty_list(root);
 820                 btrfs_tree_unlock(child);
 821                 path->locks[level] = 0;
 822                 path->nodes[level] = NULL;
 823                 clean_tree_block(trans, root, mid);
 824                 btrfs_tree_unlock(mid);
 825                 /* once for the path */
 826                 free_extent_buffer(mid);
 827                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 828                                         root->root_key.objectid,
 829                                         btrfs_header_generation(mid), 0, 0, 1);
 830                 /* once for the root ptr */
 831                 free_extent_buffer(mid);
 832                 return ret;
 833         }
 834         if (btrfs_header_nritems(mid) >
 835             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 836                 return 0;
 837
 838         if (btrfs_header_nritems(mid) < 2)
 839                 err_on_enospc = 1;
 840
 841         left = read_node_slot(root, parent, pslot - 1);
 842         if (left) {
 843                 btrfs_tree_lock(left);
 844                 wret = btrfs_cow_block(trans, root, left,
 845                                        parent, pslot - 1, &left);
 846                 if (wret) {
 847                         ret = wret;
 848                         goto enospc;
 849                 }
 850         }
 851         right = read_node_slot(root, parent, pslot + 1);
 852         if (right) {
 853                 btrfs_tree_lock(right);
 854                 wret = btrfs_cow_block(trans, root, right,
 855                                        parent, pslot + 1, &right);
 856                 if (wret) {
 857                         ret = wret;
 858                         goto enospc;
 859                 }
 860         }
 861
 862         /* first, try to make some room in the middle buffer */
 863         if (left) {
 864                 orig_slot += btrfs_header_nritems(left);
 865                 wret = push_node_left(trans, root, left, mid, 1);
 866                 if (wret < 0)
 867                         ret = wret;
 868                 if (btrfs_header_nritems(mid) < 2)
 869                         err_on_enospc = 1;
 870         }
 871
 872         /*
 873          * then try to empty the right most buffer into the middle
 874          */
 875         if (right) {
 876                 wret = push_node_left(trans, root, mid, right, 1);
 877                 if (wret < 0 && wret != -ENOSPC)
 878                         ret = wret;
 879                 if (btrfs_header_nritems(right) == 0) {
 880                         u64 bytenr = right->start;
 881                         u64 generation = btrfs_header_generation(parent);
 882                         u32 blocksize = right->len;
 883
 884                         clean_tree_block(trans, root, right);
 885                         btrfs_tree_unlock(right);
 886                         free_extent_buffer(right);
 887                         right = NULL;
 888                         wret = del_ptr(trans, root, path, level + 1, pslot +
 889                                        1);
 890                         if (wret)
 891                                 ret = wret;
 892                         wret = btrfs_free_extent(trans, root, bytenr,
 893                                                  blocksize,
 894                                                  btrfs_header_owner(parent),
 895                                                  generation, 0, 0, 1);
 896                         if (wret)
 897                                 ret = wret;
 898                 } else {
 899                         struct btrfs_disk_key right_key;
 900                         btrfs_node_key(right, &right_key, 0);
 901                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 902                         btrfs_mark_buffer_dirty(parent);
 903                 }
 904         }
 905         if (btrfs_header_nritems(mid) == 1) {
 906                 /*
 907                  * we're not allowed to leave a node with one item in the
 908                  * tree during a delete.  A deletion from lower in the tree
 909                  * could try to delete the only pointer in this node.
 910                  * So, pull some keys from the left.
 911                  * There has to be a left pointer at this point because
 912                  * otherwise we would have pulled some pointers from the
 913                  * right
 914                  */
 915                 BUG_ON(!left);
 916                 wret = balance_node_right(trans, root, mid, left);
 917                 if (wret < 0) {
 918                         ret = wret;
 919                         goto enospc;
 920                 }
 921                 if (wret == 1) {
 922                         wret = push_node_left(trans, root, left, mid, 1);
 923                         if (wret < 0)
 924                                 ret = wret;
 925                 }
 926                 BUG_ON(wret == 1);
 927         }
 928         if (btrfs_header_nritems(mid) == 0) {
 929                 /* we've managed to empty the middle node, drop it */
 930                 u64 root_gen = btrfs_header_generation(parent);
 931                 u64 bytenr = mid->start;
 932                 u32 blocksize = mid->len;
 933
 934                 clean_tree_block(trans, root, mid);
 935                 btrfs_tree_unlock(mid);
 936                 free_extent_buffer(mid);
 937                 mid = NULL;
 938                 wret = del_ptr(trans, root, path, level + 1, pslot);
 939                 if (wret)
 940                         ret = wret;
 941                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 942                                          btrfs_header_owner(parent),
 943                                          root_gen, 0, 0, 1);
 944                 if (wret)
 945                         ret = wret;
 946         } else {
 947                 /* update the parent key to reflect our changes */
 948                 struct btrfs_disk_key mid_key;
 949                 btrfs_node_key(mid, &mid_key, 0);
 950                 btrfs_set_node_key(parent, &mid_key, pslot);
 951                 btrfs_mark_buffer_dirty(parent);
 952         }
 953
 954         /* update the path */
 955         if (left) {
 956                 if (btrfs_header_nritems(left) > orig_slot) {
 957                         extent_buffer_get(left);
 958                         /* left was locked after cow */
 959                         path->nodes[level] = left;
 960                         path->slots[level + 1] -= 1;
 961                         path->slots[level] = orig_slot;
 962                         if (mid) {
 963                                 btrfs_tree_unlock(mid);
 964                                 free_extent_buffer(mid);
 965                         }
 966                 } else {
 967                         orig_slot -= btrfs_header_nritems(left);
 968                         path->slots[level] = orig_slot;
 969                 }
 970         }
 971         /* double check we haven't messed things up */
 972         check_block(root, path, level);
 973         if (orig_ptr !=
 974             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 975                 BUG();
 976 enospc:
 977         if (right) {
 978                 btrfs_tree_unlock(right);
 979                 free_extent_buffer(right);
 980         }
 981         if (left) {
 982                 if (path->nodes[level] != left)
 983                         btrfs_tree_unlock(left);
 984                 free_extent_buffer(left);
 985         }
 986         return ret;
 987 }
 988
 989 /* returns zero if the push worked, non-zero otherwise */
 990 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
 991                                           struct btrfs_root *root,
 992                                           struct btrfs_path *path, int level)
 993 {
 994         struct extent_buffer *right = NULL;
 995         struct extent_buffer *mid;
 996         struct extent_buffer *left = NULL;
 997         struct extent_buffer *parent = NULL;
 998         int ret = 0;
 999         int wret;
1000         int pslot;
1001         int orig_slot = path->slots[level];
1002         u64 orig_ptr;
1003
1004         if (level == 0)
1005                 return 1;
1006
1007         mid = path->nodes[level];
1008         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1009         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1010
1011         if (level < BTRFS_MAX_LEVEL - 1)
1012                 parent = path->nodes[level + 1];
1013         pslot = path->slots[level + 1];
1014
1015         if (!parent)
1016                 return 1;
1017
1018         left = read_node_slot(root, parent, pslot - 1);
1019
1020         /* first, try to make some room in the middle buffer */
1021         if (left) {
1022                 u32 left_nr;
1023
1024                 btrfs_tree_lock(left);
1025                 left_nr = btrfs_header_nritems(left);
1026                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1027                         wret = 1;
1028                 } else {
1029                         ret = btrfs_cow_block(trans, root, left, parent,
1030                                               pslot - 1, &left);
1031                         if (ret)
1032                                 wret = 1;
1033                         else {
1034                                 wret = push_node_left(trans, root,
1035                                                       left, mid, 0);
1036                         }
1037                 }
1038                 if (wret < 0)
1039                         ret = wret;
1040                 if (wret == 0) {
1041                         struct btrfs_disk_key disk_key;
1042                         orig_slot += left_nr;
1043                         btrfs_node_key(mid, &disk_key, 0);
1044                         btrfs_set_node_key(parent, &disk_key, pslot);
1045                         btrfs_mark_buffer_dirty(parent);
1046                         if (btrfs_header_nritems(left) > orig_slot) {
1047                                 path->nodes[level] = left;
1048                                 path->slots[level + 1] -= 1;
1049                                 path->slots[level] = orig_slot;
1050                                 btrfs_tree_unlock(mid);
1051                                 free_extent_buffer(mid);
1052                         } else {
1053                                 orig_slot -=
1054                                         btrfs_header_nritems(left);
1055                                 path->slots[level] = orig_slot;
1056                                 btrfs_tree_unlock(left);
1057                                 free_extent_buffer(left);
1058                         }
1059                         return 0;
1060                 }
1061                 btrfs_tree_unlock(left);
1062                 free_extent_buffer(left);
1063         }
1064         right = read_node_slot(root, parent, pslot + 1);
1065
1066         /*
1067          * then try to empty the right most buffer into the middle
1068          */
1069         if (right) {
1070                 u32 right_nr;
1071                 btrfs_tree_lock(right);
1072                 right_nr = btrfs_header_nritems(right);
1073                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1074                         wret = 1;
1075                 } else {
1076                         ret = btrfs_cow_block(trans, root, right,
1077                                               parent, pslot + 1,
1078                                               &right);
1079                         if (ret)
1080                                 wret = 1;
1081                         else {
1082                                 wret = balance_node_right(trans, root,
1083                                                           right, mid);
1084                         }
1085                 }
1086                 if (wret < 0)
1087                         ret = wret;
1088                 if (wret == 0) {
1089                         struct btrfs_disk_key disk_key;
1090
1091                         btrfs_node_key(right, &disk_key, 0);
1092                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
1093                         btrfs_mark_buffer_dirty(parent);
1094
1095                         if (btrfs_header_nritems(mid) <= orig_slot) {
1096                                 path->nodes[level] = right;
1097                                 path->slots[level + 1] += 1;
1098                                 path->slots[level] = orig_slot -
1099                                         btrfs_header_nritems(mid);
1100                                 btrfs_tree_unlock(mid);
1101                                 free_extent_buffer(mid);
1102                         } else {
1103                                 btrfs_tree_unlock(right);
1104                                 free_extent_buffer(right);
1105                         }
1106                         return 0;
1107                 }
1108                 btrfs_tree_unlock(right);
1109                 free_extent_buffer(right);
1110         }
1111         return 1;
1112 }
1113
1114 /*
1115  * readahead one full node of leaves
1116  */
1117 static void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
1118                              int level, int slot, u64 objectid)
1119 {
1120         struct extent_buffer *node;
1121         struct btrfs_disk_key disk_key;
1122         u32 nritems;
1123         u64 search;
1124         u64 lowest_read;
1125         u64 highest_read;
1126         u64 nread = 0;
1127         int direction = path->reada;
1128         struct extent_buffer *eb;
1129         u32 nr;
1130         u32 blocksize;
1131         u32 nscan = 0;
1132
1133         if (level != 1)
1134                 return;
1135
1136         if (!path->nodes[level])
1137                 return;
1138
1139         node = path->nodes[level];
1140         WARN_ON(!path->skip_locking && !btrfs_tree_locked(node));
1141
1142         search = btrfs_node_blockptr(node, slot);
1143         blocksize = btrfs_level_size(root, level - 1);
1144         eb = btrfs_find_tree_block(root, search, blocksize);
1145         if (eb) {
1146                 free_extent_buffer(eb);
1147                 return;
1148         }
1149
1150         highest_read = search;
1151         lowest_read = search;
1152
1153         nritems = btrfs_header_nritems(node);
1154         nr = slot;
1155         while(1) {
1156                 if (direction < 0) {
1157                         if (nr == 0)
1158                                 break;
1159                         nr--;
1160                 } else if (direction > 0) {
1161                         nr++;
1162                         if (nr >= nritems)
1163                                 break;
1164                 }
1165                 if (path->reada < 0 && objectid) {
1166                         btrfs_node_key(node, &disk_key, nr);
1167                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1168                                 break;
1169                 }
1170                 search = btrfs_node_blockptr(node, nr);
1171                 if ((search >= lowest_read && search <= highest_read) ||
1172                     (search < lowest_read && lowest_read - search <= 32768) ||
1173                     (search > highest_read && search - highest_read <= 32768)) {
1174                         readahead_tree_block(root, search, blocksize,
1175                                      btrfs_node_ptr_generation(node, nr));
1176                         nread += blocksize;
1177                 }
1178                 nscan++;
1179                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
1180                         break;
1181                 if(nread > (1024 * 1024) || nscan > 128)
1182                         break;
1183
1184                 if (search < lowest_read)
1185                         lowest_read = search;
1186                 if (search > highest_read)
1187                         highest_read = search;
1188         }
1189 }
1190
1191 static void unlock_up(struct btrfs_path *path, int level, int lowest_unlock)
1192 {
1193         int i;
1194         int skip_level = level;
1195         struct extent_buffer *t;
1196
1197         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1198                 if (!path->nodes[i])
1199                         break;
1200                 if (!path->locks[i])
1201                         break;
1202                 if (path->slots[i] == 0) {
1203                         skip_level = i + 1;
1204                         continue;
1205                 }
1206                 if (path->keep_locks) {
1207                         u32 nritems;
1208                         t = path->nodes[i];
1209                         nritems = btrfs_header_nritems(t);
1210                         if (nritems < 2 || path->slots[i] >= nritems - 2) {
1211 if (path->keep_locks) {
1212 //printk("path %p skip level now %d\n", path, skip_level);
1213 }
1214                                 skip_level = i + 1;
1215                                 continue;
1216                         }
1217                 }
1218                 t = path->nodes[i];
1219                 if (i >= lowest_unlock && i > skip_level && path->locks[i]) {
1220 if (path->keep_locks) {
1221 //printk("path %p unlocking level %d slot %d nritems %d skip_level %d\n", path, i, path->slots[i], btrfs_header_nritems(t), skip_level);
1222 }
1223                         btrfs_tree_unlock(t);
1224                         path->locks[i] = 0;
1225                 }
1226         }
1227 }
1228
1229 /*
1230  * look for key in the tree.  path is filled in with nodes along the way
1231  * if key is found, we return zero and you can find the item in the leaf
1232  * level of the path (level 0)
1233  *
1234  * If the key isn't found, the path points to the slot where it should
1235  * be inserted, and 1 is returned.  If there are other errors during the
1236  * search a negative error number is returned.
1237  *
1238  * if ins_len > 0, nodes and leaves will be split as we walk down the
1239  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1240  * possible)
1241  */
1242 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1243                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1244                       ins_len, int cow)
1245 {
1246         struct extent_buffer *b;
1247         int slot;
1248         int ret;
1249         int level;
1250         int should_reada = p->reada;
1251         int lowest_unlock = 1;
1252         u8 lowest_level = 0;
1253
1254         lowest_level = p->lowest_level;
1255         WARN_ON(lowest_level && ins_len);
1256         WARN_ON(p->nodes[0] != NULL);
1257         WARN_ON(root == root->fs_info->extent_root &&
1258                 !mutex_is_locked(&root->fs_info->alloc_mutex));
1259         WARN_ON(root == root->fs_info->chunk_root &&
1260                 !mutex_is_locked(&root->fs_info->chunk_mutex));
1261         WARN_ON(root == root->fs_info->dev_root &&
1262                 !mutex_is_locked(&root->fs_info->chunk_mutex));
1263         if (ins_len < 0)
1264                 lowest_unlock = 2;
1265 again:
1266         if (!p->skip_locking)
1267                 b = btrfs_lock_root_node(root);
1268         else
1269                 b = btrfs_root_node(root);
1270
1271         while (b) {
1272                 level = btrfs_header_level(b);
1273                 if (cow) {
1274                         int wret;
1275                         wret = btrfs_cow_block(trans, root, b,
1276                                                p->nodes[level + 1],
1277                                                p->slots[level + 1],
1278                                                &b);
1279                         if (wret) {
1280                                 free_extent_buffer(b);
1281                                 return wret;
1282                         }
1283                 }
1284                 BUG_ON(!cow && ins_len);
1285                 if (level != btrfs_header_level(b))
1286                         WARN_ON(1);
1287                 level = btrfs_header_level(b);
1288                 p->nodes[level] = b;
1289                 if (!p->skip_locking)
1290                         p->locks[level] = 1;
1291                 ret = check_block(root, p, level);
1292                 if (ret)
1293                         return -1;
1294
1295                 ret = bin_search(b, key, level, &slot);
1296                 if (level != 0) {
1297                         if (ret && slot > 0)
1298                                 slot -= 1;
1299                         p->slots[level] = slot;
1300                         if (ins_len > 0 && btrfs_header_nritems(b) >=
1301                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1302                                 int sret = split_node(trans, root, p, level);
1303                                 BUG_ON(sret > 0);
1304                                 if (sret)
1305                                         return sret;
1306                                 b = p->nodes[level];
1307                                 slot = p->slots[level];
1308                         } else if (ins_len < 0) {
1309                                 int sret = balance_level(trans, root, p,
1310                                                          level);
1311                                 if (sret)
1312                                         return sret;
1313                                 b = p->nodes[level];
1314                                 if (!b) {
1315                                         btrfs_release_path(NULL, p);
1316                                         goto again;
1317                                 }
1318                                 slot = p->slots[level];
1319                                 BUG_ON(btrfs_header_nritems(b) == 1);
1320                         }
1321                         /* this is only true while dropping a snapshot */
1322                         if (level == lowest_level) {
1323                                 unlock_up(p, level, lowest_unlock);
1324                                 break;
1325                         }
1326
1327                         if (should_reada)
1328                                 reada_for_search(root, p, level, slot,
1329                                                  key->objectid);
1330
1331                         b = read_node_slot(root, b, slot);
1332                         if (!p->skip_locking)
1333                                 btrfs_tree_lock(b);
1334                         unlock_up(p, level + 1, lowest_unlock);
1335                 } else {
1336                         p->slots[level] = slot;
1337                         if (ins_len > 0 && btrfs_leaf_free_space(root, b) <
1338                             sizeof(struct btrfs_item) + ins_len) {
1339                                 int sret = split_leaf(trans, root, key,
1340                                                       p, ins_len, ret == 0);
1341                                 BUG_ON(sret > 0);
1342                                 if (sret)
1343                                         return sret;
1344                         }
1345                         unlock_up(p, level, lowest_unlock);
1346                         return ret;
1347                 }
1348         }
1349         return 1;
1350 }
1351
1352 /*
1353  * adjust the pointers going up the tree, starting at level
1354  * making sure the right key of each node is points to 'key'.
1355  * This is used after shifting pointers to the left, so it stops
1356  * fixing up pointers when a given leaf/node is not in slot 0 of the
1357  * higher levels
1358  *
1359  * If this fails to write a tree block, it returns -1, but continues
1360  * fixing up the blocks in ram so the tree is consistent.
1361  */
1362 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1363                           struct btrfs_root *root, struct btrfs_path *path,
1364                           struct btrfs_disk_key *key, int level)
1365 {
1366         int i;
1367         int ret = 0;
1368         struct extent_buffer *t;
1369
1370         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1371                 int tslot = path->slots[i];
1372                 if (!path->nodes[i])
1373                         break;
1374                 t = path->nodes[i];
1375                 btrfs_set_node_key(t, key, tslot);
1376                 if (!btrfs_tree_locked(path->nodes[i])) {
1377                         int ii;
1378 printk("fixup without lock on level %d\n", btrfs_header_level(path->nodes[i]));
1379                         for (ii = 0; ii < BTRFS_MAX_LEVEL; ii++) {
1380 printk("level %d slot %d\n", ii, path->slots[ii]);
1381                         }
1382                 }
1383                 btrfs_mark_buffer_dirty(path->nodes[i]);
1384                 if (tslot != 0)
1385                         break;
1386         }
1387         return ret;
1388 }
1389
1390 /*
1391  * try to push data from one node into the next node left in the
1392  * tree.
1393  *
1394  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1395  * error, and > 0 if there was no room in the left hand block.
1396  */
1397 static int push_node_left(struct btrfs_trans_handle *trans,
1398                           struct btrfs_root *root, struct extent_buffer *dst,
1399                           struct extent_buffer *src, int empty)
1400 {
1401         int push_items = 0;
1402         int src_nritems;
1403         int dst_nritems;
1404         int ret = 0;
1405
1406         src_nritems = btrfs_header_nritems(src);
1407         dst_nritems = btrfs_header_nritems(dst);
1408         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1409         WARN_ON(btrfs_header_generation(src) != trans->transid);
1410         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1411
1412         if (!empty && src_nritems <= 8)
1413                 return 1;
1414
1415         if (push_items <= 0) {
1416                 return 1;
1417         }
1418
1419         if (empty) {
1420                 push_items = min(src_nritems, push_items);
1421                 if (push_items < src_nritems) {
1422                         /* leave at least 8 pointers in the node if
1423                          * we aren't going to empty it
1424                          */
1425                         if (src_nritems - push_items < 8) {
1426                                 if (push_items <= 8)
1427                                         return 1;
1428                                 push_items -= 8;
1429                         }
1430                 }
1431         } else
1432                 push_items = min(src_nritems - 8, push_items);
1433
1434         copy_extent_buffer(dst, src,
1435                            btrfs_node_key_ptr_offset(dst_nritems),
1436                            btrfs_node_key_ptr_offset(0),
1437                            push_items * sizeof(struct btrfs_key_ptr));
1438
1439         if (push_items < src_nritems) {
1440                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1441                                       btrfs_node_key_ptr_offset(push_items),
1442                                       (src_nritems - push_items) *
1443                                       sizeof(struct btrfs_key_ptr));
1444         }
1445         btrfs_set_header_nritems(src, src_nritems - push_items);
1446         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1447         btrfs_mark_buffer_dirty(src);
1448         btrfs_mark_buffer_dirty(dst);
1449         return ret;
1450 }
1451
1452 /*
1453  * try to push data from one node into the next node right in the
1454  * tree.
1455  *
1456  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1457  * error, and > 0 if there was no room in the right hand block.
1458  *
1459  * this will  only push up to 1/2 the contents of the left node over
1460  */
1461 static int balance_node_right(struct btrfs_trans_handle *trans,
1462                               struct btrfs_root *root,
1463                               struct extent_buffer *dst,
1464                               struct extent_buffer *src)
1465 {
1466         int push_items = 0;
1467         int max_push;
1468         int src_nritems;
1469         int dst_nritems;
1470         int ret = 0;
1471
1472         WARN_ON(btrfs_header_generation(src) != trans->transid);
1473         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1474
1475         src_nritems = btrfs_header_nritems(src);
1476         dst_nritems = btrfs_header_nritems(dst);
1477         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1478         if (push_items <= 0) {
1479                 return 1;
1480         }
1481
1482         if (src_nritems < 4) {
1483                 return 1;
1484         }
1485
1486         max_push = src_nritems / 2 + 1;
1487         /* don't try to empty the node */
1488         if (max_push >= src_nritems) {
1489                 return 1;
1490         }
1491
1492         if (max_push < push_items)
1493                 push_items = max_push;
1494
1495         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1496                                       btrfs_node_key_ptr_offset(0),
1497                                       (dst_nritems) *
1498                                       sizeof(struct btrfs_key_ptr));
1499
1500         copy_extent_buffer(dst, src,
1501                            btrfs_node_key_ptr_offset(0),
1502                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1503                            push_items * sizeof(struct btrfs_key_ptr));
1504
1505         btrfs_set_header_nritems(src, src_nritems - push_items);
1506         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1507
1508         btrfs_mark_buffer_dirty(src);
1509         btrfs_mark_buffer_dirty(dst);
1510         return ret;
1511 }
1512
1513 /*
1514  * helper function to insert a new root level in the tree.
1515  * A new node is allocated, and a single item is inserted to
1516  * point to the existing root
1517  *
1518  * returns zero on success or < 0 on failure.
1519  */
1520 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1521                            struct btrfs_root *root,
1522                            struct btrfs_path *path, int level)
1523 {
1524         u64 root_gen;
1525         u64 lower_gen;
1526         struct extent_buffer *lower;
1527         struct extent_buffer *c;
1528         struct extent_buffer *old;
1529         struct btrfs_disk_key lower_key;
1530
1531         BUG_ON(path->nodes[level]);
1532         BUG_ON(path->nodes[level-1] != root->node);
1533
1534         if (root->ref_cows)
1535                 root_gen = trans->transid;
1536         else
1537                 root_gen = 0;
1538
1539         lower = path->nodes[level-1];
1540         if (level == 1)
1541                 btrfs_item_key(lower, &lower_key, 0);
1542         else
1543                 btrfs_node_key(lower, &lower_key, 0);
1544
1545         c = btrfs_alloc_free_block(trans, root, root->nodesize,
1546                                    root->root_key.objectid,
1547                                    root_gen, lower_key.objectid, level,
1548                                    root->node->start, 0);
1549         if (IS_ERR(c))
1550                 return PTR_ERR(c);
1551
1552         memset_extent_buffer(c, 0, 0, root->nodesize);
1553         btrfs_set_header_nritems(c, 1);
1554         btrfs_set_header_level(c, level);
1555         btrfs_set_header_bytenr(c, c->start);
1556         btrfs_set_header_generation(c, trans->transid);
1557         btrfs_set_header_owner(c, root->root_key.objectid);
1558
1559         write_extent_buffer(c, root->fs_info->fsid,
1560                             (unsigned long)btrfs_header_fsid(c),
1561                             BTRFS_FSID_SIZE);
1562
1563         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1564                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
1565                             BTRFS_UUID_SIZE);
1566
1567         btrfs_set_node_key(c, &lower_key, 0);
1568         btrfs_set_node_blockptr(c, 0, lower->start);
1569         lower_gen = btrfs_header_generation(lower);
1570         WARN_ON(lower_gen == 0);
1571
1572         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1573
1574         btrfs_mark_buffer_dirty(c);
1575
1576         spin_lock(&root->node_lock);
1577         old = root->node;
1578         root->node = c;
1579         spin_unlock(&root->node_lock);
1580
1581         /* the super has an extra ref to root->node */
1582         free_extent_buffer(old);
1583
1584         add_root_to_dirty_list(root);
1585         extent_buffer_get(c);
1586         path->nodes[level] = c;
1587         path->locks[level] = 1;
1588         path->slots[level] = 0;
1589
1590         if (root->ref_cows && lower_gen != trans->transid) {
1591                 struct btrfs_path *back_path = btrfs_alloc_path();
1592                 int ret;
1593                 mutex_lock(&root->fs_info->alloc_mutex);
1594                 ret = btrfs_insert_extent_backref(trans,
1595                                                   root->fs_info->extent_root,
1596                                                   path, lower->start,
1597                                                   root->root_key.objectid,
1598                                                   trans->transid, 0, 0);
1599                 BUG_ON(ret);
1600                 mutex_unlock(&root->fs_info->alloc_mutex);
1601                 btrfs_free_path(back_path);
1602         }
1603         return 0;
1604 }
1605
1606 /*
1607  * worker function to insert a single pointer in a node.
1608  * the node should have enough room for the pointer already
1609  *
1610  * slot and level indicate where you want the key to go, and
1611  * blocknr is the block the key points to.
1612  *
1613  * returns zero on success and < 0 on any error
1614  */
1615 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1616                       *root, struct btrfs_path *path, struct btrfs_disk_key
1617                       *key, u64 bytenr, int slot, int level)
1618 {
1619         struct extent_buffer *lower;
1620         int nritems;
1621
1622         BUG_ON(!path->nodes[level]);
1623         lower = path->nodes[level];
1624         nritems = btrfs_header_nritems(lower);
1625         if (slot > nritems)
1626                 BUG();
1627         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1628                 BUG();
1629         if (slot != nritems) {
1630                 memmove_extent_buffer(lower,
1631                               btrfs_node_key_ptr_offset(slot + 1),
1632                               btrfs_node_key_ptr_offset(slot),
1633                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1634         }
1635         btrfs_set_node_key(lower, key, slot);
1636         btrfs_set_node_blockptr(lower, slot, bytenr);
1637         WARN_ON(trans->transid == 0);
1638         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1639         btrfs_set_header_nritems(lower, nritems + 1);
1640         btrfs_mark_buffer_dirty(lower);
1641         return 0;
1642 }
1643
1644 /*
1645  * split the node at the specified level in path in two.
1646  * The path is corrected to point to the appropriate node after the split
1647  *
1648  * Before splitting this tries to make some room in the node by pushing
1649  * left and right, if either one works, it returns right away.
1650  *
1651  * returns 0 on success and < 0 on failure
1652  */
1653 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1654                       *root, struct btrfs_path *path, int level)
1655 {
1656         u64 root_gen;
1657         struct extent_buffer *c;
1658         struct extent_buffer *split;
1659         struct btrfs_disk_key disk_key;
1660         int mid;
1661         int ret;
1662         int wret;
1663         u32 c_nritems;
1664
1665         c = path->nodes[level];
1666         WARN_ON(btrfs_header_generation(c) != trans->transid);
1667         if (c == root->node) {
1668                 /* trying to split the root, lets make a new one */
1669                 ret = insert_new_root(trans, root, path, level + 1);
1670                 if (ret)
1671                         return ret;
1672         } else {
1673                 ret = push_nodes_for_insert(trans, root, path, level);
1674                 c = path->nodes[level];
1675                 if (!ret && btrfs_header_nritems(c) <
1676                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
1677                         return 0;
1678                 if (ret < 0)
1679                         return ret;
1680         }
1681
1682         c_nritems = btrfs_header_nritems(c);
1683         if (root->ref_cows)
1684                 root_gen = trans->transid;
1685         else
1686                 root_gen = 0;
1687
1688         btrfs_node_key(c, &disk_key, 0);
1689         split = btrfs_alloc_free_block(trans, root, root->nodesize,
1690                                          root->root_key.objectid,
1691                                          root_gen,
1692                                          btrfs_disk_key_objectid(&disk_key),
1693                                          level, c->start, 0);
1694         if (IS_ERR(split))
1695                 return PTR_ERR(split);
1696
1697         btrfs_set_header_flags(split, btrfs_header_flags(c));
1698         btrfs_set_header_level(split, btrfs_header_level(c));
1699         btrfs_set_header_bytenr(split, split->start);
1700         btrfs_set_header_generation(split, trans->transid);
1701         btrfs_set_header_owner(split, root->root_key.objectid);
1702         btrfs_set_header_flags(split, 0);
1703         write_extent_buffer(split, root->fs_info->fsid,
1704                             (unsigned long)btrfs_header_fsid(split),
1705                             BTRFS_FSID_SIZE);
1706         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1707                             (unsigned long)btrfs_header_chunk_tree_uuid(split),
1708                             BTRFS_UUID_SIZE);
1709
1710         mid = (c_nritems + 1) / 2;
1711
1712         copy_extent_buffer(split, c,
1713                            btrfs_node_key_ptr_offset(0),
1714                            btrfs_node_key_ptr_offset(mid),
1715                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1716         btrfs_set_header_nritems(split, c_nritems - mid);
1717         btrfs_set_header_nritems(c, mid);
1718         ret = 0;
1719
1720         btrfs_mark_buffer_dirty(c);
1721         btrfs_mark_buffer_dirty(split);
1722
1723         btrfs_node_key(split, &disk_key, 0);
1724         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1725                           path->slots[level + 1] + 1,
1726                           level + 1);
1727         if (wret)
1728                 ret = wret;
1729
1730         if (path->slots[level] >= mid) {
1731                 path->slots[level] -= mid;
1732                 btrfs_tree_unlock(c);
1733                 free_extent_buffer(c);
1734                 path->nodes[level] = split;
1735                 path->slots[level + 1] += 1;
1736         } else {
1737                 btrfs_tree_unlock(split);
1738                 free_extent_buffer(split);
1739         }
1740         return ret;
1741 }
1742
1743 /*
1744  * how many bytes are required to store the items in a leaf.  start
1745  * and nr indicate which items in the leaf to check.  This totals up the
1746  * space used both by the item structs and the item data
1747  */
1748 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1749 {
1750         int data_len;
1751         int nritems = btrfs_header_nritems(l);
1752         int end = min(nritems, start + nr) - 1;
1753
1754         if (!nr)
1755                 return 0;
1756         data_len = btrfs_item_end_nr(l, start);
1757         data_len = data_len - btrfs_item_offset_nr(l, end);
1758         data_len += sizeof(struct btrfs_item) * nr;
1759         WARN_ON(data_len < 0);
1760         return data_len;
1761 }
1762
1763 /*
1764  * The space between the end of the leaf items and
1765  * the start of the leaf data.  IOW, how much room
1766  * the leaf has left for both items and data
1767  */
1768 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1769 {
1770         int nritems = btrfs_header_nritems(leaf);
1771         int ret;
1772         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1773         if (ret < 0) {
1774                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1775                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1776                        leaf_space_used(leaf, 0, nritems), nritems);
1777         }
1778         return ret;
1779 }
1780
1781 /*
1782  * push some data in the path leaf to the right, trying to free up at
1783  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1784  *
1785  * returns 1 if the push failed because the other node didn't have enough
1786  * room, 0 if everything worked out and < 0 if there were major errors.
1787  */
1788 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1789                            *root, struct btrfs_path *path, int data_size,
1790                            int empty)
1791 {
1792         struct extent_buffer *left = path->nodes[0];
1793         struct extent_buffer *right;
1794         struct extent_buffer *upper;
1795         struct btrfs_disk_key disk_key;
1796         int slot;
1797         u32 i;
1798         int free_space;
1799         int push_space = 0;
1800         int push_items = 0;
1801         struct btrfs_item *item;
1802         u32 left_nritems;
1803         u32 nr;
1804         u32 right_nritems;
1805         u32 data_end;
1806         u32 this_item_size;
1807         int ret;
1808
1809         slot = path->slots[1];
1810         if (!path->nodes[1]) {
1811                 return 1;
1812         }
1813         upper = path->nodes[1];
1814         if (slot >= btrfs_header_nritems(upper) - 1)
1815                 return 1;
1816
1817         WARN_ON(!btrfs_tree_locked(path->nodes[1]));
1818
1819         right = read_node_slot(root, upper, slot + 1);
1820         btrfs_tree_lock(right);
1821         free_space = btrfs_leaf_free_space(root, right);
1822         if (free_space < data_size + sizeof(struct btrfs_item))
1823                 goto out_unlock;
1824
1825         /* cow and double check */
1826         ret = btrfs_cow_block(trans, root, right, upper,
1827                               slot + 1, &right);
1828         if (ret)
1829                 goto out_unlock;
1830
1831         free_space = btrfs_leaf_free_space(root, right);
1832         if (free_space < data_size + sizeof(struct btrfs_item))
1833                 goto out_unlock;
1834
1835         left_nritems = btrfs_header_nritems(left);
1836         if (left_nritems == 0)
1837                 goto out_unlock;
1838
1839         if (empty)
1840                 nr = 0;
1841         else
1842                 nr = 1;
1843
1844         i = left_nritems - 1;
1845         while (i >= nr) {
1846                 item = btrfs_item_nr(left, i);
1847
1848                 if (path->slots[0] == i)
1849                         push_space += data_size + sizeof(*item);
1850
1851                 if (!left->map_token) {
1852                         map_extent_buffer(left, (unsigned long)item,
1853                                         sizeof(struct btrfs_item),
1854                                         &left->map_token, &left->kaddr,
1855                                         &left->map_start, &left->map_len,
1856                                         KM_USER1);
1857                 }
1858
1859                 this_item_size = btrfs_item_size(left, item);
1860                 if (this_item_size + sizeof(*item) + push_space > free_space)
1861                         break;
1862                 push_items++;
1863                 push_space += this_item_size + sizeof(*item);
1864                 if (i == 0)
1865                         break;
1866                 i--;
1867         }
1868         if (left->map_token) {
1869                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1870                 left->map_token = NULL;
1871         }
1872
1873         if (push_items == 0)
1874                 goto out_unlock;
1875
1876         if (!empty && push_items == left_nritems)
1877                 WARN_ON(1);
1878
1879         /* push left to right */
1880         right_nritems = btrfs_header_nritems(right);
1881
1882         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1883         push_space -= leaf_data_end(root, left);
1884
1885         /* make room in the right data area */
1886         data_end = leaf_data_end(root, right);
1887         memmove_extent_buffer(right,
1888                               btrfs_leaf_data(right) + data_end - push_space,
1889                               btrfs_leaf_data(right) + data_end,
1890                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1891
1892         /* copy from the left data area */
1893         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1894                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1895                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1896                      push_space);
1897
1898         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1899                               btrfs_item_nr_offset(0),
1900                               right_nritems * sizeof(struct btrfs_item));
1901
1902         /* copy the items from left to right */
1903         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1904                    btrfs_item_nr_offset(left_nritems - push_items),
1905                    push_items * sizeof(struct btrfs_item));
1906
1907         /* update the item pointers */
1908         right_nritems += push_items;
1909         btrfs_set_header_nritems(right, right_nritems);
1910         push_space = BTRFS_LEAF_DATA_SIZE(root);
1911         for (i = 0; i < right_nritems; i++) {
1912                 item = btrfs_item_nr(right, i);
1913                 if (!right->map_token) {
1914                         map_extent_buffer(right, (unsigned long)item,
1915                                         sizeof(struct btrfs_item),
1916                                         &right->map_token, &right->kaddr,
1917                                         &right->map_start, &right->map_len,
1918                                         KM_USER1);
1919                 }
1920                 push_space -= btrfs_item_size(right, item);
1921                 btrfs_set_item_offset(right, item, push_space);
1922         }
1923
1924         if (right->map_token) {
1925                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1926                 right->map_token = NULL;
1927         }
1928         left_nritems -= push_items;
1929         btrfs_set_header_nritems(left, left_nritems);
1930
1931         if (left_nritems)
1932                 btrfs_mark_buffer_dirty(left);
1933         btrfs_mark_buffer_dirty(right);
1934
1935         btrfs_item_key(right, &disk_key, 0);
1936         btrfs_set_node_key(upper, &disk_key, slot + 1);
1937         btrfs_mark_buffer_dirty(upper);
1938
1939         /* then fixup the leaf pointer in the path */
1940         if (path->slots[0] >= left_nritems) {
1941                 path->slots[0] -= left_nritems;
1942                 if (btrfs_header_nritems(path->nodes[0]) == 0)
1943                         clean_tree_block(trans, root, path->nodes[0]);
1944                 btrfs_tree_unlock(path->nodes[0]);
1945                 free_extent_buffer(path->nodes[0]);
1946                 path->nodes[0] = right;
1947                 path->slots[1] += 1;
1948         } else {
1949                 btrfs_tree_unlock(right);
1950                 free_extent_buffer(right);
1951         }
1952         return 0;
1953
1954 out_unlock:
1955         btrfs_tree_unlock(right);
1956         free_extent_buffer(right);
1957         return 1;
1958 }
1959
1960 /*
1961  * push some data in the path leaf to the left, trying to free up at
1962  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1963  */
1964 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1965                           *root, struct btrfs_path *path, int data_size,
1966                           int empty)
1967 {
1968         struct btrfs_disk_key disk_key;
1969         struct extent_buffer *right = path->nodes[0];
1970         struct extent_buffer *left;
1971         int slot;
1972         int i;
1973         int free_space;
1974         int push_space = 0;
1975         int push_items = 0;
1976         struct btrfs_item *item;
1977         u32 old_left_nritems;
1978         u32 right_nritems;
1979         u32 nr;
1980         int ret = 0;
1981         int wret;
1982         u32 this_item_size;
1983         u32 old_left_item_size;
1984
1985         slot = path->slots[1];
1986         if (slot == 0)
1987                 return 1;
1988         if (!path->nodes[1])
1989                 return 1;
1990
1991         right_nritems = btrfs_header_nritems(right);
1992         if (right_nritems == 0) {
1993                 return 1;
1994         }
1995
1996         WARN_ON(!btrfs_tree_locked(path->nodes[1]));
1997
1998         left = read_node_slot(root, path->nodes[1], slot - 1);
1999         btrfs_tree_lock(left);
2000         free_space = btrfs_leaf_free_space(root, left);
2001         if (free_space < data_size + sizeof(struct btrfs_item)) {
2002                 ret = 1;
2003                 goto out;
2004         }
2005
2006         /* cow and double check */
2007         ret = btrfs_cow_block(trans, root, left,
2008                               path->nodes[1], slot - 1, &left);
2009         if (ret) {
2010                 /* we hit -ENOSPC, but it isn't fatal here */
2011                 ret = 1;
2012                 goto out;
2013         }
2014
2015         free_space = btrfs_leaf_free_space(root, left);
2016         if (free_space < data_size + sizeof(struct btrfs_item)) {
2017                 ret = 1;
2018                 goto out;
2019         }
2020
2021         if (empty)
2022                 nr = right_nritems;
2023         else
2024                 nr = right_nritems - 1;
2025
2026         for (i = 0; i < nr; i++) {
2027                 item = btrfs_item_nr(right, i);
2028                 if (!right->map_token) {
2029                         map_extent_buffer(right, (unsigned long)item,
2030                                         sizeof(struct btrfs_item),
2031                                         &right->map_token, &right->kaddr,
2032                                         &right->map_start, &right->map_len,
2033                                         KM_USER1);
2034                 }
2035
2036                 if (path->slots[0] == i)
2037                         push_space += data_size + sizeof(*item);
2038
2039                 this_item_size = btrfs_item_size(right, item);
2040                 if (this_item_size + sizeof(*item) + push_space > free_space)
2041                         break;
2042
2043                 push_items++;
2044                 push_space += this_item_size + sizeof(*item);
2045         }
2046
2047         if (right->map_token) {
2048                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2049                 right->map_token = NULL;
2050         }
2051
2052         if (push_items == 0) {
2053                 ret = 1;
2054                 goto out;
2055         }
2056         if (!empty && push_items == btrfs_header_nritems(right))
2057                 WARN_ON(1);
2058
2059         /* push data from right to left */
2060         copy_extent_buffer(left, right,
2061                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
2062                            btrfs_item_nr_offset(0),
2063                            push_items * sizeof(struct btrfs_item));
2064
2065         push_space = BTRFS_LEAF_DATA_SIZE(root) -
2066                      btrfs_item_offset_nr(right, push_items -1);
2067
2068         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
2069                      leaf_data_end(root, left) - push_space,
2070                      btrfs_leaf_data(right) +
2071                      btrfs_item_offset_nr(right, push_items - 1),
2072                      push_space);
2073         old_left_nritems = btrfs_header_nritems(left);
2074         BUG_ON(old_left_nritems < 0);
2075
2076         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
2077         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
2078                 u32 ioff;
2079
2080                 item = btrfs_item_nr(left, i);
2081                 if (!left->map_token) {
2082                         map_extent_buffer(left, (unsigned long)item,
2083                                         sizeof(struct btrfs_item),
2084                                         &left->map_token, &left->kaddr,
2085                                         &left->map_start, &left->map_len,
2086                                         KM_USER1);
2087                 }
2088
2089                 ioff = btrfs_item_offset(left, item);
2090                 btrfs_set_item_offset(left, item,
2091                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
2092         }
2093         btrfs_set_header_nritems(left, old_left_nritems + push_items);
2094         if (left->map_token) {
2095                 unmap_extent_buffer(left, left->map_token, KM_USER1);
2096                 left->map_token = NULL;
2097         }
2098
2099         /* fixup right node */
2100         if (push_items > right_nritems) {
2101                 printk("push items %d nr %u\n", push_items, right_nritems);
2102                 WARN_ON(1);
2103         }
2104
2105         if (push_items < right_nritems) {
2106                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
2107                                                   leaf_data_end(root, right);
2108                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
2109                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
2110                                       btrfs_leaf_data(right) +
2111                                       leaf_data_end(root, right), push_space);
2112
2113                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
2114                               btrfs_item_nr_offset(push_items),
2115                              (btrfs_header_nritems(right) - push_items) *
2116                              sizeof(struct btrfs_item));
2117         }
2118         right_nritems -= push_items;
2119         btrfs_set_header_nritems(right, right_nritems);
2120         push_space = BTRFS_LEAF_DATA_SIZE(root);
2121         for (i = 0; i < right_nritems; i++) {
2122                 item = btrfs_item_nr(right, i);
2123
2124                 if (!right->map_token) {
2125                         map_extent_buffer(right, (unsigned long)item,
2126                                         sizeof(struct btrfs_item),
2127                                         &right->map_token, &right->kaddr,
2128                                         &right->map_start, &right->map_len,
2129                                         KM_USER1);
2130                 }
2131
2132                 push_space = push_space - btrfs_item_size(right, item);
2133                 btrfs_set_item_offset(right, item, push_space);
2134         }
2135         if (right->map_token) {
2136                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2137                 right->map_token = NULL;
2138         }
2139
2140         btrfs_mark_buffer_dirty(left);
2141         if (right_nritems)
2142                 btrfs_mark_buffer_dirty(right);
2143
2144         btrfs_item_key(right, &disk_key, 0);
2145         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
2146         if (wret)
2147                 ret = wret;
2148
2149         /* then fixup the leaf pointer in the path */
2150         if (path->slots[0] < push_items) {
2151                 path->slots[0] += old_left_nritems;
2152                 if (btrfs_header_nritems(path->nodes[0]) == 0)
2153                         clean_tree_block(trans, root, path->nodes[0]);
2154                 btrfs_tree_unlock(path->nodes[0]);
2155                 free_extent_buffer(path->nodes[0]);
2156                 path->nodes[0] = left;
2157                 path->slots[1] -= 1;
2158         } else {
2159                 btrfs_tree_unlock(left);
2160                 free_extent_buffer(left);
2161                 path->slots[0] -= push_items;
2162         }
2163         BUG_ON(path->slots[0] < 0);
2164         return ret;
2165 out:
2166         btrfs_tree_unlock(left);
2167         free_extent_buffer(left);
2168         return ret;
2169 }
2170
2171 /*
2172  * split the path's leaf in two, making sure there is at least data_size
2173  * available for the resulting leaf level of the path.
2174  *
2175  * returns 0 if all went well and < 0 on failure.
2176  */
2177 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
2178                       *root, struct btrfs_key *ins_key,
2179                       struct btrfs_path *path, int data_size, int extend)
2180 {
2181         u64 root_gen;
2182         struct extent_buffer *l;
2183         u32 nritems;
2184         int mid;
2185         int slot;
2186         struct extent_buffer *right;
2187         int space_needed = data_size + sizeof(struct btrfs_item);
2188         int data_copy_size;
2189         int rt_data_off;
2190         int i;
2191         int ret = 0;
2192         int wret;
2193         int double_split;
2194         int num_doubles = 0;
2195         struct btrfs_disk_key disk_key;
2196
2197         if (extend)
2198                 space_needed = data_size;
2199
2200         if (root->ref_cows)
2201                 root_gen = trans->transid;
2202         else
2203                 root_gen = 0;
2204
2205         /* first try to make some room by pushing left and right */
2206         if (ins_key->type != BTRFS_DIR_ITEM_KEY) {
2207                 wret = push_leaf_right(trans, root, path, data_size, 0);
2208                 if (wret < 0) {
2209                         return wret;
2210                 }
2211                 if (wret) {
2212                         wret = push_leaf_left(trans, root, path, data_size, 0);
2213                         if (wret < 0)
2214                                 return wret;
2215                 }
2216                 l = path->nodes[0];
2217
2218                 /* did the pushes work? */
2219                 if (btrfs_leaf_free_space(root, l) >= space_needed)
2220                         return 0;
2221         }
2222
2223         if (!path->nodes[1]) {
2224                 ret = insert_new_root(trans, root, path, 1);
2225                 if (ret)
2226                         return ret;
2227         }
2228 again:
2229         double_split = 0;
2230         l = path->nodes[0];
2231         slot = path->slots[0];
2232         nritems = btrfs_header_nritems(l);
2233         mid = (nritems + 1)/ 2;
2234
2235         btrfs_item_key(l, &disk_key, 0);
2236
2237         right = btrfs_alloc_free_block(trans, root, root->leafsize,
2238                                          root->root_key.objectid,
2239                                          root_gen, disk_key.objectid, 0,
2240                                          l->start, 0);
2241         if (IS_ERR(right)) {
2242                 BUG_ON(1);
2243                 return PTR_ERR(right);
2244         }
2245
2246         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2247         btrfs_set_header_bytenr(right, right->start);
2248         btrfs_set_header_generation(right, trans->transid);
2249         btrfs_set_header_owner(right, root->root_key.objectid);
2250         btrfs_set_header_level(right, 0);
2251         write_extent_buffer(right, root->fs_info->fsid,
2252                             (unsigned long)btrfs_header_fsid(right),
2253                             BTRFS_FSID_SIZE);
2254
2255         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2256                             (unsigned long)btrfs_header_chunk_tree_uuid(right),
2257                             BTRFS_UUID_SIZE);
2258         if (mid <= slot) {
2259                 if (nritems == 1 ||
2260                     leaf_space_used(l, mid, nritems - mid) + space_needed >
2261                         BTRFS_LEAF_DATA_SIZE(root)) {
2262                         if (slot >= nritems) {
2263                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2264                                 btrfs_set_header_nritems(right, 0);
2265                                 wret = insert_ptr(trans, root, path,
2266                                                   &disk_key, right->start,
2267                                                   path->slots[1] + 1, 1);
2268                                 if (wret)
2269                                         ret = wret;
2270
2271                                 btrfs_tree_unlock(path->nodes[0]);
2272                                 free_extent_buffer(path->nodes[0]);
2273                                 path->nodes[0] = right;
2274                                 path->slots[0] = 0;
2275                                 path->slots[1] += 1;
2276                                 btrfs_mark_buffer_dirty(right);
2277                                 return ret;
2278                         }
2279                         mid = slot;
2280                         if (mid != nritems &&
2281                             leaf_space_used(l, mid, nritems - mid) +
2282                             space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2283                                 double_split = 1;
2284                         }
2285                 }
2286         } else {
2287                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
2288                         BTRFS_LEAF_DATA_SIZE(root)) {
2289                         if (!extend && slot == 0) {
2290                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2291                                 btrfs_set_header_nritems(right, 0);
2292                                 wret = insert_ptr(trans, root, path,
2293                                                   &disk_key,
2294                                                   right->start,
2295                                                   path->slots[1], 1);
2296                                 if (wret)
2297                                         ret = wret;
2298                                 btrfs_tree_unlock(path->nodes[0]);
2299                                 free_extent_buffer(path->nodes[0]);
2300                                 path->nodes[0] = right;
2301                                 path->slots[0] = 0;
2302                                 if (path->slots[1] == 0) {
2303                                         wret = fixup_low_keys(trans, root,
2304                                                    path, &disk_key, 1);
2305                                         if (wret)
2306                                                 ret = wret;
2307                                 }
2308                                 btrfs_mark_buffer_dirty(right);
2309                                 return ret;
2310                         } else if (extend && slot == 0) {
2311                                 mid = 1;
2312                         } else {
2313                                 mid = slot;
2314                                 if (mid != nritems &&
2315                                     leaf_space_used(l, mid, nritems - mid) +
2316                                     space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2317                                         double_split = 1;
2318                                 }
2319                         }
2320                 }
2321         }
2322         nritems = nritems - mid;
2323         btrfs_set_header_nritems(right, nritems);
2324         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2325
2326         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2327                            btrfs_item_nr_offset(mid),
2328                            nritems * sizeof(struct btrfs_item));
2329
2330         copy_extent_buffer(right, l,
2331                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2332                      data_copy_size, btrfs_leaf_data(l) +
2333                      leaf_data_end(root, l), data_copy_size);
2334
2335         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2336                       btrfs_item_end_nr(l, mid);
2337
2338         for (i = 0; i < nritems; i++) {
2339                 struct btrfs_item *item = btrfs_item_nr(right, i);
2340                 u32 ioff;
2341
2342                 if (!right->map_token) {
2343                         map_extent_buffer(right, (unsigned long)item,
2344                                         sizeof(struct btrfs_item),
2345                                         &right->map_token, &right->kaddr,
2346                                         &right->map_start, &right->map_len,
2347                                         KM_USER1);
2348                 }
2349
2350                 ioff = btrfs_item_offset(right, item);
2351                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2352         }
2353
2354         if (right->map_token) {
2355                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2356                 right->map_token = NULL;
2357         }
2358
2359         btrfs_set_header_nritems(l, mid);
2360         ret = 0;
2361         btrfs_item_key(right, &disk_key, 0);
2362         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2363                           path->slots[1] + 1, 1);
2364         if (wret)
2365                 ret = wret;
2366
2367         btrfs_mark_buffer_dirty(right);
2368         btrfs_mark_buffer_dirty(l);
2369         BUG_ON(path->slots[0] != slot);
2370
2371         if (mid <= slot) {
2372                 btrfs_tree_unlock(path->nodes[0]);
2373                 free_extent_buffer(path->nodes[0]);
2374                 path->nodes[0] = right;
2375                 path->slots[0] -= mid;
2376                 path->slots[1] += 1;
2377         } else {
2378                 btrfs_tree_unlock(right);
2379                 free_extent_buffer(right);
2380         }
2381
2382         BUG_ON(path->slots[0] < 0);
2383
2384         if (double_split) {
2385                 BUG_ON(num_doubles != 0);
2386                 num_doubles++;
2387                 goto again;
2388         }
2389         return ret;
2390 }
2391
2392 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2393                         struct btrfs_root *root,
2394                         struct btrfs_path *path,
2395                         u32 new_size, int from_end)
2396 {
2397         int ret = 0;
2398         int slot;
2399         int slot_orig;
2400         struct extent_buffer *leaf;
2401         struct btrfs_item *item;
2402         u32 nritems;
2403         unsigned int data_end;
2404         unsigned int old_data_start;
2405         unsigned int old_size;
2406         unsigned int size_diff;
2407         int i;
2408
2409         slot_orig = path->slots[0];
2410         leaf = path->nodes[0];
2411         slot = path->slots[0];
2412
2413         old_size = btrfs_item_size_nr(leaf, slot);
2414         if (old_size == new_size)
2415                 return 0;
2416
2417         nritems = btrfs_header_nritems(leaf);
2418         data_end = leaf_data_end(root, leaf);
2419
2420         old_data_start = btrfs_item_offset_nr(leaf, slot);
2421
2422         size_diff = old_size - new_size;
2423
2424         BUG_ON(slot < 0);
2425         BUG_ON(slot >= nritems);
2426
2427         /*
2428          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2429          */
2430         /* first correct the data pointers */
2431         for (i = slot; i < nritems; i++) {
2432                 u32 ioff;
2433                 item = btrfs_item_nr(leaf, i);
2434
2435                 if (!leaf->map_token) {
2436                         map_extent_buffer(leaf, (unsigned long)item,
2437                                         sizeof(struct btrfs_item),
2438                                         &leaf->map_token, &leaf->kaddr,
2439                                         &leaf->map_start, &leaf->map_len,
2440                                         KM_USER1);
2441                 }
2442
2443                 ioff = btrfs_item_offset(leaf, item);
2444                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2445         }
2446
2447         if (leaf->map_token) {
2448                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2449                 leaf->map_token = NULL;
2450         }
2451
2452         /* shift the data */
2453         if (from_end) {
2454                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2455                               data_end + size_diff, btrfs_leaf_data(leaf) +
2456                               data_end, old_data_start + new_size - data_end);
2457         } else {
2458                 struct btrfs_disk_key disk_key;
2459                 u64 offset;
2460
2461                 btrfs_item_key(leaf, &disk_key, slot);
2462
2463                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2464                         unsigned long ptr;
2465                         struct btrfs_file_extent_item *fi;
2466
2467                         fi = btrfs_item_ptr(leaf, slot,
2468                                             struct btrfs_file_extent_item);
2469                         fi = (struct btrfs_file_extent_item *)(
2470                              (unsigned long)fi - size_diff);
2471
2472                         if (btrfs_file_extent_type(leaf, fi) ==
2473                             BTRFS_FILE_EXTENT_INLINE) {
2474                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2475                                 memmove_extent_buffer(leaf, ptr,
2476                                         (unsigned long)fi,
2477                                         offsetof(struct btrfs_file_extent_item,
2478                                                  disk_bytenr));
2479                         }
2480                 }
2481
2482                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2483                               data_end + size_diff, btrfs_leaf_data(leaf) +
2484                               data_end, old_data_start - data_end);
2485
2486                 offset = btrfs_disk_key_offset(&disk_key);
2487                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2488                 btrfs_set_item_key(leaf, &disk_key, slot);
2489                 if (slot == 0)
2490                         fixup_low_keys(trans, root, path, &disk_key, 1);
2491         }
2492
2493         item = btrfs_item_nr(leaf, slot);
2494         btrfs_set_item_size(leaf, item, new_size);
2495         btrfs_mark_buffer_dirty(leaf);
2496
2497         ret = 0;
2498         if (btrfs_leaf_free_space(root, leaf) < 0) {
2499                 btrfs_print_leaf(root, leaf);
2500                 BUG();
2501         }
2502         return ret;
2503 }
2504
2505 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2506                       struct btrfs_root *root, struct btrfs_path *path,
2507                       u32 data_size)
2508 {
2509         int ret = 0;
2510         int slot;
2511         int slot_orig;
2512         struct extent_buffer *leaf;
2513         struct btrfs_item *item;
2514         u32 nritems;
2515         unsigned int data_end;
2516         unsigned int old_data;
2517         unsigned int old_size;
2518         int i;
2519
2520         slot_orig = path->slots[0];
2521         leaf = path->nodes[0];
2522
2523         nritems = btrfs_header_nritems(leaf);
2524         data_end = leaf_data_end(root, leaf);
2525
2526         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2527                 btrfs_print_leaf(root, leaf);
2528                 BUG();
2529         }
2530         slot = path->slots[0];
2531         old_data = btrfs_item_end_nr(leaf, slot);
2532
2533         BUG_ON(slot < 0);
2534         if (slot >= nritems) {
2535                 btrfs_print_leaf(root, leaf);
2536                 printk("slot %d too large, nritems %d\n", slot, nritems);
2537                 BUG_ON(1);
2538         }
2539
2540         /*
2541          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2542          */
2543         /* first correct the data pointers */
2544         for (i = slot; i < nritems; i++) {
2545                 u32 ioff;
2546                 item = btrfs_item_nr(leaf, i);
2547
2548                 if (!leaf->map_token) {
2549                         map_extent_buffer(leaf, (unsigned long)item,
2550                                         sizeof(struct btrfs_item),
2551                                         &leaf->map_token, &leaf->kaddr,
2552                                         &leaf->map_start, &leaf->map_len,
2553                                         KM_USER1);
2554                 }
2555                 ioff = btrfs_item_offset(leaf, item);
2556                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2557         }
2558
2559         if (leaf->map_token) {
2560                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2561                 leaf->map_token = NULL;
2562         }
2563
2564         /* shift the data */
2565         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2566                       data_end - data_size, btrfs_leaf_data(leaf) +
2567                       data_end, old_data - data_end);
2568
2569         data_end = old_data;
2570         old_size = btrfs_item_size_nr(leaf, slot);
2571         item = btrfs_item_nr(leaf, slot);
2572         btrfs_set_item_size(leaf, item, old_size + data_size);
2573         btrfs_mark_buffer_dirty(leaf);
2574
2575         ret = 0;
2576         if (btrfs_leaf_free_space(root, leaf) < 0) {
2577                 btrfs_print_leaf(root, leaf);
2578                 BUG();
2579         }
2580         return ret;
2581 }
2582
2583 /*
2584  * Given a key and some data, insert an item into the tree.
2585  * This does all the path init required, making room in the tree if needed.
2586  */
2587 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2588                             struct btrfs_root *root,
2589                             struct btrfs_path *path,
2590                             struct btrfs_key *cpu_key, u32 *data_size,
2591                             int nr)
2592 {
2593         struct extent_buffer *leaf;
2594         struct btrfs_item *item;
2595         int ret = 0;
2596         int slot;
2597         int slot_orig;
2598         int i;
2599         u32 nritems;
2600         u32 total_size = 0;
2601         u32 total_data = 0;
2602         unsigned int data_end;
2603         struct btrfs_disk_key disk_key;
2604
2605         for (i = 0; i < nr; i++) {
2606                 total_data += data_size[i];
2607         }
2608
2609         total_size = total_data + (nr - 1) * sizeof(struct btrfs_item);
2610         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2611         if (ret == 0) {
2612                 return -EEXIST;
2613         }
2614         if (ret < 0)
2615                 goto out;
2616
2617         slot_orig = path->slots[0];
2618         leaf = path->nodes[0];
2619
2620         nritems = btrfs_header_nritems(leaf);
2621         data_end = leaf_data_end(root, leaf);
2622
2623         if (btrfs_leaf_free_space(root, leaf) <
2624             sizeof(struct btrfs_item) + total_size) {
2625                 btrfs_print_leaf(root, leaf);
2626                 printk("not enough freespace need %u have %d\n",
2627                        total_size, btrfs_leaf_free_space(root, leaf));
2628                 BUG();
2629         }
2630
2631         slot = path->slots[0];
2632         BUG_ON(slot < 0);
2633
2634         if (slot != nritems) {
2635                 int i;
2636                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2637
2638                 if (old_data < data_end) {
2639                         btrfs_print_leaf(root, leaf);
2640                         printk("slot %d old_data %d data_end %d\n",
2641                                slot, old_data, data_end);
2642                         BUG_ON(1);
2643                 }
2644                 /*
2645                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2646                  */
2647                 /* first correct the data pointers */
2648                 WARN_ON(leaf->map_token);
2649                 for (i = slot; i < nritems; i++) {
2650                         u32 ioff;
2651
2652                         item = btrfs_item_nr(leaf, i);
2653                         if (!leaf->map_token) {
2654                                 map_extent_buffer(leaf, (unsigned long)item,
2655                                         sizeof(struct btrfs_item),
2656                                         &leaf->map_token, &leaf->kaddr,
2657                                         &leaf->map_start, &leaf->map_len,
2658                                         KM_USER1);
2659                         }
2660
2661                         ioff = btrfs_item_offset(leaf, item);
2662                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2663                 }
2664                 if (leaf->map_token) {
2665                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2666                         leaf->map_token = NULL;
2667                 }
2668
2669                 /* shift the items */
2670                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2671                               btrfs_item_nr_offset(slot),
2672                               (nritems - slot) * sizeof(struct btrfs_item));
2673
2674                 /* shift the data */
2675                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2676                               data_end - total_data, btrfs_leaf_data(leaf) +
2677                               data_end, old_data - data_end);
2678                 data_end = old_data;
2679         }
2680
2681         /* setup the item for the new data */
2682         for (i = 0; i < nr; i++) {
2683                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2684                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2685                 item = btrfs_item_nr(leaf, slot + i);
2686                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2687                 data_end -= data_size[i];
2688                 btrfs_set_item_size(leaf, item, data_size[i]);
2689         }
2690         btrfs_set_header_nritems(leaf, nritems + nr);
2691         btrfs_mark_buffer_dirty(leaf);
2692
2693         ret = 0;
2694         if (slot == 0) {
2695                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2696                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2697         }
2698
2699         if (btrfs_leaf_free_space(root, leaf) < 0) {
2700                 btrfs_print_leaf(root, leaf);
2701                 BUG();
2702         }
2703 out:
2704         return ret;
2705 }
2706
2707 /*
2708  * Given a key and some data, insert an item into the tree.
2709  * This does all the path init required, making room in the tree if needed.
2710  */
2711 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2712                       *root, struct btrfs_key *cpu_key, void *data, u32
2713                       data_size)
2714 {
2715         int ret = 0;
2716         struct btrfs_path *path;
2717         struct extent_buffer *leaf;
2718         unsigned long ptr;
2719
2720         path = btrfs_alloc_path();
2721         BUG_ON(!path);
2722         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2723         if (!ret) {
2724                 leaf = path->nodes[0];
2725                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2726                 write_extent_buffer(leaf, data, ptr, data_size);
2727                 btrfs_mark_buffer_dirty(leaf);
2728         }
2729         btrfs_free_path(path);
2730         return ret;
2731 }
2732
2733 /*
2734  * delete the pointer from a given node.
2735  *
2736  * If the delete empties a node, the node is removed from the tree,
2737  * continuing all the way the root if required.  The root is converted into
2738  * a leaf if all the nodes are emptied.
2739  */
2740 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2741                    struct btrfs_path *path, int level, int slot)
2742 {
2743         struct extent_buffer *parent = path->nodes[level];
2744         u32 nritems;
2745         int ret = 0;
2746         int wret;
2747
2748         nritems = btrfs_header_nritems(parent);
2749         if (slot != nritems -1) {
2750                 memmove_extent_buffer(parent,
2751                               btrfs_node_key_ptr_offset(slot),
2752                               btrfs_node_key_ptr_offset(slot + 1),
2753                               sizeof(struct btrfs_key_ptr) *
2754                               (nritems - slot - 1));
2755         }
2756         nritems--;
2757         btrfs_set_header_nritems(parent, nritems);
2758         if (nritems == 0 && parent == root->node) {
2759                 BUG_ON(btrfs_header_level(root->node) != 1);
2760                 /* just turn the root into a leaf and break */
2761                 btrfs_set_header_level(root->node, 0);
2762         } else if (slot == 0) {
2763                 struct btrfs_disk_key disk_key;
2764
2765                 btrfs_node_key(parent, &disk_key, 0);
2766                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2767                 if (wret)
2768                         ret = wret;
2769         }
2770         btrfs_mark_buffer_dirty(parent);
2771         return ret;
2772 }
2773
2774 /*
2775  * delete the item at the leaf level in path.  If that empties
2776  * the leaf, remove it from the tree
2777  */
2778 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2779                     struct btrfs_path *path, int slot, int nr)
2780 {
2781         struct extent_buffer *leaf;
2782         struct btrfs_item *item;
2783         int last_off;
2784         int dsize = 0;
2785         int ret = 0;
2786         int wret;
2787         int i;
2788         u32 nritems;
2789
2790         leaf = path->nodes[0];
2791         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2792
2793         for (i = 0; i < nr; i++)
2794                 dsize += btrfs_item_size_nr(leaf, slot + i);
2795
2796         nritems = btrfs_header_nritems(leaf);
2797
2798         if (slot + nr != nritems) {
2799                 int i;
2800                 int data_end = leaf_data_end(root, leaf);
2801
2802                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2803                               data_end + dsize,
2804                               btrfs_leaf_data(leaf) + data_end,
2805                               last_off - data_end);
2806
2807                 for (i = slot + nr; i < nritems; i++) {
2808                         u32 ioff;
2809
2810                         item = btrfs_item_nr(leaf, i);
2811                         if (!leaf->map_token) {
2812                                 map_extent_buffer(leaf, (unsigned long)item,
2813                                         sizeof(struct btrfs_item),
2814                                         &leaf->map_token, &leaf->kaddr,
2815                                         &leaf->map_start, &leaf->map_len,
2816                                         KM_USER1);
2817                         }
2818                         ioff = btrfs_item_offset(leaf, item);
2819                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2820                 }
2821
2822                 if (leaf->map_token) {
2823                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2824                         leaf->map_token = NULL;
2825                 }
2826
2827                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2828                               btrfs_item_nr_offset(slot + nr),
2829                               sizeof(struct btrfs_item) *
2830                               (nritems - slot - nr));
2831         }
2832         btrfs_set_header_nritems(leaf, nritems - nr);
2833         nritems -= nr;
2834
2835         /* delete the leaf if we've emptied it */
2836         if (nritems == 0) {
2837                 if (leaf == root->node) {
2838                         btrfs_set_header_level(leaf, 0);
2839                 } else {
2840                         u64 root_gen = btrfs_header_generation(path->nodes[1]);
2841                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
2842                         if (wret)
2843                                 ret = wret;
2844                         wret = btrfs_free_extent(trans, root,
2845                                          leaf->start, leaf->len,
2846                                          btrfs_header_owner(path->nodes[1]),
2847                                          root_gen, 0, 0, 1);
2848                         if (wret)
2849                                 ret = wret;
2850                 }
2851         } else {
2852                 int used = leaf_space_used(leaf, 0, nritems);
2853                 if (slot == 0) {
2854                         struct btrfs_disk_key disk_key;
2855
2856                         btrfs_item_key(leaf, &disk_key, 0);
2857                         wret = fixup_low_keys(trans, root, path,
2858                                               &disk_key, 1);
2859                         if (wret)
2860                                 ret = wret;
2861                 }
2862
2863                 /* delete the leaf if it is mostly empty */
2864                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2865                         /* push_leaf_left fixes the path.
2866                          * make sure the path still points to our leaf
2867                          * for possible call to del_ptr below
2868                          */
2869                         slot = path->slots[1];
2870                         extent_buffer_get(leaf);
2871
2872                         wret = push_leaf_left(trans, root, path, 1, 1);
2873                         if (wret < 0 && wret != -ENOSPC)
2874                                 ret = wret;
2875
2876                         if (path->nodes[0] == leaf &&
2877                             btrfs_header_nritems(leaf)) {
2878                                 wret = push_leaf_right(trans, root, path, 1, 1);
2879                                 if (wret < 0 && wret != -ENOSPC)
2880                                         ret = wret;
2881                         }
2882
2883                         if (btrfs_header_nritems(leaf) == 0) {
2884                                 u64 root_gen;
2885                                 u64 bytenr = leaf->start;
2886                                 u32 blocksize = leaf->len;
2887
2888                                 root_gen = btrfs_header_generation(
2889                                                            path->nodes[1]);
2890
2891                                 wret = del_ptr(trans, root, path, 1, slot);
2892                                 if (wret)
2893                                         ret = wret;
2894
2895                                 free_extent_buffer(leaf);
2896                                 wret = btrfs_free_extent(trans, root, bytenr,
2897                                              blocksize,
2898                                              btrfs_header_owner(path->nodes[1]),
2899                                              root_gen, 0, 0, 1);
2900                                 if (wret)
2901                                         ret = wret;
2902                         } else {
2903                                 /* if we're still in the path, make sure
2904                                  * we're dirty.  Otherwise, one of the
2905                                  * push_leaf functions must have already
2906                                  * dirtied this buffer
2907                                  */
2908                                 if (path->nodes[0] == leaf)
2909                                         btrfs_mark_buffer_dirty(leaf);
2910                                 free_extent_buffer(leaf);
2911                         }
2912                 } else {
2913                         btrfs_mark_buffer_dirty(leaf);
2914                 }
2915         }
2916         return ret;
2917 }
2918
2919 /*
2920  * search the tree again to find a leaf with lesser keys
2921  * returns 0 if it found something or 1 if there are no lesser leaves.
2922  * returns < 0 on io errors.
2923  */
2924 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2925 {
2926         struct btrfs_key key;
2927         struct btrfs_disk_key found_key;
2928         int ret;
2929
2930         btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
2931
2932         if (key.offset > 0)
2933                 key.offset--;
2934         else if (key.type > 0)
2935                 key.type--;
2936         else if (key.objectid > 0)
2937                 key.objectid--;
2938         else
2939                 return 1;
2940
2941         btrfs_release_path(root, path);
2942         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2943         if (ret < 0)
2944                 return ret;
2945         btrfs_item_key(path->nodes[0], &found_key, 0);
2946         ret = comp_keys(&found_key, &key);
2947         if (ret < 0)
2948                 return 0;
2949         return 1;
2950 }
2951
2952 /*
2953  * search the tree again to find a leaf with greater keys
2954  * returns 0 if it found something or 1 if there are no greater leaves.
2955  * returns < 0 on io errors.
2956  */
2957 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2958 {
2959         int slot;
2960         int level = 1;
2961         struct extent_buffer *c;
2962         struct extent_buffer *next = NULL;
2963         struct btrfs_key key;
2964         u32 nritems;
2965         int ret;
2966
2967         nritems = btrfs_header_nritems(path->nodes[0]);
2968         if (nritems == 0) {
2969                 return 1;
2970         }
2971
2972         btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
2973
2974         btrfs_release_path(root, path);
2975         path->keep_locks = 1;
2976         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2977         path->keep_locks = 0;
2978
2979         if (ret < 0)
2980                 return ret;
2981
2982         nritems = btrfs_header_nritems(path->nodes[0]);
2983         if (nritems > 0 && path->slots[0] < nritems - 1) {
2984                 goto done;
2985         }
2986
2987         while(level < BTRFS_MAX_LEVEL) {
2988                 if (!path->nodes[level])
2989                         return 1;
2990
2991                 slot = path->slots[level] + 1;
2992                 c = path->nodes[level];
2993                 if (slot >= btrfs_header_nritems(c)) {
2994                         level++;
2995                         if (level == BTRFS_MAX_LEVEL) {
2996                                 return 1;
2997                         }
2998                         continue;
2999                 }
3000
3001                 if (next) {
3002                         btrfs_tree_unlock(next);
3003                         free_extent_buffer(next);
3004                 }
3005
3006                 if (level == 1 && path->locks[1] && path->reada)
3007                         reada_for_search(root, path, level, slot, 0);
3008
3009                 next = read_node_slot(root, c, slot);
3010                 if (!path->skip_locking) {
3011                         if (!btrfs_tree_locked(c)) {
3012                                 int i;
3013                                 WARN_ON(1);
3014 printk("path %p no lock on level %d\n", path, level);
3015 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
3016 printk("path %p level %d slot %d nritems %d\n", path, i, path->slots[i], btrfs_header_nritems(path->nodes[i]));
3017 }
3018                         }
3019                         btrfs_tree_lock(next);
3020                 }
3021                 break;
3022         }
3023         path->slots[level] = slot;
3024         while(1) {
3025                 level--;
3026                 c = path->nodes[level];
3027                 if (path->locks[level])
3028                         btrfs_tree_unlock(c);
3029                 free_extent_buffer(c);
3030                 path->nodes[level] = next;
3031                 path->slots[level] = 0;
3032                 path->locks[level] = 1;
3033                 if (!level)
3034                         break;
3035                 if (level == 1 && path->locks[1] && path->reada)
3036                         reada_for_search(root, path, level, slot, 0);
3037                 next = read_node_slot(root, next, 0);
3038                 if (!path->skip_locking) {
3039                         WARN_ON(!btrfs_tree_locked(path->nodes[level]));
3040                         btrfs_tree_lock(next);
3041                 }
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