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
  67         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  68                 p->slots[i] = 0;
  69                 if (!p->nodes[i])
  70                         continue;
  71                 if (p->locks[i]) {
  72                         btrfs_tree_unlock(p->nodes[i]);
  73                         p->locks[i] = 0;
  74                 }
  75                 free_extent_buffer(p->nodes[i]);
  76                 p->nodes[i] = NULL;
  77         }
  78 }
  79
  80 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
  81 {
  82         struct extent_buffer *eb;
  83         spin_lock(&root->node_lock);
  84         eb = root->node;
  85         extent_buffer_get(eb);
  86         spin_unlock(&root->node_lock);
  87         return eb;
  88 }
  89
  90 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
  91 {
  92         struct extent_buffer *eb;
  93
  94         while(1) {
  95                 eb = btrfs_root_node(root);
  96                 btrfs_tree_lock(eb);
  97
  98                 spin_lock(&root->node_lock);
  99                 if (eb == root->node) {
 100                         spin_unlock(&root->node_lock);
 101                         break;
 102                 }
 103                 spin_unlock(&root->node_lock);
 104
 105                 btrfs_tree_unlock(eb);
 106                 free_extent_buffer(eb);
 107         }
 108         return eb;
 109 }
 110
 111 static void add_root_to_dirty_list(struct btrfs_root *root)
 112 {
 113         if (root->track_dirty && list_empty(&root->dirty_list)) {
 114                 list_add(&root->dirty_list,
 115                          &root->fs_info->dirty_cowonly_roots);
 116         }
 117 }
 118
 119 int btrfs_copy_root(struct btrfs_trans_handle *trans,
 120                       struct btrfs_root *root,
 121                       struct extent_buffer *buf,
 122                       struct extent_buffer **cow_ret, u64 new_root_objectid)
 123 {
 124         struct extent_buffer *cow;
 125         u32 nritems;
 126         int ret = 0;
 127         int level;
 128         struct btrfs_key first_key;
 129         struct btrfs_root *new_root;
 130
 131         new_root = kmalloc(sizeof(*new_root), GFP_NOFS);
 132         if (!new_root)
 133                 return -ENOMEM;
 134
 135         memcpy(new_root, root, sizeof(*new_root));
 136         new_root->root_key.objectid = new_root_objectid;
 137
 138         WARN_ON(root->ref_cows && trans->transid !=
 139                 root->fs_info->running_transaction->transid);
 140         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 141
 142         level = btrfs_header_level(buf);
 143         nritems = btrfs_header_nritems(buf);
 144         if (nritems) {
 145                 if (level == 0)
 146                         btrfs_item_key_to_cpu(buf, &first_key, 0);
 147                 else
 148                         btrfs_node_key_to_cpu(buf, &first_key, 0);
 149         } else {
 150                 first_key.objectid = 0;
 151         }
 152         cow = btrfs_alloc_free_block(trans, new_root, buf->len,
 153                                        new_root_objectid,
 154                                        trans->transid, first_key.objectid,
 155                                        level, buf->start, 0);
 156         if (IS_ERR(cow)) {
 157                 kfree(new_root);
 158                 return PTR_ERR(cow);
 159         }
 160
 161         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 162         btrfs_set_header_bytenr(cow, cow->start);
 163         btrfs_set_header_generation(cow, trans->transid);
 164         btrfs_set_header_owner(cow, new_root_objectid);
 165         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN);
 166
 167         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 168         ret = btrfs_inc_ref(trans, new_root, buf);
 169         kfree(new_root);
 170
 171         if (ret)
 172                 return ret;
 173
 174         btrfs_mark_buffer_dirty(cow);
 175         *cow_ret = cow;
 176         return 0;
 177 }
 178
 179 int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 180                              struct btrfs_root *root,
 181                              struct extent_buffer *buf,
 182                              struct extent_buffer *parent, int parent_slot,
 183                              struct extent_buffer **cow_ret,
 184                              u64 search_start, u64 empty_size)
 185 {
 186         u64 root_gen;
 187         struct extent_buffer *cow;
 188         u32 nritems;
 189         int ret = 0;
 190         int different_trans = 0;
 191         int level;
 192         int unlock_orig = 0;
 193         struct btrfs_key first_key;
 194
 195         if (*cow_ret == buf)
 196                 unlock_orig = 1;
 197
 198         WARN_ON(!btrfs_tree_locked(buf));
 199
 200         if (root->ref_cows) {
 201                 root_gen = trans->transid;
 202         } else {
 203                 root_gen = 0;
 204         }
 205         WARN_ON(root->ref_cows && trans->transid !=
 206                 root->fs_info->running_transaction->transid);
 207         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 208
 209         level = btrfs_header_level(buf);
 210         nritems = btrfs_header_nritems(buf);
 211         if (nritems) {
 212                 if (level == 0)
 213                         btrfs_item_key_to_cpu(buf, &first_key, 0);
 214                 else
 215                         btrfs_node_key_to_cpu(buf, &first_key, 0);
 216         } else {
 217                 first_key.objectid = 0;
 218         }
 219         cow = btrfs_alloc_free_block(trans, root, buf->len,
 220                                      root->root_key.objectid,
 221                                      root_gen, first_key.objectid, level,
 222                                      search_start, empty_size);
 223         if (IS_ERR(cow))
 224                 return PTR_ERR(cow);
 225
 226         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 227         btrfs_set_header_bytenr(cow, cow->start);
 228         btrfs_set_header_generation(cow, trans->transid);
 229         btrfs_set_header_owner(cow, root->root_key.objectid);
 230         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN);
 231
 232         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 233         if (btrfs_header_generation(buf) != trans->transid) {
 234                 different_trans = 1;
 235                 ret = btrfs_inc_ref(trans, root, buf);
 236                 if (ret)
 237                         return ret;
 238         } else {
 239                 clean_tree_block(trans, root, buf);
 240         }
 241
 242         if (buf == root->node) {
 243                 WARN_ON(parent && parent != buf);
 244                 root_gen = btrfs_header_generation(buf);
 245
 246                 spin_lock(&root->node_lock);
 247                 root->node = cow;
 248                 extent_buffer_get(cow);
 249                 spin_unlock(&root->node_lock);
 250
 251                 if (buf != root->commit_root) {
 252                         btrfs_free_extent(trans, root, buf->start,
 253                                           buf->len, root->root_key.objectid,
 254                                           root_gen, 0, 0, 1);
 255                 }
 256                 free_extent_buffer(buf);
 257                 add_root_to_dirty_list(root);
 258         } else {
 259                 root_gen = btrfs_header_generation(parent);
 260                 btrfs_set_node_blockptr(parent, parent_slot,
 261                                         cow->start);
 262                 WARN_ON(trans->transid == 0);
 263                 btrfs_set_node_ptr_generation(parent, parent_slot,
 264                                               trans->transid);
 265                 btrfs_mark_buffer_dirty(parent);
 266                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 267                 btrfs_free_extent(trans, root, buf->start, buf->len,
 268                                   btrfs_header_owner(parent), root_gen,
 269                                   0, 0, 1);
 270         }
 271         if (unlock_orig)
 272                 btrfs_tree_unlock(buf);
 273         free_extent_buffer(buf);
 274         btrfs_mark_buffer_dirty(cow);
 275         *cow_ret = cow;
 276         return 0;
 277 }
 278
 279 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 280                     struct btrfs_root *root, struct extent_buffer *buf,
 281                     struct extent_buffer *parent, int parent_slot,
 282                     struct extent_buffer **cow_ret)
 283 {
 284         u64 search_start;
 285         u64 header_trans;
 286         int ret;
 287
 288         if (trans->transaction != root->fs_info->running_transaction) {
 289                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 290                        root->fs_info->running_transaction->transid);
 291                 WARN_ON(1);
 292         }
 293         if (trans->transid != root->fs_info->generation) {
 294                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 295                        root->fs_info->generation);
 296                 WARN_ON(1);
 297         }
 298
 299         header_trans = btrfs_header_generation(buf);
 300         spin_lock(&root->fs_info->hash_lock);
 301         if (header_trans == trans->transid &&
 302             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
 303                 *cow_ret = buf;
 304                 spin_unlock(&root->fs_info->hash_lock);
 305                 return 0;
 306         }
 307         spin_unlock(&root->fs_info->hash_lock);
 308         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
 309         ret = __btrfs_cow_block(trans, root, buf, parent,
 310                                  parent_slot, cow_ret, search_start, 0);
 311         return ret;
 312 }
 313
 314 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 315 {
 316         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 317                 return 1;
 318         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 319                 return 1;
 320         return 0;
 321 }
 322
 323 /*
 324  * compare two keys in a memcmp fashion
 325  */
 326 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 327 {
 328         struct btrfs_key k1;
 329
 330         btrfs_disk_key_to_cpu(&k1, disk);
 331
 332         if (k1.objectid > k2->objectid)
 333                 return 1;
 334         if (k1.objectid < k2->objectid)
 335                 return -1;
 336         if (k1.type > k2->type)
 337                 return 1;
 338         if (k1.type < k2->type)
 339                 return -1;
 340         if (k1.offset > k2->offset)
 341                 return 1;
 342         if (k1.offset < k2->offset)
 343                 return -1;
 344         return 0;
 345 }
 346
 347
 348 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 349                        struct btrfs_root *root, struct extent_buffer *parent,
 350                        int start_slot, int cache_only, u64 *last_ret,
 351                        struct btrfs_key *progress)
 352 {
 353         struct extent_buffer *cur;
 354         u64 blocknr;
 355         u64 gen;
 356         u64 search_start = *last_ret;
 357         u64 last_block = 0;
 358         u64 other;
 359         u32 parent_nritems;
 360         int end_slot;
 361         int i;
 362         int err = 0;
 363         int parent_level;
 364         int uptodate;
 365         u32 blocksize;
 366         int progress_passed = 0;
 367         struct btrfs_disk_key disk_key;
 368
 369         parent_level = btrfs_header_level(parent);
 370         if (cache_only && parent_level != 1)
 371                 return 0;
 372
 373         if (trans->transaction != root->fs_info->running_transaction) {
 374                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 375                        root->fs_info->running_transaction->transid);
 376                 WARN_ON(1);
 377         }
 378         if (trans->transid != root->fs_info->generation) {
 379                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 380                        root->fs_info->generation);
 381                 WARN_ON(1);
 382         }
 383
 384         parent_nritems = btrfs_header_nritems(parent);
 385         blocksize = btrfs_level_size(root, parent_level - 1);
 386         end_slot = parent_nritems;
 387
 388         if (parent_nritems == 1)
 389                 return 0;
 390
 391         for (i = start_slot; i < end_slot; i++) {
 392                 int close = 1;
 393
 394                 if (!parent->map_token) {
 395                         map_extent_buffer(parent,
 396                                         btrfs_node_key_ptr_offset(i),
 397                                         sizeof(struct btrfs_key_ptr),
 398                                         &parent->map_token, &parent->kaddr,
 399                                         &parent->map_start, &parent->map_len,
 400                                         KM_USER1);
 401                 }
 402                 btrfs_node_key(parent, &disk_key, i);
 403                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
 404                         continue;
 405
 406                 progress_passed = 1;
 407                 blocknr = btrfs_node_blockptr(parent, i);
 408                 gen = btrfs_node_ptr_generation(parent, i);
 409                 if (last_block == 0)
 410                         last_block = blocknr;
 411
 412                 if (i > 0) {
 413                         other = btrfs_node_blockptr(parent, i - 1);
 414                         close = close_blocks(blocknr, other, blocksize);
 415                 }
 416                 if (!close && i < end_slot - 2) {
 417                         other = btrfs_node_blockptr(parent, i + 1);
 418                         close = close_blocks(blocknr, other, blocksize);
 419                 }
 420                 if (close) {
 421                         last_block = blocknr;
 422                         continue;
 423                 }
 424                 if (parent->map_token) {
 425                         unmap_extent_buffer(parent, parent->map_token,
 426                                             KM_USER1);
 427                         parent->map_token = NULL;
 428                 }
 429
 430                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
 431                 if (cur)
 432                         uptodate = btrfs_buffer_uptodate(cur, gen);
 433                 else
 434                         uptodate = 0;
 435                 if (!cur || !uptodate) {
 436                         if (cache_only) {
 437                                 free_extent_buffer(cur);
 438                                 continue;
 439                         }
 440                         if (!cur) {
 441                                 cur = read_tree_block(root, blocknr,
 442                                                          blocksize, gen);
 443                         } else if (!uptodate) {
 444                                 btrfs_read_buffer(cur, gen);
 445                         }
 446                 }
 447                 if (search_start == 0)
 448                         search_start = last_block;
 449
 450                 btrfs_tree_lock(cur);
 451                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 452                                         &cur, search_start,
 453                                         min(16 * blocksize,
 454                                             (end_slot - i) * blocksize));
 455                 if (err) {
 456                         btrfs_tree_unlock(cur);
 457                         free_extent_buffer(cur);
 458                         break;
 459                 }
 460                 search_start = cur->start;
 461                 last_block = cur->start;
 462                 *last_ret = search_start;
 463                 btrfs_tree_unlock(cur);
 464                 free_extent_buffer(cur);
 465         }
 466         if (parent->map_token) {
 467                 unmap_extent_buffer(parent, parent->map_token,
 468                                     KM_USER1);
 469                 parent->map_token = NULL;
 470         }
 471         return err;
 472 }
 473
 474 /*
 475  * The leaf data grows from end-to-front in the node.
 476  * this returns the address of the start of the last item,
 477  * which is the stop of the leaf data stack
 478  */
 479 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 480                                          struct extent_buffer *leaf)
 481 {
 482         u32 nr = btrfs_header_nritems(leaf);
 483         if (nr == 0)
 484                 return BTRFS_LEAF_DATA_SIZE(root);
 485         return btrfs_item_offset_nr(leaf, nr - 1);
 486 }
 487
 488 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
 489                       int level)
 490 {
 491         struct extent_buffer *parent = NULL;
 492         struct extent_buffer *node = path->nodes[level];
 493         struct btrfs_disk_key parent_key;
 494         struct btrfs_disk_key node_key;
 495         int parent_slot;
 496         int slot;
 497         struct btrfs_key cpukey;
 498         u32 nritems = btrfs_header_nritems(node);
 499
 500         if (path->nodes[level + 1])
 501                 parent = path->nodes[level + 1];
 502
 503         slot = path->slots[level];
 504         BUG_ON(nritems == 0);
 505         if (parent) {
 506                 parent_slot = path->slots[level + 1];
 507                 btrfs_node_key(parent, &parent_key, parent_slot);
 508                 btrfs_node_key(node, &node_key, 0);
 509                 BUG_ON(memcmp(&parent_key, &node_key,
 510                               sizeof(struct btrfs_disk_key)));
 511                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 512                        btrfs_header_bytenr(node));
 513         }
 514         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 515         if (slot != 0) {
 516                 btrfs_node_key_to_cpu(node, &cpukey, slot - 1);
 517                 btrfs_node_key(node, &node_key, slot);
 518                 BUG_ON(comp_keys(&node_key, &cpukey) <= 0);
 519         }
 520         if (slot < nritems - 1) {
 521                 btrfs_node_key_to_cpu(node, &cpukey, slot + 1);
 522                 btrfs_node_key(node, &node_key, slot);
 523                 BUG_ON(comp_keys(&node_key, &cpukey) >= 0);
 524         }
 525         return 0;
 526 }
 527
 528 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 529                       int level)
 530 {
 531         struct extent_buffer *leaf = path->nodes[level];
 532         struct extent_buffer *parent = NULL;
 533         int parent_slot;
 534         struct btrfs_key cpukey;
 535         struct btrfs_disk_key parent_key;
 536         struct btrfs_disk_key leaf_key;
 537         int slot = path->slots[0];
 538
 539         u32 nritems = btrfs_header_nritems(leaf);
 540
 541         if (path->nodes[level + 1])
 542                 parent = path->nodes[level + 1];
 543
 544         if (nritems == 0)
 545                 return 0;
 546
 547         if (parent) {
 548                 parent_slot = path->slots[level + 1];
 549                 btrfs_node_key(parent, &parent_key, parent_slot);
 550                 btrfs_item_key(leaf, &leaf_key, 0);
 551
 552                 BUG_ON(memcmp(&parent_key, &leaf_key,
 553                        sizeof(struct btrfs_disk_key)));
 554                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 555                        btrfs_header_bytenr(leaf));
 556         }
 557 #if 0
 558         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 559                 btrfs_item_key_to_cpu(leaf, &cpukey, i + 1);
 560                 btrfs_item_key(leaf, &leaf_key, i);
 561                 if (comp_keys(&leaf_key, &cpukey) >= 0) {
 562                         btrfs_print_leaf(root, leaf);
 563                         printk("slot %d offset bad key\n", i);
 564                         BUG_ON(1);
 565                 }
 566                 if (btrfs_item_offset_nr(leaf, i) !=
 567                         btrfs_item_end_nr(leaf, i + 1)) {
 568                         btrfs_print_leaf(root, leaf);
 569                         printk("slot %d offset bad\n", i);
 570                         BUG_ON(1);
 571                 }
 572                 if (i == 0) {
 573                         if (btrfs_item_offset_nr(leaf, i) +
 574                                btrfs_item_size_nr(leaf, i) !=
 575                                BTRFS_LEAF_DATA_SIZE(root)) {
 576                                 btrfs_print_leaf(root, leaf);
 577                                 printk("slot %d first offset bad\n", i);
 578                                 BUG_ON(1);
 579                         }
 580                 }
 581         }
 582         if (nritems > 0) {
 583                 if (btrfs_item_size_nr(leaf, nritems - 1) > 4096) {
 584                                 btrfs_print_leaf(root, leaf);
 585                                 printk("slot %d bad size \n", nritems - 1);
 586                                 BUG_ON(1);
 587                 }
 588         }
 589 #endif
 590         if (slot != 0 && slot < nritems - 1) {
 591                 btrfs_item_key(leaf, &leaf_key, slot);
 592                 btrfs_item_key_to_cpu(leaf, &cpukey, slot - 1);
 593                 if (comp_keys(&leaf_key, &cpukey) <= 0) {
 594                         btrfs_print_leaf(root, leaf);
 595                         printk("slot %d offset bad key\n", slot);
 596                         BUG_ON(1);
 597                 }
 598                 if (btrfs_item_offset_nr(leaf, slot - 1) !=
 599                        btrfs_item_end_nr(leaf, slot)) {
 600                         btrfs_print_leaf(root, leaf);
 601                         printk("slot %d offset bad\n", slot);
 602                         BUG_ON(1);
 603                 }
 604         }
 605         if (slot < nritems - 1) {
 606                 btrfs_item_key(leaf, &leaf_key, slot);
 607                 btrfs_item_key_to_cpu(leaf, &cpukey, slot + 1);
 608                 BUG_ON(comp_keys(&leaf_key, &cpukey) >= 0);
 609                 if (btrfs_item_offset_nr(leaf, slot) !=
 610                         btrfs_item_end_nr(leaf, slot + 1)) {
 611                         btrfs_print_leaf(root, leaf);
 612                         printk("slot %d offset bad\n", slot);
 613                         BUG_ON(1);
 614                 }
 615         }
 616         BUG_ON(btrfs_item_offset_nr(leaf, 0) +
 617                btrfs_item_size_nr(leaf, 0) != BTRFS_LEAF_DATA_SIZE(root));
 618         return 0;
 619 }
 620
 621 static int noinline check_block(struct btrfs_root *root,
 622                                 struct btrfs_path *path, int level)
 623 {
 624         u64 found_start;
 625         return 0;
 626         if (btrfs_header_level(path->nodes[level]) != level)
 627             printk("warning: bad level %Lu wanted %d found %d\n",
 628                    path->nodes[level]->start, level,
 629                    btrfs_header_level(path->nodes[level]));
 630         found_start = btrfs_header_bytenr(path->nodes[level]);
 631         if (found_start != path->nodes[level]->start) {
 632             printk("warning: bad bytentr %Lu found %Lu\n",
 633                    path->nodes[level]->start, found_start);
 634         }
 635 #if 0
 636         struct extent_buffer *buf = path->nodes[level];
 637
 638         if (memcmp_extent_buffer(buf, root->fs_info->fsid,
 639                                  (unsigned long)btrfs_header_fsid(buf),
 640                                  BTRFS_FSID_SIZE)) {
 641                 printk("warning bad block %Lu\n", buf->start);
 642                 return 1;
 643         }
 644 #endif
 645         if (level == 0)
 646                 return check_leaf(root, path, level);
 647         return check_node(root, path, level);
 648 }
 649
 650 /*
 651  * search for key in the extent_buffer.  The items start at offset p,
 652  * and they are item_size apart.  There are 'max' items in p.
 653  *
 654  * the slot in the array is returned via slot, and it points to
 655  * the place where you would insert key if it is not found in
 656  * the array.
 657  *
 658  * slot may point to max if the key is bigger than all of the keys
 659  */
 660 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 661                               int item_size, struct btrfs_key *key,
 662                               int max, int *slot)
 663 {
 664         int low = 0;
 665         int high = max;
 666         int mid;
 667         int ret;
 668         struct btrfs_disk_key *tmp = NULL;
 669         struct btrfs_disk_key unaligned;
 670         unsigned long offset;
 671         char *map_token = NULL;
 672         char *kaddr = NULL;
 673         unsigned long map_start = 0;
 674         unsigned long map_len = 0;
 675         int err;
 676
 677         while(low < high) {
 678                 mid = (low + high) / 2;
 679                 offset = p + mid * item_size;
 680
 681                 if (!map_token || offset < map_start ||
 682                     (offset + sizeof(struct btrfs_disk_key)) >
 683                     map_start + map_len) {
 684                         if (map_token) {
 685                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 686                                 map_token = NULL;
 687                         }
 688                         err = map_extent_buffer(eb, offset,
 689                                                 sizeof(struct btrfs_disk_key),
 690                                                 &map_token, &kaddr,
 691                                                 &map_start, &map_len, KM_USER0);
 692
 693                         if (!err) {
 694                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
 695                                                         map_start);
 696                         } else {
 697                                 read_extent_buffer(eb, &unaligned,
 698                                                    offset, sizeof(unaligned));
 699                                 tmp = &unaligned;
 700                         }
 701
 702                 } else {
 703                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
 704                                                         map_start);
 705                 }
 706                 ret = comp_keys(tmp, key);
 707
 708                 if (ret < 0)
 709                         low = mid + 1;
 710                 else if (ret > 0)
 711                         high = mid;
 712                 else {
 713                         *slot = mid;
 714                         if (map_token)
 715                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 716                         return 0;
 717                 }
 718         }
 719         *slot = low;
 720         if (map_token)
 721                 unmap_extent_buffer(eb, map_token, KM_USER0);
 722         return 1;
 723 }
 724
 725 /*
 726  * simple bin_search frontend that does the right thing for
 727  * leaves vs nodes
 728  */
 729 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 730                       int level, int *slot)
 731 {
 732         if (level == 0) {
 733                 return generic_bin_search(eb,
 734                                           offsetof(struct btrfs_leaf, items),
 735                                           sizeof(struct btrfs_item),
 736                                           key, btrfs_header_nritems(eb),
 737                                           slot);
 738         } else {
 739                 return generic_bin_search(eb,
 740                                           offsetof(struct btrfs_node, ptrs),
 741                                           sizeof(struct btrfs_key_ptr),
 742                                           key, btrfs_header_nritems(eb),
 743                                           slot);
 744         }
 745         return -1;
 746 }
 747
 748 static struct extent_buffer *read_node_slot(struct btrfs_root *root,
 749                                    struct extent_buffer *parent, int slot)
 750 {
 751         int level = btrfs_header_level(parent);
 752         if (slot < 0)
 753                 return NULL;
 754         if (slot >= btrfs_header_nritems(parent))
 755                 return NULL;
 756
 757         BUG_ON(level == 0);
 758
 759         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 760                        btrfs_level_size(root, level - 1),
 761                        btrfs_node_ptr_generation(parent, slot));
 762 }
 763
 764 static int balance_level(struct btrfs_trans_handle *trans,
 765                          struct btrfs_root *root,
 766                          struct btrfs_path *path, int level)
 767 {
 768         struct extent_buffer *right = NULL;
 769         struct extent_buffer *mid;
 770         struct extent_buffer *left = NULL;
 771         struct extent_buffer *parent = NULL;
 772         int ret = 0;
 773         int wret;
 774         int pslot;
 775         int orig_slot = path->slots[level];
 776         int err_on_enospc = 0;
 777         u64 orig_ptr;
 778
 779         if (level == 0)
 780                 return 0;
 781
 782         mid = path->nodes[level];
 783         WARN_ON(!path->locks[level]);
 784         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 785
 786         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 787
 788         if (level < BTRFS_MAX_LEVEL - 1)
 789                 parent = path->nodes[level + 1];
 790         pslot = path->slots[level + 1];
 791
 792         /*
 793          * deal with the case where there is only one pointer in the root
 794          * by promoting the node below to a root
 795          */
 796         if (!parent) {
 797                 struct extent_buffer *child;
 798
 799                 if (btrfs_header_nritems(mid) != 1)
 800                         return 0;
 801
 802                 /* promote the child to a root */
 803                 child = read_node_slot(root, mid, 0);
 804                 btrfs_tree_lock(child);
 805                 BUG_ON(!child);
 806                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
 807                 BUG_ON(ret);
 808
 809                 spin_lock(&root->node_lock);
 810                 root->node = child;
 811                 spin_unlock(&root->node_lock);
 812
 813                 add_root_to_dirty_list(root);
 814                 btrfs_tree_unlock(child);
 815                 path->locks[level] = 0;
 816                 path->nodes[level] = NULL;
 817                 clean_tree_block(trans, root, mid);
 818                 btrfs_tree_unlock(mid);
 819                 /* once for the path */
 820                 free_extent_buffer(mid);
 821                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 822                                         root->root_key.objectid,
 823                                         btrfs_header_generation(mid), 0, 0, 1);
 824                 /* once for the root ptr */
 825                 free_extent_buffer(mid);
 826                 return ret;
 827         }
 828         if (btrfs_header_nritems(mid) >
 829             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 830                 return 0;
 831
 832         if (btrfs_header_nritems(mid) < 2)
 833                 err_on_enospc = 1;
 834
 835         left = read_node_slot(root, parent, pslot - 1);
 836         if (left) {
 837                 btrfs_tree_lock(left);
 838                 wret = btrfs_cow_block(trans, root, left,
 839                                        parent, pslot - 1, &left);
 840                 if (wret) {
 841                         ret = wret;
 842                         goto enospc;
 843                 }
 844         }
 845         right = read_node_slot(root, parent, pslot + 1);
 846         if (right) {
 847                 btrfs_tree_lock(right);
 848                 wret = btrfs_cow_block(trans, root, right,
 849                                        parent, pslot + 1, &right);
 850                 if (wret) {
 851                         ret = wret;
 852                         goto enospc;
 853                 }
 854         }
 855
 856         /* first, try to make some room in the middle buffer */
 857         if (left) {
 858                 orig_slot += btrfs_header_nritems(left);
 859                 wret = push_node_left(trans, root, left, mid, 1);
 860                 if (wret < 0)
 861                         ret = wret;
 862                 if (btrfs_header_nritems(mid) < 2)
 863                         err_on_enospc = 1;
 864         }
 865
 866         /*
 867          * then try to empty the right most buffer into the middle
 868          */
 869         if (right) {
 870                 wret = push_node_left(trans, root, mid, right, 1);
 871                 if (wret < 0 && wret != -ENOSPC)
 872                         ret = wret;
 873                 if (btrfs_header_nritems(right) == 0) {
 874                         u64 bytenr = right->start;
 875                         u64 generation = btrfs_header_generation(parent);
 876                         u32 blocksize = right->len;
 877
 878                         clean_tree_block(trans, root, right);
 879                         btrfs_tree_unlock(right);
 880                         free_extent_buffer(right);
 881                         right = NULL;
 882                         wret = del_ptr(trans, root, path, level + 1, pslot +
 883                                        1);
 884                         if (wret)
 885                                 ret = wret;
 886                         wret = btrfs_free_extent(trans, root, bytenr,
 887                                                  blocksize,
 888                                                  btrfs_header_owner(parent),
 889                                                  generation, 0, 0, 1);
 890                         if (wret)
 891                                 ret = wret;
 892                 } else {
 893                         struct btrfs_disk_key right_key;
 894                         btrfs_node_key(right, &right_key, 0);
 895                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 896                         btrfs_mark_buffer_dirty(parent);
 897                 }
 898         }
 899         if (btrfs_header_nritems(mid) == 1) {
 900                 /*
 901                  * we're not allowed to leave a node with one item in the
 902                  * tree during a delete.  A deletion from lower in the tree
 903                  * could try to delete the only pointer in this node.
 904                  * So, pull some keys from the left.
 905                  * There has to be a left pointer at this point because
 906                  * otherwise we would have pulled some pointers from the
 907                  * right
 908                  */
 909                 BUG_ON(!left);
 910                 wret = balance_node_right(trans, root, mid, left);
 911                 if (wret < 0) {
 912                         ret = wret;
 913                         goto enospc;
 914                 }
 915                 if (wret == 1) {
 916                         wret = push_node_left(trans, root, left, mid, 1);
 917                         if (wret < 0)
 918                                 ret = wret;
 919                 }
 920                 BUG_ON(wret == 1);
 921         }
 922         if (btrfs_header_nritems(mid) == 0) {
 923                 /* we've managed to empty the middle node, drop it */
 924                 u64 root_gen = btrfs_header_generation(parent);
 925                 u64 bytenr = mid->start;
 926                 u32 blocksize = mid->len;
 927
 928                 clean_tree_block(trans, root, mid);
 929                 btrfs_tree_unlock(mid);
 930                 free_extent_buffer(mid);
 931                 mid = NULL;
 932                 wret = del_ptr(trans, root, path, level + 1, pslot);
 933                 if (wret)
 934                         ret = wret;
 935                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 936                                          btrfs_header_owner(parent),
 937                                          root_gen, 0, 0, 1);
 938                 if (wret)
 939                         ret = wret;
 940         } else {
 941                 /* update the parent key to reflect our changes */
 942                 struct btrfs_disk_key mid_key;
 943                 btrfs_node_key(mid, &mid_key, 0);
 944                 btrfs_set_node_key(parent, &mid_key, pslot);
 945                 btrfs_mark_buffer_dirty(parent);
 946         }
 947
 948         /* update the path */
 949         if (left) {
 950                 if (btrfs_header_nritems(left) > orig_slot) {
 951                         extent_buffer_get(left);
 952                         /* left was locked after cow */
 953                         path->nodes[level] = left;
 954                         path->slots[level + 1] -= 1;
 955                         path->slots[level] = orig_slot;
 956                         if (mid) {
 957                                 btrfs_tree_unlock(mid);
 958                                 free_extent_buffer(mid);
 959                         }
 960                 } else {
 961                         orig_slot -= btrfs_header_nritems(left);
 962                         path->slots[level] = orig_slot;
 963                 }
 964         }
 965         /* double check we haven't messed things up */
 966         check_block(root, path, level);
 967         if (orig_ptr !=
 968             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 969                 BUG();
 970 enospc:
 971         if (right) {
 972                 btrfs_tree_unlock(right);
 973                 free_extent_buffer(right);
 974         }
 975         if (left) {
 976                 if (path->nodes[level] != left)
 977                         btrfs_tree_unlock(left);
 978                 free_extent_buffer(left);
 979         }
 980         return ret;
 981 }
 982
 983 /* returns zero if the push worked, non-zero otherwise */
 984 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
 985                                           struct btrfs_root *root,
 986                                           struct btrfs_path *path, int level)
 987 {
 988         struct extent_buffer *right = NULL;
 989         struct extent_buffer *mid;
 990         struct extent_buffer *left = NULL;
 991         struct extent_buffer *parent = NULL;
 992         int ret = 0;
 993         int wret;
 994         int pslot;
 995         int orig_slot = path->slots[level];
 996         u64 orig_ptr;
 997
 998         if (level == 0)
 999                 return 1;
1000
1001         mid = path->nodes[level];
1002         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1003         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1004
1005         if (level < BTRFS_MAX_LEVEL - 1)
1006                 parent = path->nodes[level + 1];
1007         pslot = path->slots[level + 1];
1008
1009         if (!parent)
1010                 return 1;
1011
1012         left = read_node_slot(root, parent, pslot - 1);
1013
1014         /* first, try to make some room in the middle buffer */
1015         if (left) {
1016                 u32 left_nr;
1017
1018                 btrfs_tree_lock(left);
1019                 left_nr = btrfs_header_nritems(left);
1020                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1021                         wret = 1;
1022                 } else {
1023                         ret = btrfs_cow_block(trans, root, left, parent,
1024                                               pslot - 1, &left);
1025                         if (ret)
1026                                 wret = 1;
1027                         else {
1028                                 wret = push_node_left(trans, root,
1029                                                       left, mid, 0);
1030                         }
1031                 }
1032                 if (wret < 0)
1033                         ret = wret;
1034                 if (wret == 0) {
1035                         struct btrfs_disk_key disk_key;
1036                         orig_slot += left_nr;
1037                         btrfs_node_key(mid, &disk_key, 0);
1038                         btrfs_set_node_key(parent, &disk_key, pslot);
1039                         btrfs_mark_buffer_dirty(parent);
1040                         if (btrfs_header_nritems(left) > orig_slot) {
1041                                 path->nodes[level] = left;
1042                                 path->slots[level + 1] -= 1;
1043                                 path->slots[level] = orig_slot;
1044                                 btrfs_tree_unlock(mid);
1045                                 free_extent_buffer(mid);
1046                         } else {
1047                                 orig_slot -=
1048                                         btrfs_header_nritems(left);
1049                                 path->slots[level] = orig_slot;
1050                                 btrfs_tree_unlock(left);
1051                                 free_extent_buffer(left);
1052                         }
1053                         return 0;
1054                 }
1055                 btrfs_tree_unlock(left);
1056                 free_extent_buffer(left);
1057         }
1058         right = read_node_slot(root, parent, pslot + 1);
1059
1060         /*
1061          * then try to empty the right most buffer into the middle
1062          */
1063         if (right) {
1064                 u32 right_nr;
1065                 btrfs_tree_lock(right);
1066                 right_nr = btrfs_header_nritems(right);
1067                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1068                         wret = 1;
1069                 } else {
1070                         ret = btrfs_cow_block(trans, root, right,
1071                                               parent, pslot + 1,
1072                                               &right);
1073                         if (ret)
1074                                 wret = 1;
1075                         else {
1076                                 wret = balance_node_right(trans, root,
1077                                                           right, mid);
1078                         }
1079                 }
1080                 if (wret < 0)
1081                         ret = wret;
1082                 if (wret == 0) {
1083                         struct btrfs_disk_key disk_key;
1084
1085                         btrfs_node_key(right, &disk_key, 0);
1086                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
1087                         btrfs_mark_buffer_dirty(parent);
1088
1089                         if (btrfs_header_nritems(mid) <= orig_slot) {
1090                                 path->nodes[level] = right;
1091                                 path->slots[level + 1] += 1;
1092                                 path->slots[level] = orig_slot -
1093                                         btrfs_header_nritems(mid);
1094                                 btrfs_tree_unlock(mid);
1095                                 free_extent_buffer(mid);
1096                         } else {
1097                                 btrfs_tree_unlock(right);
1098                                 free_extent_buffer(right);
1099                         }
1100                         return 0;
1101                 }
1102                 btrfs_tree_unlock(right);
1103                 free_extent_buffer(right);
1104         }
1105         return 1;
1106 }
1107
1108 /*
1109  * readahead one full node of leaves
1110  */
1111 static void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
1112                              int level, int slot, u64 objectid)
1113 {
1114         struct extent_buffer *node;
1115         struct btrfs_disk_key disk_key;
1116         u32 nritems;
1117         u64 search;
1118         u64 lowest_read;
1119         u64 highest_read;
1120         u64 nread = 0;
1121         int direction = path->reada;
1122         struct extent_buffer *eb;
1123         u32 nr;
1124         u32 blocksize;
1125         u32 nscan = 0;
1126
1127         if (level != 1)
1128                 return;
1129
1130         if (!path->nodes[level])
1131                 return;
1132
1133         node = path->nodes[level];
1134
1135         search = btrfs_node_blockptr(node, slot);
1136         blocksize = btrfs_level_size(root, level - 1);
1137         eb = btrfs_find_tree_block(root, search, blocksize);
1138         if (eb) {
1139                 free_extent_buffer(eb);
1140                 return;
1141         }
1142
1143         highest_read = search;
1144         lowest_read = search;
1145
1146         nritems = btrfs_header_nritems(node);
1147         nr = slot;
1148         while(1) {
1149                 if (direction < 0) {
1150                         if (nr == 0)
1151                                 break;
1152                         nr--;
1153                 } else if (direction > 0) {
1154                         nr++;
1155                         if (nr >= nritems)
1156                                 break;
1157                 }
1158                 if (path->reada < 0 && objectid) {
1159                         btrfs_node_key(node, &disk_key, nr);
1160                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1161                                 break;
1162                 }
1163                 search = btrfs_node_blockptr(node, nr);
1164                 if ((search >= lowest_read && search <= highest_read) ||
1165                     (search < lowest_read && lowest_read - search <= 32768) ||
1166                     (search > highest_read && search - highest_read <= 32768)) {
1167                         readahead_tree_block(root, search, blocksize,
1168                                      btrfs_node_ptr_generation(node, nr));
1169                         nread += blocksize;
1170                 }
1171                 nscan++;
1172                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
1173                         break;
1174                 if(nread > (1024 * 1024) || nscan > 128)
1175                         break;
1176
1177                 if (search < lowest_read)
1178                         lowest_read = search;
1179                 if (search > highest_read)
1180                         highest_read = search;
1181         }
1182 }
1183
1184 static void unlock_up(struct btrfs_path *path, int level, int lowest_unlock)
1185 {
1186         int i;
1187         int skip_level = level;
1188         int no_skips = 0;
1189         struct extent_buffer *t;
1190
1191         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1192                 if (!path->nodes[i])
1193                         break;
1194                 if (!path->locks[i])
1195                         break;
1196                 if (!no_skips && path->slots[i] == 0) {
1197                         skip_level = i + 1;
1198                         continue;
1199                 }
1200                 if (!no_skips && path->keep_locks) {
1201                         u32 nritems;
1202                         t = path->nodes[i];
1203                         nritems = btrfs_header_nritems(t);
1204                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
1205                                 skip_level = i + 1;
1206                                 continue;
1207                         }
1208                 }
1209                 if (skip_level < i && i >= lowest_unlock)
1210                         no_skips = 1;
1211
1212                 t = path->nodes[i];
1213                 if (i >= lowest_unlock && i > skip_level && path->locks[i]) {
1214                         btrfs_tree_unlock(t);
1215                         path->locks[i] = 0;
1216                 }
1217         }
1218 }
1219
1220 /*
1221  * look for key in the tree.  path is filled in with nodes along the way
1222  * if key is found, we return zero and you can find the item in the leaf
1223  * level of the path (level 0)
1224  *
1225  * If the key isn't found, the path points to the slot where it should
1226  * be inserted, and 1 is returned.  If there are other errors during the
1227  * search a negative error number is returned.
1228  *
1229  * if ins_len > 0, nodes and leaves will be split as we walk down the
1230  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1231  * possible)
1232  */
1233 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1234                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1235                       ins_len, int cow)
1236 {
1237         struct extent_buffer *b;
1238         struct extent_buffer *tmp;
1239         int slot;
1240         int ret;
1241         int level;
1242         int should_reada = p->reada;
1243         int lowest_unlock = 1;
1244         int blocksize;
1245         u8 lowest_level = 0;
1246         u64 blocknr;
1247         u64 gen;
1248
1249         lowest_level = p->lowest_level;
1250         WARN_ON(lowest_level && ins_len);
1251         WARN_ON(p->nodes[0] != NULL);
1252         WARN_ON(cow && root == root->fs_info->extent_root &&
1253                 !mutex_is_locked(&root->fs_info->alloc_mutex));
1254         WARN_ON(root == root->fs_info->chunk_root &&
1255                 !mutex_is_locked(&root->fs_info->chunk_mutex));
1256         WARN_ON(root == root->fs_info->dev_root &&
1257                 !mutex_is_locked(&root->fs_info->chunk_mutex));
1258         if (ins_len < 0)
1259                 lowest_unlock = 2;
1260 again:
1261         if (p->skip_locking)
1262                 b = btrfs_root_node(root);
1263         else
1264                 b = btrfs_lock_root_node(root);
1265
1266         while (b) {
1267                 level = btrfs_header_level(b);
1268                 if (cow) {
1269                         int wret;
1270                         wret = btrfs_cow_block(trans, root, b,
1271                                                p->nodes[level + 1],
1272                                                p->slots[level + 1],
1273                                                &b);
1274                         if (wret) {
1275                                 free_extent_buffer(b);
1276                                 return wret;
1277                         }
1278                 }
1279                 BUG_ON(!cow && ins_len);
1280                 if (level != btrfs_header_level(b))
1281                         WARN_ON(1);
1282                 level = btrfs_header_level(b);
1283                 p->nodes[level] = b;
1284                 if (!p->skip_locking)
1285                         p->locks[level] = 1;
1286                 ret = check_block(root, p, level);
1287                 if (ret)
1288                         return -1;
1289
1290                 ret = bin_search(b, key, level, &slot);
1291                 if (level != 0) {
1292                         if (ret && slot > 0)
1293                                 slot -= 1;
1294                         p->slots[level] = slot;
1295                         if (ins_len > 0 && btrfs_header_nritems(b) >=
1296                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1297                                 int sret = split_node(trans, root, p, level);
1298                                 BUG_ON(sret > 0);
1299                                 if (sret)
1300                                         return sret;
1301                                 b = p->nodes[level];
1302                                 slot = p->slots[level];
1303                         } else if (ins_len < 0) {
1304                                 int sret = balance_level(trans, root, p,
1305                                                          level);
1306                                 if (sret)
1307                                         return sret;
1308                                 b = p->nodes[level];
1309                                 if (!b) {
1310                                         btrfs_release_path(NULL, p);
1311                                         goto again;
1312                                 }
1313                                 slot = p->slots[level];
1314                                 BUG_ON(btrfs_header_nritems(b) == 1);
1315                         }
1316                         /* this is only true while dropping a snapshot */
1317                         if (level == lowest_level) {
1318                                 unlock_up(p, level, lowest_unlock);
1319                                 break;
1320                         }
1321
1322                         if (should_reada)
1323                                 reada_for_search(root, p, level, slot,
1324                                                  key->objectid);
1325
1326                         blocknr = btrfs_node_blockptr(b, slot);
1327                         gen = btrfs_node_ptr_generation(b, slot);
1328                         blocksize = btrfs_level_size(root, level - 1);
1329
1330                         tmp = btrfs_find_tree_block(root, blocknr, blocksize);
1331                         if (tmp && btrfs_buffer_uptodate(tmp, gen)) {
1332                                 b = tmp;
1333                         } else {
1334                                 /*
1335                                  * reduce lock contention at high levels
1336                                  * of the btree by dropping locks before
1337                                  * we read.
1338                                  */
1339                                 if (level > 1) {
1340                                         btrfs_release_path(NULL, p);
1341                                         if (tmp)
1342                                                 free_extent_buffer(tmp);
1343                                         tmp = read_tree_block(root, blocknr,
1344                                                          blocksize, gen);
1345                                         if (tmp)
1346                                                 free_extent_buffer(tmp);
1347                                         goto again;
1348                                 } else {
1349                                         if (tmp)
1350                                                 free_extent_buffer(tmp);
1351                                         b = read_node_slot(root, b, slot);
1352                                 }
1353                         }
1354                         if (!p->skip_locking)
1355                                 btrfs_tree_lock(b);
1356                         unlock_up(p, level, lowest_unlock);
1357                 } else {
1358                         p->slots[level] = slot;
1359                         if (ins_len > 0 && btrfs_leaf_free_space(root, b) <
1360                             sizeof(struct btrfs_item) + ins_len) {
1361                                 int sret = split_leaf(trans, root, key,
1362                                                       p, ins_len, ret == 0);
1363                                 BUG_ON(sret > 0);
1364                                 if (sret)
1365                                         return sret;
1366                         }
1367                         unlock_up(p, level, lowest_unlock);
1368                         return ret;
1369                 }
1370         }
1371         return 1;
1372 }
1373
1374 /*
1375  * adjust the pointers going up the tree, starting at level
1376  * making sure the right key of each node is points to 'key'.
1377  * This is used after shifting pointers to the left, so it stops
1378  * fixing up pointers when a given leaf/node is not in slot 0 of the
1379  * higher levels
1380  *
1381  * If this fails to write a tree block, it returns -1, but continues
1382  * fixing up the blocks in ram so the tree is consistent.
1383  */
1384 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1385                           struct btrfs_root *root, struct btrfs_path *path,
1386                           struct btrfs_disk_key *key, int level)
1387 {
1388         int i;
1389         int ret = 0;
1390         struct extent_buffer *t;
1391
1392         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1393                 int tslot = path->slots[i];
1394                 if (!path->nodes[i])
1395                         break;
1396                 t = path->nodes[i];
1397                 btrfs_set_node_key(t, key, tslot);
1398                 btrfs_mark_buffer_dirty(path->nodes[i]);
1399                 if (tslot != 0)
1400                         break;
1401         }
1402         return ret;
1403 }
1404
1405 /*
1406  * try to push data from one node into the next node left in the
1407  * tree.
1408  *
1409  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1410  * error, and > 0 if there was no room in the left hand block.
1411  */
1412 static int push_node_left(struct btrfs_trans_handle *trans,
1413                           struct btrfs_root *root, struct extent_buffer *dst,
1414                           struct extent_buffer *src, int empty)
1415 {
1416         int push_items = 0;
1417         int src_nritems;
1418         int dst_nritems;
1419         int ret = 0;
1420
1421         src_nritems = btrfs_header_nritems(src);
1422         dst_nritems = btrfs_header_nritems(dst);
1423         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1424         WARN_ON(btrfs_header_generation(src) != trans->transid);
1425         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1426
1427         if (!empty && src_nritems <= 8)
1428                 return 1;
1429
1430         if (push_items <= 0) {
1431                 return 1;
1432         }
1433
1434         if (empty) {
1435                 push_items = min(src_nritems, push_items);
1436                 if (push_items < src_nritems) {
1437                         /* leave at least 8 pointers in the node if
1438                          * we aren't going to empty it
1439                          */
1440                         if (src_nritems - push_items < 8) {
1441                                 if (push_items <= 8)
1442                                         return 1;
1443                                 push_items -= 8;
1444                         }
1445                 }
1446         } else
1447                 push_items = min(src_nritems - 8, push_items);
1448
1449         copy_extent_buffer(dst, src,
1450                            btrfs_node_key_ptr_offset(dst_nritems),
1451                            btrfs_node_key_ptr_offset(0),
1452                            push_items * sizeof(struct btrfs_key_ptr));
1453
1454         if (push_items < src_nritems) {
1455                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1456                                       btrfs_node_key_ptr_offset(push_items),
1457                                       (src_nritems - push_items) *
1458                                       sizeof(struct btrfs_key_ptr));
1459         }
1460         btrfs_set_header_nritems(src, src_nritems - push_items);
1461         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1462         btrfs_mark_buffer_dirty(src);
1463         btrfs_mark_buffer_dirty(dst);
1464         return ret;
1465 }
1466
1467 /*
1468  * try to push data from one node into the next node right in the
1469  * tree.
1470  *
1471  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1472  * error, and > 0 if there was no room in the right hand block.
1473  *
1474  * this will  only push up to 1/2 the contents of the left node over
1475  */
1476 static int balance_node_right(struct btrfs_trans_handle *trans,
1477                               struct btrfs_root *root,
1478                               struct extent_buffer *dst,
1479                               struct extent_buffer *src)
1480 {
1481         int push_items = 0;
1482         int max_push;
1483         int src_nritems;
1484         int dst_nritems;
1485         int ret = 0;
1486
1487         WARN_ON(btrfs_header_generation(src) != trans->transid);
1488         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1489
1490         src_nritems = btrfs_header_nritems(src);
1491         dst_nritems = btrfs_header_nritems(dst);
1492         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1493         if (push_items <= 0) {
1494                 return 1;
1495         }
1496
1497         if (src_nritems < 4) {
1498                 return 1;
1499         }
1500
1501         max_push = src_nritems / 2 + 1;
1502         /* don't try to empty the node */
1503         if (max_push >= src_nritems) {
1504                 return 1;
1505         }
1506
1507         if (max_push < push_items)
1508                 push_items = max_push;
1509
1510         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1511                                       btrfs_node_key_ptr_offset(0),
1512                                       (dst_nritems) *
1513                                       sizeof(struct btrfs_key_ptr));
1514
1515         copy_extent_buffer(dst, src,
1516                            btrfs_node_key_ptr_offset(0),
1517                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1518                            push_items * sizeof(struct btrfs_key_ptr));
1519
1520         btrfs_set_header_nritems(src, src_nritems - push_items);
1521         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1522
1523         btrfs_mark_buffer_dirty(src);
1524         btrfs_mark_buffer_dirty(dst);
1525         return ret;
1526 }
1527
1528 /*
1529  * helper function to insert a new root level in the tree.
1530  * A new node is allocated, and a single item is inserted to
1531  * point to the existing root
1532  *
1533  * returns zero on success or < 0 on failure.
1534  */
1535 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1536                            struct btrfs_root *root,
1537                            struct btrfs_path *path, int level)
1538 {
1539         u64 root_gen;
1540         u64 lower_gen;
1541         struct extent_buffer *lower;
1542         struct extent_buffer *c;
1543         struct extent_buffer *old;
1544         struct btrfs_disk_key lower_key;
1545
1546         BUG_ON(path->nodes[level]);
1547         BUG_ON(path->nodes[level-1] != root->node);
1548
1549         if (root->ref_cows)
1550                 root_gen = trans->transid;
1551         else
1552                 root_gen = 0;
1553
1554         lower = path->nodes[level-1];
1555         if (level == 1)
1556                 btrfs_item_key(lower, &lower_key, 0);
1557         else
1558                 btrfs_node_key(lower, &lower_key, 0);
1559
1560         c = btrfs_alloc_free_block(trans, root, root->nodesize,
1561                                    root->root_key.objectid,
1562                                    root_gen, lower_key.objectid, level,
1563                                    root->node->start, 0);
1564         if (IS_ERR(c))
1565                 return PTR_ERR(c);
1566
1567         memset_extent_buffer(c, 0, 0, root->nodesize);
1568         btrfs_set_header_nritems(c, 1);
1569         btrfs_set_header_level(c, level);
1570         btrfs_set_header_bytenr(c, c->start);
1571         btrfs_set_header_generation(c, trans->transid);
1572         btrfs_set_header_owner(c, root->root_key.objectid);
1573
1574         write_extent_buffer(c, root->fs_info->fsid,
1575                             (unsigned long)btrfs_header_fsid(c),
1576                             BTRFS_FSID_SIZE);
1577
1578         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1579                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
1580                             BTRFS_UUID_SIZE);
1581
1582         btrfs_set_node_key(c, &lower_key, 0);
1583         btrfs_set_node_blockptr(c, 0, lower->start);
1584         lower_gen = btrfs_header_generation(lower);
1585         WARN_ON(lower_gen == 0);
1586
1587         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1588
1589         btrfs_mark_buffer_dirty(c);
1590
1591         spin_lock(&root->node_lock);
1592         old = root->node;
1593         root->node = c;
1594         spin_unlock(&root->node_lock);
1595
1596         /* the super has an extra ref to root->node */
1597         free_extent_buffer(old);
1598
1599         add_root_to_dirty_list(root);
1600         extent_buffer_get(c);
1601         path->nodes[level] = c;
1602         path->locks[level] = 1;
1603         path->slots[level] = 0;
1604
1605         if (root->ref_cows && lower_gen != trans->transid) {
1606                 struct btrfs_path *back_path = btrfs_alloc_path();
1607                 int ret;
1608                 mutex_lock(&root->fs_info->alloc_mutex);
1609                 ret = btrfs_insert_extent_backref(trans,
1610                                                   root->fs_info->extent_root,
1611                                                   path, lower->start,
1612                                                   root->root_key.objectid,
1613                                                   trans->transid, 0, 0);
1614                 BUG_ON(ret);
1615                 mutex_unlock(&root->fs_info->alloc_mutex);
1616                 btrfs_free_path(back_path);
1617         }
1618         return 0;
1619 }
1620
1621 /*
1622  * worker function to insert a single pointer in a node.
1623  * the node should have enough room for the pointer already
1624  *
1625  * slot and level indicate where you want the key to go, and
1626  * blocknr is the block the key points to.
1627  *
1628  * returns zero on success and < 0 on any error
1629  */
1630 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1631                       *root, struct btrfs_path *path, struct btrfs_disk_key
1632                       *key, u64 bytenr, int slot, int level)
1633 {
1634         struct extent_buffer *lower;
1635         int nritems;
1636
1637         BUG_ON(!path->nodes[level]);
1638         lower = path->nodes[level];
1639         nritems = btrfs_header_nritems(lower);
1640         if (slot > nritems)
1641                 BUG();
1642         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1643                 BUG();
1644         if (slot != nritems) {
1645                 memmove_extent_buffer(lower,
1646                               btrfs_node_key_ptr_offset(slot + 1),
1647                               btrfs_node_key_ptr_offset(slot),
1648                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1649         }
1650         btrfs_set_node_key(lower, key, slot);
1651         btrfs_set_node_blockptr(lower, slot, bytenr);
1652         WARN_ON(trans->transid == 0);
1653         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1654         btrfs_set_header_nritems(lower, nritems + 1);
1655         btrfs_mark_buffer_dirty(lower);
1656         return 0;
1657 }
1658
1659 /*
1660  * split the node at the specified level in path in two.
1661  * The path is corrected to point to the appropriate node after the split
1662  *
1663  * Before splitting this tries to make some room in the node by pushing
1664  * left and right, if either one works, it returns right away.
1665  *
1666  * returns 0 on success and < 0 on failure
1667  */
1668 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1669                       *root, struct btrfs_path *path, int level)
1670 {
1671         u64 root_gen;
1672         struct extent_buffer *c;
1673         struct extent_buffer *split;
1674         struct btrfs_disk_key disk_key;
1675         int mid;
1676         int ret;
1677         int wret;
1678         u32 c_nritems;
1679
1680         c = path->nodes[level];
1681         WARN_ON(btrfs_header_generation(c) != trans->transid);
1682         if (c == root->node) {
1683                 /* trying to split the root, lets make a new one */
1684                 ret = insert_new_root(trans, root, path, level + 1);
1685                 if (ret)
1686                         return ret;
1687         } else {
1688                 ret = push_nodes_for_insert(trans, root, path, level);
1689                 c = path->nodes[level];
1690                 if (!ret && btrfs_header_nritems(c) <
1691                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
1692                         return 0;
1693                 if (ret < 0)
1694                         return ret;
1695         }
1696
1697         c_nritems = btrfs_header_nritems(c);
1698         if (root->ref_cows)
1699                 root_gen = trans->transid;
1700         else
1701                 root_gen = 0;
1702
1703         btrfs_node_key(c, &disk_key, 0);
1704         split = btrfs_alloc_free_block(trans, root, root->nodesize,
1705                                          root->root_key.objectid,
1706                                          root_gen,
1707                                          btrfs_disk_key_objectid(&disk_key),
1708                                          level, c->start, 0);
1709         if (IS_ERR(split))
1710                 return PTR_ERR(split);
1711
1712         btrfs_set_header_flags(split, btrfs_header_flags(c));
1713         btrfs_set_header_level(split, btrfs_header_level(c));
1714         btrfs_set_header_bytenr(split, split->start);
1715         btrfs_set_header_generation(split, trans->transid);
1716         btrfs_set_header_owner(split, root->root_key.objectid);
1717         btrfs_set_header_flags(split, 0);
1718         write_extent_buffer(split, root->fs_info->fsid,
1719                             (unsigned long)btrfs_header_fsid(split),
1720                             BTRFS_FSID_SIZE);
1721         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1722                             (unsigned long)btrfs_header_chunk_tree_uuid(split),
1723                             BTRFS_UUID_SIZE);
1724
1725         mid = (c_nritems + 1) / 2;
1726
1727         copy_extent_buffer(split, c,
1728                            btrfs_node_key_ptr_offset(0),
1729                            btrfs_node_key_ptr_offset(mid),
1730                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1731         btrfs_set_header_nritems(split, c_nritems - mid);
1732         btrfs_set_header_nritems(c, mid);
1733         ret = 0;
1734
1735         btrfs_mark_buffer_dirty(c);
1736         btrfs_mark_buffer_dirty(split);
1737
1738         btrfs_node_key(split, &disk_key, 0);
1739         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1740                           path->slots[level + 1] + 1,
1741                           level + 1);
1742         if (wret)
1743                 ret = wret;
1744
1745         if (path->slots[level] >= mid) {
1746                 path->slots[level] -= mid;
1747                 btrfs_tree_unlock(c);
1748                 free_extent_buffer(c);
1749                 path->nodes[level] = split;
1750                 path->slots[level + 1] += 1;
1751         } else {
1752                 btrfs_tree_unlock(split);
1753                 free_extent_buffer(split);
1754         }
1755         return ret;
1756 }
1757
1758 /*
1759  * how many bytes are required to store the items in a leaf.  start
1760  * and nr indicate which items in the leaf to check.  This totals up the
1761  * space used both by the item structs and the item data
1762  */
1763 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1764 {
1765         int data_len;
1766         int nritems = btrfs_header_nritems(l);
1767         int end = min(nritems, start + nr) - 1;
1768
1769         if (!nr)
1770                 return 0;
1771         data_len = btrfs_item_end_nr(l, start);
1772         data_len = data_len - btrfs_item_offset_nr(l, end);
1773         data_len += sizeof(struct btrfs_item) * nr;
1774         WARN_ON(data_len < 0);
1775         return data_len;
1776 }
1777
1778 /*
1779  * The space between the end of the leaf items and
1780  * the start of the leaf data.  IOW, how much room
1781  * the leaf has left for both items and data
1782  */
1783 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1784 {
1785         int nritems = btrfs_header_nritems(leaf);
1786         int ret;
1787         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1788         if (ret < 0) {
1789                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1790                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1791                        leaf_space_used(leaf, 0, nritems), nritems);
1792         }
1793         return ret;
1794 }
1795
1796 /*
1797  * push some data in the path leaf to the right, trying to free up at
1798  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1799  *
1800  * returns 1 if the push failed because the other node didn't have enough
1801  * room, 0 if everything worked out and < 0 if there were major errors.
1802  */
1803 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1804                            *root, struct btrfs_path *path, int data_size,
1805                            int empty)
1806 {
1807         struct extent_buffer *left = path->nodes[0];
1808         struct extent_buffer *right;
1809         struct extent_buffer *upper;
1810         struct btrfs_disk_key disk_key;
1811         int slot;
1812         u32 i;
1813         int free_space;
1814         int push_space = 0;
1815         int push_items = 0;
1816         struct btrfs_item *item;
1817         u32 left_nritems;
1818         u32 nr;
1819         u32 right_nritems;
1820         u32 data_end;
1821         u32 this_item_size;
1822         int ret;
1823
1824         slot = path->slots[1];
1825         if (!path->nodes[1]) {
1826                 return 1;
1827         }
1828         upper = path->nodes[1];
1829         if (slot >= btrfs_header_nritems(upper) - 1)
1830                 return 1;
1831
1832         WARN_ON(!btrfs_tree_locked(path->nodes[1]));
1833
1834         right = read_node_slot(root, upper, slot + 1);
1835         btrfs_tree_lock(right);
1836         free_space = btrfs_leaf_free_space(root, right);
1837         if (free_space < data_size + sizeof(struct btrfs_item))
1838                 goto out_unlock;
1839
1840         /* cow and double check */
1841         ret = btrfs_cow_block(trans, root, right, upper,
1842                               slot + 1, &right);
1843         if (ret)
1844                 goto out_unlock;
1845
1846         free_space = btrfs_leaf_free_space(root, right);
1847         if (free_space < data_size + sizeof(struct btrfs_item))
1848                 goto out_unlock;
1849
1850         left_nritems = btrfs_header_nritems(left);
1851         if (left_nritems == 0)
1852                 goto out_unlock;
1853
1854         if (empty)
1855                 nr = 0;
1856         else
1857                 nr = 1;
1858
1859         i = left_nritems - 1;
1860         while (i >= nr) {
1861                 item = btrfs_item_nr(left, i);
1862
1863                 if (path->slots[0] == i)
1864                         push_space += data_size + sizeof(*item);
1865
1866                 if (!left->map_token) {
1867                         map_extent_buffer(left, (unsigned long)item,
1868                                         sizeof(struct btrfs_item),
1869                                         &left->map_token, &left->kaddr,
1870                                         &left->map_start, &left->map_len,
1871                                         KM_USER1);
1872                 }
1873
1874                 this_item_size = btrfs_item_size(left, item);
1875                 if (this_item_size + sizeof(*item) + push_space > free_space)
1876                         break;
1877                 push_items++;
1878                 push_space += this_item_size + sizeof(*item);
1879                 if (i == 0)
1880                         break;
1881                 i--;
1882         }
1883         if (left->map_token) {
1884                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1885                 left->map_token = NULL;
1886         }
1887
1888         if (push_items == 0)
1889                 goto out_unlock;
1890
1891         if (!empty && push_items == left_nritems)
1892                 WARN_ON(1);
1893
1894         /* push left to right */
1895         right_nritems = btrfs_header_nritems(right);
1896
1897         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1898         push_space -= leaf_data_end(root, left);
1899
1900         /* make room in the right data area */
1901         data_end = leaf_data_end(root, right);
1902         memmove_extent_buffer(right,
1903                               btrfs_leaf_data(right) + data_end - push_space,
1904                               btrfs_leaf_data(right) + data_end,
1905                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1906
1907         /* copy from the left data area */
1908         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1909                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1910                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1911                      push_space);
1912
1913         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1914                               btrfs_item_nr_offset(0),
1915                               right_nritems * sizeof(struct btrfs_item));
1916
1917         /* copy the items from left to right */
1918         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1919                    btrfs_item_nr_offset(left_nritems - push_items),
1920                    push_items * sizeof(struct btrfs_item));
1921
1922         /* update the item pointers */
1923         right_nritems += push_items;
1924         btrfs_set_header_nritems(right, right_nritems);
1925         push_space = BTRFS_LEAF_DATA_SIZE(root);
1926         for (i = 0; i < right_nritems; i++) {
1927                 item = btrfs_item_nr(right, i);
1928                 if (!right->map_token) {
1929                         map_extent_buffer(right, (unsigned long)item,
1930                                         sizeof(struct btrfs_item),
1931                                         &right->map_token, &right->kaddr,
1932                                         &right->map_start, &right->map_len,
1933                                         KM_USER1);
1934                 }
1935                 push_space -= btrfs_item_size(right, item);
1936                 btrfs_set_item_offset(right, item, push_space);
1937         }
1938
1939         if (right->map_token) {
1940                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1941                 right->map_token = NULL;
1942         }
1943         left_nritems -= push_items;
1944         btrfs_set_header_nritems(left, left_nritems);
1945
1946         if (left_nritems)
1947                 btrfs_mark_buffer_dirty(left);
1948         btrfs_mark_buffer_dirty(right);
1949
1950         btrfs_item_key(right, &disk_key, 0);
1951         btrfs_set_node_key(upper, &disk_key, slot + 1);
1952         btrfs_mark_buffer_dirty(upper);
1953
1954         /* then fixup the leaf pointer in the path */
1955         if (path->slots[0] >= left_nritems) {
1956                 path->slots[0] -= left_nritems;
1957                 if (btrfs_header_nritems(path->nodes[0]) == 0)
1958                         clean_tree_block(trans, root, path->nodes[0]);
1959                 btrfs_tree_unlock(path->nodes[0]);
1960                 free_extent_buffer(path->nodes[0]);
1961                 path->nodes[0] = right;
1962                 path->slots[1] += 1;
1963         } else {
1964                 btrfs_tree_unlock(right);
1965                 free_extent_buffer(right);
1966         }
1967         return 0;
1968
1969 out_unlock:
1970         btrfs_tree_unlock(right);
1971         free_extent_buffer(right);
1972         return 1;
1973 }
1974
1975 /*
1976  * push some data in the path leaf to the left, trying to free up at
1977  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1978  */
1979 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1980                           *root, struct btrfs_path *path, int data_size,
1981                           int empty)
1982 {
1983         struct btrfs_disk_key disk_key;
1984         struct extent_buffer *right = path->nodes[0];
1985         struct extent_buffer *left;
1986         int slot;
1987         int i;
1988         int free_space;
1989         int push_space = 0;
1990         int push_items = 0;
1991         struct btrfs_item *item;
1992         u32 old_left_nritems;
1993         u32 right_nritems;
1994         u32 nr;
1995         int ret = 0;
1996         int wret;
1997         u32 this_item_size;
1998         u32 old_left_item_size;
1999
2000         slot = path->slots[1];
2001         if (slot == 0)
2002                 return 1;
2003         if (!path->nodes[1])
2004                 return 1;
2005
2006         right_nritems = btrfs_header_nritems(right);
2007         if (right_nritems == 0) {
2008                 return 1;
2009         }
2010
2011         WARN_ON(!btrfs_tree_locked(path->nodes[1]));
2012
2013         left = read_node_slot(root, path->nodes[1], slot - 1);
2014         btrfs_tree_lock(left);
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         /* cow and double check */
2022         ret = btrfs_cow_block(trans, root, left,
2023                               path->nodes[1], slot - 1, &left);
2024         if (ret) {
2025                 /* we hit -ENOSPC, but it isn't fatal here */
2026                 ret = 1;
2027                 goto out;
2028         }
2029
2030         free_space = btrfs_leaf_free_space(root, left);
2031         if (free_space < data_size + sizeof(struct btrfs_item)) {
2032                 ret = 1;
2033                 goto out;
2034         }
2035
2036         if (empty)
2037                 nr = right_nritems;
2038         else
2039                 nr = right_nritems - 1;
2040
2041         for (i = 0; i < nr; i++) {
2042                 item = btrfs_item_nr(right, i);
2043                 if (!right->map_token) {
2044                         map_extent_buffer(right, (unsigned long)item,
2045                                         sizeof(struct btrfs_item),
2046                                         &right->map_token, &right->kaddr,
2047                                         &right->map_start, &right->map_len,
2048                                         KM_USER1);
2049                 }
2050
2051                 if (path->slots[0] == i)
2052                         push_space += data_size + sizeof(*item);
2053
2054                 this_item_size = btrfs_item_size(right, item);
2055                 if (this_item_size + sizeof(*item) + push_space > free_space)
2056                         break;
2057
2058                 push_items++;
2059                 push_space += this_item_size + sizeof(*item);
2060         }
2061
2062         if (right->map_token) {
2063                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2064                 right->map_token = NULL;
2065         }
2066
2067         if (push_items == 0) {
2068                 ret = 1;
2069                 goto out;
2070         }
2071         if (!empty && push_items == btrfs_header_nritems(right))
2072                 WARN_ON(1);
2073
2074         /* push data from right to left */
2075         copy_extent_buffer(left, right,
2076                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
2077                            btrfs_item_nr_offset(0),
2078                            push_items * sizeof(struct btrfs_item));
2079
2080         push_space = BTRFS_LEAF_DATA_SIZE(root) -
2081                      btrfs_item_offset_nr(right, push_items -1);
2082
2083         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
2084                      leaf_data_end(root, left) - push_space,
2085                      btrfs_leaf_data(right) +
2086                      btrfs_item_offset_nr(right, push_items - 1),
2087                      push_space);
2088         old_left_nritems = btrfs_header_nritems(left);
2089         BUG_ON(old_left_nritems < 0);
2090
2091         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
2092         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
2093                 u32 ioff;
2094
2095                 item = btrfs_item_nr(left, i);
2096                 if (!left->map_token) {
2097                         map_extent_buffer(left, (unsigned long)item,
2098                                         sizeof(struct btrfs_item),
2099                                         &left->map_token, &left->kaddr,
2100                                         &left->map_start, &left->map_len,
2101                                         KM_USER1);
2102                 }
2103
2104                 ioff = btrfs_item_offset(left, item);
2105                 btrfs_set_item_offset(left, item,
2106                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
2107         }
2108         btrfs_set_header_nritems(left, old_left_nritems + push_items);
2109         if (left->map_token) {
2110                 unmap_extent_buffer(left, left->map_token, KM_USER1);
2111                 left->map_token = NULL;
2112         }
2113
2114         /* fixup right node */
2115         if (push_items > right_nritems) {
2116                 printk("push items %d nr %u\n", push_items, right_nritems);
2117                 WARN_ON(1);
2118         }
2119
2120         if (push_items < right_nritems) {
2121                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
2122                                                   leaf_data_end(root, right);
2123                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
2124                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
2125                                       btrfs_leaf_data(right) +
2126                                       leaf_data_end(root, right), push_space);
2127
2128                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
2129                               btrfs_item_nr_offset(push_items),
2130                              (btrfs_header_nritems(right) - push_items) *
2131                              sizeof(struct btrfs_item));
2132         }
2133         right_nritems -= push_items;
2134         btrfs_set_header_nritems(right, right_nritems);
2135         push_space = BTRFS_LEAF_DATA_SIZE(root);
2136         for (i = 0; i < right_nritems; i++) {
2137                 item = btrfs_item_nr(right, i);
2138
2139                 if (!right->map_token) {
2140                         map_extent_buffer(right, (unsigned long)item,
2141                                         sizeof(struct btrfs_item),
2142                                         &right->map_token, &right->kaddr,
2143                                         &right->map_start, &right->map_len,
2144                                         KM_USER1);
2145                 }
2146
2147                 push_space = push_space - btrfs_item_size(right, item);
2148                 btrfs_set_item_offset(right, item, push_space);
2149         }
2150         if (right->map_token) {
2151                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2152                 right->map_token = NULL;
2153         }
2154
2155         btrfs_mark_buffer_dirty(left);
2156         if (right_nritems)
2157                 btrfs_mark_buffer_dirty(right);
2158
2159         btrfs_item_key(right, &disk_key, 0);
2160         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
2161         if (wret)
2162                 ret = wret;
2163
2164         /* then fixup the leaf pointer in the path */
2165         if (path->slots[0] < push_items) {
2166                 path->slots[0] += old_left_nritems;
2167                 if (btrfs_header_nritems(path->nodes[0]) == 0)
2168                         clean_tree_block(trans, root, path->nodes[0]);
2169                 btrfs_tree_unlock(path->nodes[0]);
2170                 free_extent_buffer(path->nodes[0]);
2171                 path->nodes[0] = left;
2172                 path->slots[1] -= 1;
2173         } else {
2174                 btrfs_tree_unlock(left);
2175                 free_extent_buffer(left);
2176                 path->slots[0] -= push_items;
2177         }
2178         BUG_ON(path->slots[0] < 0);
2179         return ret;
2180 out:
2181         btrfs_tree_unlock(left);
2182         free_extent_buffer(left);
2183         return ret;
2184 }
2185
2186 /*
2187  * split the path's leaf in two, making sure there is at least data_size
2188  * available for the resulting leaf level of the path.
2189  *
2190  * returns 0 if all went well and < 0 on failure.
2191  */
2192 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
2193                       *root, struct btrfs_key *ins_key,
2194                       struct btrfs_path *path, int data_size, int extend)
2195 {
2196         u64 root_gen;
2197         struct extent_buffer *l;
2198         u32 nritems;
2199         int mid;
2200         int slot;
2201         struct extent_buffer *right;
2202         int space_needed = data_size + sizeof(struct btrfs_item);
2203         int data_copy_size;
2204         int rt_data_off;
2205         int i;
2206         int ret = 0;
2207         int wret;
2208         int double_split;
2209         int num_doubles = 0;
2210         struct btrfs_disk_key disk_key;
2211
2212         if (extend)
2213                 space_needed = data_size;
2214
2215         if (root->ref_cows)
2216                 root_gen = trans->transid;
2217         else
2218                 root_gen = 0;
2219
2220         /* first try to make some room by pushing left and right */
2221         if (ins_key->type != BTRFS_DIR_ITEM_KEY) {
2222                 wret = push_leaf_right(trans, root, path, data_size, 0);
2223                 if (wret < 0) {
2224                         return wret;
2225                 }
2226                 if (wret) {
2227                         wret = push_leaf_left(trans, root, path, data_size, 0);
2228                         if (wret < 0)
2229                                 return wret;
2230                 }
2231                 l = path->nodes[0];
2232
2233                 /* did the pushes work? */
2234                 if (btrfs_leaf_free_space(root, l) >= space_needed)
2235                         return 0;
2236         }
2237
2238         if (!path->nodes[1]) {
2239                 ret = insert_new_root(trans, root, path, 1);
2240                 if (ret)
2241                         return ret;
2242         }
2243 again:
2244         double_split = 0;
2245         l = path->nodes[0];
2246         slot = path->slots[0];
2247         nritems = btrfs_header_nritems(l);
2248         mid = (nritems + 1)/ 2;
2249
2250         btrfs_item_key(l, &disk_key, 0);
2251
2252         right = btrfs_alloc_free_block(trans, root, root->leafsize,
2253                                          root->root_key.objectid,
2254                                          root_gen, disk_key.objectid, 0,
2255                                          l->start, 0);
2256         if (IS_ERR(right)) {
2257                 BUG_ON(1);
2258                 return PTR_ERR(right);
2259         }
2260
2261         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2262         btrfs_set_header_bytenr(right, right->start);
2263         btrfs_set_header_generation(right, trans->transid);
2264         btrfs_set_header_owner(right, root->root_key.objectid);
2265         btrfs_set_header_level(right, 0);
2266         write_extent_buffer(right, root->fs_info->fsid,
2267                             (unsigned long)btrfs_header_fsid(right),
2268                             BTRFS_FSID_SIZE);
2269
2270         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2271                             (unsigned long)btrfs_header_chunk_tree_uuid(right),
2272                             BTRFS_UUID_SIZE);
2273         if (mid <= slot) {
2274                 if (nritems == 1 ||
2275                     leaf_space_used(l, mid, nritems - mid) + space_needed >
2276                         BTRFS_LEAF_DATA_SIZE(root)) {
2277                         if (slot >= nritems) {
2278                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2279                                 btrfs_set_header_nritems(right, 0);
2280                                 wret = insert_ptr(trans, root, path,
2281                                                   &disk_key, right->start,
2282                                                   path->slots[1] + 1, 1);
2283                                 if (wret)
2284                                         ret = wret;
2285
2286                                 btrfs_tree_unlock(path->nodes[0]);
2287                                 free_extent_buffer(path->nodes[0]);
2288                                 path->nodes[0] = right;
2289                                 path->slots[0] = 0;
2290                                 path->slots[1] += 1;
2291                                 btrfs_mark_buffer_dirty(right);
2292                                 return ret;
2293                         }
2294                         mid = slot;
2295                         if (mid != nritems &&
2296                             leaf_space_used(l, mid, nritems - mid) +
2297                             space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2298                                 double_split = 1;
2299                         }
2300                 }
2301         } else {
2302                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
2303                         BTRFS_LEAF_DATA_SIZE(root)) {
2304                         if (!extend && slot == 0) {
2305                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2306                                 btrfs_set_header_nritems(right, 0);
2307                                 wret = insert_ptr(trans, root, path,
2308                                                   &disk_key,
2309                                                   right->start,
2310                                                   path->slots[1], 1);
2311                                 if (wret)
2312                                         ret = wret;
2313                                 btrfs_tree_unlock(path->nodes[0]);
2314                                 free_extent_buffer(path->nodes[0]);
2315                                 path->nodes[0] = right;
2316                                 path->slots[0] = 0;
2317                                 if (path->slots[1] == 0) {
2318                                         wret = fixup_low_keys(trans, root,
2319                                                    path, &disk_key, 1);
2320                                         if (wret)
2321                                                 ret = wret;
2322                                 }
2323                                 btrfs_mark_buffer_dirty(right);
2324                                 return ret;
2325                         } else if (extend && slot == 0) {
2326                                 mid = 1;
2327                         } else {
2328                                 mid = slot;
2329                                 if (mid != nritems &&
2330                                     leaf_space_used(l, mid, nritems - mid) +
2331                                     space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
2332                                         double_split = 1;
2333                                 }
2334                         }
2335                 }
2336         }
2337         nritems = nritems - mid;
2338         btrfs_set_header_nritems(right, nritems);
2339         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2340
2341         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2342                            btrfs_item_nr_offset(mid),
2343                            nritems * sizeof(struct btrfs_item));
2344
2345         copy_extent_buffer(right, l,
2346                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2347                      data_copy_size, btrfs_leaf_data(l) +
2348                      leaf_data_end(root, l), data_copy_size);
2349
2350         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2351                       btrfs_item_end_nr(l, mid);
2352
2353         for (i = 0; i < nritems; i++) {
2354                 struct btrfs_item *item = btrfs_item_nr(right, i);
2355                 u32 ioff;
2356
2357                 if (!right->map_token) {
2358                         map_extent_buffer(right, (unsigned long)item,
2359                                         sizeof(struct btrfs_item),
2360                                         &right->map_token, &right->kaddr,
2361                                         &right->map_start, &right->map_len,
2362                                         KM_USER1);
2363                 }
2364
2365                 ioff = btrfs_item_offset(right, item);
2366                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2367         }
2368
2369         if (right->map_token) {
2370                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2371                 right->map_token = NULL;
2372         }
2373
2374         btrfs_set_header_nritems(l, mid);
2375         ret = 0;
2376         btrfs_item_key(right, &disk_key, 0);
2377         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2378                           path->slots[1] + 1, 1);
2379         if (wret)
2380                 ret = wret;
2381
2382         btrfs_mark_buffer_dirty(right);
2383         btrfs_mark_buffer_dirty(l);
2384         BUG_ON(path->slots[0] != slot);
2385
2386         if (mid <= slot) {
2387                 btrfs_tree_unlock(path->nodes[0]);
2388                 free_extent_buffer(path->nodes[0]);
2389                 path->nodes[0] = right;
2390                 path->slots[0] -= mid;
2391                 path->slots[1] += 1;
2392         } else {
2393                 btrfs_tree_unlock(right);
2394                 free_extent_buffer(right);
2395         }
2396
2397         BUG_ON(path->slots[0] < 0);
2398
2399         if (double_split) {
2400                 BUG_ON(num_doubles != 0);
2401                 num_doubles++;
2402                 goto again;
2403         }
2404         return ret;
2405 }
2406
2407 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2408                         struct btrfs_root *root,
2409                         struct btrfs_path *path,
2410                         u32 new_size, int from_end)
2411 {
2412         int ret = 0;
2413         int slot;
2414         int slot_orig;
2415         struct extent_buffer *leaf;
2416         struct btrfs_item *item;
2417         u32 nritems;
2418         unsigned int data_end;
2419         unsigned int old_data_start;
2420         unsigned int old_size;
2421         unsigned int size_diff;
2422         int i;
2423
2424         slot_orig = path->slots[0];
2425         leaf = path->nodes[0];
2426         slot = path->slots[0];
2427
2428         old_size = btrfs_item_size_nr(leaf, slot);
2429         if (old_size == new_size)
2430                 return 0;
2431
2432         nritems = btrfs_header_nritems(leaf);
2433         data_end = leaf_data_end(root, leaf);
2434
2435         old_data_start = btrfs_item_offset_nr(leaf, slot);
2436
2437         size_diff = old_size - new_size;
2438
2439         BUG_ON(slot < 0);
2440         BUG_ON(slot >= nritems);
2441
2442         /*
2443          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2444          */
2445         /* first correct the data pointers */
2446         for (i = slot; i < nritems; i++) {
2447                 u32 ioff;
2448                 item = btrfs_item_nr(leaf, i);
2449
2450                 if (!leaf->map_token) {
2451                         map_extent_buffer(leaf, (unsigned long)item,
2452                                         sizeof(struct btrfs_item),
2453                                         &leaf->map_token, &leaf->kaddr,
2454                                         &leaf->map_start, &leaf->map_len,
2455                                         KM_USER1);
2456                 }
2457
2458                 ioff = btrfs_item_offset(leaf, item);
2459                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2460         }
2461
2462         if (leaf->map_token) {
2463                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2464                 leaf->map_token = NULL;
2465         }
2466
2467         /* shift the data */
2468         if (from_end) {
2469                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2470                               data_end + size_diff, btrfs_leaf_data(leaf) +
2471                               data_end, old_data_start + new_size - data_end);
2472         } else {
2473                 struct btrfs_disk_key disk_key;
2474                 u64 offset;
2475
2476                 btrfs_item_key(leaf, &disk_key, slot);
2477
2478                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2479                         unsigned long ptr;
2480                         struct btrfs_file_extent_item *fi;
2481
2482                         fi = btrfs_item_ptr(leaf, slot,
2483                                             struct btrfs_file_extent_item);
2484                         fi = (struct btrfs_file_extent_item *)(
2485                              (unsigned long)fi - size_diff);
2486
2487                         if (btrfs_file_extent_type(leaf, fi) ==
2488                             BTRFS_FILE_EXTENT_INLINE) {
2489                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2490                                 memmove_extent_buffer(leaf, ptr,
2491                                         (unsigned long)fi,
2492                                         offsetof(struct btrfs_file_extent_item,
2493                                                  disk_bytenr));
2494                         }
2495                 }
2496
2497                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2498                               data_end + size_diff, btrfs_leaf_data(leaf) +
2499                               data_end, old_data_start - data_end);
2500
2501                 offset = btrfs_disk_key_offset(&disk_key);
2502                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2503                 btrfs_set_item_key(leaf, &disk_key, slot);
2504                 if (slot == 0)
2505                         fixup_low_keys(trans, root, path, &disk_key, 1);
2506         }
2507
2508         item = btrfs_item_nr(leaf, slot);
2509         btrfs_set_item_size(leaf, item, new_size);
2510         btrfs_mark_buffer_dirty(leaf);
2511
2512         ret = 0;
2513         if (btrfs_leaf_free_space(root, leaf) < 0) {
2514                 btrfs_print_leaf(root, leaf);
2515                 BUG();
2516         }
2517         return ret;
2518 }
2519
2520 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2521                       struct btrfs_root *root, struct btrfs_path *path,
2522                       u32 data_size)
2523 {
2524         int ret = 0;
2525         int slot;
2526         int slot_orig;
2527         struct extent_buffer *leaf;
2528         struct btrfs_item *item;
2529         u32 nritems;
2530         unsigned int data_end;
2531         unsigned int old_data;
2532         unsigned int old_size;
2533         int i;
2534
2535         slot_orig = path->slots[0];
2536         leaf = path->nodes[0];
2537
2538         nritems = btrfs_header_nritems(leaf);
2539         data_end = leaf_data_end(root, leaf);
2540
2541         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2542                 btrfs_print_leaf(root, leaf);
2543                 BUG();
2544         }
2545         slot = path->slots[0];
2546         old_data = btrfs_item_end_nr(leaf, slot);
2547
2548         BUG_ON(slot < 0);
2549         if (slot >= nritems) {
2550                 btrfs_print_leaf(root, leaf);
2551                 printk("slot %d too large, nritems %d\n", slot, nritems);
2552                 BUG_ON(1);
2553         }
2554
2555         /*
2556          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2557          */
2558         /* first correct the data pointers */
2559         for (i = slot; i < nritems; i++) {
2560                 u32 ioff;
2561                 item = btrfs_item_nr(leaf, i);
2562
2563                 if (!leaf->map_token) {
2564                         map_extent_buffer(leaf, (unsigned long)item,
2565                                         sizeof(struct btrfs_item),
2566                                         &leaf->map_token, &leaf->kaddr,
2567                                         &leaf->map_start, &leaf->map_len,
2568                                         KM_USER1);
2569                 }
2570                 ioff = btrfs_item_offset(leaf, item);
2571                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2572         }
2573
2574         if (leaf->map_token) {
2575                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2576                 leaf->map_token = NULL;
2577         }
2578
2579         /* shift the data */
2580         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2581                       data_end - data_size, btrfs_leaf_data(leaf) +
2582                       data_end, old_data - data_end);
2583
2584         data_end = old_data;
2585         old_size = btrfs_item_size_nr(leaf, slot);
2586         item = btrfs_item_nr(leaf, slot);
2587         btrfs_set_item_size(leaf, item, old_size + data_size);
2588         btrfs_mark_buffer_dirty(leaf);
2589
2590         ret = 0;
2591         if (btrfs_leaf_free_space(root, leaf) < 0) {
2592                 btrfs_print_leaf(root, leaf);
2593                 BUG();
2594         }
2595         return ret;
2596 }
2597
2598 /*
2599  * Given a key and some data, insert an item into the tree.
2600  * This does all the path init required, making room in the tree if needed.
2601  */
2602 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2603                             struct btrfs_root *root,
2604                             struct btrfs_path *path,
2605                             struct btrfs_key *cpu_key, u32 *data_size,
2606                             int nr)
2607 {
2608         struct extent_buffer *leaf;
2609         struct btrfs_item *item;
2610         int ret = 0;
2611         int slot;
2612         int slot_orig;
2613         int i;
2614         u32 nritems;
2615         u32 total_size = 0;
2616         u32 total_data = 0;
2617         unsigned int data_end;
2618         struct btrfs_disk_key disk_key;
2619
2620         for (i = 0; i < nr; i++) {
2621                 total_data += data_size[i];
2622         }
2623
2624         total_size = total_data + (nr - 1) * sizeof(struct btrfs_item);
2625         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2626         if (ret == 0) {
2627                 return -EEXIST;
2628         }
2629         if (ret < 0)
2630                 goto out;
2631
2632         slot_orig = path->slots[0];
2633         leaf = path->nodes[0];
2634
2635         nritems = btrfs_header_nritems(leaf);
2636         data_end = leaf_data_end(root, leaf);
2637
2638         if (btrfs_leaf_free_space(root, leaf) <
2639             sizeof(struct btrfs_item) + total_size) {
2640                 btrfs_print_leaf(root, leaf);
2641                 printk("not enough freespace need %u have %d\n",
2642                        total_size, btrfs_leaf_free_space(root, leaf));
2643                 BUG();
2644         }
2645
2646         slot = path->slots[0];
2647         BUG_ON(slot < 0);
2648
2649         if (slot != nritems) {
2650                 int i;
2651                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2652
2653                 if (old_data < data_end) {
2654                         btrfs_print_leaf(root, leaf);
2655                         printk("slot %d old_data %d data_end %d\n",
2656                                slot, old_data, data_end);
2657                         BUG_ON(1);
2658                 }
2659                 /*
2660                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2661                  */
2662                 /* first correct the data pointers */
2663                 WARN_ON(leaf->map_token);
2664                 for (i = slot; i < nritems; i++) {
2665                         u32 ioff;
2666
2667                         item = btrfs_item_nr(leaf, i);
2668                         if (!leaf->map_token) {
2669                                 map_extent_buffer(leaf, (unsigned long)item,
2670                                         sizeof(struct btrfs_item),
2671                                         &leaf->map_token, &leaf->kaddr,
2672                                         &leaf->map_start, &leaf->map_len,
2673                                         KM_USER1);
2674                         }
2675
2676                         ioff = btrfs_item_offset(leaf, item);
2677                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2678                 }
2679                 if (leaf->map_token) {
2680                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2681                         leaf->map_token = NULL;
2682                 }
2683
2684                 /* shift the items */
2685                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2686                               btrfs_item_nr_offset(slot),
2687                               (nritems - slot) * sizeof(struct btrfs_item));
2688
2689                 /* shift the data */
2690                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2691                               data_end - total_data, btrfs_leaf_data(leaf) +
2692                               data_end, old_data - data_end);
2693                 data_end = old_data;
2694         }
2695
2696         /* setup the item for the new data */
2697         for (i = 0; i < nr; i++) {
2698                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2699                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2700                 item = btrfs_item_nr(leaf, slot + i);
2701                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2702                 data_end -= data_size[i];
2703                 btrfs_set_item_size(leaf, item, data_size[i]);
2704         }
2705         btrfs_set_header_nritems(leaf, nritems + nr);
2706         btrfs_mark_buffer_dirty(leaf);
2707
2708         ret = 0;
2709         if (slot == 0) {
2710                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2711                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2712         }
2713
2714         if (btrfs_leaf_free_space(root, leaf) < 0) {
2715                 btrfs_print_leaf(root, leaf);
2716                 BUG();
2717         }
2718 out:
2719         return ret;
2720 }
2721
2722 /*
2723  * Given a key and some data, insert an item into the tree.
2724  * This does all the path init required, making room in the tree if needed.
2725  */
2726 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2727                       *root, struct btrfs_key *cpu_key, void *data, u32
2728                       data_size)
2729 {
2730         int ret = 0;
2731         struct btrfs_path *path;
2732         struct extent_buffer *leaf;
2733         unsigned long ptr;
2734
2735         path = btrfs_alloc_path();
2736         BUG_ON(!path);
2737         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2738         if (!ret) {
2739                 leaf = path->nodes[0];
2740                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2741                 write_extent_buffer(leaf, data, ptr, data_size);
2742                 btrfs_mark_buffer_dirty(leaf);
2743         }
2744         btrfs_free_path(path);
2745         return ret;
2746 }
2747
2748 /*
2749  * delete the pointer from a given node.
2750  *
2751  * If the delete empties a node, the node is removed from the tree,
2752  * continuing all the way the root if required.  The root is converted into
2753  * a leaf if all the nodes are emptied.
2754  */
2755 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2756                    struct btrfs_path *path, int level, int slot)
2757 {
2758         struct extent_buffer *parent = path->nodes[level];
2759         u32 nritems;
2760         int ret = 0;
2761         int wret;
2762
2763         nritems = btrfs_header_nritems(parent);
2764         if (slot != nritems -1) {
2765                 memmove_extent_buffer(parent,
2766                               btrfs_node_key_ptr_offset(slot),
2767                               btrfs_node_key_ptr_offset(slot + 1),
2768                               sizeof(struct btrfs_key_ptr) *
2769                               (nritems - slot - 1));
2770         }
2771         nritems--;
2772         btrfs_set_header_nritems(parent, nritems);
2773         if (nritems == 0 && parent == root->node) {
2774                 BUG_ON(btrfs_header_level(root->node) != 1);
2775                 /* just turn the root into a leaf and break */
2776                 btrfs_set_header_level(root->node, 0);
2777         } else if (slot == 0) {
2778                 struct btrfs_disk_key disk_key;
2779
2780                 btrfs_node_key(parent, &disk_key, 0);
2781                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2782                 if (wret)
2783                         ret = wret;
2784         }
2785         btrfs_mark_buffer_dirty(parent);
2786         return ret;
2787 }
2788
2789 /*
2790  * delete the item at the leaf level in path.  If that empties
2791  * the leaf, remove it from the tree
2792  */
2793 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2794                     struct btrfs_path *path, int slot, int nr)
2795 {
2796         struct extent_buffer *leaf;
2797         struct btrfs_item *item;
2798         int last_off;
2799         int dsize = 0;
2800         int ret = 0;
2801         int wret;
2802         int i;
2803         u32 nritems;
2804
2805         leaf = path->nodes[0];
2806         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2807
2808         for (i = 0; i < nr; i++)
2809                 dsize += btrfs_item_size_nr(leaf, slot + i);
2810
2811         nritems = btrfs_header_nritems(leaf);
2812
2813         if (slot + nr != nritems) {
2814                 int i;
2815                 int data_end = leaf_data_end(root, leaf);
2816
2817                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2818                               data_end + dsize,
2819                               btrfs_leaf_data(leaf) + data_end,
2820                               last_off - data_end);
2821
2822                 for (i = slot + nr; i < nritems; i++) {
2823                         u32 ioff;
2824
2825                         item = btrfs_item_nr(leaf, i);
2826                         if (!leaf->map_token) {
2827                                 map_extent_buffer(leaf, (unsigned long)item,
2828                                         sizeof(struct btrfs_item),
2829                                         &leaf->map_token, &leaf->kaddr,
2830                                         &leaf->map_start, &leaf->map_len,
2831                                         KM_USER1);
2832                         }
2833                         ioff = btrfs_item_offset(leaf, item);
2834                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2835                 }
2836
2837                 if (leaf->map_token) {
2838                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2839                         leaf->map_token = NULL;
2840                 }
2841
2842                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2843                               btrfs_item_nr_offset(slot + nr),
2844                               sizeof(struct btrfs_item) *
2845                               (nritems - slot - nr));
2846         }
2847         btrfs_set_header_nritems(leaf, nritems - nr);
2848         nritems -= nr;
2849
2850         /* delete the leaf if we've emptied it */
2851         if (nritems == 0) {
2852                 if (leaf == root->node) {
2853                         btrfs_set_header_level(leaf, 0);
2854                 } else {
2855                         u64 root_gen = btrfs_header_generation(path->nodes[1]);
2856                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
2857                         if (wret)
2858                                 ret = wret;
2859                         wret = btrfs_free_extent(trans, root,
2860                                          leaf->start, leaf->len,
2861                                          btrfs_header_owner(path->nodes[1]),
2862                                          root_gen, 0, 0, 1);
2863                         if (wret)
2864                                 ret = wret;
2865                 }
2866         } else {
2867                 int used = leaf_space_used(leaf, 0, nritems);
2868                 if (slot == 0) {
2869                         struct btrfs_disk_key disk_key;
2870
2871                         btrfs_item_key(leaf, &disk_key, 0);
2872                         wret = fixup_low_keys(trans, root, path,
2873                                               &disk_key, 1);
2874                         if (wret)
2875                                 ret = wret;
2876                 }
2877
2878                 /* delete the leaf if it is mostly empty */
2879                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2880                         /* push_leaf_left fixes the path.
2881                          * make sure the path still points to our leaf
2882                          * for possible call to del_ptr below
2883                          */
2884                         slot = path->slots[1];
2885                         extent_buffer_get(leaf);
2886
2887                         wret = push_leaf_left(trans, root, path, 1, 1);
2888                         if (wret < 0 && wret != -ENOSPC)
2889                                 ret = wret;
2890
2891                         if (path->nodes[0] == leaf &&
2892                             btrfs_header_nritems(leaf)) {
2893                                 wret = push_leaf_right(trans, root, path, 1, 1);
2894                                 if (wret < 0 && wret != -ENOSPC)
2895                                         ret = wret;
2896                         }
2897
2898                         if (btrfs_header_nritems(leaf) == 0) {
2899                                 u64 root_gen;
2900                                 u64 bytenr = leaf->start;
2901                                 u32 blocksize = leaf->len;
2902
2903                                 root_gen = btrfs_header_generation(
2904                                                            path->nodes[1]);
2905
2906                                 wret = del_ptr(trans, root, path, 1, slot);
2907                                 if (wret)
2908                                         ret = wret;
2909
2910                                 free_extent_buffer(leaf);
2911                                 wret = btrfs_free_extent(trans, root, bytenr,
2912                                              blocksize,
2913                                              btrfs_header_owner(path->nodes[1]),
2914                                              root_gen, 0, 0, 1);
2915                                 if (wret)
2916                                         ret = wret;
2917                         } else {
2918                                 /* if we're still in the path, make sure
2919                                  * we're dirty.  Otherwise, one of the
2920                                  * push_leaf functions must have already
2921                                  * dirtied this buffer
2922                                  */
2923                                 if (path->nodes[0] == leaf)
2924                                         btrfs_mark_buffer_dirty(leaf);
2925                                 free_extent_buffer(leaf);
2926                         }
2927                 } else {
2928                         btrfs_mark_buffer_dirty(leaf);
2929                 }
2930         }
2931         return ret;
2932 }
2933
2934 /*
2935  * search the tree again to find a leaf with lesser keys
2936  * returns 0 if it found something or 1 if there are no lesser leaves.
2937  * returns < 0 on io errors.
2938  */
2939 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2940 {
2941         struct btrfs_key key;
2942         struct btrfs_disk_key found_key;
2943         int ret;
2944
2945         btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
2946
2947         if (key.offset > 0)
2948                 key.offset--;
2949         else if (key.type > 0)
2950                 key.type--;
2951         else if (key.objectid > 0)
2952                 key.objectid--;
2953         else
2954                 return 1;
2955
2956         btrfs_release_path(root, path);
2957         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2958         if (ret < 0)
2959                 return ret;
2960         btrfs_item_key(path->nodes[0], &found_key, 0);
2961         ret = comp_keys(&found_key, &key);
2962         if (ret < 0)
2963                 return 0;
2964         return 1;
2965 }
2966
2967 /*
2968  * A helper function to walk down the tree starting at min_key, and looking
2969  * for nodes or leaves that are either in cache or have a minimum
2970  * transaction id.  This is used by the btree defrag code, but could
2971  * also be used to search for blocks that have changed since a given
2972  * transaction id.
2973  *
2974  * This does not cow, but it does stuff the starting key it finds back
2975  * into min_key, so you can call btrfs_search_slot with cow=1 on the
2976  * key and get a writable path.
2977  *
2978  * This does lock as it descends, and path->keep_locks should be set
2979  * to 1 by the caller.
2980  *
2981  * This honors path->lowest_level to prevent descent past a given level
2982  * of the tree.
2983  *
2984  * returns zero if something useful was found, < 0 on error and 1 if there
2985  * was nothing in the tree that matched the search criteria.
2986  */
2987 int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
2988                          struct btrfs_path *path, int cache_only,
2989                          u64 min_trans)
2990 {
2991         struct extent_buffer *cur;
2992         struct btrfs_key found_key;
2993         int slot;
2994         u32 nritems;
2995         int level;
2996         int ret = 1;
2997
2998 again:
2999         cur = btrfs_lock_root_node(root);
3000         level = btrfs_header_level(cur);
3001         path->nodes[level] = cur;
3002         path->locks[level] = 1;
3003
3004         if (btrfs_header_generation(cur) < min_trans) {
3005                 ret = 1;
3006                 goto out;
3007         }
3008         while(1) {
3009                 nritems = btrfs_header_nritems(cur);
3010                 level = btrfs_header_level(cur);
3011                 bin_search(cur, min_key, level, &slot);
3012
3013                 /* at level = 0, we're done, setup the path and exit */
3014                 if (level == 0) {
3015                         ret = 0;
3016                         path->slots[level] = slot;
3017                         btrfs_item_key_to_cpu(cur, &found_key, slot);
3018                         goto out;
3019                 }
3020                 /*
3021                  * check this node pointer against the cache_only and
3022                  * min_trans parameters.  If it isn't in cache or is too
3023                  * old, skip to the next one.
3024                  */
3025                 while(slot < nritems) {
3026                         u64 blockptr;
3027                         u64 gen;
3028                         struct extent_buffer *tmp;
3029                         blockptr = btrfs_node_blockptr(cur, slot);
3030                         gen = btrfs_node_ptr_generation(cur, slot);
3031                         if (gen < min_trans) {
3032                                 slot++;
3033                                 continue;
3034                         }
3035                         if (!cache_only)
3036                                 break;
3037
3038                         tmp = btrfs_find_tree_block(root, blockptr,
3039                                             btrfs_level_size(root, level - 1));
3040
3041                         if (tmp && btrfs_buffer_uptodate(tmp, gen)) {
3042                                 free_extent_buffer(tmp);
3043                                 break;
3044                         }
3045                         if (tmp)
3046                                 free_extent_buffer(tmp);
3047                         slot++;
3048                 }
3049                 /*
3050                  * we didn't find a candidate key in this node, walk forward
3051                  * and find another one
3052                  */
3053                 if (slot >= nritems) {
3054                         ret = btrfs_find_next_key(root, path, min_key, level,
3055                                                   cache_only, min_trans);
3056                         if (ret == 0) {
3057                                 btrfs_release_path(root, path);
3058                                 goto again;
3059                         } else {
3060                                 goto out;
3061                         }
3062                 }
3063                 /* save our key for returning back */
3064                 btrfs_node_key_to_cpu(cur, &found_key, slot);
3065                 path->slots[level] = slot;
3066                 if (level == path->lowest_level) {
3067                         ret = 0;
3068                         unlock_up(path, level, 1);
3069                         goto out;
3070                 }
3071                 cur = read_node_slot(root, cur, slot);
3072
3073                 btrfs_tree_lock(cur);
3074                 path->locks[level - 1] = 1;
3075                 path->nodes[level - 1] = cur;
3076                 unlock_up(path, level, 1);
3077         }
3078 out:
3079         if (ret == 0)
3080                 memcpy(min_key, &found_key, sizeof(found_key));
3081         return ret;
3082 }
3083
3084 /*
3085  * this is similar to btrfs_next_leaf, but does not try to preserve
3086  * and fixup the path.  It looks for and returns the next key in the
3087  * tree based on the current path and the cache_only and min_trans
3088  * parameters.
3089  *
3090  * 0 is returned if another key is found, < 0 if there are any errors
3091  * and 1 is returned if there are no higher keys in the tree
3092  *
3093  * path->keep_locks should be set to 1 on the search made before
3094  * calling this function.
3095  */
3096 int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
3097                         struct btrfs_key *key, int lowest_level,
3098                         int cache_only, u64 min_trans)
3099 {
3100         int level = lowest_level;
3101         int slot;
3102         struct extent_buffer *c;
3103
3104         while(level < BTRFS_MAX_LEVEL) {
3105                 if (!path->nodes[level])
3106                         return 1;
3107
3108                 slot = path->slots[level] + 1;
3109                 c = path->nodes[level];
3110 next:
3111                 if (slot >= btrfs_header_nritems(c)) {
3112                         level++;
3113                         if (level == BTRFS_MAX_LEVEL) {
3114                                 return 1;
3115                         }
3116                         continue;
3117                 }
3118                 if (level == 0)
3119                         btrfs_item_key_to_cpu(c, key, slot);
3120                 else {
3121                         u64 blockptr = btrfs_node_blockptr(c, slot);
3122                         u64 gen = btrfs_node_ptr_generation(c, slot);
3123
3124                         if (cache_only) {
3125                                 struct extent_buffer *cur;
3126                                 cur = btrfs_find_tree_block(root, blockptr,
3127                                             btrfs_level_size(root, level - 1));
3128                                 if (!cur || !btrfs_buffer_uptodate(cur, gen)) {
3129                                         slot++;
3130                                         if (cur)
3131                                                 free_extent_buffer(cur);
3132                                         goto next;
3133                                 }
3134                                 free_extent_buffer(cur);
3135                         }
3136                         if (gen < min_trans) {
3137                                 slot++;
3138                                 goto next;
3139                         }
3140                         btrfs_node_key_to_cpu(c, key, slot);
3141                 }
3142                 return 0;
3143         }
3144         return 1;
3145 }
3146
3147 /*
3148  * search the tree again to find a leaf with greater keys
3149  * returns 0 if it found something or 1 if there are no greater leaves.
3150  * returns < 0 on io errors.
3151  */
3152 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
3153 {
3154         int slot;
3155         int level = 1;
3156         struct extent_buffer *c;
3157         struct extent_buffer *next = NULL;
3158         struct btrfs_key key;
3159         u32 nritems;
3160         int ret;
3161
3162         nritems = btrfs_header_nritems(path->nodes[0]);
3163         if (nritems == 0) {
3164                 return 1;
3165         }
3166
3167         btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
3168
3169         btrfs_release_path(root, path);
3170         path->keep_locks = 1;
3171         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3172         path->keep_locks = 0;
3173
3174         if (ret < 0)
3175                 return ret;
3176
3177         nritems = btrfs_header_nritems(path->nodes[0]);
3178         /*
3179          * by releasing the path above we dropped all our locks.  A balance
3180          * could have added more items next to the key that used to be
3181          * at the very end of the block.  So, check again here and
3182          * advance the path if there are now more items available.
3183          */
3184         if (nritems > 0 && path->slots[0] < nritems - 1) {
3185                 path->slots[0]++;
3186                 goto done;
3187         }
3188
3189         while(level < BTRFS_MAX_LEVEL) {
3190                 if (!path->nodes[level])
3191                         return 1;
3192
3193                 slot = path->slots[level] + 1;
3194                 c = path->nodes[level];
3195                 if (slot >= btrfs_header_nritems(c)) {
3196                         level++;
3197                         if (level == BTRFS_MAX_LEVEL) {
3198                                 return 1;
3199                         }
3200                         continue;
3201                 }
3202
3203                 if (next) {
3204                         btrfs_tree_unlock(next);
3205                         free_extent_buffer(next);
3206                 }
3207
3208                 if (level == 1 && path->locks[1] && path->reada)
3209                         reada_for_search(root, path, level, slot, 0);
3210
3211                 next = read_node_slot(root, c, slot);
3212                 if (!path->skip_locking) {
3213                         WARN_ON(!btrfs_tree_locked(c));
3214                         btrfs_tree_lock(next);
3215                 }
3216                 break;
3217         }
3218         path->slots[level] = slot;
3219         while(1) {
3220                 level--;
3221                 c = path->nodes[level];
3222                 if (path->locks[level])
3223                         btrfs_tree_unlock(c);
3224                 free_extent_buffer(c);
3225                 path->nodes[level] = next;
3226                 path->slots[level] = 0;
3227                 if (!path->skip_locking)
3228                         path->locks[level] = 1;
3229                 if (!level)
3230                         break;
3231                 if (level == 1 && path->locks[1] && path->reada)
3232                         reada_for_search(root, path, level, slot, 0);
3233                 next = read_node_slot(root, next, 0);
3234                 if (!path->skip_locking) {
3235                         WARN_ON(!btrfs_tree_locked(path->nodes[level]));
3236                         btrfs_tree_lock(next);
3237                 }
3238         }
3239 done:
3240         unlock_up(path, 0, 1);
3241         return 0;
3242 }
3243
3244 /*
3245  * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
3246  * searching until it gets past min_objectid or finds an item of 'type'
3247  *
3248  * returns 0 if something is found, 1 if nothing was found and < 0 on error
3249  */
3250 int btrfs_previous_item(struct btrfs_root *root,
3251                         struct btrfs_path *path, u64 min_objectid,
3252                         int type)
3253 {
3254         struct btrfs_key found_key;
3255         struct extent_buffer *leaf;
3256         int ret;
3257
3258         while(1) {
3259                 if (path->slots[0] == 0) {
3260                         ret = btrfs_prev_leaf(root, path);
3261                         if (ret != 0)
3262                                 return ret;
3263                 } else {
3264                         path->slots[0]--;
3265                 }
3266                 leaf = path->nodes[0];
3267                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3268                 if (found_key.type == type)
3269                         return 0;
3270         }
3271         return 1;
3272 }
3273