fs/btrfs/ctree.c

   1 /*
   2  * Copyright (C) 2007,2008 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 <linux/slab.h>
  21 #include "ctree.h"
  22 #include "disk-io.h"
  23 #include "transaction.h"
  24 #include "print-tree.h"
  25 #include "locking.h"
  26
  27 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  28                       *root, struct btrfs_path *path, int level);
  29 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  30                       *root, struct btrfs_key *ins_key,
  31                       struct btrfs_path *path, int data_size, int extend);
  32 static int push_node_left(struct btrfs_trans_handle *trans,
  33                           struct btrfs_root *root, struct extent_buffer *dst,
  34                           struct extent_buffer *src, int empty);
  35 static int balance_node_right(struct btrfs_trans_handle *trans,
  36                               struct btrfs_root *root,
  37                               struct extent_buffer *dst_buf,
  38                               struct extent_buffer *src_buf);
  39 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  40                    struct btrfs_path *path, int level, int slot);
  41 static int setup_items_for_insert(struct btrfs_trans_handle *trans,
  42                         struct btrfs_root *root, struct btrfs_path *path,
  43                         struct btrfs_key *cpu_key, u32 *data_size,
  44                         u32 total_data, u32 total_size, int nr);
  45
  46
  47 struct btrfs_path *btrfs_alloc_path(void)
  48 {
  49         struct btrfs_path *path;
  50         path = kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
  51         if (path)
  52                 path->reada = 1;
  53         return path;
  54 }
  55
  56 /*
  57  * set all locked nodes in the path to blocking locks.  This should
  58  * be done before scheduling
  59  */
  60 noinline void btrfs_set_path_blocking(struct btrfs_path *p)
  61 {
  62         int i;
  63         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  64                 if (p->nodes[i] && p->locks[i])
  65                         btrfs_set_lock_blocking(p->nodes[i]);
  66         }
  67 }
  68
  69 /*
  70  * reset all the locked nodes in the patch to spinning locks.
  71  *
  72  * held is used to keep lockdep happy, when lockdep is enabled
  73  * we set held to a blocking lock before we go around and
  74  * retake all the spinlocks in the path.  You can safely use NULL
  75  * for held
  76  */
  77 noinline void btrfs_clear_path_blocking(struct btrfs_path *p,
  78                                         struct extent_buffer *held)
  79 {
  80         int i;
  81
  82 #ifdef CONFIG_DEBUG_LOCK_ALLOC
  83         /* lockdep really cares that we take all of these spinlocks
  84          * in the right order.  If any of the locks in the path are not
  85          * currently blocking, it is going to complain.  So, make really
  86          * really sure by forcing the path to blocking before we clear
  87          * the path blocking.
  88          */
  89         if (held)
  90                 btrfs_set_lock_blocking(held);
  91         btrfs_set_path_blocking(p);
  92 #endif
  93
  94         for (i = BTRFS_MAX_LEVEL - 1; i >= 0; i--) {
  95                 if (p->nodes[i] && p->locks[i])
  96                         btrfs_clear_lock_blocking(p->nodes[i]);
  97         }
  98
  99 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 100         if (held)
 101                 btrfs_clear_lock_blocking(held);
 102 #endif
 103 }
 104
 105 /* this also releases the path */
 106 void btrfs_free_path(struct btrfs_path *p)
 107 {
 108         if (!p)
 109                 return;
 110         btrfs_release_path(NULL, p);
 111         kmem_cache_free(btrfs_path_cachep, p);
 112 }
 113
 114 /*
 115  * path release drops references on the extent buffers in the path
 116  * and it drops any locks held by this path
 117  *
 118  * It is safe to call this on paths that no locks or extent buffers held.
 119  */
 120 noinline void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
 121 {
 122         int i;
 123
 124         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
 125                 p->slots[i] = 0;
 126                 if (!p->nodes[i])
 127                         continue;
 128                 if (p->locks[i]) {
 129                         btrfs_tree_unlock(p->nodes[i]);
 130                         p->locks[i] = 0;
 131                 }
 132                 free_extent_buffer(p->nodes[i]);
 133                 p->nodes[i] = NULL;
 134         }
 135 }
 136
 137 /*
 138  * safely gets a reference on the root node of a tree.  A lock
 139  * is not taken, so a concurrent writer may put a different node
 140  * at the root of the tree.  See btrfs_lock_root_node for the
 141  * looping required.
 142  *
 143  * The extent buffer returned by this has a reference taken, so
 144  * it won't disappear.  It may stop being the root of the tree
 145  * at any time because there are no locks held.
 146  */
 147 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
 148 {
 149         struct extent_buffer *eb;
 150         spin_lock(&root->node_lock);
 151         eb = root->node;
 152         extent_buffer_get(eb);
 153         spin_unlock(&root->node_lock);
 154         return eb;
 155 }
 156
 157 /* loop around taking references on and locking the root node of the
 158  * tree until you end up with a lock on the root.  A locked buffer
 159  * is returned, with a reference held.
 160  */
 161 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
 162 {
 163         struct extent_buffer *eb;
 164
 165         while (1) {
 166                 eb = btrfs_root_node(root);
 167                 btrfs_tree_lock(eb);
 168
 169                 spin_lock(&root->node_lock);
 170                 if (eb == root->node) {
 171                         spin_unlock(&root->node_lock);
 172                         break;
 173                 }
 174                 spin_unlock(&root->node_lock);
 175
 176                 btrfs_tree_unlock(eb);
 177                 free_extent_buffer(eb);
 178         }
 179         return eb;
 180 }
 181
 182 /* cowonly root (everything not a reference counted cow subvolume), just get
 183  * put onto a simple dirty list.  transaction.c walks this to make sure they
 184  * get properly updated on disk.
 185  */
 186 static void add_root_to_dirty_list(struct btrfs_root *root)
 187 {
 188         if (root->track_dirty && list_empty(&root->dirty_list)) {
 189                 list_add(&root->dirty_list,
 190                          &root->fs_info->dirty_cowonly_roots);
 191         }
 192 }
 193
 194 /*
 195  * used by snapshot creation to make a copy of a root for a tree with
 196  * a given objectid.  The buffer with the new root node is returned in
 197  * cow_ret, and this func returns zero on success or a negative error code.
 198  */
 199 int btrfs_copy_root(struct btrfs_trans_handle *trans,
 200                       struct btrfs_root *root,
 201                       struct extent_buffer *buf,
 202                       struct extent_buffer **cow_ret, u64 new_root_objectid)
 203 {
 204         struct extent_buffer *cow;
 205         int ret = 0;
 206         int level;
 207         struct btrfs_disk_key disk_key;
 208
 209         WARN_ON(root->ref_cows && trans->transid !=
 210                 root->fs_info->running_transaction->transid);
 211         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 212
 213         level = btrfs_header_level(buf);
 214         if (level == 0)
 215                 btrfs_item_key(buf, &disk_key, 0);
 216         else
 217                 btrfs_node_key(buf, &disk_key, 0);
 218
 219         cow = btrfs_alloc_free_block(trans, root, buf->len, 0,
 220                                      new_root_objectid, &disk_key, level,
 221                                      buf->start, 0);
 222         if (IS_ERR(cow))
 223                 return PTR_ERR(cow);
 224
 225         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 226         btrfs_set_header_bytenr(cow, cow->start);
 227         btrfs_set_header_generation(cow, trans->transid);
 228         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 229         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 230                                      BTRFS_HEADER_FLAG_RELOC);
 231         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 232                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 233         else
 234                 btrfs_set_header_owner(cow, new_root_objectid);
 235
 236         write_extent_buffer(cow, root->fs_info->fsid,
 237                             (unsigned long)btrfs_header_fsid(cow),
 238                             BTRFS_FSID_SIZE);
 239
 240         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 241         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 242                 ret = btrfs_inc_ref(trans, root, cow, 1);
 243         else
 244                 ret = btrfs_inc_ref(trans, root, cow, 0);
 245
 246         if (ret)
 247                 return ret;
 248
 249         btrfs_mark_buffer_dirty(cow);
 250         *cow_ret = cow;
 251         return 0;
 252 }
 253
 254 /*
 255  * check if the tree block can be shared by multiple trees
 256  */
 257 int btrfs_block_can_be_shared(struct btrfs_root *root,
 258                               struct extent_buffer *buf)
 259 {
 260         /*
 261          * Tree blocks not in refernece counted trees and tree roots
 262          * are never shared. If a block was allocated after the last
 263          * snapshot and the block was not allocated by tree relocation,
 264          * we know the block is not shared.
 265          */
 266         if (root->ref_cows &&
 267             buf != root->node && buf != root->commit_root &&
 268             (btrfs_header_generation(buf) <=
 269              btrfs_root_last_snapshot(&root->root_item) ||
 270              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 271                 return 1;
 272 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 273         if (root->ref_cows &&
 274             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 275                 return 1;
 276 #endif
 277         return 0;
 278 }
 279
 280 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
 281                                        struct btrfs_root *root,
 282                                        struct extent_buffer *buf,
 283                                        struct extent_buffer *cow,
 284                                        int *last_ref)
 285 {
 286         u64 refs;
 287         u64 owner;
 288         u64 flags;
 289         u64 new_flags = 0;
 290         int ret;
 291
 292         /*
 293          * Backrefs update rules:
 294          *
 295          * Always use full backrefs for extent pointers in tree block
 296          * allocated by tree relocation.
 297          *
 298          * If a shared tree block is no longer referenced by its owner
 299          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
 300          * use full backrefs for extent pointers in tree block.
 301          *
 302          * If a tree block is been relocating
 303          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
 304          * use full backrefs for extent pointers in tree block.
 305          * The reason for this is some operations (such as drop tree)
 306          * are only allowed for blocks use full backrefs.
 307          */
 308
 309         if (btrfs_block_can_be_shared(root, buf)) {
 310                 ret = btrfs_lookup_extent_info(trans, root, buf->start,
 311                                                buf->len, &refs, &flags);
 312                 BUG_ON(ret);
 313                 BUG_ON(refs == 0);
 314         } else {
 315                 refs = 1;
 316                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 317                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 318                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 319                 else
 320                         flags = 0;
 321         }
 322
 323         owner = btrfs_header_owner(buf);
 324         BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
 325                !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
 326
 327         if (refs > 1) {
 328                 if ((owner == root->root_key.objectid ||
 329                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
 330                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
 331                         ret = btrfs_inc_ref(trans, root, buf, 1);
 332                         BUG_ON(ret);
 333
 334                         if (root->root_key.objectid ==
 335                             BTRFS_TREE_RELOC_OBJECTID) {
 336                                 ret = btrfs_dec_ref(trans, root, buf, 0);
 337                                 BUG_ON(ret);
 338                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 339                                 BUG_ON(ret);
 340                         }
 341                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
 342                 } else {
 343
 344                         if (root->root_key.objectid ==
 345                             BTRFS_TREE_RELOC_OBJECTID)
 346                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 347                         else
 348                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 349                         BUG_ON(ret);
 350                 }
 351                 if (new_flags != 0) {
 352                         ret = btrfs_set_disk_extent_flags(trans, root,
 353                                                           buf->start,
 354                                                           buf->len,
 355                                                           new_flags, 0);
 356                         BUG_ON(ret);
 357                 }
 358         } else {
 359                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
 360                         if (root->root_key.objectid ==
 361                             BTRFS_TREE_RELOC_OBJECTID)
 362                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 363                         else
 364                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 365                         BUG_ON(ret);
 366                         ret = btrfs_dec_ref(trans, root, buf, 1);
 367                         BUG_ON(ret);
 368                 }
 369                 clean_tree_block(trans, root, buf);
 370                 *last_ref = 1;
 371         }
 372         return 0;
 373 }
 374
 375 /*
 376  * does the dirty work in cow of a single block.  The parent block (if
 377  * supplied) is updated to point to the new cow copy.  The new buffer is marked
 378  * dirty and returned locked.  If you modify the block it needs to be marked
 379  * dirty again.
 380  *
 381  * search_start -- an allocation hint for the new block
 382  *
 383  * empty_size -- a hint that you plan on doing more cow.  This is the size in
 384  * bytes the allocator should try to find free next to the block it returns.
 385  * This is just a hint and may be ignored by the allocator.
 386  */
 387 static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 388                              struct btrfs_root *root,
 389                              struct extent_buffer *buf,
 390                              struct extent_buffer *parent, int parent_slot,
 391                              struct extent_buffer **cow_ret,
 392                              u64 search_start, u64 empty_size)
 393 {
 394         struct btrfs_disk_key disk_key;
 395         struct extent_buffer *cow;
 396         int level;
 397         int last_ref = 0;
 398         int unlock_orig = 0;
 399         u64 parent_start;
 400
 401         if (*cow_ret == buf)
 402                 unlock_orig = 1;
 403
 404         btrfs_assert_tree_locked(buf);
 405
 406         WARN_ON(root->ref_cows && trans->transid !=
 407                 root->fs_info->running_transaction->transid);
 408         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 409
 410         level = btrfs_header_level(buf);
 411
 412         if (level == 0)
 413                 btrfs_item_key(buf, &disk_key, 0);
 414         else
 415                 btrfs_node_key(buf, &disk_key, 0);
 416
 417         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
 418                 if (parent)
 419                         parent_start = parent->start;
 420                 else
 421                         parent_start = 0;
 422         } else
 423                 parent_start = 0;
 424
 425         cow = btrfs_alloc_free_block(trans, root, buf->len, parent_start,
 426                                      root->root_key.objectid, &disk_key,
 427                                      level, search_start, empty_size);
 428         if (IS_ERR(cow))
 429                 return PTR_ERR(cow);
 430
 431         /* cow is set to blocking by btrfs_init_new_buffer */
 432
 433         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 434         btrfs_set_header_bytenr(cow, cow->start);
 435         btrfs_set_header_generation(cow, trans->transid);
 436         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 437         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 438                                      BTRFS_HEADER_FLAG_RELOC);
 439         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 440                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 441         else
 442                 btrfs_set_header_owner(cow, root->root_key.objectid);
 443
 444         write_extent_buffer(cow, root->fs_info->fsid,
 445                             (unsigned long)btrfs_header_fsid(cow),
 446                             BTRFS_FSID_SIZE);
 447
 448         update_ref_for_cow(trans, root, buf, cow, &last_ref);
 449
 450         if (root->ref_cows)
 451                 btrfs_reloc_cow_block(trans, root, buf, cow);
 452
 453         if (buf == root->node) {
 454                 WARN_ON(parent && parent != buf);
 455                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 456                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 457                         parent_start = buf->start;
 458                 else
 459                         parent_start = 0;
 460
 461                 spin_lock(&root->node_lock);
 462                 root->node = cow;
 463                 extent_buffer_get(cow);
 464                 spin_unlock(&root->node_lock);
 465
 466                 btrfs_free_tree_block(trans, root, buf, parent_start,
 467                                       last_ref);
 468                 free_extent_buffer(buf);
 469                 add_root_to_dirty_list(root);
 470         } else {
 471                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 472                         parent_start = parent->start;
 473                 else
 474                         parent_start = 0;
 475
 476                 WARN_ON(trans->transid != btrfs_header_generation(parent));
 477                 btrfs_set_node_blockptr(parent, parent_slot,
 478                                         cow->start);
 479                 btrfs_set_node_ptr_generation(parent, parent_slot,
 480                                               trans->transid);
 481                 btrfs_mark_buffer_dirty(parent);
 482                 btrfs_free_tree_block(trans, root, buf, parent_start,
 483                                       last_ref);
 484         }
 485         if (unlock_orig)
 486                 btrfs_tree_unlock(buf);
 487         free_extent_buffer(buf);
 488         btrfs_mark_buffer_dirty(cow);
 489         *cow_ret = cow;
 490         return 0;
 491 }
 492
 493 static inline int should_cow_block(struct btrfs_trans_handle *trans,
 494                                    struct btrfs_root *root,
 495                                    struct extent_buffer *buf)
 496 {
 497         if (btrfs_header_generation(buf) == trans->transid &&
 498             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
 499             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
 500               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 501                 return 0;
 502         return 1;
 503 }
 504
 505 /*
 506  * cows a single block, see __btrfs_cow_block for the real work.
 507  * This version of it has extra checks so that a block isn't cow'd more than
 508  * once per transaction, as long as it hasn't been written yet
 509  */
 510 noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
 511                     struct btrfs_root *root, struct extent_buffer *buf,
 512                     struct extent_buffer *parent, int parent_slot,
 513                     struct extent_buffer **cow_ret)
 514 {
 515         u64 search_start;
 516         int ret;
 517
 518         if (trans->transaction != root->fs_info->running_transaction) {
 519                 printk(KERN_CRIT "trans %llu running %llu\n",
 520                        (unsigned long long)trans->transid,
 521                        (unsigned long long)
 522                        root->fs_info->running_transaction->transid);
 523                 WARN_ON(1);
 524         }
 525         if (trans->transid != root->fs_info->generation) {
 526                 printk(KERN_CRIT "trans %llu running %llu\n",
 527                        (unsigned long long)trans->transid,
 528                        (unsigned long long)root->fs_info->generation);
 529                 WARN_ON(1);
 530         }
 531
 532         if (!should_cow_block(trans, root, buf)) {
 533                 *cow_ret = buf;
 534                 return 0;
 535         }
 536
 537         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
 538
 539         if (parent)
 540                 btrfs_set_lock_blocking(parent);
 541         btrfs_set_lock_blocking(buf);
 542
 543         ret = __btrfs_cow_block(trans, root, buf, parent,
 544                                  parent_slot, cow_ret, search_start, 0);
 545         return ret;
 546 }
 547
 548 /*
 549  * helper function for defrag to decide if two blocks pointed to by a
 550  * node are actually close by
 551  */
 552 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 553 {
 554         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 555                 return 1;
 556         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 557                 return 1;
 558         return 0;
 559 }
 560
 561 /*
 562  * compare two keys in a memcmp fashion
 563  */
 564 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 565 {
 566         struct btrfs_key k1;
 567
 568         btrfs_disk_key_to_cpu(&k1, disk);
 569
 570         return btrfs_comp_cpu_keys(&k1, k2);
 571 }
 572
 573 /*
 574  * same as comp_keys only with two btrfs_key's
 575  */
 576 int btrfs_comp_cpu_keys(struct btrfs_key *k1, struct btrfs_key *k2)
 577 {
 578         if (k1->objectid > k2->objectid)
 579                 return 1;
 580         if (k1->objectid < k2->objectid)
 581                 return -1;
 582         if (k1->type > k2->type)
 583                 return 1;
 584         if (k1->type < k2->type)
 585                 return -1;
 586         if (k1->offset > k2->offset)
 587                 return 1;
 588         if (k1->offset < k2->offset)
 589                 return -1;
 590         return 0;
 591 }
 592
 593 /*
 594  * this is used by the defrag code to go through all the
 595  * leaves pointed to by a node and reallocate them so that
 596  * disk order is close to key order
 597  */
 598 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 599                        struct btrfs_root *root, struct extent_buffer *parent,
 600                        int start_slot, int cache_only, u64 *last_ret,
 601                        struct btrfs_key *progress)
 602 {
 603         struct extent_buffer *cur;
 604         u64 blocknr;
 605         u64 gen;
 606         u64 search_start = *last_ret;
 607         u64 last_block = 0;
 608         u64 other;
 609         u32 parent_nritems;
 610         int end_slot;
 611         int i;
 612         int err = 0;
 613         int parent_level;
 614         int uptodate;
 615         u32 blocksize;
 616         int progress_passed = 0;
 617         struct btrfs_disk_key disk_key;
 618
 619         parent_level = btrfs_header_level(parent);
 620         if (cache_only && parent_level != 1)
 621                 return 0;
 622
 623         if (trans->transaction != root->fs_info->running_transaction)
 624                 WARN_ON(1);
 625         if (trans->transid != root->fs_info->generation)
 626                 WARN_ON(1);
 627
 628         parent_nritems = btrfs_header_nritems(parent);
 629         blocksize = btrfs_level_size(root, parent_level - 1);
 630         end_slot = parent_nritems;
 631
 632         if (parent_nritems == 1)
 633                 return 0;
 634
 635         btrfs_set_lock_blocking(parent);
 636
 637         for (i = start_slot; i < end_slot; i++) {
 638                 int close = 1;
 639
 640                 if (!parent->map_token) {
 641                         map_extent_buffer(parent,
 642                                         btrfs_node_key_ptr_offset(i),
 643                                         sizeof(struct btrfs_key_ptr),
 644                                         &parent->map_token, &parent->kaddr,
 645                                         &parent->map_start, &parent->map_len,
 646                                         KM_USER1);
 647                 }
 648                 btrfs_node_key(parent, &disk_key, i);
 649                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
 650                         continue;
 651
 652                 progress_passed = 1;
 653                 blocknr = btrfs_node_blockptr(parent, i);
 654                 gen = btrfs_node_ptr_generation(parent, i);
 655                 if (last_block == 0)
 656                         last_block = blocknr;
 657
 658                 if (i > 0) {
 659                         other = btrfs_node_blockptr(parent, i - 1);
 660                         close = close_blocks(blocknr, other, blocksize);
 661                 }
 662                 if (!close && i < end_slot - 2) {
 663                         other = btrfs_node_blockptr(parent, i + 1);
 664                         close = close_blocks(blocknr, other, blocksize);
 665                 }
 666                 if (close) {
 667                         last_block = blocknr;
 668                         continue;
 669                 }
 670                 if (parent->map_token) {
 671                         unmap_extent_buffer(parent, parent->map_token,
 672                                             KM_USER1);
 673                         parent->map_token = NULL;
 674                 }
 675
 676                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
 677                 if (cur)
 678                         uptodate = btrfs_buffer_uptodate(cur, gen);
 679                 else
 680                         uptodate = 0;
 681                 if (!cur || !uptodate) {
 682                         if (cache_only) {
 683                                 free_extent_buffer(cur);
 684                                 continue;
 685                         }
 686                         if (!cur) {
 687                                 cur = read_tree_block(root, blocknr,
 688                                                          blocksize, gen);
 689                         } else if (!uptodate) {
 690                                 btrfs_read_buffer(cur, gen);
 691                         }
 692                 }
 693                 if (search_start == 0)
 694                         search_start = last_block;
 695
 696                 btrfs_tree_lock(cur);
 697                 btrfs_set_lock_blocking(cur);
 698                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 699                                         &cur, search_start,
 700                                         min(16 * blocksize,
 701                                             (end_slot - i) * blocksize));
 702                 if (err) {
 703                         btrfs_tree_unlock(cur);
 704                         free_extent_buffer(cur);
 705                         break;
 706                 }
 707                 search_start = cur->start;
 708                 last_block = cur->start;
 709                 *last_ret = search_start;
 710                 btrfs_tree_unlock(cur);
 711                 free_extent_buffer(cur);
 712         }
 713         if (parent->map_token) {
 714                 unmap_extent_buffer(parent, parent->map_token,
 715                                     KM_USER1);
 716                 parent->map_token = NULL;
 717         }
 718         return err;
 719 }
 720
 721 /*
 722  * The leaf data grows from end-to-front in the node.
 723  * this returns the address of the start of the last item,
 724  * which is the stop of the leaf data stack
 725  */
 726 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 727                                          struct extent_buffer *leaf)
 728 {
 729         u32 nr = btrfs_header_nritems(leaf);
 730         if (nr == 0)
 731                 return BTRFS_LEAF_DATA_SIZE(root);
 732         return btrfs_item_offset_nr(leaf, nr - 1);
 733 }
 734
 735 /*
 736  * extra debugging checks to make sure all the items in a key are
 737  * well formed and in the proper order
 738  */
 739 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
 740                       int level)
 741 {
 742         struct extent_buffer *parent = NULL;
 743         struct extent_buffer *node = path->nodes[level];
 744         struct btrfs_disk_key parent_key;
 745         struct btrfs_disk_key node_key;
 746         int parent_slot;
 747         int slot;
 748         struct btrfs_key cpukey;
 749         u32 nritems = btrfs_header_nritems(node);
 750
 751         if (path->nodes[level + 1])
 752                 parent = path->nodes[level + 1];
 753
 754         slot = path->slots[level];
 755         BUG_ON(nritems == 0);
 756         if (parent) {
 757                 parent_slot = path->slots[level + 1];
 758                 btrfs_node_key(parent, &parent_key, parent_slot);
 759                 btrfs_node_key(node, &node_key, 0);
 760                 BUG_ON(memcmp(&parent_key, &node_key,
 761                               sizeof(struct btrfs_disk_key)));
 762                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 763                        btrfs_header_bytenr(node));
 764         }
 765         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 766         if (slot != 0) {
 767                 btrfs_node_key_to_cpu(node, &cpukey, slot - 1);
 768                 btrfs_node_key(node, &node_key, slot);
 769                 BUG_ON(comp_keys(&node_key, &cpukey) <= 0);
 770         }
 771         if (slot < nritems - 1) {
 772                 btrfs_node_key_to_cpu(node, &cpukey, slot + 1);
 773                 btrfs_node_key(node, &node_key, slot);
 774                 BUG_ON(comp_keys(&node_key, &cpukey) >= 0);
 775         }
 776         return 0;
 777 }
 778
 779 /*
 780  * extra checking to make sure all the items in a leaf are
 781  * well formed and in the proper order
 782  */
 783 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 784                       int level)
 785 {
 786         struct extent_buffer *leaf = path->nodes[level];
 787         struct extent_buffer *parent = NULL;
 788         int parent_slot;
 789         struct btrfs_key cpukey;
 790         struct btrfs_disk_key parent_key;
 791         struct btrfs_disk_key leaf_key;
 792         int slot = path->slots[0];
 793
 794         u32 nritems = btrfs_header_nritems(leaf);
 795
 796         if (path->nodes[level + 1])
 797                 parent = path->nodes[level + 1];
 798
 799         if (nritems == 0)
 800                 return 0;
 801
 802         if (parent) {
 803                 parent_slot = path->slots[level + 1];
 804                 btrfs_node_key(parent, &parent_key, parent_slot);
 805                 btrfs_item_key(leaf, &leaf_key, 0);
 806
 807                 BUG_ON(memcmp(&parent_key, &leaf_key,
 808                        sizeof(struct btrfs_disk_key)));
 809                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 810                        btrfs_header_bytenr(leaf));
 811         }
 812         if (slot != 0 && slot < nritems - 1) {
 813                 btrfs_item_key(leaf, &leaf_key, slot);
 814                 btrfs_item_key_to_cpu(leaf, &cpukey, slot - 1);
 815                 if (comp_keys(&leaf_key, &cpukey) <= 0) {
 816                         btrfs_print_leaf(root, leaf);
 817                         printk(KERN_CRIT "slot %d offset bad key\n", slot);
 818                         BUG_ON(1);
 819                 }
 820                 if (btrfs_item_offset_nr(leaf, slot - 1) !=
 821                        btrfs_item_end_nr(leaf, slot)) {
 822                         btrfs_print_leaf(root, leaf);
 823                         printk(KERN_CRIT "slot %d offset bad\n", slot);
 824                         BUG_ON(1);
 825                 }
 826         }
 827         if (slot < nritems - 1) {
 828                 btrfs_item_key(leaf, &leaf_key, slot);
 829                 btrfs_item_key_to_cpu(leaf, &cpukey, slot + 1);
 830                 BUG_ON(comp_keys(&leaf_key, &cpukey) >= 0);
 831                 if (btrfs_item_offset_nr(leaf, slot) !=
 832                         btrfs_item_end_nr(leaf, slot + 1)) {
 833                         btrfs_print_leaf(root, leaf);
 834                         printk(KERN_CRIT "slot %d offset bad\n", slot);
 835                         BUG_ON(1);
 836                 }
 837         }
 838         BUG_ON(btrfs_item_offset_nr(leaf, 0) +
 839                btrfs_item_size_nr(leaf, 0) != BTRFS_LEAF_DATA_SIZE(root));
 840         return 0;
 841 }
 842
 843 static noinline int check_block(struct btrfs_root *root,
 844                                 struct btrfs_path *path, int level)
 845 {
 846         return 0;
 847         if (level == 0)
 848                 return check_leaf(root, path, level);
 849         return check_node(root, path, level);
 850 }
 851
 852 /*
 853  * search for key in the extent_buffer.  The items start at offset p,
 854  * and they are item_size apart.  There are 'max' items in p.
 855  *
 856  * the slot in the array is returned via slot, and it points to
 857  * the place where you would insert key if it is not found in
 858  * the array.
 859  *
 860  * slot may point to max if the key is bigger than all of the keys
 861  */
 862 static noinline int generic_bin_search(struct extent_buffer *eb,
 863                                        unsigned long p,
 864                                        int item_size, struct btrfs_key *key,
 865                                        int max, int *slot)
 866 {
 867         int low = 0;
 868         int high = max;
 869         int mid;
 870         int ret;
 871         struct btrfs_disk_key *tmp = NULL;
 872         struct btrfs_disk_key unaligned;
 873         unsigned long offset;
 874         char *map_token = NULL;
 875         char *kaddr = NULL;
 876         unsigned long map_start = 0;
 877         unsigned long map_len = 0;
 878         int err;
 879
 880         while (low < high) {
 881                 mid = (low + high) / 2;
 882                 offset = p + mid * item_size;
 883
 884                 if (!map_token || offset < map_start ||
 885                     (offset + sizeof(struct btrfs_disk_key)) >
 886                     map_start + map_len) {
 887                         if (map_token) {
 888                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 889                                 map_token = NULL;
 890                         }
 891
 892                         err = map_private_extent_buffer(eb, offset,
 893                                                 sizeof(struct btrfs_disk_key),
 894                                                 &map_token, &kaddr,
 895                                                 &map_start, &map_len, KM_USER0);
 896
 897                         if (!err) {
 898                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
 899                                                         map_start);
 900                         } else {
 901                                 read_extent_buffer(eb, &unaligned,
 902                                                    offset, sizeof(unaligned));
 903                                 tmp = &unaligned;
 904                         }
 905
 906                 } else {
 907                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
 908                                                         map_start);
 909                 }
 910                 ret = comp_keys(tmp, key);
 911
 912                 if (ret < 0)
 913                         low = mid + 1;
 914                 else if (ret > 0)
 915                         high = mid;
 916                 else {
 917                         *slot = mid;
 918                         if (map_token)
 919                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 920                         return 0;
 921                 }
 922         }
 923         *slot = low;
 924         if (map_token)
 925                 unmap_extent_buffer(eb, map_token, KM_USER0);
 926         return 1;
 927 }
 928
 929 /*
 930  * simple bin_search frontend that does the right thing for
 931  * leaves vs nodes
 932  */
 933 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 934                       int level, int *slot)
 935 {
 936         if (level == 0) {
 937                 return generic_bin_search(eb,
 938                                           offsetof(struct btrfs_leaf, items),
 939                                           sizeof(struct btrfs_item),
 940                                           key, btrfs_header_nritems(eb),
 941                                           slot);
 942         } else {
 943                 return generic_bin_search(eb,
 944                                           offsetof(struct btrfs_node, ptrs),
 945                                           sizeof(struct btrfs_key_ptr),
 946                                           key, btrfs_header_nritems(eb),
 947                                           slot);
 948         }
 949         return -1;
 950 }
 951
 952 int btrfs_bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 953                      int level, int *slot)
 954 {
 955         return bin_search(eb, key, level, slot);
 956 }
 957
 958 static void root_add_used(struct btrfs_root *root, u32 size)
 959 {
 960         spin_lock(&root->accounting_lock);
 961         btrfs_set_root_used(&root->root_item,
 962                             btrfs_root_used(&root->root_item) + size);
 963         spin_unlock(&root->accounting_lock);
 964 }
 965
 966 static void root_sub_used(struct btrfs_root *root, u32 size)
 967 {
 968         spin_lock(&root->accounting_lock);
 969         btrfs_set_root_used(&root->root_item,
 970                             btrfs_root_used(&root->root_item) - size);
 971         spin_unlock(&root->accounting_lock);
 972 }
 973
 974 /* given a node and slot number, this reads the blocks it points to.  The
 975  * extent buffer is returned with a reference taken (but unlocked).
 976  * NULL is returned on error.
 977  */
 978 static noinline struct extent_buffer *read_node_slot(struct btrfs_root *root,
 979                                    struct extent_buffer *parent, int slot)
 980 {
 981         int level = btrfs_header_level(parent);
 982         if (slot < 0)
 983                 return NULL;
 984         if (slot >= btrfs_header_nritems(parent))
 985                 return NULL;
 986
 987         BUG_ON(level == 0);
 988
 989         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 990                        btrfs_level_size(root, level - 1),
 991                        btrfs_node_ptr_generation(parent, slot));
 992 }
 993
 994 /*
 995  * node level balancing, used to make sure nodes are in proper order for
 996  * item deletion.  We balance from the top down, so we have to make sure
 997  * that a deletion won't leave an node completely empty later on.
 998  */
 999 static noinline int balance_level(struct btrfs_trans_handle *trans,
1000                          struct btrfs_root *root,
1001                          struct btrfs_path *path, int level)
1002 {
1003         struct extent_buffer *right = NULL;
1004         struct extent_buffer *mid;
1005         struct extent_buffer *left = NULL;
1006         struct extent_buffer *parent = NULL;
1007         int ret = 0;
1008         int wret;
1009         int pslot;
1010         int orig_slot = path->slots[level];
1011         u64 orig_ptr;
1012
1013         if (level == 0)
1014                 return 0;
1015
1016         mid = path->nodes[level];
1017
1018         WARN_ON(!path->locks[level]);
1019         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1020
1021         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1022
1023         if (level < BTRFS_MAX_LEVEL - 1)
1024                 parent = path->nodes[level + 1];
1025         pslot = path->slots[level + 1];
1026
1027         /*
1028          * deal with the case where there is only one pointer in the root
1029          * by promoting the node below to a root
1030          */
1031         if (!parent) {
1032                 struct extent_buffer *child;
1033
1034                 if (btrfs_header_nritems(mid) != 1)
1035                         return 0;
1036
1037                 /* promote the child to a root */
1038                 child = read_node_slot(root, mid, 0);
1039                 BUG_ON(!child);
1040                 btrfs_tree_lock(child);
1041                 btrfs_set_lock_blocking(child);
1042                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
1043                 if (ret) {
1044                         btrfs_tree_unlock(child);
1045                         free_extent_buffer(child);
1046                         goto enospc;
1047                 }
1048
1049                 spin_lock(&root->node_lock);
1050                 root->node = child;
1051                 spin_unlock(&root->node_lock);
1052
1053                 add_root_to_dirty_list(root);
1054                 btrfs_tree_unlock(child);
1055
1056                 path->locks[level] = 0;
1057                 path->nodes[level] = NULL;
1058                 clean_tree_block(trans, root, mid);
1059                 btrfs_tree_unlock(mid);
1060                 /* once for the path */
1061                 free_extent_buffer(mid);
1062
1063                 root_sub_used(root, mid->len);
1064                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1065                 /* once for the root ptr */
1066                 free_extent_buffer(mid);
1067                 return 0;
1068         }
1069         if (btrfs_header_nritems(mid) >
1070             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
1071                 return 0;
1072
1073         btrfs_header_nritems(mid);
1074
1075         left = read_node_slot(root, parent, pslot - 1);
1076         if (left) {
1077                 btrfs_tree_lock(left);
1078                 btrfs_set_lock_blocking(left);
1079                 wret = btrfs_cow_block(trans, root, left,
1080                                        parent, pslot - 1, &left);
1081                 if (wret) {
1082                         ret = wret;
1083                         goto enospc;
1084                 }
1085         }
1086         right = read_node_slot(root, parent, pslot + 1);
1087         if (right) {
1088                 btrfs_tree_lock(right);
1089                 btrfs_set_lock_blocking(right);
1090                 wret = btrfs_cow_block(trans, root, right,
1091                                        parent, pslot + 1, &right);
1092                 if (wret) {
1093                         ret = wret;
1094                         goto enospc;
1095                 }
1096         }
1097
1098         /* first, try to make some room in the middle buffer */
1099         if (left) {
1100                 orig_slot += btrfs_header_nritems(left);
1101                 wret = push_node_left(trans, root, left, mid, 1);
1102                 if (wret < 0)
1103                         ret = wret;
1104                 btrfs_header_nritems(mid);
1105         }
1106
1107         /*
1108          * then try to empty the right most buffer into the middle
1109          */
1110         if (right) {
1111                 wret = push_node_left(trans, root, mid, right, 1);
1112                 if (wret < 0 && wret != -ENOSPC)
1113                         ret = wret;
1114                 if (btrfs_header_nritems(right) == 0) {
1115                         clean_tree_block(trans, root, right);
1116                         btrfs_tree_unlock(right);
1117                         wret = del_ptr(trans, root, path, level + 1, pslot +
1118                                        1);
1119                         if (wret)
1120                                 ret = wret;
1121                         root_sub_used(root, right->len);
1122                         btrfs_free_tree_block(trans, root, right, 0, 1);
1123                         free_extent_buffer(right);
1124                         right = NULL;
1125                 } else {
1126                         struct btrfs_disk_key right_key;
1127                         btrfs_node_key(right, &right_key, 0);
1128                         btrfs_set_node_key(parent, &right_key, pslot + 1);
1129                         btrfs_mark_buffer_dirty(parent);
1130                 }
1131         }
1132         if (btrfs_header_nritems(mid) == 1) {
1133                 /*
1134                  * we're not allowed to leave a node with one item in the
1135                  * tree during a delete.  A deletion from lower in the tree
1136                  * could try to delete the only pointer in this node.
1137                  * So, pull some keys from the left.
1138                  * There has to be a left pointer at this point because
1139                  * otherwise we would have pulled some pointers from the
1140                  * right
1141                  */
1142                 BUG_ON(!left);
1143                 wret = balance_node_right(trans, root, mid, left);
1144                 if (wret < 0) {
1145                         ret = wret;
1146                         goto enospc;
1147                 }
1148                 if (wret == 1) {
1149                         wret = push_node_left(trans, root, left, mid, 1);
1150                         if (wret < 0)
1151                                 ret = wret;
1152                 }
1153                 BUG_ON(wret == 1);
1154         }
1155         if (btrfs_header_nritems(mid) == 0) {
1156                 clean_tree_block(trans, root, mid);
1157                 btrfs_tree_unlock(mid);
1158                 wret = del_ptr(trans, root, path, level + 1, pslot);
1159                 if (wret)
1160                         ret = wret;
1161                 root_sub_used(root, mid->len);
1162                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1163                 free_extent_buffer(mid);
1164                 mid = NULL;
1165         } else {
1166                 /* update the parent key to reflect our changes */
1167                 struct btrfs_disk_key mid_key;
1168                 btrfs_node_key(mid, &mid_key, 0);
1169                 btrfs_set_node_key(parent, &mid_key, pslot);
1170                 btrfs_mark_buffer_dirty(parent);
1171         }
1172
1173         /* update the path */
1174         if (left) {
1175                 if (btrfs_header_nritems(left) > orig_slot) {
1176                         extent_buffer_get(left);
1177                         /* left was locked after cow */
1178                         path->nodes[level] = left;
1179                         path->slots[level + 1] -= 1;
1180                         path->slots[level] = orig_slot;
1181                         if (mid) {
1182                                 btrfs_tree_unlock(mid);
1183                                 free_extent_buffer(mid);
1184                         }
1185                 } else {
1186                         orig_slot -= btrfs_header_nritems(left);
1187                         path->slots[level] = orig_slot;
1188                 }
1189         }
1190         /* double check we haven't messed things up */
1191         check_block(root, path, level);
1192         if (orig_ptr !=
1193             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
1194                 BUG();
1195 enospc:
1196         if (right) {
1197                 btrfs_tree_unlock(right);
1198                 free_extent_buffer(right);
1199         }
1200         if (left) {
1201                 if (path->nodes[level] != left)
1202                         btrfs_tree_unlock(left);
1203                 free_extent_buffer(left);
1204         }
1205         return ret;
1206 }
1207
1208 /* Node balancing for insertion.  Here we only split or push nodes around
1209  * when they are completely full.  This is also done top down, so we
1210  * have to be pessimistic.
1211  */
1212 static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
1213                                           struct btrfs_root *root,
1214                                           struct btrfs_path *path, int level)
1215 {
1216         struct extent_buffer *right = NULL;
1217         struct extent_buffer *mid;
1218         struct extent_buffer *left = NULL;
1219         struct extent_buffer *parent = NULL;
1220         int ret = 0;
1221         int wret;
1222         int pslot;
1223         int orig_slot = path->slots[level];
1224
1225         if (level == 0)
1226                 return 1;
1227
1228         mid = path->nodes[level];
1229         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1230
1231         if (level < BTRFS_MAX_LEVEL - 1)
1232                 parent = path->nodes[level + 1];
1233         pslot = path->slots[level + 1];
1234
1235         if (!parent)
1236                 return 1;
1237
1238         left = read_node_slot(root, parent, pslot - 1);
1239
1240         /* first, try to make some room in the middle buffer */
1241         if (left) {
1242                 u32 left_nr;
1243
1244                 btrfs_tree_lock(left);
1245                 btrfs_set_lock_blocking(left);
1246
1247                 left_nr = btrfs_header_nritems(left);
1248                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1249                         wret = 1;
1250                 } else {
1251                         ret = btrfs_cow_block(trans, root, left, parent,
1252                                               pslot - 1, &left);
1253                         if (ret)
1254                                 wret = 1;
1255                         else {
1256                                 wret = push_node_left(trans, root,
1257                                                       left, mid, 0);
1258                         }
1259                 }
1260                 if (wret < 0)
1261                         ret = wret;
1262                 if (wret == 0) {
1263                         struct btrfs_disk_key disk_key;
1264                         orig_slot += left_nr;
1265                         btrfs_node_key(mid, &disk_key, 0);
1266                         btrfs_set_node_key(parent, &disk_key, pslot);
1267                         btrfs_mark_buffer_dirty(parent);
1268                         if (btrfs_header_nritems(left) > orig_slot) {
1269                                 path->nodes[level] = left;
1270                                 path->slots[level + 1] -= 1;
1271                                 path->slots[level] = orig_slot;
1272                                 btrfs_tree_unlock(mid);
1273                                 free_extent_buffer(mid);
1274                         } else {
1275                                 orig_slot -=
1276                                         btrfs_header_nritems(left);
1277                                 path->slots[level] = orig_slot;
1278                                 btrfs_tree_unlock(left);
1279                                 free_extent_buffer(left);
1280                         }
1281                         return 0;
1282                 }
1283                 btrfs_tree_unlock(left);
1284                 free_extent_buffer(left);
1285         }
1286         right = read_node_slot(root, parent, pslot + 1);
1287
1288         /*
1289          * then try to empty the right most buffer into the middle
1290          */
1291         if (right) {
1292                 u32 right_nr;
1293
1294                 btrfs_tree_lock(right);
1295                 btrfs_set_lock_blocking(right);
1296
1297                 right_nr = btrfs_header_nritems(right);
1298                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1299                         wret = 1;
1300                 } else {
1301                         ret = btrfs_cow_block(trans, root, right,
1302                                               parent, pslot + 1,
1303                                               &right);
1304                         if (ret)
1305                                 wret = 1;
1306                         else {
1307                                 wret = balance_node_right(trans, root,
1308                                                           right, mid);
1309                         }
1310                 }
1311                 if (wret < 0)
1312                         ret = wret;
1313                 if (wret == 0) {
1314                         struct btrfs_disk_key disk_key;
1315
1316                         btrfs_node_key(right, &disk_key, 0);
1317                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
1318                         btrfs_mark_buffer_dirty(parent);
1319
1320                         if (btrfs_header_nritems(mid) <= orig_slot) {
1321                                 path->nodes[level] = right;
1322                                 path->slots[level + 1] += 1;
1323                                 path->slots[level] = orig_slot -
1324                                         btrfs_header_nritems(mid);
1325                                 btrfs_tree_unlock(mid);
1326                                 free_extent_buffer(mid);
1327                         } else {
1328                                 btrfs_tree_unlock(right);
1329                                 free_extent_buffer(right);
1330                         }
1331                         return 0;
1332                 }
1333                 btrfs_tree_unlock(right);
1334                 free_extent_buffer(right);
1335         }
1336         return 1;
1337 }
1338
1339 /*
1340  * readahead one full node of leaves, finding things that are close
1341  * to the block in 'slot', and triggering ra on them.
1342  */
1343 static void reada_for_search(struct btrfs_root *root,
1344                              struct btrfs_path *path,
1345                              int level, int slot, u64 objectid)
1346 {
1347         struct extent_buffer *node;
1348         struct btrfs_disk_key disk_key;
1349         u32 nritems;
1350         u64 search;
1351         u64 target;
1352         u64 nread = 0;
1353         int direction = path->reada;
1354         struct extent_buffer *eb;
1355         u32 nr;
1356         u32 blocksize;
1357         u32 nscan = 0;
1358
1359         if (level != 1)
1360                 return;
1361
1362         if (!path->nodes[level])
1363                 return;
1364
1365         node = path->nodes[level];
1366
1367         search = btrfs_node_blockptr(node, slot);
1368         blocksize = btrfs_level_size(root, level - 1);
1369         eb = btrfs_find_tree_block(root, search, blocksize);
1370         if (eb) {
1371                 free_extent_buffer(eb);
1372                 return;
1373         }
1374
1375         target = search;
1376
1377         nritems = btrfs_header_nritems(node);
1378         nr = slot;
1379         while (1) {
1380                 if (direction < 0) {
1381                         if (nr == 0)
1382                                 break;
1383                         nr--;
1384                 } else if (direction > 0) {
1385                         nr++;
1386                         if (nr >= nritems)
1387                                 break;
1388                 }
1389                 if (path->reada < 0 && objectid) {
1390                         btrfs_node_key(node, &disk_key, nr);
1391                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1392                                 break;
1393                 }
1394                 search = btrfs_node_blockptr(node, nr);
1395                 if ((search <= target && target - search <= 65536) ||
1396                     (search > target && search - target <= 65536)) {
1397                         readahead_tree_block(root, search, blocksize,
1398                                      btrfs_node_ptr_generation(node, nr));
1399                         nread += blocksize;
1400                 }
1401                 nscan++;
1402                 if ((nread > 65536 || nscan > 32))
1403                         break;
1404         }
1405 }
1406
1407 /*
1408  * returns -EAGAIN if it had to drop the path, or zero if everything was in
1409  * cache
1410  */
1411 static noinline int reada_for_balance(struct btrfs_root *root,
1412                                       struct btrfs_path *path, int level)
1413 {
1414         int slot;
1415         int nritems;
1416         struct extent_buffer *parent;
1417         struct extent_buffer *eb;
1418         u64 gen;
1419         u64 block1 = 0;
1420         u64 block2 = 0;
1421         int ret = 0;
1422         int blocksize;
1423
1424         parent = path->nodes[level + 1];
1425         if (!parent)
1426                 return 0;
1427
1428         nritems = btrfs_header_nritems(parent);
1429         slot = path->slots[level + 1];
1430         blocksize = btrfs_level_size(root, level);
1431
1432         if (slot > 0) {
1433                 block1 = btrfs_node_blockptr(parent, slot - 1);
1434                 gen = btrfs_node_ptr_generation(parent, slot - 1);
1435                 eb = btrfs_find_tree_block(root, block1, blocksize);
1436                 if (eb && btrfs_buffer_uptodate(eb, gen))
1437                         block1 = 0;
1438                 free_extent_buffer(eb);
1439         }
1440         if (slot + 1 < nritems) {
1441                 block2 = btrfs_node_blockptr(parent, slot + 1);
1442                 gen = btrfs_node_ptr_generation(parent, slot + 1);
1443                 eb = btrfs_find_tree_block(root, block2, blocksize);
1444                 if (eb && btrfs_buffer_uptodate(eb, gen))
1445                         block2 = 0;
1446                 free_extent_buffer(eb);
1447         }
1448         if (block1 || block2) {
1449                 ret = -EAGAIN;
1450
1451                 /* release the whole path */
1452                 btrfs_release_path(root, path);
1453
1454                 /* read the blocks */
1455                 if (block1)
1456                         readahead_tree_block(root, block1, blocksize, 0);
1457                 if (block2)
1458                         readahead_tree_block(root, block2, blocksize, 0);
1459
1460                 if (block1) {
1461                         eb = read_tree_block(root, block1, blocksize, 0);
1462                         free_extent_buffer(eb);
1463                 }
1464                 if (block2) {
1465                         eb = read_tree_block(root, block2, blocksize, 0);
1466                         free_extent_buffer(eb);
1467                 }
1468         }
1469         return ret;
1470 }
1471
1472
1473 /*
1474  * when we walk down the tree, it is usually safe to unlock the higher layers
1475  * in the tree.  The exceptions are when our path goes through slot 0, because
1476  * operations on the tree might require changing key pointers higher up in the
1477  * tree.
1478  *
1479  * callers might also have set path->keep_locks, which tells this code to keep
1480  * the lock if the path points to the last slot in the block.  This is part of
1481  * walking through the tree, and selecting the next slot in the higher block.
1482  *
1483  * lowest_unlock sets the lowest level in the tree we're allowed to unlock.  so
1484  * if lowest_unlock is 1, level 0 won't be unlocked
1485  */
1486 static noinline void unlock_up(struct btrfs_path *path, int level,
1487                                int lowest_unlock)
1488 {
1489         int i;
1490         int skip_level = level;
1491         int no_skips = 0;
1492         struct extent_buffer *t;
1493
1494         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1495                 if (!path->nodes[i])
1496                         break;
1497                 if (!path->locks[i])
1498                         break;
1499                 if (!no_skips && path->slots[i] == 0) {
1500                         skip_level = i + 1;
1501                         continue;
1502                 }
1503                 if (!no_skips && path->keep_locks) {
1504                         u32 nritems;
1505                         t = path->nodes[i];
1506                         nritems = btrfs_header_nritems(t);
1507                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
1508                                 skip_level = i + 1;
1509                                 continue;
1510                         }
1511                 }
1512                 if (skip_level < i && i >= lowest_unlock)
1513                         no_skips = 1;
1514
1515                 t = path->nodes[i];
1516                 if (i >= lowest_unlock && i > skip_level && path->locks[i]) {
1517                         btrfs_tree_unlock(t);
1518                         path->locks[i] = 0;
1519                 }
1520         }
1521 }
1522
1523 /*
1524  * This releases any locks held in the path starting at level and
1525  * going all the way up to the root.
1526  *
1527  * btrfs_search_slot will keep the lock held on higher nodes in a few
1528  * corner cases, such as COW of the block at slot zero in the node.  This
1529  * ignores those rules, and it should only be called when there are no
1530  * more updates to be done higher up in the tree.
1531  */
1532 noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
1533 {
1534         int i;
1535
1536         if (path->keep_locks)
1537                 return;
1538
1539         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1540                 if (!path->nodes[i])
1541                         continue;
1542                 if (!path->locks[i])
1543                         continue;
1544                 btrfs_tree_unlock(path->nodes[i]);
1545                 path->locks[i] = 0;
1546         }
1547 }
1548
1549 /*
1550  * helper function for btrfs_search_slot.  The goal is to find a block
1551  * in cache without setting the path to blocking.  If we find the block
1552  * we return zero and the path is unchanged.
1553  *
1554  * If we can't find the block, we set the path blocking and do some
1555  * reada.  -EAGAIN is returned and the search must be repeated.
1556  */
1557 static int
1558 read_block_for_search(struct btrfs_trans_handle *trans,
1559                        struct btrfs_root *root, struct btrfs_path *p,
1560                        struct extent_buffer **eb_ret, int level, int slot,
1561                        struct btrfs_key *key)
1562 {
1563         u64 blocknr;
1564         u64 gen;
1565         u32 blocksize;
1566         struct extent_buffer *b = *eb_ret;
1567         struct extent_buffer *tmp;
1568         int ret;
1569
1570         blocknr = btrfs_node_blockptr(b, slot);
1571         gen = btrfs_node_ptr_generation(b, slot);
1572         blocksize = btrfs_level_size(root, level - 1);
1573
1574         tmp = btrfs_find_tree_block(root, blocknr, blocksize);
1575         if (tmp) {
1576                 if (btrfs_buffer_uptodate(tmp, 0)) {
1577                         if (btrfs_buffer_uptodate(tmp, gen)) {
1578                                 /*
1579                                  * we found an up to date block without
1580                                  * sleeping, return
1581                                  * right away
1582                                  */
1583                                 *eb_ret = tmp;
1584                                 return 0;
1585                         }
1586                         /* the pages were up to date, but we failed
1587                          * the generation number check.  Do a full
1588                          * read for the generation number that is correct.
1589                          * We must do this without dropping locks so
1590                          * we can trust our generation number
1591                          */
1592                         free_extent_buffer(tmp);
1593                         tmp = read_tree_block(root, blocknr, blocksize, gen);
1594                         if (tmp && btrfs_buffer_uptodate(tmp, gen)) {
1595                                 *eb_ret = tmp;
1596                                 return 0;
1597                         }
1598                         free_extent_buffer(tmp);
1599                         btrfs_release_path(NULL, p);
1600                         return -EIO;
1601                 }
1602         }
1603
1604         /*
1605          * reduce lock contention at high levels
1606          * of the btree by dropping locks before
1607          * we read.  Don't release the lock on the current
1608          * level because we need to walk this node to figure
1609          * out which blocks to read.
1610          */
1611         btrfs_unlock_up_safe(p, level + 1);
1612         btrfs_set_path_blocking(p);
1613
1614         free_extent_buffer(tmp);
1615         if (p->reada)
1616                 reada_for_search(root, p, level, slot, key->objectid);
1617
1618         btrfs_release_path(NULL, p);
1619
1620         ret = -EAGAIN;
1621         tmp = read_tree_block(root, blocknr, blocksize, 0);
1622         if (tmp) {
1623                 /*
1624                  * If the read above didn't mark this buffer up to date,
1625                  * it will never end up being up to date.  Set ret to EIO now
1626                  * and give up so that our caller doesn't loop forever
1627                  * on our EAGAINs.
1628                  */
1629                 if (!btrfs_buffer_uptodate(tmp, 0))
1630                         ret = -EIO;
1631                 free_extent_buffer(tmp);
1632         }
1633         return ret;
1634 }
1635
1636 /*
1637  * helper function for btrfs_search_slot.  This does all of the checks
1638  * for node-level blocks and does any balancing required based on
1639  * the ins_len.
1640  *
1641  * If no extra work was required, zero is returned.  If we had to
1642  * drop the path, -EAGAIN is returned and btrfs_search_slot must
1643  * start over
1644  */
1645 static int
1646 setup_nodes_for_search(struct btrfs_trans_handle *trans,
1647                        struct btrfs_root *root, struct btrfs_path *p,
1648                        struct extent_buffer *b, int level, int ins_len)
1649 {
1650         int ret;
1651         if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
1652             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1653                 int sret;
1654
1655                 sret = reada_for_balance(root, p, level);
1656                 if (sret)
1657                         goto again;
1658
1659                 btrfs_set_path_blocking(p);
1660                 sret = split_node(trans, root, p, level);
1661                 btrfs_clear_path_blocking(p, NULL);
1662
1663                 BUG_ON(sret > 0);
1664                 if (sret) {
1665                         ret = sret;
1666                         goto done;
1667                 }
1668                 b = p->nodes[level];
1669         } else if (ins_len < 0 && btrfs_header_nritems(b) <
1670                    BTRFS_NODEPTRS_PER_BLOCK(root) / 2) {
1671                 int sret;
1672
1673                 sret = reada_for_balance(root, p, level);
1674                 if (sret)
1675                         goto again;
1676
1677                 btrfs_set_path_blocking(p);
1678                 sret = balance_level(trans, root, p, level);
1679                 btrfs_clear_path_blocking(p, NULL);
1680
1681                 if (sret) {
1682                         ret = sret;
1683                         goto done;
1684                 }
1685                 b = p->nodes[level];
1686                 if (!b) {
1687                         btrfs_release_path(NULL, p);
1688                         goto again;
1689                 }
1690                 BUG_ON(btrfs_header_nritems(b) == 1);
1691         }
1692         return 0;
1693
1694 again:
1695         ret = -EAGAIN;
1696 done:
1697         return ret;
1698 }
1699
1700 /*
1701  * look for key in the tree.  path is filled in with nodes along the way
1702  * if key is found, we return zero and you can find the item in the leaf
1703  * level of the path (level 0)
1704  *
1705  * If the key isn't found, the path points to the slot where it should
1706  * be inserted, and 1 is returned.  If there are other errors during the
1707  * search a negative error number is returned.
1708  *
1709  * if ins_len > 0, nodes and leaves will be split as we walk down the
1710  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1711  * possible)
1712  */
1713 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1714                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1715                       ins_len, int cow)
1716 {
1717         struct extent_buffer *b;
1718         int slot;
1719         int ret;
1720         int err;
1721         int level;
1722         int lowest_unlock = 1;
1723         u8 lowest_level = 0;
1724
1725         lowest_level = p->lowest_level;
1726         WARN_ON(lowest_level && ins_len > 0);
1727         WARN_ON(p->nodes[0] != NULL);
1728
1729         if (ins_len < 0)
1730                 lowest_unlock = 2;
1731
1732 again:
1733         if (p->search_commit_root) {
1734                 b = root->commit_root;
1735                 extent_buffer_get(b);
1736                 if (!p->skip_locking)
1737                         btrfs_tree_lock(b);
1738         } else {
1739                 if (p->skip_locking)
1740                         b = btrfs_root_node(root);
1741                 else
1742                         b = btrfs_lock_root_node(root);
1743         }
1744
1745         while (b) {
1746                 level = btrfs_header_level(b);
1747
1748                 /*
1749                  * setup the path here so we can release it under lock
1750                  * contention with the cow code
1751                  */
1752                 p->nodes[level] = b;
1753                 if (!p->skip_locking)
1754                         p->locks[level] = 1;
1755
1756                 if (cow) {
1757                         /*
1758                          * if we don't really need to cow this block
1759                          * then we don't want to set the path blocking,
1760                          * so we test it here
1761                          */
1762                         if (!should_cow_block(trans, root, b))
1763                                 goto cow_done;
1764
1765                         btrfs_set_path_blocking(p);
1766
1767                         err = btrfs_cow_block(trans, root, b,
1768                                               p->nodes[level + 1],
1769                                               p->slots[level + 1], &b);
1770                         if (err) {
1771                                 ret = err;
1772                                 goto done;
1773                         }
1774                 }
1775 cow_done:
1776                 BUG_ON(!cow && ins_len);
1777                 if (level != btrfs_header_level(b))
1778                         WARN_ON(1);
1779                 level = btrfs_header_level(b);
1780
1781                 p->nodes[level] = b;
1782                 if (!p->skip_locking)
1783                         p->locks[level] = 1;
1784
1785                 btrfs_clear_path_blocking(p, NULL);
1786
1787                 /*
1788                  * we have a lock on b and as long as we aren't changing
1789                  * the tree, there is no way to for the items in b to change.
1790                  * It is safe to drop the lock on our parent before we
1791                  * go through the expensive btree search on b.
1792                  *
1793                  * If cow is true, then we might be changing slot zero,
1794                  * which may require changing the parent.  So, we can't
1795                  * drop the lock until after we know which slot we're
1796                  * operating on.
1797                  */
1798                 if (!cow)
1799                         btrfs_unlock_up_safe(p, level + 1);
1800
1801                 ret = check_block(root, p, level);
1802                 if (ret) {
1803                         ret = -1;
1804                         goto done;
1805                 }
1806
1807                 ret = bin_search(b, key, level, &slot);
1808
1809                 if (level != 0) {
1810                         int dec = 0;
1811                         if (ret && slot > 0) {
1812                                 dec = 1;
1813                                 slot -= 1;
1814                         }
1815                         p->slots[level] = slot;
1816                         err = setup_nodes_for_search(trans, root, p, b, level,
1817                                                      ins_len);
1818                         if (err == -EAGAIN)
1819                                 goto again;
1820                         if (err) {
1821                                 ret = err;
1822                                 goto done;
1823                         }
1824                         b = p->nodes[level];
1825                         slot = p->slots[level];
1826
1827                         unlock_up(p, level, lowest_unlock);
1828
1829                         if (level == lowest_level) {
1830                                 if (dec)
1831                                         p->slots[level]++;
1832                                 goto done;
1833                         }
1834
1835                         err = read_block_for_search(trans, root, p,
1836                                                     &b, level, slot, key);
1837                         if (err == -EAGAIN)
1838                                 goto again;
1839                         if (err) {
1840                                 ret = err;
1841                                 goto done;
1842                         }
1843
1844                         if (!p->skip_locking) {
1845                                 btrfs_clear_path_blocking(p, NULL);
1846                                 err = btrfs_try_spin_lock(b);
1847
1848                                 if (!err) {
1849                                         btrfs_set_path_blocking(p);
1850                                         btrfs_tree_lock(b);
1851                                         btrfs_clear_path_blocking(p, b);
1852                                 }
1853                         }
1854                 } else {
1855                         p->slots[level] = slot;
1856                         if (ins_len > 0 &&
1857                             btrfs_leaf_free_space(root, b) < ins_len) {
1858                                 btrfs_set_path_blocking(p);
1859                                 err = split_leaf(trans, root, key,
1860                                                  p, ins_len, ret == 0);
1861                                 btrfs_clear_path_blocking(p, NULL);
1862
1863                                 BUG_ON(err > 0);
1864                                 if (err) {
1865                                         ret = err;
1866                                         goto done;
1867                                 }
1868                         }
1869                         if (!p->search_for_split)
1870                                 unlock_up(p, level, lowest_unlock);
1871                         goto done;
1872                 }
1873         }
1874         ret = 1;
1875 done:
1876         /*
1877          * we don't really know what they plan on doing with the path
1878          * from here on, so for now just mark it as blocking
1879          */
1880         if (!p->leave_spinning)
1881                 btrfs_set_path_blocking(p);
1882         if (ret < 0)
1883                 btrfs_release_path(root, p);
1884         return ret;
1885 }
1886
1887 /*
1888  * adjust the pointers going up the tree, starting at level
1889  * making sure the right key of each node is points to 'key'.
1890  * This is used after shifting pointers to the left, so it stops
1891  * fixing up pointers when a given leaf/node is not in slot 0 of the
1892  * higher levels
1893  *
1894  * If this fails to write a tree block, it returns -1, but continues
1895  * fixing up the blocks in ram so the tree is consistent.
1896  */
1897 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1898                           struct btrfs_root *root, struct btrfs_path *path,
1899                           struct btrfs_disk_key *key, int level)
1900 {
1901         int i;
1902         int ret = 0;
1903         struct extent_buffer *t;
1904
1905         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1906                 int tslot = path->slots[i];
1907                 if (!path->nodes[i])
1908                         break;
1909                 t = path->nodes[i];
1910                 btrfs_set_node_key(t, key, tslot);
1911                 btrfs_mark_buffer_dirty(path->nodes[i]);
1912                 if (tslot != 0)
1913                         break;
1914         }
1915         return ret;
1916 }
1917
1918 /*
1919  * update item key.
1920  *
1921  * This function isn't completely safe. It's the caller's responsibility
1922  * that the new key won't break the order
1923  */
1924 int btrfs_set_item_key_safe(struct btrfs_trans_handle *trans,
1925                             struct btrfs_root *root, struct btrfs_path *path,
1926                             struct btrfs_key *new_key)
1927 {
1928         struct btrfs_disk_key disk_key;
1929         struct extent_buffer *eb;
1930         int slot;
1931
1932         eb = path->nodes[0];
1933         slot = path->slots[0];
1934         if (slot > 0) {
1935                 btrfs_item_key(eb, &disk_key, slot - 1);
1936                 if (comp_keys(&disk_key, new_key) >= 0)
1937                         return -1;
1938         }
1939         if (slot < btrfs_header_nritems(eb) - 1) {
1940                 btrfs_item_key(eb, &disk_key, slot + 1);
1941                 if (comp_keys(&disk_key, new_key) <= 0)
1942                         return -1;
1943         }
1944
1945         btrfs_cpu_key_to_disk(&disk_key, new_key);
1946         btrfs_set_item_key(eb, &disk_key, slot);
1947         btrfs_mark_buffer_dirty(eb);
1948         if (slot == 0)
1949                 fixup_low_keys(trans, root, path, &disk_key, 1);
1950         return 0;
1951 }
1952
1953 /*
1954  * try to push data from one node into the next node left in the
1955  * tree.
1956  *
1957  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1958  * error, and > 0 if there was no room in the left hand block.
1959  */
1960 static int push_node_left(struct btrfs_trans_handle *trans,
1961                           struct btrfs_root *root, struct extent_buffer *dst,
1962                           struct extent_buffer *src, int empty)
1963 {
1964         int push_items = 0;
1965         int src_nritems;
1966         int dst_nritems;
1967         int ret = 0;
1968
1969         src_nritems = btrfs_header_nritems(src);
1970         dst_nritems = btrfs_header_nritems(dst);
1971         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1972         WARN_ON(btrfs_header_generation(src) != trans->transid);
1973         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1974
1975         if (!empty && src_nritems <= 8)
1976                 return 1;
1977
1978         if (push_items <= 0)
1979                 return 1;
1980
1981         if (empty) {
1982                 push_items = min(src_nritems, push_items);
1983                 if (push_items < src_nritems) {
1984                         /* leave at least 8 pointers in the node if
1985                          * we aren't going to empty it
1986                          */
1987                         if (src_nritems - push_items < 8) {
1988                                 if (push_items <= 8)
1989                                         return 1;
1990                                 push_items -= 8;
1991                         }
1992                 }
1993         } else
1994                 push_items = min(src_nritems - 8, push_items);
1995
1996         copy_extent_buffer(dst, src,
1997                            btrfs_node_key_ptr_offset(dst_nritems),
1998                            btrfs_node_key_ptr_offset(0),
1999                            push_items * sizeof(struct btrfs_key_ptr));
2000
2001         if (push_items < src_nritems) {
2002                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
2003                                       btrfs_node_key_ptr_offset(push_items),
2004                                       (src_nritems - push_items) *
2005                                       sizeof(struct btrfs_key_ptr));
2006         }
2007         btrfs_set_header_nritems(src, src_nritems - push_items);
2008         btrfs_set_header_nritems(dst, dst_nritems + push_items);
2009         btrfs_mark_buffer_dirty(src);
2010         btrfs_mark_buffer_dirty(dst);
2011
2012         return ret;
2013 }
2014
2015 /*
2016  * try to push data from one node into the next node right in the
2017  * tree.
2018  *
2019  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
2020  * error, and > 0 if there was no room in the right hand block.
2021  *
2022  * this will  only push up to 1/2 the contents of the left node over
2023  */
2024 static int balance_node_right(struct btrfs_trans_handle *trans,
2025                               struct btrfs_root *root,
2026                               struct extent_buffer *dst,
2027                               struct extent_buffer *src)
2028 {
2029         int push_items = 0;
2030         int max_push;
2031         int src_nritems;
2032         int dst_nritems;
2033         int ret = 0;
2034
2035         WARN_ON(btrfs_header_generation(src) != trans->transid);
2036         WARN_ON(btrfs_header_generation(dst) != trans->transid);
2037
2038         src_nritems = btrfs_header_nritems(src);
2039         dst_nritems = btrfs_header_nritems(dst);
2040         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
2041         if (push_items <= 0)
2042                 return 1;
2043
2044         if (src_nritems < 4)
2045                 return 1;
2046
2047         max_push = src_nritems / 2 + 1;
2048         /* don't try to empty the node */
2049         if (max_push >= src_nritems)
2050                 return 1;
2051
2052         if (max_push < push_items)
2053                 push_items = max_push;
2054
2055         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
2056                                       btrfs_node_key_ptr_offset(0),
2057                                       (dst_nritems) *
2058                                       sizeof(struct btrfs_key_ptr));
2059
2060         copy_extent_buffer(dst, src,
2061                            btrfs_node_key_ptr_offset(0),
2062                            btrfs_node_key_ptr_offset(src_nritems - push_items),
2063                            push_items * sizeof(struct btrfs_key_ptr));
2064
2065         btrfs_set_header_nritems(src, src_nritems - push_items);
2066         btrfs_set_header_nritems(dst, dst_nritems + push_items);
2067
2068         btrfs_mark_buffer_dirty(src);
2069         btrfs_mark_buffer_dirty(dst);
2070
2071         return ret;
2072 }
2073
2074 /*
2075  * helper function to insert a new root level in the tree.
2076  * A new node is allocated, and a single item is inserted to
2077  * point to the existing root
2078  *
2079  * returns zero on success or < 0 on failure.
2080  */
2081 static noinline int insert_new_root(struct btrfs_trans_handle *trans,
2082                            struct btrfs_root *root,
2083                            struct btrfs_path *path, int level)
2084 {
2085         u64 lower_gen;
2086         struct extent_buffer *lower;
2087         struct extent_buffer *c;
2088         struct extent_buffer *old;
2089         struct btrfs_disk_key lower_key;
2090
2091         BUG_ON(path->nodes[level]);
2092         BUG_ON(path->nodes[level-1] != root->node);
2093
2094         lower = path->nodes[level-1];
2095         if (level == 1)
2096                 btrfs_item_key(lower, &lower_key, 0);
2097         else
2098                 btrfs_node_key(lower, &lower_key, 0);
2099
2100         c = btrfs_alloc_free_block(trans, root, root->nodesize, 0,
2101                                    root->root_key.objectid, &lower_key,
2102                                    level, root->node->start, 0);
2103         if (IS_ERR(c))
2104                 return PTR_ERR(c);
2105
2106         root_add_used(root, root->nodesize);
2107
2108         memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
2109         btrfs_set_header_nritems(c, 1);
2110         btrfs_set_header_level(c, level);
2111         btrfs_set_header_bytenr(c, c->start);
2112         btrfs_set_header_generation(c, trans->transid);
2113         btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
2114         btrfs_set_header_owner(c, root->root_key.objectid);
2115
2116         write_extent_buffer(c, root->fs_info->fsid,
2117                             (unsigned long)btrfs_header_fsid(c),
2118                             BTRFS_FSID_SIZE);
2119
2120         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
2121                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
2122                             BTRFS_UUID_SIZE);
2123
2124         btrfs_set_node_key(c, &lower_key, 0);
2125         btrfs_set_node_blockptr(c, 0, lower->start);
2126         lower_gen = btrfs_header_generation(lower);
2127         WARN_ON(lower_gen != trans->transid);
2128
2129         btrfs_set_node_ptr_generation(c, 0, lower_gen);
2130
2131         btrfs_mark_buffer_dirty(c);
2132
2133         spin_lock(&root->node_lock);
2134         old = root->node;
2135         root->node = c;
2136         spin_unlock(&root->node_lock);
2137
2138         /* the super has an extra ref to root->node */
2139         free_extent_buffer(old);
2140
2141         add_root_to_dirty_list(root);
2142         extent_buffer_get(c);
2143         path->nodes[level] = c;
2144         path->locks[level] = 1;
2145         path->slots[level] = 0;
2146         return 0;
2147 }
2148
2149 /*
2150  * worker function to insert a single pointer in a node.
2151  * the node should have enough room for the pointer already
2152  *
2153  * slot and level indicate where you want the key to go, and
2154  * blocknr is the block the key points to.
2155  *
2156  * returns zero on success and < 0 on any error
2157  */
2158 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
2159                       *root, struct btrfs_path *path, struct btrfs_disk_key
2160                       *key, u64 bytenr, int slot, int level)
2161 {
2162         struct extent_buffer *lower;
2163         int nritems;
2164
2165         BUG_ON(!path->nodes[level]);
2166         btrfs_assert_tree_locked(path->nodes[level]);
2167         lower = path->nodes[level];
2168         nritems = btrfs_header_nritems(lower);
2169         BUG_ON(slot > nritems);
2170         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
2171                 BUG();
2172         if (slot != nritems) {
2173                 memmove_extent_buffer(lower,
2174                               btrfs_node_key_ptr_offset(slot + 1),
2175                               btrfs_node_key_ptr_offset(slot),
2176                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
2177         }
2178         btrfs_set_node_key(lower, key, slot);
2179         btrfs_set_node_blockptr(lower, slot, bytenr);
2180         WARN_ON(trans->transid == 0);
2181         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
2182         btrfs_set_header_nritems(lower, nritems + 1);
2183         btrfs_mark_buffer_dirty(lower);
2184         return 0;
2185 }
2186
2187 /*
2188  * split the node at the specified level in path in two.
2189  * The path is corrected to point to the appropriate node after the split
2190  *
2191  * Before splitting this tries to make some room in the node by pushing
2192  * left and right, if either one works, it returns right away.
2193  *
2194  * returns 0 on success and < 0 on failure
2195  */
2196 static noinline int split_node(struct btrfs_trans_handle *trans,
2197                                struct btrfs_root *root,
2198                                struct btrfs_path *path, int level)
2199 {
2200         struct extent_buffer *c;
2201         struct extent_buffer *split;
2202         struct btrfs_disk_key disk_key;
2203         int mid;
2204         int ret;
2205         int wret;
2206         u32 c_nritems;
2207
2208         c = path->nodes[level];
2209         WARN_ON(btrfs_header_generation(c) != trans->transid);
2210         if (c == root->node) {
2211                 /* trying to split the root, lets make a new one */
2212                 ret = insert_new_root(trans, root, path, level + 1);
2213                 if (ret)
2214                         return ret;
2215         } else {
2216                 ret = push_nodes_for_insert(trans, root, path, level);
2217                 c = path->nodes[level];
2218                 if (!ret && btrfs_header_nritems(c) <
2219                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
2220                         return 0;
2221                 if (ret < 0)
2222                         return ret;
2223         }
2224
2225         c_nritems = btrfs_header_nritems(c);
2226         mid = (c_nritems + 1) / 2;
2227         btrfs_node_key(c, &disk_key, mid);
2228
2229         split = btrfs_alloc_free_block(trans, root, root->nodesize, 0,
2230                                         root->root_key.objectid,
2231                                         &disk_key, level, c->start, 0);
2232         if (IS_ERR(split))
2233                 return PTR_ERR(split);
2234
2235         root_add_used(root, root->nodesize);
2236
2237         memset_extent_buffer(split, 0, 0, sizeof(struct btrfs_header));
2238         btrfs_set_header_level(split, btrfs_header_level(c));
2239         btrfs_set_header_bytenr(split, split->start);
2240         btrfs_set_header_generation(split, trans->transid);
2241         btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
2242         btrfs_set_header_owner(split, root->root_key.objectid);
2243         write_extent_buffer(split, root->fs_info->fsid,
2244                             (unsigned long)btrfs_header_fsid(split),
2245                             BTRFS_FSID_SIZE);
2246         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
2247                             (unsigned long)btrfs_header_chunk_tree_uuid(split),
2248                             BTRFS_UUID_SIZE);
2249
2250
2251         copy_extent_buffer(split, c,
2252                            btrfs_node_key_ptr_offset(0),
2253                            btrfs_node_key_ptr_offset(mid),
2254                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
2255         btrfs_set_header_nritems(split, c_nritems - mid);
2256         btrfs_set_header_nritems(c, mid);
2257         ret = 0;
2258
2259         btrfs_mark_buffer_dirty(c);
2260         btrfs_mark_buffer_dirty(split);
2261
2262         wret = insert_ptr(trans, root, path, &disk_key, split->start,
2263                           path->slots[level + 1] + 1,
2264                           level + 1);
2265         if (wret)
2266                 ret = wret;
2267
2268         if (path->slots[level] >= mid) {
2269                 path->slots[level] -= mid;
2270                 btrfs_tree_unlock(c);
2271                 free_extent_buffer(c);
2272                 path->nodes[level] = split;
2273                 path->slots[level + 1] += 1;
2274         } else {
2275                 btrfs_tree_unlock(split);
2276                 free_extent_buffer(split);
2277         }
2278         return ret;
2279 }
2280
2281 /*
2282  * how many bytes are required to store the items in a leaf.  start
2283  * and nr indicate which items in the leaf to check.  This totals up the
2284  * space used both by the item structs and the item data
2285  */
2286 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
2287 {
2288         int data_len;
2289         int nritems = btrfs_header_nritems(l);
2290         int end = min(nritems, start + nr) - 1;
2291
2292         if (!nr)
2293                 return 0;
2294         data_len = btrfs_item_end_nr(l, start);
2295         data_len = data_len - btrfs_item_offset_nr(l, end);
2296         data_len += sizeof(struct btrfs_item) * nr;
2297         WARN_ON(data_len < 0);
2298         return data_len;
2299 }
2300
2301 /*
2302  * The space between the end of the leaf items and
2303  * the start of the leaf data.  IOW, how much room
2304  * the leaf has left for both items and data
2305  */
2306 noinline int btrfs_leaf_free_space(struct btrfs_root *root,
2307                                    struct extent_buffer *leaf)
2308 {
2309         int nritems = btrfs_header_nritems(leaf);
2310         int ret;
2311         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
2312         if (ret < 0) {
2313                 printk(KERN_CRIT "leaf free space ret %d, leaf data size %lu, "
2314                        "used %d nritems %d\n",
2315                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
2316                        leaf_space_used(leaf, 0, nritems), nritems);
2317         }
2318         return ret;
2319 }
2320
2321 /*
2322  * min slot controls the lowest index we're willing to push to the
2323  * right.  We'll push up to and including min_slot, but no lower
2324  */
2325 static noinline int __push_leaf_right(struct btrfs_trans_handle *trans,
2326                                       struct btrfs_root *root,
2327                                       struct btrfs_path *path,
2328                                       int data_size, int empty,
2329                                       struct extent_buffer *right,
2330                                       int free_space, u32 left_nritems,
2331                                       u32 min_slot)
2332 {
2333         struct extent_buffer *left = path->nodes[0];
2334         struct extent_buffer *upper = path->nodes[1];
2335         struct btrfs_disk_key disk_key;
2336         int slot;
2337         u32 i;
2338         int push_space = 0;
2339         int push_items = 0;
2340         struct btrfs_item *item;
2341         u32 nr;
2342         u32 right_nritems;
2343         u32 data_end;
2344         u32 this_item_size;
2345
2346         if (empty)
2347                 nr = 0;
2348         else
2349                 nr = max_t(u32, 1, min_slot);
2350
2351         if (path->slots[0] >= left_nritems)
2352                 push_space += data_size;
2353
2354         slot = path->slots[1];
2355         i = left_nritems - 1;
2356         while (i >= nr) {
2357                 item = btrfs_item_nr(left, i);
2358
2359                 if (!empty && push_items > 0) {
2360                         if (path->slots[0] > i)
2361                                 break;
2362                         if (path->slots[0] == i) {
2363                                 int space = btrfs_leaf_free_space(root, left);
2364                                 if (space + push_space * 2 > free_space)
2365                                         break;
2366                         }
2367                 }
2368
2369                 if (path->slots[0] == i)
2370                         push_space += data_size;
2371
2372                 if (!left->map_token) {
2373                         map_extent_buffer(left, (unsigned long)item,
2374                                         sizeof(struct btrfs_item),
2375                                         &left->map_token, &left->kaddr,
2376                                         &left->map_start, &left->map_len,
2377                                         KM_USER1);
2378                 }
2379
2380                 this_item_size = btrfs_item_size(left, item);
2381                 if (this_item_size + sizeof(*item) + push_space > free_space)
2382                         break;
2383
2384                 push_items++;
2385                 push_space += this_item_size + sizeof(*item);
2386                 if (i == 0)
2387                         break;
2388                 i--;
2389         }
2390         if (left->map_token) {
2391                 unmap_extent_buffer(left, left->map_token, KM_USER1);
2392                 left->map_token = NULL;
2393         }
2394
2395         if (push_items == 0)
2396                 goto out_unlock;
2397
2398         if (!empty && push_items == left_nritems)
2399                 WARN_ON(1);
2400
2401         /* push left to right */
2402         right_nritems = btrfs_header_nritems(right);
2403
2404         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
2405         push_space -= leaf_data_end(root, left);
2406
2407         /* make room in the right data area */
2408         data_end = leaf_data_end(root, right);
2409         memmove_extent_buffer(right,
2410                               btrfs_leaf_data(right) + data_end - push_space,
2411                               btrfs_leaf_data(right) + data_end,
2412                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
2413
2414         /* copy from the left data area */
2415         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
2416                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
2417                      btrfs_leaf_data(left) + leaf_data_end(root, left),
2418                      push_space);
2419
2420         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
2421                               btrfs_item_nr_offset(0),
2422                               right_nritems * sizeof(struct btrfs_item));
2423
2424         /* copy the items from left to right */
2425         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
2426                    btrfs_item_nr_offset(left_nritems - push_items),
2427                    push_items * sizeof(struct btrfs_item));
2428
2429         /* update the item pointers */
2430         right_nritems += push_items;
2431         btrfs_set_header_nritems(right, right_nritems);
2432         push_space = BTRFS_LEAF_DATA_SIZE(root);
2433         for (i = 0; i < right_nritems; i++) {
2434                 item = btrfs_item_nr(right, i);
2435                 if (!right->map_token) {
2436                         map_extent_buffer(right, (unsigned long)item,
2437                                         sizeof(struct btrfs_item),
2438                                         &right->map_token, &right->kaddr,
2439                                         &right->map_start, &right->map_len,
2440                                         KM_USER1);
2441                 }
2442                 push_space -= btrfs_item_size(right, item);
2443                 btrfs_set_item_offset(right, item, push_space);
2444         }
2445
2446         if (right->map_token) {
2447                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2448                 right->map_token = NULL;
2449         }
2450         left_nritems -= push_items;
2451         btrfs_set_header_nritems(left, left_nritems);
2452
2453         if (left_nritems)
2454                 btrfs_mark_buffer_dirty(left);
2455         else
2456                 clean_tree_block(trans, root, left);
2457
2458         btrfs_mark_buffer_dirty(right);
2459
2460         btrfs_item_key(right, &disk_key, 0);
2461         btrfs_set_node_key(upper, &disk_key, slot + 1);
2462         btrfs_mark_buffer_dirty(upper);
2463
2464         /* then fixup the leaf pointer in the path */
2465         if (path->slots[0] >= left_nritems) {
2466                 path->slots[0] -= left_nritems;
2467                 if (btrfs_header_nritems(path->nodes[0]) == 0)
2468                         clean_tree_block(trans, root, path->nodes[0]);
2469                 btrfs_tree_unlock(path->nodes[0]);
2470                 free_extent_buffer(path->nodes[0]);
2471                 path->nodes[0] = right;
2472                 path->slots[1] += 1;
2473         } else {
2474                 btrfs_tree_unlock(right);
2475                 free_extent_buffer(right);
2476         }
2477         return 0;
2478
2479 out_unlock:
2480         btrfs_tree_unlock(right);
2481         free_extent_buffer(right);
2482         return 1;
2483 }
2484
2485 /*
2486  * push some data in the path leaf to the right, trying to free up at
2487  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
2488  *
2489  * returns 1 if the push failed because the other node didn't have enough
2490  * room, 0 if everything worked out and < 0 if there were major errors.
2491  *
2492  * this will push starting from min_slot to the end of the leaf.  It won't
2493  * push any slot lower than min_slot
2494  */
2495 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
2496                            *root, struct btrfs_path *path,
2497                            int min_data_size, int data_size,
2498                            int empty, u32 min_slot)
2499 {
2500         struct extent_buffer *left = path->nodes[0];
2501         struct extent_buffer *right;
2502         struct extent_buffer *upper;
2503         int slot;
2504         int free_space;
2505         u32 left_nritems;
2506         int ret;
2507
2508         if (!path->nodes[1])
2509                 return 1;
2510
2511         slot = path->slots[1];
2512         upper = path->nodes[1];
2513         if (slot >= btrfs_header_nritems(upper) - 1)
2514                 return 1;
2515
2516         btrfs_assert_tree_locked(path->nodes[1]);
2517
2518         right = read_node_slot(root, upper, slot + 1);
2519         if (right == NULL)
2520                 return 1;
2521
2522         btrfs_tree_lock(right);
2523         btrfs_set_lock_blocking(right);
2524
2525         free_space = btrfs_leaf_free_space(root, right);
2526         if (free_space < data_size)
2527                 goto out_unlock;
2528
2529         /* cow and double check */
2530         ret = btrfs_cow_block(trans, root, right, upper,
2531                               slot + 1, &right);
2532         if (ret)
2533                 goto out_unlock;
2534
2535         free_space = btrfs_leaf_free_space(root, right);
2536         if (free_space < data_size)
2537                 goto out_unlock;
2538
2539         left_nritems = btrfs_header_nritems(left);
2540         if (left_nritems == 0)
2541                 goto out_unlock;
2542
2543         return __push_leaf_right(trans, root, path, min_data_size, empty,
2544                                 right, free_space, left_nritems, min_slot);
2545 out_unlock:
2546         btrfs_tree_unlock(right);
2547         free_extent_buffer(right);
2548         return 1;
2549 }
2550
2551 /*
2552  * push some data in the path leaf to the left, trying to free up at
2553  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
2554  *
2555  * max_slot can put a limit on how far into the leaf we'll push items.  The
2556  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us do all the
2557  * items
2558  */
2559 static noinline int __push_leaf_left(struct btrfs_trans_handle *trans,
2560                                      struct btrfs_root *root,
2561                                      struct btrfs_path *path, int data_size,
2562                                      int empty, struct extent_buffer *left,
2563                                      int free_space, u32 right_nritems,
2564                                      u32 max_slot)
2565 {
2566         struct btrfs_disk_key disk_key;
2567         struct extent_buffer *right = path->nodes[0];
2568         int i;
2569         int push_space = 0;
2570         int push_items = 0;
2571         struct btrfs_item *item;
2572         u32 old_left_nritems;
2573         u32 nr;
2574         int ret = 0;
2575         int wret;
2576         u32 this_item_size;
2577         u32 old_left_item_size;
2578
2579         if (empty)
2580                 nr = min(right_nritems, max_slot);
2581         else
2582                 nr = min(right_nritems - 1, max_slot);
2583
2584         for (i = 0; i < nr; i++) {
2585                 item = btrfs_item_nr(right, i);
2586                 if (!right->map_token) {
2587                         map_extent_buffer(right, (unsigned long)item,
2588                                         sizeof(struct btrfs_item),
2589                                         &right->map_token, &right->kaddr,
2590                                         &right->map_start, &right->map_len,
2591                                         KM_USER1);
2592                 }
2593
2594                 if (!empty && push_items > 0) {
2595                         if (path->slots[0] < i)
2596                                 break;
2597                         if (path->slots[0] == i) {
2598                                 int space = btrfs_leaf_free_space(root, right);
2599                                 if (space + push_space * 2 > free_space)
2600                                         break;
2601                         }
2602                 }
2603
2604                 if (path->slots[0] == i)
2605                         push_space += data_size;
2606
2607                 this_item_size = btrfs_item_size(right, item);
2608                 if (this_item_size + sizeof(*item) + push_space > free_space)
2609                         break;
2610
2611                 push_items++;
2612                 push_space += this_item_size + sizeof(*item);
2613         }
2614
2615         if (right->map_token) {
2616                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2617                 right->map_token = NULL;
2618         }
2619
2620         if (push_items == 0) {
2621                 ret = 1;
2622                 goto out;
2623         }
2624         if (!empty && push_items == btrfs_header_nritems(right))
2625                 WARN_ON(1);
2626
2627         /* push data from right to left */
2628         copy_extent_buffer(left, right,
2629                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
2630                            btrfs_item_nr_offset(0),
2631                            push_items * sizeof(struct btrfs_item));
2632
2633         push_space = BTRFS_LEAF_DATA_SIZE(root) -
2634                      btrfs_item_offset_nr(right, push_items - 1);
2635
2636         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
2637                      leaf_data_end(root, left) - push_space,
2638                      btrfs_leaf_data(right) +
2639                      btrfs_item_offset_nr(right, push_items - 1),
2640                      push_space);
2641         old_left_nritems = btrfs_header_nritems(left);
2642         BUG_ON(old_left_nritems <= 0);
2643
2644         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
2645         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
2646                 u32 ioff;
2647
2648                 item = btrfs_item_nr(left, i);
2649                 if (!left->map_token) {
2650                         map_extent_buffer(left, (unsigned long)item,
2651                                         sizeof(struct btrfs_item),
2652                                         &left->map_token, &left->kaddr,
2653                                         &left->map_start, &left->map_len,
2654                                         KM_USER1);
2655                 }
2656
2657                 ioff = btrfs_item_offset(left, item);
2658                 btrfs_set_item_offset(left, item,
2659                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
2660         }
2661         btrfs_set_header_nritems(left, old_left_nritems + push_items);
2662         if (left->map_token) {
2663                 unmap_extent_buffer(left, left->map_token, KM_USER1);
2664                 left->map_token = NULL;
2665         }
2666
2667         /* fixup right node */
2668         if (push_items > right_nritems) {
2669                 printk(KERN_CRIT "push items %d nr %u\n", push_items,
2670                        right_nritems);
2671                 WARN_ON(1);
2672         }
2673
2674         if (push_items < right_nritems) {
2675                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
2676                                                   leaf_data_end(root, right);
2677                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
2678                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
2679                                       btrfs_leaf_data(right) +
2680                                       leaf_data_end(root, right), push_space);
2681
2682                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
2683                               btrfs_item_nr_offset(push_items),
2684                              (btrfs_header_nritems(right) - push_items) *
2685                              sizeof(struct btrfs_item));
2686         }
2687         right_nritems -= push_items;
2688         btrfs_set_header_nritems(right, right_nritems);
2689         push_space = BTRFS_LEAF_DATA_SIZE(root);
2690         for (i = 0; i < right_nritems; i++) {
2691                 item = btrfs_item_nr(right, i);
2692
2693                 if (!right->map_token) {
2694                         map_extent_buffer(right, (unsigned long)item,
2695                                         sizeof(struct btrfs_item),
2696                                         &right->map_token, &right->kaddr,
2697                                         &right->map_start, &right->map_len,
2698                                         KM_USER1);
2699                 }
2700
2701                 push_space = push_space - btrfs_item_size(right, item);
2702                 btrfs_set_item_offset(right, item, push_space);
2703         }
2704         if (right->map_token) {
2705                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2706                 right->map_token = NULL;
2707         }
2708
2709         btrfs_mark_buffer_dirty(left);
2710         if (right_nritems)
2711                 btrfs_mark_buffer_dirty(right);
2712         else
2713                 clean_tree_block(trans, root, right);
2714
2715         btrfs_item_key(right, &disk_key, 0);
2716         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
2717         if (wret)
2718                 ret = wret;
2719
2720         /* then fixup the leaf pointer in the path */
2721         if (path->slots[0] < push_items) {
2722                 path->slots[0] += old_left_nritems;
2723                 btrfs_tree_unlock(path->nodes[0]);
2724                 free_extent_buffer(path->nodes[0]);
2725                 path->nodes[0] = left;
2726                 path->slots[1] -= 1;
2727         } else {
2728                 btrfs_tree_unlock(left);
2729                 free_extent_buffer(left);
2730                 path->slots[0] -= push_items;
2731         }
2732         BUG_ON(path->slots[0] < 0);
2733         return ret;
2734 out:
2735         btrfs_tree_unlock(left);
2736         free_extent_buffer(left);
2737         return ret;
2738 }
2739
2740 /*
2741  * push some data in the path leaf to the left, trying to free up at
2742  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
2743  *
2744  * max_slot can put a limit on how far into the leaf we'll push items.  The
2745  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us push all the
2746  * items
2747  */
2748 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
2749                           *root, struct btrfs_path *path, int min_data_size,
2750                           int data_size, int empty, u32 max_slot)
2751 {
2752         struct extent_buffer *right = path->nodes[0];
2753         struct extent_buffer *left;
2754         int slot;
2755         int free_space;
2756         u32 right_nritems;
2757         int ret = 0;
2758
2759         slot = path->slots[1];
2760         if (slot == 0)
2761                 return 1;
2762         if (!path->nodes[1])
2763                 return 1;
2764
2765         right_nritems = btrfs_header_nritems(right);
2766         if (right_nritems == 0)
2767                 return 1;
2768
2769         btrfs_assert_tree_locked(path->nodes[1]);
2770
2771         left = read_node_slot(root, path->nodes[1], slot - 1);
2772         if (left == NULL)
2773                 return 1;
2774
2775         btrfs_tree_lock(left);
2776         btrfs_set_lock_blocking(left);
2777
2778         free_space = btrfs_leaf_free_space(root, left);
2779         if (free_space < data_size) {
2780                 ret = 1;
2781                 goto out;
2782         }
2783
2784         /* cow and double check */
2785         ret = btrfs_cow_block(trans, root, left,
2786                               path->nodes[1], slot - 1, &left);
2787         if (ret) {
2788                 /* we hit -ENOSPC, but it isn't fatal here */
2789                 ret = 1;
2790                 goto out;
2791         }
2792
2793         free_space = btrfs_leaf_free_space(root, left);
2794         if (free_space < data_size) {
2795                 ret = 1;
2796                 goto out;
2797         }
2798
2799         return __push_leaf_left(trans, root, path, min_data_size,
2800                                empty, left, free_space, right_nritems,
2801                                max_slot);
2802 out:
2803         btrfs_tree_unlock(left);
2804         free_extent_buffer(left);
2805         return ret;
2806 }
2807
2808 /*
2809  * split the path's leaf in two, making sure there is at least data_size
2810  * available for the resulting leaf level of the path.
2811  *
2812  * returns 0 if all went well and < 0 on failure.
2813  */
2814 static noinline int copy_for_split(struct btrfs_trans_handle *trans,
2815                                struct btrfs_root *root,
2816                                struct btrfs_path *path,
2817                                struct extent_buffer *l,
2818                                struct extent_buffer *right,
2819                                int slot, int mid, int nritems)
2820 {
2821         int data_copy_size;
2822         int rt_data_off;
2823         int i;
2824         int ret = 0;
2825         int wret;
2826         struct btrfs_disk_key disk_key;
2827
2828         nritems = nritems - mid;
2829         btrfs_set_header_nritems(right, nritems);
2830         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2831
2832         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2833                            btrfs_item_nr_offset(mid),
2834                            nritems * sizeof(struct btrfs_item));
2835
2836         copy_extent_buffer(right, l,
2837                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2838                      data_copy_size, btrfs_leaf_data(l) +
2839                      leaf_data_end(root, l), data_copy_size);
2840
2841         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2842                       btrfs_item_end_nr(l, mid);
2843
2844         for (i = 0; i < nritems; i++) {
2845                 struct btrfs_item *item = btrfs_item_nr(right, i);
2846                 u32 ioff;
2847
2848                 if (!right->map_token) {
2849                         map_extent_buffer(right, (unsigned long)item,
2850                                         sizeof(struct btrfs_item),
2851                                         &right->map_token, &right->kaddr,
2852                                         &right->map_start, &right->map_len,
2853                                         KM_USER1);
2854                 }
2855
2856                 ioff = btrfs_item_offset(right, item);
2857                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2858         }
2859
2860         if (right->map_token) {
2861                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2862                 right->map_token = NULL;
2863         }
2864
2865         btrfs_set_header_nritems(l, mid);
2866         ret = 0;
2867         btrfs_item_key(right, &disk_key, 0);
2868         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2869                           path->slots[1] + 1, 1);
2870         if (wret)
2871                 ret = wret;
2872
2873         btrfs_mark_buffer_dirty(right);
2874         btrfs_mark_buffer_dirty(l);
2875         BUG_ON(path->slots[0] != slot);
2876
2877         if (mid <= slot) {
2878                 btrfs_tree_unlock(path->nodes[0]);
2879                 free_extent_buffer(path->nodes[0]);
2880                 path->nodes[0] = right;
2881                 path->slots[0] -= mid;
2882                 path->slots[1] += 1;
2883         } else {
2884                 btrfs_tree_unlock(right);
2885                 free_extent_buffer(right);
2886         }
2887
2888         BUG_ON(path->slots[0] < 0);
2889
2890         return ret;
2891 }
2892
2893 /*
2894  * double splits happen when we need to insert a big item in the middle
2895  * of a leaf.  A double split can leave us with 3 mostly empty leaves:
2896  * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
2897  *          A                 B                 C
2898  *
2899  * We avoid this by trying to push the items on either side of our target
2900  * into the adjacent leaves.  If all goes well we can avoid the double split
2901  * completely.
2902  */
2903 static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
2904                                           struct btrfs_root *root,
2905                                           struct btrfs_path *path,
2906                                           int data_size)
2907 {
2908         int ret;
2909         int progress = 0;
2910         int slot;
2911         u32 nritems;
2912
2913         slot = path->slots[0];
2914
2915         /*
2916          * try to push all the items after our slot into the
2917          * right leaf
2918          */
2919         ret = push_leaf_right(trans, root, path, 1, data_size, 0, slot);
2920         if (ret < 0)
2921                 return ret;
2922
2923         if (ret == 0)
2924                 progress++;
2925
2926         nritems = btrfs_header_nritems(path->nodes[0]);
2927         /*
2928          * our goal is to get our slot at the start or end of a leaf.  If
2929          * we've done so we're done
2930          */
2931         if (path->slots[0] == 0 || path->slots[0] == nritems)
2932                 return 0;
2933
2934         if (btrfs_leaf_free_space(root, path->nodes[0]) >= data_size)
2935                 return 0;
2936
2937         /* try to push all the items before our slot into the next leaf */
2938         slot = path->slots[0];
2939         ret = push_leaf_left(trans, root, path, 1, data_size, 0, slot);
2940         if (ret < 0)
2941                 return ret;
2942
2943         if (ret == 0)
2944                 progress++;
2945
2946         if (progress)
2947                 return 0;
2948         return 1;
2949 }
2950
2951 /*
2952  * split the path's leaf in two, making sure there is at least data_size
2953  * available for the resulting leaf level of the path.
2954  *
2955  * returns 0 if all went well and < 0 on failure.
2956  */
2957 static noinline int split_leaf(struct btrfs_trans_handle *trans,
2958                                struct btrfs_root *root,
2959                                struct btrfs_key *ins_key,
2960                                struct btrfs_path *path, int data_size,
2961                                int extend)
2962 {
2963         struct btrfs_disk_key disk_key;
2964         struct extent_buffer *l;
2965         u32 nritems;
2966         int mid;
2967         int slot;
2968         struct extent_buffer *right;
2969         int ret = 0;
2970         int wret;
2971         int split;
2972         int num_doubles = 0;
2973         int tried_avoid_double = 0;
2974
2975         l = path->nodes[0];
2976         slot = path->slots[0];
2977         if (extend && data_size + btrfs_item_size_nr(l, slot) +
2978             sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(root))
2979                 return -EOVERFLOW;
2980
2981         /* first try to make some room by pushing left and right */
2982         if (data_size) {
2983                 wret = push_leaf_right(trans, root, path, data_size,
2984                                        data_size, 0, 0);
2985                 if (wret < 0)
2986                         return wret;
2987                 if (wret) {
2988                         wret = push_leaf_left(trans, root, path, data_size,
2989                                               data_size, 0, (u32)-1);
2990                         if (wret < 0)
2991                                 return wret;
2992                 }
2993                 l = path->nodes[0];
2994
2995                 /* did the pushes work? */
2996                 if (btrfs_leaf_free_space(root, l) >= data_size)
2997                         return 0;
2998         }
2999
3000         if (!path->nodes[1]) {
3001                 ret = insert_new_root(trans, root, path, 1);
3002                 if (ret)
3003                         return ret;
3004         }
3005 again:
3006         split = 1;
3007         l = path->nodes[0];
3008         slot = path->slots[0];
3009         nritems = btrfs_header_nritems(l);
3010         mid = (nritems + 1) / 2;
3011
3012         if (mid <= slot) {
3013                 if (nritems == 1 ||
3014                     leaf_space_used(l, mid, nritems - mid) + data_size >
3015                         BTRFS_LEAF_DATA_SIZE(root)) {
3016                         if (slot >= nritems) {
3017                                 split = 0;
3018                         } else {
3019                                 mid = slot;
3020                                 if (mid != nritems &&
3021                                     leaf_space_used(l, mid, nritems - mid) +
3022                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
3023                                         if (data_size && !tried_avoid_double)
3024                                                 goto push_for_double;
3025                                         split = 2;
3026                                 }
3027                         }
3028                 }
3029         } else {
3030                 if (leaf_space_used(l, 0, mid) + data_size >
3031                         BTRFS_LEAF_DATA_SIZE(root)) {
3032                         if (!extend && data_size && slot == 0) {
3033                                 split = 0;
3034                         } else if ((extend || !data_size) && slot == 0) {
3035                                 mid = 1;
3036                         } else {
3037                                 mid = slot;
3038                                 if (mid != nritems &&
3039                                     leaf_space_used(l, mid, nritems - mid) +
3040                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
3041                                         if (data_size && !tried_avoid_double)
3042                                                 goto push_for_double;
3043                                         split = 2 ;
3044                                 }
3045                         }
3046                 }
3047         }
3048
3049         if (split == 0)
3050                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
3051         else
3052                 btrfs_item_key(l, &disk_key, mid);
3053
3054         right = btrfs_alloc_free_block(trans, root, root->leafsize, 0,
3055                                         root->root_key.objectid,
3056                                         &disk_key, 0, l->start, 0);
3057         if (IS_ERR(right))
3058                 return PTR_ERR(right);
3059
3060         root_add_used(root, root->leafsize);
3061
3062         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
3063         btrfs_set_header_bytenr(right, right->start);
3064         btrfs_set_header_generation(right, trans->transid);
3065         btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
3066         btrfs_set_header_owner(right, root->root_key.objectid);
3067         btrfs_set_header_level(right, 0);
3068         write_extent_buffer(right, root->fs_info->fsid,
3069                             (unsigned long)btrfs_header_fsid(right),
3070                             BTRFS_FSID_SIZE);
3071
3072         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
3073                             (unsigned long)btrfs_header_chunk_tree_uuid(right),
3074                             BTRFS_UUID_SIZE);
3075
3076         if (split == 0) {
3077                 if (mid <= slot) {
3078                         btrfs_set_header_nritems(right, 0);
3079                         wret = insert_ptr(trans, root, path,
3080                                           &disk_key, right->start,
3081                                           path->slots[1] + 1, 1);
3082                         if (wret)
3083                                 ret = wret;
3084
3085                         btrfs_tree_unlock(path->nodes[0]);
3086                         free_extent_buffer(path->nodes[0]);
3087                         path->nodes[0] = right;
3088                         path->slots[0] = 0;
3089                         path->slots[1] += 1;
3090                 } else {
3091                         btrfs_set_header_nritems(right, 0);
3092                         wret = insert_ptr(trans, root, path,
3093                                           &disk_key,
3094                                           right->start,
3095                                           path->slots[1], 1);
3096                         if (wret)
3097                                 ret = wret;
3098                         btrfs_tree_unlock(path->nodes[0]);
3099                         free_extent_buffer(path->nodes[0]);
3100                         path->nodes[0] = right;
3101                         path->slots[0] = 0;
3102                         if (path->slots[1] == 0) {
3103                                 wret = fixup_low_keys(trans, root,
3104                                                 path, &disk_key, 1);
3105                                 if (wret)
3106                                         ret = wret;
3107                         }
3108                 }
3109                 btrfs_mark_buffer_dirty(right);
3110                 return ret;
3111         }
3112
3113         ret = copy_for_split(trans, root, path, l, right, slot, mid, nritems);
3114         BUG_ON(ret);
3115
3116         if (split == 2) {
3117                 BUG_ON(num_doubles != 0);
3118                 num_doubles++;
3119                 goto again;
3120         }
3121
3122         return ret;
3123
3124 push_for_double:
3125         push_for_double_split(trans, root, path, data_size);
3126         tried_avoid_double = 1;
3127         if (btrfs_leaf_free_space(root, path->nodes[0]) >= data_size)
3128                 return 0;
3129         goto again;
3130 }
3131
3132 static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
3133                                          struct btrfs_root *root,
3134                                          struct btrfs_path *path, int ins_len)
3135 {
3136         struct btrfs_key key;
3137         struct extent_buffer *leaf;
3138         struct btrfs_file_extent_item *fi;
3139         u64 extent_len = 0;
3140         u32 item_size;
3141         int ret;
3142
3143         leaf = path->nodes[0];
3144         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3145
3146         BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
3147                key.type != BTRFS_EXTENT_CSUM_KEY);
3148
3149         if (btrfs_leaf_free_space(root, leaf) >= ins_len)
3150                 return 0;
3151
3152         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3153         if (key.type == BTRFS_EXTENT_DATA_KEY) {
3154                 fi = btrfs_item_ptr(leaf, path->slots[0],
3155                                     struct btrfs_file_extent_item);
3156                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
3157         }
3158         btrfs_release_path(root, path);
3159
3160         path->keep_locks = 1;
3161         path->search_for_split = 1;
3162         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
3163         path->search_for_split = 0;
3164         if (ret < 0)
3165                 goto err;
3166
3167         ret = -EAGAIN;
3168         leaf = path->nodes[0];
3169         /* if our item isn't there or got smaller, return now */
3170         if (ret > 0 || item_size != btrfs_item_size_nr(leaf, path->slots[0]))
3171                 goto err;
3172
3173         /* the leaf has  changed, it now has room.  return now */
3174         if (btrfs_leaf_free_space(root, path->nodes[0]) >= ins_len)
3175                 goto err;
3176
3177         if (key.type == BTRFS_EXTENT_DATA_KEY) {
3178                 fi = btrfs_item_ptr(leaf, path->slots[0],
3179                                     struct btrfs_file_extent_item);
3180                 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
3181                         goto err;
3182         }
3183
3184         btrfs_set_path_blocking(path);
3185         ret = split_leaf(trans, root, &key, path, ins_len, 1);
3186         if (ret)
3187                 goto err;
3188
3189         path->keep_locks = 0;
3190         btrfs_unlock_up_safe(path, 1);
3191         return 0;
3192 err:
3193         path->keep_locks = 0;
3194         return ret;
3195 }
3196
3197 static noinline int split_item(struct btrfs_trans_handle *trans,
3198                                struct btrfs_root *root,
3199                                struct btrfs_path *path,
3200                                struct btrfs_key *new_key,
3201                                unsigned long split_offset)
3202 {
3203         struct extent_buffer *leaf;
3204         struct btrfs_item *item;
3205         struct btrfs_item *new_item;
3206         int slot;
3207         char *buf;
3208         u32 nritems;
3209         u32 item_size;
3210         u32 orig_offset;
3211         struct btrfs_disk_key disk_key;
3212
3213         leaf = path->nodes[0];
3214         BUG_ON(btrfs_leaf_free_space(root, leaf) < sizeof(struct btrfs_item));
3215
3216         btrfs_set_path_blocking(path);
3217
3218         item = btrfs_item_nr(leaf, path->slots[0]);
3219         orig_offset = btrfs_item_offset(leaf, item);
3220         item_size = btrfs_item_size(leaf, item);
3221
3222         buf = kmalloc(item_size, GFP_NOFS);
3223         if (!buf)
3224                 return -ENOMEM;
3225
3226         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
3227                             path->slots[0]), item_size);
3228
3229         slot = path->slots[0] + 1;
3230         nritems = btrfs_header_nritems(leaf);
3231         if (slot != nritems) {
3232                 /* shift the items */
3233                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
3234                                 btrfs_item_nr_offset(slot),
3235                                 (nritems - slot) * sizeof(struct btrfs_item));
3236         }
3237
3238         btrfs_cpu_key_to_disk(&disk_key, new_key);
3239         btrfs_set_item_key(leaf, &disk_key, slot);
3240
3241         new_item = btrfs_item_nr(leaf, slot);
3242
3243         btrfs_set_item_offset(leaf, new_item, orig_offset);
3244         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
3245
3246         btrfs_set_item_offset(leaf, item,
3247                               orig_offset + item_size - split_offset);
3248         btrfs_set_item_size(leaf, item, split_offset);
3249
3250         btrfs_set_header_nritems(leaf, nritems + 1);
3251
3252         /* write the data for the start of the original item */
3253         write_extent_buffer(leaf, buf,
3254                             btrfs_item_ptr_offset(leaf, path->slots[0]),
3255                             split_offset);
3256
3257         /* write the data for the new item */
3258         write_extent_buffer(leaf, buf + split_offset,
3259                             btrfs_item_ptr_offset(leaf, slot),
3260                             item_size - split_offset);
3261         btrfs_mark_buffer_dirty(leaf);
3262
3263         BUG_ON(btrfs_leaf_free_space(root, leaf) < 0);
3264         kfree(buf);
3265         return 0;
3266 }
3267
3268 /*
3269  * This function splits a single item into two items,
3270  * giving 'new_key' to the new item and splitting the
3271  * old one at split_offset (from the start of the item).
3272  *
3273  * The path may be released by this operation.  After
3274  * the split, the path is pointing to the old item.  The
3275  * new item is going to be in the same node as the old one.
3276  *
3277  * Note, the item being split must be smaller enough to live alone on
3278  * a tree block with room for one extra struct btrfs_item
3279  *
3280  * This allows us to split the item in place, keeping a lock on the
3281  * leaf the entire time.
3282  */
3283 int btrfs_split_item(struct btrfs_trans_handle *trans,
3284                      struct btrfs_root *root,
3285                      struct btrfs_path *path,
3286                      struct btrfs_key *new_key,
3287                      unsigned long split_offset)
3288 {
3289         int ret;
3290         ret = setup_leaf_for_split(trans, root, path,
3291                                    sizeof(struct btrfs_item));
3292         if (ret)
3293                 return ret;
3294
3295         ret = split_item(trans, root, path, new_key, split_offset);
3296         return ret;
3297 }
3298
3299 /*
3300  * This function duplicate a item, giving 'new_key' to the new item.
3301  * It guarantees both items live in the same tree leaf and the new item
3302  * is contiguous with the original item.
3303  *
3304  * This allows us to split file extent in place, keeping a lock on the
3305  * leaf the entire time.
3306  */
3307 int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
3308                          struct btrfs_root *root,
3309                          struct btrfs_path *path,
3310                          struct btrfs_key *new_key)
3311 {
3312         struct extent_buffer *leaf;
3313         int ret;
3314         u32 item_size;
3315
3316         leaf = path->nodes[0];
3317         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3318         ret = setup_leaf_for_split(trans, root, path,
3319                                    item_size + sizeof(struct btrfs_item));
3320         if (ret)
3321                 return ret;
3322
3323         path->slots[0]++;
3324         ret = setup_items_for_insert(trans, root, path, new_key, &item_size,
3325                                      item_size, item_size +
3326                                      sizeof(struct btrfs_item), 1);
3327         BUG_ON(ret);
3328
3329         leaf = path->nodes[0];
3330         memcpy_extent_buffer(leaf,
3331                              btrfs_item_ptr_offset(leaf, path->slots[0]),
3332                              btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
3333                              item_size);
3334         return 0;
3335 }
3336
3337 /*
3338  * make the item pointed to by the path smaller.  new_size indicates
3339  * how small to make it, and from_end tells us if we just chop bytes
3340  * off the end of the item or if we shift the item to chop bytes off
3341  * the front.
3342  */
3343 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
3344                         struct btrfs_root *root,
3345                         struct btrfs_path *path,
3346                         u32 new_size, int from_end)
3347 {
3348         int ret = 0;
3349         int slot;
3350         struct extent_buffer *leaf;
3351         struct btrfs_item *item;
3352         u32 nritems;
3353         unsigned int data_end;
3354         unsigned int old_data_start;
3355         unsigned int old_size;
3356         unsigned int size_diff;
3357         int i;
3358
3359         leaf = path->nodes[0];
3360         slot = path->slots[0];
3361
3362         old_size = btrfs_item_size_nr(leaf, slot);
3363         if (old_size == new_size)
3364                 return 0;
3365
3366         nritems = btrfs_header_nritems(leaf);
3367         data_end = leaf_data_end(root, leaf);
3368
3369         old_data_start = btrfs_item_offset_nr(leaf, slot);
3370
3371         size_diff = old_size - new_size;
3372
3373         BUG_ON(slot < 0);
3374         BUG_ON(slot >= nritems);
3375
3376         /*
3377          * item0..itemN ... dataN.offset..dataN.size .. data0.size
3378          */
3379         /* first correct the data pointers */
3380         for (i = slot; i < nritems; i++) {
3381                 u32 ioff;
3382                 item = btrfs_item_nr(leaf, i);
3383
3384                 if (!leaf->map_token) {
3385                         map_extent_buffer(leaf, (unsigned long)item,
3386                                         sizeof(struct btrfs_item),
3387                                         &leaf->map_token, &leaf->kaddr,
3388                                         &leaf->map_start, &leaf->map_len,
3389                                         KM_USER1);
3390                 }
3391
3392                 ioff = btrfs_item_offset(leaf, item);
3393                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
3394         }
3395
3396         if (leaf->map_token) {
3397                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
3398                 leaf->map_token = NULL;
3399         }
3400
3401         /* shift the data */
3402         if (from_end) {
3403                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3404                               data_end + size_diff, btrfs_leaf_data(leaf) +
3405                               data_end, old_data_start + new_size - data_end);
3406         } else {
3407                 struct btrfs_disk_key disk_key;
3408                 u64 offset;
3409
3410                 btrfs_item_key(leaf, &disk_key, slot);
3411
3412                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
3413                         unsigned long ptr;
3414                         struct btrfs_file_extent_item *fi;
3415
3416                         fi = btrfs_item_ptr(leaf, slot,
3417                                             struct btrfs_file_extent_item);
3418                         fi = (struct btrfs_file_extent_item *)(
3419                              (unsigned long)fi - size_diff);
3420
3421                         if (btrfs_file_extent_type(leaf, fi) ==
3422                             BTRFS_FILE_EXTENT_INLINE) {
3423                                 ptr = btrfs_item_ptr_offset(leaf, slot);
3424                                 memmove_extent_buffer(leaf, ptr,
3425                                       (unsigned long)fi,
3426                                       offsetof(struct btrfs_file_extent_item,
3427                                                  disk_bytenr));
3428                         }
3429                 }
3430
3431                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3432                               data_end + size_diff, btrfs_leaf_data(leaf) +
3433                               data_end, old_data_start - data_end);
3434
3435                 offset = btrfs_disk_key_offset(&disk_key);
3436                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
3437                 btrfs_set_item_key(leaf, &disk_key, slot);
3438                 if (slot == 0)
3439                         fixup_low_keys(trans, root, path, &disk_key, 1);
3440         }
3441
3442         item = btrfs_item_nr(leaf, slot);
3443         btrfs_set_item_size(leaf, item, new_size);
3444         btrfs_mark_buffer_dirty(leaf);
3445
3446         ret = 0;
3447         if (btrfs_leaf_free_space(root, leaf) < 0) {
3448                 btrfs_print_leaf(root, leaf);
3449                 BUG();
3450         }
3451         return ret;
3452 }
3453
3454 /*
3455  * make the item pointed to by the path bigger, data_size is the new size.
3456  */
3457 int btrfs_extend_item(struct btrfs_trans_handle *trans,
3458                       struct btrfs_root *root, struct btrfs_path *path,
3459                       u32 data_size)
3460 {
3461         int ret = 0;
3462         int slot;
3463         struct extent_buffer *leaf;
3464         struct btrfs_item *item;
3465         u32 nritems;
3466         unsigned int data_end;
3467         unsigned int old_data;
3468         unsigned int old_size;
3469         int i;
3470
3471         leaf = path->nodes[0];
3472
3473         nritems = btrfs_header_nritems(leaf);
3474         data_end = leaf_data_end(root, leaf);
3475
3476         if (btrfs_leaf_free_space(root, leaf) < data_size) {
3477                 btrfs_print_leaf(root, leaf);
3478                 BUG();
3479         }
3480         slot = path->slots[0];
3481         old_data = btrfs_item_end_nr(leaf, slot);
3482
3483         BUG_ON(slot < 0);
3484         if (slot >= nritems) {
3485                 btrfs_print_leaf(root, leaf);
3486                 printk(KERN_CRIT "slot %d too large, nritems %d\n",
3487                        slot, nritems);
3488                 BUG_ON(1);
3489         }
3490
3491         /*
3492          * item0..itemN ... dataN.offset..dataN.size .. data0.size
3493          */
3494         /* first correct the data pointers */
3495         for (i = slot; i < nritems; i++) {
3496                 u32 ioff;
3497                 item = btrfs_item_nr(leaf, i);
3498
3499                 if (!leaf->map_token) {
3500                         map_extent_buffer(leaf, (unsigned long)item,
3501                                         sizeof(struct btrfs_item),
3502                                         &leaf->map_token, &leaf->kaddr,
3503                                         &leaf->map_start, &leaf->map_len,
3504                                         KM_USER1);
3505                 }
3506                 ioff = btrfs_item_offset(leaf, item);
3507                 btrfs_set_item_offset(leaf, item, ioff - data_size);
3508         }
3509
3510         if (leaf->map_token) {
3511                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
3512                 leaf->map_token = NULL;
3513         }
3514
3515         /* shift the data */
3516         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3517                       data_end - data_size, btrfs_leaf_data(leaf) +
3518                       data_end, old_data - data_end);
3519
3520         data_end = old_data;
3521         old_size = btrfs_item_size_nr(leaf, slot);
3522         item = btrfs_item_nr(leaf, slot);
3523         btrfs_set_item_size(leaf, item, old_size + data_size);
3524         btrfs_mark_buffer_dirty(leaf);
3525
3526         ret = 0;
3527         if (btrfs_leaf_free_space(root, leaf) < 0) {
3528                 btrfs_print_leaf(root, leaf);
3529                 BUG();
3530         }
3531         return ret;
3532 }
3533
3534 /*
3535  * Given a key and some data, insert items into the tree.
3536  * This does all the path init required, making room in the tree if needed.
3537  * Returns the number of keys that were inserted.
3538  */
3539 int btrfs_insert_some_items(struct btrfs_trans_handle *trans,
3540                             struct btrfs_root *root,
3541                             struct btrfs_path *path,
3542                             struct btrfs_key *cpu_key, u32 *data_size,
3543                             int nr)
3544 {
3545         struct extent_buffer *leaf;
3546         struct btrfs_item *item;
3547         int ret = 0;
3548         int slot;
3549         int i;
3550         u32 nritems;
3551         u32 total_data = 0;
3552         u32 total_size = 0;
3553         unsigned int data_end;
3554         struct btrfs_disk_key disk_key;
3555         struct btrfs_key found_key;
3556
3557         for (i = 0; i < nr; i++) {
3558                 if (total_size + data_size[i] + sizeof(struct btrfs_item) >
3559                     BTRFS_LEAF_DATA_SIZE(root)) {
3560                         break;
3561                         nr = i;
3562                 }
3563                 total_data += data_size[i];
3564                 total_size += data_size[i] + sizeof(struct btrfs_item);
3565         }
3566         BUG_ON(nr == 0);
3567
3568         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
3569         if (ret == 0)
3570                 return -EEXIST;
3571         if (ret < 0)
3572                 goto out;
3573
3574         leaf = path->nodes[0];
3575
3576         nritems = btrfs_header_nritems(leaf);
3577         data_end = leaf_data_end(root, leaf);
3578
3579         if (btrfs_leaf_free_space(root, leaf) < total_size) {
3580                 for (i = nr; i >= 0; i--) {
3581                         total_data -= data_size[i];
3582                         total_size -= data_size[i] + sizeof(struct btrfs_item);
3583                         if (total_size < btrfs_leaf_free_space(root, leaf))
3584                                 break;
3585                 }
3586                 nr = i;
3587         }
3588
3589         slot = path->slots[0];
3590         BUG_ON(slot < 0);
3591
3592         if (slot != nritems) {
3593                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
3594
3595                 item = btrfs_item_nr(leaf, slot);
3596                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3597
3598                 /* figure out how many keys we can insert in here */
3599                 total_data = data_size[0];
3600                 for (i = 1; i < nr; i++) {
3601                         if (btrfs_comp_cpu_keys(&found_key, cpu_key + i) <= 0)
3602                                 break;
3603                         total_data += data_size[i];
3604                 }
3605                 nr = i;
3606
3607                 if (old_data < data_end) {
3608                         btrfs_print_leaf(root, leaf);
3609                         printk(KERN_CRIT "slot %d old_data %d data_end %d\n",
3610                                slot, old_data, data_end);
3611                         BUG_ON(1);
3612                 }
3613                 /*
3614                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
3615                  */
3616                 /* first correct the data pointers */
3617                 WARN_ON(leaf->map_token);
3618                 for (i = slot; i < nritems; i++) {
3619                         u32 ioff;
3620
3621                         item = btrfs_item_nr(leaf, i);
3622                         if (!leaf->map_token) {
3623                                 map_extent_buffer(leaf, (unsigned long)item,
3624                                         sizeof(struct btrfs_item),
3625                                         &leaf->map_token, &leaf->kaddr,
3626                                         &leaf->map_start, &leaf->map_len,
3627                                         KM_USER1);
3628                         }
3629
3630                         ioff = btrfs_item_offset(leaf, item);
3631                         btrfs_set_item_offset(leaf, item, ioff - total_data);
3632                 }
3633                 if (leaf->map_token) {
3634                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
3635                         leaf->map_token = NULL;
3636                 }
3637
3638                 /* shift the items */
3639                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
3640                               btrfs_item_nr_offset(slot),
3641                               (nritems - slot) * sizeof(struct btrfs_item));
3642
3643                 /* shift the data */
3644                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3645                               data_end - total_data, btrfs_leaf_data(leaf) +
3646                               data_end, old_data - data_end);
3647                 data_end = old_data;
3648         } else {
3649                 /*
3650                  * this sucks but it has to be done, if we are inserting at
3651                  * the end of the leaf only insert 1 of the items, since we
3652                  * have no way of knowing whats on the next leaf and we'd have
3653                  * to drop our current locks to figure it out
3654                  */
3655                 nr = 1;
3656         }
3657
3658         /* setup the item for the new data */
3659         for (i = 0; i < nr; i++) {
3660                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
3661                 btrfs_set_item_key(leaf, &disk_key, slot + i);
3662                 item = btrfs_item_nr(leaf, slot + i);
3663                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
3664                 data_end -= data_size[i];
3665                 btrfs_set_item_size(leaf, item, data_size[i]);
3666         }
3667         btrfs_set_header_nritems(leaf, nritems + nr);
3668         btrfs_mark_buffer_dirty(leaf);
3669
3670         ret = 0;
3671         if (slot == 0) {
3672                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
3673                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
3674         }
3675
3676         if (btrfs_leaf_free_space(root, leaf) < 0) {
3677                 btrfs_print_leaf(root, leaf);
3678                 BUG();
3679         }
3680 out:
3681         if (!ret)
3682                 ret = nr;
3683         return ret;
3684 }
3685
3686 /*
3687  * this is a helper for btrfs_insert_empty_items, the main goal here is
3688  * to save stack depth by doing the bulk of the work in a function
3689  * that doesn't call btrfs_search_slot
3690  */
3691 static noinline_for_stack int
3692 setup_items_for_insert(struct btrfs_trans_handle *trans,
3693                       struct btrfs_root *root, struct btrfs_path *path,
3694                       struct btrfs_key *cpu_key, u32 *data_size,
3695                       u32 total_data, u32 total_size, int nr)
3696 {
3697         struct btrfs_item *item;
3698         int i;
3699         u32 nritems;
3700         unsigned int data_end;
3701         struct btrfs_disk_key disk_key;
3702         int ret;
3703         struct extent_buffer *leaf;
3704         int slot;
3705
3706         leaf = path->nodes[0];
3707         slot = path->slots[0];
3708
3709         nritems = btrfs_header_nritems(leaf);
3710         data_end = leaf_data_end(root, leaf);
3711
3712         if (btrfs_leaf_free_space(root, leaf) < total_size) {
3713                 btrfs_print_leaf(root, leaf);
3714                 printk(KERN_CRIT "not enough freespace need %u have %d\n",
3715                        total_size, btrfs_leaf_free_space(root, leaf));
3716                 BUG();
3717         }
3718
3719         if (slot != nritems) {
3720                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
3721
3722                 if (old_data < data_end) {
3723                         btrfs_print_leaf(root, leaf);
3724                         printk(KERN_CRIT "slot %d old_data %d data_end %d\n",
3725                                slot, old_data, data_end);
3726                         BUG_ON(1);
3727                 }
3728                 /*
3729                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
3730                  */
3731                 /* first correct the data pointers */
3732                 WARN_ON(leaf->map_token);
3733                 for (i = slot; i < nritems; i++) {
3734                         u32 ioff;
3735
3736                         item = btrfs_item_nr(leaf, i);
3737                         if (!leaf->map_token) {
3738                                 map_extent_buffer(leaf, (unsigned long)item,
3739                                         sizeof(struct btrfs_item),
3740                                         &leaf->map_token, &leaf->kaddr,
3741                                         &leaf->map_start, &leaf->map_len,
3742                                         KM_USER1);
3743                         }
3744
3745                         ioff = btrfs_item_offset(leaf, item);
3746                         btrfs_set_item_offset(leaf, item, ioff - total_data);
3747                 }
3748                 if (leaf->map_token) {
3749                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
3750                         leaf->map_token = NULL;
3751                 }
3752
3753                 /* shift the items */
3754                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
3755                               btrfs_item_nr_offset(slot),
3756                               (nritems - slot) * sizeof(struct btrfs_item));
3757
3758                 /* shift the data */
3759                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3760                               data_end - total_data, btrfs_leaf_data(leaf) +
3761                               data_end, old_data - data_end);
3762                 data_end = old_data;
3763         }
3764
3765         /* setup the item for the new data */
3766         for (i = 0; i < nr; i++) {
3767                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
3768                 btrfs_set_item_key(leaf, &disk_key, slot + i);
3769                 item = btrfs_item_nr(leaf, slot + i);
3770                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
3771                 data_end -= data_size[i];
3772                 btrfs_set_item_size(leaf, item, data_size[i]);
3773         }
3774
3775         btrfs_set_header_nritems(leaf, nritems + nr);
3776
3777         ret = 0;
3778         if (slot == 0) {
3779                 struct btrfs_disk_key disk_key;
3780                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
3781                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
3782         }
3783         btrfs_unlock_up_safe(path, 1);
3784         btrfs_mark_buffer_dirty(leaf);
3785
3786         if (btrfs_leaf_free_space(root, leaf) < 0) {
3787                 btrfs_print_leaf(root, leaf);
3788                 BUG();
3789         }
3790         return ret;
3791 }
3792
3793 /*
3794  * Given a key and some data, insert items into the tree.
3795  * This does all the path init required, making room in the tree if needed.
3796  */
3797 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
3798                             struct btrfs_root *root,
3799                             struct btrfs_path *path,
3800                             struct btrfs_key *cpu_key, u32 *data_size,
3801                             int nr)
3802 {
3803         int ret = 0;
3804         int slot;
3805         int i;
3806         u32 total_size = 0;
3807         u32 total_data = 0;
3808
3809         for (i = 0; i < nr; i++)
3810                 total_data += data_size[i];
3811
3812         total_size = total_data + (nr * sizeof(struct btrfs_item));
3813         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
3814         if (ret == 0)
3815                 return -EEXIST;
3816         if (ret < 0)
3817                 goto out;
3818
3819         slot = path->slots[0];
3820         BUG_ON(slot < 0);
3821
3822         ret = setup_items_for_insert(trans, root, path, cpu_key, data_size,
3823                                total_data, total_size, nr);
3824
3825 out:
3826         return ret;
3827 }
3828
3829 /*
3830  * Given a key and some data, insert an item into the tree.
3831  * This does all the path init required, making room in the tree if needed.
3832  */
3833 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
3834                       *root, struct btrfs_key *cpu_key, void *data, u32
3835                       data_size)
3836 {
3837         int ret = 0;
3838         struct btrfs_path *path;
3839         struct extent_buffer *leaf;
3840         unsigned long ptr;
3841
3842         path = btrfs_alloc_path();
3843         BUG_ON(!path);
3844         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
3845         if (!ret) {
3846                 leaf = path->nodes[0];
3847                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3848                 write_extent_buffer(leaf, data, ptr, data_size);
3849                 btrfs_mark_buffer_dirty(leaf);
3850         }
3851         btrfs_free_path(path);
3852         return ret;
3853 }
3854
3855 /*
3856  * delete the pointer from a given node.
3857  *
3858  * the tree should have been previously balanced so the deletion does not
3859  * empty a node.
3860  */
3861 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3862                    struct btrfs_path *path, int level, int slot)
3863 {
3864         struct extent_buffer *parent = path->nodes[level];
3865         u32 nritems;
3866         int ret = 0;
3867         int wret;
3868
3869         nritems = btrfs_header_nritems(parent);
3870         if (slot != nritems - 1) {
3871                 memmove_extent_buffer(parent,
3872                               btrfs_node_key_ptr_offset(slot),
3873                               btrfs_node_key_ptr_offset(slot + 1),
3874                               sizeof(struct btrfs_key_ptr) *
3875                               (nritems - slot - 1));
3876         }
3877         nritems--;
3878         btrfs_set_header_nritems(parent, nritems);
3879         if (nritems == 0 && parent == root->node) {
3880                 BUG_ON(btrfs_header_level(root->node) != 1);
3881                 /* just turn the root into a leaf and break */
3882                 btrfs_set_header_level(root->node, 0);
3883         } else if (slot == 0) {
3884                 struct btrfs_disk_key disk_key;
3885
3886                 btrfs_node_key(parent, &disk_key, 0);
3887                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
3888                 if (wret)
3889                         ret = wret;
3890         }
3891         btrfs_mark_buffer_dirty(parent);
3892         return ret;
3893 }
3894
3895 /*
3896  * a helper function to delete the leaf pointed to by path->slots[1] and
3897  * path->nodes[1].
3898  *
3899  * This deletes the pointer in path->nodes[1] and frees the leaf
3900  * block extent.  zero is returned if it all worked out, < 0 otherwise.
3901  *
3902  * The path must have already been setup for deleting the leaf, including
3903  * all the proper balancing.  path->nodes[1] must be locked.
3904  */
3905 static noinline int btrfs_del_leaf(struct btrfs_trans_handle *trans,
3906                                    struct btrfs_root *root,
3907                                    struct btrfs_path *path,
3908                                    struct extent_buffer *leaf)
3909 {
3910         int ret;
3911
3912         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
3913         ret = del_ptr(trans, root, path, 1, path->slots[1]);
3914         if (ret)
3915                 return ret;
3916
3917         /*
3918          * btrfs_free_extent is expensive, we want to make sure we
3919          * aren't holding any locks when we call it
3920          */
3921         btrfs_unlock_up_safe(path, 0);
3922
3923         root_sub_used(root, leaf->len);
3924
3925         btrfs_free_tree_block(trans, root, leaf, 0, 1);
3926         return 0;
3927 }
3928 /*
3929  * delete the item at the leaf level in path.  If that empties
3930  * the leaf, remove it from the tree
3931  */
3932 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3933                     struct btrfs_path *path, int slot, int nr)
3934 {
3935         struct extent_buffer *leaf;
3936         struct btrfs_item *item;
3937         int last_off;
3938         int dsize = 0;
3939         int ret = 0;
3940         int wret;
3941         int i;
3942         u32 nritems;
3943
3944         leaf = path->nodes[0];
3945         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
3946
3947         for (i = 0; i < nr; i++)
3948                 dsize += btrfs_item_size_nr(leaf, slot + i);
3949
3950         nritems = btrfs_header_nritems(leaf);
3951
3952         if (slot + nr != nritems) {
3953                 int data_end = leaf_data_end(root, leaf);
3954
3955                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3956                               data_end + dsize,
3957                               btrfs_leaf_data(leaf) + data_end,
3958                               last_off - data_end);
3959
3960                 for (i = slot + nr; i < nritems; i++) {
3961                         u32 ioff;
3962
3963                         item = btrfs_item_nr(leaf, i);
3964                         if (!leaf->map_token) {
3965                                 map_extent_buffer(leaf, (unsigned long)item,
3966                                         sizeof(struct btrfs_item),
3967                                         &leaf->map_token, &leaf->kaddr,
3968                                         &leaf->map_start, &leaf->map_len,
3969                                         KM_USER1);
3970                         }
3971                         ioff = btrfs_item_offset(leaf, item);
3972                         btrfs_set_item_offset(leaf, item, ioff + dsize);
3973                 }
3974
3975                 if (leaf->map_token) {
3976                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
3977                         leaf->map_token = NULL;
3978                 }
3979
3980                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
3981                               btrfs_item_nr_offset(slot + nr),
3982                               sizeof(struct btrfs_item) *
3983                               (nritems - slot - nr));
3984         }
3985         btrfs_set_header_nritems(leaf, nritems - nr);
3986         nritems -= nr;
3987
3988         /* delete the leaf if we've emptied it */
3989         if (nritems == 0) {
3990                 if (leaf == root->node) {
3991                         btrfs_set_header_level(leaf, 0);
3992                 } else {
3993                         btrfs_set_path_blocking(path);
3994                         clean_tree_block(trans, root, leaf);
3995                         ret = btrfs_del_leaf(trans, root, path, leaf);
3996                         BUG_ON(ret);
3997                 }
3998         } else {
3999                 int used = leaf_space_used(leaf, 0, nritems);
4000                 if (slot == 0) {
4001                         struct btrfs_disk_key disk_key;
4002
4003                         btrfs_item_key(leaf, &disk_key, 0);
4004                         wret = fixup_low_keys(trans, root, path,
4005                                               &disk_key, 1);
4006                         if (wret)
4007                                 ret = wret;
4008                 }
4009
4010                 /* delete the leaf if it is mostly empty */
4011                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 3) {
4012                         /* push_leaf_left fixes the path.
4013                          * make sure the path still points to our leaf
4014                          * for possible call to del_ptr below
4015                          */
4016                         slot = path->slots[1];
4017                         extent_buffer_get(leaf);
4018
4019                         btrfs_set_path_blocking(path);
4020                         wret = push_leaf_left(trans, root, path, 1, 1,
4021                                               1, (u32)-1);
4022                         if (wret < 0 && wret != -ENOSPC)
4023                                 ret = wret;
4024
4025                         if (path->nodes[0] == leaf &&
4026                             btrfs_header_nritems(leaf)) {
4027                                 wret = push_leaf_right(trans, root, path, 1,
4028                                                        1, 1, 0);
4029                                 if (wret < 0 && wret != -ENOSPC)
4030                                         ret = wret;
4031                         }
4032
4033                         if (btrfs_header_nritems(leaf) == 0) {
4034                                 path->slots[1] = slot;
4035                                 ret = btrfs_del_leaf(trans, root, path, leaf);
4036                                 BUG_ON(ret);
4037                                 free_extent_buffer(leaf);
4038                         } else {
4039                                 /* if we're still in the path, make sure
4040                                  * we're dirty.  Otherwise, one of the
4041                                  * push_leaf functions must have already
4042                                  * dirtied this buffer
4043                                  */
4044                                 if (path->nodes[0] == leaf)
4045                                         btrfs_mark_buffer_dirty(leaf);
4046                                 free_extent_buffer(leaf);
4047                         }
4048                 } else {
4049                         btrfs_mark_buffer_dirty(leaf);
4050                 }
4051         }
4052         return ret;
4053 }
4054
4055 /*
4056  * search the tree again to find a leaf with lesser keys
4057  * returns 0 if it found something or 1 if there are no lesser leaves.
4058  * returns < 0 on io errors.
4059  *
4060  * This may release the path, and so you may lose any locks held at the
4061  * time you call it.
4062  */
4063 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
4064 {
4065         struct btrfs_key key;
4066         struct btrfs_disk_key found_key;
4067         int ret;
4068
4069         btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
4070
4071         if (key.offset > 0)
4072                 key.offset--;
4073         else if (key.type > 0)
4074                 key.type--;
4075         else if (key.objectid > 0)
4076                 key.objectid--;
4077         else
4078                 return 1;
4079
4080         btrfs_release_path(root, path);
4081         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4082         if (ret < 0)
4083                 return ret;
4084         btrfs_item_key(path->nodes[0], &found_key, 0);
4085         ret = comp_keys(&found_key, &key);
4086         if (ret < 0)
4087                 return 0;
4088         return 1;
4089 }
4090
4091 /*
4092  * A helper function to walk down the tree starting at min_key, and looking
4093  * for nodes or leaves that are either in cache or have a minimum
4094  * transaction id.  This is used by the btree defrag code, and tree logging
4095  *
4096  * This does not cow, but it does stuff the starting key it finds back
4097  * into min_key, so you can call btrfs_search_slot with cow=1 on the
4098  * key and get a writable path.
4099  *
4100  * This does lock as it descends, and path->keep_locks should be set
4101  * to 1 by the caller.
4102  *
4103  * This honors path->lowest_level to prevent descent past a given level
4104  * of the tree.
4105  *
4106  * min_trans indicates the oldest transaction that you are interested
4107  * in walking through.  Any nodes or leaves older than min_trans are
4108  * skipped over (without reading them).
4109  *
4110  * returns zero if something useful was found, < 0 on error and 1 if there
4111  * was nothing in the tree that matched the search criteria.
4112  */
4113 int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
4114                          struct btrfs_key *max_key,
4115                          struct btrfs_path *path, int cache_only,
4116                          u64 min_trans)
4117 {
4118         struct extent_buffer *cur;
4119         struct btrfs_key found_key;
4120         int slot;
4121         int sret;
4122         u32 nritems;
4123         int level;
4124         int ret = 1;
4125
4126         WARN_ON(!path->keep_locks);
4127 again:
4128         cur = btrfs_lock_root_node(root);
4129         level = btrfs_header_level(cur);
4130         WARN_ON(path->nodes[level]);
4131         path->nodes[level] = cur;
4132         path->locks[level] = 1;
4133
4134         if (btrfs_header_generation(cur) < min_trans) {
4135                 ret = 1;
4136                 goto out;
4137         }
4138         while (1) {
4139                 nritems = btrfs_header_nritems(cur);
4140                 level = btrfs_header_level(cur);
4141                 sret = bin_search(cur, min_key, level, &slot);
4142
4143                 /* at the lowest level, we're done, setup the path and exit */
4144                 if (level == path->lowest_level) {
4145                         if (slot >= nritems)
4146                                 goto find_next_key;
4147                         ret = 0;
4148                         path->slots[level] = slot;
4149                         btrfs_item_key_to_cpu(cur, &found_key, slot);
4150                         goto out;
4151                 }
4152                 if (sret && slot > 0)
4153                         slot--;
4154                 /*
4155                  * check this node pointer against the cache_only and
4156                  * min_trans parameters.  If it isn't in cache or is too
4157                  * old, skip to the next one.
4158                  */
4159                 while (slot < nritems) {
4160                         u64 blockptr;
4161                         u64 gen;
4162                         struct extent_buffer *tmp;
4163                         struct btrfs_disk_key disk_key;
4164
4165                         blockptr = btrfs_node_blockptr(cur, slot);
4166                         gen = btrfs_node_ptr_generation(cur, slot);
4167                         if (gen < min_trans) {
4168                                 slot++;
4169                                 continue;
4170                         }
4171                         if (!cache_only)
4172                                 break;
4173
4174                         if (max_key) {
4175                                 btrfs_node_key(cur, &disk_key, slot);
4176                                 if (comp_keys(&disk_key, max_key) >= 0) {
4177                                         ret = 1;
4178                                         goto out;
4179                                 }
4180                         }
4181
4182                         tmp = btrfs_find_tree_block(root, blockptr,
4183                                             btrfs_level_size(root, level - 1));
4184
4185                         if (tmp && btrfs_buffer_uptodate(tmp, gen)) {
4186                                 free_extent_buffer(tmp);
4187                                 break;
4188                         }
4189                         if (tmp)
4190                                 free_extent_buffer(tmp);
4191                         slot++;
4192                 }
4193 find_next_key:
4194                 /*
4195                  * we didn't find a candidate key in this node, walk forward
4196                  * and find another one
4197                  */
4198                 if (slot >= nritems) {
4199                         path->slots[level] = slot;
4200                         btrfs_set_path_blocking(path);
4201                         sret = btrfs_find_next_key(root, path, min_key, level,
4202                                                   cache_only, min_trans);
4203                         if (sret == 0) {
4204                                 btrfs_release_path(root, path);
4205                                 goto again;
4206                         } else {
4207                                 goto out;
4208                         }
4209                 }
4210                 /* save our key for returning back */
4211                 btrfs_node_key_to_cpu(cur, &found_key, slot);
4212                 path->slots[level] = slot;
4213                 if (level == path->lowest_level) {
4214                         ret = 0;
4215                         unlock_up(path, level, 1);
4216                         goto out;
4217                 }
4218                 btrfs_set_path_blocking(path);
4219                 cur = read_node_slot(root, cur, slot);
4220
4221                 btrfs_tree_lock(cur);
4222
4223                 path->locks[level - 1] = 1;
4224                 path->nodes[level - 1] = cur;
4225                 unlock_up(path, level, 1);
4226                 btrfs_clear_path_blocking(path, NULL);
4227         }
4228 out:
4229         if (ret == 0)
4230                 memcpy(min_key, &found_key, sizeof(found_key));
4231         btrfs_set_path_blocking(path);
4232         return ret;
4233 }
4234
4235 /*
4236  * this is similar to btrfs_next_leaf, but does not try to preserve
4237  * and fixup the path.  It looks for and returns the next key in the
4238  * tree based on the current path and the cache_only and min_trans
4239  * parameters.
4240  *
4241  * 0 is returned if another key is found, < 0 if there are any errors
4242  * and 1 is returned if there are no higher keys in the tree
4243  *
4244  * path->keep_locks should be set to 1 on the search made before
4245  * calling this function.
4246  */
4247 int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
4248                         struct btrfs_key *key, int level,
4249                         int cache_only, u64 min_trans)
4250 {
4251         int slot;
4252         struct extent_buffer *c;
4253
4254         WARN_ON(!path->keep_locks);
4255         while (level < BTRFS_MAX_LEVEL) {
4256                 if (!path->nodes[level])
4257                         return 1;
4258
4259                 slot = path->slots[level] + 1;
4260                 c = path->nodes[level];
4261 next:
4262                 if (slot >= btrfs_header_nritems(c)) {
4263                         int ret;
4264                         int orig_lowest;
4265                         struct btrfs_key cur_key;
4266                         if (level + 1 >= BTRFS_MAX_LEVEL ||
4267                             !path->nodes[level + 1])
4268                                 return 1;
4269
4270                         if (path->locks[level + 1]) {
4271                                 level++;
4272                                 continue;
4273                         }
4274
4275                         slot = btrfs_header_nritems(c) - 1;
4276                         if (level == 0)
4277                                 btrfs_item_key_to_cpu(c, &cur_key, slot);
4278                         else
4279                                 btrfs_node_key_to_cpu(c, &cur_key, slot);
4280
4281                         orig_lowest = path->lowest_level;
4282                         btrfs_release_path(root, path);
4283                         path->lowest_level = level;
4284                         ret = btrfs_search_slot(NULL, root, &cur_key, path,
4285                                                 0, 0);
4286                         path->lowest_level = orig_lowest;
4287                         if (ret < 0)
4288                                 return ret;
4289
4290                         c = path->nodes[level];
4291                         slot = path->slots[level];
4292                         if (ret == 0)
4293                                 slot++;
4294                         goto next;
4295                 }
4296
4297                 if (level == 0)
4298                         btrfs_item_key_to_cpu(c, key, slot);
4299                 else {
4300                         u64 blockptr = btrfs_node_blockptr(c, slot);
4301                         u64 gen = btrfs_node_ptr_generation(c, slot);
4302
4303                         if (cache_only) {
4304                                 struct extent_buffer *cur;
4305                                 cur = btrfs_find_tree_block(root, blockptr,
4306                                             btrfs_level_size(root, level - 1));
4307                                 if (!cur || !btrfs_buffer_uptodate(cur, gen)) {
4308                                         slot++;
4309                                         if (cur)
4310                                                 free_extent_buffer(cur);
4311                                         goto next;
4312                                 }
4313                                 free_extent_buffer(cur);
4314                         }
4315                         if (gen < min_trans) {
4316                                 slot++;
4317                                 goto next;
4318                         }
4319                         btrfs_node_key_to_cpu(c, key, slot);
4320                 }
4321                 return 0;
4322         }
4323         return 1;
4324 }
4325
4326 /*
4327  * search the tree again to find a leaf with greater keys
4328  * returns 0 if it found something or 1 if there are no greater leaves.
4329  * returns < 0 on io errors.
4330  */
4331 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
4332 {
4333         int slot;
4334         int level;
4335         struct extent_buffer *c;
4336         struct extent_buffer *next;
4337         struct btrfs_key key;
4338         u32 nritems;
4339         int ret;
4340         int old_spinning = path->leave_spinning;
4341         int force_blocking = 0;
4342
4343         nritems = btrfs_header_nritems(path->nodes[0]);
4344         if (nritems == 0)
4345                 return 1;
4346
4347         /*
4348          * we take the blocks in an order that upsets lockdep.  Using
4349          * blocking mode is the only way around it.
4350          */
4351 #ifdef CONFIG_DEBUG_LOCK_ALLOC
4352         force_blocking = 1;
4353 #endif
4354
4355         btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
4356 again:
4357         level = 1;
4358         next = NULL;
4359         btrfs_release_path(root, path);
4360
4361         path->keep_locks = 1;
4362
4363         if (!force_blocking)
4364                 path->leave_spinning = 1;
4365
4366         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4367         path->keep_locks = 0;
4368
4369         if (ret < 0)
4370                 return ret;
4371
4372         nritems = btrfs_header_nritems(path->nodes[0]);
4373         /*
4374          * by releasing the path above we dropped all our locks.  A balance
4375          * could have added more items next to the key that used to be
4376          * at the very end of the block.  So, check again here and
4377          * advance the path if there are now more items available.
4378          */
4379         if (nritems > 0 && path->slots[0] < nritems - 1) {
4380                 if (ret == 0)
4381                         path->slots[0]++;
4382                 ret = 0;
4383                 goto done;
4384         }
4385
4386         while (level < BTRFS_MAX_LEVEL) {
4387                 if (!path->nodes[level]) {
4388                         ret = 1;
4389                         goto done;
4390                 }
4391
4392                 slot = path->slots[level] + 1;
4393                 c = path->nodes[level];
4394                 if (slot >= btrfs_header_nritems(c)) {
4395                         level++;
4396                         if (level == BTRFS_MAX_LEVEL) {
4397                                 ret = 1;
4398                                 goto done;
4399                         }
4400                         continue;
4401                 }
4402
4403                 if (next) {
4404                         btrfs_tree_unlock(next);
4405                         free_extent_buffer(next);
4406                 }
4407
4408                 next = c;
4409                 ret = read_block_for_search(NULL, root, path, &next, level,
4410                                             slot, &key);
4411                 if (ret == -EAGAIN)
4412                         goto again;
4413
4414                 if (ret < 0) {
4415                         btrfs_release_path(root, path);
4416                         goto done;
4417                 }
4418
4419                 if (!path->skip_locking) {
4420                         ret = btrfs_try_spin_lock(next);
4421                         if (!ret) {
4422                                 btrfs_set_path_blocking(path);
4423                                 btrfs_tree_lock(next);
4424                                 if (!force_blocking)
4425                                         btrfs_clear_path_blocking(path, next);
4426                         }
4427                         if (force_blocking)
4428                                 btrfs_set_lock_blocking(next);
4429                 }
4430                 break;
4431         }
4432         path->slots[level] = slot;
4433         while (1) {
4434                 level--;
4435                 c = path->nodes[level];
4436                 if (path->locks[level])
4437                         btrfs_tree_unlock(c);
4438
4439                 free_extent_buffer(c);
4440                 path->nodes[level] = next;
4441                 path->slots[level] = 0;
4442                 if (!path->skip_locking)
4443                         path->locks[level] = 1;
4444
4445                 if (!level)
4446                         break;
4447
4448                 ret = read_block_for_search(NULL, root, path, &next, level,
4449                                             0, &key);
4450                 if (ret == -EAGAIN)
4451                         goto again;
4452
4453                 if (ret < 0) {
4454                         btrfs_release_path(root, path);
4455                         goto done;
4456                 }
4457
4458                 if (!path->skip_locking) {
4459                         btrfs_assert_tree_locked(path->nodes[level]);
4460                         ret = btrfs_try_spin_lock(next);
4461                         if (!ret) {
4462                                 btrfs_set_path_blocking(path);
4463                                 btrfs_tree_lock(next);
4464                                 if (!force_blocking)
4465                                         btrfs_clear_path_blocking(path, next);
4466                         }
4467                         if (force_blocking)
4468                                 btrfs_set_lock_blocking(next);
4469                 }
4470         }
4471         ret = 0;
4472 done:
4473         unlock_up(path, 0, 1);
4474         path->leave_spinning = old_spinning;
4475         if (!old_spinning)
4476                 btrfs_set_path_blocking(path);
4477
4478         return ret;
4479 }
4480
4481 /*
4482  * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
4483  * searching until it gets past min_objectid or finds an item of 'type'
4484  *
4485  * returns 0 if something is found, 1 if nothing was found and < 0 on error
4486  */
4487 int btrfs_previous_item(struct btrfs_root *root,
4488                         struct btrfs_path *path, u64 min_objectid,
4489                         int type)
4490 {
4491         struct btrfs_key found_key;
4492         struct extent_buffer *leaf;
4493         u32 nritems;
4494         int ret;
4495
4496         while (1) {
4497                 if (path->slots[0] == 0) {
4498                         btrfs_set_path_blocking(path);
4499                         ret = btrfs_prev_leaf(root, path);
4500                         if (ret != 0)
4501                                 return ret;
4502                 } else {
4503                         path->slots[0]--;
4504                 }
4505                 leaf = path->nodes[0];
4506                 nritems = btrfs_header_nritems(leaf);
4507                 if (nritems == 0)
4508                         return 1;
4509                 if (path->slots[0] == nritems)
4510                         path->slots[0]--;
4511
4512                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4513                 if (found_key.objectid < min_objectid)
4514                         break;
4515                 if (found_key.type == type)
4516                         return 0;
4517                 if (found_key.objectid == min_objectid &&
4518                     found_key.type < type)
4519                         break;
4520         }
4521         return 1;
4522 }