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