fs/btrfs/ctree.c

   1 /*
   2  * Copyright (C) 2007,2008 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/sched.h>
  20 #include <linux/slab.h>
  21 #include "ctree.h"
  22 #include "disk-io.h"
  23 #include "transaction.h"
  24 #include "print-tree.h"
  25 #include "locking.h"
  26
  27 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  28                       *root, struct btrfs_path *path, int level);
  29 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  30                       *root, struct btrfs_key *ins_key,
  31                       struct btrfs_path *path, int data_size, int extend);
  32 static int push_node_left(struct btrfs_trans_handle *trans,
  33                           struct btrfs_root *root, struct extent_buffer *dst,
  34                           struct extent_buffer *src, int empty);
  35 static int balance_node_right(struct btrfs_trans_handle *trans,
  36                               struct btrfs_root *root,
  37                               struct extent_buffer *dst_buf,
  38                               struct extent_buffer *src_buf);
  39 static void del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  40                    struct btrfs_path *path, int level, int slot);
  41
  42 struct btrfs_path *btrfs_alloc_path(void)
  43 {
  44         struct btrfs_path *path;
  45         path = kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
  46         return path;
  47 }
  48
  49 /*
  50  * set all locked nodes in the path to blocking locks.  This should
  51  * be done before scheduling
  52  */
  53 noinline void btrfs_set_path_blocking(struct btrfs_path *p)
  54 {
  55         int i;
  56         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  57                 if (!p->nodes[i] || !p->locks[i])
  58                         continue;
  59                 btrfs_set_lock_blocking_rw(p->nodes[i], p->locks[i]);
  60                 if (p->locks[i] == BTRFS_READ_LOCK)
  61                         p->locks[i] = BTRFS_READ_LOCK_BLOCKING;
  62                 else if (p->locks[i] == BTRFS_WRITE_LOCK)
  63                         p->locks[i] = BTRFS_WRITE_LOCK_BLOCKING;
  64         }
  65 }
  66
  67 /*
  68  * reset all the locked nodes in the patch to spinning locks.
  69  *
  70  * held is used to keep lockdep happy, when lockdep is enabled
  71  * we set held to a blocking lock before we go around and
  72  * retake all the spinlocks in the path.  You can safely use NULL
  73  * for held
  74  */
  75 noinline void btrfs_clear_path_blocking(struct btrfs_path *p,
  76                                         struct extent_buffer *held, int held_rw)
  77 {
  78         int i;
  79
  80 #ifdef CONFIG_DEBUG_LOCK_ALLOC
  81         /* lockdep really cares that we take all of these spinlocks
  82          * in the right order.  If any of the locks in the path are not
  83          * currently blocking, it is going to complain.  So, make really
  84          * really sure by forcing the path to blocking before we clear
  85          * the path blocking.
  86          */
  87         if (held) {
  88                 btrfs_set_lock_blocking_rw(held, held_rw);
  89                 if (held_rw == BTRFS_WRITE_LOCK)
  90                         held_rw = BTRFS_WRITE_LOCK_BLOCKING;
  91                 else if (held_rw == BTRFS_READ_LOCK)
  92                         held_rw = BTRFS_READ_LOCK_BLOCKING;
  93         }
  94         btrfs_set_path_blocking(p);
  95 #endif
  96
  97         for (i = BTRFS_MAX_LEVEL - 1; i >= 0; i--) {
  98                 if (p->nodes[i] && p->locks[i]) {
  99                         btrfs_clear_lock_blocking_rw(p->nodes[i], p->locks[i]);
 100                         if (p->locks[i] == BTRFS_WRITE_LOCK_BLOCKING)
 101                                 p->locks[i] = BTRFS_WRITE_LOCK;
 102                         else if (p->locks[i] == BTRFS_READ_LOCK_BLOCKING)
 103                                 p->locks[i] = BTRFS_READ_LOCK;
 104                 }
 105         }
 106
 107 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 108         if (held)
 109                 btrfs_clear_lock_blocking_rw(held, held_rw);
 110 #endif
 111 }
 112
 113 /* this also releases the path */
 114 void btrfs_free_path(struct btrfs_path *p)
 115 {
 116         if (!p)
 117                 return;
 118         btrfs_release_path(p);
 119         kmem_cache_free(btrfs_path_cachep, p);
 120 }
 121
 122 /*
 123  * path release drops references on the extent buffers in the path
 124  * and it drops any locks held by this path
 125  *
 126  * It is safe to call this on paths that no locks or extent buffers held.
 127  */
 128 noinline void btrfs_release_path(struct btrfs_path *p)
 129 {
 130         int i;
 131
 132         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
 133                 p->slots[i] = 0;
 134                 if (!p->nodes[i])
 135                         continue;
 136                 if (p->locks[i]) {
 137                         btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
 138                         p->locks[i] = 0;
 139                 }
 140                 free_extent_buffer(p->nodes[i]);
 141                 p->nodes[i] = NULL;
 142         }
 143 }
 144
 145 /*
 146  * safely gets a reference on the root node of a tree.  A lock
 147  * is not taken, so a concurrent writer may put a different node
 148  * at the root of the tree.  See btrfs_lock_root_node for the
 149  * looping required.
 150  *
 151  * The extent buffer returned by this has a reference taken, so
 152  * it won't disappear.  It may stop being the root of the tree
 153  * at any time because there are no locks held.
 154  */
 155 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
 156 {
 157         struct extent_buffer *eb;
 158
 159         while (1) {
 160                 rcu_read_lock();
 161                 eb = rcu_dereference(root->node);
 162
 163                 /*
 164                  * RCU really hurts here, we could free up the root node because
 165                  * it was cow'ed but we may not get the new root node yet so do
 166                  * the inc_not_zero dance and if it doesn't work then
 167                  * synchronize_rcu and try again.
 168                  */
 169                 if (atomic_inc_not_zero(&eb->refs)) {
 170                         rcu_read_unlock();
 171                         break;
 172                 }
 173                 rcu_read_unlock();
 174                 synchronize_rcu();
 175         }
 176         return eb;
 177 }
 178
 179 /* loop around taking references on and locking the root node of the
 180  * tree until you end up with a lock on the root.  A locked buffer
 181  * is returned, with a reference held.
 182  */
 183 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
 184 {
 185         struct extent_buffer *eb;
 186
 187         while (1) {
 188                 eb = btrfs_root_node(root);
 189                 btrfs_tree_lock(eb);
 190                 if (eb == root->node)
 191                         break;
 192                 btrfs_tree_unlock(eb);
 193                 free_extent_buffer(eb);
 194         }
 195         return eb;
 196 }
 197
 198 /* loop around taking references on and locking the root node of the
 199  * tree until you end up with a lock on the root.  A locked buffer
 200  * is returned, with a reference held.
 201  */
 202 struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
 203 {
 204         struct extent_buffer *eb;
 205
 206         while (1) {
 207                 eb = btrfs_root_node(root);
 208                 btrfs_tree_read_lock(eb);
 209                 if (eb == root->node)
 210                         break;
 211                 btrfs_tree_read_unlock(eb);
 212                 free_extent_buffer(eb);
 213         }
 214         return eb;
 215 }
 216
 217 /* cowonly root (everything not a reference counted cow subvolume), just get
 218  * put onto a simple dirty list.  transaction.c walks this to make sure they
 219  * get properly updated on disk.
 220  */
 221 static void add_root_to_dirty_list(struct btrfs_root *root)
 222 {
 223         if (root->track_dirty && list_empty(&root->dirty_list)) {
 224                 list_add(&root->dirty_list,
 225                          &root->fs_info->dirty_cowonly_roots);
 226         }
 227 }
 228
 229 /*
 230  * used by snapshot creation to make a copy of a root for a tree with
 231  * a given objectid.  The buffer with the new root node is returned in
 232  * cow_ret, and this func returns zero on success or a negative error code.
 233  */
 234 int btrfs_copy_root(struct btrfs_trans_handle *trans,
 235                       struct btrfs_root *root,
 236                       struct extent_buffer *buf,
 237                       struct extent_buffer **cow_ret, u64 new_root_objectid)
 238 {
 239         struct extent_buffer *cow;
 240         int ret = 0;
 241         int level;
 242         struct btrfs_disk_key disk_key;
 243
 244         WARN_ON(root->ref_cows && trans->transid !=
 245                 root->fs_info->running_transaction->transid);
 246         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 247
 248         level = btrfs_header_level(buf);
 249         if (level == 0)
 250                 btrfs_item_key(buf, &disk_key, 0);
 251         else
 252                 btrfs_node_key(buf, &disk_key, 0);
 253
 254         cow = btrfs_alloc_free_block(trans, root, buf->len, 0,
 255                                      new_root_objectid, &disk_key, level,
 256                                      buf->start, 0, 1);
 257         if (IS_ERR(cow))
 258                 return PTR_ERR(cow);
 259
 260         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 261         btrfs_set_header_bytenr(cow, cow->start);
 262         btrfs_set_header_generation(cow, trans->transid);
 263         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 264         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 265                                      BTRFS_HEADER_FLAG_RELOC);
 266         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 267                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 268         else
 269                 btrfs_set_header_owner(cow, new_root_objectid);
 270
 271         write_extent_buffer(cow, root->fs_info->fsid,
 272                             (unsigned long)btrfs_header_fsid(cow),
 273                             BTRFS_FSID_SIZE);
 274
 275         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 276         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 277                 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
 278         else
 279                 ret = btrfs_inc_ref(trans, root, cow, 0, 1);
 280
 281         if (ret)
 282                 return ret;
 283
 284         btrfs_mark_buffer_dirty(cow);
 285         *cow_ret = cow;
 286         return 0;
 287 }
 288
 289 /*
 290  * check if the tree block can be shared by multiple trees
 291  */
 292 int btrfs_block_can_be_shared(struct btrfs_root *root,
 293                               struct extent_buffer *buf)
 294 {
 295         /*
 296          * Tree blocks not in refernece counted trees and tree roots
 297          * are never shared. If a block was allocated after the last
 298          * snapshot and the block was not allocated by tree relocation,
 299          * we know the block is not shared.
 300          */
 301         if (root->ref_cows &&
 302             buf != root->node && buf != root->commit_root &&
 303             (btrfs_header_generation(buf) <=
 304              btrfs_root_last_snapshot(&root->root_item) ||
 305              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 306                 return 1;
 307 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 308         if (root->ref_cows &&
 309             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 310                 return 1;
 311 #endif
 312         return 0;
 313 }
 314
 315 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
 316                                        struct btrfs_root *root,
 317                                        struct extent_buffer *buf,
 318                                        struct extent_buffer *cow,
 319                                        int *last_ref)
 320 {
 321         u64 refs;
 322         u64 owner;
 323         u64 flags;
 324         u64 new_flags = 0;
 325         int ret;
 326
 327         /*
 328          * Backrefs update rules:
 329          *
 330          * Always use full backrefs for extent pointers in tree block
 331          * allocated by tree relocation.
 332          *
 333          * If a shared tree block is no longer referenced by its owner
 334          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
 335          * use full backrefs for extent pointers in tree block.
 336          *
 337          * If a tree block is been relocating
 338          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
 339          * use full backrefs for extent pointers in tree block.
 340          * The reason for this is some operations (such as drop tree)
 341          * are only allowed for blocks use full backrefs.
 342          */
 343
 344         if (btrfs_block_can_be_shared(root, buf)) {
 345                 ret = btrfs_lookup_extent_info(trans, root, buf->start,
 346                                                buf->len, &refs, &flags);
 347                 if (ret)
 348                         return ret;
 349                 if (refs == 0) {
 350                         ret = -EROFS;
 351                         btrfs_std_error(root->fs_info, ret);
 352                         return ret;
 353                 }
 354         } else {
 355                 refs = 1;
 356                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 357                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 358                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 359                 else
 360                         flags = 0;
 361         }
 362
 363         owner = btrfs_header_owner(buf);
 364         BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
 365                !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
 366
 367         if (refs > 1) {
 368                 if ((owner == root->root_key.objectid ||
 369                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
 370                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
 371                         ret = btrfs_inc_ref(trans, root, buf, 1, 1);
 372                         BUG_ON(ret); /* -ENOMEM */
 373
 374                         if (root->root_key.objectid ==
 375                             BTRFS_TREE_RELOC_OBJECTID) {
 376                                 ret = btrfs_dec_ref(trans, root, buf, 0, 1);
 377                                 BUG_ON(ret); /* -ENOMEM */
 378                                 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
 379                                 BUG_ON(ret); /* -ENOMEM */
 380                         }
 381                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
 382                 } else {
 383
 384                         if (root->root_key.objectid ==
 385                             BTRFS_TREE_RELOC_OBJECTID)
 386                                 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
 387                         else
 388                                 ret = btrfs_inc_ref(trans, root, cow, 0, 1);
 389                         BUG_ON(ret); /* -ENOMEM */
 390                 }
 391                 if (new_flags != 0) {
 392                         ret = btrfs_set_disk_extent_flags(trans, root,
 393                                                           buf->start,
 394                                                           buf->len,
 395                                                           new_flags, 0);
 396                         if (ret)
 397                                 return ret;
 398                 }
 399         } else {
 400                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
 401                         if (root->root_key.objectid ==
 402                             BTRFS_TREE_RELOC_OBJECTID)
 403                                 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
 404                         else
 405                                 ret = btrfs_inc_ref(trans, root, cow, 0, 1);
 406                         BUG_ON(ret); /* -ENOMEM */
 407                         ret = btrfs_dec_ref(trans, root, buf, 1, 1);
 408                         BUG_ON(ret); /* -ENOMEM */
 409                 }
 410                 clean_tree_block(trans, root, buf);
 411                 *last_ref = 1;
 412         }
 413         return 0;
 414 }
 415
 416 /*
 417  * does the dirty work in cow of a single block.  The parent block (if
 418  * supplied) is updated to point to the new cow copy.  The new buffer is marked
 419  * dirty and returned locked.  If you modify the block it needs to be marked
 420  * dirty again.
 421  *
 422  * search_start -- an allocation hint for the new block
 423  *
 424  * empty_size -- a hint that you plan on doing more cow.  This is the size in
 425  * bytes the allocator should try to find free next to the block it returns.
 426  * This is just a hint and may be ignored by the allocator.
 427  */
 428 static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 429                              struct btrfs_root *root,
 430                              struct extent_buffer *buf,
 431                              struct extent_buffer *parent, int parent_slot,
 432                              struct extent_buffer **cow_ret,
 433                              u64 search_start, u64 empty_size)
 434 {
 435         struct btrfs_disk_key disk_key;
 436         struct extent_buffer *cow;
 437         int level, ret;
 438         int last_ref = 0;
 439         int unlock_orig = 0;
 440         u64 parent_start;
 441
 442         if (*cow_ret == buf)
 443                 unlock_orig = 1;
 444
 445         btrfs_assert_tree_locked(buf);
 446
 447         WARN_ON(root->ref_cows && trans->transid !=
 448                 root->fs_info->running_transaction->transid);
 449         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 450
 451         level = btrfs_header_level(buf);
 452
 453         if (level == 0)
 454                 btrfs_item_key(buf, &disk_key, 0);
 455         else
 456                 btrfs_node_key(buf, &disk_key, 0);
 457
 458         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
 459                 if (parent)
 460                         parent_start = parent->start;
 461                 else
 462                         parent_start = 0;
 463         } else
 464                 parent_start = 0;
 465
 466         cow = btrfs_alloc_free_block(trans, root, buf->len, parent_start,
 467                                      root->root_key.objectid, &disk_key,
 468                                      level, search_start, empty_size, 1);
 469         if (IS_ERR(cow))
 470                 return PTR_ERR(cow);
 471
 472         /* cow is set to blocking by btrfs_init_new_buffer */
 473
 474         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 475         btrfs_set_header_bytenr(cow, cow->start);
 476         btrfs_set_header_generation(cow, trans->transid);
 477         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 478         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 479                                      BTRFS_HEADER_FLAG_RELOC);
 480         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 481                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 482         else
 483                 btrfs_set_header_owner(cow, root->root_key.objectid);
 484
 485         write_extent_buffer(cow, root->fs_info->fsid,
 486                             (unsigned long)btrfs_header_fsid(cow),
 487                             BTRFS_FSID_SIZE);
 488
 489         ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
 490         if (ret) {
 491                 btrfs_abort_transaction(trans, root, ret);
 492                 return ret;
 493         }
 494
 495         if (root->ref_cows)
 496                 btrfs_reloc_cow_block(trans, root, buf, cow);
 497
 498         if (buf == root->node) {
 499                 WARN_ON(parent && parent != buf);
 500                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 501                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 502                         parent_start = buf->start;
 503                 else
 504                         parent_start = 0;
 505
 506                 extent_buffer_get(cow);
 507                 rcu_assign_pointer(root->node, cow);
 508
 509                 btrfs_free_tree_block(trans, root, buf, parent_start,
 510                                       last_ref, 1);
 511                 free_extent_buffer(buf);
 512                 add_root_to_dirty_list(root);
 513         } else {
 514                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 515                         parent_start = parent->start;
 516                 else
 517                         parent_start = 0;
 518
 519                 WARN_ON(trans->transid != btrfs_header_generation(parent));
 520                 btrfs_set_node_blockptr(parent, parent_slot,
 521                                         cow->start);
 522                 btrfs_set_node_ptr_generation(parent, parent_slot,
 523                                               trans->transid);
 524                 btrfs_mark_buffer_dirty(parent);
 525                 btrfs_free_tree_block(trans, root, buf, parent_start,
 526                                       last_ref, 1);
 527         }
 528         if (unlock_orig)
 529                 btrfs_tree_unlock(buf);
 530         free_extent_buffer_stale(buf);
 531         btrfs_mark_buffer_dirty(cow);
 532         *cow_ret = cow;
 533         return 0;
 534 }
 535
 536 static inline int should_cow_block(struct btrfs_trans_handle *trans,
 537                                    struct btrfs_root *root,
 538                                    struct extent_buffer *buf)
 539 {
 540         /* ensure we can see the force_cow */
 541         smp_rmb();
 542
 543         /*
 544          * We do not need to cow a block if
 545          * 1) this block is not created or changed in this transaction;
 546          * 2) this block does not belong to TREE_RELOC tree;
 547          * 3) the root is not forced COW.
 548          *
 549          * What is forced COW:
 550          *    when we create snapshot during commiting the transaction,
 551          *    after we've finished coping src root, we must COW the shared
 552          *    block to ensure the metadata consistency.
 553          */
 554         if (btrfs_header_generation(buf) == trans->transid &&
 555             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
 556             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
 557               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
 558             !root->force_cow)
 559                 return 0;
 560         return 1;
 561 }
 562
 563 /*
 564  * cows a single block, see __btrfs_cow_block for the real work.
 565  * This version of it has extra checks so that a block isn't cow'd more than
 566  * once per transaction, as long as it hasn't been written yet
 567  */
 568 noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
 569                     struct btrfs_root *root, struct extent_buffer *buf,
 570                     struct extent_buffer *parent, int parent_slot,
 571                     struct extent_buffer **cow_ret)
 572 {
 573         u64 search_start;
 574         int ret;
 575
 576         if (trans->transaction != root->fs_info->running_transaction) {
 577                 printk(KERN_CRIT "trans %llu running %llu\n",
 578                        (unsigned long long)trans->transid,
 579                        (unsigned long long)
 580                        root->fs_info->running_transaction->transid);
 581                 WARN_ON(1);
 582         }
 583         if (trans->transid != root->fs_info->generation) {
 584                 printk(KERN_CRIT "trans %llu running %llu\n",
 585                        (unsigned long long)trans->transid,
 586                        (unsigned long long)root->fs_info->generation);
 587                 WARN_ON(1);
 588         }
 589
 590         if (!should_cow_block(trans, root, buf)) {
 591                 *cow_ret = buf;
 592                 return 0;
 593         }
 594
 595         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
 596
 597         if (parent)
 598                 btrfs_set_lock_blocking(parent);
 599         btrfs_set_lock_blocking(buf);
 600
 601         ret = __btrfs_cow_block(trans, root, buf, parent,
 602                                  parent_slot, cow_ret, search_start, 0);
 603
 604         trace_btrfs_cow_block(root, buf, *cow_ret);
 605
 606         return ret;
 607 }
 608
 609 /*
 610  * helper function for defrag to decide if two blocks pointed to by a
 611  * node are actually close by
 612  */
 613 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 614 {
 615         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 616                 return 1;
 617         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 618                 return 1;
 619         return 0;
 620 }
 621
 622 /*
 623  * compare two keys in a memcmp fashion
 624  */
 625 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 626 {
 627         struct btrfs_key k1;
 628
 629         btrfs_disk_key_to_cpu(&k1, disk);
 630
 631         return btrfs_comp_cpu_keys(&k1, k2);
 632 }
 633
 634 /*
 635  * same as comp_keys only with two btrfs_key's
 636  */
 637 int btrfs_comp_cpu_keys(struct btrfs_key *k1, struct btrfs_key *k2)
 638 {
 639         if (k1->objectid > k2->objectid)
 640                 return 1;
 641         if (k1->objectid < k2->objectid)
 642                 return -1;
 643         if (k1->type > k2->type)
 644                 return 1;
 645         if (k1->type < k2->type)
 646                 return -1;
 647         if (k1->offset > k2->offset)
 648                 return 1;
 649         if (k1->offset < k2->offset)
 650                 return -1;
 651         return 0;
 652 }
 653
 654 /*
 655  * this is used by the defrag code to go through all the
 656  * leaves pointed to by a node and reallocate them so that
 657  * disk order is close to key order
 658  */
 659 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 660                        struct btrfs_root *root, struct extent_buffer *parent,
 661                        int start_slot, int cache_only, u64 *last_ret,
 662                        struct btrfs_key *progress)
 663 {
 664         struct extent_buffer *cur;
 665         u64 blocknr;
 666         u64 gen;
 667         u64 search_start = *last_ret;
 668         u64 last_block = 0;
 669         u64 other;
 670         u32 parent_nritems;
 671         int end_slot;
 672         int i;
 673         int err = 0;
 674         int parent_level;
 675         int uptodate;
 676         u32 blocksize;
 677         int progress_passed = 0;
 678         struct btrfs_disk_key disk_key;
 679
 680         parent_level = btrfs_header_level(parent);
 681         if (cache_only && parent_level != 1)
 682                 return 0;
 683
 684         if (trans->transaction != root->fs_info->running_transaction)
 685                 WARN_ON(1);
 686         if (trans->transid != root->fs_info->generation)
 687                 WARN_ON(1);
 688
 689         parent_nritems = btrfs_header_nritems(parent);
 690         blocksize = btrfs_level_size(root, parent_level - 1);
 691         end_slot = parent_nritems;
 692
 693         if (parent_nritems == 1)
 694                 return 0;
 695
 696         btrfs_set_lock_blocking(parent);
 697
 698         for (i = start_slot; i < end_slot; i++) {
 699                 int close = 1;
 700
 701                 btrfs_node_key(parent, &disk_key, i);
 702                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
 703                         continue;
 704
 705                 progress_passed = 1;
 706                 blocknr = btrfs_node_blockptr(parent, i);
 707                 gen = btrfs_node_ptr_generation(parent, i);
 708                 if (last_block == 0)
 709                         last_block = blocknr;
 710
 711                 if (i > 0) {
 712                         other = btrfs_node_blockptr(parent, i - 1);
 713                         close = close_blocks(blocknr, other, blocksize);
 714                 }
 715                 if (!close && i < end_slot - 2) {
 716                         other = btrfs_node_blockptr(parent, i + 1);
 717                         close = close_blocks(blocknr, other, blocksize);
 718                 }
 719                 if (close) {
 720                         last_block = blocknr;
 721                         continue;
 722                 }
 723
 724                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
 725                 if (cur)
 726                         uptodate = btrfs_buffer_uptodate(cur, gen);
 727                 else
 728                         uptodate = 0;
 729                 if (!cur || !uptodate) {
 730                         if (cache_only) {
 731                                 free_extent_buffer(cur);
 732                                 continue;
 733                         }
 734                         if (!cur) {
 735                                 cur = read_tree_block(root, blocknr,
 736                                                          blocksize, gen);
 737                                 if (!cur)
 738                                         return -EIO;
 739                         } else if (!uptodate) {
 740                                 btrfs_read_buffer(cur, gen);
 741                         }
 742                 }
 743                 if (search_start == 0)
 744                         search_start = last_block;
 745
 746                 btrfs_tree_lock(cur);
 747                 btrfs_set_lock_blocking(cur);
 748                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 749                                         &cur, search_start,
 750                                         min(16 * blocksize,
 751                                             (end_slot - i) * blocksize));
 752                 if (err) {
 753                         btrfs_tree_unlock(cur);
 754                         free_extent_buffer(cur);
 755                         break;
 756                 }
 757                 search_start = cur->start;
 758                 last_block = cur->start;
 759                 *last_ret = search_start;
 760                 btrfs_tree_unlock(cur);
 761                 free_extent_buffer(cur);
 762         }
 763         return err;
 764 }
 765
 766 /*
 767  * The leaf data grows from end-to-front in the node.
 768  * this returns the address of the start of the last item,
 769  * which is the stop of the leaf data stack
 770  */
 771 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 772                                          struct extent_buffer *leaf)
 773 {
 774         u32 nr = btrfs_header_nritems(leaf);
 775         if (nr == 0)
 776                 return BTRFS_LEAF_DATA_SIZE(root);
 777         return btrfs_item_offset_nr(leaf, nr - 1);
 778 }
 779
 780
 781 /*
 782  * search for key in the extent_buffer.  The items start at offset p,
 783  * and they are item_size apart.  There are 'max' items in p.
 784  *
 785  * the slot in the array is returned via slot, and it points to
 786  * the place where you would insert key if it is not found in
 787  * the array.
 788  *
 789  * slot may point to max if the key is bigger than all of the keys
 790  */
 791 static noinline int generic_bin_search(struct extent_buffer *eb,
 792                                        unsigned long p,
 793                                        int item_size, struct btrfs_key *key,
 794                                        int max, int *slot)
 795 {
 796         int low = 0;
 797         int high = max;
 798         int mid;
 799         int ret;
 800         struct btrfs_disk_key *tmp = NULL;
 801         struct btrfs_disk_key unaligned;
 802         unsigned long offset;
 803         char *kaddr = NULL;
 804         unsigned long map_start = 0;
 805         unsigned long map_len = 0;
 806         int err;
 807
 808         while (low < high) {
 809                 mid = (low + high) / 2;
 810                 offset = p + mid * item_size;
 811
 812                 if (!kaddr || offset < map_start ||
 813                     (offset + sizeof(struct btrfs_disk_key)) >
 814                     map_start + map_len) {
 815
 816                         err = map_private_extent_buffer(eb, offset,
 817                                                 sizeof(struct btrfs_disk_key),
 818                                                 &kaddr, &map_start, &map_len);
 819
 820                         if (!err) {
 821                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
 822                                                         map_start);
 823                         } else {
 824                                 read_extent_buffer(eb, &unaligned,
 825                                                    offset, sizeof(unaligned));
 826                                 tmp = &unaligned;
 827                         }
 828
 829                 } else {
 830                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
 831                                                         map_start);
 832                 }
 833                 ret = comp_keys(tmp, key);
 834
 835                 if (ret < 0)
 836                         low = mid + 1;
 837                 else if (ret > 0)
 838                         high = mid;
 839                 else {
 840                         *slot = mid;
 841                         return 0;
 842                 }
 843         }
 844         *slot = low;
 845         return 1;
 846 }
 847
 848 /*
 849  * simple bin_search frontend that does the right thing for
 850  * leaves vs nodes
 851  */
 852 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 853                       int level, int *slot)
 854 {
 855         if (level == 0) {
 856                 return generic_bin_search(eb,
 857                                           offsetof(struct btrfs_leaf, items),
 858                                           sizeof(struct btrfs_item),
 859                                           key, btrfs_header_nritems(eb),
 860                                           slot);
 861         } else {
 862                 return generic_bin_search(eb,
 863                                           offsetof(struct btrfs_node, ptrs),
 864                                           sizeof(struct btrfs_key_ptr),
 865                                           key, btrfs_header_nritems(eb),
 866                                           slot);
 867         }
 868         return -1;
 869 }
 870
 871 int btrfs_bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 872                      int level, int *slot)
 873 {
 874         return bin_search(eb, key, level, slot);
 875 }
 876
 877 static void root_add_used(struct btrfs_root *root, u32 size)
 878 {
 879         spin_lock(&root->accounting_lock);
 880         btrfs_set_root_used(&root->root_item,
 881                             btrfs_root_used(&root->root_item) + size);
 882         spin_unlock(&root->accounting_lock);
 883 }
 884
 885 static void root_sub_used(struct btrfs_root *root, u32 size)
 886 {
 887         spin_lock(&root->accounting_lock);
 888         btrfs_set_root_used(&root->root_item,
 889                             btrfs_root_used(&root->root_item) - size);
 890         spin_unlock(&root->accounting_lock);
 891 }
 892
 893 /* given a node and slot number, this reads the blocks it points to.  The
 894  * extent buffer is returned with a reference taken (but unlocked).
 895  * NULL is returned on error.
 896  */
 897 static noinline struct extent_buffer *read_node_slot(struct btrfs_root *root,
 898                                    struct extent_buffer *parent, int slot)
 899 {
 900         int level = btrfs_header_level(parent);
 901         if (slot < 0)
 902                 return NULL;
 903         if (slot >= btrfs_header_nritems(parent))
 904                 return NULL;
 905
 906         BUG_ON(level == 0);
 907
 908         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 909                        btrfs_level_size(root, level - 1),
 910                        btrfs_node_ptr_generation(parent, slot));
 911 }
 912
 913 /*
 914  * node level balancing, used to make sure nodes are in proper order for
 915  * item deletion.  We balance from the top down, so we have to make sure
 916  * that a deletion won't leave an node completely empty later on.
 917  */
 918 static noinline int balance_level(struct btrfs_trans_handle *trans,
 919                          struct btrfs_root *root,
 920                          struct btrfs_path *path, int level)
 921 {
 922         struct extent_buffer *right = NULL;
 923         struct extent_buffer *mid;
 924         struct extent_buffer *left = NULL;
 925         struct extent_buffer *parent = NULL;
 926         int ret = 0;
 927         int wret;
 928         int pslot;
 929         int orig_slot = path->slots[level];
 930         u64 orig_ptr;
 931
 932         if (level == 0)
 933                 return 0;
 934
 935         mid = path->nodes[level];
 936
 937         WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK &&
 938                 path->locks[level] != BTRFS_WRITE_LOCK_BLOCKING);
 939         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 940
 941         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 942
 943         if (level < BTRFS_MAX_LEVEL - 1) {
 944                 parent = path->nodes[level + 1];
 945                 pslot = path->slots[level + 1];
 946         }
 947
 948         /*
 949          * deal with the case where there is only one pointer in the root
 950          * by promoting the node below to a root
 951          */
 952         if (!parent) {
 953                 struct extent_buffer *child;
 954
 955                 if (btrfs_header_nritems(mid) != 1)
 956                         return 0;
 957
 958                 /* promote the child to a root */
 959                 child = read_node_slot(root, mid, 0);
 960                 if (!child) {
 961                         ret = -EROFS;
 962                         btrfs_std_error(root->fs_info, ret);
 963                         goto enospc;
 964                 }
 965
 966                 btrfs_tree_lock(child);
 967                 btrfs_set_lock_blocking(child);
 968                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
 969                 if (ret) {
 970                         btrfs_tree_unlock(child);
 971                         free_extent_buffer(child);
 972                         goto enospc;
 973                 }
 974
 975                 rcu_assign_pointer(root->node, child);
 976
 977                 add_root_to_dirty_list(root);
 978                 btrfs_tree_unlock(child);
 979
 980                 path->locks[level] = 0;
 981                 path->nodes[level] = NULL;
 982                 clean_tree_block(trans, root, mid);
 983                 btrfs_tree_unlock(mid);
 984                 /* once for the path */
 985                 free_extent_buffer(mid);
 986
 987                 root_sub_used(root, mid->len);
 988                 btrfs_free_tree_block(trans, root, mid, 0, 1, 0);
 989                 /* once for the root ptr */
 990                 free_extent_buffer_stale(mid);
 991                 return 0;
 992         }
 993         if (btrfs_header_nritems(mid) >
 994             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 995                 return 0;
 996
 997         btrfs_header_nritems(mid);
 998
 999         left = read_node_slot(root, parent, pslot - 1);
1000         if (left) {
1001                 btrfs_tree_lock(left);
1002                 btrfs_set_lock_blocking(left);
1003                 wret = btrfs_cow_block(trans, root, left,
1004                                        parent, pslot - 1, &left);
1005                 if (wret) {
1006                         ret = wret;
1007                         goto enospc;
1008                 }
1009         }
1010         right = read_node_slot(root, parent, pslot + 1);
1011         if (right) {
1012                 btrfs_tree_lock(right);
1013                 btrfs_set_lock_blocking(right);
1014                 wret = btrfs_cow_block(trans, root, right,
1015                                        parent, pslot + 1, &right);
1016                 if (wret) {
1017                         ret = wret;
1018                         goto enospc;
1019                 }
1020         }
1021
1022         /* first, try to make some room in the middle buffer */
1023         if (left) {
1024                 orig_slot += btrfs_header_nritems(left);
1025                 wret = push_node_left(trans, root, left, mid, 1);
1026                 if (wret < 0)
1027                         ret = wret;
1028                 btrfs_header_nritems(mid);
1029         }
1030
1031         /*
1032          * then try to empty the right most buffer into the middle
1033          */
1034         if (right) {
1035                 wret = push_node_left(trans, root, mid, right, 1);
1036                 if (wret < 0 && wret != -ENOSPC)
1037                         ret = wret;
1038                 if (btrfs_header_nritems(right) == 0) {
1039                         clean_tree_block(trans, root, right);
1040                         btrfs_tree_unlock(right);
1041                         del_ptr(trans, root, path, level + 1, pslot + 1);
1042                         root_sub_used(root, right->len);
1043                         btrfs_free_tree_block(trans, root, right, 0, 1, 0);
1044                         free_extent_buffer_stale(right);
1045                         right = NULL;
1046                 } else {
1047                         struct btrfs_disk_key right_key;
1048                         btrfs_node_key(right, &right_key, 0);
1049                         btrfs_set_node_key(parent, &right_key, pslot + 1);
1050                         btrfs_mark_buffer_dirty(parent);
1051                 }
1052         }
1053         if (btrfs_header_nritems(mid) == 1) {
1054                 /*
1055                  * we're not allowed to leave a node with one item in the
1056                  * tree during a delete.  A deletion from lower in the tree
1057                  * could try to delete the only pointer in this node.
1058                  * So, pull some keys from the left.
1059                  * There has to be a left pointer at this point because
1060                  * otherwise we would have pulled some pointers from the
1061                  * right
1062                  */
1063                 if (!left) {
1064                         ret = -EROFS;
1065                         btrfs_std_error(root->fs_info, ret);
1066                         goto enospc;
1067                 }
1068                 wret = balance_node_right(trans, root, mid, left);
1069                 if (wret < 0) {
1070                         ret = wret;
1071                         goto enospc;
1072                 }
1073                 if (wret == 1) {
1074                         wret = push_node_left(trans, root, left, mid, 1);
1075                         if (wret < 0)
1076                                 ret = wret;
1077                 }
1078                 BUG_ON(wret == 1);
1079         }
1080         if (btrfs_header_nritems(mid) == 0) {
1081                 clean_tree_block(trans, root, mid);
1082                 btrfs_tree_unlock(mid);
1083                 del_ptr(trans, root, path, level + 1, pslot);
1084                 root_sub_used(root, mid->len);
1085                 btrfs_free_tree_block(trans, root, mid, 0, 1, 0);
1086                 free_extent_buffer_stale(mid);
1087                 mid = NULL;
1088         } else {
1089                 /* update the parent key to reflect our changes */
1090                 struct btrfs_disk_key mid_key;
1091                 btrfs_node_key(mid, &mid_key, 0);
1092                 btrfs_set_node_key(parent, &mid_key, pslot);
1093                 btrfs_mark_buffer_dirty(parent);
1094         }
1095
1096         /* update the path */
1097         if (left) {
1098                 if (btrfs_header_nritems(left) > orig_slot) {
1099                         extent_buffer_get(left);
1100                         /* left was locked after cow */
1101                         path->nodes[level] = left;
1102                         path->slots[level + 1] -= 1;
1103                         path->slots[level] = orig_slot;
1104                         if (mid) {
1105                                 btrfs_tree_unlock(mid);
1106                                 free_extent_buffer(mid);
1107                         }
1108                 } else {
1109                         orig_slot -= btrfs_header_nritems(left);
1110                         path->slots[level] = orig_slot;
1111                 }
1112         }
1113         /* double check we haven't messed things up */
1114         if (orig_ptr !=
1115             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
1116                 BUG();
1117 enospc:
1118         if (right) {
1119                 btrfs_tree_unlock(right);
1120                 free_extent_buffer(right);
1121         }
1122         if (left) {
1123                 if (path->nodes[level] != left)
1124                         btrfs_tree_unlock(left);
1125                 free_extent_buffer(left);
1126         }
1127         return ret;
1128 }
1129
1130 /* Node balancing for insertion.  Here we only split or push nodes around
1131  * when they are completely full.  This is also done top down, so we
1132  * have to be pessimistic.
1133  */
1134 static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
1135                                           struct btrfs_root *root,
1136                                           struct btrfs_path *path, int level)
1137 {
1138         struct extent_buffer *right = NULL;
1139         struct extent_buffer *mid;
1140         struct extent_buffer *left = NULL;
1141         struct extent_buffer *parent = NULL;
1142         int ret = 0;
1143         int wret;
1144         int pslot;
1145         int orig_slot = path->slots[level];
1146
1147         if (level == 0)
1148                 return 1;
1149
1150         mid = path->nodes[level];
1151         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1152
1153         if (level < BTRFS_MAX_LEVEL - 1) {
1154                 parent = path->nodes[level + 1];
1155                 pslot = path->slots[level + 1];
1156         }
1157
1158         if (!parent)
1159                 return 1;
1160
1161         left = read_node_slot(root, parent, pslot - 1);
1162
1163         /* first, try to make some room in the middle buffer */
1164         if (left) {
1165                 u32 left_nr;
1166
1167                 btrfs_tree_lock(left);
1168                 btrfs_set_lock_blocking(left);
1169
1170                 left_nr = btrfs_header_nritems(left);
1171                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1172                         wret = 1;
1173                 } else {
1174                         ret = btrfs_cow_block(trans, root, left, parent,
1175                                               pslot - 1, &left);
1176                         if (ret)
1177                                 wret = 1;
1178                         else {
1179                                 wret = push_node_left(trans, root,
1180                                                       left, mid, 0);
1181                         }
1182                 }
1183                 if (wret < 0)
1184                         ret = wret;
1185                 if (wret == 0) {
1186                         struct btrfs_disk_key disk_key;
1187                         orig_slot += left_nr;
1188                         btrfs_node_key(mid, &disk_key, 0);
1189                         btrfs_set_node_key(parent, &disk_key, pslot);
1190                         btrfs_mark_buffer_dirty(parent);
1191                         if (btrfs_header_nritems(left) > orig_slot) {
1192                                 path->nodes[level] = left;
1193                                 path->slots[level + 1] -= 1;
1194                                 path->slots[level] = orig_slot;
1195                                 btrfs_tree_unlock(mid);
1196                                 free_extent_buffer(mid);
1197                         } else {
1198                                 orig_slot -=
1199                                         btrfs_header_nritems(left);
1200                                 path->slots[level] = orig_slot;
1201                                 btrfs_tree_unlock(left);
1202                                 free_extent_buffer(left);
1203                         }
1204                         return 0;
1205                 }
1206                 btrfs_tree_unlock(left);
1207                 free_extent_buffer(left);
1208         }
1209         right = read_node_slot(root, parent, pslot + 1);
1210
1211         /*
1212          * then try to empty the right most buffer into the middle
1213          */
1214         if (right) {
1215                 u32 right_nr;
1216
1217                 btrfs_tree_lock(right);
1218                 btrfs_set_lock_blocking(right);
1219
1220                 right_nr = btrfs_header_nritems(right);
1221                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1222                         wret = 1;
1223                 } else {
1224                         ret = btrfs_cow_block(trans, root, right,
1225                                               parent, pslot + 1,
1226                                               &right);
1227                         if (ret)
1228                                 wret = 1;
1229                         else {
1230                                 wret = balance_node_right(trans, root,
1231                                                           right, mid);
1232                         }
1233                 }
1234                 if (wret < 0)
1235                         ret = wret;
1236                 if (wret == 0) {
1237                         struct btrfs_disk_key disk_key;
1238
1239                         btrfs_node_key(right, &disk_key, 0);
1240                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
1241                         btrfs_mark_buffer_dirty(parent);
1242
1243                         if (btrfs_header_nritems(mid) <= orig_slot) {
1244                                 path->nodes[level] = right;
1245                                 path->slots[level + 1] += 1;
1246                                 path->slots[level] = orig_slot -
1247                                         btrfs_header_nritems(mid);
1248                                 btrfs_tree_unlock(mid);
1249                                 free_extent_buffer(mid);
1250                         } else {
1251                                 btrfs_tree_unlock(right);
1252                                 free_extent_buffer(right);
1253                         }
1254                         return 0;
1255                 }
1256                 btrfs_tree_unlock(right);
1257                 free_extent_buffer(right);
1258         }
1259         return 1;
1260 }
1261
1262 /*
1263  * readahead one full node of leaves, finding things that are close
1264  * to the block in 'slot', and triggering ra on them.
1265  */
1266 static void reada_for_search(struct btrfs_root *root,
1267                              struct btrfs_path *path,
1268                              int level, int slot, u64 objectid)
1269 {
1270         struct extent_buffer *node;
1271         struct btrfs_disk_key disk_key;
1272         u32 nritems;
1273         u64 search;
1274         u64 target;
1275         u64 nread = 0;
1276         u64 gen;
1277         int direction = path->reada;
1278         struct extent_buffer *eb;
1279         u32 nr;
1280         u32 blocksize;
1281         u32 nscan = 0;
1282
1283         if (level != 1)
1284                 return;
1285
1286         if (!path->nodes[level])
1287                 return;
1288
1289         node = path->nodes[level];
1290
1291         search = btrfs_node_blockptr(node, slot);
1292         blocksize = btrfs_level_size(root, level - 1);
1293         eb = btrfs_find_tree_block(root, search, blocksize);
1294         if (eb) {
1295                 free_extent_buffer(eb);
1296                 return;
1297         }
1298
1299         target = search;
1300
1301         nritems = btrfs_header_nritems(node);
1302         nr = slot;
1303
1304         while (1) {
1305                 if (direction < 0) {
1306                         if (nr == 0)
1307                                 break;
1308                         nr--;
1309                 } else if (direction > 0) {
1310                         nr++;
1311                         if (nr >= nritems)
1312                                 break;
1313                 }
1314                 if (path->reada < 0 && objectid) {
1315                         btrfs_node_key(node, &disk_key, nr);
1316                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1317                                 break;
1318                 }
1319                 search = btrfs_node_blockptr(node, nr);
1320                 if ((search <= target && target - search <= 65536) ||
1321                     (search > target && search - target <= 65536)) {
1322                         gen = btrfs_node_ptr_generation(node, nr);
1323                         readahead_tree_block(root, search, blocksize, gen);
1324                         nread += blocksize;
1325                 }
1326                 nscan++;
1327                 if ((nread > 65536 || nscan > 32))
1328                         break;
1329         }
1330 }
1331
1332 /*
1333  * returns -EAGAIN if it had to drop the path, or zero if everything was in
1334  * cache
1335  */
1336 static noinline int reada_for_balance(struct btrfs_root *root,
1337                                       struct btrfs_path *path, int level)
1338 {
1339         int slot;
1340         int nritems;
1341         struct extent_buffer *parent;
1342         struct extent_buffer *eb;
1343         u64 gen;
1344         u64 block1 = 0;
1345         u64 block2 = 0;
1346         int ret = 0;
1347         int blocksize;
1348
1349         parent = path->nodes[level + 1];
1350         if (!parent)
1351                 return 0;
1352
1353         nritems = btrfs_header_nritems(parent);
1354         slot = path->slots[level + 1];
1355         blocksize = btrfs_level_size(root, level);
1356
1357         if (slot > 0) {
1358                 block1 = btrfs_node_blockptr(parent, slot - 1);
1359                 gen = btrfs_node_ptr_generation(parent, slot - 1);
1360                 eb = btrfs_find_tree_block(root, block1, blocksize);
1361                 if (eb && btrfs_buffer_uptodate(eb, gen))
1362                         block1 = 0;
1363                 free_extent_buffer(eb);
1364         }
1365         if (slot + 1 < nritems) {
1366                 block2 = btrfs_node_blockptr(parent, slot + 1);
1367                 gen = btrfs_node_ptr_generation(parent, slot + 1);
1368                 eb = btrfs_find_tree_block(root, block2, blocksize);
1369                 if (eb && btrfs_buffer_uptodate(eb, gen))
1370                         block2 = 0;
1371                 free_extent_buffer(eb);
1372         }
1373         if (block1 || block2) {
1374                 ret = -EAGAIN;
1375
1376                 /* release the whole path */
1377                 btrfs_release_path(path);
1378
1379                 /* read the blocks */
1380                 if (block1)
1381                         readahead_tree_block(root, block1, blocksize, 0);
1382                 if (block2)
1383                         readahead_tree_block(root, block2, blocksize, 0);
1384
1385                 if (block1) {
1386                         eb = read_tree_block(root, block1, blocksize, 0);
1387                         free_extent_buffer(eb);
1388                 }
1389                 if (block2) {
1390                         eb = read_tree_block(root, block2, blocksize, 0);
1391                         free_extent_buffer(eb);
1392                 }
1393         }
1394         return ret;
1395 }
1396
1397
1398 /*
1399  * when we walk down the tree, it is usually safe to unlock the higher layers
1400  * in the tree.  The exceptions are when our path goes through slot 0, because
1401  * operations on the tree might require changing key pointers higher up in the
1402  * tree.
1403  *
1404  * callers might also have set path->keep_locks, which tells this code to keep
1405  * the lock if the path points to the last slot in the block.  This is part of
1406  * walking through the tree, and selecting the next slot in the higher block.
1407  *
1408  * lowest_unlock sets the lowest level in the tree we're allowed to unlock.  so
1409  * if lowest_unlock is 1, level 0 won't be unlocked
1410  */
1411 static noinline void unlock_up(struct btrfs_path *path, int level,
1412                                int lowest_unlock, int min_write_lock_level,
1413                                int *write_lock_level)
1414 {
1415         int i;
1416         int skip_level = level;
1417         int no_skips = 0;
1418         struct extent_buffer *t;
1419
1420         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1421                 if (!path->nodes[i])
1422                         break;
1423                 if (!path->locks[i])
1424                         break;
1425                 if (!no_skips && path->slots[i] == 0) {
1426                         skip_level = i + 1;
1427                         continue;
1428                 }
1429                 if (!no_skips && path->keep_locks) {
1430                         u32 nritems;
1431                         t = path->nodes[i];
1432                         nritems = btrfs_header_nritems(t);
1433                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
1434                                 skip_level = i + 1;
1435                                 continue;
1436                         }
1437                 }
1438                 if (skip_level < i && i >= lowest_unlock)
1439                         no_skips = 1;
1440
1441                 t = path->nodes[i];
1442                 if (i >= lowest_unlock && i > skip_level && path->locks[i]) {
1443                         btrfs_tree_unlock_rw(t, path->locks[i]);
1444                         path->locks[i] = 0;
1445                         if (write_lock_level &&
1446                             i > min_write_lock_level &&
1447                             i <= *write_lock_level) {
1448                                 *write_lock_level = i - 1;
1449                         }
1450                 }
1451         }
1452 }
1453
1454 /*
1455  * This releases any locks held in the path starting at level and
1456  * going all the way up to the root.
1457  *
1458  * btrfs_search_slot will keep the lock held on higher nodes in a few
1459  * corner cases, such as COW of the block at slot zero in the node.  This
1460  * ignores those rules, and it should only be called when there are no
1461  * more updates to be done higher up in the tree.
1462  */
1463 noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
1464 {
1465         int i;
1466
1467         if (path->keep_locks)
1468                 return;
1469
1470         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1471                 if (!path->nodes[i])
1472                         continue;
1473                 if (!path->locks[i])
1474                         continue;
1475                 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
1476                 path->locks[i] = 0;
1477         }
1478 }
1479
1480 /*
1481  * helper function for btrfs_search_slot.  The goal is to find a block
1482  * in cache without setting the path to blocking.  If we find the block
1483  * we return zero and the path is unchanged.
1484  *
1485  * If we can't find the block, we set the path blocking and do some
1486  * reada.  -EAGAIN is returned and the search must be repeated.
1487  */
1488 static int
1489 read_block_for_search(struct btrfs_trans_handle *trans,
1490                        struct btrfs_root *root, struct btrfs_path *p,
1491                        struct extent_buffer **eb_ret, int level, int slot,
1492                        struct btrfs_key *key)
1493 {
1494         u64 blocknr;
1495         u64 gen;
1496         u32 blocksize;
1497         struct extent_buffer *b = *eb_ret;
1498         struct extent_buffer *tmp;
1499         int ret;
1500
1501         blocknr = btrfs_node_blockptr(b, slot);
1502         gen = btrfs_node_ptr_generation(b, slot);
1503         blocksize = btrfs_level_size(root, level - 1);
1504
1505         tmp = btrfs_find_tree_block(root, blocknr, blocksize);
1506         if (tmp) {
1507                 if (btrfs_buffer_uptodate(tmp, 0)) {
1508                         if (btrfs_buffer_uptodate(tmp, gen)) {
1509                                 /*
1510                                  * we found an up to date block without
1511                                  * sleeping, return
1512                                  * right away
1513                                  */
1514                                 *eb_ret = tmp;
1515                                 return 0;
1516                         }
1517                         /* the pages were up to date, but we failed
1518                          * the generation number check.  Do a full
1519                          * read for the generation number that is correct.
1520                          * We must do this without dropping locks so
1521                          * we can trust our generation number
1522                          */
1523                         free_extent_buffer(tmp);
1524                         btrfs_set_path_blocking(p);
1525
1526                         tmp = read_tree_block(root, blocknr, blocksize, gen);
1527                         if (tmp && btrfs_buffer_uptodate(tmp, gen)) {
1528                                 *eb_ret = tmp;
1529                                 return 0;
1530                         }
1531                         free_extent_buffer(tmp);
1532                         btrfs_release_path(p);
1533                         return -EIO;
1534                 }
1535         }
1536
1537         /*
1538          * reduce lock contention at high levels
1539          * of the btree by dropping locks before
1540          * we read.  Don't release the lock on the current
1541          * level because we need to walk this node to figure
1542          * out which blocks to read.
1543          */
1544         btrfs_unlock_up_safe(p, level + 1);
1545         btrfs_set_path_blocking(p);
1546
1547         free_extent_buffer(tmp);
1548         if (p->reada)
1549                 reada_for_search(root, p, level, slot, key->objectid);
1550
1551         btrfs_release_path(p);
1552
1553         ret = -EAGAIN;
1554         tmp = read_tree_block(root, blocknr, blocksize, 0);
1555         if (tmp) {
1556                 /*
1557                  * If the read above didn't mark this buffer up to date,
1558                  * it will never end up being up to date.  Set ret to EIO now
1559                  * and give up so that our caller doesn't loop forever
1560                  * on our EAGAINs.
1561                  */
1562                 if (!btrfs_buffer_uptodate(tmp, 0))
1563                         ret = -EIO;
1564                 free_extent_buffer(tmp);
1565         }
1566         return ret;
1567 }
1568
1569 /*
1570  * helper function for btrfs_search_slot.  This does all of the checks
1571  * for node-level blocks and does any balancing required based on
1572  * the ins_len.
1573  *
1574  * If no extra work was required, zero is returned.  If we had to
1575  * drop the path, -EAGAIN is returned and btrfs_search_slot must
1576  * start over
1577  */
1578 static int
1579 setup_nodes_for_search(struct btrfs_trans_handle *trans,
1580                        struct btrfs_root *root, struct btrfs_path *p,
1581                        struct extent_buffer *b, int level, int ins_len,
1582                        int *write_lock_level)
1583 {
1584         int ret;
1585         if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
1586             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1587                 int sret;
1588
1589                 if (*write_lock_level < level + 1) {
1590                         *write_lock_level = level + 1;
1591                         btrfs_release_path(p);
1592                         goto again;
1593                 }
1594
1595                 sret = reada_for_balance(root, p, level);
1596                 if (sret)
1597                         goto again;
1598
1599                 btrfs_set_path_blocking(p);
1600                 sret = split_node(trans, root, p, level);
1601                 btrfs_clear_path_blocking(p, NULL, 0);
1602
1603                 BUG_ON(sret > 0);
1604                 if (sret) {
1605                         ret = sret;
1606                         goto done;
1607                 }
1608                 b = p->nodes[level];
1609         } else if (ins_len < 0 && btrfs_header_nritems(b) <
1610                    BTRFS_NODEPTRS_PER_BLOCK(root) / 2) {
1611                 int sret;
1612
1613                 if (*write_lock_level < level + 1) {
1614                         *write_lock_level = level + 1;
1615                         btrfs_release_path(p);
1616                         goto again;
1617                 }
1618
1619                 sret = reada_for_balance(root, p, level);
1620                 if (sret)
1621                         goto again;
1622
1623                 btrfs_set_path_blocking(p);
1624                 sret = balance_level(trans, root, p, level);
1625                 btrfs_clear_path_blocking(p, NULL, 0);
1626
1627                 if (sret) {
1628                         ret = sret;
1629                         goto done;
1630                 }
1631                 b = p->nodes[level];
1632                 if (!b) {
1633                         btrfs_release_path(p);
1634                         goto again;
1635                 }
1636                 BUG_ON(btrfs_header_nritems(b) == 1);
1637         }
1638         return 0;
1639
1640 again:
1641         ret = -EAGAIN;
1642 done:
1643         return ret;
1644 }
1645
1646 /*
1647  * look for key in the tree.  path is filled in with nodes along the way
1648  * if key is found, we return zero and you can find the item in the leaf
1649  * level of the path (level 0)
1650  *
1651  * If the key isn't found, the path points to the slot where it should
1652  * be inserted, and 1 is returned.  If there are other errors during the
1653  * search a negative error number is returned.
1654  *
1655  * if ins_len > 0, nodes and leaves will be split as we walk down the
1656  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1657  * possible)
1658  */
1659 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1660                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1661                       ins_len, int cow)
1662 {
1663         struct extent_buffer *b;
1664         int slot;
1665         int ret;
1666         int err;
1667         int level;
1668         int lowest_unlock = 1;
1669         int root_lock;
1670         /* everything at write_lock_level or lower must be write locked */
1671         int write_lock_level = 0;
1672         u8 lowest_level = 0;
1673         int min_write_lock_level;
1674
1675         lowest_level = p->lowest_level;
1676         WARN_ON(lowest_level && ins_len > 0);
1677         WARN_ON(p->nodes[0] != NULL);
1678
1679         if (ins_len < 0) {
1680                 lowest_unlock = 2;
1681
1682                 /* when we are removing items, we might have to go up to level
1683                  * two as we update tree pointers  Make sure we keep write
1684                  * for those levels as well
1685                  */
1686                 write_lock_level = 2;
1687         } else if (ins_len > 0) {
1688                 /*
1689                  * for inserting items, make sure we have a write lock on
1690                  * level 1 so we can update keys
1691                  */
1692                 write_lock_level = 1;
1693         }
1694
1695         if (!cow)
1696                 write_lock_level = -1;
1697
1698         if (cow && (p->keep_locks || p->lowest_level))
1699                 write_lock_level = BTRFS_MAX_LEVEL;
1700
1701         min_write_lock_level = write_lock_level;
1702
1703 again:
1704         /*
1705          * we try very hard to do read locks on the root
1706          */
1707         root_lock = BTRFS_READ_LOCK;
1708         level = 0;
1709         if (p->search_commit_root) {
1710                 /*
1711                  * the commit roots are read only
1712                  * so we always do read locks
1713                  */
1714                 b = root->commit_root;
1715                 extent_buffer_get(b);
1716                 level = btrfs_header_level(b);
1717                 if (!p->skip_locking)
1718                         btrfs_tree_read_lock(b);
1719         } else {
1720                 if (p->skip_locking) {
1721                         b = btrfs_root_node(root);
1722                         level = btrfs_header_level(b);
1723                 } else {
1724                         /* we don't know the level of the root node
1725                          * until we actually have it read locked
1726                          */
1727                         b = btrfs_read_lock_root_node(root);
1728                         level = btrfs_header_level(b);
1729                         if (level <= write_lock_level) {
1730                                 /* whoops, must trade for write lock */
1731                                 btrfs_tree_read_unlock(b);
1732                                 free_extent_buffer(b);
1733                                 b = btrfs_lock_root_node(root);
1734                                 root_lock = BTRFS_WRITE_LOCK;
1735
1736                                 /* the level might have changed, check again */
1737                                 level = btrfs_header_level(b);
1738                         }
1739                 }
1740         }
1741         p->nodes[level] = b;
1742         if (!p->skip_locking)
1743                 p->locks[level] = root_lock;
1744
1745         while (b) {
1746                 level = btrfs_header_level(b);
1747
1748                 /*
1749                  * setup the path here so we can release it under lock
1750                  * contention with the cow code
1751                  */
1752                 if (cow) {
1753                         /*
1754                          * if we don't really need to cow this block
1755                          * then we don't want to set the path blocking,
1756                          * so we test it here
1757                          */
1758                         if (!should_cow_block(trans, root, b))
1759                                 goto cow_done;
1760
1761                         btrfs_set_path_blocking(p);
1762
1763                         /*
1764                          * must have write locks on this node and the
1765                          * parent
1766                          */
1767                         if (level + 1 > write_lock_level) {
1768                                 write_lock_level = level + 1;
1769                                 btrfs_release_path(p);
1770                                 goto again;
1771                         }
1772
1773                         err = btrfs_cow_block(trans, root, b,
1774                                               p->nodes[level + 1],
1775                                               p->slots[level + 1], &b);
1776                         if (err) {
1777                                 ret = err;
1778                                 goto done;
1779                         }
1780                 }
1781 cow_done:
1782                 BUG_ON(!cow && ins_len);
1783
1784                 p->nodes[level] = b;
1785                 btrfs_clear_path_blocking(p, NULL, 0);
1786
1787                 /*
1788                  * we have a lock on b and as long as we aren't changing
1789                  * the tree, there is no way to for the items in b to change.
1790                  * It is safe to drop the lock on our parent before we
1791                  * go through the expensive btree search on b.
1792                  *
1793                  * If cow is true, then we might be changing slot zero,
1794                  * which may require changing the parent.  So, we can't
1795                  * drop the lock until after we know which slot we're
1796                  * operating on.
1797                  */
1798                 if (!cow)
1799                         btrfs_unlock_up_safe(p, level + 1);
1800
1801                 ret = bin_search(b, key, level, &slot);
1802
1803                 if (level != 0) {
1804                         int dec = 0;
1805                         if (ret && slot > 0) {
1806                                 dec = 1;
1807                                 slot -= 1;
1808                         }
1809                         p->slots[level] = slot;
1810                         err = setup_nodes_for_search(trans, root, p, b, level,
1811                                              ins_len, &write_lock_level);
1812                         if (err == -EAGAIN)
1813                                 goto again;
1814                         if (err) {
1815                                 ret = err;
1816                                 goto done;
1817                         }
1818                         b = p->nodes[level];
1819                         slot = p->slots[level];
1820
1821                         /*
1822                          * slot 0 is special, if we change the key
1823                          * we have to update the parent pointer
1824                          * which means we must have a write lock
1825                          * on the parent
1826                          */
1827                         if (slot == 0 && cow &&
1828                             write_lock_level < level + 1) {
1829                                 write_lock_level = level + 1;
1830                                 btrfs_release_path(p);
1831                                 goto again;
1832                         }
1833
1834                         unlock_up(p, level, lowest_unlock,
1835                                   min_write_lock_level, &write_lock_level);
1836
1837                         if (level == lowest_level) {
1838                                 if (dec)
1839                                         p->slots[level]++;
1840                                 goto done;
1841                         }
1842
1843                         err = read_block_for_search(trans, root, p,
1844                                                     &b, level, slot, key);
1845                         if (err == -EAGAIN)
1846                                 goto again;
1847                         if (err) {
1848                                 ret = err;
1849                                 goto done;
1850                         }
1851
1852                         if (!p->skip_locking) {
1853                                 level = btrfs_header_level(b);
1854                                 if (level <= write_lock_level) {
1855                                         err = btrfs_try_tree_write_lock(b);
1856                                         if (!err) {
1857                                                 btrfs_set_path_blocking(p);
1858                                                 btrfs_tree_lock(b);
1859                                                 btrfs_clear_path_blocking(p, b,
1860                                                                   BTRFS_WRITE_LOCK);
1861                                         }
1862                                         p->locks[level] = BTRFS_WRITE_LOCK;
1863                                 } else {
1864                                         err = btrfs_try_tree_read_lock(b);
1865                                         if (!err) {
1866                                                 btrfs_set_path_blocking(p);
1867                                                 btrfs_tree_read_lock(b);
1868                                                 btrfs_clear_path_blocking(p, b,
1869                                                                   BTRFS_READ_LOCK);
1870                                         }
1871                                         p->locks[level] = BTRFS_READ_LOCK;
1872                                 }
1873                                 p->nodes[level] = b;
1874                         }
1875                 } else {
1876                         p->slots[level] = slot;
1877                         if (ins_len > 0 &&
1878                             btrfs_leaf_free_space(root, b) < ins_len) {
1879                                 if (write_lock_level < 1) {
1880                                         write_lock_level = 1;
1881                                         btrfs_release_path(p);
1882                                         goto again;
1883                                 }
1884
1885                                 btrfs_set_path_blocking(p);
1886                                 err = split_leaf(trans, root, key,
1887                                                  p, ins_len, ret == 0);
1888                                 btrfs_clear_path_blocking(p, NULL, 0);
1889
1890                                 BUG_ON(err > 0);
1891                                 if (err) {
1892                                         ret = err;
1893                                         goto done;
1894                                 }
1895                         }
1896                         if (!p->search_for_split)
1897                                 unlock_up(p, level, lowest_unlock,
1898                                           min_write_lock_level, &write_lock_level);
1899                         goto done;
1900                 }
1901         }
1902         ret = 1;
1903 done:
1904         /*
1905          * we don't really know what they plan on doing with the path
1906          * from here on, so for now just mark it as blocking
1907          */
1908         if (!p->leave_spinning)
1909                 btrfs_set_path_blocking(p);
1910         if (ret < 0)
1911                 btrfs_release_path(p);
1912         return ret;
1913 }
1914
1915 /*
1916  * adjust the pointers going up the tree, starting at level
1917  * making sure the right key of each node is points to 'key'.
1918  * This is used after shifting pointers to the left, so it stops
1919  * fixing up pointers when a given leaf/node is not in slot 0 of the
1920  * higher levels
1921  *
1922  */
1923 static void fixup_low_keys(struct btrfs_trans_handle *trans,
1924                            struct btrfs_root *root, struct btrfs_path *path,
1925                            struct btrfs_disk_key *key, int level)
1926 {
1927         int i;
1928         struct extent_buffer *t;
1929
1930         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1931                 int tslot = path->slots[i];
1932                 if (!path->nodes[i])
1933                         break;
1934                 t = path->nodes[i];
1935                 btrfs_set_node_key(t, key, tslot);
1936                 btrfs_mark_buffer_dirty(path->nodes[i]);
1937                 if (tslot != 0)
1938                         break;
1939         }
1940 }
1941
1942 /*
1943  * update item key.
1944  *
1945  * This function isn't completely safe. It's the caller's responsibility
1946  * that the new key won't break the order
1947  */
1948 void btrfs_set_item_key_safe(struct btrfs_trans_handle *trans,
1949                              struct btrfs_root *root, struct btrfs_path *path,
1950                              struct btrfs_key *new_key)
1951 {
1952         struct btrfs_disk_key disk_key;
1953         struct extent_buffer *eb;
1954         int slot;
1955
1956         eb = path->nodes[0];
1957         slot = path->slots[0];
1958         if (slot > 0) {
1959                 btrfs_item_key(eb, &disk_key, slot - 1);
1960                 BUG_ON(comp_keys(&disk_key, new_key) >= 0);
1961         }
1962         if (slot < btrfs_header_nritems(eb) - 1) {
1963                 btrfs_item_key(eb, &disk_key, slot + 1);
1964                 BUG_ON(comp_keys(&disk_key, new_key) <= 0);
1965         }
1966
1967         btrfs_cpu_key_to_disk(&disk_key, new_key);
1968         btrfs_set_item_key(eb, &disk_key, slot);
1969         btrfs_mark_buffer_dirty(eb);
1970         if (slot == 0)
1971                 fixup_low_keys(trans, root, path, &disk_key, 1);
1972 }
1973
1974 /*
1975  * try to push data from one node into the next node left in the
1976  * tree.
1977  *
1978  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1979  * error, and > 0 if there was no room in the left hand block.
1980  */
1981 static int push_node_left(struct btrfs_trans_handle *trans,
1982                           struct btrfs_root *root, struct extent_buffer *dst,
1983                           struct extent_buffer *src, int empty)
1984 {
1985         int push_items = 0;
1986         int src_nritems;
1987         int dst_nritems;
1988         int ret = 0;
1989
1990         src_nritems = btrfs_header_nritems(src);
1991         dst_nritems = btrfs_header_nritems(dst);
1992         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1993         WARN_ON(btrfs_header_generation(src) != trans->transid);
1994         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1995
1996         if (!empty && src_nritems <= 8)
1997                 return 1;
1998
1999         if (push_items <= 0)
2000                 return 1;
2001
2002         if (empty) {
2003                 push_items = min(src_nritems, push_items);
2004                 if (push_items < src_nritems) {
2005                         /* leave at least 8 pointers in the node if
2006                          * we aren't going to empty it
2007                          */
2008                         if (src_nritems - push_items < 8) {
2009                                 if (push_items <= 8)
2010                                         return 1;
2011                                 push_items -= 8;
2012                         }
2013                 }
2014         } else
2015                 push_items = min(src_nritems - 8, push_items);
2016
2017         copy_extent_buffer(dst, src,
2018                            btrfs_node_key_ptr_offset(dst_nritems),
2019                            btrfs_node_key_ptr_offset(0),
2020                            push_items * sizeof(struct btrfs_key_ptr));
2021
2022         if (push_items < src_nritems) {
2023                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
2024                                       btrfs_node_key_ptr_offset(push_items),
2025                                       (src_nritems - push_items) *
2026                                       sizeof(struct btrfs_key_ptr));
2027         }
2028         btrfs_set_header_nritems(src, src_nritems - push_items);
2029         btrfs_set_header_nritems(dst, dst_nritems + push_items);
2030         btrfs_mark_buffer_dirty(src);
2031         btrfs_mark_buffer_dirty(dst);
2032
2033         return ret;
2034 }
2035
2036 /*
2037  * try to push data from one node into the next node right in the
2038  * tree.
2039  *
2040  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
2041  * error, and > 0 if there was no room in the right hand block.
2042  *
2043  * this will  only push up to 1/2 the contents of the left node over
2044  */
2045 static int balance_node_right(struct btrfs_trans_handle *trans,
2046                               struct btrfs_root *root,
2047                               struct extent_buffer *dst,
2048                               struct extent_buffer *src)
2049 {
2050         int push_items = 0;
2051         int max_push;
2052         int src_nritems;
2053         int dst_nritems;
2054         int ret = 0;
2055
2056         WARN_ON(btrfs_header_generation(src) != trans->transid);
2057         WARN_ON(btrfs_header_generation(dst) != trans->transid);
2058
2059         src_nritems = btrfs_header_nritems(src);
2060         dst_nritems = btrfs_header_nritems(dst);
2061         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
2062         if (push_items <= 0)
2063                 return 1;
2064
2065         if (src_nritems < 4)
2066                 return 1;
2067
2068         max_push = src_nritems / 2 + 1;
2069         /* don't try to empty the node */
2070         if (max_push >= src_nritems)
2071                 return 1;
2072
2073         if (max_push < push_items)
2074                 push_items = max_push;
2075
2076         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
2077                                       btrfs_node_key_ptr_offset(0),
2078                                       (dst_nritems) *
2079                                       sizeof(struct btrfs_key_ptr));
2080
2081         copy_extent_buffer(dst, src,
2082                            btrfs_node_key_ptr_offset(0),
2083                            btrfs_node_key_ptr_offset(src_nritems - push_items),
2084                            push_items * sizeof(struct btrfs_key_ptr));
2085
2086         btrfs_set_header_nritems(src, src_nritems - push_items);
2087         btrfs_set_header_nritems(dst, dst_nritems + push_items);
2088
2089         btrfs_mark_buffer_dirty(src);
2090         btrfs_mark_buffer_dirty(dst);
2091
2092         return ret;
2093 }
2094
2095 /*
2096  * helper function to insert a new root level in the tree.
2097  * A new node is allocated, and a single item is inserted to
2098  * point to the existing root
2099  *
2100  * returns zero on success or < 0 on failure.
2101  */
2102 static noinline int insert_new_root(struct btrfs_trans_handle *trans,
2103                            struct btrfs_root *root,
2104                            struct btrfs_path *path, int level)
2105 {
2106         u64 lower_gen;
2107         struct extent_buffer *lower;
2108         struct extent_buffer *c;
2109         struct extent_buffer *old;
2110         struct btrfs_disk_key lower_key;
2111
2112         BUG_ON(path->nodes[level]);
2113         BUG_ON(path->nodes[level-1] != root->node);
2114
2115         lower = path->nodes[level-1];
2116         if (level == 1)
2117                 btrfs_item_key(lower, &lower_key, 0);
2118         else
2119                 btrfs_node_key(lower, &lower_key, 0);
2120
2121         c = btrfs_alloc_free_block(trans, root, root->nodesize, 0,
2122                                    root->root_key.objectid, &lower_key,
2123                                    level, root->node->start, 0, 0);
2124         if (IS_ERR(c))
2125                 return PTR_ERR(c);
2126
2127         root_add_used(root, root->nodesize);
2128
2129         memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
2130         btrfs_set_header_nritems(c, 1);
2131         btrfs_set_header_level(c, level);
2132         btrfs_set_header_bytenr(c, c->start);
2133         btrfs_set_header_generation(c, trans->transid);
2134         btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
2135         btrfs_set_header_owner(c, root->root_key.objectid);
2136
2137         write_extent_buffer(c, root->fs_info->fsid,
2138                             (unsigned long)btrfs_header_fsid(c),
2139                             BTRFS_FSID_SIZE);
2140
2141         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
2142                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
2143                             BTRFS_UUID_SIZE);
2144
2145         btrfs_set_node_key(c, &lower_key, 0);
2146         btrfs_set_node_blockptr(c, 0, lower->start);
2147         lower_gen = btrfs_header_generation(lower);
2148         WARN_ON(lower_gen != trans->transid);
2149
2150         btrfs_set_node_ptr_generation(c, 0, lower_gen);
2151
2152         btrfs_mark_buffer_dirty(c);
2153
2154         old = root->node;
2155         rcu_assign_pointer(root->node, c);
2156
2157         /* the super has an extra ref to root->node */
2158         free_extent_buffer(old);
2159
2160         add_root_to_dirty_list(root);
2161         extent_buffer_get(c);
2162         path->nodes[level] = c;
2163         path->locks[level] = BTRFS_WRITE_LOCK;
2164         path->slots[level] = 0;
2165         return 0;
2166 }
2167
2168 /*
2169  * worker function to insert a single pointer in a node.
2170  * the node should have enough room for the pointer already
2171  *
2172  * slot and level indicate where you want the key to go, and
2173  * blocknr is the block the key points to.
2174  */
2175 static void insert_ptr(struct btrfs_trans_handle *trans,
2176                        struct btrfs_root *root, struct btrfs_path *path,
2177                        struct btrfs_disk_key *key, u64 bytenr,
2178                        int slot, int level)
2179 {
2180         struct extent_buffer *lower;
2181         int nritems;
2182
2183         BUG_ON(!path->nodes[level]);
2184         btrfs_assert_tree_locked(path->nodes[level]);
2185         lower = path->nodes[level];
2186         nritems = btrfs_header_nritems(lower);
2187         BUG_ON(slot > nritems);
2188         BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(root));
2189         if (slot != nritems) {
2190                 memmove_extent_buffer(lower,
2191                               btrfs_node_key_ptr_offset(slot + 1),
2192                               btrfs_node_key_ptr_offset(slot),
2193                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
2194         }
2195         btrfs_set_node_key(lower, key, slot);
2196         btrfs_set_node_blockptr(lower, slot, bytenr);
2197         WARN_ON(trans->transid == 0);
2198         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
2199         btrfs_set_header_nritems(lower, nritems + 1);
2200         btrfs_mark_buffer_dirty(lower);
2201 }
2202
2203 /*
2204  * split the node at the specified level in path in two.
2205  * The path is corrected to point to the appropriate node after the split
2206  *
2207  * Before splitting this tries to make some room in the node by pushing
2208  * left and right, if either one works, it returns right away.
2209  *
2210  * returns 0 on success and < 0 on failure
2211  */
2212 static noinline int split_node(struct btrfs_trans_handle *trans,
2213                                struct btrfs_root *root,
2214                                struct btrfs_path *path, int level)
2215 {
2216         struct extent_buffer *c;
2217         struct extent_buffer *split;
2218         struct btrfs_disk_key disk_key;
2219         int mid;
2220         int ret;
2221         u32 c_nritems;
2222
2223         c = path->nodes[level];
2224         WARN_ON(btrfs_header_generation(c) != trans->transid);
2225         if (c == root->node) {
2226                 /* trying to split the root, lets make a new one */
2227                 ret = insert_new_root(trans, root, path, level + 1);
2228                 if (ret)
2229                         return ret;
2230         } else {
2231                 ret = push_nodes_for_insert(trans, root, path, level);
2232                 c = path->nodes[level];
2233                 if (!ret && btrfs_header_nritems(c) <
2234                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
2235                         return 0;
2236                 if (ret < 0)
2237                         return ret;
2238         }
2239
2240         c_nritems = btrfs_header_nritems(c);
2241         mid = (c_nritems + 1) / 2;
2242         btrfs_node_key(c, &disk_key, mid);
2243
2244         split = btrfs_alloc_free_block(trans, root, root->nodesize, 0,
2245                                         root->root_key.objectid,
2246                                         &disk_key, level, c->start, 0, 0);
2247         if (IS_ERR(split))
2248                 return PTR_ERR(split);
2249
2250         root_add_used(root, root->nodesize);
2251
2252         memset_extent_buffer(split, 0, 0, sizeof(struct btrfs_header));
2253         btrfs_set_header_level(split, btrfs_header_level(c));
2254         btrfs_set_header_bytenr(split, split->start);
2255         btrfs_set_header_generation(split, trans->transid);
2256         btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
2257         btrfs_set_header_owner(split, root->root_key.objectid);
2258         write_extent_buffer(split, root->fs_info->fsid,
2259                             (unsigned long)btrfs_header_fsid(split),
2260                             BTRFS_FSID_SIZE);
2261         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
2262                             (unsigned long)btrfs_header_chunk_tree_uuid(split),
2263                             BTRFS_UUID_SIZE);
2264
2265
2266         copy_extent_buffer(split, c,
2267                            btrfs_node_key_ptr_offset(0),
2268                            btrfs_node_key_ptr_offset(mid),
2269                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
2270         btrfs_set_header_nritems(split, c_nritems - mid);
2271         btrfs_set_header_nritems(c, mid);
2272         ret = 0;
2273
2274         btrfs_mark_buffer_dirty(c);
2275         btrfs_mark_buffer_dirty(split);
2276
2277         insert_ptr(trans, root, path, &disk_key, split->start,
2278                    path->slots[level + 1] + 1, level + 1);
2279
2280         if (path->slots[level] >= mid) {
2281                 path->slots[level] -= mid;
2282                 btrfs_tree_unlock(c);
2283                 free_extent_buffer(c);
2284                 path->nodes[level] = split;
2285                 path->slots[level + 1] += 1;
2286         } else {
2287                 btrfs_tree_unlock(split);
2288                 free_extent_buffer(split);
2289         }
2290         return ret;
2291 }
2292
2293 /*
2294  * how many bytes are required to store the items in a leaf.  start
2295  * and nr indicate which items in the leaf to check.  This totals up the
2296  * space used both by the item structs and the item data
2297  */
2298 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
2299 {
2300         int data_len;
2301         int nritems = btrfs_header_nritems(l);
2302         int end = min(nritems, start + nr) - 1;
2303
2304         if (!nr)
2305                 return 0;
2306         data_len = btrfs_item_end_nr(l, start);
2307         data_len = data_len - btrfs_item_offset_nr(l, end);
2308         data_len += sizeof(struct btrfs_item) * nr;
2309         WARN_ON(data_len < 0);
2310         return data_len;
2311 }
2312
2313 /*
2314  * The space between the end of the leaf items and
2315  * the start of the leaf data.  IOW, how much room
2316  * the leaf has left for both items and data
2317  */
2318 noinline int btrfs_leaf_free_space(struct btrfs_root *root,
2319                                    struct extent_buffer *leaf)
2320 {
2321         int nritems = btrfs_header_nritems(leaf);
2322         int ret;
2323         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
2324         if (ret < 0) {
2325                 printk(KERN_CRIT "leaf free space ret %d, leaf data size %lu, "
2326                        "used %d nritems %d\n",
2327                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
2328                        leaf_space_used(leaf, 0, nritems), nritems);
2329         }
2330         return ret;
2331 }
2332
2333 /*
2334  * min slot controls the lowest index we're willing to push to the
2335  * right.  We'll push up to and including min_slot, but no lower
2336  */
2337 static noinline int __push_leaf_right(struct btrfs_trans_handle *trans,
2338                                       struct btrfs_root *root,
2339                                       struct btrfs_path *path,
2340                                       int data_size, int empty,
2341                                       struct extent_buffer *right,
2342                                       int free_space, u32 left_nritems,
2343                                       u32 min_slot)
2344 {
2345         struct extent_buffer *left = path->nodes[0];
2346         struct extent_buffer *upper = path->nodes[1];
2347         struct btrfs_map_token token;
2348         struct btrfs_disk_key disk_key;
2349         int slot;
2350         u32 i;
2351         int push_space = 0;
2352         int push_items = 0;
2353         struct btrfs_item *item;
2354         u32 nr;
2355         u32 right_nritems;
2356         u32 data_end;
2357         u32 this_item_size;
2358
2359         btrfs_init_map_token(&token);
2360
2361         if (empty)
2362                 nr = 0;
2363         else
2364                 nr = max_t(u32, 1, min_slot);
2365
2366         if (path->slots[0] >= left_nritems)
2367                 push_space += data_size;
2368
2369         slot = path->slots[1];
2370         i = left_nritems - 1;
2371         while (i >= nr) {
2372                 item = btrfs_item_nr(left, i);
2373
2374                 if (!empty && push_items > 0) {
2375                         if (path->slots[0] > i)
2376                                 break;
2377                         if (path->slots[0] == i) {
2378                                 int space = btrfs_leaf_free_space(root, left);
2379                                 if (space + push_space * 2 > free_space)
2380                                         break;
2381                         }
2382                 }
2383
2384                 if (path->slots[0] == i)
2385                         push_space += data_size;
2386
2387                 this_item_size = btrfs_item_size(left, item);
2388                 if (this_item_size + sizeof(*item) + push_space > free_space)
2389                         break;
2390
2391                 push_items++;
2392                 push_space += this_item_size + sizeof(*item);
2393                 if (i == 0)
2394                         break;
2395                 i--;
2396         }
2397
2398         if (push_items == 0)
2399                 goto out_unlock;
2400
2401         if (!empty && push_items == left_nritems)
2402                 WARN_ON(1);
2403
2404         /* push left to right */
2405         right_nritems = btrfs_header_nritems(right);
2406
2407         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
2408         push_space -= leaf_data_end(root, left);
2409
2410         /* make room in the right data area */
2411         data_end = leaf_data_end(root, right);
2412         memmove_extent_buffer(right,
2413                               btrfs_leaf_data(right) + data_end - push_space,
2414                               btrfs_leaf_data(right) + data_end,
2415                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
2416
2417         /* copy from the left data area */
2418         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
2419                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
2420                      btrfs_leaf_data(left) + leaf_data_end(root, left),
2421                      push_space);
2422
2423         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
2424                               btrfs_item_nr_offset(0),
2425                               right_nritems * sizeof(struct btrfs_item));
2426
2427         /* copy the items from left to right */
2428         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
2429                    btrfs_item_nr_offset(left_nritems - push_items),
2430                    push_items * sizeof(struct btrfs_item));
2431
2432         /* update the item pointers */
2433         right_nritems += push_items;
2434         btrfs_set_header_nritems(right, right_nritems);
2435         push_space = BTRFS_LEAF_DATA_SIZE(root);
2436         for (i = 0; i < right_nritems; i++) {
2437                 item = btrfs_item_nr(right, i);
2438                 push_space -= btrfs_token_item_size(right, item, &token);
2439                 btrfs_set_token_item_offset(right, item, push_space, &token);
2440         }
2441
2442         left_nritems -= push_items;
2443         btrfs_set_header_nritems(left, left_nritems);
2444
2445         if (left_nritems)
2446                 btrfs_mark_buffer_dirty(left);
2447         else
2448                 clean_tree_block(trans, root, left);
2449
2450         btrfs_mark_buffer_dirty(right);
2451
2452         btrfs_item_key(right, &disk_key, 0);
2453         btrfs_set_node_key(upper, &disk_key, slot + 1);
2454         btrfs_mark_buffer_dirty(upper);
2455
2456         /* then fixup the leaf pointer in the path */
2457         if (path->slots[0] >= left_nritems) {
2458                 path->slots[0] -= left_nritems;
2459                 if (btrfs_header_nritems(path->nodes[0]) == 0)
2460                         clean_tree_block(trans, root, path->nodes[0]);
2461                 btrfs_tree_unlock(path->nodes[0]);
2462                 free_extent_buffer(path->nodes[0]);
2463                 path->nodes[0] = right;
2464                 path->slots[1] += 1;
2465         } else {
2466                 btrfs_tree_unlock(right);
2467                 free_extent_buffer(right);
2468         }
2469         return 0;
2470
2471 out_unlock:
2472         btrfs_tree_unlock(right);
2473         free_extent_buffer(right);
2474         return 1;
2475 }
2476
2477 /*
2478  * push some data in the path leaf to the right, trying to free up at
2479  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
2480  *
2481  * returns 1 if the push failed because the other node didn't have enough
2482  * room, 0 if everything worked out and < 0 if there were major errors.
2483  *
2484  * this will push starting from min_slot to the end of the leaf.  It won't
2485  * push any slot lower than min_slot
2486  */
2487 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
2488                            *root, struct btrfs_path *path,
2489                            int min_data_size, int data_size,
2490                            int empty, u32 min_slot)
2491 {
2492         struct extent_buffer *left = path->nodes[0];
2493         struct extent_buffer *right;
2494         struct extent_buffer *upper;
2495         int slot;
2496         int free_space;
2497         u32 left_nritems;
2498         int ret;
2499
2500         if (!path->nodes[1])
2501                 return 1;
2502
2503         slot = path->slots[1];
2504         upper = path->nodes[1];
2505         if (slot >= btrfs_header_nritems(upper) - 1)
2506                 return 1;
2507
2508         btrfs_assert_tree_locked(path->nodes[1]);
2509
2510         right = read_node_slot(root, upper, slot + 1);
2511         if (right == NULL)
2512                 return 1;
2513
2514         btrfs_tree_lock(right);
2515         btrfs_set_lock_blocking(right);
2516
2517         free_space = btrfs_leaf_free_space(root, right);
2518         if (free_space < data_size)
2519                 goto out_unlock;
2520
2521         /* cow and double check */
2522         ret = btrfs_cow_block(trans, root, right, upper,
2523                               slot + 1, &right);
2524         if (ret)
2525                 goto out_unlock;
2526
2527         free_space = btrfs_leaf_free_space(root, right);
2528         if (free_space < data_size)
2529                 goto out_unlock;
2530
2531         left_nritems = btrfs_header_nritems(left);
2532         if (left_nritems == 0)
2533                 goto out_unlock;
2534
2535         return __push_leaf_right(trans, root, path, min_data_size, empty,
2536                                 right, free_space, left_nritems, min_slot);
2537 out_unlock:
2538         btrfs_tree_unlock(right);
2539         free_extent_buffer(right);
2540         return 1;
2541 }
2542
2543 /*
2544  * push some data in the path leaf to the left, trying to free up at
2545  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
2546  *
2547  * max_slot can put a limit on how far into the leaf we'll push items.  The
2548  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us do all the
2549  * items
2550  */
2551 static noinline int __push_leaf_left(struct btrfs_trans_handle *trans,
2552                                      struct btrfs_root *root,
2553                                      struct btrfs_path *path, int data_size,
2554                                      int empty, struct extent_buffer *left,
2555                                      int free_space, u32 right_nritems,
2556                                      u32 max_slot)
2557 {
2558         struct btrfs_disk_key disk_key;
2559         struct extent_buffer *right = path->nodes[0];
2560         int i;
2561         int push_space = 0;
2562         int push_items = 0;
2563         struct btrfs_item *item;
2564         u32 old_left_nritems;
2565         u32 nr;
2566         int ret = 0;
2567         u32 this_item_size;
2568         u32 old_left_item_size;
2569         struct btrfs_map_token token;
2570
2571         btrfs_init_map_token(&token);
2572
2573         if (empty)
2574                 nr = min(right_nritems, max_slot);
2575         else
2576                 nr = min(right_nritems - 1, max_slot);
2577
2578         for (i = 0; i < nr; i++) {
2579                 item = btrfs_item_nr(right, i);
2580
2581                 if (!empty && push_items > 0) {
2582                         if (path->slots[0] < i)
2583                                 break;
2584                         if (path->slots[0] == i) {
2585                                 int space = btrfs_leaf_free_space(root, right);
2586                                 if (space + push_space * 2 > free_space)
2587                                         break;
2588                         }
2589                 }
2590
2591                 if (path->slots[0] == i)
2592                         push_space += data_size;
2593
2594                 this_item_size = btrfs_item_size(right, item);
2595                 if (this_item_size + sizeof(*item) + push_space > free_space)
2596                         break;
2597
2598                 push_items++;
2599                 push_space += this_item_size + sizeof(*item);
2600         }
2601
2602         if (push_items == 0) {
2603                 ret = 1;
2604                 goto out;
2605         }
2606         if (!empty && push_items == btrfs_header_nritems(right))
2607                 WARN_ON(1);
2608
2609         /* push data from right to left */
2610         copy_extent_buffer(left, right,
2611                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
2612                            btrfs_item_nr_offset(0),
2613                            push_items * sizeof(struct btrfs_item));
2614
2615         push_space = BTRFS_LEAF_DATA_SIZE(root) -
2616                      btrfs_item_offset_nr(right, push_items - 1);
2617
2618         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
2619                      leaf_data_end(root, left) - push_space,
2620                      btrfs_leaf_data(right) +
2621                      btrfs_item_offset_nr(right, push_items - 1),
2622                      push_space);
2623         old_left_nritems = btrfs_header_nritems(left);
2624         BUG_ON(old_left_nritems <= 0);
2625
2626         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
2627         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
2628                 u32 ioff;
2629
2630                 item = btrfs_item_nr(left, i);
2631
2632                 ioff = btrfs_token_item_offset(left, item, &token);
2633                 btrfs_set_token_item_offset(left, item,
2634                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size),
2635                       &token);
2636         }
2637         btrfs_set_header_nritems(left, old_left_nritems + push_items);
2638
2639         /* fixup right node */
2640         if (push_items > right_nritems) {
2641                 printk(KERN_CRIT "push items %d nr %u\n", push_items,
2642                        right_nritems);
2643                 WARN_ON(1);
2644         }
2645
2646         if (push_items < right_nritems) {
2647                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
2648                                                   leaf_data_end(root, right);
2649                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
2650                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
2651                                       btrfs_leaf_data(right) +
2652                                       leaf_data_end(root, right), push_space);
2653
2654                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
2655                               btrfs_item_nr_offset(push_items),
2656                              (btrfs_header_nritems(right) - push_items) *
2657                              sizeof(struct btrfs_item));
2658         }
2659         right_nritems -= push_items;
2660         btrfs_set_header_nritems(right, right_nritems);
2661         push_space = BTRFS_LEAF_DATA_SIZE(root);
2662         for (i = 0; i < right_nritems; i++) {
2663                 item = btrfs_item_nr(right, i);
2664
2665                 push_space = push_space - btrfs_token_item_size(right,
2666                                                                 item, &token);
2667                 btrfs_set_token_item_offset(right, item, push_space, &token);
2668         }
2669
2670         btrfs_mark_buffer_dirty(left);
2671         if (right_nritems)
2672                 btrfs_mark_buffer_dirty(right);
2673         else
2674                 clean_tree_block(trans, root, right);
2675
2676         btrfs_item_key(right, &disk_key, 0);
2677         fixup_low_keys(trans, root, path, &disk_key, 1);
2678
2679         /* then fixup the leaf pointer in the path */
2680         if (path->slots[0] < push_items) {
2681                 path->slots[0] += old_left_nritems;
2682                 btrfs_tree_unlock(path->nodes[0]);
2683                 free_extent_buffer(path->nodes[0]);
2684                 path->nodes[0] = left;
2685                 path->slots[1] -= 1;
2686         } else {
2687                 btrfs_tree_unlock(left);
2688                 free_extent_buffer(left);
2689                 path->slots[0] -= push_items;
2690         }
2691         BUG_ON(path->slots[0] < 0);
2692         return ret;
2693 out:
2694         btrfs_tree_unlock(left);
2695         free_extent_buffer(left);
2696         return ret;
2697 }
2698
2699 /*
2700  * push some data in the path leaf to the left, trying to free up at
2701  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
2702  *
2703  * max_slot can put a limit on how far into the leaf we'll push items.  The
2704  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us push all the
2705  * items
2706  */
2707 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
2708                           *root, struct btrfs_path *path, int min_data_size,
2709                           int data_size, int empty, u32 max_slot)
2710 {
2711         struct extent_buffer *right = path->nodes[0];
2712         struct extent_buffer *left;
2713         int slot;
2714         int free_space;
2715         u32 right_nritems;
2716         int ret = 0;
2717
2718         slot = path->slots[1];
2719         if (slot == 0)
2720                 return 1;
2721         if (!path->nodes[1])
2722                 return 1;
2723
2724         right_nritems = btrfs_header_nritems(right);
2725         if (right_nritems == 0)
2726                 return 1;
2727
2728         btrfs_assert_tree_locked(path->nodes[1]);
2729
2730         left = read_node_slot(root, path->nodes[1], slot - 1);
2731         if (left == NULL)
2732                 return 1;
2733
2734         btrfs_tree_lock(left);
2735         btrfs_set_lock_blocking(left);
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         /* cow and double check */
2744         ret = btrfs_cow_block(trans, root, left,
2745                               path->nodes[1], slot - 1, &left);
2746         if (ret) {
2747                 /* we hit -ENOSPC, but it isn't fatal here */
2748                 if (ret == -ENOSPC)
2749                         ret = 1;
2750                 goto out;
2751         }
2752
2753         free_space = btrfs_leaf_free_space(root, left);
2754         if (free_space < data_size) {
2755                 ret = 1;
2756                 goto out;
2757         }
2758
2759         return __push_leaf_left(trans, root, path, min_data_size,
2760                                empty, left, free_space, right_nritems,
2761                                max_slot);
2762 out:
2763         btrfs_tree_unlock(left);
2764         free_extent_buffer(left);
2765         return ret;
2766 }
2767
2768 /*
2769  * split the path's leaf in two, making sure there is at least data_size
2770  * available for the resulting leaf level of the path.
2771  */
2772 static noinline void copy_for_split(struct btrfs_trans_handle *trans,
2773                                     struct btrfs_root *root,
2774                                     struct btrfs_path *path,
2775                                     struct extent_buffer *l,
2776                                     struct extent_buffer *right,
2777                                     int slot, int mid, int nritems)
2778 {
2779         int data_copy_size;
2780         int rt_data_off;
2781         int i;
2782         struct btrfs_disk_key disk_key;
2783         struct btrfs_map_token token;
2784
2785         btrfs_init_map_token(&token);
2786
2787         nritems = nritems - mid;
2788         btrfs_set_header_nritems(right, nritems);
2789         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2790
2791         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2792                            btrfs_item_nr_offset(mid),
2793                            nritems * sizeof(struct btrfs_item));
2794
2795         copy_extent_buffer(right, l,
2796                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2797                      data_copy_size, btrfs_leaf_data(l) +
2798                      leaf_data_end(root, l), data_copy_size);
2799
2800         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2801                       btrfs_item_end_nr(l, mid);
2802
2803         for (i = 0; i < nritems; i++) {
2804                 struct btrfs_item *item = btrfs_item_nr(right, i);
2805                 u32 ioff;
2806
2807                 ioff = btrfs_token_item_offset(right, item, &token);
2808                 btrfs_set_token_item_offset(right, item,
2809                                             ioff + rt_data_off, &token);
2810         }
2811
2812         btrfs_set_header_nritems(l, mid);
2813         btrfs_item_key(right, &disk_key, 0);
2814         insert_ptr(trans, root, path, &disk_key, right->start,
2815                    path->slots[1] + 1, 1);
2816
2817         btrfs_mark_buffer_dirty(right);
2818         btrfs_mark_buffer_dirty(l);
2819         BUG_ON(path->slots[0] != slot);
2820
2821         if (mid <= slot) {
2822                 btrfs_tree_unlock(path->nodes[0]);
2823                 free_extent_buffer(path->nodes[0]);
2824                 path->nodes[0] = right;
2825                 path->slots[0] -= mid;
2826                 path->slots[1] += 1;
2827         } else {
2828                 btrfs_tree_unlock(right);
2829                 free_extent_buffer(right);
2830         }
2831
2832         BUG_ON(path->slots[0] < 0);
2833 }
2834
2835 /*
2836  * double splits happen when we need to insert a big item in the middle
2837  * of a leaf.  A double split can leave us with 3 mostly empty leaves:
2838  * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
2839  *          A                 B                 C
2840  *
2841  * We avoid this by trying to push the items on either side of our target
2842  * into the adjacent leaves.  If all goes well we can avoid the double split
2843  * completely.
2844  */
2845 static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
2846                                           struct btrfs_root *root,
2847                                           struct btrfs_path *path,
2848                                           int data_size)
2849 {
2850         int ret;
2851         int progress = 0;
2852         int slot;
2853         u32 nritems;
2854
2855         slot = path->slots[0];
2856
2857         /*
2858          * try to push all the items after our slot into the
2859          * right leaf
2860          */
2861         ret = push_leaf_right(trans, root, path, 1, data_size, 0, slot);
2862         if (ret < 0)
2863                 return ret;
2864
2865         if (ret == 0)
2866                 progress++;
2867
2868         nritems = btrfs_header_nritems(path->nodes[0]);
2869         /*
2870          * our goal is to get our slot at the start or end of a leaf.  If
2871          * we've done so we're done
2872          */
2873         if (path->slots[0] == 0 || path->slots[0] == nritems)
2874                 return 0;
2875
2876         if (btrfs_leaf_free_space(root, path->nodes[0]) >= data_size)
2877                 return 0;
2878
2879         /* try to push all the items before our slot into the next leaf */
2880         slot = path->slots[0];
2881         ret = push_leaf_left(trans, root, path, 1, data_size, 0, slot);
2882         if (ret < 0)
2883                 return ret;
2884
2885         if (ret == 0)
2886                 progress++;
2887
2888         if (progress)
2889                 return 0;
2890         return 1;
2891 }
2892
2893 /*
2894  * split the path's leaf in two, making sure there is at least data_size
2895  * available for the resulting leaf level of the path.
2896  *
2897  * returns 0 if all went well and < 0 on failure.
2898  */
2899 static noinline int split_leaf(struct btrfs_trans_handle *trans,
2900                                struct btrfs_root *root,
2901                                struct btrfs_key *ins_key,
2902                                struct btrfs_path *path, int data_size,
2903                                int extend)
2904 {
2905         struct btrfs_disk_key disk_key;
2906         struct extent_buffer *l;
2907         u32 nritems;
2908         int mid;
2909         int slot;
2910         struct extent_buffer *right;
2911         int ret = 0;
2912         int wret;
2913         int split;
2914         int num_doubles = 0;
2915         int tried_avoid_double = 0;
2916
2917         l = path->nodes[0];
2918         slot = path->slots[0];
2919         if (extend && data_size + btrfs_item_size_nr(l, slot) +
2920             sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(root))
2921                 return -EOVERFLOW;
2922
2923         /* first try to make some room by pushing left and right */
2924         if (data_size) {
2925                 wret = push_leaf_right(trans, root, path, data_size,
2926                                        data_size, 0, 0);
2927                 if (wret < 0)
2928                         return wret;
2929                 if (wret) {
2930                         wret = push_leaf_left(trans, root, path, data_size,
2931                                               data_size, 0, (u32)-1);
2932                         if (wret < 0)
2933                                 return wret;
2934                 }
2935                 l = path->nodes[0];
2936
2937                 /* did the pushes work? */
2938                 if (btrfs_leaf_free_space(root, l) >= data_size)
2939                         return 0;
2940         }
2941
2942         if (!path->nodes[1]) {
2943                 ret = insert_new_root(trans, root, path, 1);
2944                 if (ret)
2945                         return ret;
2946         }
2947 again:
2948         split = 1;
2949         l = path->nodes[0];
2950         slot = path->slots[0];
2951         nritems = btrfs_header_nritems(l);
2952         mid = (nritems + 1) / 2;
2953
2954         if (mid <= slot) {
2955                 if (nritems == 1 ||
2956                     leaf_space_used(l, mid, nritems - mid) + data_size >
2957                         BTRFS_LEAF_DATA_SIZE(root)) {
2958                         if (slot >= nritems) {
2959                                 split = 0;
2960                         } else {
2961                                 mid = slot;
2962                                 if (mid != nritems &&
2963                                     leaf_space_used(l, mid, nritems - mid) +
2964                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2965                                         if (data_size && !tried_avoid_double)
2966                                                 goto push_for_double;
2967                                         split = 2;
2968                                 }
2969                         }
2970                 }
2971         } else {
2972                 if (leaf_space_used(l, 0, mid) + data_size >
2973                         BTRFS_LEAF_DATA_SIZE(root)) {
2974                         if (!extend && data_size && slot == 0) {
2975                                 split = 0;
2976                         } else if ((extend || !data_size) && slot == 0) {
2977                                 mid = 1;
2978                         } else {
2979                                 mid = slot;
2980                                 if (mid != nritems &&
2981                                     leaf_space_used(l, mid, nritems - mid) +
2982                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2983                                         if (data_size && !tried_avoid_double)
2984                                                 goto push_for_double;
2985                                         split = 2 ;
2986                                 }
2987                         }
2988                 }
2989         }
2990
2991         if (split == 0)
2992                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2993         else
2994                 btrfs_item_key(l, &disk_key, mid);
2995
2996         right = btrfs_alloc_free_block(trans, root, root->leafsize, 0,
2997                                         root->root_key.objectid,
2998                                         &disk_key, 0, l->start, 0, 0);
2999         if (IS_ERR(right))
3000                 return PTR_ERR(right);
3001
3002         root_add_used(root, root->leafsize);
3003
3004         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
3005         btrfs_set_header_bytenr(right, right->start);
3006         btrfs_set_header_generation(right, trans->transid);
3007         btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
3008         btrfs_set_header_owner(right, root->root_key.objectid);
3009         btrfs_set_header_level(right, 0);
3010         write_extent_buffer(right, root->fs_info->fsid,
3011                             (unsigned long)btrfs_header_fsid(right),
3012                             BTRFS_FSID_SIZE);
3013
3014         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
3015                             (unsigned long)btrfs_header_chunk_tree_uuid(right),
3016                             BTRFS_UUID_SIZE);
3017
3018         if (split == 0) {
3019                 if (mid <= slot) {
3020                         btrfs_set_header_nritems(right, 0);
3021                         insert_ptr(trans, root, path, &disk_key, right->start,
3022                                    path->slots[1] + 1, 1);
3023                         btrfs_tree_unlock(path->nodes[0]);
3024                         free_extent_buffer(path->nodes[0]);
3025                         path->nodes[0] = right;
3026                         path->slots[0] = 0;
3027                         path->slots[1] += 1;
3028                 } else {
3029                         btrfs_set_header_nritems(right, 0);
3030                         insert_ptr(trans, root, path, &disk_key, right->start,
3031                                           path->slots[1], 1);
3032                         btrfs_tree_unlock(path->nodes[0]);
3033                         free_extent_buffer(path->nodes[0]);
3034                         path->nodes[0] = right;
3035                         path->slots[0] = 0;
3036                         if (path->slots[1] == 0)
3037                                 fixup_low_keys(trans, root, path,
3038                                                &disk_key, 1);
3039                 }
3040                 btrfs_mark_buffer_dirty(right);
3041                 return ret;
3042         }
3043
3044         copy_for_split(trans, root, path, l, right, slot, mid, nritems);
3045
3046         if (split == 2) {
3047                 BUG_ON(num_doubles != 0);
3048                 num_doubles++;
3049                 goto again;
3050         }
3051
3052         return 0;
3053
3054 push_for_double:
3055         push_for_double_split(trans, root, path, data_size);
3056         tried_avoid_double = 1;
3057         if (btrfs_leaf_free_space(root, path->nodes[0]) >= data_size)
3058                 return 0;
3059         goto again;
3060 }
3061
3062 static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
3063                                          struct btrfs_root *root,
3064                                          struct btrfs_path *path, int ins_len)
3065 {
3066         struct btrfs_key key;
3067         struct extent_buffer *leaf;
3068         struct btrfs_file_extent_item *fi;
3069         u64 extent_len = 0;
3070         u32 item_size;
3071         int ret;
3072
3073         leaf = path->nodes[0];
3074         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3075
3076         BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
3077                key.type != BTRFS_EXTENT_CSUM_KEY);
3078
3079         if (btrfs_leaf_free_space(root, leaf) >= ins_len)
3080                 return 0;
3081
3082         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3083         if (key.type == BTRFS_EXTENT_DATA_KEY) {
3084                 fi = btrfs_item_ptr(leaf, path->slots[0],
3085                                     struct btrfs_file_extent_item);
3086                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
3087         }
3088         btrfs_release_path(path);
3089
3090         path->keep_locks = 1;
3091         path->search_for_split = 1;
3092         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
3093         path->search_for_split = 0;
3094         if (ret < 0)
3095                 goto err;
3096
3097         ret = -EAGAIN;
3098         leaf = path->nodes[0];
3099         /* if our item isn't there or got smaller, return now */
3100         if (ret > 0 || item_size != btrfs_item_size_nr(leaf, path->slots[0]))
3101                 goto err;
3102
3103         /* the leaf has  changed, it now has room.  return now */
3104         if (btrfs_leaf_free_space(root, path->nodes[0]) >= ins_len)
3105                 goto err;
3106
3107         if (key.type == BTRFS_EXTENT_DATA_KEY) {
3108                 fi = btrfs_item_ptr(leaf, path->slots[0],
3109                                     struct btrfs_file_extent_item);
3110                 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
3111                         goto err;
3112         }
3113
3114         btrfs_set_path_blocking(path);
3115         ret = split_leaf(trans, root, &key, path, ins_len, 1);
3116         if (ret)
3117                 goto err;
3118
3119         path->keep_locks = 0;
3120         btrfs_unlock_up_safe(path, 1);
3121         return 0;
3122 err:
3123         path->keep_locks = 0;
3124         return ret;
3125 }
3126
3127 static noinline int split_item(struct btrfs_trans_handle *trans,
3128                                struct btrfs_root *root,
3129                                struct btrfs_path *path,
3130                                struct btrfs_key *new_key,
3131                                unsigned long split_offset)
3132 {
3133         struct extent_buffer *leaf;
3134         struct btrfs_item *item;
3135         struct btrfs_item *new_item;
3136         int slot;
3137         char *buf;
3138         u32 nritems;
3139         u32 item_size;
3140         u32 orig_offset;
3141         struct btrfs_disk_key disk_key;
3142
3143         leaf = path->nodes[0];
3144         BUG_ON(btrfs_leaf_free_space(root, leaf) < sizeof(struct btrfs_item));
3145
3146         btrfs_set_path_blocking(path);
3147
3148         item = btrfs_item_nr(leaf, path->slots[0]);
3149         orig_offset = btrfs_item_offset(leaf, item);
3150         item_size = btrfs_item_size(leaf, item);
3151
3152         buf = kmalloc(item_size, GFP_NOFS);
3153         if (!buf)
3154                 return -ENOMEM;
3155
3156         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
3157                             path->slots[0]), item_size);
3158
3159         slot = path->slots[0] + 1;
3160         nritems = btrfs_header_nritems(leaf);
3161         if (slot != nritems) {
3162                 /* shift the items */
3163                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
3164                                 btrfs_item_nr_offset(slot),
3165                                 (nritems - slot) * sizeof(struct btrfs_item));
3166         }
3167
3168         btrfs_cpu_key_to_disk(&disk_key, new_key);
3169         btrfs_set_item_key(leaf, &disk_key, slot);
3170
3171         new_item = btrfs_item_nr(leaf, slot);
3172
3173         btrfs_set_item_offset(leaf, new_item, orig_offset);
3174         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
3175
3176         btrfs_set_item_offset(leaf, item,
3177                               orig_offset + item_size - split_offset);
3178         btrfs_set_item_size(leaf, item, split_offset);
3179
3180         btrfs_set_header_nritems(leaf, nritems + 1);
3181
3182         /* write the data for the start of the original item */
3183         write_extent_buffer(leaf, buf,
3184                             btrfs_item_ptr_offset(leaf, path->slots[0]),
3185                             split_offset);
3186
3187         /* write the data for the new item */
3188         write_extent_buffer(leaf, buf + split_offset,
3189                             btrfs_item_ptr_offset(leaf, slot),
3190                             item_size - split_offset);
3191         btrfs_mark_buffer_dirty(leaf);
3192
3193         BUG_ON(btrfs_leaf_free_space(root, leaf) < 0);
3194         kfree(buf);
3195         return 0;
3196 }
3197
3198 /*
3199  * This function splits a single item into two items,
3200  * giving 'new_key' to the new item and splitting the
3201  * old one at split_offset (from the start of the item).
3202  *
3203  * The path may be released by this operation.  After
3204  * the split, the path is pointing to the old item.  The
3205  * new item is going to be in the same node as the old one.
3206  *
3207  * Note, the item being split must be smaller enough to live alone on
3208  * a tree block with room for one extra struct btrfs_item
3209  *
3210  * This allows us to split the item in place, keeping a lock on the
3211  * leaf the entire time.
3212  */
3213 int btrfs_split_item(struct btrfs_trans_handle *trans,
3214                      struct btrfs_root *root,
3215                      struct btrfs_path *path,
3216                      struct btrfs_key *new_key,
3217                      unsigned long split_offset)
3218 {
3219         int ret;
3220         ret = setup_leaf_for_split(trans, root, path,
3221                                    sizeof(struct btrfs_item));
3222         if (ret)
3223                 return ret;
3224
3225         ret = split_item(trans, root, path, new_key, split_offset);
3226         return ret;
3227 }
3228
3229 /*
3230  * This function duplicate a item, giving 'new_key' to the new item.
3231  * It guarantees both items live in the same tree leaf and the new item
3232  * is contiguous with the original item.
3233  *
3234  * This allows us to split file extent in place, keeping a lock on the
3235  * leaf the entire time.
3236  */
3237 int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
3238                          struct btrfs_root *root,
3239                          struct btrfs_path *path,
3240                          struct btrfs_key *new_key)
3241 {
3242         struct extent_buffer *leaf;
3243         int ret;
3244         u32 item_size;
3245
3246         leaf = path->nodes[0];
3247         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3248         ret = setup_leaf_for_split(trans, root, path,
3249                                    item_size + sizeof(struct btrfs_item));
3250         if (ret)
3251                 return ret;
3252
3253         path->slots[0]++;
3254         setup_items_for_insert(trans, root, path, new_key, &item_size,
3255                                item_size, item_size +
3256                                sizeof(struct btrfs_item), 1);
3257         leaf = path->nodes[0];
3258         memcpy_extent_buffer(leaf,
3259                              btrfs_item_ptr_offset(leaf, path->slots[0]),
3260                              btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
3261                              item_size);
3262         return 0;
3263 }
3264
3265 /*
3266  * make the item pointed to by the path smaller.  new_size indicates
3267  * how small to make it, and from_end tells us if we just chop bytes
3268  * off the end of the item or if we shift the item to chop bytes off
3269  * the front.
3270  */
3271 void btrfs_truncate_item(struct btrfs_trans_handle *trans,
3272                          struct btrfs_root *root,
3273                          struct btrfs_path *path,
3274                          u32 new_size, int from_end)
3275 {
3276         int slot;
3277         struct extent_buffer *leaf;
3278         struct btrfs_item *item;
3279         u32 nritems;
3280         unsigned int data_end;
3281         unsigned int old_data_start;
3282         unsigned int old_size;
3283         unsigned int size_diff;
3284         int i;
3285         struct btrfs_map_token token;
3286
3287         btrfs_init_map_token(&token);
3288
3289         leaf = path->nodes[0];
3290         slot = path->slots[0];
3291
3292         old_size = btrfs_item_size_nr(leaf, slot);
3293         if (old_size == new_size)
3294                 return;
3295
3296         nritems = btrfs_header_nritems(leaf);
3297         data_end = leaf_data_end(root, leaf);
3298
3299         old_data_start = btrfs_item_offset_nr(leaf, slot);
3300
3301         size_diff = old_size - new_size;
3302
3303         BUG_ON(slot < 0);
3304         BUG_ON(slot >= nritems);
3305
3306         /*
3307          * item0..itemN ... dataN.offset..dataN.size .. data0.size
3308          */
3309         /* first correct the data pointers */
3310         for (i = slot; i < nritems; i++) {
3311                 u32 ioff;
3312                 item = btrfs_item_nr(leaf, i);
3313
3314                 ioff = btrfs_token_item_offset(leaf, item, &token);
3315                 btrfs_set_token_item_offset(leaf, item,
3316                                             ioff + size_diff, &token);
3317         }
3318
3319         /* shift the data */
3320         if (from_end) {
3321                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3322                               data_end + size_diff, btrfs_leaf_data(leaf) +
3323                               data_end, old_data_start + new_size - data_end);
3324         } else {
3325                 struct btrfs_disk_key disk_key;
3326                 u64 offset;
3327
3328                 btrfs_item_key(leaf, &disk_key, slot);
3329
3330                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
3331                         unsigned long ptr;
3332                         struct btrfs_file_extent_item *fi;
3333
3334                         fi = btrfs_item_ptr(leaf, slot,
3335                                             struct btrfs_file_extent_item);
3336                         fi = (struct btrfs_file_extent_item *)(
3337                              (unsigned long)fi - size_diff);
3338
3339                         if (btrfs_file_extent_type(leaf, fi) ==
3340                             BTRFS_FILE_EXTENT_INLINE) {
3341                                 ptr = btrfs_item_ptr_offset(leaf, slot);
3342                                 memmove_extent_buffer(leaf, ptr,
3343                                       (unsigned long)fi,
3344                                       offsetof(struct btrfs_file_extent_item,
3345                                                  disk_bytenr));
3346                         }
3347                 }
3348
3349                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3350                               data_end + size_diff, btrfs_leaf_data(leaf) +
3351                               data_end, old_data_start - data_end);
3352
3353                 offset = btrfs_disk_key_offset(&disk_key);
3354                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
3355                 btrfs_set_item_key(leaf, &disk_key, slot);
3356                 if (slot == 0)
3357                         fixup_low_keys(trans, root, path, &disk_key, 1);
3358         }
3359
3360         item = btrfs_item_nr(leaf, slot);
3361         btrfs_set_item_size(leaf, item, new_size);
3362         btrfs_mark_buffer_dirty(leaf);
3363
3364         if (btrfs_leaf_free_space(root, leaf) < 0) {
3365                 btrfs_print_leaf(root, leaf);
3366                 BUG();
3367         }
3368 }
3369
3370 /*
3371  * make the item pointed to by the path bigger, data_size is the new size.
3372  */
3373 void btrfs_extend_item(struct btrfs_trans_handle *trans,
3374                        struct btrfs_root *root, struct btrfs_path *path,
3375                        u32 data_size)
3376 {
3377         int slot;
3378         struct extent_buffer *leaf;
3379         struct btrfs_item *item;
3380         u32 nritems;
3381         unsigned int data_end;
3382         unsigned int old_data;
3383         unsigned int old_size;
3384         int i;
3385         struct btrfs_map_token token;
3386
3387         btrfs_init_map_token(&token);
3388
3389         leaf = path->nodes[0];
3390
3391         nritems = btrfs_header_nritems(leaf);
3392         data_end = leaf_data_end(root, leaf);
3393
3394         if (btrfs_leaf_free_space(root, leaf) < data_size) {
3395                 btrfs_print_leaf(root, leaf);
3396                 BUG();
3397         }
3398         slot = path->slots[0];
3399         old_data = btrfs_item_end_nr(leaf, slot);
3400
3401         BUG_ON(slot < 0);
3402         if (slot >= nritems) {
3403                 btrfs_print_leaf(root, leaf);
3404                 printk(KERN_CRIT "slot %d too large, nritems %d\n",
3405                        slot, nritems);
3406                 BUG_ON(1);
3407         }
3408
3409         /*
3410          * item0..itemN ... dataN.offset..dataN.size .. data0.size
3411          */
3412         /* first correct the data pointers */
3413         for (i = slot; i < nritems; i++) {
3414                 u32 ioff;
3415                 item = btrfs_item_nr(leaf, i);
3416
3417                 ioff = btrfs_token_item_offset(leaf, item, &token);
3418                 btrfs_set_token_item_offset(leaf, item,
3419                                             ioff - data_size, &token);
3420         }
3421
3422         /* shift the data */
3423         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3424                       data_end - data_size, btrfs_leaf_data(leaf) +
3425                       data_end, old_data - data_end);
3426
3427         data_end = old_data;
3428         old_size = btrfs_item_size_nr(leaf, slot);
3429         item = btrfs_item_nr(leaf, slot);
3430         btrfs_set_item_size(leaf, item, old_size + data_size);
3431         btrfs_mark_buffer_dirty(leaf);
3432
3433         if (btrfs_leaf_free_space(root, leaf) < 0) {
3434                 btrfs_print_leaf(root, leaf);
3435                 BUG();
3436         }
3437 }
3438
3439 /*
3440  * Given a key and some data, insert items into the tree.
3441  * This does all the path init required, making room in the tree if needed.
3442  * Returns the number of keys that were inserted.
3443  */
3444 int btrfs_insert_some_items(struct btrfs_trans_handle *trans,
3445                             struct btrfs_root *root,
3446                             struct btrfs_path *path,
3447                             struct btrfs_key *cpu_key, u32 *data_size,
3448                             int nr)
3449 {
3450         struct extent_buffer *leaf;
3451         struct btrfs_item *item;
3452         int ret = 0;
3453         int slot;
3454         int i;
3455         u32 nritems;
3456         u32 total_data = 0;
3457         u32 total_size = 0;
3458         unsigned int data_end;
3459         struct btrfs_disk_key disk_key;
3460         struct btrfs_key found_key;
3461         struct btrfs_map_token token;
3462
3463         btrfs_init_map_token(&token);
3464
3465         for (i = 0; i < nr; i++) {
3466                 if (total_size + data_size[i] + sizeof(struct btrfs_item) >
3467                     BTRFS_LEAF_DATA_SIZE(root)) {
3468                         break;
3469                         nr = i;
3470                 }
3471                 total_data += data_size[i];
3472                 total_size += data_size[i] + sizeof(struct btrfs_item);
3473         }
3474         BUG_ON(nr == 0);
3475
3476         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
3477         if (ret == 0)
3478                 return -EEXIST;
3479         if (ret < 0)
3480                 goto out;
3481
3482         leaf = path->nodes[0];
3483
3484         nritems = btrfs_header_nritems(leaf);
3485         data_end = leaf_data_end(root, leaf);
3486
3487         if (btrfs_leaf_free_space(root, leaf) < total_size) {
3488                 for (i = nr; i >= 0; i--) {
3489                         total_data -= data_size[i];
3490                         total_size -= data_size[i] + sizeof(struct btrfs_item);
3491                         if (total_size < btrfs_leaf_free_space(root, leaf))
3492                                 break;
3493                 }
3494                 nr = i;
3495         }
3496
3497         slot = path->slots[0];
3498         BUG_ON(slot < 0);
3499
3500         if (slot != nritems) {
3501                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
3502
3503                 item = btrfs_item_nr(leaf, slot);
3504                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3505
3506                 /* figure out how many keys we can insert in here */
3507                 total_data = data_size[0];
3508                 for (i = 1; i < nr; i++) {
3509                         if (btrfs_comp_cpu_keys(&found_key, cpu_key + i) <= 0)
3510                                 break;
3511                         total_data += data_size[i];
3512                 }
3513                 nr = i;
3514
3515                 if (old_data < data_end) {
3516                         btrfs_print_leaf(root, leaf);
3517                         printk(KERN_CRIT "slot %d old_data %d data_end %d\n",
3518                                slot, old_data, data_end);
3519                         BUG_ON(1);
3520                 }
3521                 /*
3522                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
3523                  */
3524                 /* first correct the data pointers */
3525                 for (i = slot; i < nritems; i++) {
3526                         u32 ioff;
3527
3528                         item = btrfs_item_nr(leaf, i);
3529                         ioff = btrfs_token_item_offset(leaf, item, &token);
3530                         btrfs_set_token_item_offset(leaf, item,
3531                                                     ioff - total_data, &token);
3532                 }
3533                 /* shift the items */
3534                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
3535                               btrfs_item_nr_offset(slot),
3536                               (nritems - slot) * sizeof(struct btrfs_item));
3537
3538                 /* shift the data */
3539                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3540                               data_end - total_data, btrfs_leaf_data(leaf) +
3541                               data_end, old_data - data_end);
3542                 data_end = old_data;
3543         } else {
3544                 /*
3545                  * this sucks but it has to be done, if we are inserting at
3546                  * the end of the leaf only insert 1 of the items, since we
3547                  * have no way of knowing whats on the next leaf and we'd have
3548                  * to drop our current locks to figure it out
3549                  */
3550                 nr = 1;
3551         }
3552
3553         /* setup the item for the new data */
3554         for (i = 0; i < nr; i++) {
3555                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
3556                 btrfs_set_item_key(leaf, &disk_key, slot + i);
3557                 item = btrfs_item_nr(leaf, slot + i);
3558                 btrfs_set_token_item_offset(leaf, item,
3559                                             data_end - data_size[i], &token);
3560                 data_end -= data_size[i];
3561                 btrfs_set_token_item_size(leaf, item, data_size[i], &token);
3562         }
3563         btrfs_set_header_nritems(leaf, nritems + nr);
3564         btrfs_mark_buffer_dirty(leaf);
3565
3566         ret = 0;
3567         if (slot == 0) {
3568                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
3569                 fixup_low_keys(trans, root, path, &disk_key, 1);
3570         }
3571
3572         if (btrfs_leaf_free_space(root, leaf) < 0) {
3573                 btrfs_print_leaf(root, leaf);
3574                 BUG();
3575         }
3576 out:
3577         if (!ret)
3578                 ret = nr;
3579         return ret;
3580 }
3581
3582 /*
3583  * this is a helper for btrfs_insert_empty_items, the main goal here is
3584  * to save stack depth by doing the bulk of the work in a function
3585  * that doesn't call btrfs_search_slot
3586  */
3587 void setup_items_for_insert(struct btrfs_trans_handle *trans,
3588                             struct btrfs_root *root, struct btrfs_path *path,
3589                             struct btrfs_key *cpu_key, u32 *data_size,
3590                             u32 total_data, u32 total_size, int nr)
3591 {
3592         struct btrfs_item *item;
3593         int i;
3594         u32 nritems;
3595         unsigned int data_end;
3596         struct btrfs_disk_key disk_key;
3597         struct extent_buffer *leaf;
3598         int slot;
3599         struct btrfs_map_token token;
3600
3601         btrfs_init_map_token(&token);
3602
3603         leaf = path->nodes[0];
3604         slot = path->slots[0];
3605
3606         nritems = btrfs_header_nritems(leaf);
3607         data_end = leaf_data_end(root, leaf);
3608
3609         if (btrfs_leaf_free_space(root, leaf) < total_size) {
3610                 btrfs_print_leaf(root, leaf);
3611                 printk(KERN_CRIT "not enough freespace need %u have %d\n",
3612                        total_size, btrfs_leaf_free_space(root, leaf));
3613                 BUG();
3614         }
3615
3616         if (slot != nritems) {
3617                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
3618
3619                 if (old_data < data_end) {
3620                         btrfs_print_leaf(root, leaf);
3621                         printk(KERN_CRIT "slot %d old_data %d data_end %d\n",
3622                                slot, old_data, data_end);
3623                         BUG_ON(1);
3624                 }
3625                 /*
3626                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
3627                  */
3628                 /* first correct the data pointers */
3629                 for (i = slot; i < nritems; i++) {
3630                         u32 ioff;
3631
3632                         item = btrfs_item_nr(leaf, i);
3633                         ioff = btrfs_token_item_offset(leaf, item, &token);
3634                         btrfs_set_token_item_offset(leaf, item,
3635                                                     ioff - total_data, &token);
3636                 }
3637                 /* shift the items */
3638                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
3639                               btrfs_item_nr_offset(slot),
3640                               (nritems - slot) * sizeof(struct btrfs_item));
3641
3642                 /* shift the data */
3643                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3644                               data_end - total_data, btrfs_leaf_data(leaf) +
3645                               data_end, old_data - data_end);
3646                 data_end = old_data;
3647         }
3648
3649         /* setup the item for the new data */
3650         for (i = 0; i < nr; i++) {
3651                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
3652                 btrfs_set_item_key(leaf, &disk_key, slot + i);
3653                 item = btrfs_item_nr(leaf, slot + i);
3654                 btrfs_set_token_item_offset(leaf, item,
3655                                             data_end - data_size[i], &token);
3656                 data_end -= data_size[i];
3657                 btrfs_set_token_item_size(leaf, item, data_size[i], &token);
3658         }
3659
3660         btrfs_set_header_nritems(leaf, nritems + nr);
3661
3662         if (slot == 0) {
3663                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
3664                 fixup_low_keys(trans, root, path, &disk_key, 1);
3665         }
3666         btrfs_unlock_up_safe(path, 1);
3667         btrfs_mark_buffer_dirty(leaf);
3668
3669         if (btrfs_leaf_free_space(root, leaf) < 0) {
3670                 btrfs_print_leaf(root, leaf);
3671                 BUG();
3672         }
3673 }
3674
3675 /*
3676  * Given a key and some data, insert items into the tree.
3677  * This does all the path init required, making room in the tree if needed.
3678  */
3679 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
3680                             struct btrfs_root *root,
3681                             struct btrfs_path *path,
3682                             struct btrfs_key *cpu_key, u32 *data_size,
3683                             int nr)
3684 {
3685         int ret = 0;
3686         int slot;
3687         int i;
3688         u32 total_size = 0;
3689         u32 total_data = 0;
3690
3691         for (i = 0; i < nr; i++)
3692                 total_data += data_size[i];
3693
3694         total_size = total_data + (nr * sizeof(struct btrfs_item));
3695         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
3696         if (ret == 0)
3697                 return -EEXIST;
3698         if (ret < 0)
3699                 return ret;
3700
3701         slot = path->slots[0];
3702         BUG_ON(slot < 0);
3703
3704         setup_items_for_insert(trans, root, path, cpu_key, data_size,
3705                                total_data, total_size, nr);
3706         return 0;
3707 }
3708
3709 /*
3710  * Given a key and some data, insert an item into the tree.
3711  * This does all the path init required, making room in the tree if needed.
3712  */
3713 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
3714                       *root, struct btrfs_key *cpu_key, void *data, u32
3715                       data_size)
3716 {
3717         int ret = 0;
3718         struct btrfs_path *path;
3719         struct extent_buffer *leaf;
3720         unsigned long ptr;
3721
3722         path = btrfs_alloc_path();
3723         if (!path)
3724                 return -ENOMEM;
3725         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
3726         if (!ret) {
3727                 leaf = path->nodes[0];
3728                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3729                 write_extent_buffer(leaf, data, ptr, data_size);
3730                 btrfs_mark_buffer_dirty(leaf);
3731         }
3732         btrfs_free_path(path);
3733         return ret;
3734 }
3735
3736 /*
3737  * delete the pointer from a given node.
3738  *
3739  * the tree should have been previously balanced so the deletion does not
3740  * empty a node.
3741  */
3742 static void del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3743                     struct btrfs_path *path, int level, int slot)
3744 {
3745         struct extent_buffer *parent = path->nodes[level];
3746         u32 nritems;
3747
3748         nritems = btrfs_header_nritems(parent);
3749         if (slot != nritems - 1) {
3750                 memmove_extent_buffer(parent,
3751                               btrfs_node_key_ptr_offset(slot),
3752                               btrfs_node_key_ptr_offset(slot + 1),
3753                               sizeof(struct btrfs_key_ptr) *
3754                               (nritems - slot - 1));
3755         }
3756         nritems--;
3757         btrfs_set_header_nritems(parent, nritems);
3758         if (nritems == 0 && parent == root->node) {
3759                 BUG_ON(btrfs_header_level(root->node) != 1);
3760                 /* just turn the root into a leaf and break */
3761                 btrfs_set_header_level(root->node, 0);
3762         } else if (slot == 0) {
3763                 struct btrfs_disk_key disk_key;
3764
3765                 btrfs_node_key(parent, &disk_key, 0);
3766                 fixup_low_keys(trans, root, path, &disk_key, level + 1);
3767         }
3768         btrfs_mark_buffer_dirty(parent);
3769 }
3770
3771 /*
3772  * a helper function to delete the leaf pointed to by path->slots[1] and
3773  * path->nodes[1].
3774  *
3775  * This deletes the pointer in path->nodes[1] and frees the leaf
3776  * block extent.  zero is returned if it all worked out, < 0 otherwise.
3777  *
3778  * The path must have already been setup for deleting the leaf, including
3779  * all the proper balancing.  path->nodes[1] must be locked.
3780  */
3781 static noinline void btrfs_del_leaf(struct btrfs_trans_handle *trans,
3782                                     struct btrfs_root *root,
3783                                     struct btrfs_path *path,
3784                                     struct extent_buffer *leaf)
3785 {
3786         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
3787         del_ptr(trans, root, path, 1, path->slots[1]);
3788
3789         /*
3790          * btrfs_free_extent is expensive, we want to make sure we
3791          * aren't holding any locks when we call it
3792          */
3793         btrfs_unlock_up_safe(path, 0);
3794
3795         root_sub_used(root, leaf->len);
3796
3797         extent_buffer_get(leaf);
3798         btrfs_free_tree_block(trans, root, leaf, 0, 1, 0);
3799         free_extent_buffer_stale(leaf);
3800 }
3801 /*
3802  * delete the item at the leaf level in path.  If that empties
3803  * the leaf, remove it from the tree
3804  */
3805 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3806                     struct btrfs_path *path, int slot, int nr)
3807 {
3808         struct extent_buffer *leaf;
3809         struct btrfs_item *item;
3810         int last_off;
3811         int dsize = 0;
3812         int ret = 0;
3813         int wret;
3814         int i;
3815         u32 nritems;
3816         struct btrfs_map_token token;
3817
3818         btrfs_init_map_token(&token);
3819
3820         leaf = path->nodes[0];
3821         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
3822
3823         for (i = 0; i < nr; i++)
3824                 dsize += btrfs_item_size_nr(leaf, slot + i);
3825
3826         nritems = btrfs_header_nritems(leaf);
3827
3828         if (slot + nr != nritems) {
3829                 int data_end = leaf_data_end(root, leaf);
3830
3831                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3832                               data_end + dsize,
3833                               btrfs_leaf_data(leaf) + data_end,
3834                               last_off - data_end);
3835
3836                 for (i = slot + nr; i < nritems; i++) {
3837                         u32 ioff;
3838
3839                         item = btrfs_item_nr(leaf, i);
3840                         ioff = btrfs_token_item_offset(leaf, item, &token);
3841                         btrfs_set_token_item_offset(leaf, item,
3842                                                     ioff + dsize, &token);
3843                 }
3844
3845                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
3846                               btrfs_item_nr_offset(slot + nr),
3847                               sizeof(struct btrfs_item) *
3848                               (nritems - slot - nr));
3849         }
3850         btrfs_set_header_nritems(leaf, nritems - nr);
3851         nritems -= nr;
3852
3853         /* delete the leaf if we've emptied it */
3854         if (nritems == 0) {
3855                 if (leaf == root->node) {
3856                         btrfs_set_header_level(leaf, 0);
3857                 } else {
3858                         btrfs_set_path_blocking(path);
3859                         clean_tree_block(trans, root, leaf);
3860                         btrfs_del_leaf(trans, root, path, leaf);
3861                 }
3862         } else {
3863                 int used = leaf_space_used(leaf, 0, nritems);
3864                 if (slot == 0) {
3865                         struct btrfs_disk_key disk_key;
3866
3867                         btrfs_item_key(leaf, &disk_key, 0);
3868                         fixup_low_keys(trans, root, path, &disk_key, 1);
3869                 }
3870
3871                 /* delete the leaf if it is mostly empty */
3872                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 3) {
3873                         /* push_leaf_left fixes the path.
3874                          * make sure the path still points to our leaf
3875                          * for possible call to del_ptr below
3876                          */
3877                         slot = path->slots[1];
3878                         extent_buffer_get(leaf);
3879
3880                         btrfs_set_path_blocking(path);
3881                         wret = push_leaf_left(trans, root, path, 1, 1,
3882                                               1, (u32)-1);
3883                         if (wret < 0 && wret != -ENOSPC)
3884                                 ret = wret;
3885
3886                         if (path->nodes[0] == leaf &&
3887                             btrfs_header_nritems(leaf)) {
3888                                 wret = push_leaf_right(trans, root, path, 1,
3889                                                        1, 1, 0);
3890                                 if (wret < 0 && wret != -ENOSPC)
3891                                         ret = wret;
3892                         }
3893
3894                         if (btrfs_header_nritems(leaf) == 0) {
3895                                 path->slots[1] = slot;
3896                                 btrfs_del_leaf(trans, root, path, leaf);
3897                                 free_extent_buffer(leaf);
3898                                 ret = 0;
3899                         } else {
3900                                 /* if we're still in the path, make sure
3901                                  * we're dirty.  Otherwise, one of the
3902                                  * push_leaf functions must have already
3903                                  * dirtied this buffer
3904                                  */
3905                                 if (path->nodes[0] == leaf)
3906                                         btrfs_mark_buffer_dirty(leaf);
3907                                 free_extent_buffer(leaf);
3908                         }
3909                 } else {
3910                         btrfs_mark_buffer_dirty(leaf);
3911                 }
3912         }
3913         return ret;
3914 }
3915
3916 /*
3917  * search the tree again to find a leaf with lesser keys
3918  * returns 0 if it found something or 1 if there are no lesser leaves.
3919  * returns < 0 on io errors.
3920  *
3921  * This may release the path, and so you may lose any locks held at the
3922  * time you call it.
3923  */
3924 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
3925 {
3926         struct btrfs_key key;
3927         struct btrfs_disk_key found_key;
3928         int ret;
3929
3930         btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
3931
3932         if (key.offset > 0)
3933                 key.offset--;
3934         else if (key.type > 0)
3935                 key.type--;
3936         else if (key.objectid > 0)
3937                 key.objectid--;
3938         else
3939                 return 1;
3940
3941         btrfs_release_path(path);
3942         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3943         if (ret < 0)
3944                 return ret;
3945         btrfs_item_key(path->nodes[0], &found_key, 0);
3946         ret = comp_keys(&found_key, &key);
3947         if (ret < 0)
3948                 return 0;
3949         return 1;
3950 }
3951
3952 /*
3953  * A helper function to walk down the tree starting at min_key, and looking
3954  * for nodes or leaves that are either in cache or have a minimum
3955  * transaction id.  This is used by the btree defrag code, and tree logging
3956  *
3957  * This does not cow, but it does stuff the starting key it finds back
3958  * into min_key, so you can call btrfs_search_slot with cow=1 on the
3959  * key and get a writable path.
3960  *
3961  * This does lock as it descends, and path->keep_locks should be set
3962  * to 1 by the caller.
3963  *
3964  * This honors path->lowest_level to prevent descent past a given level
3965  * of the tree.
3966  *
3967  * min_trans indicates the oldest transaction that you are interested
3968  * in walking through.  Any nodes or leaves older than min_trans are
3969  * skipped over (without reading them).
3970  *
3971  * returns zero if something useful was found, < 0 on error and 1 if there
3972  * was nothing in the tree that matched the search criteria.
3973  */
3974 int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
3975                          struct btrfs_key *max_key,
3976                          struct btrfs_path *path, int cache_only,
3977                          u64 min_trans)
3978 {
3979         struct extent_buffer *cur;
3980         struct btrfs_key found_key;
3981         int slot;
3982         int sret;
3983         u32 nritems;
3984         int level;
3985         int ret = 1;
3986
3987         WARN_ON(!path->keep_locks);
3988 again:
3989         cur = btrfs_read_lock_root_node(root);
3990         level = btrfs_header_level(cur);
3991         WARN_ON(path->nodes[level]);
3992         path->nodes[level] = cur;
3993         path->locks[level] = BTRFS_READ_LOCK;
3994
3995         if (btrfs_header_generation(cur) < min_trans) {
3996                 ret = 1;
3997                 goto out;
3998         }
3999         while (1) {
4000                 nritems = btrfs_header_nritems(cur);
4001                 level = btrfs_header_level(cur);
4002                 sret = bin_search(cur, min_key, level, &slot);
4003
4004                 /* at the lowest level, we're done, setup the path and exit */
4005                 if (level == path->lowest_level) {
4006                         if (slot >= nritems)
4007                                 goto find_next_key;
4008                         ret = 0;
4009                         path->slots[level] = slot;
4010                         btrfs_item_key_to_cpu(cur, &found_key, slot);
4011                         goto out;
4012                 }
4013                 if (sret && slot > 0)
4014                         slot--;
4015                 /*
4016                  * check this node pointer against the cache_only and
4017                  * min_trans parameters.  If it isn't in cache or is too
4018                  * old, skip to the next one.
4019                  */
4020                 while (slot < nritems) {
4021                         u64 blockptr;
4022                         u64 gen;
4023                         struct extent_buffer *tmp;
4024                         struct btrfs_disk_key disk_key;
4025
4026                         blockptr = btrfs_node_blockptr(cur, slot);
4027                         gen = btrfs_node_ptr_generation(cur, slot);
4028                         if (gen < min_trans) {
4029                                 slot++;
4030                                 continue;
4031                         }
4032                         if (!cache_only)
4033                                 break;
4034
4035                         if (max_key) {
4036                                 btrfs_node_key(cur, &disk_key, slot);
4037                                 if (comp_keys(&disk_key, max_key) >= 0) {
4038                                         ret = 1;
4039                                         goto out;
4040                                 }
4041                         }
4042
4043                         tmp = btrfs_find_tree_block(root, blockptr,
4044                                             btrfs_level_size(root, level - 1));
4045
4046                         if (tmp && btrfs_buffer_uptodate(tmp, gen)) {
4047                                 free_extent_buffer(tmp);
4048                                 break;
4049                         }
4050                         if (tmp)
4051                                 free_extent_buffer(tmp);
4052                         slot++;
4053                 }
4054 find_next_key:
4055                 /*
4056                  * we didn't find a candidate key in this node, walk forward
4057                  * and find another one
4058                  */
4059                 if (slot >= nritems) {
4060                         path->slots[level] = slot;
4061                         btrfs_set_path_blocking(path);
4062                         sret = btrfs_find_next_key(root, path, min_key, level,
4063                                                   cache_only, min_trans);
4064                         if (sret == 0) {
4065                                 btrfs_release_path(path);
4066                                 goto again;
4067                         } else {
4068                                 goto out;
4069                         }
4070                 }
4071                 /* save our key for returning back */
4072                 btrfs_node_key_to_cpu(cur, &found_key, slot);
4073                 path->slots[level] = slot;
4074                 if (level == path->lowest_level) {
4075                         ret = 0;
4076                         unlock_up(path, level, 1, 0, NULL);
4077                         goto out;
4078                 }
4079                 btrfs_set_path_blocking(path);
4080                 cur = read_node_slot(root, cur, slot);
4081                 BUG_ON(!cur); /* -ENOMEM */
4082
4083                 btrfs_tree_read_lock(cur);
4084
4085                 path->locks[level - 1] = BTRFS_READ_LOCK;
4086                 path->nodes[level - 1] = cur;
4087                 unlock_up(path, level, 1, 0, NULL);
4088                 btrfs_clear_path_blocking(path, NULL, 0);
4089         }
4090 out:
4091         if (ret == 0)
4092                 memcpy(min_key, &found_key, sizeof(found_key));
4093         btrfs_set_path_blocking(path);
4094         return ret;
4095 }
4096
4097 /*
4098  * this is similar to btrfs_next_leaf, but does not try to preserve
4099  * and fixup the path.  It looks for and returns the next key in the
4100  * tree based on the current path and the cache_only and min_trans
4101  * parameters.
4102  *
4103  * 0 is returned if another key is found, < 0 if there are any errors
4104  * and 1 is returned if there are no higher keys in the tree
4105  *
4106  * path->keep_locks should be set to 1 on the search made before
4107  * calling this function.
4108  */
4109 int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
4110                         struct btrfs_key *key, int level,
4111                         int cache_only, u64 min_trans)
4112 {
4113         int slot;
4114         struct extent_buffer *c;
4115
4116         WARN_ON(!path->keep_locks);
4117         while (level < BTRFS_MAX_LEVEL) {
4118                 if (!path->nodes[level])
4119                         return 1;
4120
4121                 slot = path->slots[level] + 1;
4122                 c = path->nodes[level];
4123 next:
4124                 if (slot >= btrfs_header_nritems(c)) {
4125                         int ret;
4126                         int orig_lowest;
4127                         struct btrfs_key cur_key;
4128                         if (level + 1 >= BTRFS_MAX_LEVEL ||
4129                             !path->nodes[level + 1])
4130                                 return 1;
4131
4132                         if (path->locks[level + 1]) {
4133                                 level++;
4134                                 continue;
4135                         }
4136
4137                         slot = btrfs_header_nritems(c) - 1;
4138                         if (level == 0)
4139                                 btrfs_item_key_to_cpu(c, &cur_key, slot);
4140                         else
4141                                 btrfs_node_key_to_cpu(c, &cur_key, slot);
4142
4143                         orig_lowest = path->lowest_level;
4144                         btrfs_release_path(path);
4145                         path->lowest_level = level;
4146                         ret = btrfs_search_slot(NULL, root, &cur_key, path,
4147                                                 0, 0);
4148                         path->lowest_level = orig_lowest;
4149                         if (ret < 0)
4150                                 return ret;
4151
4152                         c = path->nodes[level];
4153                         slot = path->slots[level];
4154                         if (ret == 0)
4155                                 slot++;
4156                         goto next;
4157                 }
4158
4159                 if (level == 0)
4160                         btrfs_item_key_to_cpu(c, key, slot);
4161                 else {
4162                         u64 blockptr = btrfs_node_blockptr(c, slot);
4163                         u64 gen = btrfs_node_ptr_generation(c, slot);
4164
4165                         if (cache_only) {
4166                                 struct extent_buffer *cur;
4167                                 cur = btrfs_find_tree_block(root, blockptr,
4168                                             btrfs_level_size(root, level - 1));
4169                                 if (!cur || !btrfs_buffer_uptodate(cur, gen)) {
4170                                         slot++;
4171                                         if (cur)
4172                                                 free_extent_buffer(cur);
4173                                         goto next;
4174                                 }
4175                                 free_extent_buffer(cur);
4176                         }
4177                         if (gen < min_trans) {
4178                                 slot++;
4179                                 goto next;
4180                         }
4181                         btrfs_node_key_to_cpu(c, key, slot);
4182                 }
4183                 return 0;
4184         }
4185         return 1;
4186 }
4187
4188 /*
4189  * search the tree again to find a leaf with greater keys
4190  * returns 0 if it found something or 1 if there are no greater leaves.
4191  * returns < 0 on io errors.
4192  */
4193 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
4194 {
4195         int slot;
4196         int level;
4197         struct extent_buffer *c;
4198         struct extent_buffer *next;
4199         struct btrfs_key key;
4200         u32 nritems;
4201         int ret;
4202         int old_spinning = path->leave_spinning;
4203         int next_rw_lock = 0;
4204
4205         nritems = btrfs_header_nritems(path->nodes[0]);
4206         if (nritems == 0)
4207                 return 1;
4208
4209         btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
4210 again:
4211         level = 1;
4212         next = NULL;
4213         next_rw_lock = 0;
4214         btrfs_release_path(path);
4215
4216         path->keep_locks = 1;
4217         path->leave_spinning = 1;
4218
4219         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4220         path->keep_locks = 0;
4221
4222         if (ret < 0)
4223                 return ret;
4224
4225         nritems = btrfs_header_nritems(path->nodes[0]);
4226         /*
4227          * by releasing the path above we dropped all our locks.  A balance
4228          * could have added more items next to the key that used to be
4229          * at the very end of the block.  So, check again here and
4230          * advance the path if there are now more items available.
4231          */
4232         if (nritems > 0 && path->slots[0] < nritems - 1) {
4233                 if (ret == 0)
4234                         path->slots[0]++;
4235                 ret = 0;
4236                 goto done;
4237         }
4238
4239         while (level < BTRFS_MAX_LEVEL) {
4240                 if (!path->nodes[level]) {
4241                         ret = 1;
4242                         goto done;
4243                 }
4244
4245                 slot = path->slots[level] + 1;
4246                 c = path->nodes[level];
4247                 if (slot >= btrfs_header_nritems(c)) {
4248                         level++;
4249                         if (level == BTRFS_MAX_LEVEL) {
4250                                 ret = 1;
4251                                 goto done;
4252                         }
4253                         continue;
4254                 }
4255
4256                 if (next) {
4257                         btrfs_tree_unlock_rw(next, next_rw_lock);
4258                         free_extent_buffer(next);
4259                 }
4260
4261                 next = c;
4262                 next_rw_lock = path->locks[level];
4263                 ret = read_block_for_search(NULL, root, path, &next, level,
4264                                             slot, &key);
4265                 if (ret == -EAGAIN)
4266                         goto again;
4267
4268                 if (ret < 0) {
4269                         btrfs_release_path(path);
4270                         goto done;
4271                 }
4272
4273                 if (!path->skip_locking) {
4274                         ret = btrfs_try_tree_read_lock(next);
4275                         if (!ret) {
4276                                 btrfs_set_path_blocking(path);
4277                                 btrfs_tree_read_lock(next);
4278                                 btrfs_clear_path_blocking(path, next,
4279                                                           BTRFS_READ_LOCK);
4280                         }
4281                         next_rw_lock = BTRFS_READ_LOCK;
4282                 }
4283                 break;
4284         }
4285         path->slots[level] = slot;
4286         while (1) {
4287                 level--;
4288                 c = path->nodes[level];
4289                 if (path->locks[level])
4290                         btrfs_tree_unlock_rw(c, path->locks[level]);
4291
4292                 free_extent_buffer(c);
4293                 path->nodes[level] = next;
4294                 path->slots[level] = 0;
4295                 if (!path->skip_locking)
4296                         path->locks[level] = next_rw_lock;
4297                 if (!level)
4298                         break;
4299
4300                 ret = read_block_for_search(NULL, root, path, &next, level,
4301                                             0, &key);
4302                 if (ret == -EAGAIN)
4303                         goto again;
4304
4305                 if (ret < 0) {
4306                         btrfs_release_path(path);
4307                         goto done;
4308                 }
4309
4310                 if (!path->skip_locking) {
4311                         ret = btrfs_try_tree_read_lock(next);
4312                         if (!ret) {
4313                                 btrfs_set_path_blocking(path);
4314                                 btrfs_tree_read_lock(next);
4315                                 btrfs_clear_path_blocking(path, next,
4316                                                           BTRFS_READ_LOCK);
4317                         }
4318                         next_rw_lock = BTRFS_READ_LOCK;
4319                 }
4320         }
4321         ret = 0;
4322 done:
4323         unlock_up(path, 0, 1, 0, NULL);
4324         path->leave_spinning = old_spinning;
4325         if (!old_spinning)
4326                 btrfs_set_path_blocking(path);
4327
4328         return ret;
4329 }
4330
4331 /*
4332  * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
4333  * searching until it gets past min_objectid or finds an item of 'type'
4334  *
4335  * returns 0 if something is found, 1 if nothing was found and < 0 on error
4336  */
4337 int btrfs_previous_item(struct btrfs_root *root,
4338                         struct btrfs_path *path, u64 min_objectid,
4339                         int type)
4340 {
4341         struct btrfs_key found_key;
4342         struct extent_buffer *leaf;
4343         u32 nritems;
4344         int ret;
4345
4346         while (1) {
4347                 if (path->slots[0] == 0) {
4348                         btrfs_set_path_blocking(path);
4349                         ret = btrfs_prev_leaf(root, path);
4350                         if (ret != 0)
4351                                 return ret;
4352                 } else {
4353                         path->slots[0]--;
4354                 }
4355                 leaf = path->nodes[0];
4356                 nritems = btrfs_header_nritems(leaf);
4357                 if (nritems == 0)
4358                         return 1;
4359                 if (path->slots[0] == nritems)
4360                         path->slots[0]--;
4361
4362                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4363                 if (found_key.objectid < min_objectid)
4364                         break;
4365                 if (found_key.type == type)
4366                         return 0;
4367                 if (found_key.objectid == min_objectid &&
4368                     found_key.type < type)
4369                         break;
4370         }
4371         return 1;
4372 }