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