fs/btrfs/ctree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007,2008 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/slab.h>
   8 #include <linux/rbtree.h>
   9 #include <linux/mm.h>
  10 #include "ctree.h"
  11 #include "disk-io.h"
  12 #include "transaction.h"
  13 #include "print-tree.h"
  14 #include "locking.h"
  15 #include "volumes.h"
  16 #include "qgroup.h"
  17 #include "tree-mod-log.h"
  18
  19 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  20                       *root, struct btrfs_path *path, int level);
  21 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  22                       const struct btrfs_key *ins_key, struct btrfs_path *path,
  23                       int data_size, int extend);
  24 static int push_node_left(struct btrfs_trans_handle *trans,
  25                           struct extent_buffer *dst,
  26                           struct extent_buffer *src, int empty);
  27 static int balance_node_right(struct btrfs_trans_handle *trans,
  28                               struct extent_buffer *dst_buf,
  29                               struct extent_buffer *src_buf);
  30 static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
  31                     int level, int slot);
  32
  33 static const struct btrfs_csums {
  34         u16             size;
  35         const char      name[10];
  36         const char      driver[12];
  37 } btrfs_csums[] = {
  38         [BTRFS_CSUM_TYPE_CRC32] = { .size = 4, .name = "crc32c" },
  39         [BTRFS_CSUM_TYPE_XXHASH] = { .size = 8, .name = "xxhash64" },
  40         [BTRFS_CSUM_TYPE_SHA256] = { .size = 32, .name = "sha256" },
  41         [BTRFS_CSUM_TYPE_BLAKE2] = { .size = 32, .name = "blake2b",
  42                                      .driver = "blake2b-256" },
  43 };
  44
  45 int btrfs_super_csum_size(const struct btrfs_super_block *s)
  46 {
  47         u16 t = btrfs_super_csum_type(s);
  48         /*
  49          * csum type is validated at mount time
  50          */
  51         return btrfs_csums[t].size;
  52 }
  53
  54 const char *btrfs_super_csum_name(u16 csum_type)
  55 {
  56         /* csum type is validated at mount time */
  57         return btrfs_csums[csum_type].name;
  58 }
  59
  60 /*
  61  * Return driver name if defined, otherwise the name that's also a valid driver
  62  * name
  63  */
  64 const char *btrfs_super_csum_driver(u16 csum_type)
  65 {
  66         /* csum type is validated at mount time */
  67         return btrfs_csums[csum_type].driver[0] ?
  68                 btrfs_csums[csum_type].driver :
  69                 btrfs_csums[csum_type].name;
  70 }
  71
  72 size_t __attribute_const__ btrfs_get_num_csums(void)
  73 {
  74         return ARRAY_SIZE(btrfs_csums);
  75 }
  76
  77 struct btrfs_path *btrfs_alloc_path(void)
  78 {
  79         return kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
  80 }
  81
  82 /* this also releases the path */
  83 void btrfs_free_path(struct btrfs_path *p)
  84 {
  85         if (!p)
  86                 return;
  87         btrfs_release_path(p);
  88         kmem_cache_free(btrfs_path_cachep, p);
  89 }
  90
  91 /*
  92  * path release drops references on the extent buffers in the path
  93  * and it drops any locks held by this path
  94  *
  95  * It is safe to call this on paths that no locks or extent buffers held.
  96  */
  97 noinline void btrfs_release_path(struct btrfs_path *p)
  98 {
  99         int i;
 100
 101         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
 102                 p->slots[i] = 0;
 103                 if (!p->nodes[i])
 104                         continue;
 105                 if (p->locks[i]) {
 106                         btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
 107                         p->locks[i] = 0;
 108                 }
 109                 free_extent_buffer(p->nodes[i]);
 110                 p->nodes[i] = NULL;
 111         }
 112 }
 113
 114 /*
 115  * safely gets a reference on the root node of a tree.  A lock
 116  * is not taken, so a concurrent writer may put a different node
 117  * at the root of the tree.  See btrfs_lock_root_node for the
 118  * looping required.
 119  *
 120  * The extent buffer returned by this has a reference taken, so
 121  * it won't disappear.  It may stop being the root of the tree
 122  * at any time because there are no locks held.
 123  */
 124 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
 125 {
 126         struct extent_buffer *eb;
 127
 128         while (1) {
 129                 rcu_read_lock();
 130                 eb = rcu_dereference(root->node);
 131
 132                 /*
 133                  * RCU really hurts here, we could free up the root node because
 134                  * it was COWed but we may not get the new root node yet so do
 135                  * the inc_not_zero dance and if it doesn't work then
 136                  * synchronize_rcu and try again.
 137                  */
 138                 if (atomic_inc_not_zero(&eb->refs)) {
 139                         rcu_read_unlock();
 140                         break;
 141                 }
 142                 rcu_read_unlock();
 143                 synchronize_rcu();
 144         }
 145         return eb;
 146 }
 147
 148 /*
 149  * Cowonly root (not-shareable trees, everything not subvolume or reloc roots),
 150  * just get put onto a simple dirty list.  Transaction walks this list to make
 151  * sure they get properly updated on disk.
 152  */
 153 static void add_root_to_dirty_list(struct btrfs_root *root)
 154 {
 155         struct btrfs_fs_info *fs_info = root->fs_info;
 156
 157         if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
 158             !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
 159                 return;
 160
 161         spin_lock(&fs_info->trans_lock);
 162         if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
 163                 /* Want the extent tree to be the last on the list */
 164                 if (root->root_key.objectid == BTRFS_EXTENT_TREE_OBJECTID)
 165                         list_move_tail(&root->dirty_list,
 166                                        &fs_info->dirty_cowonly_roots);
 167                 else
 168                         list_move(&root->dirty_list,
 169                                   &fs_info->dirty_cowonly_roots);
 170         }
 171         spin_unlock(&fs_info->trans_lock);
 172 }
 173
 174 /*
 175  * used by snapshot creation to make a copy of a root for a tree with
 176  * a given objectid.  The buffer with the new root node is returned in
 177  * cow_ret, and this func returns zero on success or a negative error code.
 178  */
 179 int btrfs_copy_root(struct btrfs_trans_handle *trans,
 180                       struct btrfs_root *root,
 181                       struct extent_buffer *buf,
 182                       struct extent_buffer **cow_ret, u64 new_root_objectid)
 183 {
 184         struct btrfs_fs_info *fs_info = root->fs_info;
 185         struct extent_buffer *cow;
 186         int ret = 0;
 187         int level;
 188         struct btrfs_disk_key disk_key;
 189
 190         WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 191                 trans->transid != fs_info->running_transaction->transid);
 192         WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 193                 trans->transid != root->last_trans);
 194
 195         level = btrfs_header_level(buf);
 196         if (level == 0)
 197                 btrfs_item_key(buf, &disk_key, 0);
 198         else
 199                 btrfs_node_key(buf, &disk_key, 0);
 200
 201         cow = btrfs_alloc_tree_block(trans, root, 0, new_root_objectid,
 202                                      &disk_key, level, buf->start, 0,
 203                                      BTRFS_NESTING_NEW_ROOT);
 204         if (IS_ERR(cow))
 205                 return PTR_ERR(cow);
 206
 207         copy_extent_buffer_full(cow, buf);
 208         btrfs_set_header_bytenr(cow, cow->start);
 209         btrfs_set_header_generation(cow, trans->transid);
 210         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 211         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 212                                      BTRFS_HEADER_FLAG_RELOC);
 213         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 214                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 215         else
 216                 btrfs_set_header_owner(cow, new_root_objectid);
 217
 218         write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
 219
 220         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 221         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 222                 ret = btrfs_inc_ref(trans, root, cow, 1);
 223         else
 224                 ret = btrfs_inc_ref(trans, root, cow, 0);
 225         if (ret) {
 226                 btrfs_tree_unlock(cow);
 227                 free_extent_buffer(cow);
 228                 btrfs_abort_transaction(trans, ret);
 229                 return ret;
 230         }
 231
 232         btrfs_mark_buffer_dirty(cow);
 233         *cow_ret = cow;
 234         return 0;
 235 }
 236
 237 /*
 238  * check if the tree block can be shared by multiple trees
 239  */
 240 int btrfs_block_can_be_shared(struct btrfs_root *root,
 241                               struct extent_buffer *buf)
 242 {
 243         /*
 244          * Tree blocks not in shareable trees and tree roots are never shared.
 245          * If a block was allocated after the last snapshot and the block was
 246          * not allocated by tree relocation, we know the block is not shared.
 247          */
 248         if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 249             buf != root->node && buf != root->commit_root &&
 250             (btrfs_header_generation(buf) <=
 251              btrfs_root_last_snapshot(&root->root_item) ||
 252              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 253                 return 1;
 254
 255         return 0;
 256 }
 257
 258 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
 259                                        struct btrfs_root *root,
 260                                        struct extent_buffer *buf,
 261                                        struct extent_buffer *cow,
 262                                        int *last_ref)
 263 {
 264         struct btrfs_fs_info *fs_info = root->fs_info;
 265         u64 refs;
 266         u64 owner;
 267         u64 flags;
 268         u64 new_flags = 0;
 269         int ret;
 270
 271         /*
 272          * Backrefs update rules:
 273          *
 274          * Always use full backrefs for extent pointers in tree block
 275          * allocated by tree relocation.
 276          *
 277          * If a shared tree block is no longer referenced by its owner
 278          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
 279          * use full backrefs for extent pointers in tree block.
 280          *
 281          * If a tree block is been relocating
 282          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
 283          * use full backrefs for extent pointers in tree block.
 284          * The reason for this is some operations (such as drop tree)
 285          * are only allowed for blocks use full backrefs.
 286          */
 287
 288         if (btrfs_block_can_be_shared(root, buf)) {
 289                 ret = btrfs_lookup_extent_info(trans, fs_info, buf->start,
 290                                                btrfs_header_level(buf), 1,
 291                                                &refs, &flags);
 292                 if (ret)
 293                         return ret;
 294                 if (refs == 0) {
 295                         ret = -EROFS;
 296                         btrfs_handle_fs_error(fs_info, ret, NULL);
 297                         return ret;
 298                 }
 299         } else {
 300                 refs = 1;
 301                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 302                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 303                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 304                 else
 305                         flags = 0;
 306         }
 307
 308         owner = btrfs_header_owner(buf);
 309         BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
 310                !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
 311
 312         if (refs > 1) {
 313                 if ((owner == root->root_key.objectid ||
 314                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
 315                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
 316                         ret = btrfs_inc_ref(trans, root, buf, 1);
 317                         if (ret)
 318                                 return ret;
 319
 320                         if (root->root_key.objectid ==
 321                             BTRFS_TREE_RELOC_OBJECTID) {
 322                                 ret = btrfs_dec_ref(trans, root, buf, 0);
 323                                 if (ret)
 324                                         return ret;
 325                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 326                                 if (ret)
 327                                         return ret;
 328                         }
 329                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
 330                 } else {
 331
 332                         if (root->root_key.objectid ==
 333                             BTRFS_TREE_RELOC_OBJECTID)
 334                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 335                         else
 336                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 337                         if (ret)
 338                                 return ret;
 339                 }
 340                 if (new_flags != 0) {
 341                         int level = btrfs_header_level(buf);
 342
 343                         ret = btrfs_set_disk_extent_flags(trans, buf,
 344                                                           new_flags, level, 0);
 345                         if (ret)
 346                                 return ret;
 347                 }
 348         } else {
 349                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
 350                         if (root->root_key.objectid ==
 351                             BTRFS_TREE_RELOC_OBJECTID)
 352                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 353                         else
 354                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 355                         if (ret)
 356                                 return ret;
 357                         ret = btrfs_dec_ref(trans, root, buf, 1);
 358                         if (ret)
 359                                 return ret;
 360                 }
 361                 btrfs_clean_tree_block(buf);
 362                 *last_ref = 1;
 363         }
 364         return 0;
 365 }
 366
 367 static struct extent_buffer *alloc_tree_block_no_bg_flush(
 368                                           struct btrfs_trans_handle *trans,
 369                                           struct btrfs_root *root,
 370                                           u64 parent_start,
 371                                           const struct btrfs_disk_key *disk_key,
 372                                           int level,
 373                                           u64 hint,
 374                                           u64 empty_size,
 375                                           enum btrfs_lock_nesting nest)
 376 {
 377         struct btrfs_fs_info *fs_info = root->fs_info;
 378         struct extent_buffer *ret;
 379
 380         /*
 381          * If we are COWing a node/leaf from the extent, chunk, device or free
 382          * space trees, make sure that we do not finish block group creation of
 383          * pending block groups. We do this to avoid a deadlock.
 384          * COWing can result in allocation of a new chunk, and flushing pending
 385          * block groups (btrfs_create_pending_block_groups()) can be triggered
 386          * when finishing allocation of a new chunk. Creation of a pending block
 387          * group modifies the extent, chunk, device and free space trees,
 388          * therefore we could deadlock with ourselves since we are holding a
 389          * lock on an extent buffer that btrfs_create_pending_block_groups() may
 390          * try to COW later.
 391          * For similar reasons, we also need to delay flushing pending block
 392          * groups when splitting a leaf or node, from one of those trees, since
 393          * we are holding a write lock on it and its parent or when inserting a
 394          * new root node for one of those trees.
 395          */
 396         if (root == fs_info->extent_root ||
 397             root == fs_info->chunk_root ||
 398             root == fs_info->dev_root ||
 399             root == fs_info->free_space_root)
 400                 trans->can_flush_pending_bgs = false;
 401
 402         ret = btrfs_alloc_tree_block(trans, root, parent_start,
 403                                      root->root_key.objectid, disk_key, level,
 404                                      hint, empty_size, nest);
 405         trans->can_flush_pending_bgs = true;
 406
 407         return ret;
 408 }
 409
 410 /*
 411  * does the dirty work in cow of a single block.  The parent block (if
 412  * supplied) is updated to point to the new cow copy.  The new buffer is marked
 413  * dirty and returned locked.  If you modify the block it needs to be marked
 414  * dirty again.
 415  *
 416  * search_start -- an allocation hint for the new block
 417  *
 418  * empty_size -- a hint that you plan on doing more cow.  This is the size in
 419  * bytes the allocator should try to find free next to the block it returns.
 420  * This is just a hint and may be ignored by the allocator.
 421  */
 422 static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 423                              struct btrfs_root *root,
 424                              struct extent_buffer *buf,
 425                              struct extent_buffer *parent, int parent_slot,
 426                              struct extent_buffer **cow_ret,
 427                              u64 search_start, u64 empty_size,
 428                              enum btrfs_lock_nesting nest)
 429 {
 430         struct btrfs_fs_info *fs_info = root->fs_info;
 431         struct btrfs_disk_key disk_key;
 432         struct extent_buffer *cow;
 433         int level, ret;
 434         int last_ref = 0;
 435         int unlock_orig = 0;
 436         u64 parent_start = 0;
 437
 438         if (*cow_ret == buf)
 439                 unlock_orig = 1;
 440
 441         btrfs_assert_tree_locked(buf);
 442
 443         WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 444                 trans->transid != fs_info->running_transaction->transid);
 445         WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 446                 trans->transid != root->last_trans);
 447
 448         level = btrfs_header_level(buf);
 449
 450         if (level == 0)
 451                 btrfs_item_key(buf, &disk_key, 0);
 452         else
 453                 btrfs_node_key(buf, &disk_key, 0);
 454
 455         if ((root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) && parent)
 456                 parent_start = parent->start;
 457
 458         cow = alloc_tree_block_no_bg_flush(trans, root, parent_start, &disk_key,
 459                                            level, search_start, empty_size, nest);
 460         if (IS_ERR(cow))
 461                 return PTR_ERR(cow);
 462
 463         /* cow is set to blocking by btrfs_init_new_buffer */
 464
 465         copy_extent_buffer_full(cow, buf);
 466         btrfs_set_header_bytenr(cow, cow->start);
 467         btrfs_set_header_generation(cow, trans->transid);
 468         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 469         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 470                                      BTRFS_HEADER_FLAG_RELOC);
 471         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 472                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 473         else
 474                 btrfs_set_header_owner(cow, root->root_key.objectid);
 475
 476         write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
 477
 478         ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
 479         if (ret) {
 480                 btrfs_tree_unlock(cow);
 481                 free_extent_buffer(cow);
 482                 btrfs_abort_transaction(trans, ret);
 483                 return ret;
 484         }
 485
 486         if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
 487                 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
 488                 if (ret) {
 489                         btrfs_tree_unlock(cow);
 490                         free_extent_buffer(cow);
 491                         btrfs_abort_transaction(trans, ret);
 492                         return ret;
 493                 }
 494         }
 495
 496         if (buf == root->node) {
 497                 WARN_ON(parent && parent != buf);
 498                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 499                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 500                         parent_start = buf->start;
 501
 502                 atomic_inc(&cow->refs);
 503                 ret = btrfs_tree_mod_log_insert_root(root->node, cow, 1);
 504                 BUG_ON(ret < 0);
 505                 rcu_assign_pointer(root->node, cow);
 506
 507                 btrfs_free_tree_block(trans, root, buf, parent_start,
 508                                       last_ref);
 509                 free_extent_buffer(buf);
 510                 add_root_to_dirty_list(root);
 511         } else {
 512                 WARN_ON(trans->transid != btrfs_header_generation(parent));
 513                 btrfs_tree_mod_log_insert_key(parent, parent_slot,
 514                                               BTRFS_MOD_LOG_KEY_REPLACE, GFP_NOFS);
 515                 btrfs_set_node_blockptr(parent, parent_slot,
 516                                         cow->start);
 517                 btrfs_set_node_ptr_generation(parent, parent_slot,
 518                                               trans->transid);
 519                 btrfs_mark_buffer_dirty(parent);
 520                 if (last_ref) {
 521                         ret = btrfs_tree_mod_log_free_eb(buf);
 522                         if (ret) {
 523                                 btrfs_tree_unlock(cow);
 524                                 free_extent_buffer(cow);
 525                                 btrfs_abort_transaction(trans, ret);
 526                                 return ret;
 527                         }
 528                 }
 529                 btrfs_free_tree_block(trans, root, buf, parent_start,
 530                                       last_ref);
 531         }
 532         if (unlock_orig)
 533                 btrfs_tree_unlock(buf);
 534         free_extent_buffer_stale(buf);
 535         btrfs_mark_buffer_dirty(cow);
 536         *cow_ret = cow;
 537         return 0;
 538 }
 539
 540 static inline int should_cow_block(struct btrfs_trans_handle *trans,
 541                                    struct btrfs_root *root,
 542                                    struct extent_buffer *buf)
 543 {
 544         if (btrfs_is_testing(root->fs_info))
 545                 return 0;
 546
 547         /* Ensure we can see the FORCE_COW bit */
 548         smp_mb__before_atomic();
 549
 550         /*
 551          * We do not need to cow a block if
 552          * 1) this block is not created or changed in this transaction;
 553          * 2) this block does not belong to TREE_RELOC tree;
 554          * 3) the root is not forced COW.
 555          *
 556          * What is forced COW:
 557          *    when we create snapshot during committing the transaction,
 558          *    after we've finished copying src root, we must COW the shared
 559          *    block to ensure the metadata consistency.
 560          */
 561         if (btrfs_header_generation(buf) == trans->transid &&
 562             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
 563             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
 564               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
 565             !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
 566                 return 0;
 567         return 1;
 568 }
 569
 570 /*
 571  * cows a single block, see __btrfs_cow_block for the real work.
 572  * This version of it has extra checks so that a block isn't COWed more than
 573  * once per transaction, as long as it hasn't been written yet
 574  */
 575 noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
 576                     struct btrfs_root *root, struct extent_buffer *buf,
 577                     struct extent_buffer *parent, int parent_slot,
 578                     struct extent_buffer **cow_ret,
 579                     enum btrfs_lock_nesting nest)
 580 {
 581         struct btrfs_fs_info *fs_info = root->fs_info;
 582         u64 search_start;
 583         int ret;
 584
 585         if (test_bit(BTRFS_ROOT_DELETING, &root->state))
 586                 btrfs_err(fs_info,
 587                         "COW'ing blocks on a fs root that's being dropped");
 588
 589         if (trans->transaction != fs_info->running_transaction)
 590                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
 591                        trans->transid,
 592                        fs_info->running_transaction->transid);
 593
 594         if (trans->transid != fs_info->generation)
 595                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
 596                        trans->transid, fs_info->generation);
 597
 598         if (!should_cow_block(trans, root, buf)) {
 599                 trans->dirty = true;
 600                 *cow_ret = buf;
 601                 return 0;
 602         }
 603
 604         search_start = buf->start & ~((u64)SZ_1G - 1);
 605
 606         /*
 607          * Before CoWing this block for later modification, check if it's
 608          * the subtree root and do the delayed subtree trace if needed.
 609          *
 610          * Also We don't care about the error, as it's handled internally.
 611          */
 612         btrfs_qgroup_trace_subtree_after_cow(trans, root, buf);
 613         ret = __btrfs_cow_block(trans, root, buf, parent,
 614                                  parent_slot, cow_ret, search_start, 0, nest);
 615
 616         trace_btrfs_cow_block(root, buf, *cow_ret);
 617
 618         return ret;
 619 }
 620 ALLOW_ERROR_INJECTION(btrfs_cow_block, ERRNO);
 621
 622 /*
 623  * helper function for defrag to decide if two blocks pointed to by a
 624  * node are actually close by
 625  */
 626 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 627 {
 628         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 629                 return 1;
 630         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 631                 return 1;
 632         return 0;
 633 }
 634
 635 #ifdef __LITTLE_ENDIAN
 636
 637 /*
 638  * Compare two keys, on little-endian the disk order is same as CPU order and
 639  * we can avoid the conversion.
 640  */
 641 static int comp_keys(const struct btrfs_disk_key *disk_key,
 642                      const struct btrfs_key *k2)
 643 {
 644         const struct btrfs_key *k1 = (const struct btrfs_key *)disk_key;
 645
 646         return btrfs_comp_cpu_keys(k1, k2);
 647 }
 648
 649 #else
 650
 651 /*
 652  * compare two keys in a memcmp fashion
 653  */
 654 static int comp_keys(const struct btrfs_disk_key *disk,
 655                      const struct btrfs_key *k2)
 656 {
 657         struct btrfs_key k1;
 658
 659         btrfs_disk_key_to_cpu(&k1, disk);
 660
 661         return btrfs_comp_cpu_keys(&k1, k2);
 662 }
 663 #endif
 664
 665 /*
 666  * same as comp_keys only with two btrfs_key's
 667  */
 668 int __pure btrfs_comp_cpu_keys(const struct btrfs_key *k1, const struct btrfs_key *k2)
 669 {
 670         if (k1->objectid > k2->objectid)
 671                 return 1;
 672         if (k1->objectid < k2->objectid)
 673                 return -1;
 674         if (k1->type > k2->type)
 675                 return 1;
 676         if (k1->type < k2->type)
 677                 return -1;
 678         if (k1->offset > k2->offset)
 679                 return 1;
 680         if (k1->offset < k2->offset)
 681                 return -1;
 682         return 0;
 683 }
 684
 685 /*
 686  * this is used by the defrag code to go through all the
 687  * leaves pointed to by a node and reallocate them so that
 688  * disk order is close to key order
 689  */
 690 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 691                        struct btrfs_root *root, struct extent_buffer *parent,
 692                        int start_slot, u64 *last_ret,
 693                        struct btrfs_key *progress)
 694 {
 695         struct btrfs_fs_info *fs_info = root->fs_info;
 696         struct extent_buffer *cur;
 697         u64 blocknr;
 698         u64 search_start = *last_ret;
 699         u64 last_block = 0;
 700         u64 other;
 701         u32 parent_nritems;
 702         int end_slot;
 703         int i;
 704         int err = 0;
 705         u32 blocksize;
 706         int progress_passed = 0;
 707         struct btrfs_disk_key disk_key;
 708
 709         WARN_ON(trans->transaction != fs_info->running_transaction);
 710         WARN_ON(trans->transid != fs_info->generation);
 711
 712         parent_nritems = btrfs_header_nritems(parent);
 713         blocksize = fs_info->nodesize;
 714         end_slot = parent_nritems - 1;
 715
 716         if (parent_nritems <= 1)
 717                 return 0;
 718
 719         for (i = start_slot; i <= end_slot; i++) {
 720                 int close = 1;
 721
 722                 btrfs_node_key(parent, &disk_key, i);
 723                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
 724                         continue;
 725
 726                 progress_passed = 1;
 727                 blocknr = btrfs_node_blockptr(parent, i);
 728                 if (last_block == 0)
 729                         last_block = blocknr;
 730
 731                 if (i > 0) {
 732                         other = btrfs_node_blockptr(parent, i - 1);
 733                         close = close_blocks(blocknr, other, blocksize);
 734                 }
 735                 if (!close && i < end_slot) {
 736                         other = btrfs_node_blockptr(parent, i + 1);
 737                         close = close_blocks(blocknr, other, blocksize);
 738                 }
 739                 if (close) {
 740                         last_block = blocknr;
 741                         continue;
 742                 }
 743
 744                 cur = btrfs_read_node_slot(parent, i);
 745                 if (IS_ERR(cur))
 746                         return PTR_ERR(cur);
 747                 if (search_start == 0)
 748                         search_start = last_block;
 749
 750                 btrfs_tree_lock(cur);
 751                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 752                                         &cur, search_start,
 753                                         min(16 * blocksize,
 754                                             (end_slot - i) * blocksize),
 755                                         BTRFS_NESTING_COW);
 756                 if (err) {
 757                         btrfs_tree_unlock(cur);
 758                         free_extent_buffer(cur);
 759                         break;
 760                 }
 761                 search_start = cur->start;
 762                 last_block = cur->start;
 763                 *last_ret = search_start;
 764                 btrfs_tree_unlock(cur);
 765                 free_extent_buffer(cur);
 766         }
 767         return err;
 768 }
 769
 770 /*
 771  * search for key in the extent_buffer.  The items start at offset p,
 772  * and they are item_size apart.  There are 'max' items in p.
 773  *
 774  * the slot in the array is returned via slot, and it points to
 775  * the place where you would insert key if it is not found in
 776  * the array.
 777  *
 778  * slot may point to max if the key is bigger than all of the keys
 779  */
 780 static noinline int generic_bin_search(struct extent_buffer *eb,
 781                                        unsigned long p, int item_size,
 782                                        const struct btrfs_key *key,
 783                                        int max, int *slot)
 784 {
 785         int low = 0;
 786         int high = max;
 787         int ret;
 788         const int key_size = sizeof(struct btrfs_disk_key);
 789
 790         if (low > high) {
 791                 btrfs_err(eb->fs_info,
 792                  "%s: low (%d) > high (%d) eb %llu owner %llu level %d",
 793                           __func__, low, high, eb->start,
 794                           btrfs_header_owner(eb), btrfs_header_level(eb));
 795                 return -EINVAL;
 796         }
 797
 798         while (low < high) {
 799                 unsigned long oip;
 800                 unsigned long offset;
 801                 struct btrfs_disk_key *tmp;
 802                 struct btrfs_disk_key unaligned;
 803                 int mid;
 804
 805                 mid = (low + high) / 2;
 806                 offset = p + mid * item_size;
 807                 oip = offset_in_page(offset);
 808
 809                 if (oip + key_size <= PAGE_SIZE) {
 810                         const unsigned long idx = get_eb_page_index(offset);
 811                         char *kaddr = page_address(eb->pages[idx]);
 812
 813                         oip = get_eb_offset_in_page(eb, offset);
 814                         tmp = (struct btrfs_disk_key *)(kaddr + oip);
 815                 } else {
 816                         read_extent_buffer(eb, &unaligned, offset, key_size);
 817                         tmp = &unaligned;
 818                 }
 819
 820                 ret = comp_keys(tmp, key);
 821
 822                 if (ret < 0)
 823                         low = mid + 1;
 824                 else if (ret > 0)
 825                         high = mid;
 826                 else {
 827                         *slot = mid;
 828                         return 0;
 829                 }
 830         }
 831         *slot = low;
 832         return 1;
 833 }
 834
 835 /*
 836  * simple bin_search frontend that does the right thing for
 837  * leaves vs nodes
 838  */
 839 int btrfs_bin_search(struct extent_buffer *eb, const struct btrfs_key *key,
 840                      int *slot)
 841 {
 842         if (btrfs_header_level(eb) == 0)
 843                 return generic_bin_search(eb,
 844                                           offsetof(struct btrfs_leaf, items),
 845                                           sizeof(struct btrfs_item),
 846                                           key, btrfs_header_nritems(eb),
 847                                           slot);
 848         else
 849                 return generic_bin_search(eb,
 850                                           offsetof(struct btrfs_node, ptrs),
 851                                           sizeof(struct btrfs_key_ptr),
 852                                           key, btrfs_header_nritems(eb),
 853                                           slot);
 854 }
 855
 856 static void root_add_used(struct btrfs_root *root, u32 size)
 857 {
 858         spin_lock(&root->accounting_lock);
 859         btrfs_set_root_used(&root->root_item,
 860                             btrfs_root_used(&root->root_item) + size);
 861         spin_unlock(&root->accounting_lock);
 862 }
 863
 864 static void root_sub_used(struct btrfs_root *root, u32 size)
 865 {
 866         spin_lock(&root->accounting_lock);
 867         btrfs_set_root_used(&root->root_item,
 868                             btrfs_root_used(&root->root_item) - size);
 869         spin_unlock(&root->accounting_lock);
 870 }
 871
 872 /* given a node and slot number, this reads the blocks it points to.  The
 873  * extent buffer is returned with a reference taken (but unlocked).
 874  */
 875 struct extent_buffer *btrfs_read_node_slot(struct extent_buffer *parent,
 876                                            int slot)
 877 {
 878         int level = btrfs_header_level(parent);
 879         struct extent_buffer *eb;
 880         struct btrfs_key first_key;
 881
 882         if (slot < 0 || slot >= btrfs_header_nritems(parent))
 883                 return ERR_PTR(-ENOENT);
 884
 885         BUG_ON(level == 0);
 886
 887         btrfs_node_key_to_cpu(parent, &first_key, slot);
 888         eb = read_tree_block(parent->fs_info, btrfs_node_blockptr(parent, slot),
 889                              btrfs_header_owner(parent),
 890                              btrfs_node_ptr_generation(parent, slot),
 891                              level - 1, &first_key);
 892         if (!IS_ERR(eb) && !extent_buffer_uptodate(eb)) {
 893                 free_extent_buffer(eb);
 894                 eb = ERR_PTR(-EIO);
 895         }
 896
 897         return eb;
 898 }
 899
 900 /*
 901  * node level balancing, used to make sure nodes are in proper order for
 902  * item deletion.  We balance from the top down, so we have to make sure
 903  * that a deletion won't leave an node completely empty later on.
 904  */
 905 static noinline int balance_level(struct btrfs_trans_handle *trans,
 906                          struct btrfs_root *root,
 907                          struct btrfs_path *path, int level)
 908 {
 909         struct btrfs_fs_info *fs_info = root->fs_info;
 910         struct extent_buffer *right = NULL;
 911         struct extent_buffer *mid;
 912         struct extent_buffer *left = NULL;
 913         struct extent_buffer *parent = NULL;
 914         int ret = 0;
 915         int wret;
 916         int pslot;
 917         int orig_slot = path->slots[level];
 918         u64 orig_ptr;
 919
 920         ASSERT(level > 0);
 921
 922         mid = path->nodes[level];
 923
 924         WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK);
 925         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 926
 927         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 928
 929         if (level < BTRFS_MAX_LEVEL - 1) {
 930                 parent = path->nodes[level + 1];
 931                 pslot = path->slots[level + 1];
 932         }
 933
 934         /*
 935          * deal with the case where there is only one pointer in the root
 936          * by promoting the node below to a root
 937          */
 938         if (!parent) {
 939                 struct extent_buffer *child;
 940
 941                 if (btrfs_header_nritems(mid) != 1)
 942                         return 0;
 943
 944                 /* promote the child to a root */
 945                 child = btrfs_read_node_slot(mid, 0);
 946                 if (IS_ERR(child)) {
 947                         ret = PTR_ERR(child);
 948                         btrfs_handle_fs_error(fs_info, ret, NULL);
 949                         goto enospc;
 950                 }
 951
 952                 btrfs_tree_lock(child);
 953                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child,
 954                                       BTRFS_NESTING_COW);
 955                 if (ret) {
 956                         btrfs_tree_unlock(child);
 957                         free_extent_buffer(child);
 958                         goto enospc;
 959                 }
 960
 961                 ret = btrfs_tree_mod_log_insert_root(root->node, child, 1);
 962                 BUG_ON(ret < 0);
 963                 rcu_assign_pointer(root->node, child);
 964
 965                 add_root_to_dirty_list(root);
 966                 btrfs_tree_unlock(child);
 967
 968                 path->locks[level] = 0;
 969                 path->nodes[level] = NULL;
 970                 btrfs_clean_tree_block(mid);
 971                 btrfs_tree_unlock(mid);
 972                 /* once for the path */
 973                 free_extent_buffer(mid);
 974
 975                 root_sub_used(root, mid->len);
 976                 btrfs_free_tree_block(trans, root, mid, 0, 1);
 977                 /* once for the root ptr */
 978                 free_extent_buffer_stale(mid);
 979                 return 0;
 980         }
 981         if (btrfs_header_nritems(mid) >
 982             BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 4)
 983                 return 0;
 984
 985         left = btrfs_read_node_slot(parent, pslot - 1);
 986         if (IS_ERR(left))
 987                 left = NULL;
 988
 989         if (left) {
 990                 __btrfs_tree_lock(left, BTRFS_NESTING_LEFT);
 991                 wret = btrfs_cow_block(trans, root, left,
 992                                        parent, pslot - 1, &left,
 993                                        BTRFS_NESTING_LEFT_COW);
 994                 if (wret) {
 995                         ret = wret;
 996                         goto enospc;
 997                 }
 998         }
 999
1000         right = btrfs_read_node_slot(parent, pslot + 1);
1001         if (IS_ERR(right))
1002                 right = NULL;
1003
1004         if (right) {
1005                 __btrfs_tree_lock(right, BTRFS_NESTING_RIGHT);
1006                 wret = btrfs_cow_block(trans, root, right,
1007                                        parent, pslot + 1, &right,
1008                                        BTRFS_NESTING_RIGHT_COW);
1009                 if (wret) {
1010                         ret = wret;
1011                         goto enospc;
1012                 }
1013         }
1014
1015         /* first, try to make some room in the middle buffer */
1016         if (left) {
1017                 orig_slot += btrfs_header_nritems(left);
1018                 wret = push_node_left(trans, left, mid, 1);
1019                 if (wret < 0)
1020                         ret = wret;
1021         }
1022
1023         /*
1024          * then try to empty the right most buffer into the middle
1025          */
1026         if (right) {
1027                 wret = push_node_left(trans, mid, right, 1);
1028                 if (wret < 0 && wret != -ENOSPC)
1029                         ret = wret;
1030                 if (btrfs_header_nritems(right) == 0) {
1031                         btrfs_clean_tree_block(right);
1032                         btrfs_tree_unlock(right);
1033                         del_ptr(root, path, level + 1, pslot + 1);
1034                         root_sub_used(root, right->len);
1035                         btrfs_free_tree_block(trans, root, right, 0, 1);
1036                         free_extent_buffer_stale(right);
1037                         right = NULL;
1038                 } else {
1039                         struct btrfs_disk_key right_key;
1040                         btrfs_node_key(right, &right_key, 0);
1041                         ret = btrfs_tree_mod_log_insert_key(parent, pslot + 1,
1042                                         BTRFS_MOD_LOG_KEY_REPLACE, GFP_NOFS);
1043                         BUG_ON(ret < 0);
1044                         btrfs_set_node_key(parent, &right_key, pslot + 1);
1045                         btrfs_mark_buffer_dirty(parent);
1046                 }
1047         }
1048         if (btrfs_header_nritems(mid) == 1) {
1049                 /*
1050                  * we're not allowed to leave a node with one item in the
1051                  * tree during a delete.  A deletion from lower in the tree
1052                  * could try to delete the only pointer in this node.
1053                  * So, pull some keys from the left.
1054                  * There has to be a left pointer at this point because
1055                  * otherwise we would have pulled some pointers from the
1056                  * right
1057                  */
1058                 if (!left) {
1059                         ret = -EROFS;
1060                         btrfs_handle_fs_error(fs_info, ret, NULL);
1061                         goto enospc;
1062                 }
1063                 wret = balance_node_right(trans, mid, left);
1064                 if (wret < 0) {
1065                         ret = wret;
1066                         goto enospc;
1067                 }
1068                 if (wret == 1) {
1069                         wret = push_node_left(trans, left, mid, 1);
1070                         if (wret < 0)
1071                                 ret = wret;
1072                 }
1073                 BUG_ON(wret == 1);
1074         }
1075         if (btrfs_header_nritems(mid) == 0) {
1076                 btrfs_clean_tree_block(mid);
1077                 btrfs_tree_unlock(mid);
1078                 del_ptr(root, path, level + 1, pslot);
1079                 root_sub_used(root, mid->len);
1080                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1081                 free_extent_buffer_stale(mid);
1082                 mid = NULL;
1083         } else {
1084                 /* update the parent key to reflect our changes */
1085                 struct btrfs_disk_key mid_key;
1086                 btrfs_node_key(mid, &mid_key, 0);
1087                 ret = btrfs_tree_mod_log_insert_key(parent, pslot,
1088                                 BTRFS_MOD_LOG_KEY_REPLACE, GFP_NOFS);
1089                 BUG_ON(ret < 0);
1090                 btrfs_set_node_key(parent, &mid_key, pslot);
1091                 btrfs_mark_buffer_dirty(parent);
1092         }
1093
1094         /* update the path */
1095         if (left) {
1096                 if (btrfs_header_nritems(left) > orig_slot) {
1097                         atomic_inc(&left->refs);
1098                         /* left was locked after cow */
1099                         path->nodes[level] = left;
1100                         path->slots[level + 1] -= 1;
1101                         path->slots[level] = orig_slot;
1102                         if (mid) {
1103                                 btrfs_tree_unlock(mid);
1104                                 free_extent_buffer(mid);
1105                         }
1106                 } else {
1107                         orig_slot -= btrfs_header_nritems(left);
1108                         path->slots[level] = orig_slot;
1109                 }
1110         }
1111         /* double check we haven't messed things up */
1112         if (orig_ptr !=
1113             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
1114                 BUG();
1115 enospc:
1116         if (right) {
1117                 btrfs_tree_unlock(right);
1118                 free_extent_buffer(right);
1119         }
1120         if (left) {
1121                 if (path->nodes[level] != left)
1122                         btrfs_tree_unlock(left);
1123                 free_extent_buffer(left);
1124         }
1125         return ret;
1126 }
1127
1128 /* Node balancing for insertion.  Here we only split or push nodes around
1129  * when they are completely full.  This is also done top down, so we
1130  * have to be pessimistic.
1131  */
1132 static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
1133                                           struct btrfs_root *root,
1134                                           struct btrfs_path *path, int level)
1135 {
1136         struct btrfs_fs_info *fs_info = root->fs_info;
1137         struct extent_buffer *right = NULL;
1138         struct extent_buffer *mid;
1139         struct extent_buffer *left = NULL;
1140         struct extent_buffer *parent = NULL;
1141         int ret = 0;
1142         int wret;
1143         int pslot;
1144         int orig_slot = path->slots[level];
1145
1146         if (level == 0)
1147                 return 1;
1148
1149         mid = path->nodes[level];
1150         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1151
1152         if (level < BTRFS_MAX_LEVEL - 1) {
1153                 parent = path->nodes[level + 1];
1154                 pslot = path->slots[level + 1];
1155         }
1156
1157         if (!parent)
1158                 return 1;
1159
1160         left = btrfs_read_node_slot(parent, pslot - 1);
1161         if (IS_ERR(left))
1162                 left = NULL;
1163
1164         /* first, try to make some room in the middle buffer */
1165         if (left) {
1166                 u32 left_nr;
1167
1168                 __btrfs_tree_lock(left, BTRFS_NESTING_LEFT);
1169
1170                 left_nr = btrfs_header_nritems(left);
1171                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
1172                         wret = 1;
1173                 } else {
1174                         ret = btrfs_cow_block(trans, root, left, parent,
1175                                               pslot - 1, &left,
1176                                               BTRFS_NESTING_LEFT_COW);
1177                         if (ret)
1178                                 wret = 1;
1179                         else {
1180                                 wret = push_node_left(trans, left, mid, 0);
1181                         }
1182                 }
1183                 if (wret < 0)
1184                         ret = wret;
1185                 if (wret == 0) {
1186                         struct btrfs_disk_key disk_key;
1187                         orig_slot += left_nr;
1188                         btrfs_node_key(mid, &disk_key, 0);
1189                         ret = btrfs_tree_mod_log_insert_key(parent, pslot,
1190                                         BTRFS_MOD_LOG_KEY_REPLACE, GFP_NOFS);
1191                         BUG_ON(ret < 0);
1192                         btrfs_set_node_key(parent, &disk_key, pslot);
1193                         btrfs_mark_buffer_dirty(parent);
1194                         if (btrfs_header_nritems(left) > orig_slot) {
1195                                 path->nodes[level] = left;
1196                                 path->slots[level + 1] -= 1;
1197                                 path->slots[level] = orig_slot;
1198                                 btrfs_tree_unlock(mid);
1199                                 free_extent_buffer(mid);
1200                         } else {
1201                                 orig_slot -=
1202                                         btrfs_header_nritems(left);
1203                                 path->slots[level] = orig_slot;
1204                                 btrfs_tree_unlock(left);
1205                                 free_extent_buffer(left);
1206                         }
1207                         return 0;
1208                 }
1209                 btrfs_tree_unlock(left);
1210                 free_extent_buffer(left);
1211         }
1212         right = btrfs_read_node_slot(parent, pslot + 1);
1213         if (IS_ERR(right))
1214                 right = NULL;
1215
1216         /*
1217          * then try to empty the right most buffer into the middle
1218          */
1219         if (right) {
1220                 u32 right_nr;
1221
1222                 __btrfs_tree_lock(right, BTRFS_NESTING_RIGHT);
1223
1224                 right_nr = btrfs_header_nritems(right);
1225                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
1226                         wret = 1;
1227                 } else {
1228                         ret = btrfs_cow_block(trans, root, right,
1229                                               parent, pslot + 1,
1230                                               &right, BTRFS_NESTING_RIGHT_COW);
1231                         if (ret)
1232                                 wret = 1;
1233                         else {
1234                                 wret = balance_node_right(trans, right, mid);
1235                         }
1236                 }
1237                 if (wret < 0)
1238                         ret = wret;
1239                 if (wret == 0) {
1240                         struct btrfs_disk_key disk_key;
1241
1242                         btrfs_node_key(right, &disk_key, 0);
1243                         ret = btrfs_tree_mod_log_insert_key(parent, pslot + 1,
1244                                         BTRFS_MOD_LOG_KEY_REPLACE, GFP_NOFS);
1245                         BUG_ON(ret < 0);
1246                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
1247                         btrfs_mark_buffer_dirty(parent);
1248
1249                         if (btrfs_header_nritems(mid) <= orig_slot) {
1250                                 path->nodes[level] = right;
1251                                 path->slots[level + 1] += 1;
1252                                 path->slots[level] = orig_slot -
1253                                         btrfs_header_nritems(mid);
1254                                 btrfs_tree_unlock(mid);
1255                                 free_extent_buffer(mid);
1256                         } else {
1257                                 btrfs_tree_unlock(right);
1258                                 free_extent_buffer(right);
1259                         }
1260                         return 0;
1261                 }
1262                 btrfs_tree_unlock(right);
1263                 free_extent_buffer(right);
1264         }
1265         return 1;
1266 }
1267
1268 /*
1269  * readahead one full node of leaves, finding things that are close
1270  * to the block in 'slot', and triggering ra on them.
1271  */
1272 static void reada_for_search(struct btrfs_fs_info *fs_info,
1273                              struct btrfs_path *path,
1274                              int level, int slot, u64 objectid)
1275 {
1276         struct extent_buffer *node;
1277         struct btrfs_disk_key disk_key;
1278         u32 nritems;
1279         u64 search;
1280         u64 target;
1281         u64 nread = 0;
1282         struct extent_buffer *eb;
1283         u32 nr;
1284         u32 blocksize;
1285         u32 nscan = 0;
1286
1287         if (level != 1)
1288                 return;
1289
1290         if (!path->nodes[level])
1291                 return;
1292
1293         node = path->nodes[level];
1294
1295         search = btrfs_node_blockptr(node, slot);
1296         blocksize = fs_info->nodesize;
1297         eb = find_extent_buffer(fs_info, search);
1298         if (eb) {
1299                 free_extent_buffer(eb);
1300                 return;
1301         }
1302
1303         target = search;
1304
1305         nritems = btrfs_header_nritems(node);
1306         nr = slot;
1307
1308         while (1) {
1309                 if (path->reada == READA_BACK) {
1310                         if (nr == 0)
1311                                 break;
1312                         nr--;
1313                 } else if (path->reada == READA_FORWARD) {
1314                         nr++;
1315                         if (nr >= nritems)
1316                                 break;
1317                 }
1318                 if (path->reada == READA_BACK && objectid) {
1319                         btrfs_node_key(node, &disk_key, nr);
1320                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1321                                 break;
1322                 }
1323                 search = btrfs_node_blockptr(node, nr);
1324                 if ((search <= target && target - search <= 65536) ||
1325                     (search > target && search - target <= 65536)) {
1326                         btrfs_readahead_node_child(node, nr);
1327                         nread += blocksize;
1328                 }
1329                 nscan++;
1330                 if ((nread > 65536 || nscan > 32))
1331                         break;
1332         }
1333 }
1334
1335 static noinline void reada_for_balance(struct btrfs_path *path, int level)
1336 {
1337         struct extent_buffer *parent;
1338         int slot;
1339         int nritems;
1340
1341         parent = path->nodes[level + 1];
1342         if (!parent)
1343                 return;
1344
1345         nritems = btrfs_header_nritems(parent);
1346         slot = path->slots[level + 1];
1347
1348         if (slot > 0)
1349                 btrfs_readahead_node_child(parent, slot - 1);
1350         if (slot + 1 < nritems)
1351                 btrfs_readahead_node_child(parent, slot + 1);
1352 }
1353
1354
1355 /*
1356  * when we walk down the tree, it is usually safe to unlock the higher layers
1357  * in the tree.  The exceptions are when our path goes through slot 0, because
1358  * operations on the tree might require changing key pointers higher up in the
1359  * tree.
1360  *
1361  * callers might also have set path->keep_locks, which tells this code to keep
1362  * the lock if the path points to the last slot in the block.  This is part of
1363  * walking through the tree, and selecting the next slot in the higher block.
1364  *
1365  * lowest_unlock sets the lowest level in the tree we're allowed to unlock.  so
1366  * if lowest_unlock is 1, level 0 won't be unlocked
1367  */
1368 static noinline void unlock_up(struct btrfs_path *path, int level,
1369                                int lowest_unlock, int min_write_lock_level,
1370                                int *write_lock_level)
1371 {
1372         int i;
1373         int skip_level = level;
1374         int no_skips = 0;
1375         struct extent_buffer *t;
1376
1377         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1378                 if (!path->nodes[i])
1379                         break;
1380                 if (!path->locks[i])
1381                         break;
1382                 if (!no_skips && path->slots[i] == 0) {
1383                         skip_level = i + 1;
1384                         continue;
1385                 }
1386                 if (!no_skips && path->keep_locks) {
1387                         u32 nritems;
1388                         t = path->nodes[i];
1389                         nritems = btrfs_header_nritems(t);
1390                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
1391                                 skip_level = i + 1;
1392                                 continue;
1393                         }
1394                 }
1395                 if (skip_level < i && i >= lowest_unlock)
1396                         no_skips = 1;
1397
1398                 t = path->nodes[i];
1399                 if (i >= lowest_unlock && i > skip_level) {
1400                         btrfs_tree_unlock_rw(t, path->locks[i]);
1401                         path->locks[i] = 0;
1402                         if (write_lock_level &&
1403                             i > min_write_lock_level &&
1404                             i <= *write_lock_level) {
1405                                 *write_lock_level = i - 1;
1406                         }
1407                 }
1408         }
1409 }
1410
1411 /*
1412  * helper function for btrfs_search_slot.  The goal is to find a block
1413  * in cache without setting the path to blocking.  If we find the block
1414  * we return zero and the path is unchanged.
1415  *
1416  * If we can't find the block, we set the path blocking and do some
1417  * reada.  -EAGAIN is returned and the search must be repeated.
1418  */
1419 static int
1420 read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
1421                       struct extent_buffer **eb_ret, int level, int slot,
1422                       const struct btrfs_key *key)
1423 {
1424         struct btrfs_fs_info *fs_info = root->fs_info;
1425         u64 blocknr;
1426         u64 gen;
1427         struct extent_buffer *tmp;
1428         struct btrfs_key first_key;
1429         int ret;
1430         int parent_level;
1431
1432         blocknr = btrfs_node_blockptr(*eb_ret, slot);
1433         gen = btrfs_node_ptr_generation(*eb_ret, slot);
1434         parent_level = btrfs_header_level(*eb_ret);
1435         btrfs_node_key_to_cpu(*eb_ret, &first_key, slot);
1436
1437         tmp = find_extent_buffer(fs_info, blocknr);
1438         if (tmp) {
1439                 /* first we do an atomic uptodate check */
1440                 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
1441                         /*
1442                          * Do extra check for first_key, eb can be stale due to
1443                          * being cached, read from scrub, or have multiple
1444                          * parents (shared tree blocks).
1445                          */
1446                         if (btrfs_verify_level_key(tmp,
1447                                         parent_level - 1, &first_key, gen)) {
1448                                 free_extent_buffer(tmp);
1449                                 return -EUCLEAN;
1450                         }
1451                         *eb_ret = tmp;
1452                         return 0;
1453                 }
1454
1455                 /* now we're allowed to do a blocking uptodate check */
1456                 ret = btrfs_read_buffer(tmp, gen, parent_level - 1, &first_key);
1457                 if (!ret) {
1458                         *eb_ret = tmp;
1459                         return 0;
1460                 }
1461                 free_extent_buffer(tmp);
1462                 btrfs_release_path(p);
1463                 return -EIO;
1464         }
1465
1466         /*
1467          * reduce lock contention at high levels
1468          * of the btree by dropping locks before
1469          * we read.  Don't release the lock on the current
1470          * level because we need to walk this node to figure
1471          * out which blocks to read.
1472          */
1473         btrfs_unlock_up_safe(p, level + 1);
1474
1475         if (p->reada != READA_NONE)
1476                 reada_for_search(fs_info, p, level, slot, key->objectid);
1477
1478         ret = -EAGAIN;
1479         tmp = read_tree_block(fs_info, blocknr, root->root_key.objectid,
1480                               gen, parent_level - 1, &first_key);
1481         if (!IS_ERR(tmp)) {
1482                 /*
1483                  * If the read above didn't mark this buffer up to date,
1484                  * it will never end up being up to date.  Set ret to EIO now
1485                  * and give up so that our caller doesn't loop forever
1486                  * on our EAGAINs.
1487                  */
1488                 if (!extent_buffer_uptodate(tmp))
1489                         ret = -EIO;
1490                 free_extent_buffer(tmp);
1491         } else {
1492                 ret = PTR_ERR(tmp);
1493         }
1494
1495         btrfs_release_path(p);
1496         return ret;
1497 }
1498
1499 /*
1500  * helper function for btrfs_search_slot.  This does all of the checks
1501  * for node-level blocks and does any balancing required based on
1502  * the ins_len.
1503  *
1504  * If no extra work was required, zero is returned.  If we had to
1505  * drop the path, -EAGAIN is returned and btrfs_search_slot must
1506  * start over
1507  */
1508 static int
1509 setup_nodes_for_search(struct btrfs_trans_handle *trans,
1510                        struct btrfs_root *root, struct btrfs_path *p,
1511                        struct extent_buffer *b, int level, int ins_len,
1512                        int *write_lock_level)
1513 {
1514         struct btrfs_fs_info *fs_info = root->fs_info;
1515         int ret = 0;
1516
1517         if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
1518             BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3) {
1519
1520                 if (*write_lock_level < level + 1) {
1521                         *write_lock_level = level + 1;
1522                         btrfs_release_path(p);
1523                         return -EAGAIN;
1524                 }
1525
1526                 reada_for_balance(p, level);
1527                 ret = split_node(trans, root, p, level);
1528
1529                 b = p->nodes[level];
1530         } else if (ins_len < 0 && btrfs_header_nritems(b) <
1531                    BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 2) {
1532
1533                 if (*write_lock_level < level + 1) {
1534                         *write_lock_level = level + 1;
1535                         btrfs_release_path(p);
1536                         return -EAGAIN;
1537                 }
1538
1539                 reada_for_balance(p, level);
1540                 ret = balance_level(trans, root, p, level);
1541                 if (ret)
1542                         return ret;
1543
1544                 b = p->nodes[level];
1545                 if (!b) {
1546                         btrfs_release_path(p);
1547                         return -EAGAIN;
1548                 }
1549                 BUG_ON(btrfs_header_nritems(b) == 1);
1550         }
1551         return ret;
1552 }
1553
1554 int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
1555                 u64 iobjectid, u64 ioff, u8 key_type,
1556                 struct btrfs_key *found_key)
1557 {
1558         int ret;
1559         struct btrfs_key key;
1560         struct extent_buffer *eb;
1561
1562         ASSERT(path);
1563         ASSERT(found_key);
1564
1565         key.type = key_type;
1566         key.objectid = iobjectid;
1567         key.offset = ioff;
1568
1569         ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
1570         if (ret < 0)
1571                 return ret;
1572
1573         eb = path->nodes[0];
1574         if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
1575                 ret = btrfs_next_leaf(fs_root, path);
1576                 if (ret)
1577                         return ret;
1578                 eb = path->nodes[0];
1579         }
1580
1581         btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
1582         if (found_key->type != key.type ||
1583                         found_key->objectid != key.objectid)
1584                 return 1;
1585
1586         return 0;
1587 }
1588
1589 static struct extent_buffer *btrfs_search_slot_get_root(struct btrfs_root *root,
1590                                                         struct btrfs_path *p,
1591                                                         int write_lock_level)
1592 {
1593         struct btrfs_fs_info *fs_info = root->fs_info;
1594         struct extent_buffer *b;
1595         int root_lock;
1596         int level = 0;
1597
1598         /* We try very hard to do read locks on the root */
1599         root_lock = BTRFS_READ_LOCK;
1600
1601         if (p->search_commit_root) {
1602                 /*
1603                  * The commit roots are read only so we always do read locks,
1604                  * and we always must hold the commit_root_sem when doing
1605                  * searches on them, the only exception is send where we don't
1606                  * want to block transaction commits for a long time, so
1607                  * we need to clone the commit root in order to avoid races
1608                  * with transaction commits that create a snapshot of one of
1609                  * the roots used by a send operation.
1610                  */
1611                 if (p->need_commit_sem) {
1612                         down_read(&fs_info->commit_root_sem);
1613                         b = btrfs_clone_extent_buffer(root->commit_root);
1614                         up_read(&fs_info->commit_root_sem);
1615                         if (!b)
1616                                 return ERR_PTR(-ENOMEM);
1617
1618                 } else {
1619                         b = root->commit_root;
1620                         atomic_inc(&b->refs);
1621                 }
1622                 level = btrfs_header_level(b);
1623                 /*
1624                  * Ensure that all callers have set skip_locking when
1625                  * p->search_commit_root = 1.
1626                  */
1627                 ASSERT(p->skip_locking == 1);
1628
1629                 goto out;
1630         }
1631
1632         if (p->skip_locking) {
1633                 b = btrfs_root_node(root);
1634                 level = btrfs_header_level(b);
1635                 goto out;
1636         }
1637
1638         /*
1639          * If the level is set to maximum, we can skip trying to get the read
1640          * lock.
1641          */
1642         if (write_lock_level < BTRFS_MAX_LEVEL) {
1643                 /*
1644                  * We don't know the level of the root node until we actually
1645                  * have it read locked
1646                  */
1647                 b = btrfs_read_lock_root_node(root);
1648                 level = btrfs_header_level(b);
1649                 if (level > write_lock_level)
1650                         goto out;
1651
1652                 /* Whoops, must trade for write lock */
1653                 btrfs_tree_read_unlock(b);
1654                 free_extent_buffer(b);
1655         }
1656
1657         b = btrfs_lock_root_node(root);
1658         root_lock = BTRFS_WRITE_LOCK;
1659
1660         /* The level might have changed, check again */
1661         level = btrfs_header_level(b);
1662
1663 out:
1664         p->nodes[level] = b;
1665         if (!p->skip_locking)
1666                 p->locks[level] = root_lock;
1667         /*
1668          * Callers are responsible for dropping b's references.
1669          */
1670         return b;
1671 }
1672
1673
1674 /*
1675  * btrfs_search_slot - look for a key in a tree and perform necessary
1676  * modifications to preserve tree invariants.
1677  *
1678  * @trans:      Handle of transaction, used when modifying the tree
1679  * @p:          Holds all btree nodes along the search path
1680  * @root:       The root node of the tree
1681  * @key:        The key we are looking for
1682  * @ins_len:    Indicates purpose of search:
1683  *              >0  for inserts it's size of item inserted (*)
1684  *              <0  for deletions
1685  *               0  for plain searches, not modifying the tree
1686  *
1687  *              (*) If size of item inserted doesn't include
1688  *              sizeof(struct btrfs_item), then p->search_for_extension must
1689  *              be set.
1690  * @cow:        boolean should CoW operations be performed. Must always be 1
1691  *              when modifying the tree.
1692  *
1693  * If @ins_len > 0, nodes and leaves will be split as we walk down the tree.
1694  * If @ins_len < 0, nodes will be merged as we walk down the tree (if possible)
1695  *
1696  * If @key is found, 0 is returned and you can find the item in the leaf level
1697  * of the path (level 0)
1698  *
1699  * If @key isn't found, 1 is returned and the leaf level of the path (level 0)
1700  * points to the slot where it should be inserted
1701  *
1702  * If an error is encountered while searching the tree a negative error number
1703  * is returned
1704  */
1705 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1706                       const struct btrfs_key *key, struct btrfs_path *p,
1707                       int ins_len, int cow)
1708 {
1709         struct extent_buffer *b;
1710         int slot;
1711         int ret;
1712         int err;
1713         int level;
1714         int lowest_unlock = 1;
1715         /* everything at write_lock_level or lower must be write locked */
1716         int write_lock_level = 0;
1717         u8 lowest_level = 0;
1718         int min_write_lock_level;
1719         int prev_cmp;
1720
1721         lowest_level = p->lowest_level;
1722         WARN_ON(lowest_level && ins_len > 0);
1723         WARN_ON(p->nodes[0] != NULL);
1724         BUG_ON(!cow && ins_len);
1725
1726         if (ins_len < 0) {
1727                 lowest_unlock = 2;
1728
1729                 /* when we are removing items, we might have to go up to level
1730                  * two as we update tree pointers  Make sure we keep write
1731                  * for those levels as well
1732                  */
1733                 write_lock_level = 2;
1734         } else if (ins_len > 0) {
1735                 /*
1736                  * for inserting items, make sure we have a write lock on
1737                  * level 1 so we can update keys
1738                  */
1739                 write_lock_level = 1;
1740         }
1741
1742         if (!cow)
1743                 write_lock_level = -1;
1744
1745         if (cow && (p->keep_locks || p->lowest_level))
1746                 write_lock_level = BTRFS_MAX_LEVEL;
1747
1748         min_write_lock_level = write_lock_level;
1749
1750 again:
1751         prev_cmp = -1;
1752         b = btrfs_search_slot_get_root(root, p, write_lock_level);
1753         if (IS_ERR(b)) {
1754                 ret = PTR_ERR(b);
1755                 goto done;
1756         }
1757
1758         while (b) {
1759                 int dec = 0;
1760
1761                 level = btrfs_header_level(b);
1762
1763                 if (cow) {
1764                         bool last_level = (level == (BTRFS_MAX_LEVEL - 1));
1765
1766                         /*
1767                          * if we don't really need to cow this block
1768                          * then we don't want to set the path blocking,
1769                          * so we test it here
1770                          */
1771                         if (!should_cow_block(trans, root, b)) {
1772                                 trans->dirty = true;
1773                                 goto cow_done;
1774                         }
1775
1776                         /*
1777                          * must have write locks on this node and the
1778                          * parent
1779                          */
1780                         if (level > write_lock_level ||
1781                             (level + 1 > write_lock_level &&
1782                             level + 1 < BTRFS_MAX_LEVEL &&
1783                             p->nodes[level + 1])) {
1784                                 write_lock_level = level + 1;
1785                                 btrfs_release_path(p);
1786                                 goto again;
1787                         }
1788
1789                         if (last_level)
1790                                 err = btrfs_cow_block(trans, root, b, NULL, 0,
1791                                                       &b,
1792                                                       BTRFS_NESTING_COW);
1793                         else
1794                                 err = btrfs_cow_block(trans, root, b,
1795                                                       p->nodes[level + 1],
1796                                                       p->slots[level + 1], &b,
1797                                                       BTRFS_NESTING_COW);
1798                         if (err) {
1799                                 ret = err;
1800                                 goto done;
1801                         }
1802                 }
1803 cow_done:
1804                 p->nodes[level] = b;
1805                 /*
1806                  * Leave path with blocking locks to avoid massive
1807                  * lock context switch, this is made on purpose.
1808                  */
1809
1810                 /*
1811                  * we have a lock on b and as long as we aren't changing
1812                  * the tree, there is no way to for the items in b to change.
1813                  * It is safe to drop the lock on our parent before we
1814                  * go through the expensive btree search on b.
1815                  *
1816                  * If we're inserting or deleting (ins_len != 0), then we might
1817                  * be changing slot zero, which may require changing the parent.
1818                  * So, we can't drop the lock until after we know which slot
1819                  * we're operating on.
1820                  */
1821                 if (!ins_len && !p->keep_locks) {
1822                         int u = level + 1;
1823
1824                         if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
1825                                 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
1826                                 p->locks[u] = 0;
1827                         }
1828                 }
1829
1830                 /*
1831                  * If btrfs_bin_search returns an exact match (prev_cmp == 0)
1832                  * we can safely assume the target key will always be in slot 0
1833                  * on lower levels due to the invariants BTRFS' btree provides,
1834                  * namely that a btrfs_key_ptr entry always points to the
1835                  * lowest key in the child node, thus we can skip searching
1836                  * lower levels
1837                  */
1838                 if (prev_cmp == 0) {
1839                         slot = 0;
1840                         ret = 0;
1841                 } else {
1842                         ret = btrfs_bin_search(b, key, &slot);
1843                         prev_cmp = ret;
1844                         if (ret < 0)
1845                                 goto done;
1846                 }
1847
1848                 if (level == 0) {
1849                         p->slots[level] = slot;
1850                         /*
1851                          * Item key already exists. In this case, if we are
1852                          * allowed to insert the item (for example, in dir_item
1853                          * case, item key collision is allowed), it will be
1854                          * merged with the original item. Only the item size
1855                          * grows, no new btrfs item will be added. If
1856                          * search_for_extension is not set, ins_len already
1857                          * accounts the size btrfs_item, deduct it here so leaf
1858                          * space check will be correct.
1859                          */
1860                         if (ret == 0 && ins_len > 0 && !p->search_for_extension) {
1861                                 ASSERT(ins_len >= sizeof(struct btrfs_item));
1862                                 ins_len -= sizeof(struct btrfs_item);
1863                         }
1864                         if (ins_len > 0 &&
1865                             btrfs_leaf_free_space(b) < ins_len) {
1866                                 if (write_lock_level < 1) {
1867                                         write_lock_level = 1;
1868                                         btrfs_release_path(p);
1869                                         goto again;
1870                                 }
1871
1872                                 err = split_leaf(trans, root, key,
1873                                                  p, ins_len, ret == 0);
1874
1875                                 BUG_ON(err > 0);
1876                                 if (err) {
1877                                         ret = err;
1878                                         goto done;
1879                                 }
1880                         }
1881                         if (!p->search_for_split)
1882                                 unlock_up(p, level, lowest_unlock,
1883                                           min_write_lock_level, NULL);
1884                         goto done;
1885                 }
1886                 if (ret && slot > 0) {
1887                         dec = 1;
1888                         slot--;
1889                 }
1890                 p->slots[level] = slot;
1891                 err = setup_nodes_for_search(trans, root, p, b, level, ins_len,
1892                                              &write_lock_level);
1893                 if (err == -EAGAIN)
1894                         goto again;
1895                 if (err) {
1896                         ret = err;
1897                         goto done;
1898                 }
1899                 b = p->nodes[level];
1900                 slot = p->slots[level];
1901
1902                 /*
1903                  * Slot 0 is special, if we change the key we have to update
1904                  * the parent pointer which means we must have a write lock on
1905                  * the parent
1906                  */
1907                 if (slot == 0 && ins_len && write_lock_level < level + 1) {
1908                         write_lock_level = level + 1;
1909                         btrfs_release_path(p);
1910                         goto again;
1911                 }
1912
1913                 unlock_up(p, level, lowest_unlock, min_write_lock_level,
1914                           &write_lock_level);
1915
1916                 if (level == lowest_level) {
1917                         if (dec)
1918                                 p->slots[level]++;
1919                         goto done;
1920                 }
1921
1922                 err = read_block_for_search(root, p, &b, level, slot, key);
1923                 if (err == -EAGAIN)
1924                         goto again;
1925                 if (err) {
1926                         ret = err;
1927                         goto done;
1928                 }
1929
1930                 if (!p->skip_locking) {
1931                         level = btrfs_header_level(b);
1932                         if (level <= write_lock_level) {
1933                                 btrfs_tree_lock(b);
1934                                 p->locks[level] = BTRFS_WRITE_LOCK;
1935                         } else {
1936                                 btrfs_tree_read_lock(b);
1937                                 p->locks[level] = BTRFS_READ_LOCK;
1938                         }
1939                         p->nodes[level] = b;
1940                 }
1941         }
1942         ret = 1;
1943 done:
1944         if (ret < 0 && !p->skip_release_on_error)
1945                 btrfs_release_path(p);
1946         return ret;
1947 }
1948 ALLOW_ERROR_INJECTION(btrfs_search_slot, ERRNO);
1949
1950 /*
1951  * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
1952  * current state of the tree together with the operations recorded in the tree
1953  * modification log to search for the key in a previous version of this tree, as
1954  * denoted by the time_seq parameter.
1955  *
1956  * Naturally, there is no support for insert, delete or cow operations.
1957  *
1958  * The resulting path and return value will be set up as if we called
1959  * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
1960  */
1961 int btrfs_search_old_slot(struct btrfs_root *root, const struct btrfs_key *key,
1962                           struct btrfs_path *p, u64 time_seq)
1963 {
1964         struct btrfs_fs_info *fs_info = root->fs_info;
1965         struct extent_buffer *b;
1966         int slot;
1967         int ret;
1968         int err;
1969         int level;
1970         int lowest_unlock = 1;
1971         u8 lowest_level = 0;
1972
1973         lowest_level = p->lowest_level;
1974         WARN_ON(p->nodes[0] != NULL);
1975
1976         if (p->search_commit_root) {
1977                 BUG_ON(time_seq);
1978                 return btrfs_search_slot(NULL, root, key, p, 0, 0);
1979         }
1980
1981 again:
1982         b = btrfs_get_old_root(root, time_seq);
1983         if (!b) {
1984                 ret = -EIO;
1985                 goto done;
1986         }
1987         level = btrfs_header_level(b);
1988         p->locks[level] = BTRFS_READ_LOCK;
1989
1990         while (b) {
1991                 int dec = 0;
1992
1993                 level = btrfs_header_level(b);
1994                 p->nodes[level] = b;
1995
1996                 /*
1997                  * we have a lock on b and as long as we aren't changing
1998                  * the tree, there is no way to for the items in b to change.
1999                  * It is safe to drop the lock on our parent before we
2000                  * go through the expensive btree search on b.
2001                  */
2002                 btrfs_unlock_up_safe(p, level + 1);
2003
2004                 ret = btrfs_bin_search(b, key, &slot);
2005                 if (ret < 0)
2006                         goto done;
2007
2008                 if (level == 0) {
2009                         p->slots[level] = slot;
2010                         unlock_up(p, level, lowest_unlock, 0, NULL);
2011                         goto done;
2012                 }
2013
2014                 if (ret && slot > 0) {
2015                         dec = 1;
2016                         slot--;
2017                 }
2018                 p->slots[level] = slot;
2019                 unlock_up(p, level, lowest_unlock, 0, NULL);
2020
2021                 if (level == lowest_level) {
2022                         if (dec)
2023                                 p->slots[level]++;
2024                         goto done;
2025                 }
2026
2027                 err = read_block_for_search(root, p, &b, level, slot, key);
2028                 if (err == -EAGAIN)
2029                         goto again;
2030                 if (err) {
2031                         ret = err;
2032                         goto done;
2033                 }
2034
2035                 level = btrfs_header_level(b);
2036                 btrfs_tree_read_lock(b);
2037                 b = btrfs_tree_mod_log_rewind(fs_info, p, b, time_seq);
2038                 if (!b) {
2039                         ret = -ENOMEM;
2040                         goto done;
2041                 }
2042                 p->locks[level] = BTRFS_READ_LOCK;
2043                 p->nodes[level] = b;
2044         }
2045         ret = 1;
2046 done:
2047         if (ret < 0)
2048                 btrfs_release_path(p);
2049
2050         return ret;
2051 }
2052
2053 /*
2054  * helper to use instead of search slot if no exact match is needed but
2055  * instead the next or previous item should be returned.
2056  * When find_higher is true, the next higher item is returned, the next lower
2057  * otherwise.
2058  * When return_any and find_higher are both true, and no higher item is found,
2059  * return the next lower instead.
2060  * When return_any is true and find_higher is false, and no lower item is found,
2061  * return the next higher instead.
2062  * It returns 0 if any item is found, 1 if none is found (tree empty), and
2063  * < 0 on error
2064  */
2065 int btrfs_search_slot_for_read(struct btrfs_root *root,
2066                                const struct btrfs_key *key,
2067                                struct btrfs_path *p, int find_higher,
2068                                int return_any)
2069 {
2070         int ret;
2071         struct extent_buffer *leaf;
2072
2073 again:
2074         ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
2075         if (ret <= 0)
2076                 return ret;
2077         /*
2078          * a return value of 1 means the path is at the position where the
2079          * item should be inserted. Normally this is the next bigger item,
2080          * but in case the previous item is the last in a leaf, path points
2081          * to the first free slot in the previous leaf, i.e. at an invalid
2082          * item.
2083          */
2084         leaf = p->nodes[0];
2085
2086         if (find_higher) {
2087                 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
2088                         ret = btrfs_next_leaf(root, p);
2089                         if (ret <= 0)
2090                                 return ret;
2091                         if (!return_any)
2092                                 return 1;
2093                         /*
2094                          * no higher item found, return the next
2095                          * lower instead
2096                          */
2097                         return_any = 0;
2098                         find_higher = 0;
2099                         btrfs_release_path(p);
2100                         goto again;
2101                 }
2102         } else {
2103                 if (p->slots[0] == 0) {
2104                         ret = btrfs_prev_leaf(root, p);
2105                         if (ret < 0)
2106                                 return ret;
2107                         if (!ret) {
2108                                 leaf = p->nodes[0];
2109                                 if (p->slots[0] == btrfs_header_nritems(leaf))
2110                                         p->slots[0]--;
2111                                 return 0;
2112                         }
2113                         if (!return_any)
2114                                 return 1;
2115                         /*
2116                          * no lower item found, return the next
2117                          * higher instead
2118                          */
2119                         return_any = 0;
2120                         find_higher = 1;
2121                         btrfs_release_path(p);
2122                         goto again;
2123                 } else {
2124                         --p->slots[0];
2125                 }
2126         }
2127         return 0;
2128 }
2129
2130 /*
2131  * adjust the pointers going up the tree, starting at level
2132  * making sure the right key of each node is points to 'key'.
2133  * This is used after shifting pointers to the left, so it stops
2134  * fixing up pointers when a given leaf/node is not in slot 0 of the
2135  * higher levels
2136  *
2137  */
2138 static void fixup_low_keys(struct btrfs_path *path,
2139                            struct btrfs_disk_key *key, int level)
2140 {
2141         int i;
2142         struct extent_buffer *t;
2143         int ret;
2144
2145         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2146                 int tslot = path->slots[i];
2147
2148                 if (!path->nodes[i])
2149                         break;
2150                 t = path->nodes[i];
2151                 ret = btrfs_tree_mod_log_insert_key(t, tslot,
2152                                 BTRFS_MOD_LOG_KEY_REPLACE, GFP_ATOMIC);
2153                 BUG_ON(ret < 0);
2154                 btrfs_set_node_key(t, key, tslot);
2155                 btrfs_mark_buffer_dirty(path->nodes[i]);
2156                 if (tslot != 0)
2157                         break;
2158         }
2159 }
2160
2161 /*
2162  * update item key.
2163  *
2164  * This function isn't completely safe. It's the caller's responsibility
2165  * that the new key won't break the order
2166  */
2167 void btrfs_set_item_key_safe(struct btrfs_fs_info *fs_info,
2168                              struct btrfs_path *path,
2169                              const struct btrfs_key *new_key)
2170 {
2171         struct btrfs_disk_key disk_key;
2172         struct extent_buffer *eb;
2173         int slot;
2174
2175         eb = path->nodes[0];
2176         slot = path->slots[0];
2177         if (slot > 0) {
2178                 btrfs_item_key(eb, &disk_key, slot - 1);
2179                 if (unlikely(comp_keys(&disk_key, new_key) >= 0)) {
2180                         btrfs_crit(fs_info,
2181                 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
2182                                    slot, btrfs_disk_key_objectid(&disk_key),
2183                                    btrfs_disk_key_type(&disk_key),
2184                                    btrfs_disk_key_offset(&disk_key),
2185                                    new_key->objectid, new_key->type,
2186                                    new_key->offset);
2187                         btrfs_print_leaf(eb);
2188                         BUG();
2189                 }
2190         }
2191         if (slot < btrfs_header_nritems(eb) - 1) {
2192                 btrfs_item_key(eb, &disk_key, slot + 1);
2193                 if (unlikely(comp_keys(&disk_key, new_key) <= 0)) {
2194                         btrfs_crit(fs_info,
2195                 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
2196                                    slot, btrfs_disk_key_objectid(&disk_key),
2197                                    btrfs_disk_key_type(&disk_key),
2198                                    btrfs_disk_key_offset(&disk_key),
2199                                    new_key->objectid, new_key->type,
2200                                    new_key->offset);
2201                         btrfs_print_leaf(eb);
2202                         BUG();
2203                 }
2204         }
2205
2206         btrfs_cpu_key_to_disk(&disk_key, new_key);
2207         btrfs_set_item_key(eb, &disk_key, slot);
2208         btrfs_mark_buffer_dirty(eb);
2209         if (slot == 0)
2210                 fixup_low_keys(path, &disk_key, 1);
2211 }
2212
2213 /*
2214  * Check key order of two sibling extent buffers.
2215  *
2216  * Return true if something is wrong.
2217  * Return false if everything is fine.
2218  *
2219  * Tree-checker only works inside one tree block, thus the following
2220  * corruption can not be detected by tree-checker:
2221  *
2222  * Leaf @left                   | Leaf @right
2223  * --------------------------------------------------------------
2224  * | 1 | 2 | 3 | 4 | 5 | f6 |   | 7 | 8 |
2225  *
2226  * Key f6 in leaf @left itself is valid, but not valid when the next
2227  * key in leaf @right is 7.
2228  * This can only be checked at tree block merge time.
2229  * And since tree checker has ensured all key order in each tree block
2230  * is correct, we only need to bother the last key of @left and the first
2231  * key of @right.
2232  */
2233 static bool check_sibling_keys(struct extent_buffer *left,
2234                                struct extent_buffer *right)
2235 {
2236         struct btrfs_key left_last;
2237         struct btrfs_key right_first;
2238         int level = btrfs_header_level(left);
2239         int nr_left = btrfs_header_nritems(left);
2240         int nr_right = btrfs_header_nritems(right);
2241
2242         /* No key to check in one of the tree blocks */
2243         if (!nr_left || !nr_right)
2244                 return false;
2245
2246         if (level) {
2247                 btrfs_node_key_to_cpu(left, &left_last, nr_left - 1);
2248                 btrfs_node_key_to_cpu(right, &right_first, 0);
2249         } else {
2250                 btrfs_item_key_to_cpu(left, &left_last, nr_left - 1);
2251                 btrfs_item_key_to_cpu(right, &right_first, 0);
2252         }
2253
2254         if (btrfs_comp_cpu_keys(&left_last, &right_first) >= 0) {
2255                 btrfs_crit(left->fs_info,
2256 "bad key order, sibling blocks, left last (%llu %u %llu) right first (%llu %u %llu)",
2257                            left_last.objectid, left_last.type,
2258                            left_last.offset, right_first.objectid,
2259                            right_first.type, right_first.offset);
2260                 return true;
2261         }
2262         return false;
2263 }
2264
2265 /*
2266  * try to push data from one node into the next node left in the
2267  * tree.
2268  *
2269  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
2270  * error, and > 0 if there was no room in the left hand block.
2271  */
2272 static int push_node_left(struct btrfs_trans_handle *trans,
2273                           struct extent_buffer *dst,
2274                           struct extent_buffer *src, int empty)
2275 {
2276         struct btrfs_fs_info *fs_info = trans->fs_info;
2277         int push_items = 0;
2278         int src_nritems;
2279         int dst_nritems;
2280         int ret = 0;
2281
2282         src_nritems = btrfs_header_nritems(src);
2283         dst_nritems = btrfs_header_nritems(dst);
2284         push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
2285         WARN_ON(btrfs_header_generation(src) != trans->transid);
2286         WARN_ON(btrfs_header_generation(dst) != trans->transid);
2287
2288         if (!empty && src_nritems <= 8)
2289                 return 1;
2290
2291         if (push_items <= 0)
2292                 return 1;
2293
2294         if (empty) {
2295                 push_items = min(src_nritems, push_items);
2296                 if (push_items < src_nritems) {
2297                         /* leave at least 8 pointers in the node if
2298                          * we aren't going to empty it
2299                          */
2300                         if (src_nritems - push_items < 8) {
2301                                 if (push_items <= 8)
2302                                         return 1;
2303                                 push_items -= 8;
2304                         }
2305                 }
2306         } else
2307                 push_items = min(src_nritems - 8, push_items);
2308
2309         /* dst is the left eb, src is the middle eb */
2310         if (check_sibling_keys(dst, src)) {
2311                 ret = -EUCLEAN;
2312                 btrfs_abort_transaction(trans, ret);
2313                 return ret;
2314         }
2315         ret = btrfs_tree_mod_log_eb_copy(dst, src, dst_nritems, 0, push_items);
2316         if (ret) {
2317                 btrfs_abort_transaction(trans, ret);
2318                 return ret;
2319         }
2320         copy_extent_buffer(dst, src,
2321                            btrfs_node_key_ptr_offset(dst_nritems),
2322                            btrfs_node_key_ptr_offset(0),
2323                            push_items * sizeof(struct btrfs_key_ptr));
2324
2325         if (push_items < src_nritems) {
2326                 /*
2327                  * Don't call btrfs_tree_mod_log_insert_move() here, key removal
2328                  * was already fully logged by btrfs_tree_mod_log_eb_copy() above.
2329                  */
2330                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
2331                                       btrfs_node_key_ptr_offset(push_items),
2332                                       (src_nritems - push_items) *
2333                                       sizeof(struct btrfs_key_ptr));
2334         }
2335         btrfs_set_header_nritems(src, src_nritems - push_items);
2336         btrfs_set_header_nritems(dst, dst_nritems + push_items);
2337         btrfs_mark_buffer_dirty(src);
2338         btrfs_mark_buffer_dirty(dst);
2339
2340         return ret;
2341 }
2342
2343 /*
2344  * try to push data from one node into the next node right in the
2345  * tree.
2346  *
2347  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
2348  * error, and > 0 if there was no room in the right hand block.
2349  *
2350  * this will  only push up to 1/2 the contents of the left node over
2351  */
2352 static int balance_node_right(struct btrfs_trans_handle *trans,
2353                               struct extent_buffer *dst,
2354                               struct extent_buffer *src)
2355 {
2356         struct btrfs_fs_info *fs_info = trans->fs_info;
2357         int push_items = 0;
2358         int max_push;
2359         int src_nritems;
2360         int dst_nritems;
2361         int ret = 0;
2362
2363         WARN_ON(btrfs_header_generation(src) != trans->transid);
2364         WARN_ON(btrfs_header_generation(dst) != trans->transid);
2365
2366         src_nritems = btrfs_header_nritems(src);
2367         dst_nritems = btrfs_header_nritems(dst);
2368         push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
2369         if (push_items <= 0)
2370                 return 1;
2371
2372         if (src_nritems < 4)
2373                 return 1;
2374
2375         max_push = src_nritems / 2 + 1;
2376         /* don't try to empty the node */
2377         if (max_push >= src_nritems)
2378                 return 1;
2379
2380         if (max_push < push_items)
2381                 push_items = max_push;
2382
2383         /* dst is the right eb, src is the middle eb */
2384         if (check_sibling_keys(src, dst)) {
2385                 ret = -EUCLEAN;
2386                 btrfs_abort_transaction(trans, ret);
2387                 return ret;
2388         }
2389         ret = btrfs_tree_mod_log_insert_move(dst, push_items, 0, dst_nritems);
2390         BUG_ON(ret < 0);
2391         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
2392                                       btrfs_node_key_ptr_offset(0),
2393                                       (dst_nritems) *
2394                                       sizeof(struct btrfs_key_ptr));
2395
2396         ret = btrfs_tree_mod_log_eb_copy(dst, src, 0, src_nritems - push_items,
2397                                          push_items);
2398         if (ret) {
2399                 btrfs_abort_transaction(trans, ret);
2400                 return ret;
2401         }
2402         copy_extent_buffer(dst, src,
2403                            btrfs_node_key_ptr_offset(0),
2404                            btrfs_node_key_ptr_offset(src_nritems - push_items),
2405                            push_items * sizeof(struct btrfs_key_ptr));
2406
2407         btrfs_set_header_nritems(src, src_nritems - push_items);
2408         btrfs_set_header_nritems(dst, dst_nritems + push_items);
2409
2410         btrfs_mark_buffer_dirty(src);
2411         btrfs_mark_buffer_dirty(dst);
2412
2413         return ret;
2414 }
2415
2416 /*
2417  * helper function to insert a new root level in the tree.
2418  * A new node is allocated, and a single item is inserted to
2419  * point to the existing root
2420  *
2421  * returns zero on success or < 0 on failure.
2422  */
2423 static noinline int insert_new_root(struct btrfs_trans_handle *trans,
2424                            struct btrfs_root *root,
2425                            struct btrfs_path *path, int level)
2426 {
2427         struct btrfs_fs_info *fs_info = root->fs_info;
2428         u64 lower_gen;
2429         struct extent_buffer *lower;
2430         struct extent_buffer *c;
2431         struct extent_buffer *old;
2432         struct btrfs_disk_key lower_key;
2433         int ret;
2434
2435         BUG_ON(path->nodes[level]);
2436         BUG_ON(path->nodes[level-1] != root->node);
2437
2438         lower = path->nodes[level-1];
2439         if (level == 1)
2440                 btrfs_item_key(lower, &lower_key, 0);
2441         else
2442                 btrfs_node_key(lower, &lower_key, 0);
2443
2444         c = alloc_tree_block_no_bg_flush(trans, root, 0, &lower_key, level,
2445                                          root->node->start, 0,
2446                                          BTRFS_NESTING_NEW_ROOT);
2447         if (IS_ERR(c))
2448                 return PTR_ERR(c);
2449
2450         root_add_used(root, fs_info->nodesize);
2451
2452         btrfs_set_header_nritems(c, 1);
2453         btrfs_set_node_key(c, &lower_key, 0);
2454         btrfs_set_node_blockptr(c, 0, lower->start);
2455         lower_gen = btrfs_header_generation(lower);
2456         WARN_ON(lower_gen != trans->transid);
2457
2458         btrfs_set_node_ptr_generation(c, 0, lower_gen);
2459
2460         btrfs_mark_buffer_dirty(c);
2461
2462         old = root->node;
2463         ret = btrfs_tree_mod_log_insert_root(root->node, c, 0);
2464         BUG_ON(ret < 0);
2465         rcu_assign_pointer(root->node, c);
2466
2467         /* the super has an extra ref to root->node */
2468         free_extent_buffer(old);
2469
2470         add_root_to_dirty_list(root);
2471         atomic_inc(&c->refs);
2472         path->nodes[level] = c;
2473         path->locks[level] = BTRFS_WRITE_LOCK;
2474         path->slots[level] = 0;
2475         return 0;
2476 }
2477
2478 /*
2479  * worker function to insert a single pointer in a node.
2480  * the node should have enough room for the pointer already
2481  *
2482  * slot and level indicate where you want the key to go, and
2483  * blocknr is the block the key points to.
2484  */
2485 static void insert_ptr(struct btrfs_trans_handle *trans,
2486                        struct btrfs_path *path,
2487                        struct btrfs_disk_key *key, u64 bytenr,
2488                        int slot, int level)
2489 {
2490         struct extent_buffer *lower;
2491         int nritems;
2492         int ret;
2493
2494         BUG_ON(!path->nodes[level]);
2495         btrfs_assert_tree_locked(path->nodes[level]);
2496         lower = path->nodes[level];
2497         nritems = btrfs_header_nritems(lower);
2498         BUG_ON(slot > nritems);
2499         BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(trans->fs_info));
2500         if (slot != nritems) {
2501                 if (level) {
2502                         ret = btrfs_tree_mod_log_insert_move(lower, slot + 1,
2503                                         slot, nritems - slot);
2504                         BUG_ON(ret < 0);
2505                 }
2506                 memmove_extent_buffer(lower,
2507                               btrfs_node_key_ptr_offset(slot + 1),
2508                               btrfs_node_key_ptr_offset(slot),
2509                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
2510         }
2511         if (level) {
2512                 ret = btrfs_tree_mod_log_insert_key(lower, slot,
2513                                             BTRFS_MOD_LOG_KEY_ADD, GFP_NOFS);
2514                 BUG_ON(ret < 0);
2515         }
2516         btrfs_set_node_key(lower, key, slot);
2517         btrfs_set_node_blockptr(lower, slot, bytenr);
2518         WARN_ON(trans->transid == 0);
2519         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
2520         btrfs_set_header_nritems(lower, nritems + 1);
2521         btrfs_mark_buffer_dirty(lower);
2522 }
2523
2524 /*
2525  * split the node at the specified level in path in two.
2526  * The path is corrected to point to the appropriate node after the split
2527  *
2528  * Before splitting this tries to make some room in the node by pushing
2529  * left and right, if either one works, it returns right away.
2530  *
2531  * returns 0 on success and < 0 on failure
2532  */
2533 static noinline int split_node(struct btrfs_trans_handle *trans,
2534                                struct btrfs_root *root,
2535                                struct btrfs_path *path, int level)
2536 {
2537         struct btrfs_fs_info *fs_info = root->fs_info;
2538         struct extent_buffer *c;
2539         struct extent_buffer *split;
2540         struct btrfs_disk_key disk_key;
2541         int mid;
2542         int ret;
2543         u32 c_nritems;
2544
2545         c = path->nodes[level];
2546         WARN_ON(btrfs_header_generation(c) != trans->transid);
2547         if (c == root->node) {
2548                 /*
2549                  * trying to split the root, lets make a new one
2550                  *
2551                  * tree mod log: We don't log_removal old root in
2552                  * insert_new_root, because that root buffer will be kept as a
2553                  * normal node. We are going to log removal of half of the
2554                  * elements below with btrfs_tree_mod_log_eb_copy(). We're
2555                  * holding a tree lock on the buffer, which is why we cannot
2556                  * race with other tree_mod_log users.
2557                  */
2558                 ret = insert_new_root(trans, root, path, level + 1);
2559                 if (ret)
2560                         return ret;
2561         } else {
2562                 ret = push_nodes_for_insert(trans, root, path, level);
2563                 c = path->nodes[level];
2564                 if (!ret && btrfs_header_nritems(c) <
2565                     BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3)
2566                         return 0;
2567                 if (ret < 0)
2568                         return ret;
2569         }
2570
2571         c_nritems = btrfs_header_nritems(c);
2572         mid = (c_nritems + 1) / 2;
2573         btrfs_node_key(c, &disk_key, mid);
2574
2575         split = alloc_tree_block_no_bg_flush(trans, root, 0, &disk_key, level,
2576                                              c->start, 0, BTRFS_NESTING_SPLIT);
2577         if (IS_ERR(split))
2578                 return PTR_ERR(split);
2579
2580         root_add_used(root, fs_info->nodesize);
2581         ASSERT(btrfs_header_level(c) == level);
2582
2583         ret = btrfs_tree_mod_log_eb_copy(split, c, 0, mid, c_nritems - mid);
2584         if (ret) {
2585                 btrfs_abort_transaction(trans, ret);
2586                 return ret;
2587         }
2588         copy_extent_buffer(split, c,
2589                            btrfs_node_key_ptr_offset(0),
2590                            btrfs_node_key_ptr_offset(mid),
2591                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
2592         btrfs_set_header_nritems(split, c_nritems - mid);
2593         btrfs_set_header_nritems(c, mid);
2594
2595         btrfs_mark_buffer_dirty(c);
2596         btrfs_mark_buffer_dirty(split);
2597
2598         insert_ptr(trans, path, &disk_key, split->start,
2599                    path->slots[level + 1] + 1, level + 1);
2600
2601         if (path->slots[level] >= mid) {
2602                 path->slots[level] -= mid;
2603                 btrfs_tree_unlock(c);
2604                 free_extent_buffer(c);
2605                 path->nodes[level] = split;
2606                 path->slots[level + 1] += 1;
2607         } else {
2608                 btrfs_tree_unlock(split);
2609                 free_extent_buffer(split);
2610         }
2611         return 0;
2612 }
2613
2614 /*
2615  * how many bytes are required to store the items in a leaf.  start
2616  * and nr indicate which items in the leaf to check.  This totals up the
2617  * space used both by the item structs and the item data
2618  */
2619 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
2620 {
2621         struct btrfs_item *start_item;
2622         struct btrfs_item *end_item;
2623         int data_len;
2624         int nritems = btrfs_header_nritems(l);
2625         int end = min(nritems, start + nr) - 1;
2626
2627         if (!nr)
2628                 return 0;
2629         start_item = btrfs_item_nr(start);
2630         end_item = btrfs_item_nr(end);
2631         data_len = btrfs_item_offset(l, start_item) +
2632                    btrfs_item_size(l, start_item);
2633         data_len = data_len - btrfs_item_offset(l, end_item);
2634         data_len += sizeof(struct btrfs_item) * nr;
2635         WARN_ON(data_len < 0);
2636         return data_len;
2637 }
2638
2639 /*
2640  * The space between the end of the leaf items and
2641  * the start of the leaf data.  IOW, how much room
2642  * the leaf has left for both items and data
2643  */
2644 noinline int btrfs_leaf_free_space(struct extent_buffer *leaf)
2645 {
2646         struct btrfs_fs_info *fs_info = leaf->fs_info;
2647         int nritems = btrfs_header_nritems(leaf);
2648         int ret;
2649
2650         ret = BTRFS_LEAF_DATA_SIZE(fs_info) - leaf_space_used(leaf, 0, nritems);
2651         if (ret < 0) {
2652                 btrfs_crit(fs_info,
2653                            "leaf free space ret %d, leaf data size %lu, used %d nritems %d",
2654                            ret,
2655                            (unsigned long) BTRFS_LEAF_DATA_SIZE(fs_info),
2656                            leaf_space_used(leaf, 0, nritems), nritems);
2657         }
2658         return ret;
2659 }
2660
2661 /*
2662  * min slot controls the lowest index we're willing to push to the
2663  * right.  We'll push up to and including min_slot, but no lower
2664  */
2665 static noinline int __push_leaf_right(struct btrfs_path *path,
2666                                       int data_size, int empty,
2667                                       struct extent_buffer *right,
2668                                       int free_space, u32 left_nritems,
2669                                       u32 min_slot)
2670 {
2671         struct btrfs_fs_info *fs_info = right->fs_info;
2672         struct extent_buffer *left = path->nodes[0];
2673         struct extent_buffer *upper = path->nodes[1];
2674         struct btrfs_map_token token;
2675         struct btrfs_disk_key disk_key;
2676         int slot;
2677         u32 i;
2678         int push_space = 0;
2679         int push_items = 0;
2680         struct btrfs_item *item;
2681         u32 nr;
2682         u32 right_nritems;
2683         u32 data_end;
2684         u32 this_item_size;
2685
2686         if (empty)
2687                 nr = 0;
2688         else
2689                 nr = max_t(u32, 1, min_slot);
2690
2691         if (path->slots[0] >= left_nritems)
2692                 push_space += data_size;
2693
2694         slot = path->slots[1];
2695         i = left_nritems - 1;
2696         while (i >= nr) {
2697                 item = btrfs_item_nr(i);
2698
2699                 if (!empty && push_items > 0) {
2700                         if (path->slots[0] > i)
2701                                 break;
2702                         if (path->slots[0] == i) {
2703                                 int space = btrfs_leaf_free_space(left);
2704
2705                                 if (space + push_space * 2 > free_space)
2706                                         break;
2707                         }
2708                 }
2709
2710                 if (path->slots[0] == i)
2711                         push_space += data_size;
2712
2713                 this_item_size = btrfs_item_size(left, item);
2714                 if (this_item_size + sizeof(*item) + push_space > free_space)
2715                         break;
2716
2717                 push_items++;
2718                 push_space += this_item_size + sizeof(*item);
2719                 if (i == 0)
2720                         break;
2721                 i--;
2722         }
2723
2724         if (push_items == 0)
2725                 goto out_unlock;
2726
2727         WARN_ON(!empty && push_items == left_nritems);
2728
2729         /* push left to right */
2730         right_nritems = btrfs_header_nritems(right);
2731
2732         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
2733         push_space -= leaf_data_end(left);
2734
2735         /* make room in the right data area */
2736         data_end = leaf_data_end(right);
2737         memmove_extent_buffer(right,
2738                               BTRFS_LEAF_DATA_OFFSET + data_end - push_space,
2739                               BTRFS_LEAF_DATA_OFFSET + data_end,
2740                               BTRFS_LEAF_DATA_SIZE(fs_info) - data_end);
2741
2742         /* copy from the left data area */
2743         copy_extent_buffer(right, left, BTRFS_LEAF_DATA_OFFSET +
2744                      BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
2745                      BTRFS_LEAF_DATA_OFFSET + leaf_data_end(left),
2746                      push_space);
2747
2748         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
2749                               btrfs_item_nr_offset(0),
2750                               right_nritems * sizeof(struct btrfs_item));
2751
2752         /* copy the items from left to right */
2753         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
2754                    btrfs_item_nr_offset(left_nritems - push_items),
2755                    push_items * sizeof(struct btrfs_item));
2756
2757         /* update the item pointers */
2758         btrfs_init_map_token(&token, right);
2759         right_nritems += push_items;
2760         btrfs_set_header_nritems(right, right_nritems);
2761         push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
2762         for (i = 0; i < right_nritems; i++) {
2763                 item = btrfs_item_nr(i);
2764                 push_space -= btrfs_token_item_size(&token, item);
2765                 btrfs_set_token_item_offset(&token, item, push_space);
2766         }
2767
2768         left_nritems -= push_items;
2769         btrfs_set_header_nritems(left, left_nritems);
2770
2771         if (left_nritems)
2772                 btrfs_mark_buffer_dirty(left);
2773         else
2774                 btrfs_clean_tree_block(left);
2775
2776         btrfs_mark_buffer_dirty(right);
2777
2778         btrfs_item_key(right, &disk_key, 0);
2779         btrfs_set_node_key(upper, &disk_key, slot + 1);
2780         btrfs_mark_buffer_dirty(upper);
2781
2782         /* then fixup the leaf pointer in the path */
2783         if (path->slots[0] >= left_nritems) {
2784                 path->slots[0] -= left_nritems;
2785                 if (btrfs_header_nritems(path->nodes[0]) == 0)
2786                         btrfs_clean_tree_block(path->nodes[0]);
2787                 btrfs_tree_unlock(path->nodes[0]);
2788                 free_extent_buffer(path->nodes[0]);
2789                 path->nodes[0] = right;
2790                 path->slots[1] += 1;
2791         } else {
2792                 btrfs_tree_unlock(right);
2793                 free_extent_buffer(right);
2794         }
2795         return 0;
2796
2797 out_unlock:
2798         btrfs_tree_unlock(right);
2799         free_extent_buffer(right);
2800         return 1;
2801 }
2802
2803 /*
2804  * push some data in the path leaf to the right, trying to free up at
2805  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
2806  *
2807  * returns 1 if the push failed because the other node didn't have enough
2808  * room, 0 if everything worked out and < 0 if there were major errors.
2809  *
2810  * this will push starting from min_slot to the end of the leaf.  It won't
2811  * push any slot lower than min_slot
2812  */
2813 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
2814                            *root, struct btrfs_path *path,
2815                            int min_data_size, int data_size,
2816                            int empty, u32 min_slot)
2817 {
2818         struct extent_buffer *left = path->nodes[0];
2819         struct extent_buffer *right;
2820         struct extent_buffer *upper;
2821         int slot;
2822         int free_space;
2823         u32 left_nritems;
2824         int ret;
2825
2826         if (!path->nodes[1])
2827                 return 1;
2828
2829         slot = path->slots[1];
2830         upper = path->nodes[1];
2831         if (slot >= btrfs_header_nritems(upper) - 1)
2832                 return 1;
2833
2834         btrfs_assert_tree_locked(path->nodes[1]);
2835
2836         right = btrfs_read_node_slot(upper, slot + 1);
2837         /*
2838          * slot + 1 is not valid or we fail to read the right node,
2839          * no big deal, just return.
2840          */
2841         if (IS_ERR(right))
2842                 return 1;
2843
2844         __btrfs_tree_lock(right, BTRFS_NESTING_RIGHT);
2845
2846         free_space = btrfs_leaf_free_space(right);
2847         if (free_space < data_size)
2848                 goto out_unlock;
2849
2850         /* cow and double check */
2851         ret = btrfs_cow_block(trans, root, right, upper,
2852                               slot + 1, &right, BTRFS_NESTING_RIGHT_COW);
2853         if (ret)
2854                 goto out_unlock;
2855
2856         free_space = btrfs_leaf_free_space(right);
2857         if (free_space < data_size)
2858                 goto out_unlock;
2859
2860         left_nritems = btrfs_header_nritems(left);
2861         if (left_nritems == 0)
2862                 goto out_unlock;
2863
2864         if (check_sibling_keys(left, right)) {
2865                 ret = -EUCLEAN;
2866                 btrfs_tree_unlock(right);
2867                 free_extent_buffer(right);
2868                 return ret;
2869         }
2870         if (path->slots[0] == left_nritems && !empty) {
2871                 /* Key greater than all keys in the leaf, right neighbor has
2872                  * enough room for it and we're not emptying our leaf to delete
2873                  * it, therefore use right neighbor to insert the new item and
2874                  * no need to touch/dirty our left leaf. */
2875                 btrfs_tree_unlock(left);
2876                 free_extent_buffer(left);
2877                 path->nodes[0] = right;
2878                 path->slots[0] = 0;
2879                 path->slots[1]++;
2880                 return 0;
2881         }
2882
2883         return __push_leaf_right(path, min_data_size, empty,
2884                                 right, free_space, left_nritems, min_slot);
2885 out_unlock:
2886         btrfs_tree_unlock(right);
2887         free_extent_buffer(right);
2888         return 1;
2889 }
2890
2891 /*
2892  * push some data in the path leaf to the left, trying to free up at
2893  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
2894  *
2895  * max_slot can put a limit on how far into the leaf we'll push items.  The
2896  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us do all the
2897  * items
2898  */
2899 static noinline int __push_leaf_left(struct btrfs_path *path, int data_size,
2900                                      int empty, struct extent_buffer *left,
2901                                      int free_space, u32 right_nritems,
2902                                      u32 max_slot)
2903 {
2904         struct btrfs_fs_info *fs_info = left->fs_info;
2905         struct btrfs_disk_key disk_key;
2906         struct extent_buffer *right = path->nodes[0];
2907         int i;
2908         int push_space = 0;
2909         int push_items = 0;
2910         struct btrfs_item *item;
2911         u32 old_left_nritems;
2912         u32 nr;
2913         int ret = 0;
2914         u32 this_item_size;
2915         u32 old_left_item_size;
2916         struct btrfs_map_token token;
2917
2918         if (empty)
2919                 nr = min(right_nritems, max_slot);
2920         else
2921                 nr = min(right_nritems - 1, max_slot);
2922
2923         for (i = 0; i < nr; i++) {
2924                 item = btrfs_item_nr(i);
2925
2926                 if (!empty && push_items > 0) {
2927                         if (path->slots[0] < i)
2928                                 break;
2929                         if (path->slots[0] == i) {
2930                                 int space = btrfs_leaf_free_space(right);
2931
2932                                 if (space + push_space * 2 > free_space)
2933                                         break;
2934                         }
2935                 }
2936
2937                 if (path->slots[0] == i)
2938                         push_space += data_size;
2939
2940                 this_item_size = btrfs_item_size(right, item);
2941                 if (this_item_size + sizeof(*item) + push_space > free_space)
2942                         break;
2943
2944                 push_items++;
2945                 push_space += this_item_size + sizeof(*item);
2946         }
2947
2948         if (push_items == 0) {
2949                 ret = 1;
2950                 goto out;
2951         }
2952         WARN_ON(!empty && push_items == btrfs_header_nritems(right));
2953
2954         /* push data from right to left */
2955         copy_extent_buffer(left, right,
2956                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
2957                            btrfs_item_nr_offset(0),
2958                            push_items * sizeof(struct btrfs_item));
2959
2960         push_space = BTRFS_LEAF_DATA_SIZE(fs_info) -
2961                      btrfs_item_offset_nr(right, push_items - 1);
2962
2963         copy_extent_buffer(left, right, BTRFS_LEAF_DATA_OFFSET +
2964                      leaf_data_end(left) - push_space,
2965                      BTRFS_LEAF_DATA_OFFSET +
2966                      btrfs_item_offset_nr(right, push_items - 1),
2967                      push_space);
2968         old_left_nritems = btrfs_header_nritems(left);
2969         BUG_ON(old_left_nritems <= 0);
2970
2971         btrfs_init_map_token(&token, left);
2972         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
2973         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
2974                 u32 ioff;
2975
2976                 item = btrfs_item_nr(i);
2977
2978                 ioff = btrfs_token_item_offset(&token, item);
2979                 btrfs_set_token_item_offset(&token, item,
2980                       ioff - (BTRFS_LEAF_DATA_SIZE(fs_info) - old_left_item_size));
2981         }
2982         btrfs_set_header_nritems(left, old_left_nritems + push_items);
2983
2984         /* fixup right node */
2985         if (push_items > right_nritems)
2986                 WARN(1, KERN_CRIT "push items %d nr %u\n", push_items,
2987                        right_nritems);
2988
2989         if (push_items < right_nritems) {
2990                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
2991                                                   leaf_data_end(right);
2992                 memmove_extent_buffer(right, BTRFS_LEAF_DATA_OFFSET +
2993                                       BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
2994                                       BTRFS_LEAF_DATA_OFFSET +
2995                                       leaf_data_end(right), push_space);
2996
2997                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
2998                               btrfs_item_nr_offset(push_items),
2999                              (btrfs_header_nritems(right) - push_items) *
3000                              sizeof(struct btrfs_item));
3001         }
3002
3003         btrfs_init_map_token(&token, right);
3004         right_nritems -= push_items;
3005         btrfs_set_header_nritems(right, right_nritems);
3006         push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3007         for (i = 0; i < right_nritems; i++) {
3008                 item = btrfs_item_nr(i);
3009
3010                 push_space = push_space - btrfs_token_item_size(&token, item);
3011                 btrfs_set_token_item_offset(&token, item, push_space);
3012         }
3013
3014         btrfs_mark_buffer_dirty(left);
3015         if (right_nritems)
3016                 btrfs_mark_buffer_dirty(right);
3017         else
3018                 btrfs_clean_tree_block(right);
3019
3020         btrfs_item_key(right, &disk_key, 0);
3021         fixup_low_keys(path, &disk_key, 1);
3022
3023         /* then fixup the leaf pointer in the path */
3024         if (path->slots[0] < push_items) {
3025                 path->slots[0] += old_left_nritems;
3026                 btrfs_tree_unlock(path->nodes[0]);
3027                 free_extent_buffer(path->nodes[0]);
3028                 path->nodes[0] = left;
3029                 path->slots[1] -= 1;
3030         } else {
3031                 btrfs_tree_unlock(left);
3032                 free_extent_buffer(left);
3033                 path->slots[0] -= push_items;
3034         }
3035         BUG_ON(path->slots[0] < 0);
3036         return ret;
3037 out:
3038         btrfs_tree_unlock(left);
3039         free_extent_buffer(left);
3040         return ret;
3041 }
3042
3043 /*
3044  * push some data in the path leaf to the left, trying to free up at
3045  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
3046  *
3047  * max_slot can put a limit on how far into the leaf we'll push items.  The
3048  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us push all the
3049  * items
3050  */
3051 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
3052                           *root, struct btrfs_path *path, int min_data_size,
3053                           int data_size, int empty, u32 max_slot)
3054 {
3055         struct extent_buffer *right = path->nodes[0];
3056         struct extent_buffer *left;
3057         int slot;
3058         int free_space;
3059         u32 right_nritems;
3060         int ret = 0;
3061
3062         slot = path->slots[1];
3063         if (slot == 0)
3064                 return 1;
3065         if (!path->nodes[1])
3066                 return 1;
3067
3068         right_nritems = btrfs_header_nritems(right);
3069         if (right_nritems == 0)
3070                 return 1;
3071
3072         btrfs_assert_tree_locked(path->nodes[1]);
3073
3074         left = btrfs_read_node_slot(path->nodes[1], slot - 1);
3075         /*
3076          * slot - 1 is not valid or we fail to read the left node,
3077          * no big deal, just return.
3078          */
3079         if (IS_ERR(left))
3080                 return 1;
3081
3082         __btrfs_tree_lock(left, BTRFS_NESTING_LEFT);
3083
3084         free_space = btrfs_leaf_free_space(left);
3085         if (free_space < data_size) {
3086                 ret = 1;
3087                 goto out;
3088         }
3089
3090         /* cow and double check */
3091         ret = btrfs_cow_block(trans, root, left,
3092                               path->nodes[1], slot - 1, &left,
3093                               BTRFS_NESTING_LEFT_COW);
3094         if (ret) {
3095                 /* we hit -ENOSPC, but it isn't fatal here */
3096                 if (ret == -ENOSPC)
3097                         ret = 1;
3098                 goto out;
3099         }
3100
3101         free_space = btrfs_leaf_free_space(left);
3102         if (free_space < data_size) {
3103                 ret = 1;
3104                 goto out;
3105         }
3106
3107         if (check_sibling_keys(left, right)) {
3108                 ret = -EUCLEAN;
3109                 goto out;
3110         }
3111         return __push_leaf_left(path, min_data_size,
3112                                empty, left, free_space, right_nritems,
3113                                max_slot);
3114 out:
3115         btrfs_tree_unlock(left);
3116         free_extent_buffer(left);
3117         return ret;
3118 }
3119
3120 /*
3121  * split the path's leaf in two, making sure there is at least data_size
3122  * available for the resulting leaf level of the path.
3123  */
3124 static noinline void copy_for_split(struct btrfs_trans_handle *trans,
3125                                     struct btrfs_path *path,
3126                                     struct extent_buffer *l,
3127                                     struct extent_buffer *right,
3128                                     int slot, int mid, int nritems)
3129 {
3130         struct btrfs_fs_info *fs_info = trans->fs_info;
3131         int data_copy_size;
3132         int rt_data_off;
3133         int i;
3134         struct btrfs_disk_key disk_key;
3135         struct btrfs_map_token token;
3136
3137         nritems = nritems - mid;
3138         btrfs_set_header_nritems(right, nritems);
3139         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(l);
3140
3141         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
3142                            btrfs_item_nr_offset(mid),
3143                            nritems * sizeof(struct btrfs_item));
3144
3145         copy_extent_buffer(right, l,
3146                      BTRFS_LEAF_DATA_OFFSET + BTRFS_LEAF_DATA_SIZE(fs_info) -
3147                      data_copy_size, BTRFS_LEAF_DATA_OFFSET +
3148                      leaf_data_end(l), data_copy_size);
3149
3150         rt_data_off = BTRFS_LEAF_DATA_SIZE(fs_info) - btrfs_item_end_nr(l, mid);
3151
3152         btrfs_init_map_token(&token, right);
3153         for (i = 0; i < nritems; i++) {
3154                 struct btrfs_item *item = btrfs_item_nr(i);
3155                 u32 ioff;
3156
3157                 ioff = btrfs_token_item_offset(&token, item);
3158                 btrfs_set_token_item_offset(&token, item, ioff + rt_data_off);
3159         }
3160
3161         btrfs_set_header_nritems(l, mid);
3162         btrfs_item_key(right, &disk_key, 0);
3163         insert_ptr(trans, path, &disk_key, right->start, path->slots[1] + 1, 1);
3164
3165         btrfs_mark_buffer_dirty(right);
3166         btrfs_mark_buffer_dirty(l);
3167         BUG_ON(path->slots[0] != slot);
3168
3169         if (mid <= slot) {
3170                 btrfs_tree_unlock(path->nodes[0]);
3171                 free_extent_buffer(path->nodes[0]);
3172                 path->nodes[0] = right;
3173                 path->slots[0] -= mid;
3174                 path->slots[1] += 1;
3175         } else {
3176                 btrfs_tree_unlock(right);
3177                 free_extent_buffer(right);
3178         }
3179
3180         BUG_ON(path->slots[0] < 0);
3181 }
3182
3183 /*
3184  * double splits happen when we need to insert a big item in the middle
3185  * of a leaf.  A double split can leave us with 3 mostly empty leaves:
3186  * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
3187  *          A                 B                 C
3188  *
3189  * We avoid this by trying to push the items on either side of our target
3190  * into the adjacent leaves.  If all goes well we can avoid the double split
3191  * completely.
3192  */
3193 static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
3194                                           struct btrfs_root *root,
3195                                           struct btrfs_path *path,
3196                                           int data_size)
3197 {
3198         int ret;
3199         int progress = 0;
3200         int slot;
3201         u32 nritems;
3202         int space_needed = data_size;
3203
3204         slot = path->slots[0];
3205         if (slot < btrfs_header_nritems(path->nodes[0]))
3206                 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
3207
3208         /*
3209          * try to push all the items after our slot into the
3210          * right leaf
3211          */
3212         ret = push_leaf_right(trans, root, path, 1, space_needed, 0, slot);
3213         if (ret < 0)
3214                 return ret;
3215
3216         if (ret == 0)
3217                 progress++;
3218
3219         nritems = btrfs_header_nritems(path->nodes[0]);
3220         /*
3221          * our goal is to get our slot at the start or end of a leaf.  If
3222          * we've done so we're done
3223          */
3224         if (path->slots[0] == 0 || path->slots[0] == nritems)
3225                 return 0;
3226
3227         if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
3228                 return 0;
3229
3230         /* try to push all the items before our slot into the next leaf */
3231         slot = path->slots[0];
3232         space_needed = data_size;
3233         if (slot > 0)
3234                 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
3235         ret = push_leaf_left(trans, root, path, 1, space_needed, 0, slot);
3236         if (ret < 0)
3237                 return ret;
3238
3239         if (ret == 0)
3240                 progress++;
3241
3242         if (progress)
3243                 return 0;
3244         return 1;
3245 }
3246
3247 /*
3248  * split the path's leaf in two, making sure there is at least data_size
3249  * available for the resulting leaf level of the path.
3250  *
3251  * returns 0 if all went well and < 0 on failure.
3252  */
3253 static noinline int split_leaf(struct btrfs_trans_handle *trans,
3254                                struct btrfs_root *root,
3255                                const struct btrfs_key *ins_key,
3256                                struct btrfs_path *path, int data_size,
3257                                int extend)
3258 {
3259         struct btrfs_disk_key disk_key;
3260         struct extent_buffer *l;
3261         u32 nritems;
3262         int mid;
3263         int slot;
3264         struct extent_buffer *right;
3265         struct btrfs_fs_info *fs_info = root->fs_info;
3266         int ret = 0;
3267         int wret;
3268         int split;
3269         int num_doubles = 0;
3270         int tried_avoid_double = 0;
3271
3272         l = path->nodes[0];
3273         slot = path->slots[0];
3274         if (extend && data_size + btrfs_item_size_nr(l, slot) +
3275             sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(fs_info))
3276                 return -EOVERFLOW;
3277
3278         /* first try to make some room by pushing left and right */
3279         if (data_size && path->nodes[1]) {
3280                 int space_needed = data_size;
3281
3282                 if (slot < btrfs_header_nritems(l))
3283                         space_needed -= btrfs_leaf_free_space(l);
3284
3285                 wret = push_leaf_right(trans, root, path, space_needed,
3286                                        space_needed, 0, 0);
3287                 if (wret < 0)
3288                         return wret;
3289                 if (wret) {
3290                         space_needed = data_size;
3291                         if (slot > 0)
3292                                 space_needed -= btrfs_leaf_free_space(l);
3293                         wret = push_leaf_left(trans, root, path, space_needed,
3294                                               space_needed, 0, (u32)-1);
3295                         if (wret < 0)
3296                                 return wret;
3297                 }
3298                 l = path->nodes[0];
3299
3300                 /* did the pushes work? */
3301                 if (btrfs_leaf_free_space(l) >= data_size)
3302                         return 0;
3303         }
3304
3305         if (!path->nodes[1]) {
3306                 ret = insert_new_root(trans, root, path, 1);
3307                 if (ret)
3308                         return ret;
3309         }
3310 again:
3311         split = 1;
3312         l = path->nodes[0];
3313         slot = path->slots[0];
3314         nritems = btrfs_header_nritems(l);
3315         mid = (nritems + 1) / 2;
3316
3317         if (mid <= slot) {
3318                 if (nritems == 1 ||
3319                     leaf_space_used(l, mid, nritems - mid) + data_size >
3320                         BTRFS_LEAF_DATA_SIZE(fs_info)) {
3321                         if (slot >= nritems) {
3322                                 split = 0;
3323                         } else {
3324                                 mid = slot;
3325                                 if (mid != nritems &&
3326                                     leaf_space_used(l, mid, nritems - mid) +
3327                                     data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
3328                                         if (data_size && !tried_avoid_double)
3329                                                 goto push_for_double;
3330                                         split = 2;
3331                                 }
3332                         }
3333                 }
3334         } else {
3335                 if (leaf_space_used(l, 0, mid) + data_size >
3336                         BTRFS_LEAF_DATA_SIZE(fs_info)) {
3337                         if (!extend && data_size && slot == 0) {
3338                                 split = 0;
3339                         } else if ((extend || !data_size) && slot == 0) {
3340                                 mid = 1;
3341                         } else {
3342                                 mid = slot;
3343                                 if (mid != nritems &&
3344                                     leaf_space_used(l, mid, nritems - mid) +
3345                                     data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
3346                                         if (data_size && !tried_avoid_double)
3347                                                 goto push_for_double;
3348                                         split = 2;
3349                                 }
3350                         }
3351                 }
3352         }
3353
3354         if (split == 0)
3355                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
3356         else
3357                 btrfs_item_key(l, &disk_key, mid);
3358
3359         /*
3360          * We have to about BTRFS_NESTING_NEW_ROOT here if we've done a double
3361          * split, because we're only allowed to have MAX_LOCKDEP_SUBCLASSES
3362          * subclasses, which is 8 at the time of this patch, and we've maxed it
3363          * out.  In the future we could add a
3364          * BTRFS_NESTING_SPLIT_THE_SPLITTENING if we need to, but for now just
3365          * use BTRFS_NESTING_NEW_ROOT.
3366          */
3367         right = alloc_tree_block_no_bg_flush(trans, root, 0, &disk_key, 0,
3368                                              l->start, 0, num_doubles ?
3369                                              BTRFS_NESTING_NEW_ROOT :
3370                                              BTRFS_NESTING_SPLIT);
3371         if (IS_ERR(right))
3372                 return PTR_ERR(right);
3373
3374         root_add_used(root, fs_info->nodesize);
3375
3376         if (split == 0) {
3377                 if (mid <= slot) {
3378                         btrfs_set_header_nritems(right, 0);
3379                         insert_ptr(trans, path, &disk_key,
3380                                    right->start, path->slots[1] + 1, 1);
3381                         btrfs_tree_unlock(path->nodes[0]);
3382                         free_extent_buffer(path->nodes[0]);
3383                         path->nodes[0] = right;
3384                         path->slots[0] = 0;
3385                         path->slots[1] += 1;
3386                 } else {
3387                         btrfs_set_header_nritems(right, 0);
3388                         insert_ptr(trans, path, &disk_key,
3389                                    right->start, path->slots[1], 1);
3390                         btrfs_tree_unlock(path->nodes[0]);
3391                         free_extent_buffer(path->nodes[0]);
3392                         path->nodes[0] = right;
3393                         path->slots[0] = 0;
3394                         if (path->slots[1] == 0)
3395                                 fixup_low_keys(path, &disk_key, 1);
3396                 }
3397                 /*
3398                  * We create a new leaf 'right' for the required ins_len and
3399                  * we'll do btrfs_mark_buffer_dirty() on this leaf after copying
3400                  * the content of ins_len to 'right'.
3401                  */
3402                 return ret;
3403         }
3404
3405         copy_for_split(trans, path, l, right, slot, mid, nritems);
3406
3407         if (split == 2) {
3408                 BUG_ON(num_doubles != 0);
3409                 num_doubles++;
3410                 goto again;
3411         }
3412
3413         return 0;
3414
3415 push_for_double:
3416         push_for_double_split(trans, root, path, data_size);
3417         tried_avoid_double = 1;
3418         if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
3419                 return 0;
3420         goto again;
3421 }
3422
3423 static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
3424                                          struct btrfs_root *root,
3425                                          struct btrfs_path *path, int ins_len)
3426 {
3427         struct btrfs_key key;
3428         struct extent_buffer *leaf;
3429         struct btrfs_file_extent_item *fi;
3430         u64 extent_len = 0;
3431         u32 item_size;
3432         int ret;
3433
3434         leaf = path->nodes[0];
3435         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3436
3437         BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
3438                key.type != BTRFS_EXTENT_CSUM_KEY);
3439
3440         if (btrfs_leaf_free_space(leaf) >= ins_len)
3441                 return 0;
3442
3443         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3444         if (key.type == BTRFS_EXTENT_DATA_KEY) {
3445                 fi = btrfs_item_ptr(leaf, path->slots[0],
3446                                     struct btrfs_file_extent_item);
3447                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
3448         }
3449         btrfs_release_path(path);
3450
3451         path->keep_locks = 1;
3452         path->search_for_split = 1;
3453         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
3454         path->search_for_split = 0;
3455         if (ret > 0)
3456                 ret = -EAGAIN;
3457         if (ret < 0)
3458                 goto err;
3459
3460         ret = -EAGAIN;
3461         leaf = path->nodes[0];
3462         /* if our item isn't there, return now */
3463         if (item_size != btrfs_item_size_nr(leaf, path->slots[0]))
3464                 goto err;
3465
3466         /* the leaf has  changed, it now has room.  return now */
3467         if (btrfs_leaf_free_space(path->nodes[0]) >= ins_len)
3468                 goto err;
3469
3470         if (key.type == BTRFS_EXTENT_DATA_KEY) {
3471                 fi = btrfs_item_ptr(leaf, path->slots[0],
3472                                     struct btrfs_file_extent_item);
3473                 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
3474                         goto err;
3475         }
3476
3477         ret = split_leaf(trans, root, &key, path, ins_len, 1);
3478         if (ret)
3479                 goto err;
3480
3481         path->keep_locks = 0;
3482         btrfs_unlock_up_safe(path, 1);
3483         return 0;
3484 err:
3485         path->keep_locks = 0;
3486         return ret;
3487 }
3488
3489 static noinline int split_item(struct btrfs_path *path,
3490                                const struct btrfs_key *new_key,
3491                                unsigned long split_offset)
3492 {
3493         struct extent_buffer *leaf;
3494         struct btrfs_item *item;
3495         struct btrfs_item *new_item;
3496         int slot;
3497         char *buf;
3498         u32 nritems;
3499         u32 item_size;
3500         u32 orig_offset;
3501         struct btrfs_disk_key disk_key;
3502
3503         leaf = path->nodes[0];
3504         BUG_ON(btrfs_leaf_free_space(leaf) < sizeof(struct btrfs_item));
3505
3506         item = btrfs_item_nr(path->slots[0]);
3507         orig_offset = btrfs_item_offset(leaf, item);
3508         item_size = btrfs_item_size(leaf, item);
3509
3510         buf = kmalloc(item_size, GFP_NOFS);
3511         if (!buf)
3512                 return -ENOMEM;
3513
3514         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
3515                             path->slots[0]), item_size);
3516
3517         slot = path->slots[0] + 1;
3518         nritems = btrfs_header_nritems(leaf);
3519         if (slot != nritems) {
3520                 /* shift the items */
3521                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
3522                                 btrfs_item_nr_offset(slot),
3523                                 (nritems - slot) * sizeof(struct btrfs_item));
3524         }
3525
3526         btrfs_cpu_key_to_disk(&disk_key, new_key);
3527         btrfs_set_item_key(leaf, &disk_key, slot);
3528
3529         new_item = btrfs_item_nr(slot);
3530
3531         btrfs_set_item_offset(leaf, new_item, orig_offset);
3532         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
3533
3534         btrfs_set_item_offset(leaf, item,
3535                               orig_offset + item_size - split_offset);
3536         btrfs_set_item_size(leaf, item, split_offset);
3537
3538         btrfs_set_header_nritems(leaf, nritems + 1);
3539
3540         /* write the data for the start of the original item */
3541         write_extent_buffer(leaf, buf,
3542                             btrfs_item_ptr_offset(leaf, path->slots[0]),
3543                             split_offset);
3544
3545         /* write the data for the new item */
3546         write_extent_buffer(leaf, buf + split_offset,
3547                             btrfs_item_ptr_offset(leaf, slot),
3548                             item_size - split_offset);
3549         btrfs_mark_buffer_dirty(leaf);
3550
3551         BUG_ON(btrfs_leaf_free_space(leaf) < 0);
3552         kfree(buf);
3553         return 0;
3554 }
3555
3556 /*
3557  * This function splits a single item into two items,
3558  * giving 'new_key' to the new item and splitting the
3559  * old one at split_offset (from the start of the item).
3560  *
3561  * The path may be released by this operation.  After
3562  * the split, the path is pointing to the old item.  The
3563  * new item is going to be in the same node as the old one.
3564  *
3565  * Note, the item being split must be smaller enough to live alone on
3566  * a tree block with room for one extra struct btrfs_item
3567  *
3568  * This allows us to split the item in place, keeping a lock on the
3569  * leaf the entire time.
3570  */
3571 int btrfs_split_item(struct btrfs_trans_handle *trans,
3572                      struct btrfs_root *root,
3573                      struct btrfs_path *path,
3574                      const struct btrfs_key *new_key,
3575                      unsigned long split_offset)
3576 {
3577         int ret;
3578         ret = setup_leaf_for_split(trans, root, path,
3579                                    sizeof(struct btrfs_item));
3580         if (ret)
3581                 return ret;
3582
3583         ret = split_item(path, new_key, split_offset);
3584         return ret;
3585 }
3586
3587 /*
3588  * This function duplicate a item, giving 'new_key' to the new item.
3589  * It guarantees both items live in the same tree leaf and the new item
3590  * is contiguous with the original item.
3591  *
3592  * This allows us to split file extent in place, keeping a lock on the
3593  * leaf the entire time.
3594  */
3595 int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
3596                          struct btrfs_root *root,
3597                          struct btrfs_path *path,
3598                          const struct btrfs_key *new_key)
3599 {
3600         struct extent_buffer *leaf;
3601         int ret;
3602         u32 item_size;
3603
3604         leaf = path->nodes[0];
3605         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3606         ret = setup_leaf_for_split(trans, root, path,
3607                                    item_size + sizeof(struct btrfs_item));
3608         if (ret)
3609                 return ret;
3610
3611         path->slots[0]++;
3612         setup_items_for_insert(root, path, new_key, &item_size, 1);
3613         leaf = path->nodes[0];
3614         memcpy_extent_buffer(leaf,
3615                              btrfs_item_ptr_offset(leaf, path->slots[0]),
3616                              btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
3617                              item_size);
3618         return 0;
3619 }
3620
3621 /*
3622  * make the item pointed to by the path smaller.  new_size indicates
3623  * how small to make it, and from_end tells us if we just chop bytes
3624  * off the end of the item or if we shift the item to chop bytes off
3625  * the front.
3626  */
3627 void btrfs_truncate_item(struct btrfs_path *path, u32 new_size, int from_end)
3628 {
3629         int slot;
3630         struct extent_buffer *leaf;
3631         struct btrfs_item *item;
3632         u32 nritems;
3633         unsigned int data_end;
3634         unsigned int old_data_start;
3635         unsigned int old_size;
3636         unsigned int size_diff;
3637         int i;
3638         struct btrfs_map_token token;
3639
3640         leaf = path->nodes[0];
3641         slot = path->slots[0];
3642
3643         old_size = btrfs_item_size_nr(leaf, slot);
3644         if (old_size == new_size)
3645                 return;
3646
3647         nritems = btrfs_header_nritems(leaf);
3648         data_end = leaf_data_end(leaf);
3649
3650         old_data_start = btrfs_item_offset_nr(leaf, slot);
3651
3652         size_diff = old_size - new_size;
3653
3654         BUG_ON(slot < 0);
3655         BUG_ON(slot >= nritems);
3656
3657         /*
3658          * item0..itemN ... dataN.offset..dataN.size .. data0.size
3659          */
3660         /* first correct the data pointers */
3661         btrfs_init_map_token(&token, leaf);
3662         for (i = slot; i < nritems; i++) {
3663                 u32 ioff;
3664                 item = btrfs_item_nr(i);
3665
3666                 ioff = btrfs_token_item_offset(&token, item);
3667                 btrfs_set_token_item_offset(&token, item, ioff + size_diff);
3668         }
3669
3670         /* shift the data */
3671         if (from_end) {
3672                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
3673                               data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
3674                               data_end, old_data_start + new_size - data_end);
3675         } else {
3676                 struct btrfs_disk_key disk_key;
3677                 u64 offset;
3678
3679                 btrfs_item_key(leaf, &disk_key, slot);
3680
3681                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
3682                         unsigned long ptr;
3683                         struct btrfs_file_extent_item *fi;
3684
3685                         fi = btrfs_item_ptr(leaf, slot,
3686                                             struct btrfs_file_extent_item);
3687                         fi = (struct btrfs_file_extent_item *)(
3688                              (unsigned long)fi - size_diff);
3689
3690                         if (btrfs_file_extent_type(leaf, fi) ==
3691                             BTRFS_FILE_EXTENT_INLINE) {
3692                                 ptr = btrfs_item_ptr_offset(leaf, slot);
3693                                 memmove_extent_buffer(leaf, ptr,
3694                                       (unsigned long)fi,
3695                                       BTRFS_FILE_EXTENT_INLINE_DATA_START);
3696                         }
3697                 }
3698
3699                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
3700                               data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
3701                               data_end, old_data_start - data_end);
3702
3703                 offset = btrfs_disk_key_offset(&disk_key);
3704                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
3705                 btrfs_set_item_key(leaf, &disk_key, slot);
3706                 if (slot == 0)
3707                         fixup_low_keys(path, &disk_key, 1);
3708         }
3709
3710         item = btrfs_item_nr(slot);
3711         btrfs_set_item_size(leaf, item, new_size);
3712         btrfs_mark_buffer_dirty(leaf);
3713
3714         if (btrfs_leaf_free_space(leaf) < 0) {
3715                 btrfs_print_leaf(leaf);
3716                 BUG();
3717         }
3718 }
3719
3720 /*
3721  * make the item pointed to by the path bigger, data_size is the added size.
3722  */
3723 void btrfs_extend_item(struct btrfs_path *path, u32 data_size)
3724 {
3725         int slot;
3726         struct extent_buffer *leaf;
3727         struct btrfs_item *item;
3728         u32 nritems;
3729         unsigned int data_end;
3730         unsigned int old_data;
3731         unsigned int old_size;
3732         int i;
3733         struct btrfs_map_token token;
3734
3735         leaf = path->nodes[0];
3736
3737         nritems = btrfs_header_nritems(leaf);
3738         data_end = leaf_data_end(leaf);
3739
3740         if (btrfs_leaf_free_space(leaf) < data_size) {
3741                 btrfs_print_leaf(leaf);
3742                 BUG();
3743         }
3744         slot = path->slots[0];
3745         old_data = btrfs_item_end_nr(leaf, slot);
3746
3747         BUG_ON(slot < 0);
3748         if (slot >= nritems) {
3749                 btrfs_print_leaf(leaf);
3750                 btrfs_crit(leaf->fs_info, "slot %d too large, nritems %d",
3751                            slot, nritems);
3752                 BUG();
3753         }
3754
3755         /*
3756          * item0..itemN ... dataN.offset..dataN.size .. data0.size
3757          */
3758         /* first correct the data pointers */
3759         btrfs_init_map_token(&token, leaf);
3760         for (i = slot; i < nritems; i++) {
3761                 u32 ioff;
3762                 item = btrfs_item_nr(i);
3763
3764                 ioff = btrfs_token_item_offset(&token, item);
3765                 btrfs_set_token_item_offset(&token, item, ioff - data_size);
3766         }
3767
3768         /* shift the data */
3769         memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
3770                       data_end - data_size, BTRFS_LEAF_DATA_OFFSET +
3771                       data_end, old_data - data_end);
3772
3773         data_end = old_data;
3774         old_size = btrfs_item_size_nr(leaf, slot);
3775         item = btrfs_item_nr(slot);
3776         btrfs_set_item_size(leaf, item, old_size + data_size);
3777         btrfs_mark_buffer_dirty(leaf);
3778
3779         if (btrfs_leaf_free_space(leaf) < 0) {
3780                 btrfs_print_leaf(leaf);
3781                 BUG();
3782         }
3783 }
3784
3785 /**
3786  * setup_items_for_insert - Helper called before inserting one or more items
3787  * to a leaf. Main purpose is to save stack depth by doing the bulk of the work
3788  * in a function that doesn't call btrfs_search_slot
3789  *
3790  * @root:       root we are inserting items to
3791  * @path:       points to the leaf/slot where we are going to insert new items
3792  * @cpu_key:    array of keys for items to be inserted
3793  * @data_size:  size of the body of each item we are going to insert
3794  * @nr:         size of @cpu_key/@data_size arrays
3795  */
3796 void setup_items_for_insert(struct btrfs_root *root, struct btrfs_path *path,
3797                             const struct btrfs_key *cpu_key, u32 *data_size,
3798                             int nr)
3799 {
3800         struct btrfs_fs_info *fs_info = root->fs_info;
3801         struct btrfs_item *item;
3802         int i;
3803         u32 nritems;
3804         unsigned int data_end;
3805         struct btrfs_disk_key disk_key;
3806         struct extent_buffer *leaf;
3807         int slot;
3808         struct btrfs_map_token token;
3809         u32 total_size;
3810         u32 total_data = 0;
3811
3812         for (i = 0; i < nr; i++)
3813                 total_data += data_size[i];
3814         total_size = total_data + (nr * sizeof(struct btrfs_item));
3815
3816         if (path->slots[0] == 0) {
3817                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
3818                 fixup_low_keys(path, &disk_key, 1);
3819         }
3820         btrfs_unlock_up_safe(path, 1);
3821
3822         leaf = path->nodes[0];
3823         slot = path->slots[0];
3824
3825         nritems = btrfs_header_nritems(leaf);
3826         data_end = leaf_data_end(leaf);
3827
3828         if (btrfs_leaf_free_space(leaf) < total_size) {
3829                 btrfs_print_leaf(leaf);
3830                 btrfs_crit(fs_info, "not enough freespace need %u have %d",
3831                            total_size, btrfs_leaf_free_space(leaf));
3832                 BUG();
3833         }
3834
3835         btrfs_init_map_token(&token, leaf);
3836         if (slot != nritems) {
3837                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
3838
3839                 if (old_data < data_end) {
3840                         btrfs_print_leaf(leaf);
3841                         btrfs_crit(fs_info,
3842                 "item at slot %d with data offset %u beyond data end of leaf %u",
3843                                    slot, old_data, data_end);
3844                         BUG();
3845                 }
3846                 /*
3847                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
3848                  */
3849                 /* first correct the data pointers */
3850                 for (i = slot; i < nritems; i++) {
3851                         u32 ioff;
3852
3853                         item = btrfs_item_nr(i);
3854                         ioff = btrfs_token_item_offset(&token, item);
3855                         btrfs_set_token_item_offset(&token, item,
3856                                                     ioff - total_data);
3857                 }
3858                 /* shift the items */
3859                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
3860                               btrfs_item_nr_offset(slot),
3861                               (nritems - slot) * sizeof(struct btrfs_item));
3862
3863                 /* shift the data */
3864                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
3865                               data_end - total_data, BTRFS_LEAF_DATA_OFFSET +
3866                               data_end, old_data - data_end);
3867                 data_end = old_data;
3868         }
3869
3870         /* setup the item for the new data */
3871         for (i = 0; i < nr; i++) {
3872                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
3873                 btrfs_set_item_key(leaf, &disk_key, slot + i);
3874                 item = btrfs_item_nr(slot + i);
3875                 data_end -= data_size[i];
3876                 btrfs_set_token_item_offset(&token, item, data_end);
3877                 btrfs_set_token_item_size(&token, item, data_size[i]);
3878         }
3879
3880         btrfs_set_header_nritems(leaf, nritems + nr);
3881         btrfs_mark_buffer_dirty(leaf);
3882
3883         if (btrfs_leaf_free_space(leaf) < 0) {
3884                 btrfs_print_leaf(leaf);
3885                 BUG();
3886         }
3887 }
3888
3889 /*
3890  * Given a key and some data, insert items into the tree.
3891  * This does all the path init required, making room in the tree if needed.
3892  */
3893 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
3894                             struct btrfs_root *root,
3895                             struct btrfs_path *path,
3896                             const struct btrfs_key *cpu_key, u32 *data_size,
3897                             int nr)
3898 {
3899         int ret = 0;
3900         int slot;
3901         int i;
3902         u32 total_size = 0;
3903         u32 total_data = 0;
3904
3905         for (i = 0; i < nr; i++)
3906                 total_data += data_size[i];
3907
3908         total_size = total_data + (nr * sizeof(struct btrfs_item));
3909         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
3910         if (ret == 0)
3911                 return -EEXIST;
3912         if (ret < 0)
3913                 return ret;
3914
3915         slot = path->slots[0];
3916         BUG_ON(slot < 0);
3917
3918         setup_items_for_insert(root, path, cpu_key, data_size, nr);
3919         return 0;
3920 }
3921
3922 /*
3923  * Given a key and some data, insert an item into the tree.
3924  * This does all the path init required, making room in the tree if needed.
3925  */
3926 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3927                       const struct btrfs_key *cpu_key, void *data,
3928                       u32 data_size)
3929 {
3930         int ret = 0;
3931         struct btrfs_path *path;
3932         struct extent_buffer *leaf;
3933         unsigned long ptr;
3934
3935         path = btrfs_alloc_path();
3936         if (!path)
3937                 return -ENOMEM;
3938         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
3939         if (!ret) {
3940                 leaf = path->nodes[0];
3941                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3942                 write_extent_buffer(leaf, data, ptr, data_size);
3943                 btrfs_mark_buffer_dirty(leaf);
3944         }
3945         btrfs_free_path(path);
3946         return ret;
3947 }
3948
3949 /*
3950  * delete the pointer from a given node.
3951  *
3952  * the tree should have been previously balanced so the deletion does not
3953  * empty a node.
3954  */
3955 static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
3956                     int level, int slot)
3957 {
3958         struct extent_buffer *parent = path->nodes[level];
3959         u32 nritems;
3960         int ret;
3961
3962         nritems = btrfs_header_nritems(parent);
3963         if (slot != nritems - 1) {
3964                 if (level) {
3965                         ret = btrfs_tree_mod_log_insert_move(parent, slot,
3966                                         slot + 1, nritems - slot - 1);
3967                         BUG_ON(ret < 0);
3968                 }
3969                 memmove_extent_buffer(parent,
3970                               btrfs_node_key_ptr_offset(slot),
3971                               btrfs_node_key_ptr_offset(slot + 1),
3972                               sizeof(struct btrfs_key_ptr) *
3973                               (nritems - slot - 1));
3974         } else if (level) {
3975                 ret = btrfs_tree_mod_log_insert_key(parent, slot,
3976                                 BTRFS_MOD_LOG_KEY_REMOVE, GFP_NOFS);
3977                 BUG_ON(ret < 0);
3978         }
3979
3980         nritems--;
3981         btrfs_set_header_nritems(parent, nritems);
3982         if (nritems == 0 && parent == root->node) {
3983                 BUG_ON(btrfs_header_level(root->node) != 1);
3984                 /* just turn the root into a leaf and break */
3985                 btrfs_set_header_level(root->node, 0);
3986         } else if (slot == 0) {
3987                 struct btrfs_disk_key disk_key;
3988
3989                 btrfs_node_key(parent, &disk_key, 0);
3990                 fixup_low_keys(path, &disk_key, level + 1);
3991         }
3992         btrfs_mark_buffer_dirty(parent);
3993 }
3994
3995 /*
3996  * a helper function to delete the leaf pointed to by path->slots[1] and
3997  * path->nodes[1].
3998  *
3999  * This deletes the pointer in path->nodes[1] and frees the leaf
4000  * block extent.  zero is returned if it all worked out, < 0 otherwise.
4001  *
4002  * The path must have already been setup for deleting the leaf, including
4003  * all the proper balancing.  path->nodes[1] must be locked.
4004  */
4005 static noinline void btrfs_del_leaf(struct btrfs_trans_handle *trans,
4006                                     struct btrfs_root *root,
4007                                     struct btrfs_path *path,
4008                                     struct extent_buffer *leaf)
4009 {
4010         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
4011         del_ptr(root, path, 1, path->slots[1]);
4012
4013         /*
4014          * btrfs_free_extent is expensive, we want to make sure we
4015          * aren't holding any locks when we call it
4016          */
4017         btrfs_unlock_up_safe(path, 0);
4018
4019         root_sub_used(root, leaf->len);
4020
4021         atomic_inc(&leaf->refs);
4022         btrfs_free_tree_block(trans, root, leaf, 0, 1);
4023         free_extent_buffer_stale(leaf);
4024 }
4025 /*
4026  * delete the item at the leaf level in path.  If that empties
4027  * the leaf, remove it from the tree
4028  */
4029 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4030                     struct btrfs_path *path, int slot, int nr)
4031 {
4032         struct btrfs_fs_info *fs_info = root->fs_info;
4033         struct extent_buffer *leaf;
4034         struct btrfs_item *item;
4035         u32 last_off;
4036         u32 dsize = 0;
4037         int ret = 0;
4038         int wret;
4039         int i;
4040         u32 nritems;
4041
4042         leaf = path->nodes[0];
4043         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
4044
4045         for (i = 0; i < nr; i++)
4046                 dsize += btrfs_item_size_nr(leaf, slot + i);
4047
4048         nritems = btrfs_header_nritems(leaf);
4049
4050         if (slot + nr != nritems) {
4051                 int data_end = leaf_data_end(leaf);
4052                 struct btrfs_map_token token;
4053
4054                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4055                               data_end + dsize,
4056                               BTRFS_LEAF_DATA_OFFSET + data_end,
4057                               last_off - data_end);
4058
4059                 btrfs_init_map_token(&token, leaf);
4060                 for (i = slot + nr; i < nritems; i++) {
4061                         u32 ioff;
4062
4063                         item = btrfs_item_nr(i);
4064                         ioff = btrfs_token_item_offset(&token, item);
4065                         btrfs_set_token_item_offset(&token, item, ioff + dsize);
4066                 }
4067
4068                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
4069                               btrfs_item_nr_offset(slot + nr),
4070                               sizeof(struct btrfs_item) *
4071                               (nritems - slot - nr));
4072         }
4073         btrfs_set_header_nritems(leaf, nritems - nr);
4074         nritems -= nr;
4075
4076         /* delete the leaf if we've emptied it */
4077         if (nritems == 0) {
4078                 if (leaf == root->node) {
4079                         btrfs_set_header_level(leaf, 0);
4080                 } else {
4081                         btrfs_clean_tree_block(leaf);
4082                         btrfs_del_leaf(trans, root, path, leaf);
4083                 }
4084         } else {
4085                 int used = leaf_space_used(leaf, 0, nritems);
4086                 if (slot == 0) {
4087                         struct btrfs_disk_key disk_key;
4088
4089                         btrfs_item_key(leaf, &disk_key, 0);
4090                         fixup_low_keys(path, &disk_key, 1);
4091                 }
4092
4093                 /* delete the leaf if it is mostly empty */
4094                 if (used < BTRFS_LEAF_DATA_SIZE(fs_info) / 3) {
4095                         /* push_leaf_left fixes the path.
4096                          * make sure the path still points to our leaf
4097                          * for possible call to del_ptr below
4098                          */
4099                         slot = path->slots[1];
4100                         atomic_inc(&leaf->refs);
4101
4102                         wret = push_leaf_left(trans, root, path, 1, 1,
4103                                               1, (u32)-1);
4104                         if (wret < 0 && wret != -ENOSPC)
4105                                 ret = wret;
4106
4107                         if (path->nodes[0] == leaf &&
4108                             btrfs_header_nritems(leaf)) {
4109                                 wret = push_leaf_right(trans, root, path, 1,
4110                                                        1, 1, 0);
4111                                 if (wret < 0 && wret != -ENOSPC)
4112                                         ret = wret;
4113                         }
4114
4115                         if (btrfs_header_nritems(leaf) == 0) {
4116                                 path->slots[1] = slot;
4117                                 btrfs_del_leaf(trans, root, path, leaf);
4118                                 free_extent_buffer(leaf);
4119                                 ret = 0;
4120                         } else {
4121                                 /* if we're still in the path, make sure
4122                                  * we're dirty.  Otherwise, one of the
4123                                  * push_leaf functions must have already
4124                                  * dirtied this buffer
4125                                  */
4126                                 if (path->nodes[0] == leaf)
4127                                         btrfs_mark_buffer_dirty(leaf);
4128                                 free_extent_buffer(leaf);
4129                         }
4130                 } else {
4131                         btrfs_mark_buffer_dirty(leaf);
4132                 }
4133         }
4134         return ret;
4135 }
4136
4137 /*
4138  * search the tree again to find a leaf with lesser keys
4139  * returns 0 if it found something or 1 if there are no lesser leaves.
4140  * returns < 0 on io errors.
4141  *
4142  * This may release the path, and so you may lose any locks held at the
4143  * time you call it.
4144  */
4145 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
4146 {
4147         struct btrfs_key key;
4148         struct btrfs_disk_key found_key;
4149         int ret;
4150
4151         btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
4152
4153         if (key.offset > 0) {
4154                 key.offset--;
4155         } else if (key.type > 0) {
4156                 key.type--;
4157                 key.offset = (u64)-1;
4158         } else if (key.objectid > 0) {
4159                 key.objectid--;
4160                 key.type = (u8)-1;
4161                 key.offset = (u64)-1;
4162         } else {
4163                 return 1;
4164         }
4165
4166         btrfs_release_path(path);
4167         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4168         if (ret < 0)
4169                 return ret;
4170         btrfs_item_key(path->nodes[0], &found_key, 0);
4171         ret = comp_keys(&found_key, &key);
4172         /*
4173          * We might have had an item with the previous key in the tree right
4174          * before we released our path. And after we released our path, that
4175          * item might have been pushed to the first slot (0) of the leaf we
4176          * were holding due to a tree balance. Alternatively, an item with the
4177          * previous key can exist as the only element of a leaf (big fat item).
4178          * Therefore account for these 2 cases, so that our callers (like
4179          * btrfs_previous_item) don't miss an existing item with a key matching
4180          * the previous key we computed above.
4181          */
4182         if (ret <= 0)
4183                 return 0;
4184         return 1;
4185 }
4186
4187 /*
4188  * A helper function to walk down the tree starting at min_key, and looking
4189  * for nodes or leaves that are have a minimum transaction id.
4190  * This is used by the btree defrag code, and tree logging
4191  *
4192  * This does not cow, but it does stuff the starting key it finds back
4193  * into min_key, so you can call btrfs_search_slot with cow=1 on the
4194  * key and get a writable path.
4195  *
4196  * This honors path->lowest_level to prevent descent past a given level
4197  * of the tree.
4198  *
4199  * min_trans indicates the oldest transaction that you are interested
4200  * in walking through.  Any nodes or leaves older than min_trans are
4201  * skipped over (without reading them).
4202  *
4203  * returns zero if something useful was found, < 0 on error and 1 if there
4204  * was nothing in the tree that matched the search criteria.
4205  */
4206 int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
4207                          struct btrfs_path *path,
4208                          u64 min_trans)
4209 {
4210         struct extent_buffer *cur;
4211         struct btrfs_key found_key;
4212         int slot;
4213         int sret;
4214         u32 nritems;
4215         int level;
4216         int ret = 1;
4217         int keep_locks = path->keep_locks;
4218
4219         path->keep_locks = 1;
4220 again:
4221         cur = btrfs_read_lock_root_node(root);
4222         level = btrfs_header_level(cur);
4223         WARN_ON(path->nodes[level]);
4224         path->nodes[level] = cur;
4225         path->locks[level] = BTRFS_READ_LOCK;
4226
4227         if (btrfs_header_generation(cur) < min_trans) {
4228                 ret = 1;
4229                 goto out;
4230         }
4231         while (1) {
4232                 nritems = btrfs_header_nritems(cur);
4233                 level = btrfs_header_level(cur);
4234                 sret = btrfs_bin_search(cur, min_key, &slot);
4235                 if (sret < 0) {
4236                         ret = sret;
4237                         goto out;
4238                 }
4239
4240                 /* at the lowest level, we're done, setup the path and exit */
4241                 if (level == path->lowest_level) {
4242                         if (slot >= nritems)
4243                                 goto find_next_key;
4244                         ret = 0;
4245                         path->slots[level] = slot;
4246                         btrfs_item_key_to_cpu(cur, &found_key, slot);
4247                         goto out;
4248                 }
4249                 if (sret && slot > 0)
4250                         slot--;
4251                 /*
4252                  * check this node pointer against the min_trans parameters.
4253                  * If it is too old, skip to the next one.
4254                  */
4255                 while (slot < nritems) {
4256                         u64 gen;
4257
4258                         gen = btrfs_node_ptr_generation(cur, slot);
4259                         if (gen < min_trans) {
4260                                 slot++;
4261                                 continue;
4262                         }
4263                         break;
4264                 }
4265 find_next_key:
4266                 /*
4267                  * we didn't find a candidate key in this node, walk forward
4268                  * and find another one
4269                  */
4270                 if (slot >= nritems) {
4271                         path->slots[level] = slot;
4272                         sret = btrfs_find_next_key(root, path, min_key, level,
4273                                                   min_trans);
4274                         if (sret == 0) {
4275                                 btrfs_release_path(path);
4276                                 goto again;
4277                         } else {
4278                                 goto out;
4279                         }
4280                 }
4281                 /* save our key for returning back */
4282                 btrfs_node_key_to_cpu(cur, &found_key, slot);
4283                 path->slots[level] = slot;
4284                 if (level == path->lowest_level) {
4285                         ret = 0;
4286                         goto out;
4287                 }
4288                 cur = btrfs_read_node_slot(cur, slot);
4289                 if (IS_ERR(cur)) {
4290                         ret = PTR_ERR(cur);
4291                         goto out;
4292                 }
4293
4294                 btrfs_tree_read_lock(cur);
4295
4296                 path->locks[level - 1] = BTRFS_READ_LOCK;
4297                 path->nodes[level - 1] = cur;
4298                 unlock_up(path, level, 1, 0, NULL);
4299         }
4300 out:
4301         path->keep_locks = keep_locks;
4302         if (ret == 0) {
4303                 btrfs_unlock_up_safe(path, path->lowest_level + 1);
4304                 memcpy(min_key, &found_key, sizeof(found_key));
4305         }
4306         return ret;
4307 }
4308
4309 /*
4310  * this is similar to btrfs_next_leaf, but does not try to preserve
4311  * and fixup the path.  It looks for and returns the next key in the
4312  * tree based on the current path and the min_trans parameters.
4313  *
4314  * 0 is returned if another key is found, < 0 if there are any errors
4315  * and 1 is returned if there are no higher keys in the tree
4316  *
4317  * path->keep_locks should be set to 1 on the search made before
4318  * calling this function.
4319  */
4320 int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
4321                         struct btrfs_key *key, int level, u64 min_trans)
4322 {
4323         int slot;
4324         struct extent_buffer *c;
4325
4326         WARN_ON(!path->keep_locks && !path->skip_locking);
4327         while (level < BTRFS_MAX_LEVEL) {
4328                 if (!path->nodes[level])
4329                         return 1;
4330
4331                 slot = path->slots[level] + 1;
4332                 c = path->nodes[level];
4333 next:
4334                 if (slot >= btrfs_header_nritems(c)) {
4335                         int ret;
4336                         int orig_lowest;
4337                         struct btrfs_key cur_key;
4338                         if (level + 1 >= BTRFS_MAX_LEVEL ||
4339                             !path->nodes[level + 1])
4340                                 return 1;
4341
4342                         if (path->locks[level + 1] || path->skip_locking) {
4343                                 level++;
4344                                 continue;
4345                         }
4346
4347                         slot = btrfs_header_nritems(c) - 1;
4348                         if (level == 0)
4349                                 btrfs_item_key_to_cpu(c, &cur_key, slot);
4350                         else
4351                                 btrfs_node_key_to_cpu(c, &cur_key, slot);
4352
4353                         orig_lowest = path->lowest_level;
4354                         btrfs_release_path(path);
4355                         path->lowest_level = level;
4356                         ret = btrfs_search_slot(NULL, root, &cur_key, path,
4357                                                 0, 0);
4358                         path->lowest_level = orig_lowest;
4359                         if (ret < 0)
4360                                 return ret;
4361
4362                         c = path->nodes[level];
4363                         slot = path->slots[level];
4364                         if (ret == 0)
4365                                 slot++;
4366                         goto next;
4367                 }
4368
4369                 if (level == 0)
4370                         btrfs_item_key_to_cpu(c, key, slot);
4371                 else {
4372                         u64 gen = btrfs_node_ptr_generation(c, slot);
4373
4374                         if (gen < min_trans) {
4375                                 slot++;
4376                                 goto next;
4377                         }
4378                         btrfs_node_key_to_cpu(c, key, slot);
4379                 }
4380                 return 0;
4381         }
4382         return 1;
4383 }
4384
4385 /*
4386  * search the tree again to find a leaf with greater keys
4387  * returns 0 if it found something or 1 if there are no greater leaves.
4388  * returns < 0 on io errors.
4389  */
4390 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
4391 {
4392         return btrfs_next_old_leaf(root, path, 0);
4393 }
4394
4395 int btrfs_next_old_leaf(struct btrfs_root *root, struct btrfs_path *path,
4396                         u64 time_seq)
4397 {
4398         int slot;
4399         int level;
4400         struct extent_buffer *c;
4401         struct extent_buffer *next;
4402         struct btrfs_key key;
4403         u32 nritems;
4404         int ret;
4405         int i;
4406
4407         nritems = btrfs_header_nritems(path->nodes[0]);
4408         if (nritems == 0)
4409                 return 1;
4410
4411         btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
4412 again:
4413         level = 1;
4414         next = NULL;
4415         btrfs_release_path(path);
4416
4417         path->keep_locks = 1;
4418
4419         if (time_seq)
4420                 ret = btrfs_search_old_slot(root, &key, path, time_seq);
4421         else
4422                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4423         path->keep_locks = 0;
4424
4425         if (ret < 0)
4426                 return ret;
4427
4428         nritems = btrfs_header_nritems(path->nodes[0]);
4429         /*
4430          * by releasing the path above we dropped all our locks.  A balance
4431          * could have added more items next to the key that used to be
4432          * at the very end of the block.  So, check again here and
4433          * advance the path if there are now more items available.
4434          */
4435         if (nritems > 0 && path->slots[0] < nritems - 1) {
4436                 if (ret == 0)
4437                         path->slots[0]++;
4438                 ret = 0;
4439                 goto done;
4440         }
4441         /*
4442          * So the above check misses one case:
4443          * - after releasing the path above, someone has removed the item that
4444          *   used to be at the very end of the block, and balance between leafs
4445          *   gets another one with bigger key.offset to replace it.
4446          *
4447          * This one should be returned as well, or we can get leaf corruption
4448          * later(esp. in __btrfs_drop_extents()).
4449          *
4450          * And a bit more explanation about this check,
4451          * with ret > 0, the key isn't found, the path points to the slot
4452          * where it should be inserted, so the path->slots[0] item must be the
4453          * bigger one.
4454          */
4455         if (nritems > 0 && ret > 0 && path->slots[0] == nritems - 1) {
4456                 ret = 0;
4457                 goto done;
4458         }
4459
4460         while (level < BTRFS_MAX_LEVEL) {
4461                 if (!path->nodes[level]) {
4462                         ret = 1;
4463                         goto done;
4464                 }
4465
4466                 slot = path->slots[level] + 1;
4467                 c = path->nodes[level];
4468                 if (slot >= btrfs_header_nritems(c)) {
4469                         level++;
4470                         if (level == BTRFS_MAX_LEVEL) {
4471                                 ret = 1;
4472                                 goto done;
4473                         }
4474                         continue;
4475                 }
4476
4477
4478                 /*
4479                  * Our current level is where we're going to start from, and to
4480                  * make sure lockdep doesn't complain we need to drop our locks
4481                  * and nodes from 0 to our current level.
4482                  */
4483                 for (i = 0; i < level; i++) {
4484                         if (path->locks[level]) {
4485                                 btrfs_tree_read_unlock(path->nodes[i]);
4486                                 path->locks[i] = 0;
4487                         }
4488                         free_extent_buffer(path->nodes[i]);
4489                         path->nodes[i] = NULL;
4490                 }
4491
4492                 next = c;
4493                 ret = read_block_for_search(root, path, &next, level,
4494                                             slot, &key);
4495                 if (ret == -EAGAIN)
4496                         goto again;
4497
4498                 if (ret < 0) {
4499                         btrfs_release_path(path);
4500                         goto done;
4501                 }
4502
4503                 if (!path->skip_locking) {
4504                         ret = btrfs_try_tree_read_lock(next);
4505                         if (!ret && time_seq) {
4506                                 /*
4507                                  * If we don't get the lock, we may be racing
4508                                  * with push_leaf_left, holding that lock while
4509                                  * itself waiting for the leaf we've currently
4510                                  * locked. To solve this situation, we give up
4511                                  * on our lock and cycle.
4512                                  */
4513                                 free_extent_buffer(next);
4514                                 btrfs_release_path(path);
4515                                 cond_resched();
4516                                 goto again;
4517                         }
4518                         if (!ret)
4519                                 btrfs_tree_read_lock(next);
4520                 }
4521                 break;
4522         }
4523         path->slots[level] = slot;
4524         while (1) {
4525                 level--;
4526                 path->nodes[level] = next;
4527                 path->slots[level] = 0;
4528                 if (!path->skip_locking)
4529                         path->locks[level] = BTRFS_READ_LOCK;
4530                 if (!level)
4531                         break;
4532
4533                 ret = read_block_for_search(root, path, &next, level,
4534                                             0, &key);
4535                 if (ret == -EAGAIN)
4536                         goto again;
4537
4538                 if (ret < 0) {
4539                         btrfs_release_path(path);
4540                         goto done;
4541                 }
4542
4543                 if (!path->skip_locking)
4544                         btrfs_tree_read_lock(next);
4545         }
4546         ret = 0;
4547 done:
4548         unlock_up(path, 0, 1, 0, NULL);
4549
4550         return ret;
4551 }
4552
4553 /*
4554  * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
4555  * searching until it gets past min_objectid or finds an item of 'type'
4556  *
4557  * returns 0 if something is found, 1 if nothing was found and < 0 on error
4558  */
4559 int btrfs_previous_item(struct btrfs_root *root,
4560                         struct btrfs_path *path, u64 min_objectid,
4561                         int type)
4562 {
4563         struct btrfs_key found_key;
4564         struct extent_buffer *leaf;
4565         u32 nritems;
4566         int ret;
4567
4568         while (1) {
4569                 if (path->slots[0] == 0) {
4570                         ret = btrfs_prev_leaf(root, path);
4571                         if (ret != 0)
4572                                 return ret;
4573                 } else {
4574                         path->slots[0]--;
4575                 }
4576                 leaf = path->nodes[0];
4577                 nritems = btrfs_header_nritems(leaf);
4578                 if (nritems == 0)
4579                         return 1;
4580                 if (path->slots[0] == nritems)
4581                         path->slots[0]--;
4582
4583                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4584                 if (found_key.objectid < min_objectid)
4585                         break;
4586                 if (found_key.type == type)
4587                         return 0;
4588                 if (found_key.objectid == min_objectid &&
4589                     found_key.type < type)
4590                         break;
4591         }
4592         return 1;
4593 }
4594
4595 /*
4596  * search in extent tree to find a previous Metadata/Data extent item with
4597  * min objecitd.
4598  *
4599  * returns 0 if something is found, 1 if nothing was found and < 0 on error
4600  */
4601 int btrfs_previous_extent_item(struct btrfs_root *root,
4602                         struct btrfs_path *path, u64 min_objectid)
4603 {
4604         struct btrfs_key found_key;
4605         struct extent_buffer *leaf;
4606         u32 nritems;
4607         int ret;
4608
4609         while (1) {
4610                 if (path->slots[0] == 0) {
4611                         ret = btrfs_prev_leaf(root, path);
4612                         if (ret != 0)
4613                                 return ret;
4614                 } else {
4615                         path->slots[0]--;
4616                 }
4617                 leaf = path->nodes[0];
4618                 nritems = btrfs_header_nritems(leaf);
4619                 if (nritems == 0)
4620                         return 1;
4621                 if (path->slots[0] == nritems)
4622                         path->slots[0]--;
4623
4624                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4625                 if (found_key.objectid < min_objectid)
4626                         break;
4627                 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
4628                     found_key.type == BTRFS_METADATA_ITEM_KEY)
4629                         return 0;
4630                 if (found_key.objectid == min_objectid &&
4631                     found_key.type < BTRFS_EXTENT_ITEM_KEY)
4632                         break;
4633         }
4634         return 1;
4635 }