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 <linux/error-injection.h>
  11 #include "ctree.h"
  12 #include "disk-io.h"
  13 #include "transaction.h"
  14 #include "print-tree.h"
  15 #include "locking.h"
  16 #include "volumes.h"
  17 #include "qgroup.h"
  18 #include "tree-mod-log.h"
  19
  20 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  21                       *root, struct btrfs_path *path, int level);
  22 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  23                       const struct btrfs_key *ins_key, struct btrfs_path *path,
  24                       int data_size, int extend);
  25 static int push_node_left(struct btrfs_trans_handle *trans,
  26                           struct extent_buffer *dst,
  27                           struct extent_buffer *src, int empty);
  28 static int balance_node_right(struct btrfs_trans_handle *trans,
  29                               struct extent_buffer *dst_buf,
  30                               struct extent_buffer *src_buf);
  31 static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
  32                     int level, int slot);
  33
  34 static const struct btrfs_csums {
  35         u16             size;
  36         const char      name[10];
  37         const char      driver[12];
  38 } btrfs_csums[] = {
  39         [BTRFS_CSUM_TYPE_CRC32] = { .size = 4, .name = "crc32c" },
  40         [BTRFS_CSUM_TYPE_XXHASH] = { .size = 8, .name = "xxhash64" },
  41         [BTRFS_CSUM_TYPE_SHA256] = { .size = 32, .name = "sha256" },
  42         [BTRFS_CSUM_TYPE_BLAKE2] = { .size = 32, .name = "blake2b",
  43                                      .driver = "blake2b-256" },
  44 };
  45
  46 int btrfs_super_csum_size(const struct btrfs_super_block *s)
  47 {
  48         u16 t = btrfs_super_csum_type(s);
  49         /*
  50          * csum type is validated at mount time
  51          */
  52         return btrfs_csums[t].size;
  53 }
  54
  55 const char *btrfs_super_csum_name(u16 csum_type)
  56 {
  57         /* csum type is validated at mount time */
  58         return btrfs_csums[csum_type].name;
  59 }
  60
  61 /*
  62  * Return driver name if defined, otherwise the name that's also a valid driver
  63  * name
  64  */
  65 const char *btrfs_super_csum_driver(u16 csum_type)
  66 {
  67         /* csum type is validated at mount time */
  68         return btrfs_csums[csum_type].driver[0] ?
  69                 btrfs_csums[csum_type].driver :
  70                 btrfs_csums[csum_type].name;
  71 }
  72
  73 size_t __attribute_const__ btrfs_get_num_csums(void)
  74 {
  75         return ARRAY_SIZE(btrfs_csums);
  76 }
  77
  78 struct btrfs_path *btrfs_alloc_path(void)
  79 {
  80         return kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
  81 }
  82
  83 /* this also releases the path */
  84 void btrfs_free_path(struct btrfs_path *p)
  85 {
  86         if (!p)
  87                 return;
  88         btrfs_release_path(p);
  89         kmem_cache_free(btrfs_path_cachep, p);
  90 }
  91
  92 /*
  93  * path release drops references on the extent buffers in the path
  94  * and it drops any locks held by this path
  95  *
  96  * It is safe to call this on paths that no locks or extent buffers held.
  97  */
  98 noinline void btrfs_release_path(struct btrfs_path *p)
  99 {
 100         int i;
 101
 102         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
 103                 p->slots[i] = 0;
 104                 if (!p->nodes[i])
 105                         continue;
 106                 if (p->locks[i]) {
 107                         btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
 108                         p->locks[i] = 0;
 109                 }
 110                 free_extent_buffer(p->nodes[i]);
 111                 p->nodes[i] = NULL;
 112         }
 113 }
 114
 115 /*
 116  * safely gets a reference on the root node of a tree.  A lock
 117  * is not taken, so a concurrent writer may put a different node
 118  * at the root of the tree.  See btrfs_lock_root_node for the
 119  * looping required.
 120  *
 121  * The extent buffer returned by this has a reference taken, so
 122  * it won't disappear.  It may stop being the root of the tree
 123  * at any time because there are no locks held.
 124  */
 125 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
 126 {
 127         struct extent_buffer *eb;
 128
 129         while (1) {
 130                 rcu_read_lock();
 131                 eb = rcu_dereference(root->node);
 132
 133                 /*
 134                  * RCU really hurts here, we could free up the root node because
 135                  * it was COWed but we may not get the new root node yet so do
 136                  * the inc_not_zero dance and if it doesn't work then
 137                  * synchronize_rcu and try again.
 138                  */
 139                 if (atomic_inc_not_zero(&eb->refs)) {
 140                         rcu_read_unlock();
 141                         break;
 142                 }
 143                 rcu_read_unlock();
 144                 synchronize_rcu();
 145         }
 146         return eb;
 147 }
 148
 149 /*
 150  * Cowonly root (not-shareable trees, everything not subvolume or reloc roots),
 151  * just get put onto a simple dirty list.  Transaction walks this list to make
 152  * sure they get properly updated on disk.
 153  */
 154 static void add_root_to_dirty_list(struct btrfs_root *root)
 155 {
 156         struct btrfs_fs_info *fs_info = root->fs_info;
 157
 158         if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
 159             !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
 160                 return;
 161
 162         spin_lock(&fs_info->trans_lock);
 163         if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
 164                 /* Want the extent tree to be the last on the list */
 165                 if (root->root_key.objectid == BTRFS_EXTENT_TREE_OBJECTID)
 166                         list_move_tail(&root->dirty_list,
 167                                        &fs_info->dirty_cowonly_roots);
 168                 else
 169                         list_move(&root->dirty_list,
 170                                   &fs_info->dirty_cowonly_roots);
 171         }
 172         spin_unlock(&fs_info->trans_lock);
 173 }
 174
 175 /*
 176  * used by snapshot creation to make a copy of a root for a tree with
 177  * a given objectid.  The buffer with the new root node is returned in
 178  * cow_ret, and this func returns zero on success or a negative error code.
 179  */
 180 int btrfs_copy_root(struct btrfs_trans_handle *trans,
 181                       struct btrfs_root *root,
 182                       struct extent_buffer *buf,
 183                       struct extent_buffer **cow_ret, u64 new_root_objectid)
 184 {
 185         struct btrfs_fs_info *fs_info = root->fs_info;
 186         struct extent_buffer *cow;
 187         int ret = 0;
 188         int level;
 189         struct btrfs_disk_key disk_key;
 190
 191         WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 192                 trans->transid != fs_info->running_transaction->transid);
 193         WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 194                 trans->transid != root->last_trans);
 195
 196         level = btrfs_header_level(buf);
 197         if (level == 0)
 198                 btrfs_item_key(buf, &disk_key, 0);
 199         else
 200                 btrfs_node_key(buf, &disk_key, 0);
 201
 202         cow = btrfs_alloc_tree_block(trans, root, 0, new_root_objectid,
 203                                      &disk_key, level, buf->start, 0,
 204                                      BTRFS_NESTING_NEW_ROOT);
 205         if (IS_ERR(cow))
 206                 return PTR_ERR(cow);
 207
 208         copy_extent_buffer_full(cow, buf);
 209         btrfs_set_header_bytenr(cow, cow->start);
 210         btrfs_set_header_generation(cow, trans->transid);
 211         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 212         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 213                                      BTRFS_HEADER_FLAG_RELOC);
 214         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 215                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 216         else
 217                 btrfs_set_header_owner(cow, new_root_objectid);
 218
 219         write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
 220
 221         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 222         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 223                 ret = btrfs_inc_ref(trans, root, cow, 1);
 224         else
 225                 ret = btrfs_inc_ref(trans, root, cow, 0);
 226         if (ret) {
 227                 btrfs_tree_unlock(cow);
 228                 free_extent_buffer(cow);
 229                 btrfs_abort_transaction(trans, ret);
 230                 return ret;
 231         }
 232
 233         btrfs_mark_buffer_dirty(cow);
 234         *cow_ret = cow;
 235         return 0;
 236 }
 237
 238 /*
 239  * check if the tree block can be shared by multiple trees
 240  */
 241 int btrfs_block_can_be_shared(struct btrfs_root *root,
 242                               struct extent_buffer *buf)
 243 {
 244         /*
 245          * Tree blocks not in shareable trees and tree roots are never shared.
 246          * If a block was allocated after the last snapshot and the block was
 247          * not allocated by tree relocation, we know the block is not shared.
 248          */
 249         if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 250             buf != root->node && buf != root->commit_root &&
 251             (btrfs_header_generation(buf) <=
 252              btrfs_root_last_snapshot(&root->root_item) ||
 253              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 254                 return 1;
 255
 256         return 0;
 257 }
 258
 259 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
 260                                        struct btrfs_root *root,
 261                                        struct extent_buffer *buf,
 262                                        struct extent_buffer *cow,
 263                                        int *last_ref)
 264 {
 265         struct btrfs_fs_info *fs_info = root->fs_info;
 266         u64 refs;
 267         u64 owner;
 268         u64 flags;
 269         u64 new_flags = 0;
 270         int ret;
 271
 272         /*
 273          * Backrefs update rules:
 274          *
 275          * Always use full backrefs for extent pointers in tree block
 276          * allocated by tree relocation.
 277          *
 278          * If a shared tree block is no longer referenced by its owner
 279          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
 280          * use full backrefs for extent pointers in tree block.
 281          *
 282          * If a tree block is been relocating
 283          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
 284          * use full backrefs for extent pointers in tree block.
 285          * The reason for this is some operations (such as drop tree)
 286          * are only allowed for blocks use full backrefs.
 287          */
 288
 289         if (btrfs_block_can_be_shared(root, buf)) {
 290                 ret = btrfs_lookup_extent_info(trans, fs_info, buf->start,
 291                                                btrfs_header_level(buf), 1,
 292                                                &refs, &flags);
 293                 if (ret)
 294                         return ret;
 295                 if (refs == 0) {
 296                         ret = -EROFS;
 297                         btrfs_handle_fs_error(fs_info, ret, NULL);
 298                         return ret;
 299                 }
 300         } else {
 301                 refs = 1;
 302                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 303                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 304                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 305                 else
 306                         flags = 0;
 307         }
 308
 309         owner = btrfs_header_owner(buf);
 310         BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
 311                !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
 312
 313         if (refs > 1) {
 314                 if ((owner == root->root_key.objectid ||
 315                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
 316                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
 317                         ret = btrfs_inc_ref(trans, root, buf, 1);
 318                         if (ret)
 319                                 return ret;
 320
 321                         if (root->root_key.objectid ==
 322                             BTRFS_TREE_RELOC_OBJECTID) {
 323                                 ret = btrfs_dec_ref(trans, root, buf, 0);
 324                                 if (ret)
 325                                         return ret;
 326                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 327                                 if (ret)
 328                                         return ret;
 329                         }
 330                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
 331                 } else {
 332
 333                         if (root->root_key.objectid ==
 334                             BTRFS_TREE_RELOC_OBJECTID)
 335                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 336                         else
 337                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 338                         if (ret)
 339                                 return ret;
 340                 }
 341                 if (new_flags != 0) {
 342                         int level = btrfs_header_level(buf);
 343
 344                         ret = btrfs_set_disk_extent_flags(trans, buf,
 345                                                           new_flags, level, 0);
 346                         if (ret)
 347                                 return ret;
 348                 }
 349         } else {
 350                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
 351                         if (root->root_key.objectid ==
 352                             BTRFS_TREE_RELOC_OBJECTID)
 353                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 354                         else
 355                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 356                         if (ret)
 357                                 return ret;
 358                         ret = btrfs_dec_ref(trans, root, buf, 1);
 359                         if (ret)
 360                                 return ret;
 361                 }
 362                 btrfs_clean_tree_block(buf);
 363                 *last_ref = 1;
 364         }
 365         return 0;
 366 }
 367
 368 /*
 369  * does the dirty work in cow of a single block.  The parent block (if
 370  * supplied) is updated to point to the new cow copy.  The new buffer is marked
 371  * dirty and returned locked.  If you modify the block it needs to be marked
 372  * dirty again.
 373  *
 374  * search_start -- an allocation hint for the new block
 375  *
 376  * empty_size -- a hint that you plan on doing more cow.  This is the size in
 377  * bytes the allocator should try to find free next to the block it returns.
 378  * This is just a hint and may be ignored by the allocator.
 379  */
 380 static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 381                              struct btrfs_root *root,
 382                              struct extent_buffer *buf,
 383                              struct extent_buffer *parent, int parent_slot,
 384                              struct extent_buffer **cow_ret,
 385                              u64 search_start, u64 empty_size,
 386                              enum btrfs_lock_nesting nest)
 387 {
 388         struct btrfs_fs_info *fs_info = root->fs_info;
 389         struct btrfs_disk_key disk_key;
 390         struct extent_buffer *cow;
 391         int level, ret;
 392         int last_ref = 0;
 393         int unlock_orig = 0;
 394         u64 parent_start = 0;
 395
 396         if (*cow_ret == buf)
 397                 unlock_orig = 1;
 398
 399         btrfs_assert_tree_write_locked(buf);
 400
 401         WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 402                 trans->transid != fs_info->running_transaction->transid);
 403         WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 404                 trans->transid != root->last_trans);
 405
 406         level = btrfs_header_level(buf);
 407
 408         if (level == 0)
 409                 btrfs_item_key(buf, &disk_key, 0);
 410         else
 411                 btrfs_node_key(buf, &disk_key, 0);
 412
 413         if ((root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) && parent)
 414                 parent_start = parent->start;
 415
 416         cow = btrfs_alloc_tree_block(trans, root, parent_start,
 417                                      root->root_key.objectid, &disk_key, level,
 418                                      search_start, empty_size, nest);
 419         if (IS_ERR(cow))
 420                 return PTR_ERR(cow);
 421
 422         /* cow is set to blocking by btrfs_init_new_buffer */
 423
 424         copy_extent_buffer_full(cow, buf);
 425         btrfs_set_header_bytenr(cow, cow->start);
 426         btrfs_set_header_generation(cow, trans->transid);
 427         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 428         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 429                                      BTRFS_HEADER_FLAG_RELOC);
 430         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 431                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 432         else
 433                 btrfs_set_header_owner(cow, root->root_key.objectid);
 434
 435         write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
 436
 437         ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
 438         if (ret) {
 439                 btrfs_tree_unlock(cow);
 440                 free_extent_buffer(cow);
 441                 btrfs_abort_transaction(trans, ret);
 442                 return ret;
 443         }
 444
 445         if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
 446                 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
 447                 if (ret) {
 448                         btrfs_tree_unlock(cow);
 449                         free_extent_buffer(cow);
 450                         btrfs_abort_transaction(trans, ret);
 451                         return ret;
 452                 }
 453         }
 454
 455         if (buf == root->node) {
 456                 WARN_ON(parent && parent != buf);
 457                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 458                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 459                         parent_start = buf->start;
 460
 461                 atomic_inc(&cow->refs);
 462                 ret = btrfs_tree_mod_log_insert_root(root->node, cow, true);
 463                 BUG_ON(ret < 0);
 464                 rcu_assign_pointer(root->node, cow);
 465
 466                 btrfs_free_tree_block(trans, btrfs_root_id(root), buf,
 467                                       parent_start, last_ref);
 468                 free_extent_buffer(buf);
 469                 add_root_to_dirty_list(root);
 470         } else {
 471                 WARN_ON(trans->transid != btrfs_header_generation(parent));
 472                 btrfs_tree_mod_log_insert_key(parent, parent_slot,
 473                                               BTRFS_MOD_LOG_KEY_REPLACE, GFP_NOFS);
 474                 btrfs_set_node_blockptr(parent, parent_slot,
 475                                         cow->start);
 476                 btrfs_set_node_ptr_generation(parent, parent_slot,
 477                                               trans->transid);
 478                 btrfs_mark_buffer_dirty(parent);
 479                 if (last_ref) {
 480                         ret = btrfs_tree_mod_log_free_eb(buf);
 481                         if (ret) {
 482                                 btrfs_tree_unlock(cow);
 483                                 free_extent_buffer(cow);
 484                                 btrfs_abort_transaction(trans, ret);
 485                                 return ret;
 486                         }
 487                 }
 488                 btrfs_free_tree_block(trans, btrfs_root_id(root), buf,
 489                                       parent_start, last_ref);
 490         }
 491         if (unlock_orig)
 492                 btrfs_tree_unlock(buf);
 493         free_extent_buffer_stale(buf);
 494         btrfs_mark_buffer_dirty(cow);
 495         *cow_ret = cow;
 496         return 0;
 497 }
 498
 499 static inline int should_cow_block(struct btrfs_trans_handle *trans,
 500                                    struct btrfs_root *root,
 501                                    struct extent_buffer *buf)
 502 {
 503         if (btrfs_is_testing(root->fs_info))
 504                 return 0;
 505
 506         /* Ensure we can see the FORCE_COW bit */
 507         smp_mb__before_atomic();
 508
 509         /*
 510          * We do not need to cow a block if
 511          * 1) this block is not created or changed in this transaction;
 512          * 2) this block does not belong to TREE_RELOC tree;
 513          * 3) the root is not forced COW.
 514          *
 515          * What is forced COW:
 516          *    when we create snapshot during committing the transaction,
 517          *    after we've finished copying src root, we must COW the shared
 518          *    block to ensure the metadata consistency.
 519          */
 520         if (btrfs_header_generation(buf) == trans->transid &&
 521             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
 522             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
 523               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
 524             !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
 525                 return 0;
 526         return 1;
 527 }
 528
 529 /*
 530  * cows a single block, see __btrfs_cow_block for the real work.
 531  * This version of it has extra checks so that a block isn't COWed more than
 532  * once per transaction, as long as it hasn't been written yet
 533  */
 534 noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
 535                     struct btrfs_root *root, struct extent_buffer *buf,
 536                     struct extent_buffer *parent, int parent_slot,
 537                     struct extent_buffer **cow_ret,
 538                     enum btrfs_lock_nesting nest)
 539 {
 540         struct btrfs_fs_info *fs_info = root->fs_info;
 541         u64 search_start;
 542         int ret;
 543
 544         if (test_bit(BTRFS_ROOT_DELETING, &root->state))
 545                 btrfs_err(fs_info,
 546                         "COW'ing blocks on a fs root that's being dropped");
 547
 548         if (trans->transaction != fs_info->running_transaction)
 549                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
 550                        trans->transid,
 551                        fs_info->running_transaction->transid);
 552
 553         if (trans->transid != fs_info->generation)
 554                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
 555                        trans->transid, fs_info->generation);
 556
 557         if (!should_cow_block(trans, root, buf)) {
 558                 *cow_ret = buf;
 559                 return 0;
 560         }
 561
 562         search_start = buf->start & ~((u64)SZ_1G - 1);
 563
 564         /*
 565          * Before CoWing this block for later modification, check if it's
 566          * the subtree root and do the delayed subtree trace if needed.
 567          *
 568          * Also We don't care about the error, as it's handled internally.
 569          */
 570         btrfs_qgroup_trace_subtree_after_cow(trans, root, buf);
 571         ret = __btrfs_cow_block(trans, root, buf, parent,
 572                                  parent_slot, cow_ret, search_start, 0, nest);
 573
 574         trace_btrfs_cow_block(root, buf, *cow_ret);
 575
 576         return ret;
 577 }
 578 ALLOW_ERROR_INJECTION(btrfs_cow_block, ERRNO);
 579
 580 /*
 581  * helper function for defrag to decide if two blocks pointed to by a
 582  * node are actually close by
 583  */
 584 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 585 {
 586         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 587                 return 1;
 588         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 589                 return 1;
 590         return 0;
 591 }
 592
 593 #ifdef __LITTLE_ENDIAN
 594
 595 /*
 596  * Compare two keys, on little-endian the disk order is same as CPU order and
 597  * we can avoid the conversion.
 598  */
 599 static int comp_keys(const struct btrfs_disk_key *disk_key,
 600                      const struct btrfs_key *k2)
 601 {
 602         const struct btrfs_key *k1 = (const struct btrfs_key *)disk_key;
 603
 604         return btrfs_comp_cpu_keys(k1, k2);
 605 }
 606
 607 #else
 608
 609 /*
 610  * compare two keys in a memcmp fashion
 611  */
 612 static int comp_keys(const struct btrfs_disk_key *disk,
 613                      const struct btrfs_key *k2)
 614 {
 615         struct btrfs_key k1;
 616
 617         btrfs_disk_key_to_cpu(&k1, disk);
 618
 619         return btrfs_comp_cpu_keys(&k1, k2);
 620 }
 621 #endif
 622
 623 /*
 624  * same as comp_keys only with two btrfs_key's
 625  */
 626 int __pure btrfs_comp_cpu_keys(const struct btrfs_key *k1, const struct btrfs_key *k2)
 627 {
 628         if (k1->objectid > k2->objectid)
 629                 return 1;
 630         if (k1->objectid < k2->objectid)
 631                 return -1;
 632         if (k1->type > k2->type)
 633                 return 1;
 634         if (k1->type < k2->type)
 635                 return -1;
 636         if (k1->offset > k2->offset)
 637                 return 1;
 638         if (k1->offset < k2->offset)
 639                 return -1;
 640         return 0;
 641 }
 642
 643 /*
 644  * this is used by the defrag code to go through all the
 645  * leaves pointed to by a node and reallocate them so that
 646  * disk order is close to key order
 647  */
 648 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 649                        struct btrfs_root *root, struct extent_buffer *parent,
 650                        int start_slot, u64 *last_ret,
 651                        struct btrfs_key *progress)
 652 {
 653         struct btrfs_fs_info *fs_info = root->fs_info;
 654         struct extent_buffer *cur;
 655         u64 blocknr;
 656         u64 search_start = *last_ret;
 657         u64 last_block = 0;
 658         u64 other;
 659         u32 parent_nritems;
 660         int end_slot;
 661         int i;
 662         int err = 0;
 663         u32 blocksize;
 664         int progress_passed = 0;
 665         struct btrfs_disk_key disk_key;
 666
 667         WARN_ON(trans->transaction != fs_info->running_transaction);
 668         WARN_ON(trans->transid != fs_info->generation);
 669
 670         parent_nritems = btrfs_header_nritems(parent);
 671         blocksize = fs_info->nodesize;
 672         end_slot = parent_nritems - 1;
 673
 674         if (parent_nritems <= 1)
 675                 return 0;
 676
 677         for (i = start_slot; i <= end_slot; i++) {
 678                 int close = 1;
 679
 680                 btrfs_node_key(parent, &disk_key, i);
 681                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
 682                         continue;
 683
 684                 progress_passed = 1;
 685                 blocknr = btrfs_node_blockptr(parent, i);
 686                 if (last_block == 0)
 687                         last_block = blocknr;
 688
 689                 if (i > 0) {
 690                         other = btrfs_node_blockptr(parent, i - 1);
 691                         close = close_blocks(blocknr, other, blocksize);
 692                 }
 693                 if (!close && i < end_slot) {
 694                         other = btrfs_node_blockptr(parent, i + 1);
 695                         close = close_blocks(blocknr, other, blocksize);
 696                 }
 697                 if (close) {
 698                         last_block = blocknr;
 699                         continue;
 700                 }
 701
 702                 cur = btrfs_read_node_slot(parent, i);
 703                 if (IS_ERR(cur))
 704                         return PTR_ERR(cur);
 705                 if (search_start == 0)
 706                         search_start = last_block;
 707
 708                 btrfs_tree_lock(cur);
 709                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 710                                         &cur, search_start,
 711                                         min(16 * blocksize,
 712                                             (end_slot - i) * blocksize),
 713                                         BTRFS_NESTING_COW);
 714                 if (err) {
 715                         btrfs_tree_unlock(cur);
 716                         free_extent_buffer(cur);
 717                         break;
 718                 }
 719                 search_start = cur->start;
 720                 last_block = cur->start;
 721                 *last_ret = search_start;
 722                 btrfs_tree_unlock(cur);
 723                 free_extent_buffer(cur);
 724         }
 725         return err;
 726 }
 727
 728 /*
 729  * search for key in the extent_buffer.  The items start at offset p,
 730  * and they are item_size apart.
 731  *
 732  * the slot in the array is returned via slot, and it points to
 733  * the place where you would insert key if it is not found in
 734  * the array.
 735  *
 736  * Slot may point to total number of items if the key is bigger than
 737  * all of the keys
 738  */
 739 static noinline int generic_bin_search(struct extent_buffer *eb,
 740                                        unsigned long p, int item_size,
 741                                        const struct btrfs_key *key, int *slot)
 742 {
 743         int low = 0;
 744         int high = btrfs_header_nritems(eb);
 745         int ret;
 746         const int key_size = sizeof(struct btrfs_disk_key);
 747
 748         if (low > high) {
 749                 btrfs_err(eb->fs_info,
 750                  "%s: low (%d) > high (%d) eb %llu owner %llu level %d",
 751                           __func__, low, high, eb->start,
 752                           btrfs_header_owner(eb), btrfs_header_level(eb));
 753                 return -EINVAL;
 754         }
 755
 756         while (low < high) {
 757                 unsigned long oip;
 758                 unsigned long offset;
 759                 struct btrfs_disk_key *tmp;
 760                 struct btrfs_disk_key unaligned;
 761                 int mid;
 762
 763                 mid = (low + high) / 2;
 764                 offset = p + mid * item_size;
 765                 oip = offset_in_page(offset);
 766
 767                 if (oip + key_size <= PAGE_SIZE) {
 768                         const unsigned long idx = get_eb_page_index(offset);
 769                         char *kaddr = page_address(eb->pages[idx]);
 770
 771                         oip = get_eb_offset_in_page(eb, offset);
 772                         tmp = (struct btrfs_disk_key *)(kaddr + oip);
 773                 } else {
 774                         read_extent_buffer(eb, &unaligned, offset, key_size);
 775                         tmp = &unaligned;
 776                 }
 777
 778                 ret = comp_keys(tmp, key);
 779
 780                 if (ret < 0)
 781                         low = mid + 1;
 782                 else if (ret > 0)
 783                         high = mid;
 784                 else {
 785                         *slot = mid;
 786                         return 0;
 787                 }
 788         }
 789         *slot = low;
 790         return 1;
 791 }
 792
 793 /*
 794  * simple bin_search frontend that does the right thing for
 795  * leaves vs nodes
 796  */
 797 int btrfs_bin_search(struct extent_buffer *eb, const struct btrfs_key *key,
 798                      int *slot)
 799 {
 800         if (btrfs_header_level(eb) == 0)
 801                 return generic_bin_search(eb,
 802                                           offsetof(struct btrfs_leaf, items),
 803                                           sizeof(struct btrfs_item), key, slot);
 804         else
 805                 return generic_bin_search(eb,
 806                                           offsetof(struct btrfs_node, ptrs),
 807                                           sizeof(struct btrfs_key_ptr), key, slot);
 808 }
 809
 810 static void root_add_used(struct btrfs_root *root, u32 size)
 811 {
 812         spin_lock(&root->accounting_lock);
 813         btrfs_set_root_used(&root->root_item,
 814                             btrfs_root_used(&root->root_item) + size);
 815         spin_unlock(&root->accounting_lock);
 816 }
 817
 818 static void root_sub_used(struct btrfs_root *root, u32 size)
 819 {
 820         spin_lock(&root->accounting_lock);
 821         btrfs_set_root_used(&root->root_item,
 822                             btrfs_root_used(&root->root_item) - size);
 823         spin_unlock(&root->accounting_lock);
 824 }
 825
 826 /* given a node and slot number, this reads the blocks it points to.  The
 827  * extent buffer is returned with a reference taken (but unlocked).
 828  */
 829 struct extent_buffer *btrfs_read_node_slot(struct extent_buffer *parent,
 830                                            int slot)
 831 {
 832         int level = btrfs_header_level(parent);
 833         struct extent_buffer *eb;
 834         struct btrfs_key first_key;
 835
 836         if (slot < 0 || slot >= btrfs_header_nritems(parent))
 837                 return ERR_PTR(-ENOENT);
 838
 839         BUG_ON(level == 0);
 840
 841         btrfs_node_key_to_cpu(parent, &first_key, slot);
 842         eb = read_tree_block(parent->fs_info, btrfs_node_blockptr(parent, slot),
 843                              btrfs_header_owner(parent),
 844                              btrfs_node_ptr_generation(parent, slot),
 845                              level - 1, &first_key);
 846         if (!IS_ERR(eb) && !extent_buffer_uptodate(eb)) {
 847                 free_extent_buffer(eb);
 848                 eb = ERR_PTR(-EIO);
 849         }
 850
 851         return eb;
 852 }
 853
 854 /*
 855  * node level balancing, used to make sure nodes are in proper order for
 856  * item deletion.  We balance from the top down, so we have to make sure
 857  * that a deletion won't leave an node completely empty later on.
 858  */
 859 static noinline int balance_level(struct btrfs_trans_handle *trans,
 860                          struct btrfs_root *root,
 861                          struct btrfs_path *path, int level)
 862 {
 863         struct btrfs_fs_info *fs_info = root->fs_info;
 864         struct extent_buffer *right = NULL;
 865         struct extent_buffer *mid;
 866         struct extent_buffer *left = NULL;
 867         struct extent_buffer *parent = NULL;
 868         int ret = 0;
 869         int wret;
 870         int pslot;
 871         int orig_slot = path->slots[level];
 872         u64 orig_ptr;
 873
 874         ASSERT(level > 0);
 875
 876         mid = path->nodes[level];
 877
 878         WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK);
 879         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 880
 881         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 882
 883         if (level < BTRFS_MAX_LEVEL - 1) {
 884                 parent = path->nodes[level + 1];
 885                 pslot = path->slots[level + 1];
 886         }
 887
 888         /*
 889          * deal with the case where there is only one pointer in the root
 890          * by promoting the node below to a root
 891          */
 892         if (!parent) {
 893                 struct extent_buffer *child;
 894
 895                 if (btrfs_header_nritems(mid) != 1)
 896                         return 0;
 897
 898                 /* promote the child to a root */
 899                 child = btrfs_read_node_slot(mid, 0);
 900                 if (IS_ERR(child)) {
 901                         ret = PTR_ERR(child);
 902                         btrfs_handle_fs_error(fs_info, ret, NULL);
 903                         goto enospc;
 904                 }
 905
 906                 btrfs_tree_lock(child);
 907                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child,
 908                                       BTRFS_NESTING_COW);
 909                 if (ret) {
 910                         btrfs_tree_unlock(child);
 911                         free_extent_buffer(child);
 912                         goto enospc;
 913                 }
 914
 915                 ret = btrfs_tree_mod_log_insert_root(root->node, child, true);
 916                 BUG_ON(ret < 0);
 917                 rcu_assign_pointer(root->node, child);
 918
 919                 add_root_to_dirty_list(root);
 920                 btrfs_tree_unlock(child);
 921
 922                 path->locks[level] = 0;
 923                 path->nodes[level] = NULL;
 924                 btrfs_clean_tree_block(mid);
 925                 btrfs_tree_unlock(mid);
 926                 /* once for the path */
 927                 free_extent_buffer(mid);
 928
 929                 root_sub_used(root, mid->len);
 930                 btrfs_free_tree_block(trans, btrfs_root_id(root), mid, 0, 1);
 931                 /* once for the root ptr */
 932                 free_extent_buffer_stale(mid);
 933                 return 0;
 934         }
 935         if (btrfs_header_nritems(mid) >
 936             BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 4)
 937                 return 0;
 938
 939         left = btrfs_read_node_slot(parent, pslot - 1);
 940         if (IS_ERR(left))
 941                 left = NULL;
 942
 943         if (left) {
 944                 __btrfs_tree_lock(left, BTRFS_NESTING_LEFT);
 945                 wret = btrfs_cow_block(trans, root, left,
 946                                        parent, pslot - 1, &left,
 947                                        BTRFS_NESTING_LEFT_COW);
 948                 if (wret) {
 949                         ret = wret;
 950                         goto enospc;
 951                 }
 952         }
 953
 954         right = btrfs_read_node_slot(parent, pslot + 1);
 955         if (IS_ERR(right))
 956                 right = NULL;
 957
 958         if (right) {
 959                 __btrfs_tree_lock(right, BTRFS_NESTING_RIGHT);
 960                 wret = btrfs_cow_block(trans, root, right,
 961                                        parent, pslot + 1, &right,
 962                                        BTRFS_NESTING_RIGHT_COW);
 963                 if (wret) {
 964                         ret = wret;
 965                         goto enospc;
 966                 }
 967         }
 968
 969         /* first, try to make some room in the middle buffer */
 970         if (left) {
 971                 orig_slot += btrfs_header_nritems(left);
 972                 wret = push_node_left(trans, left, mid, 1);
 973                 if (wret < 0)
 974                         ret = wret;
 975         }
 976
 977         /*
 978          * then try to empty the right most buffer into the middle
 979          */
 980         if (right) {
 981                 wret = push_node_left(trans, mid, right, 1);
 982                 if (wret < 0 && wret != -ENOSPC)
 983                         ret = wret;
 984                 if (btrfs_header_nritems(right) == 0) {
 985                         btrfs_clean_tree_block(right);
 986                         btrfs_tree_unlock(right);
 987                         del_ptr(root, path, level + 1, pslot + 1);
 988                         root_sub_used(root, right->len);
 989                         btrfs_free_tree_block(trans, btrfs_root_id(root), right,
 990                                               0, 1);
 991                         free_extent_buffer_stale(right);
 992                         right = NULL;
 993                 } else {
 994                         struct btrfs_disk_key right_key;
 995                         btrfs_node_key(right, &right_key, 0);
 996                         ret = btrfs_tree_mod_log_insert_key(parent, pslot + 1,
 997                                         BTRFS_MOD_LOG_KEY_REPLACE, GFP_NOFS);
 998                         BUG_ON(ret < 0);
 999                         btrfs_set_node_key(parent, &right_key, pslot + 1);
1000                         btrfs_mark_buffer_dirty(parent);
1001                 }
1002         }
1003         if (btrfs_header_nritems(mid) == 1) {
1004                 /*
1005                  * we're not allowed to leave a node with one item in the
1006                  * tree during a delete.  A deletion from lower in the tree
1007                  * could try to delete the only pointer in this node.
1008                  * So, pull some keys from the left.
1009                  * There has to be a left pointer at this point because
1010                  * otherwise we would have pulled some pointers from the
1011                  * right
1012                  */
1013                 if (!left) {
1014                         ret = -EROFS;
1015                         btrfs_handle_fs_error(fs_info, ret, NULL);
1016                         goto enospc;
1017                 }
1018                 wret = balance_node_right(trans, mid, left);
1019                 if (wret < 0) {
1020                         ret = wret;
1021                         goto enospc;
1022                 }
1023                 if (wret == 1) {
1024                         wret = push_node_left(trans, left, mid, 1);
1025                         if (wret < 0)
1026                                 ret = wret;
1027                 }
1028                 BUG_ON(wret == 1);
1029         }
1030         if (btrfs_header_nritems(mid) == 0) {
1031                 btrfs_clean_tree_block(mid);
1032                 btrfs_tree_unlock(mid);
1033                 del_ptr(root, path, level + 1, pslot);
1034                 root_sub_used(root, mid->len);
1035                 btrfs_free_tree_block(trans, btrfs_root_id(root), mid, 0, 1);
1036                 free_extent_buffer_stale(mid);
1037                 mid = NULL;
1038         } else {
1039                 /* update the parent key to reflect our changes */
1040                 struct btrfs_disk_key mid_key;
1041                 btrfs_node_key(mid, &mid_key, 0);
1042                 ret = btrfs_tree_mod_log_insert_key(parent, pslot,
1043                                 BTRFS_MOD_LOG_KEY_REPLACE, GFP_NOFS);
1044                 BUG_ON(ret < 0);
1045                 btrfs_set_node_key(parent, &mid_key, pslot);
1046                 btrfs_mark_buffer_dirty(parent);
1047         }
1048
1049         /* update the path */
1050         if (left) {
1051                 if (btrfs_header_nritems(left) > orig_slot) {
1052                         atomic_inc(&left->refs);
1053                         /* left was locked after cow */
1054                         path->nodes[level] = left;
1055                         path->slots[level + 1] -= 1;
1056                         path->slots[level] = orig_slot;
1057                         if (mid) {
1058                                 btrfs_tree_unlock(mid);
1059                                 free_extent_buffer(mid);
1060                         }
1061                 } else {
1062                         orig_slot -= btrfs_header_nritems(left);
1063                         path->slots[level] = orig_slot;
1064                 }
1065         }
1066         /* double check we haven't messed things up */
1067         if (orig_ptr !=
1068             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
1069                 BUG();
1070 enospc:
1071         if (right) {
1072                 btrfs_tree_unlock(right);
1073                 free_extent_buffer(right);
1074         }
1075         if (left) {
1076                 if (path->nodes[level] != left)
1077                         btrfs_tree_unlock(left);
1078                 free_extent_buffer(left);
1079         }
1080         return ret;
1081 }
1082
1083 /* Node balancing for insertion.  Here we only split or push nodes around
1084  * when they are completely full.  This is also done top down, so we
1085  * have to be pessimistic.
1086  */
1087 static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
1088                                           struct btrfs_root *root,
1089                                           struct btrfs_path *path, int level)
1090 {
1091         struct btrfs_fs_info *fs_info = root->fs_info;
1092         struct extent_buffer *right = NULL;
1093         struct extent_buffer *mid;
1094         struct extent_buffer *left = NULL;
1095         struct extent_buffer *parent = NULL;
1096         int ret = 0;
1097         int wret;
1098         int pslot;
1099         int orig_slot = path->slots[level];
1100
1101         if (level == 0)
1102                 return 1;
1103
1104         mid = path->nodes[level];
1105         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1106
1107         if (level < BTRFS_MAX_LEVEL - 1) {
1108                 parent = path->nodes[level + 1];
1109                 pslot = path->slots[level + 1];
1110         }
1111
1112         if (!parent)
1113                 return 1;
1114
1115         left = btrfs_read_node_slot(parent, pslot - 1);
1116         if (IS_ERR(left))
1117                 left = NULL;
1118
1119         /* first, try to make some room in the middle buffer */
1120         if (left) {
1121                 u32 left_nr;
1122
1123                 __btrfs_tree_lock(left, BTRFS_NESTING_LEFT);
1124
1125                 left_nr = btrfs_header_nritems(left);
1126                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
1127                         wret = 1;
1128                 } else {
1129                         ret = btrfs_cow_block(trans, root, left, parent,
1130                                               pslot - 1, &left,
1131                                               BTRFS_NESTING_LEFT_COW);
1132                         if (ret)
1133                                 wret = 1;
1134                         else {
1135                                 wret = push_node_left(trans, left, mid, 0);
1136                         }
1137                 }
1138                 if (wret < 0)
1139                         ret = wret;
1140                 if (wret == 0) {
1141                         struct btrfs_disk_key disk_key;
1142                         orig_slot += left_nr;
1143                         btrfs_node_key(mid, &disk_key, 0);
1144                         ret = btrfs_tree_mod_log_insert_key(parent, pslot,
1145                                         BTRFS_MOD_LOG_KEY_REPLACE, GFP_NOFS);
1146                         BUG_ON(ret < 0);
1147                         btrfs_set_node_key(parent, &disk_key, pslot);
1148                         btrfs_mark_buffer_dirty(parent);
1149                         if (btrfs_header_nritems(left) > orig_slot) {
1150                                 path->nodes[level] = left;
1151                                 path->slots[level + 1] -= 1;
1152                                 path->slots[level] = orig_slot;
1153                                 btrfs_tree_unlock(mid);
1154                                 free_extent_buffer(mid);
1155                         } else {
1156                                 orig_slot -=
1157                                         btrfs_header_nritems(left);
1158                                 path->slots[level] = orig_slot;
1159                                 btrfs_tree_unlock(left);
1160                                 free_extent_buffer(left);
1161                         }
1162                         return 0;
1163                 }
1164                 btrfs_tree_unlock(left);
1165                 free_extent_buffer(left);
1166         }
1167         right = btrfs_read_node_slot(parent, pslot + 1);
1168         if (IS_ERR(right))
1169                 right = NULL;
1170
1171         /*
1172          * then try to empty the right most buffer into the middle
1173          */
1174         if (right) {
1175                 u32 right_nr;
1176
1177                 __btrfs_tree_lock(right, BTRFS_NESTING_RIGHT);
1178
1179                 right_nr = btrfs_header_nritems(right);
1180                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
1181                         wret = 1;
1182                 } else {
1183                         ret = btrfs_cow_block(trans, root, right,
1184                                               parent, pslot + 1,
1185                                               &right, BTRFS_NESTING_RIGHT_COW);
1186                         if (ret)
1187                                 wret = 1;
1188                         else {
1189                                 wret = balance_node_right(trans, right, mid);
1190                         }
1191                 }
1192                 if (wret < 0)
1193                         ret = wret;
1194                 if (wret == 0) {
1195                         struct btrfs_disk_key disk_key;
1196
1197                         btrfs_node_key(right, &disk_key, 0);
1198                         ret = btrfs_tree_mod_log_insert_key(parent, pslot + 1,
1199                                         BTRFS_MOD_LOG_KEY_REPLACE, GFP_NOFS);
1200                         BUG_ON(ret < 0);
1201                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
1202                         btrfs_mark_buffer_dirty(parent);
1203
1204                         if (btrfs_header_nritems(mid) <= orig_slot) {
1205                                 path->nodes[level] = right;
1206                                 path->slots[level + 1] += 1;
1207                                 path->slots[level] = orig_slot -
1208                                         btrfs_header_nritems(mid);
1209                                 btrfs_tree_unlock(mid);
1210                                 free_extent_buffer(mid);
1211                         } else {
1212                                 btrfs_tree_unlock(right);
1213                                 free_extent_buffer(right);
1214                         }
1215                         return 0;
1216                 }
1217                 btrfs_tree_unlock(right);
1218                 free_extent_buffer(right);
1219         }
1220         return 1;
1221 }
1222
1223 /*
1224  * readahead one full node of leaves, finding things that are close
1225  * to the block in 'slot', and triggering ra on them.
1226  */
1227 static void reada_for_search(struct btrfs_fs_info *fs_info,
1228                              struct btrfs_path *path,
1229                              int level, int slot, u64 objectid)
1230 {
1231         struct extent_buffer *node;
1232         struct btrfs_disk_key disk_key;
1233         u32 nritems;
1234         u64 search;
1235         u64 target;
1236         u64 nread = 0;
1237         u64 nread_max;
1238         u32 nr;
1239         u32 blocksize;
1240         u32 nscan = 0;
1241
1242         if (level != 1 && path->reada != READA_FORWARD_ALWAYS)
1243                 return;
1244
1245         if (!path->nodes[level])
1246                 return;
1247
1248         node = path->nodes[level];
1249
1250         /*
1251          * Since the time between visiting leaves is much shorter than the time
1252          * between visiting nodes, limit read ahead of nodes to 1, to avoid too
1253          * much IO at once (possibly random).
1254          */
1255         if (path->reada == READA_FORWARD_ALWAYS) {
1256                 if (level > 1)
1257                         nread_max = node->fs_info->nodesize;
1258                 else
1259                         nread_max = SZ_128K;
1260         } else {
1261                 nread_max = SZ_64K;
1262         }
1263
1264         search = btrfs_node_blockptr(node, slot);
1265         blocksize = fs_info->nodesize;
1266         if (path->reada != READA_FORWARD_ALWAYS) {
1267                 struct extent_buffer *eb;
1268
1269                 eb = find_extent_buffer(fs_info, search);
1270                 if (eb) {
1271                         free_extent_buffer(eb);
1272                         return;
1273                 }
1274         }
1275
1276         target = search;
1277
1278         nritems = btrfs_header_nritems(node);
1279         nr = slot;
1280
1281         while (1) {
1282                 if (path->reada == READA_BACK) {
1283                         if (nr == 0)
1284                                 break;
1285                         nr--;
1286                 } else if (path->reada == READA_FORWARD ||
1287                            path->reada == READA_FORWARD_ALWAYS) {
1288                         nr++;
1289                         if (nr >= nritems)
1290                                 break;
1291                 }
1292                 if (path->reada == READA_BACK && objectid) {
1293                         btrfs_node_key(node, &disk_key, nr);
1294                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1295                                 break;
1296                 }
1297                 search = btrfs_node_blockptr(node, nr);
1298                 if (path->reada == READA_FORWARD_ALWAYS ||
1299                     (search <= target && target - search <= 65536) ||
1300                     (search > target && search - target <= 65536)) {
1301                         btrfs_readahead_node_child(node, nr);
1302                         nread += blocksize;
1303                 }
1304                 nscan++;
1305                 if (nread > nread_max || nscan > 32)
1306                         break;
1307         }
1308 }
1309
1310 static noinline void reada_for_balance(struct btrfs_path *path, int level)
1311 {
1312         struct extent_buffer *parent;
1313         int slot;
1314         int nritems;
1315
1316         parent = path->nodes[level + 1];
1317         if (!parent)
1318                 return;
1319
1320         nritems = btrfs_header_nritems(parent);
1321         slot = path->slots[level + 1];
1322
1323         if (slot > 0)
1324                 btrfs_readahead_node_child(parent, slot - 1);
1325         if (slot + 1 < nritems)
1326                 btrfs_readahead_node_child(parent, slot + 1);
1327 }
1328
1329
1330 /*
1331  * when we walk down the tree, it is usually safe to unlock the higher layers
1332  * in the tree.  The exceptions are when our path goes through slot 0, because
1333  * operations on the tree might require changing key pointers higher up in the
1334  * tree.
1335  *
1336  * callers might also have set path->keep_locks, which tells this code to keep
1337  * the lock if the path points to the last slot in the block.  This is part of
1338  * walking through the tree, and selecting the next slot in the higher block.
1339  *
1340  * lowest_unlock sets the lowest level in the tree we're allowed to unlock.  so
1341  * if lowest_unlock is 1, level 0 won't be unlocked
1342  */
1343 static noinline void unlock_up(struct btrfs_path *path, int level,
1344                                int lowest_unlock, int min_write_lock_level,
1345                                int *write_lock_level)
1346 {
1347         int i;
1348         int skip_level = level;
1349         int no_skips = 0;
1350         struct extent_buffer *t;
1351
1352         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1353                 if (!path->nodes[i])
1354                         break;
1355                 if (!path->locks[i])
1356                         break;
1357                 if (!no_skips && path->slots[i] == 0) {
1358                         skip_level = i + 1;
1359                         continue;
1360                 }
1361                 if (!no_skips && path->keep_locks) {
1362                         u32 nritems;
1363                         t = path->nodes[i];
1364                         nritems = btrfs_header_nritems(t);
1365                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
1366                                 skip_level = i + 1;
1367                                 continue;
1368                         }
1369                 }
1370                 if (skip_level < i && i >= lowest_unlock)
1371                         no_skips = 1;
1372
1373                 t = path->nodes[i];
1374                 if (i >= lowest_unlock && i > skip_level) {
1375                         btrfs_tree_unlock_rw(t, path->locks[i]);
1376                         path->locks[i] = 0;
1377                         if (write_lock_level &&
1378                             i > min_write_lock_level &&
1379                             i <= *write_lock_level) {
1380                                 *write_lock_level = i - 1;
1381                         }
1382                 }
1383         }
1384 }
1385
1386 /*
1387  * helper function for btrfs_search_slot.  The goal is to find a block
1388  * in cache without setting the path to blocking.  If we find the block
1389  * we return zero and the path is unchanged.
1390  *
1391  * If we can't find the block, we set the path blocking and do some
1392  * reada.  -EAGAIN is returned and the search must be repeated.
1393  */
1394 static int
1395 read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
1396                       struct extent_buffer **eb_ret, int level, int slot,
1397                       const struct btrfs_key *key)
1398 {
1399         struct btrfs_fs_info *fs_info = root->fs_info;
1400         u64 blocknr;
1401         u64 gen;
1402         struct extent_buffer *tmp;
1403         struct btrfs_key first_key;
1404         int ret;
1405         int parent_level;
1406
1407         blocknr = btrfs_node_blockptr(*eb_ret, slot);
1408         gen = btrfs_node_ptr_generation(*eb_ret, slot);
1409         parent_level = btrfs_header_level(*eb_ret);
1410         btrfs_node_key_to_cpu(*eb_ret, &first_key, slot);
1411
1412         tmp = find_extent_buffer(fs_info, blocknr);
1413         if (tmp) {
1414                 if (p->reada == READA_FORWARD_ALWAYS)
1415                         reada_for_search(fs_info, p, level, slot, key->objectid);
1416
1417                 /* first we do an atomic uptodate check */
1418                 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
1419                         /*
1420                          * Do extra check for first_key, eb can be stale due to
1421                          * being cached, read from scrub, or have multiple
1422                          * parents (shared tree blocks).
1423                          */
1424                         if (btrfs_verify_level_key(tmp,
1425                                         parent_level - 1, &first_key, gen)) {
1426                                 free_extent_buffer(tmp);
1427                                 return -EUCLEAN;
1428                         }
1429                         *eb_ret = tmp;
1430                         return 0;
1431                 }
1432
1433                 /* now we're allowed to do a blocking uptodate check */
1434                 ret = btrfs_read_buffer(tmp, gen, parent_level - 1, &first_key);
1435                 if (!ret) {
1436                         *eb_ret = tmp;
1437                         return 0;
1438                 }
1439                 free_extent_buffer(tmp);
1440                 btrfs_release_path(p);
1441                 return -EIO;
1442         }
1443
1444         /*
1445          * reduce lock contention at high levels
1446          * of the btree by dropping locks before
1447          * we read.  Don't release the lock on the current
1448          * level because we need to walk this node to figure
1449          * out which blocks to read.
1450          */
1451         btrfs_unlock_up_safe(p, level + 1);
1452
1453         if (p->reada != READA_NONE)
1454                 reada_for_search(fs_info, p, level, slot, key->objectid);
1455
1456         ret = -EAGAIN;
1457         tmp = read_tree_block(fs_info, blocknr, root->root_key.objectid,
1458                               gen, parent_level - 1, &first_key);
1459         if (!IS_ERR(tmp)) {
1460                 /*
1461                  * If the read above didn't mark this buffer up to date,
1462                  * it will never end up being up to date.  Set ret to EIO now
1463                  * and give up so that our caller doesn't loop forever
1464                  * on our EAGAINs.
1465                  */
1466                 if (!extent_buffer_uptodate(tmp))
1467                         ret = -EIO;
1468                 free_extent_buffer(tmp);
1469         } else {
1470                 ret = PTR_ERR(tmp);
1471         }
1472
1473         btrfs_release_path(p);
1474         return ret;
1475 }
1476
1477 /*
1478  * helper function for btrfs_search_slot.  This does all of the checks
1479  * for node-level blocks and does any balancing required based on
1480  * the ins_len.
1481  *
1482  * If no extra work was required, zero is returned.  If we had to
1483  * drop the path, -EAGAIN is returned and btrfs_search_slot must
1484  * start over
1485  */
1486 static int
1487 setup_nodes_for_search(struct btrfs_trans_handle *trans,
1488                        struct btrfs_root *root, struct btrfs_path *p,
1489                        struct extent_buffer *b, int level, int ins_len,
1490                        int *write_lock_level)
1491 {
1492         struct btrfs_fs_info *fs_info = root->fs_info;
1493         int ret = 0;
1494
1495         if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
1496             BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3) {
1497
1498                 if (*write_lock_level < level + 1) {
1499                         *write_lock_level = level + 1;
1500                         btrfs_release_path(p);
1501                         return -EAGAIN;
1502                 }
1503
1504                 reada_for_balance(p, level);
1505                 ret = split_node(trans, root, p, level);
1506
1507                 b = p->nodes[level];
1508         } else if (ins_len < 0 && btrfs_header_nritems(b) <
1509                    BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 2) {
1510
1511                 if (*write_lock_level < level + 1) {
1512                         *write_lock_level = level + 1;
1513                         btrfs_release_path(p);
1514                         return -EAGAIN;
1515                 }
1516
1517                 reada_for_balance(p, level);
1518                 ret = balance_level(trans, root, p, level);
1519                 if (ret)
1520                         return ret;
1521
1522                 b = p->nodes[level];
1523                 if (!b) {
1524                         btrfs_release_path(p);
1525                         return -EAGAIN;
1526                 }
1527                 BUG_ON(btrfs_header_nritems(b) == 1);
1528         }
1529         return ret;
1530 }
1531
1532 int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
1533                 u64 iobjectid, u64 ioff, u8 key_type,
1534                 struct btrfs_key *found_key)
1535 {
1536         int ret;
1537         struct btrfs_key key;
1538         struct extent_buffer *eb;
1539
1540         ASSERT(path);
1541         ASSERT(found_key);
1542
1543         key.type = key_type;
1544         key.objectid = iobjectid;
1545         key.offset = ioff;
1546
1547         ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
1548         if (ret < 0)
1549                 return ret;
1550
1551         eb = path->nodes[0];
1552         if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
1553                 ret = btrfs_next_leaf(fs_root, path);
1554                 if (ret)
1555                         return ret;
1556                 eb = path->nodes[0];
1557         }
1558
1559         btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
1560         if (found_key->type != key.type ||
1561                         found_key->objectid != key.objectid)
1562                 return 1;
1563
1564         return 0;
1565 }
1566
1567 static struct extent_buffer *btrfs_search_slot_get_root(struct btrfs_root *root,
1568                                                         struct btrfs_path *p,
1569                                                         int write_lock_level)
1570 {
1571         struct btrfs_fs_info *fs_info = root->fs_info;
1572         struct extent_buffer *b;
1573         int root_lock;
1574         int level = 0;
1575
1576         /* We try very hard to do read locks on the root */
1577         root_lock = BTRFS_READ_LOCK;
1578
1579         if (p->search_commit_root) {
1580                 /*
1581                  * The commit roots are read only so we always do read locks,
1582                  * and we always must hold the commit_root_sem when doing
1583                  * searches on them, the only exception is send where we don't
1584                  * want to block transaction commits for a long time, so
1585                  * we need to clone the commit root in order to avoid races
1586                  * with transaction commits that create a snapshot of one of
1587                  * the roots used by a send operation.
1588                  */
1589                 if (p->need_commit_sem) {
1590                         down_read(&fs_info->commit_root_sem);
1591                         b = btrfs_clone_extent_buffer(root->commit_root);
1592                         up_read(&fs_info->commit_root_sem);
1593                         if (!b)
1594                                 return ERR_PTR(-ENOMEM);
1595
1596                 } else {
1597                         b = root->commit_root;
1598                         atomic_inc(&b->refs);
1599                 }
1600                 level = btrfs_header_level(b);
1601                 /*
1602                  * Ensure that all callers have set skip_locking when
1603                  * p->search_commit_root = 1.
1604                  */
1605                 ASSERT(p->skip_locking == 1);
1606
1607                 goto out;
1608         }
1609
1610         if (p->skip_locking) {
1611                 b = btrfs_root_node(root);
1612                 level = btrfs_header_level(b);
1613                 goto out;
1614         }
1615
1616         /*
1617          * If the level is set to maximum, we can skip trying to get the read
1618          * lock.
1619          */
1620         if (write_lock_level < BTRFS_MAX_LEVEL) {
1621                 /*
1622                  * We don't know the level of the root node until we actually
1623                  * have it read locked
1624                  */
1625                 b = btrfs_read_lock_root_node(root);
1626                 level = btrfs_header_level(b);
1627                 if (level > write_lock_level)
1628                         goto out;
1629
1630                 /* Whoops, must trade for write lock */
1631                 btrfs_tree_read_unlock(b);
1632                 free_extent_buffer(b);
1633         }
1634
1635         b = btrfs_lock_root_node(root);
1636         root_lock = BTRFS_WRITE_LOCK;
1637
1638         /* The level might have changed, check again */
1639         level = btrfs_header_level(b);
1640
1641 out:
1642         p->nodes[level] = b;
1643         if (!p->skip_locking)
1644                 p->locks[level] = root_lock;
1645         /*
1646          * Callers are responsible for dropping b's references.
1647          */
1648         return b;
1649 }
1650
1651
1652 /*
1653  * btrfs_search_slot - look for a key in a tree and perform necessary
1654  * modifications to preserve tree invariants.
1655  *
1656  * @trans:      Handle of transaction, used when modifying the tree
1657  * @p:          Holds all btree nodes along the search path
1658  * @root:       The root node of the tree
1659  * @key:        The key we are looking for
1660  * @ins_len:    Indicates purpose of search:
1661  *              >0  for inserts it's size of item inserted (*)
1662  *              <0  for deletions
1663  *               0  for plain searches, not modifying the tree
1664  *
1665  *              (*) If size of item inserted doesn't include
1666  *              sizeof(struct btrfs_item), then p->search_for_extension must
1667  *              be set.
1668  * @cow:        boolean should CoW operations be performed. Must always be 1
1669  *              when modifying the tree.
1670  *
1671  * If @ins_len > 0, nodes and leaves will be split as we walk down the tree.
1672  * If @ins_len < 0, nodes will be merged as we walk down the tree (if possible)
1673  *
1674  * If @key is found, 0 is returned and you can find the item in the leaf level
1675  * of the path (level 0)
1676  *
1677  * If @key isn't found, 1 is returned and the leaf level of the path (level 0)
1678  * points to the slot where it should be inserted
1679  *
1680  * If an error is encountered while searching the tree a negative error number
1681  * is returned
1682  */
1683 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1684                       const struct btrfs_key *key, struct btrfs_path *p,
1685                       int ins_len, int cow)
1686 {
1687         struct extent_buffer *b;
1688         int slot;
1689         int ret;
1690         int err;
1691         int level;
1692         int lowest_unlock = 1;
1693         /* everything at write_lock_level or lower must be write locked */
1694         int write_lock_level = 0;
1695         u8 lowest_level = 0;
1696         int min_write_lock_level;
1697         int prev_cmp;
1698
1699         lowest_level = p->lowest_level;
1700         WARN_ON(lowest_level && ins_len > 0);
1701         WARN_ON(p->nodes[0] != NULL);
1702         BUG_ON(!cow && ins_len);
1703
1704         if (ins_len < 0) {
1705                 lowest_unlock = 2;
1706
1707                 /* when we are removing items, we might have to go up to level
1708                  * two as we update tree pointers  Make sure we keep write
1709                  * for those levels as well
1710                  */
1711                 write_lock_level = 2;
1712         } else if (ins_len > 0) {
1713                 /*
1714                  * for inserting items, make sure we have a write lock on
1715                  * level 1 so we can update keys
1716                  */
1717                 write_lock_level = 1;
1718         }
1719
1720         if (!cow)
1721                 write_lock_level = -1;
1722
1723         if (cow && (p->keep_locks || p->lowest_level))
1724                 write_lock_level = BTRFS_MAX_LEVEL;
1725
1726         min_write_lock_level = write_lock_level;
1727
1728 again:
1729         prev_cmp = -1;
1730         b = btrfs_search_slot_get_root(root, p, write_lock_level);
1731         if (IS_ERR(b)) {
1732                 ret = PTR_ERR(b);
1733                 goto done;
1734         }
1735
1736         while (b) {
1737                 int dec = 0;
1738
1739                 level = btrfs_header_level(b);
1740
1741                 if (cow) {
1742                         bool last_level = (level == (BTRFS_MAX_LEVEL - 1));
1743
1744                         /*
1745                          * if we don't really need to cow this block
1746                          * then we don't want to set the path blocking,
1747                          * so we test it here
1748                          */
1749                         if (!should_cow_block(trans, root, b))
1750                                 goto cow_done;
1751
1752                         /*
1753                          * must have write locks on this node and the
1754                          * parent
1755                          */
1756                         if (level > write_lock_level ||
1757                             (level + 1 > write_lock_level &&
1758                             level + 1 < BTRFS_MAX_LEVEL &&
1759                             p->nodes[level + 1])) {
1760                                 write_lock_level = level + 1;
1761                                 btrfs_release_path(p);
1762                                 goto again;
1763                         }
1764
1765                         if (last_level)
1766                                 err = btrfs_cow_block(trans, root, b, NULL, 0,
1767                                                       &b,
1768                                                       BTRFS_NESTING_COW);
1769                         else
1770                                 err = btrfs_cow_block(trans, root, b,
1771                                                       p->nodes[level + 1],
1772                                                       p->slots[level + 1], &b,
1773                                                       BTRFS_NESTING_COW);
1774                         if (err) {
1775                                 ret = err;
1776                                 goto done;
1777                         }
1778                 }
1779 cow_done:
1780                 p->nodes[level] = b;
1781                 /*
1782                  * Leave path with blocking locks to avoid massive
1783                  * lock context switch, this is made on purpose.
1784                  */
1785
1786                 /*
1787                  * we have a lock on b and as long as we aren't changing
1788                  * the tree, there is no way to for the items in b to change.
1789                  * It is safe to drop the lock on our parent before we
1790                  * go through the expensive btree search on b.
1791                  *
1792                  * If we're inserting or deleting (ins_len != 0), then we might
1793                  * be changing slot zero, which may require changing the parent.
1794                  * So, we can't drop the lock until after we know which slot
1795                  * we're operating on.
1796                  */
1797                 if (!ins_len && !p->keep_locks) {
1798                         int u = level + 1;
1799
1800                         if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
1801                                 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
1802                                 p->locks[u] = 0;
1803                         }
1804                 }
1805
1806                 /*
1807                  * If btrfs_bin_search returns an exact match (prev_cmp == 0)
1808                  * we can safely assume the target key will always be in slot 0
1809                  * on lower levels due to the invariants BTRFS' btree provides,
1810                  * namely that a btrfs_key_ptr entry always points to the
1811                  * lowest key in the child node, thus we can skip searching
1812                  * lower levels
1813                  */
1814                 if (prev_cmp == 0) {
1815                         slot = 0;
1816                         ret = 0;
1817                 } else {
1818                         ret = btrfs_bin_search(b, key, &slot);
1819                         prev_cmp = ret;
1820                         if (ret < 0)
1821                                 goto done;
1822                 }
1823
1824                 if (level == 0) {
1825                         p->slots[level] = slot;
1826                         /*
1827                          * Item key already exists. In this case, if we are
1828                          * allowed to insert the item (for example, in dir_item
1829                          * case, item key collision is allowed), it will be
1830                          * merged with the original item. Only the item size
1831                          * grows, no new btrfs item will be added. If
1832                          * search_for_extension is not set, ins_len already
1833                          * accounts the size btrfs_item, deduct it here so leaf
1834                          * space check will be correct.
1835                          */
1836                         if (ret == 0 && ins_len > 0 && !p->search_for_extension) {
1837                                 ASSERT(ins_len >= sizeof(struct btrfs_item));
1838                                 ins_len -= sizeof(struct btrfs_item);
1839                         }
1840                         if (ins_len > 0 &&
1841                             btrfs_leaf_free_space(b) < ins_len) {
1842                                 if (write_lock_level < 1) {
1843                                         write_lock_level = 1;
1844                                         btrfs_release_path(p);
1845                                         goto again;
1846                                 }
1847
1848                                 err = split_leaf(trans, root, key,
1849                                                  p, ins_len, ret == 0);
1850
1851                                 BUG_ON(err > 0);
1852                                 if (err) {
1853                                         ret = err;
1854                                         goto done;
1855                                 }
1856                         }
1857                         if (!p->search_for_split)
1858                                 unlock_up(p, level, lowest_unlock,
1859                                           min_write_lock_level, NULL);
1860                         goto done;
1861                 }
1862                 if (ret && slot > 0) {
1863                         dec = 1;
1864                         slot--;
1865                 }
1866                 p->slots[level] = slot;
1867                 err = setup_nodes_for_search(trans, root, p, b, level, ins_len,
1868                                              &write_lock_level);
1869                 if (err == -EAGAIN)
1870                         goto again;
1871                 if (err) {
1872                         ret = err;
1873                         goto done;
1874                 }
1875                 b = p->nodes[level];
1876                 slot = p->slots[level];
1877
1878                 /*
1879                  * Slot 0 is special, if we change the key we have to update
1880                  * the parent pointer which means we must have a write lock on
1881                  * the parent
1882                  */
1883                 if (slot == 0 && ins_len && write_lock_level < level + 1) {
1884                         write_lock_level = level + 1;
1885                         btrfs_release_path(p);
1886                         goto again;
1887                 }
1888
1889                 unlock_up(p, level, lowest_unlock, min_write_lock_level,
1890                           &write_lock_level);
1891
1892                 if (level == lowest_level) {
1893                         if (dec)
1894                                 p->slots[level]++;
1895                         goto done;
1896                 }
1897
1898                 err = read_block_for_search(root, p, &b, level, slot, key);
1899                 if (err == -EAGAIN)
1900                         goto again;
1901                 if (err) {
1902                         ret = err;
1903                         goto done;
1904                 }
1905
1906                 if (!p->skip_locking) {
1907                         level = btrfs_header_level(b);
1908                         if (level <= write_lock_level) {
1909                                 btrfs_tree_lock(b);
1910                                 p->locks[level] = BTRFS_WRITE_LOCK;
1911                         } else {
1912                                 btrfs_tree_read_lock(b);
1913                                 p->locks[level] = BTRFS_READ_LOCK;
1914                         }
1915                         p->nodes[level] = b;
1916                 }
1917         }
1918         ret = 1;
1919 done:
1920         if (ret < 0 && !p->skip_release_on_error)
1921                 btrfs_release_path(p);
1922         return ret;
1923 }
1924 ALLOW_ERROR_INJECTION(btrfs_search_slot, ERRNO);
1925
1926 /*
1927  * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
1928  * current state of the tree together with the operations recorded in the tree
1929  * modification log to search for the key in a previous version of this tree, as
1930  * denoted by the time_seq parameter.
1931  *
1932  * Naturally, there is no support for insert, delete or cow operations.
1933  *
1934  * The resulting path and return value will be set up as if we called
1935  * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
1936  */
1937 int btrfs_search_old_slot(struct btrfs_root *root, const struct btrfs_key *key,
1938                           struct btrfs_path *p, u64 time_seq)
1939 {
1940         struct btrfs_fs_info *fs_info = root->fs_info;
1941         struct extent_buffer *b;
1942         int slot;
1943         int ret;
1944         int err;
1945         int level;
1946         int lowest_unlock = 1;
1947         u8 lowest_level = 0;
1948
1949         lowest_level = p->lowest_level;
1950         WARN_ON(p->nodes[0] != NULL);
1951
1952         if (p->search_commit_root) {
1953                 BUG_ON(time_seq);
1954                 return btrfs_search_slot(NULL, root, key, p, 0, 0);
1955         }
1956
1957 again:
1958         b = btrfs_get_old_root(root, time_seq);
1959         if (!b) {
1960                 ret = -EIO;
1961                 goto done;
1962         }
1963         level = btrfs_header_level(b);
1964         p->locks[level] = BTRFS_READ_LOCK;
1965
1966         while (b) {
1967                 int dec = 0;
1968
1969                 level = btrfs_header_level(b);
1970                 p->nodes[level] = b;
1971
1972                 /*
1973                  * we have a lock on b and as long as we aren't changing
1974                  * the tree, there is no way to for the items in b to change.
1975                  * It is safe to drop the lock on our parent before we
1976                  * go through the expensive btree search on b.
1977                  */
1978                 btrfs_unlock_up_safe(p, level + 1);
1979
1980                 ret = btrfs_bin_search(b, key, &slot);
1981                 if (ret < 0)
1982                         goto done;
1983
1984                 if (level == 0) {
1985                         p->slots[level] = slot;
1986                         unlock_up(p, level, lowest_unlock, 0, NULL);
1987                         goto done;
1988                 }
1989
1990                 if (ret && slot > 0) {
1991                         dec = 1;
1992                         slot--;
1993                 }
1994                 p->slots[level] = slot;
1995                 unlock_up(p, level, lowest_unlock, 0, NULL);
1996
1997                 if (level == lowest_level) {
1998                         if (dec)
1999                                 p->slots[level]++;
2000                         goto done;
2001                 }
2002
2003                 err = read_block_for_search(root, p, &b, level, slot, key);
2004                 if (err == -EAGAIN)
2005                         goto again;
2006                 if (err) {
2007                         ret = err;
2008                         goto done;
2009                 }
2010
2011                 level = btrfs_header_level(b);
2012                 btrfs_tree_read_lock(b);
2013                 b = btrfs_tree_mod_log_rewind(fs_info, p, b, time_seq);
2014                 if (!b) {
2015                         ret = -ENOMEM;
2016                         goto done;
2017                 }
2018                 p->locks[level] = BTRFS_READ_LOCK;
2019                 p->nodes[level] = b;
2020         }
2021         ret = 1;
2022 done:
2023         if (ret < 0)
2024                 btrfs_release_path(p);
2025
2026         return ret;
2027 }
2028
2029 /*
2030  * helper to use instead of search slot if no exact match is needed but
2031  * instead the next or previous item should be returned.
2032  * When find_higher is true, the next higher item is returned, the next lower
2033  * otherwise.
2034  * When return_any and find_higher are both true, and no higher item is found,
2035  * return the next lower instead.
2036  * When return_any is true and find_higher is false, and no lower item is found,
2037  * return the next higher instead.
2038  * It returns 0 if any item is found, 1 if none is found (tree empty), and
2039  * < 0 on error
2040  */
2041 int btrfs_search_slot_for_read(struct btrfs_root *root,
2042                                const struct btrfs_key *key,
2043                                struct btrfs_path *p, int find_higher,
2044                                int return_any)
2045 {
2046         int ret;
2047         struct extent_buffer *leaf;
2048
2049 again:
2050         ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
2051         if (ret <= 0)
2052                 return ret;
2053         /*
2054          * a return value of 1 means the path is at the position where the
2055          * item should be inserted. Normally this is the next bigger item,
2056          * but in case the previous item is the last in a leaf, path points
2057          * to the first free slot in the previous leaf, i.e. at an invalid
2058          * item.
2059          */
2060         leaf = p->nodes[0];
2061
2062         if (find_higher) {
2063                 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
2064                         ret = btrfs_next_leaf(root, p);
2065                         if (ret <= 0)
2066                                 return ret;
2067                         if (!return_any)
2068                                 return 1;
2069                         /*
2070                          * no higher item found, return the next
2071                          * lower instead
2072                          */
2073                         return_any = 0;
2074                         find_higher = 0;
2075                         btrfs_release_path(p);
2076                         goto again;
2077                 }
2078         } else {
2079                 if (p->slots[0] == 0) {
2080                         ret = btrfs_prev_leaf(root, p);
2081                         if (ret < 0)
2082                                 return ret;
2083                         if (!ret) {
2084                                 leaf = p->nodes[0];
2085                                 if (p->slots[0] == btrfs_header_nritems(leaf))
2086                                         p->slots[0]--;
2087                                 return 0;
2088                         }
2089                         if (!return_any)
2090                                 return 1;
2091                         /*
2092                          * no lower item found, return the next
2093                          * higher instead
2094                          */
2095                         return_any = 0;
2096                         find_higher = 1;
2097                         btrfs_release_path(p);
2098                         goto again;
2099                 } else {
2100                         --p->slots[0];
2101                 }
2102         }
2103         return 0;
2104 }
2105
2106 /*
2107  * Execute search and call btrfs_previous_item to traverse backwards if the item
2108  * was not found.
2109  *
2110  * Return 0 if found, 1 if not found and < 0 if error.
2111  */
2112 int btrfs_search_backwards(struct btrfs_root *root, struct btrfs_key *key,
2113                            struct btrfs_path *path)
2114 {
2115         int ret;
2116
2117         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
2118         if (ret > 0)
2119                 ret = btrfs_previous_item(root, path, key->objectid, key->type);
2120
2121         if (ret == 0)
2122                 btrfs_item_key_to_cpu(path->nodes[0], key, path->slots[0]);
2123
2124         return ret;
2125 }
2126
2127 /*
2128  * adjust the pointers going up the tree, starting at level
2129  * making sure the right key of each node is points to 'key'.
2130  * This is used after shifting pointers to the left, so it stops
2131  * fixing up pointers when a given leaf/node is not in slot 0 of the
2132  * higher levels
2133  *
2134  */
2135 static void fixup_low_keys(struct btrfs_path *path,
2136                            struct btrfs_disk_key *key, int level)
2137 {
2138         int i;
2139         struct extent_buffer *t;
2140         int ret;
2141
2142         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2143                 int tslot = path->slots[i];
2144
2145                 if (!path->nodes[i])
2146                         break;
2147                 t = path->nodes[i];
2148                 ret = btrfs_tree_mod_log_insert_key(t, tslot,
2149                                 BTRFS_MOD_LOG_KEY_REPLACE, GFP_ATOMIC);
2150                 BUG_ON(ret < 0);
2151                 btrfs_set_node_key(t, key, tslot);
2152                 btrfs_mark_buffer_dirty(path->nodes[i]);
2153                 if (tslot != 0)
2154                         break;
2155         }
2156 }
2157
2158 /*
2159  * update item key.
2160  *
2161  * This function isn't completely safe. It's the caller's responsibility
2162  * that the new key won't break the order
2163  */
2164 void btrfs_set_item_key_safe(struct btrfs_fs_info *fs_info,
2165                              struct btrfs_path *path,
2166                              const struct btrfs_key *new_key)
2167 {
2168         struct btrfs_disk_key disk_key;
2169         struct extent_buffer *eb;
2170         int slot;
2171
2172         eb = path->nodes[0];
2173         slot = path->slots[0];
2174         if (slot > 0) {
2175                 btrfs_item_key(eb, &disk_key, slot - 1);
2176                 if (unlikely(comp_keys(&disk_key, new_key) >= 0)) {
2177                         btrfs_crit(fs_info,
2178                 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
2179                                    slot, btrfs_disk_key_objectid(&disk_key),
2180                                    btrfs_disk_key_type(&disk_key),
2181                                    btrfs_disk_key_offset(&disk_key),
2182                                    new_key->objectid, new_key->type,
2183                                    new_key->offset);
2184                         btrfs_print_leaf(eb);
2185                         BUG();
2186                 }
2187         }
2188         if (slot < btrfs_header_nritems(eb) - 1) {
2189                 btrfs_item_key(eb, &disk_key, slot + 1);
2190                 if (unlikely(comp_keys(&disk_key, new_key) <= 0)) {
2191                         btrfs_crit(fs_info,
2192                 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
2193                                    slot, btrfs_disk_key_objectid(&disk_key),
2194                                    btrfs_disk_key_type(&disk_key),
2195                                    btrfs_disk_key_offset(&disk_key),
2196                                    new_key->objectid, new_key->type,
2197                                    new_key->offset);
2198                         btrfs_print_leaf(eb);
2199                         BUG();
2200                 }
2201         }
2202
2203         btrfs_cpu_key_to_disk(&disk_key, new_key);
2204         btrfs_set_item_key(eb, &disk_key, slot);
2205         btrfs_mark_buffer_dirty(eb);
2206         if (slot == 0)
2207                 fixup_low_keys(path, &disk_key, 1);
2208 }
2209
2210 /*
2211  * Check key order of two sibling extent buffers.
2212  *
2213  * Return true if something is wrong.
2214  * Return false if everything is fine.
2215  *
2216  * Tree-checker only works inside one tree block, thus the following
2217  * corruption can not be detected by tree-checker:
2218  *
2219  * Leaf @left                   | Leaf @right
2220  * --------------------------------------------------------------
2221  * | 1 | 2 | 3 | 4 | 5 | f6 |   | 7 | 8 |
2222  *
2223  * Key f6 in leaf @left itself is valid, but not valid when the next
2224  * key in leaf @right is 7.
2225  * This can only be checked at tree block merge time.
2226  * And since tree checker has ensured all key order in each tree block
2227  * is correct, we only need to bother the last key of @left and the first
2228  * key of @right.
2229  */
2230 static bool check_sibling_keys(struct extent_buffer *left,
2231                                struct extent_buffer *right)
2232 {
2233         struct btrfs_key left_last;
2234         struct btrfs_key right_first;
2235         int level = btrfs_header_level(left);
2236         int nr_left = btrfs_header_nritems(left);
2237         int nr_right = btrfs_header_nritems(right);
2238
2239         /* No key to check in one of the tree blocks */
2240         if (!nr_left || !nr_right)
2241                 return false;
2242
2243         if (level) {
2244                 btrfs_node_key_to_cpu(left, &left_last, nr_left - 1);
2245                 btrfs_node_key_to_cpu(right, &right_first, 0);
2246         } else {
2247                 btrfs_item_key_to_cpu(left, &left_last, nr_left - 1);
2248                 btrfs_item_key_to_cpu(right, &right_first, 0);
2249         }
2250
2251         if (btrfs_comp_cpu_keys(&left_last, &right_first) >= 0) {
2252                 btrfs_crit(left->fs_info,
2253 "bad key order, sibling blocks, left last (%llu %u %llu) right first (%llu %u %llu)",
2254                            left_last.objectid, left_last.type,
2255                            left_last.offset, right_first.objectid,
2256                            right_first.type, right_first.offset);
2257                 return true;
2258         }
2259         return false;
2260 }
2261
2262 /*
2263  * try to push data from one node into the next node left in the
2264  * tree.
2265  *
2266  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
2267  * error, and > 0 if there was no room in the left hand block.
2268  */
2269 static int push_node_left(struct btrfs_trans_handle *trans,
2270                           struct extent_buffer *dst,
2271                           struct extent_buffer *src, int empty)
2272 {
2273         struct btrfs_fs_info *fs_info = trans->fs_info;
2274         int push_items = 0;
2275         int src_nritems;
2276         int dst_nritems;
2277         int ret = 0;
2278
2279         src_nritems = btrfs_header_nritems(src);
2280         dst_nritems = btrfs_header_nritems(dst);
2281         push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
2282         WARN_ON(btrfs_header_generation(src) != trans->transid);
2283         WARN_ON(btrfs_header_generation(dst) != trans->transid);
2284
2285         if (!empty && src_nritems <= 8)
2286                 return 1;
2287
2288         if (push_items <= 0)
2289                 return 1;
2290
2291         if (empty) {
2292                 push_items = min(src_nritems, push_items);
2293                 if (push_items < src_nritems) {
2294                         /* leave at least 8 pointers in the node if
2295                          * we aren't going to empty it
2296                          */
2297                         if (src_nritems - push_items < 8) {
2298                                 if (push_items <= 8)
2299                                         return 1;
2300                                 push_items -= 8;
2301                         }
2302                 }
2303         } else
2304                 push_items = min(src_nritems - 8, push_items);
2305
2306         /* dst is the left eb, src is the middle eb */
2307         if (check_sibling_keys(dst, src)) {
2308                 ret = -EUCLEAN;
2309                 btrfs_abort_transaction(trans, ret);
2310                 return ret;
2311         }
2312         ret = btrfs_tree_mod_log_eb_copy(dst, src, dst_nritems, 0, push_items);
2313         if (ret) {
2314                 btrfs_abort_transaction(trans, ret);
2315                 return ret;
2316         }
2317         copy_extent_buffer(dst, src,
2318                            btrfs_node_key_ptr_offset(dst_nritems),
2319                            btrfs_node_key_ptr_offset(0),
2320                            push_items * sizeof(struct btrfs_key_ptr));
2321
2322         if (push_items < src_nritems) {
2323                 /*
2324                  * Don't call btrfs_tree_mod_log_insert_move() here, key removal
2325                  * was already fully logged by btrfs_tree_mod_log_eb_copy() above.
2326                  */
2327                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
2328                                       btrfs_node_key_ptr_offset(push_items),
2329                                       (src_nritems - push_items) *
2330                                       sizeof(struct btrfs_key_ptr));
2331         }
2332         btrfs_set_header_nritems(src, src_nritems - push_items);
2333         btrfs_set_header_nritems(dst, dst_nritems + push_items);
2334         btrfs_mark_buffer_dirty(src);
2335         btrfs_mark_buffer_dirty(dst);
2336
2337         return ret;
2338 }
2339
2340 /*
2341  * try to push data from one node into the next node right in the
2342  * tree.
2343  *
2344  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
2345  * error, and > 0 if there was no room in the right hand block.
2346  *
2347  * this will  only push up to 1/2 the contents of the left node over
2348  */
2349 static int balance_node_right(struct btrfs_trans_handle *trans,
2350                               struct extent_buffer *dst,
2351                               struct extent_buffer *src)
2352 {
2353         struct btrfs_fs_info *fs_info = trans->fs_info;
2354         int push_items = 0;
2355         int max_push;
2356         int src_nritems;
2357         int dst_nritems;
2358         int ret = 0;
2359
2360         WARN_ON(btrfs_header_generation(src) != trans->transid);
2361         WARN_ON(btrfs_header_generation(dst) != trans->transid);
2362
2363         src_nritems = btrfs_header_nritems(src);
2364         dst_nritems = btrfs_header_nritems(dst);
2365         push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
2366         if (push_items <= 0)
2367                 return 1;
2368
2369         if (src_nritems < 4)
2370                 return 1;
2371
2372         max_push = src_nritems / 2 + 1;
2373         /* don't try to empty the node */
2374         if (max_push >= src_nritems)
2375                 return 1;
2376
2377         if (max_push < push_items)
2378                 push_items = max_push;
2379
2380         /* dst is the right eb, src is the middle eb */
2381         if (check_sibling_keys(src, dst)) {
2382                 ret = -EUCLEAN;
2383                 btrfs_abort_transaction(trans, ret);
2384                 return ret;
2385         }
2386         ret = btrfs_tree_mod_log_insert_move(dst, push_items, 0, dst_nritems);
2387         BUG_ON(ret < 0);
2388         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
2389                                       btrfs_node_key_ptr_offset(0),
2390                                       (dst_nritems) *
2391                                       sizeof(struct btrfs_key_ptr));
2392
2393         ret = btrfs_tree_mod_log_eb_copy(dst, src, 0, src_nritems - push_items,
2394                                          push_items);
2395         if (ret) {
2396                 btrfs_abort_transaction(trans, ret);
2397                 return ret;
2398         }
2399         copy_extent_buffer(dst, src,
2400                            btrfs_node_key_ptr_offset(0),
2401                            btrfs_node_key_ptr_offset(src_nritems - push_items),
2402                            push_items * sizeof(struct btrfs_key_ptr));
2403
2404         btrfs_set_header_nritems(src, src_nritems - push_items);
2405         btrfs_set_header_nritems(dst, dst_nritems + push_items);
2406
2407         btrfs_mark_buffer_dirty(src);
2408         btrfs_mark_buffer_dirty(dst);
2409
2410         return ret;
2411 }
2412
2413 /*
2414  * helper function to insert a new root level in the tree.
2415  * A new node is allocated, and a single item is inserted to
2416  * point to the existing root
2417  *
2418  * returns zero on success or < 0 on failure.
2419  */
2420 static noinline int insert_new_root(struct btrfs_trans_handle *trans,
2421                            struct btrfs_root *root,
2422                            struct btrfs_path *path, int level)
2423 {
2424         struct btrfs_fs_info *fs_info = root->fs_info;
2425         u64 lower_gen;
2426         struct extent_buffer *lower;
2427         struct extent_buffer *c;
2428         struct extent_buffer *old;
2429         struct btrfs_disk_key lower_key;
2430         int ret;
2431
2432         BUG_ON(path->nodes[level]);
2433         BUG_ON(path->nodes[level-1] != root->node);
2434
2435         lower = path->nodes[level-1];
2436         if (level == 1)
2437                 btrfs_item_key(lower, &lower_key, 0);
2438         else
2439                 btrfs_node_key(lower, &lower_key, 0);
2440
2441         c = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
2442                                    &lower_key, level, root->node->start, 0,
2443                                    BTRFS_NESTING_NEW_ROOT);
2444         if (IS_ERR(c))
2445                 return PTR_ERR(c);
2446
2447         root_add_used(root, fs_info->nodesize);
2448
2449         btrfs_set_header_nritems(c, 1);
2450         btrfs_set_node_key(c, &lower_key, 0);
2451         btrfs_set_node_blockptr(c, 0, lower->start);
2452         lower_gen = btrfs_header_generation(lower);
2453         WARN_ON(lower_gen != trans->transid);
2454
2455         btrfs_set_node_ptr_generation(c, 0, lower_gen);
2456
2457         btrfs_mark_buffer_dirty(c);
2458
2459         old = root->node;
2460         ret = btrfs_tree_mod_log_insert_root(root->node, c, false);
2461         BUG_ON(ret < 0);
2462         rcu_assign_pointer(root->node, c);
2463
2464         /* the super has an extra ref to root->node */
2465         free_extent_buffer(old);
2466
2467         add_root_to_dirty_list(root);
2468         atomic_inc(&c->refs);
2469         path->nodes[level] = c;
2470         path->locks[level] = BTRFS_WRITE_LOCK;
2471         path->slots[level] = 0;
2472         return 0;
2473 }
2474
2475 /*
2476  * worker function to insert a single pointer in a node.
2477  * the node should have enough room for the pointer already
2478  *
2479  * slot and level indicate where you want the key to go, and
2480  * blocknr is the block the key points to.
2481  */
2482 static void insert_ptr(struct btrfs_trans_handle *trans,
2483                        struct btrfs_path *path,
2484                        struct btrfs_disk_key *key, u64 bytenr,
2485                        int slot, int level)
2486 {
2487         struct extent_buffer *lower;
2488         int nritems;
2489         int ret;
2490
2491         BUG_ON(!path->nodes[level]);
2492         btrfs_assert_tree_write_locked(path->nodes[level]);
2493         lower = path->nodes[level];
2494         nritems = btrfs_header_nritems(lower);
2495         BUG_ON(slot > nritems);
2496         BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(trans->fs_info));
2497         if (slot != nritems) {
2498                 if (level) {
2499                         ret = btrfs_tree_mod_log_insert_move(lower, slot + 1,
2500                                         slot, nritems - slot);
2501                         BUG_ON(ret < 0);
2502                 }
2503                 memmove_extent_buffer(lower,
2504                               btrfs_node_key_ptr_offset(slot + 1),
2505                               btrfs_node_key_ptr_offset(slot),
2506                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
2507         }
2508         if (level) {
2509                 ret = btrfs_tree_mod_log_insert_key(lower, slot,
2510                                             BTRFS_MOD_LOG_KEY_ADD, GFP_NOFS);
2511                 BUG_ON(ret < 0);
2512         }
2513         btrfs_set_node_key(lower, key, slot);
2514         btrfs_set_node_blockptr(lower, slot, bytenr);
2515         WARN_ON(trans->transid == 0);
2516         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
2517         btrfs_set_header_nritems(lower, nritems + 1);
2518         btrfs_mark_buffer_dirty(lower);
2519 }
2520
2521 /*
2522  * split the node at the specified level in path in two.
2523  * The path is corrected to point to the appropriate node after the split
2524  *
2525  * Before splitting this tries to make some room in the node by pushing
2526  * left and right, if either one works, it returns right away.
2527  *
2528  * returns 0 on success and < 0 on failure
2529  */
2530 static noinline int split_node(struct btrfs_trans_handle *trans,
2531                                struct btrfs_root *root,
2532                                struct btrfs_path *path, int level)
2533 {
2534         struct btrfs_fs_info *fs_info = root->fs_info;
2535         struct extent_buffer *c;
2536         struct extent_buffer *split;
2537         struct btrfs_disk_key disk_key;
2538         int mid;
2539         int ret;
2540         u32 c_nritems;
2541
2542         c = path->nodes[level];
2543         WARN_ON(btrfs_header_generation(c) != trans->transid);
2544         if (c == root->node) {
2545                 /*
2546                  * trying to split the root, lets make a new one
2547                  *
2548                  * tree mod log: We don't log_removal old root in
2549                  * insert_new_root, because that root buffer will be kept as a
2550                  * normal node. We are going to log removal of half of the
2551                  * elements below with btrfs_tree_mod_log_eb_copy(). We're
2552                  * holding a tree lock on the buffer, which is why we cannot
2553                  * race with other tree_mod_log users.
2554                  */
2555                 ret = insert_new_root(trans, root, path, level + 1);
2556                 if (ret)
2557                         return ret;
2558         } else {
2559                 ret = push_nodes_for_insert(trans, root, path, level);
2560                 c = path->nodes[level];
2561                 if (!ret && btrfs_header_nritems(c) <
2562                     BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3)
2563                         return 0;
2564                 if (ret < 0)
2565                         return ret;
2566         }
2567
2568         c_nritems = btrfs_header_nritems(c);
2569         mid = (c_nritems + 1) / 2;
2570         btrfs_node_key(c, &disk_key, mid);
2571
2572         split = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
2573                                        &disk_key, level, c->start, 0,
2574                                        BTRFS_NESTING_SPLIT);
2575         if (IS_ERR(split))
2576                 return PTR_ERR(split);
2577
2578         root_add_used(root, fs_info->nodesize);
2579         ASSERT(btrfs_header_level(c) == level);
2580
2581         ret = btrfs_tree_mod_log_eb_copy(split, c, 0, mid, c_nritems - mid);
2582         if (ret) {
2583                 btrfs_abort_transaction(trans, ret);
2584                 return ret;
2585         }
2586         copy_extent_buffer(split, c,
2587                            btrfs_node_key_ptr_offset(0),
2588                            btrfs_node_key_ptr_offset(mid),
2589                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
2590         btrfs_set_header_nritems(split, c_nritems - mid);
2591         btrfs_set_header_nritems(c, mid);
2592
2593         btrfs_mark_buffer_dirty(c);
2594         btrfs_mark_buffer_dirty(split);
2595
2596         insert_ptr(trans, path, &disk_key, split->start,
2597                    path->slots[level + 1] + 1, level + 1);
2598
2599         if (path->slots[level] >= mid) {
2600                 path->slots[level] -= mid;
2601                 btrfs_tree_unlock(c);
2602                 free_extent_buffer(c);
2603                 path->nodes[level] = split;
2604                 path->slots[level + 1] += 1;
2605         } else {
2606                 btrfs_tree_unlock(split);
2607                 free_extent_buffer(split);
2608         }
2609         return 0;
2610 }
2611
2612 /*
2613  * how many bytes are required to store the items in a leaf.  start
2614  * and nr indicate which items in the leaf to check.  This totals up the
2615  * space used both by the item structs and the item data
2616  */
2617 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
2618 {
2619         struct btrfs_item *start_item;
2620         struct btrfs_item *end_item;
2621         int data_len;
2622         int nritems = btrfs_header_nritems(l);
2623         int end = min(nritems, start + nr) - 1;
2624
2625         if (!nr)
2626                 return 0;
2627         start_item = btrfs_item_nr(start);
2628         end_item = btrfs_item_nr(end);
2629         data_len = btrfs_item_offset(l, start_item) +
2630                    btrfs_item_size(l, start_item);
2631         data_len = data_len - btrfs_item_offset(l, end_item);
2632         data_len += sizeof(struct btrfs_item) * nr;
2633         WARN_ON(data_len < 0);
2634         return data_len;
2635 }
2636
2637 /*
2638  * The space between the end of the leaf items and
2639  * the start of the leaf data.  IOW, how much room
2640  * the leaf has left for both items and data
2641  */
2642 noinline int btrfs_leaf_free_space(struct extent_buffer *leaf)
2643 {
2644         struct btrfs_fs_info *fs_info = leaf->fs_info;
2645         int nritems = btrfs_header_nritems(leaf);
2646         int ret;
2647
2648         ret = BTRFS_LEAF_DATA_SIZE(fs_info) - leaf_space_used(leaf, 0, nritems);
2649         if (ret < 0) {
2650                 btrfs_crit(fs_info,
2651                            "leaf free space ret %d, leaf data size %lu, used %d nritems %d",
2652                            ret,
2653                            (unsigned long) BTRFS_LEAF_DATA_SIZE(fs_info),
2654                            leaf_space_used(leaf, 0, nritems), nritems);
2655         }
2656         return ret;
2657 }
2658
2659 /*
2660  * min slot controls the lowest index we're willing to push to the
2661  * right.  We'll push up to and including min_slot, but no lower
2662  */
2663 static noinline int __push_leaf_right(struct btrfs_path *path,
2664                                       int data_size, int empty,
2665                                       struct extent_buffer *right,
2666                                       int free_space, u32 left_nritems,
2667                                       u32 min_slot)
2668 {
2669         struct btrfs_fs_info *fs_info = right->fs_info;
2670         struct extent_buffer *left = path->nodes[0];
2671         struct extent_buffer *upper = path->nodes[1];
2672         struct btrfs_map_token token;
2673         struct btrfs_disk_key disk_key;
2674         int slot;
2675         u32 i;
2676         int push_space = 0;
2677         int push_items = 0;
2678         struct btrfs_item *item;
2679         u32 nr;
2680         u32 right_nritems;
2681         u32 data_end;
2682         u32 this_item_size;
2683
2684         if (empty)
2685                 nr = 0;
2686         else
2687                 nr = max_t(u32, 1, min_slot);
2688
2689         if (path->slots[0] >= left_nritems)
2690                 push_space += data_size;
2691
2692         slot = path->slots[1];
2693         i = left_nritems - 1;
2694         while (i >= nr) {
2695                 item = btrfs_item_nr(i);
2696
2697                 if (!empty && push_items > 0) {
2698                         if (path->slots[0] > i)
2699                                 break;
2700                         if (path->slots[0] == i) {
2701                                 int space = btrfs_leaf_free_space(left);
2702
2703                                 if (space + push_space * 2 > free_space)
2704                                         break;
2705                         }
2706                 }
2707
2708                 if (path->slots[0] == i)
2709                         push_space += data_size;
2710
2711                 this_item_size = btrfs_item_size(left, item);
2712                 if (this_item_size + sizeof(*item) + push_space > free_space)
2713                         break;
2714
2715                 push_items++;
2716                 push_space += this_item_size + sizeof(*item);
2717                 if (i == 0)
2718                         break;
2719                 i--;
2720         }
2721
2722         if (push_items == 0)
2723                 goto out_unlock;
2724
2725         WARN_ON(!empty && push_items == left_nritems);
2726
2727         /* push left to right */
2728         right_nritems = btrfs_header_nritems(right);
2729
2730         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
2731         push_space -= leaf_data_end(left);
2732
2733         /* make room in the right data area */
2734         data_end = leaf_data_end(right);
2735         memmove_extent_buffer(right,
2736                               BTRFS_LEAF_DATA_OFFSET + data_end - push_space,
2737                               BTRFS_LEAF_DATA_OFFSET + data_end,
2738                               BTRFS_LEAF_DATA_SIZE(fs_info) - data_end);
2739
2740         /* copy from the left data area */
2741         copy_extent_buffer(right, left, BTRFS_LEAF_DATA_OFFSET +
2742                      BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
2743                      BTRFS_LEAF_DATA_OFFSET + leaf_data_end(left),
2744                      push_space);
2745
2746         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
2747                               btrfs_item_nr_offset(0),
2748                               right_nritems * sizeof(struct btrfs_item));
2749
2750         /* copy the items from left to right */
2751         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
2752                    btrfs_item_nr_offset(left_nritems - push_items),
2753                    push_items * sizeof(struct btrfs_item));
2754
2755         /* update the item pointers */
2756         btrfs_init_map_token(&token, right);
2757         right_nritems += push_items;
2758         btrfs_set_header_nritems(right, right_nritems);
2759         push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
2760         for (i = 0; i < right_nritems; i++) {
2761                 item = btrfs_item_nr(i);
2762                 push_space -= btrfs_token_item_size(&token, item);
2763                 btrfs_set_token_item_offset(&token, item, push_space);
2764         }
2765
2766         left_nritems -= push_items;
2767         btrfs_set_header_nritems(left, left_nritems);
2768
2769         if (left_nritems)
2770                 btrfs_mark_buffer_dirty(left);
2771         else
2772                 btrfs_clean_tree_block(left);
2773
2774         btrfs_mark_buffer_dirty(right);
2775
2776         btrfs_item_key(right, &disk_key, 0);
2777         btrfs_set_node_key(upper, &disk_key, slot + 1);
2778         btrfs_mark_buffer_dirty(upper);
2779
2780         /* then fixup the leaf pointer in the path */
2781         if (path->slots[0] >= left_nritems) {
2782                 path->slots[0] -= left_nritems;
2783                 if (btrfs_header_nritems(path->nodes[0]) == 0)
2784                         btrfs_clean_tree_block(path->nodes[0]);
2785                 btrfs_tree_unlock(path->nodes[0]);
2786                 free_extent_buffer(path->nodes[0]);
2787                 path->nodes[0] = right;
2788                 path->slots[1] += 1;
2789         } else {
2790                 btrfs_tree_unlock(right);
2791                 free_extent_buffer(right);
2792         }
2793         return 0;
2794
2795 out_unlock:
2796         btrfs_tree_unlock(right);
2797         free_extent_buffer(right);
2798         return 1;
2799 }
2800
2801 /*
2802  * push some data in the path leaf to the right, trying to free up at
2803  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
2804  *
2805  * returns 1 if the push failed because the other node didn't have enough
2806  * room, 0 if everything worked out and < 0 if there were major errors.
2807  *
2808  * this will push starting from min_slot to the end of the leaf.  It won't
2809  * push any slot lower than min_slot
2810  */
2811 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
2812                            *root, struct btrfs_path *path,
2813                            int min_data_size, int data_size,
2814                            int empty, u32 min_slot)
2815 {
2816         struct extent_buffer *left = path->nodes[0];
2817         struct extent_buffer *right;
2818         struct extent_buffer *upper;
2819         int slot;
2820         int free_space;
2821         u32 left_nritems;
2822         int ret;
2823
2824         if (!path->nodes[1])
2825                 return 1;
2826
2827         slot = path->slots[1];
2828         upper = path->nodes[1];
2829         if (slot >= btrfs_header_nritems(upper) - 1)
2830                 return 1;
2831
2832         btrfs_assert_tree_write_locked(path->nodes[1]);
2833
2834         right = btrfs_read_node_slot(upper, slot + 1);
2835         /*
2836          * slot + 1 is not valid or we fail to read the right node,
2837          * no big deal, just return.
2838          */
2839         if (IS_ERR(right))
2840                 return 1;
2841
2842         __btrfs_tree_lock(right, BTRFS_NESTING_RIGHT);
2843
2844         free_space = btrfs_leaf_free_space(right);
2845         if (free_space < data_size)
2846                 goto out_unlock;
2847
2848         /* cow and double check */
2849         ret = btrfs_cow_block(trans, root, right, upper,
2850                               slot + 1, &right, BTRFS_NESTING_RIGHT_COW);
2851         if (ret)
2852                 goto out_unlock;
2853
2854         free_space = btrfs_leaf_free_space(right);
2855         if (free_space < data_size)
2856                 goto out_unlock;
2857
2858         left_nritems = btrfs_header_nritems(left);
2859         if (left_nritems == 0)
2860                 goto out_unlock;
2861
2862         if (check_sibling_keys(left, right)) {
2863                 ret = -EUCLEAN;
2864                 btrfs_tree_unlock(right);
2865                 free_extent_buffer(right);
2866                 return ret;
2867         }
2868         if (path->slots[0] == left_nritems && !empty) {
2869                 /* Key greater than all keys in the leaf, right neighbor has
2870                  * enough room for it and we're not emptying our leaf to delete
2871                  * it, therefore use right neighbor to insert the new item and
2872                  * no need to touch/dirty our left leaf. */
2873                 btrfs_tree_unlock(left);
2874                 free_extent_buffer(left);
2875                 path->nodes[0] = right;
2876                 path->slots[0] = 0;
2877                 path->slots[1]++;
2878                 return 0;
2879         }
2880
2881         return __push_leaf_right(path, min_data_size, empty,
2882                                 right, free_space, left_nritems, min_slot);
2883 out_unlock:
2884         btrfs_tree_unlock(right);
2885         free_extent_buffer(right);
2886         return 1;
2887 }
2888
2889 /*
2890  * push some data in the path leaf to the left, trying to free up at
2891  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
2892  *
2893  * max_slot can put a limit on how far into the leaf we'll push items.  The
2894  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us do all the
2895  * items
2896  */
2897 static noinline int __push_leaf_left(struct btrfs_path *path, int data_size,
2898                                      int empty, struct extent_buffer *left,
2899                                      int free_space, u32 right_nritems,
2900                                      u32 max_slot)
2901 {
2902         struct btrfs_fs_info *fs_info = left->fs_info;
2903         struct btrfs_disk_key disk_key;
2904         struct extent_buffer *right = path->nodes[0];
2905         int i;
2906         int push_space = 0;
2907         int push_items = 0;
2908         struct btrfs_item *item;
2909         u32 old_left_nritems;
2910         u32 nr;
2911         int ret = 0;
2912         u32 this_item_size;
2913         u32 old_left_item_size;
2914         struct btrfs_map_token token;
2915
2916         if (empty)
2917                 nr = min(right_nritems, max_slot);
2918         else
2919                 nr = min(right_nritems - 1, max_slot);
2920
2921         for (i = 0; i < nr; i++) {
2922                 item = btrfs_item_nr(i);
2923
2924                 if (!empty && push_items > 0) {
2925                         if (path->slots[0] < i)
2926                                 break;
2927                         if (path->slots[0] == i) {
2928                                 int space = btrfs_leaf_free_space(right);
2929
2930                                 if (space + push_space * 2 > free_space)
2931                                         break;
2932                         }
2933                 }
2934
2935                 if (path->slots[0] == i)
2936                         push_space += data_size;
2937
2938                 this_item_size = btrfs_item_size(right, item);
2939                 if (this_item_size + sizeof(*item) + push_space > free_space)
2940                         break;
2941
2942                 push_items++;
2943                 push_space += this_item_size + sizeof(*item);
2944         }
2945
2946         if (push_items == 0) {
2947                 ret = 1;
2948                 goto out;
2949         }
2950         WARN_ON(!empty && push_items == btrfs_header_nritems(right));
2951
2952         /* push data from right to left */
2953         copy_extent_buffer(left, right,
2954                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
2955                            btrfs_item_nr_offset(0),
2956                            push_items * sizeof(struct btrfs_item));
2957
2958         push_space = BTRFS_LEAF_DATA_SIZE(fs_info) -
2959                      btrfs_item_offset_nr(right, push_items - 1);
2960
2961         copy_extent_buffer(left, right, BTRFS_LEAF_DATA_OFFSET +
2962                      leaf_data_end(left) - push_space,
2963                      BTRFS_LEAF_DATA_OFFSET +
2964                      btrfs_item_offset_nr(right, push_items - 1),
2965                      push_space);
2966         old_left_nritems = btrfs_header_nritems(left);
2967         BUG_ON(old_left_nritems <= 0);
2968
2969         btrfs_init_map_token(&token, left);
2970         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
2971         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
2972                 u32 ioff;
2973
2974                 item = btrfs_item_nr(i);
2975
2976                 ioff = btrfs_token_item_offset(&token, item);
2977                 btrfs_set_token_item_offset(&token, item,
2978                       ioff - (BTRFS_LEAF_DATA_SIZE(fs_info) - old_left_item_size));
2979         }
2980         btrfs_set_header_nritems(left, old_left_nritems + push_items);
2981
2982         /* fixup right node */
2983         if (push_items > right_nritems)
2984                 WARN(1, KERN_CRIT "push items %d nr %u\n", push_items,
2985                        right_nritems);
2986
2987         if (push_items < right_nritems) {
2988                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
2989                                                   leaf_data_end(right);
2990                 memmove_extent_buffer(right, BTRFS_LEAF_DATA_OFFSET +
2991                                       BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
2992                                       BTRFS_LEAF_DATA_OFFSET +
2993                                       leaf_data_end(right), push_space);
2994
2995                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
2996                               btrfs_item_nr_offset(push_items),
2997                              (btrfs_header_nritems(right) - push_items) *
2998                              sizeof(struct btrfs_item));
2999         }
3000
3001         btrfs_init_map_token(&token, right);
3002         right_nritems -= push_items;
3003         btrfs_set_header_nritems(right, right_nritems);
3004         push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3005         for (i = 0; i < right_nritems; i++) {
3006                 item = btrfs_item_nr(i);
3007
3008                 push_space = push_space - btrfs_token_item_size(&token, item);
3009                 btrfs_set_token_item_offset(&token, item, push_space);
3010         }
3011
3012         btrfs_mark_buffer_dirty(left);
3013         if (right_nritems)
3014                 btrfs_mark_buffer_dirty(right);
3015         else
3016                 btrfs_clean_tree_block(right);
3017
3018         btrfs_item_key(right, &disk_key, 0);
3019         fixup_low_keys(path, &disk_key, 1);
3020
3021         /* then fixup the leaf pointer in the path */
3022         if (path->slots[0] < push_items) {
3023                 path->slots[0] += old_left_nritems;
3024                 btrfs_tree_unlock(path->nodes[0]);
3025                 free_extent_buffer(path->nodes[0]);
3026                 path->nodes[0] = left;
3027                 path->slots[1] -= 1;
3028         } else {
3029                 btrfs_tree_unlock(left);
3030                 free_extent_buffer(left);
3031                 path->slots[0] -= push_items;
3032         }
3033         BUG_ON(path->slots[0] < 0);
3034         return ret;
3035 out:
3036         btrfs_tree_unlock(left);
3037         free_extent_buffer(left);
3038         return ret;
3039 }
3040
3041 /*
3042  * push some data in the path leaf to the left, trying to free up at
3043  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
3044  *
3045  * max_slot can put a limit on how far into the leaf we'll push items.  The
3046  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us push all the
3047  * items
3048  */
3049 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
3050                           *root, struct btrfs_path *path, int min_data_size,
3051                           int data_size, int empty, u32 max_slot)
3052 {
3053         struct extent_buffer *right = path->nodes[0];
3054         struct extent_buffer *left;
3055         int slot;
3056         int free_space;
3057         u32 right_nritems;
3058         int ret = 0;
3059
3060         slot = path->slots[1];
3061         if (slot == 0)
3062                 return 1;
3063         if (!path->nodes[1])
3064                 return 1;
3065
3066         right_nritems = btrfs_header_nritems(right);
3067         if (right_nritems == 0)
3068                 return 1;
3069
3070         btrfs_assert_tree_write_locked(path->nodes[1]);
3071
3072         left = btrfs_read_node_slot(path->nodes[1], slot - 1);
3073         /*
3074          * slot - 1 is not valid or we fail to read the left node,
3075          * no big deal, just return.
3076          */
3077         if (IS_ERR(left))
3078                 return 1;
3079
3080         __btrfs_tree_lock(left, BTRFS_NESTING_LEFT);
3081
3082         free_space = btrfs_leaf_free_space(left);
3083         if (free_space < data_size) {
3084                 ret = 1;
3085                 goto out;
3086         }
3087
3088         /* cow and double check */
3089         ret = btrfs_cow_block(trans, root, left,
3090                               path->nodes[1], slot - 1, &left,
3091                               BTRFS_NESTING_LEFT_COW);
3092         if (ret) {
3093                 /* we hit -ENOSPC, but it isn't fatal here */
3094                 if (ret == -ENOSPC)
3095                         ret = 1;
3096                 goto out;
3097         }
3098
3099         free_space = btrfs_leaf_free_space(left);
3100         if (free_space < data_size) {
3101                 ret = 1;
3102                 goto out;
3103         }
3104
3105         if (check_sibling_keys(left, right)) {
3106                 ret = -EUCLEAN;
3107                 goto out;
3108         }
3109         return __push_leaf_left(path, min_data_size,
3110                                empty, left, free_space, right_nritems,
3111                                max_slot);
3112 out:
3113         btrfs_tree_unlock(left);
3114         free_extent_buffer(left);
3115         return ret;
3116 }
3117
3118 /*
3119  * split the path's leaf in two, making sure there is at least data_size
3120  * available for the resulting leaf level of the path.
3121  */
3122 static noinline void copy_for_split(struct btrfs_trans_handle *trans,
3123                                     struct btrfs_path *path,
3124                                     struct extent_buffer *l,
3125                                     struct extent_buffer *right,
3126                                     int slot, int mid, int nritems)
3127 {
3128         struct btrfs_fs_info *fs_info = trans->fs_info;
3129         int data_copy_size;
3130         int rt_data_off;
3131         int i;
3132         struct btrfs_disk_key disk_key;
3133         struct btrfs_map_token token;
3134
3135         nritems = nritems - mid;
3136         btrfs_set_header_nritems(right, nritems);
3137         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(l);
3138
3139         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
3140                            btrfs_item_nr_offset(mid),
3141                            nritems * sizeof(struct btrfs_item));
3142
3143         copy_extent_buffer(right, l,
3144                      BTRFS_LEAF_DATA_OFFSET + BTRFS_LEAF_DATA_SIZE(fs_info) -
3145                      data_copy_size, BTRFS_LEAF_DATA_OFFSET +
3146                      leaf_data_end(l), data_copy_size);
3147
3148         rt_data_off = BTRFS_LEAF_DATA_SIZE(fs_info) - btrfs_item_end_nr(l, mid);
3149
3150         btrfs_init_map_token(&token, right);
3151         for (i = 0; i < nritems; i++) {
3152                 struct btrfs_item *item = btrfs_item_nr(i);
3153                 u32 ioff;
3154
3155                 ioff = btrfs_token_item_offset(&token, item);
3156                 btrfs_set_token_item_offset(&token, item, ioff + rt_data_off);
3157         }
3158
3159         btrfs_set_header_nritems(l, mid);
3160         btrfs_item_key(right, &disk_key, 0);
3161         insert_ptr(trans, path, &disk_key, right->start, path->slots[1] + 1, 1);
3162
3163         btrfs_mark_buffer_dirty(right);
3164         btrfs_mark_buffer_dirty(l);
3165         BUG_ON(path->slots[0] != slot);
3166
3167         if (mid <= slot) {
3168                 btrfs_tree_unlock(path->nodes[0]);
3169                 free_extent_buffer(path->nodes[0]);
3170                 path->nodes[0] = right;
3171                 path->slots[0] -= mid;
3172                 path->slots[1] += 1;
3173         } else {
3174                 btrfs_tree_unlock(right);
3175                 free_extent_buffer(right);
3176         }
3177
3178         BUG_ON(path->slots[0] < 0);
3179 }
3180
3181 /*
3182  * double splits happen when we need to insert a big item in the middle
3183  * of a leaf.  A double split can leave us with 3 mostly empty leaves:
3184  * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
3185  *          A                 B                 C
3186  *
3187  * We avoid this by trying to push the items on either side of our target
3188  * into the adjacent leaves.  If all goes well we can avoid the double split
3189  * completely.
3190  */
3191 static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
3192                                           struct btrfs_root *root,
3193                                           struct btrfs_path *path,
3194                                           int data_size)
3195 {
3196         int ret;
3197         int progress = 0;
3198         int slot;
3199         u32 nritems;
3200         int space_needed = data_size;
3201
3202         slot = path->slots[0];
3203         if (slot < btrfs_header_nritems(path->nodes[0]))
3204                 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
3205
3206         /*
3207          * try to push all the items after our slot into the
3208          * right leaf
3209          */
3210         ret = push_leaf_right(trans, root, path, 1, space_needed, 0, slot);
3211         if (ret < 0)
3212                 return ret;
3213
3214         if (ret == 0)
3215                 progress++;
3216
3217         nritems = btrfs_header_nritems(path->nodes[0]);
3218         /*
3219          * our goal is to get our slot at the start or end of a leaf.  If
3220          * we've done so we're done
3221          */
3222         if (path->slots[0] == 0 || path->slots[0] == nritems)
3223                 return 0;
3224
3225         if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
3226                 return 0;
3227
3228         /* try to push all the items before our slot into the next leaf */
3229         slot = path->slots[0];
3230         space_needed = data_size;
3231         if (slot > 0)
3232                 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
3233         ret = push_leaf_left(trans, root, path, 1, space_needed, 0, slot);
3234         if (ret < 0)
3235                 return ret;
3236
3237         if (ret == 0)
3238                 progress++;
3239
3240         if (progress)
3241                 return 0;
3242         return 1;
3243 }
3244
3245 /*
3246  * split the path's leaf in two, making sure there is at least data_size
3247  * available for the resulting leaf level of the path.
3248  *
3249  * returns 0 if all went well and < 0 on failure.
3250  */
3251 static noinline int split_leaf(struct btrfs_trans_handle *trans,
3252                                struct btrfs_root *root,
3253                                const struct btrfs_key *ins_key,
3254                                struct btrfs_path *path, int data_size,
3255                                int extend)
3256 {
3257         struct btrfs_disk_key disk_key;
3258         struct extent_buffer *l;
3259         u32 nritems;
3260         int mid;
3261         int slot;
3262         struct extent_buffer *right;
3263         struct btrfs_fs_info *fs_info = root->fs_info;
3264         int ret = 0;
3265         int wret;
3266         int split;
3267         int num_doubles = 0;
3268         int tried_avoid_double = 0;
3269
3270         l = path->nodes[0];
3271         slot = path->slots[0];
3272         if (extend && data_size + btrfs_item_size_nr(l, slot) +
3273             sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(fs_info))
3274                 return -EOVERFLOW;
3275
3276         /* first try to make some room by pushing left and right */
3277         if (data_size && path->nodes[1]) {
3278                 int space_needed = data_size;
3279
3280                 if (slot < btrfs_header_nritems(l))
3281                         space_needed -= btrfs_leaf_free_space(l);
3282
3283                 wret = push_leaf_right(trans, root, path, space_needed,
3284                                        space_needed, 0, 0);
3285                 if (wret < 0)
3286                         return wret;
3287                 if (wret) {
3288                         space_needed = data_size;
3289                         if (slot > 0)
3290                                 space_needed -= btrfs_leaf_free_space(l);
3291                         wret = push_leaf_left(trans, root, path, space_needed,
3292                                               space_needed, 0, (u32)-1);
3293                         if (wret < 0)
3294                                 return wret;
3295                 }
3296                 l = path->nodes[0];
3297
3298                 /* did the pushes work? */
3299                 if (btrfs_leaf_free_space(l) >= data_size)
3300                         return 0;
3301         }
3302
3303         if (!path->nodes[1]) {
3304                 ret = insert_new_root(trans, root, path, 1);
3305                 if (ret)
3306                         return ret;
3307         }
3308 again:
3309         split = 1;
3310         l = path->nodes[0];
3311         slot = path->slots[0];
3312         nritems = btrfs_header_nritems(l);
3313         mid = (nritems + 1) / 2;
3314
3315         if (mid <= slot) {
3316                 if (nritems == 1 ||
3317                     leaf_space_used(l, mid, nritems - mid) + data_size >
3318                         BTRFS_LEAF_DATA_SIZE(fs_info)) {
3319                         if (slot >= nritems) {
3320                                 split = 0;
3321                         } else {
3322                                 mid = slot;
3323                                 if (mid != nritems &&
3324                                     leaf_space_used(l, mid, nritems - mid) +
3325                                     data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
3326                                         if (data_size && !tried_avoid_double)
3327                                                 goto push_for_double;
3328                                         split = 2;
3329                                 }
3330                         }
3331                 }
3332         } else {
3333                 if (leaf_space_used(l, 0, mid) + data_size >
3334                         BTRFS_LEAF_DATA_SIZE(fs_info)) {
3335                         if (!extend && data_size && slot == 0) {
3336                                 split = 0;
3337                         } else if ((extend || !data_size) && slot == 0) {
3338                                 mid = 1;
3339                         } else {
3340                                 mid = slot;
3341                                 if (mid != nritems &&
3342                                     leaf_space_used(l, mid, nritems - mid) +
3343                                     data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
3344                                         if (data_size && !tried_avoid_double)
3345                                                 goto push_for_double;
3346                                         split = 2;
3347                                 }
3348                         }
3349                 }
3350         }
3351
3352         if (split == 0)
3353                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
3354         else
3355                 btrfs_item_key(l, &disk_key, mid);
3356
3357         /*
3358          * We have to about BTRFS_NESTING_NEW_ROOT here if we've done a double
3359          * split, because we're only allowed to have MAX_LOCKDEP_SUBCLASSES
3360          * subclasses, which is 8 at the time of this patch, and we've maxed it
3361          * out.  In the future we could add a
3362          * BTRFS_NESTING_SPLIT_THE_SPLITTENING if we need to, but for now just
3363          * use BTRFS_NESTING_NEW_ROOT.
3364          */
3365         right = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
3366                                        &disk_key, 0, l->start, 0,
3367                                        num_doubles ? BTRFS_NESTING_NEW_ROOT :
3368                                        BTRFS_NESTING_SPLIT);
3369         if (IS_ERR(right))
3370                 return PTR_ERR(right);
3371
3372         root_add_used(root, fs_info->nodesize);
3373
3374         if (split == 0) {
3375                 if (mid <= slot) {
3376                         btrfs_set_header_nritems(right, 0);
3377                         insert_ptr(trans, path, &disk_key,
3378                                    right->start, path->slots[1] + 1, 1);
3379                         btrfs_tree_unlock(path->nodes[0]);
3380                         free_extent_buffer(path->nodes[0]);
3381                         path->nodes[0] = right;
3382                         path->slots[0] = 0;
3383                         path->slots[1] += 1;
3384                 } else {
3385                         btrfs_set_header_nritems(right, 0);
3386                         insert_ptr(trans, path, &disk_key,
3387                                    right->start, path->slots[1], 1);
3388                         btrfs_tree_unlock(path->nodes[0]);
3389                         free_extent_buffer(path->nodes[0]);
3390                         path->nodes[0] = right;
3391                         path->slots[0] = 0;
3392                         if (path->slots[1] == 0)
3393                                 fixup_low_keys(path, &disk_key, 1);
3394                 }
3395                 /*
3396                  * We create a new leaf 'right' for the required ins_len and
3397                  * we'll do btrfs_mark_buffer_dirty() on this leaf after copying
3398                  * the content of ins_len to 'right'.
3399                  */
3400                 return ret;
3401         }
3402
3403         copy_for_split(trans, path, l, right, slot, mid, nritems);
3404
3405         if (split == 2) {
3406                 BUG_ON(num_doubles != 0);
3407                 num_doubles++;
3408                 goto again;
3409         }
3410
3411         return 0;
3412
3413 push_for_double:
3414         push_for_double_split(trans, root, path, data_size);
3415         tried_avoid_double = 1;
3416         if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
3417                 return 0;
3418         goto again;
3419 }
3420
3421 static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
3422                                          struct btrfs_root *root,
3423                                          struct btrfs_path *path, int ins_len)
3424 {
3425         struct btrfs_key key;
3426         struct extent_buffer *leaf;
3427         struct btrfs_file_extent_item *fi;
3428         u64 extent_len = 0;
3429         u32 item_size;
3430         int ret;
3431
3432         leaf = path->nodes[0];
3433         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3434
3435         BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
3436                key.type != BTRFS_EXTENT_CSUM_KEY);
3437
3438         if (btrfs_leaf_free_space(leaf) >= ins_len)
3439                 return 0;
3440
3441         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3442         if (key.type == BTRFS_EXTENT_DATA_KEY) {
3443                 fi = btrfs_item_ptr(leaf, path->slots[0],
3444                                     struct btrfs_file_extent_item);
3445                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
3446         }
3447         btrfs_release_path(path);
3448
3449         path->keep_locks = 1;
3450         path->search_for_split = 1;
3451         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
3452         path->search_for_split = 0;
3453         if (ret > 0)
3454                 ret = -EAGAIN;
3455         if (ret < 0)
3456                 goto err;
3457
3458         ret = -EAGAIN;
3459         leaf = path->nodes[0];
3460         /* if our item isn't there, return now */
3461         if (item_size != btrfs_item_size_nr(leaf, path->slots[0]))
3462                 goto err;
3463
3464         /* the leaf has  changed, it now has room.  return now */
3465         if (btrfs_leaf_free_space(path->nodes[0]) >= ins_len)
3466                 goto err;
3467
3468         if (key.type == BTRFS_EXTENT_DATA_KEY) {
3469                 fi = btrfs_item_ptr(leaf, path->slots[0],
3470                                     struct btrfs_file_extent_item);
3471                 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
3472                         goto err;
3473         }
3474
3475         ret = split_leaf(trans, root, &key, path, ins_len, 1);
3476         if (ret)
3477                 goto err;
3478
3479         path->keep_locks = 0;
3480         btrfs_unlock_up_safe(path, 1);
3481         return 0;
3482 err:
3483         path->keep_locks = 0;
3484         return ret;
3485 }
3486
3487 static noinline int split_item(struct btrfs_path *path,
3488                                const struct btrfs_key *new_key,
3489                                unsigned long split_offset)
3490 {
3491         struct extent_buffer *leaf;
3492         struct btrfs_item *item;
3493         struct btrfs_item *new_item;
3494         int slot;
3495         char *buf;
3496         u32 nritems;
3497         u32 item_size;
3498         u32 orig_offset;
3499         struct btrfs_disk_key disk_key;
3500
3501         leaf = path->nodes[0];
3502         BUG_ON(btrfs_leaf_free_space(leaf) < sizeof(struct btrfs_item));
3503
3504         item = btrfs_item_nr(path->slots[0]);
3505         orig_offset = btrfs_item_offset(leaf, item);
3506         item_size = btrfs_item_size(leaf, item);
3507
3508         buf = kmalloc(item_size, GFP_NOFS);
3509         if (!buf)
3510                 return -ENOMEM;
3511
3512         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
3513                             path->slots[0]), item_size);
3514
3515         slot = path->slots[0] + 1;
3516         nritems = btrfs_header_nritems(leaf);
3517         if (slot != nritems) {
3518                 /* shift the items */
3519                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
3520                                 btrfs_item_nr_offset(slot),
3521                                 (nritems - slot) * sizeof(struct btrfs_item));
3522         }
3523
3524         btrfs_cpu_key_to_disk(&disk_key, new_key);
3525         btrfs_set_item_key(leaf, &disk_key, slot);
3526
3527         new_item = btrfs_item_nr(slot);
3528
3529         btrfs_set_item_offset(leaf, new_item, orig_offset);
3530         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
3531
3532         btrfs_set_item_offset(leaf, item,
3533                               orig_offset + item_size - split_offset);
3534         btrfs_set_item_size(leaf, item, split_offset);
3535
3536         btrfs_set_header_nritems(leaf, nritems + 1);
3537
3538         /* write the data for the start of the original item */
3539         write_extent_buffer(leaf, buf,
3540                             btrfs_item_ptr_offset(leaf, path->slots[0]),
3541                             split_offset);
3542
3543         /* write the data for the new item */
3544         write_extent_buffer(leaf, buf + split_offset,
3545                             btrfs_item_ptr_offset(leaf, slot),
3546                             item_size - split_offset);
3547         btrfs_mark_buffer_dirty(leaf);
3548
3549         BUG_ON(btrfs_leaf_free_space(leaf) < 0);
3550         kfree(buf);
3551         return 0;
3552 }
3553
3554 /*
3555  * This function splits a single item into two items,
3556  * giving 'new_key' to the new item and splitting the
3557  * old one at split_offset (from the start of the item).
3558  *
3559  * The path may be released by this operation.  After
3560  * the split, the path is pointing to the old item.  The
3561  * new item is going to be in the same node as the old one.
3562  *
3563  * Note, the item being split must be smaller enough to live alone on
3564  * a tree block with room for one extra struct btrfs_item
3565  *
3566  * This allows us to split the item in place, keeping a lock on the
3567  * leaf the entire time.
3568  */
3569 int btrfs_split_item(struct btrfs_trans_handle *trans,
3570                      struct btrfs_root *root,
3571                      struct btrfs_path *path,
3572                      const struct btrfs_key *new_key,
3573                      unsigned long split_offset)
3574 {
3575         int ret;
3576         ret = setup_leaf_for_split(trans, root, path,
3577                                    sizeof(struct btrfs_item));
3578         if (ret)
3579                 return ret;
3580
3581         ret = split_item(path, new_key, split_offset);
3582         return ret;
3583 }
3584
3585 /*
3586  * make the item pointed to by the path smaller.  new_size indicates
3587  * how small to make it, and from_end tells us if we just chop bytes
3588  * off the end of the item or if we shift the item to chop bytes off
3589  * the front.
3590  */
3591 void btrfs_truncate_item(struct btrfs_path *path, u32 new_size, int from_end)
3592 {
3593         int slot;
3594         struct extent_buffer *leaf;
3595         struct btrfs_item *item;
3596         u32 nritems;
3597         unsigned int data_end;
3598         unsigned int old_data_start;
3599         unsigned int old_size;
3600         unsigned int size_diff;
3601         int i;
3602         struct btrfs_map_token token;
3603
3604         leaf = path->nodes[0];
3605         slot = path->slots[0];
3606
3607         old_size = btrfs_item_size_nr(leaf, slot);
3608         if (old_size == new_size)
3609                 return;
3610
3611         nritems = btrfs_header_nritems(leaf);
3612         data_end = leaf_data_end(leaf);
3613
3614         old_data_start = btrfs_item_offset_nr(leaf, slot);
3615
3616         size_diff = old_size - new_size;
3617
3618         BUG_ON(slot < 0);
3619         BUG_ON(slot >= nritems);
3620
3621         /*
3622          * item0..itemN ... dataN.offset..dataN.size .. data0.size
3623          */
3624         /* first correct the data pointers */
3625         btrfs_init_map_token(&token, leaf);
3626         for (i = slot; i < nritems; i++) {
3627                 u32 ioff;
3628                 item = btrfs_item_nr(i);
3629
3630                 ioff = btrfs_token_item_offset(&token, item);
3631                 btrfs_set_token_item_offset(&token, item, ioff + size_diff);
3632         }
3633
3634         /* shift the data */
3635         if (from_end) {
3636                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
3637                               data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
3638                               data_end, old_data_start + new_size - data_end);
3639         } else {
3640                 struct btrfs_disk_key disk_key;
3641                 u64 offset;
3642
3643                 btrfs_item_key(leaf, &disk_key, slot);
3644
3645                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
3646                         unsigned long ptr;
3647                         struct btrfs_file_extent_item *fi;
3648
3649                         fi = btrfs_item_ptr(leaf, slot,
3650                                             struct btrfs_file_extent_item);
3651                         fi = (struct btrfs_file_extent_item *)(
3652                              (unsigned long)fi - size_diff);
3653
3654                         if (btrfs_file_extent_type(leaf, fi) ==
3655                             BTRFS_FILE_EXTENT_INLINE) {
3656                                 ptr = btrfs_item_ptr_offset(leaf, slot);
3657                                 memmove_extent_buffer(leaf, ptr,
3658                                       (unsigned long)fi,
3659                                       BTRFS_FILE_EXTENT_INLINE_DATA_START);
3660                         }
3661                 }
3662
3663                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
3664                               data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
3665                               data_end, old_data_start - data_end);
3666
3667                 offset = btrfs_disk_key_offset(&disk_key);
3668                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
3669                 btrfs_set_item_key(leaf, &disk_key, slot);
3670                 if (slot == 0)
3671                         fixup_low_keys(path, &disk_key, 1);
3672         }
3673
3674         item = btrfs_item_nr(slot);
3675         btrfs_set_item_size(leaf, item, new_size);
3676         btrfs_mark_buffer_dirty(leaf);
3677
3678         if (btrfs_leaf_free_space(leaf) < 0) {
3679                 btrfs_print_leaf(leaf);
3680                 BUG();
3681         }
3682 }
3683
3684 /*
3685  * make the item pointed to by the path bigger, data_size is the added size.
3686  */
3687 void btrfs_extend_item(struct btrfs_path *path, u32 data_size)
3688 {
3689         int slot;
3690         struct extent_buffer *leaf;
3691         struct btrfs_item *item;
3692         u32 nritems;
3693         unsigned int data_end;
3694         unsigned int old_data;
3695         unsigned int old_size;
3696         int i;
3697         struct btrfs_map_token token;
3698
3699         leaf = path->nodes[0];
3700
3701         nritems = btrfs_header_nritems(leaf);
3702         data_end = leaf_data_end(leaf);
3703
3704         if (btrfs_leaf_free_space(leaf) < data_size) {
3705                 btrfs_print_leaf(leaf);
3706                 BUG();
3707         }
3708         slot = path->slots[0];
3709         old_data = btrfs_item_end_nr(leaf, slot);
3710
3711         BUG_ON(slot < 0);
3712         if (slot >= nritems) {
3713                 btrfs_print_leaf(leaf);
3714                 btrfs_crit(leaf->fs_info, "slot %d too large, nritems %d",
3715                            slot, nritems);
3716                 BUG();
3717         }
3718
3719         /*
3720          * item0..itemN ... dataN.offset..dataN.size .. data0.size
3721          */
3722         /* first correct the data pointers */
3723         btrfs_init_map_token(&token, leaf);
3724         for (i = slot; i < nritems; i++) {
3725                 u32 ioff;
3726                 item = btrfs_item_nr(i);
3727
3728                 ioff = btrfs_token_item_offset(&token, item);
3729                 btrfs_set_token_item_offset(&token, item, ioff - data_size);
3730         }
3731
3732         /* shift the data */
3733         memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
3734                       data_end - data_size, BTRFS_LEAF_DATA_OFFSET +
3735                       data_end, old_data - data_end);
3736
3737         data_end = old_data;
3738         old_size = btrfs_item_size_nr(leaf, slot);
3739         item = btrfs_item_nr(slot);
3740         btrfs_set_item_size(leaf, item, old_size + data_size);
3741         btrfs_mark_buffer_dirty(leaf);
3742
3743         if (btrfs_leaf_free_space(leaf) < 0) {
3744                 btrfs_print_leaf(leaf);
3745                 BUG();
3746         }
3747 }
3748
3749 /**
3750  * setup_items_for_insert - Helper called before inserting one or more items
3751  * to a leaf. Main purpose is to save stack depth by doing the bulk of the work
3752  * in a function that doesn't call btrfs_search_slot
3753  *
3754  * @root:       root we are inserting items to
3755  * @path:       points to the leaf/slot where we are going to insert new items
3756  * @batch:      information about the batch of items to insert
3757  */
3758 static void setup_items_for_insert(struct btrfs_root *root, struct btrfs_path *path,
3759                                    const struct btrfs_item_batch *batch)
3760 {
3761         struct btrfs_fs_info *fs_info = root->fs_info;
3762         struct btrfs_item *item;
3763         int i;
3764         u32 nritems;
3765         unsigned int data_end;
3766         struct btrfs_disk_key disk_key;
3767         struct extent_buffer *leaf;
3768         int slot;
3769         struct btrfs_map_token token;
3770         u32 total_size;
3771
3772         /*
3773          * Before anything else, update keys in the parent and other ancestors
3774          * if needed, then release the write locks on them, so that other tasks
3775          * can use them while we modify the leaf.
3776          */
3777         if (path->slots[0] == 0) {
3778                 btrfs_cpu_key_to_disk(&disk_key, &batch->keys[0]);
3779                 fixup_low_keys(path, &disk_key, 1);
3780         }
3781         btrfs_unlock_up_safe(path, 1);
3782
3783         leaf = path->nodes[0];
3784         slot = path->slots[0];
3785
3786         nritems = btrfs_header_nritems(leaf);
3787         data_end = leaf_data_end(leaf);
3788         total_size = batch->total_data_size + (batch->nr * sizeof(struct btrfs_item));
3789
3790         if (btrfs_leaf_free_space(leaf) < total_size) {
3791                 btrfs_print_leaf(leaf);
3792                 btrfs_crit(fs_info, "not enough freespace need %u have %d",
3793                            total_size, btrfs_leaf_free_space(leaf));
3794                 BUG();
3795         }
3796
3797         btrfs_init_map_token(&token, leaf);
3798         if (slot != nritems) {
3799                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
3800
3801                 if (old_data < data_end) {
3802                         btrfs_print_leaf(leaf);
3803                         btrfs_crit(fs_info,
3804                 "item at slot %d with data offset %u beyond data end of leaf %u",
3805                                    slot, old_data, data_end);
3806                         BUG();
3807                 }
3808                 /*
3809                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
3810                  */
3811                 /* first correct the data pointers */
3812                 for (i = slot; i < nritems; i++) {
3813                         u32 ioff;
3814
3815                         item = btrfs_item_nr(i);
3816                         ioff = btrfs_token_item_offset(&token, item);
3817                         btrfs_set_token_item_offset(&token, item,
3818                                                     ioff - batch->total_data_size);
3819                 }
3820                 /* shift the items */
3821                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + batch->nr),
3822                               btrfs_item_nr_offset(slot),
3823                               (nritems - slot) * sizeof(struct btrfs_item));
3824
3825                 /* shift the data */
3826                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
3827                                       data_end - batch->total_data_size,
3828                                       BTRFS_LEAF_DATA_OFFSET + data_end,
3829                                       old_data - data_end);
3830                 data_end = old_data;
3831         }
3832
3833         /* setup the item for the new data */
3834         for (i = 0; i < batch->nr; i++) {
3835                 btrfs_cpu_key_to_disk(&disk_key, &batch->keys[i]);
3836                 btrfs_set_item_key(leaf, &disk_key, slot + i);
3837                 item = btrfs_item_nr(slot + i);
3838                 data_end -= batch->data_sizes[i];
3839                 btrfs_set_token_item_offset(&token, item, data_end);
3840                 btrfs_set_token_item_size(&token, item, batch->data_sizes[i]);
3841         }
3842
3843         btrfs_set_header_nritems(leaf, nritems + batch->nr);
3844         btrfs_mark_buffer_dirty(leaf);
3845
3846         if (btrfs_leaf_free_space(leaf) < 0) {
3847                 btrfs_print_leaf(leaf);
3848                 BUG();
3849         }
3850 }
3851
3852 /*
3853  * Insert a new item into a leaf.
3854  *
3855  * @root:      The root of the btree.
3856  * @path:      A path pointing to the target leaf and slot.
3857  * @key:       The key of the new item.
3858  * @data_size: The size of the data associated with the new key.
3859  */
3860 void btrfs_setup_item_for_insert(struct btrfs_root *root,
3861                                  struct btrfs_path *path,
3862                                  const struct btrfs_key *key,
3863                                  u32 data_size)
3864 {
3865         struct btrfs_item_batch batch;
3866
3867         batch.keys = key;
3868         batch.data_sizes = &data_size;
3869         batch.total_data_size = data_size;
3870         batch.nr = 1;
3871
3872         setup_items_for_insert(root, path, &batch);
3873 }
3874
3875 /*
3876  * Given a key and some data, insert items into the tree.
3877  * This does all the path init required, making room in the tree if needed.
3878  */
3879 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
3880                             struct btrfs_root *root,
3881                             struct btrfs_path *path,
3882                             const struct btrfs_item_batch *batch)
3883 {
3884         int ret = 0;
3885         int slot;
3886         u32 total_size;
3887
3888         total_size = batch->total_data_size + (batch->nr * sizeof(struct btrfs_item));
3889         ret = btrfs_search_slot(trans, root, &batch->keys[0], path, total_size, 1);
3890         if (ret == 0)
3891                 return -EEXIST;
3892         if (ret < 0)
3893                 return ret;
3894
3895         slot = path->slots[0];
3896         BUG_ON(slot < 0);
3897
3898         setup_items_for_insert(root, path, batch);
3899         return 0;
3900 }
3901
3902 /*
3903  * Given a key and some data, insert an item into the tree.
3904  * This does all the path init required, making room in the tree if needed.
3905  */
3906 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3907                       const struct btrfs_key *cpu_key, void *data,
3908                       u32 data_size)
3909 {
3910         int ret = 0;
3911         struct btrfs_path *path;
3912         struct extent_buffer *leaf;
3913         unsigned long ptr;
3914
3915         path = btrfs_alloc_path();
3916         if (!path)
3917                 return -ENOMEM;
3918         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
3919         if (!ret) {
3920                 leaf = path->nodes[0];
3921                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3922                 write_extent_buffer(leaf, data, ptr, data_size);
3923                 btrfs_mark_buffer_dirty(leaf);
3924         }
3925         btrfs_free_path(path);
3926         return ret;
3927 }
3928
3929 /*
3930  * This function duplicates an item, giving 'new_key' to the new item.
3931  * It guarantees both items live in the same tree leaf and the new item is
3932  * contiguous with the original item.
3933  *
3934  * This allows us to split a file extent in place, keeping a lock on the leaf
3935  * the entire time.
3936  */
3937 int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
3938                          struct btrfs_root *root,
3939                          struct btrfs_path *path,
3940                          const struct btrfs_key *new_key)
3941 {
3942         struct extent_buffer *leaf;
3943         int ret;
3944         u32 item_size;
3945
3946         leaf = path->nodes[0];
3947         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3948         ret = setup_leaf_for_split(trans, root, path,
3949                                    item_size + sizeof(struct btrfs_item));
3950         if (ret)
3951                 return ret;
3952
3953         path->slots[0]++;
3954         btrfs_setup_item_for_insert(root, path, new_key, item_size);
3955         leaf = path->nodes[0];
3956         memcpy_extent_buffer(leaf,
3957                              btrfs_item_ptr_offset(leaf, path->slots[0]),
3958                              btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
3959                              item_size);
3960         return 0;
3961 }
3962
3963 /*
3964  * delete the pointer from a given node.
3965  *
3966  * the tree should have been previously balanced so the deletion does not
3967  * empty a node.
3968  */
3969 static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
3970                     int level, int slot)
3971 {
3972         struct extent_buffer *parent = path->nodes[level];
3973         u32 nritems;
3974         int ret;
3975
3976         nritems = btrfs_header_nritems(parent);
3977         if (slot != nritems - 1) {
3978                 if (level) {
3979                         ret = btrfs_tree_mod_log_insert_move(parent, slot,
3980                                         slot + 1, nritems - slot - 1);
3981                         BUG_ON(ret < 0);
3982                 }
3983                 memmove_extent_buffer(parent,
3984                               btrfs_node_key_ptr_offset(slot),
3985                               btrfs_node_key_ptr_offset(slot + 1),
3986                               sizeof(struct btrfs_key_ptr) *
3987                               (nritems - slot - 1));
3988         } else if (level) {
3989                 ret = btrfs_tree_mod_log_insert_key(parent, slot,
3990                                 BTRFS_MOD_LOG_KEY_REMOVE, GFP_NOFS);
3991                 BUG_ON(ret < 0);
3992         }
3993
3994         nritems--;
3995         btrfs_set_header_nritems(parent, nritems);
3996         if (nritems == 0 && parent == root->node) {
3997                 BUG_ON(btrfs_header_level(root->node) != 1);
3998                 /* just turn the root into a leaf and break */
3999                 btrfs_set_header_level(root->node, 0);
4000         } else if (slot == 0) {
4001                 struct btrfs_disk_key disk_key;
4002
4003                 btrfs_node_key(parent, &disk_key, 0);
4004                 fixup_low_keys(path, &disk_key, level + 1);
4005         }
4006         btrfs_mark_buffer_dirty(parent);
4007 }
4008
4009 /*
4010  * a helper function to delete the leaf pointed to by path->slots[1] and
4011  * path->nodes[1].
4012  *
4013  * This deletes the pointer in path->nodes[1] and frees the leaf
4014  * block extent.  zero is returned if it all worked out, < 0 otherwise.
4015  *
4016  * The path must have already been setup for deleting the leaf, including
4017  * all the proper balancing.  path->nodes[1] must be locked.
4018  */
4019 static noinline void btrfs_del_leaf(struct btrfs_trans_handle *trans,
4020                                     struct btrfs_root *root,
4021                                     struct btrfs_path *path,
4022                                     struct extent_buffer *leaf)
4023 {
4024         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
4025         del_ptr(root, path, 1, path->slots[1]);
4026
4027         /*
4028          * btrfs_free_extent is expensive, we want to make sure we
4029          * aren't holding any locks when we call it
4030          */
4031         btrfs_unlock_up_safe(path, 0);
4032
4033         root_sub_used(root, leaf->len);
4034
4035         atomic_inc(&leaf->refs);
4036         btrfs_free_tree_block(trans, btrfs_root_id(root), leaf, 0, 1);
4037         free_extent_buffer_stale(leaf);
4038 }
4039 /*
4040  * delete the item at the leaf level in path.  If that empties
4041  * the leaf, remove it from the tree
4042  */
4043 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4044                     struct btrfs_path *path, int slot, int nr)
4045 {
4046         struct btrfs_fs_info *fs_info = root->fs_info;
4047         struct extent_buffer *leaf;
4048         struct btrfs_item *item;
4049         u32 last_off;
4050         u32 dsize = 0;
4051         int ret = 0;
4052         int wret;
4053         int i;
4054         u32 nritems;
4055
4056         leaf = path->nodes[0];
4057         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
4058
4059         for (i = 0; i < nr; i++)
4060                 dsize += btrfs_item_size_nr(leaf, slot + i);
4061
4062         nritems = btrfs_header_nritems(leaf);
4063
4064         if (slot + nr != nritems) {
4065                 int data_end = leaf_data_end(leaf);
4066                 struct btrfs_map_token token;
4067
4068                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4069                               data_end + dsize,
4070                               BTRFS_LEAF_DATA_OFFSET + data_end,
4071                               last_off - data_end);
4072
4073                 btrfs_init_map_token(&token, leaf);
4074                 for (i = slot + nr; i < nritems; i++) {
4075                         u32 ioff;
4076
4077                         item = btrfs_item_nr(i);
4078                         ioff = btrfs_token_item_offset(&token, item);
4079                         btrfs_set_token_item_offset(&token, item, ioff + dsize);
4080                 }
4081
4082                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
4083                               btrfs_item_nr_offset(slot + nr),
4084                               sizeof(struct btrfs_item) *
4085                               (nritems - slot - nr));
4086         }
4087         btrfs_set_header_nritems(leaf, nritems - nr);
4088         nritems -= nr;
4089
4090         /* delete the leaf if we've emptied it */
4091         if (nritems == 0) {
4092                 if (leaf == root->node) {
4093                         btrfs_set_header_level(leaf, 0);
4094                 } else {
4095                         btrfs_clean_tree_block(leaf);
4096                         btrfs_del_leaf(trans, root, path, leaf);
4097                 }
4098         } else {
4099                 int used = leaf_space_used(leaf, 0, nritems);
4100                 if (slot == 0) {
4101                         struct btrfs_disk_key disk_key;
4102
4103                         btrfs_item_key(leaf, &disk_key, 0);
4104                         fixup_low_keys(path, &disk_key, 1);
4105                 }
4106
4107                 /* delete the leaf if it is mostly empty */
4108                 if (used < BTRFS_LEAF_DATA_SIZE(fs_info) / 3) {
4109                         /* push_leaf_left fixes the path.
4110                          * make sure the path still points to our leaf
4111                          * for possible call to del_ptr below
4112                          */
4113                         slot = path->slots[1];
4114                         atomic_inc(&leaf->refs);
4115
4116                         wret = push_leaf_left(trans, root, path, 1, 1,
4117                                               1, (u32)-1);
4118                         if (wret < 0 && wret != -ENOSPC)
4119                                 ret = wret;
4120
4121                         if (path->nodes[0] == leaf &&
4122                             btrfs_header_nritems(leaf)) {
4123                                 wret = push_leaf_right(trans, root, path, 1,
4124                                                        1, 1, 0);
4125                                 if (wret < 0 && wret != -ENOSPC)
4126                                         ret = wret;
4127                         }
4128
4129                         if (btrfs_header_nritems(leaf) == 0) {
4130                                 path->slots[1] = slot;
4131                                 btrfs_del_leaf(trans, root, path, leaf);
4132                                 free_extent_buffer(leaf);
4133                                 ret = 0;
4134                         } else {
4135                                 /* if we're still in the path, make sure
4136                                  * we're dirty.  Otherwise, one of the
4137                                  * push_leaf functions must have already
4138                                  * dirtied this buffer
4139                                  */
4140                                 if (path->nodes[0] == leaf)
4141                                         btrfs_mark_buffer_dirty(leaf);
4142                                 free_extent_buffer(leaf);
4143                         }
4144                 } else {
4145                         btrfs_mark_buffer_dirty(leaf);
4146                 }
4147         }
4148         return ret;
4149 }
4150
4151 /*
4152  * search the tree again to find a leaf with lesser keys
4153  * returns 0 if it found something or 1 if there are no lesser leaves.
4154  * returns < 0 on io errors.
4155  *
4156  * This may release the path, and so you may lose any locks held at the
4157  * time you call it.
4158  */
4159 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
4160 {
4161         struct btrfs_key key;
4162         struct btrfs_disk_key found_key;
4163         int ret;
4164
4165         btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
4166
4167         if (key.offset > 0) {
4168                 key.offset--;
4169         } else if (key.type > 0) {
4170                 key.type--;
4171                 key.offset = (u64)-1;
4172         } else if (key.objectid > 0) {
4173                 key.objectid--;
4174                 key.type = (u8)-1;
4175                 key.offset = (u64)-1;
4176         } else {
4177                 return 1;
4178         }
4179
4180         btrfs_release_path(path);
4181         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4182         if (ret < 0)
4183                 return ret;
4184         btrfs_item_key(path->nodes[0], &found_key, 0);
4185         ret = comp_keys(&found_key, &key);
4186         /*
4187          * We might have had an item with the previous key in the tree right
4188          * before we released our path. And after we released our path, that
4189          * item might have been pushed to the first slot (0) of the leaf we
4190          * were holding due to a tree balance. Alternatively, an item with the
4191          * previous key can exist as the only element of a leaf (big fat item).
4192          * Therefore account for these 2 cases, so that our callers (like
4193          * btrfs_previous_item) don't miss an existing item with a key matching
4194          * the previous key we computed above.
4195          */
4196         if (ret <= 0)
4197                 return 0;
4198         return 1;
4199 }
4200
4201 /*
4202  * A helper function to walk down the tree starting at min_key, and looking
4203  * for nodes or leaves that are have a minimum transaction id.
4204  * This is used by the btree defrag code, and tree logging
4205  *
4206  * This does not cow, but it does stuff the starting key it finds back
4207  * into min_key, so you can call btrfs_search_slot with cow=1 on the
4208  * key and get a writable path.
4209  *
4210  * This honors path->lowest_level to prevent descent past a given level
4211  * of the tree.
4212  *
4213  * min_trans indicates the oldest transaction that you are interested
4214  * in walking through.  Any nodes or leaves older than min_trans are
4215  * skipped over (without reading them).
4216  *
4217  * returns zero if something useful was found, < 0 on error and 1 if there
4218  * was nothing in the tree that matched the search criteria.
4219  */
4220 int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
4221                          struct btrfs_path *path,
4222                          u64 min_trans)
4223 {
4224         struct extent_buffer *cur;
4225         struct btrfs_key found_key;
4226         int slot;
4227         int sret;
4228         u32 nritems;
4229         int level;
4230         int ret = 1;
4231         int keep_locks = path->keep_locks;
4232
4233         path->keep_locks = 1;
4234 again:
4235         cur = btrfs_read_lock_root_node(root);
4236         level = btrfs_header_level(cur);
4237         WARN_ON(path->nodes[level]);
4238         path->nodes[level] = cur;
4239         path->locks[level] = BTRFS_READ_LOCK;
4240
4241         if (btrfs_header_generation(cur) < min_trans) {
4242                 ret = 1;
4243                 goto out;
4244         }
4245         while (1) {
4246                 nritems = btrfs_header_nritems(cur);
4247                 level = btrfs_header_level(cur);
4248                 sret = btrfs_bin_search(cur, min_key, &slot);
4249                 if (sret < 0) {
4250                         ret = sret;
4251                         goto out;
4252                 }
4253
4254                 /* at the lowest level, we're done, setup the path and exit */
4255                 if (level == path->lowest_level) {
4256                         if (slot >= nritems)
4257                                 goto find_next_key;
4258                         ret = 0;
4259                         path->slots[level] = slot;
4260                         btrfs_item_key_to_cpu(cur, &found_key, slot);
4261                         goto out;
4262                 }
4263                 if (sret && slot > 0)
4264                         slot--;
4265                 /*
4266                  * check this node pointer against the min_trans parameters.
4267                  * If it is too old, skip to the next one.
4268                  */
4269                 while (slot < nritems) {
4270                         u64 gen;
4271
4272                         gen = btrfs_node_ptr_generation(cur, slot);
4273                         if (gen < min_trans) {
4274                                 slot++;
4275                                 continue;
4276                         }
4277                         break;
4278                 }
4279 find_next_key:
4280                 /*
4281                  * we didn't find a candidate key in this node, walk forward
4282                  * and find another one
4283                  */
4284                 if (slot >= nritems) {
4285                         path->slots[level] = slot;
4286                         sret = btrfs_find_next_key(root, path, min_key, level,
4287                                                   min_trans);
4288                         if (sret == 0) {
4289                                 btrfs_release_path(path);
4290                                 goto again;
4291                         } else {
4292                                 goto out;
4293                         }
4294                 }
4295                 /* save our key for returning back */
4296                 btrfs_node_key_to_cpu(cur, &found_key, slot);
4297                 path->slots[level] = slot;
4298                 if (level == path->lowest_level) {
4299                         ret = 0;
4300                         goto out;
4301                 }
4302                 cur = btrfs_read_node_slot(cur, slot);
4303                 if (IS_ERR(cur)) {
4304                         ret = PTR_ERR(cur);
4305                         goto out;
4306                 }
4307
4308                 btrfs_tree_read_lock(cur);
4309
4310                 path->locks[level - 1] = BTRFS_READ_LOCK;
4311                 path->nodes[level - 1] = cur;
4312                 unlock_up(path, level, 1, 0, NULL);
4313         }
4314 out:
4315         path->keep_locks = keep_locks;
4316         if (ret == 0) {
4317                 btrfs_unlock_up_safe(path, path->lowest_level + 1);
4318                 memcpy(min_key, &found_key, sizeof(found_key));
4319         }
4320         return ret;
4321 }
4322
4323 /*
4324  * this is similar to btrfs_next_leaf, but does not try to preserve
4325  * and fixup the path.  It looks for and returns the next key in the
4326  * tree based on the current path and the min_trans parameters.
4327  *
4328  * 0 is returned if another key is found, < 0 if there are any errors
4329  * and 1 is returned if there are no higher keys in the tree
4330  *
4331  * path->keep_locks should be set to 1 on the search made before
4332  * calling this function.
4333  */
4334 int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
4335                         struct btrfs_key *key, int level, u64 min_trans)
4336 {
4337         int slot;
4338         struct extent_buffer *c;
4339
4340         WARN_ON(!path->keep_locks && !path->skip_locking);
4341         while (level < BTRFS_MAX_LEVEL) {
4342                 if (!path->nodes[level])
4343                         return 1;
4344
4345                 slot = path->slots[level] + 1;
4346                 c = path->nodes[level];
4347 next:
4348                 if (slot >= btrfs_header_nritems(c)) {
4349                         int ret;
4350                         int orig_lowest;
4351                         struct btrfs_key cur_key;
4352                         if (level + 1 >= BTRFS_MAX_LEVEL ||
4353                             !path->nodes[level + 1])
4354                                 return 1;
4355
4356                         if (path->locks[level + 1] || path->skip_locking) {
4357                                 level++;
4358                                 continue;
4359                         }
4360
4361                         slot = btrfs_header_nritems(c) - 1;
4362                         if (level == 0)
4363                                 btrfs_item_key_to_cpu(c, &cur_key, slot);
4364                         else
4365                                 btrfs_node_key_to_cpu(c, &cur_key, slot);
4366
4367                         orig_lowest = path->lowest_level;
4368                         btrfs_release_path(path);
4369                         path->lowest_level = level;
4370                         ret = btrfs_search_slot(NULL, root, &cur_key, path,
4371                                                 0, 0);
4372                         path->lowest_level = orig_lowest;
4373                         if (ret < 0)
4374                                 return ret;
4375
4376                         c = path->nodes[level];
4377                         slot = path->slots[level];
4378                         if (ret == 0)
4379                                 slot++;
4380                         goto next;
4381                 }
4382
4383                 if (level == 0)
4384                         btrfs_item_key_to_cpu(c, key, slot);
4385                 else {
4386                         u64 gen = btrfs_node_ptr_generation(c, slot);
4387
4388                         if (gen < min_trans) {
4389                                 slot++;
4390                                 goto next;
4391                         }
4392                         btrfs_node_key_to_cpu(c, key, slot);
4393                 }
4394                 return 0;
4395         }
4396         return 1;
4397 }
4398
4399 int btrfs_next_old_leaf(struct btrfs_root *root, struct btrfs_path *path,
4400                         u64 time_seq)
4401 {
4402         int slot;
4403         int level;
4404         struct extent_buffer *c;
4405         struct extent_buffer *next;
4406         struct btrfs_key key;
4407         u32 nritems;
4408         int ret;
4409         int i;
4410
4411         nritems = btrfs_header_nritems(path->nodes[0]);
4412         if (nritems == 0)
4413                 return 1;
4414
4415         btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
4416 again:
4417         level = 1;
4418         next = NULL;
4419         btrfs_release_path(path);
4420
4421         path->keep_locks = 1;
4422
4423         if (time_seq)
4424                 ret = btrfs_search_old_slot(root, &key, path, time_seq);
4425         else
4426                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4427         path->keep_locks = 0;
4428
4429         if (ret < 0)
4430                 return ret;
4431
4432         nritems = btrfs_header_nritems(path->nodes[0]);
4433         /*
4434          * by releasing the path above we dropped all our locks.  A balance
4435          * could have added more items next to the key that used to be
4436          * at the very end of the block.  So, check again here and
4437          * advance the path if there are now more items available.
4438          */
4439         if (nritems > 0 && path->slots[0] < nritems - 1) {
4440                 if (ret == 0)
4441                         path->slots[0]++;
4442                 ret = 0;
4443                 goto done;
4444         }
4445         /*
4446          * So the above check misses one case:
4447          * - after releasing the path above, someone has removed the item that
4448          *   used to be at the very end of the block, and balance between leafs
4449          *   gets another one with bigger key.offset to replace it.
4450          *
4451          * This one should be returned as well, or we can get leaf corruption
4452          * later(esp. in __btrfs_drop_extents()).
4453          *
4454          * And a bit more explanation about this check,
4455          * with ret > 0, the key isn't found, the path points to the slot
4456          * where it should be inserted, so the path->slots[0] item must be the
4457          * bigger one.
4458          */
4459         if (nritems > 0 && ret > 0 && path->slots[0] == nritems - 1) {
4460                 ret = 0;
4461                 goto done;
4462         }
4463
4464         while (level < BTRFS_MAX_LEVEL) {
4465                 if (!path->nodes[level]) {
4466                         ret = 1;
4467                         goto done;
4468                 }
4469
4470                 slot = path->slots[level] + 1;
4471                 c = path->nodes[level];
4472                 if (slot >= btrfs_header_nritems(c)) {
4473                         level++;
4474                         if (level == BTRFS_MAX_LEVEL) {
4475                                 ret = 1;
4476                                 goto done;
4477                         }
4478                         continue;
4479                 }
4480
4481
4482                 /*
4483                  * Our current level is where we're going to start from, and to
4484                  * make sure lockdep doesn't complain we need to drop our locks
4485                  * and nodes from 0 to our current level.
4486                  */
4487                 for (i = 0; i < level; i++) {
4488                         if (path->locks[level]) {
4489                                 btrfs_tree_read_unlock(path->nodes[i]);
4490                                 path->locks[i] = 0;
4491                         }
4492                         free_extent_buffer(path->nodes[i]);
4493                         path->nodes[i] = NULL;
4494                 }
4495
4496                 next = c;
4497                 ret = read_block_for_search(root, path, &next, level,
4498                                             slot, &key);
4499                 if (ret == -EAGAIN)
4500                         goto again;
4501
4502                 if (ret < 0) {
4503                         btrfs_release_path(path);
4504                         goto done;
4505                 }
4506
4507                 if (!path->skip_locking) {
4508                         ret = btrfs_try_tree_read_lock(next);
4509                         if (!ret && time_seq) {
4510                                 /*
4511                                  * If we don't get the lock, we may be racing
4512                                  * with push_leaf_left, holding that lock while
4513                                  * itself waiting for the leaf we've currently
4514                                  * locked. To solve this situation, we give up
4515                                  * on our lock and cycle.
4516                                  */
4517                                 free_extent_buffer(next);
4518                                 btrfs_release_path(path);
4519                                 cond_resched();
4520                                 goto again;
4521                         }
4522                         if (!ret)
4523                                 btrfs_tree_read_lock(next);
4524                 }
4525                 break;
4526         }
4527         path->slots[level] = slot;
4528         while (1) {
4529                 level--;
4530                 path->nodes[level] = next;
4531                 path->slots[level] = 0;
4532                 if (!path->skip_locking)
4533                         path->locks[level] = BTRFS_READ_LOCK;
4534                 if (!level)
4535                         break;
4536
4537                 ret = read_block_for_search(root, path, &next, level,
4538                                             0, &key);
4539                 if (ret == -EAGAIN)
4540                         goto again;
4541
4542                 if (ret < 0) {
4543                         btrfs_release_path(path);
4544                         goto done;
4545                 }
4546
4547                 if (!path->skip_locking)
4548                         btrfs_tree_read_lock(next);
4549         }
4550         ret = 0;
4551 done:
4552         unlock_up(path, 0, 1, 0, NULL);
4553
4554         return ret;
4555 }
4556
4557 /*
4558  * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
4559  * searching until it gets past min_objectid or finds an item of 'type'
4560  *
4561  * returns 0 if something is found, 1 if nothing was found and < 0 on error
4562  */
4563 int btrfs_previous_item(struct btrfs_root *root,
4564                         struct btrfs_path *path, u64 min_objectid,
4565                         int type)
4566 {
4567         struct btrfs_key found_key;
4568         struct extent_buffer *leaf;
4569         u32 nritems;
4570         int ret;
4571
4572         while (1) {
4573                 if (path->slots[0] == 0) {
4574                         ret = btrfs_prev_leaf(root, path);
4575                         if (ret != 0)
4576                                 return ret;
4577                 } else {
4578                         path->slots[0]--;
4579                 }
4580                 leaf = path->nodes[0];
4581                 nritems = btrfs_header_nritems(leaf);
4582                 if (nritems == 0)
4583                         return 1;
4584                 if (path->slots[0] == nritems)
4585                         path->slots[0]--;
4586
4587                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4588                 if (found_key.objectid < min_objectid)
4589                         break;
4590                 if (found_key.type == type)
4591                         return 0;
4592                 if (found_key.objectid == min_objectid &&
4593                     found_key.type < type)
4594                         break;
4595         }
4596         return 1;
4597 }
4598
4599 /*
4600  * search in extent tree to find a previous Metadata/Data extent item with
4601  * min objecitd.
4602  *
4603  * returns 0 if something is found, 1 if nothing was found and < 0 on error
4604  */
4605 int btrfs_previous_extent_item(struct btrfs_root *root,
4606                         struct btrfs_path *path, u64 min_objectid)
4607 {
4608         struct btrfs_key found_key;
4609         struct extent_buffer *leaf;
4610         u32 nritems;
4611         int ret;
4612
4613         while (1) {
4614                 if (path->slots[0] == 0) {
4615                         ret = btrfs_prev_leaf(root, path);
4616                         if (ret != 0)
4617                                 return ret;
4618                 } else {
4619                         path->slots[0]--;
4620                 }
4621                 leaf = path->nodes[0];
4622                 nritems = btrfs_header_nritems(leaf);
4623                 if (nritems == 0)
4624                         return 1;
4625                 if (path->slots[0] == nritems)
4626                         path->slots[0]--;
4627
4628                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4629                 if (found_key.objectid < min_objectid)
4630                         break;
4631                 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
4632                     found_key.type == BTRFS_METADATA_ITEM_KEY)
4633                         return 0;
4634                 if (found_key.objectid == min_objectid &&
4635                     found_key.type < BTRFS_EXTENT_ITEM_KEY)
4636                         break;
4637         }
4638         return 1;
4639 }