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 #include "tree-checker.h"
  20
  21 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  22                       *root, struct btrfs_path *path, int level);
  23 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  24                       const struct btrfs_key *ins_key, struct btrfs_path *path,
  25                       int data_size, int extend);
  26 static int push_node_left(struct btrfs_trans_handle *trans,
  27                           struct extent_buffer *dst,
  28                           struct extent_buffer *src, int empty);
  29 static int balance_node_right(struct btrfs_trans_handle *trans,
  30                               struct extent_buffer *dst_buf,
  31                               struct extent_buffer *src_buf);
  32 static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
  33                     int level, int slot);
  34
  35 static const struct btrfs_csums {
  36         u16             size;
  37         const char      name[10];
  38         const char      driver[12];
  39 } btrfs_csums[] = {
  40         [BTRFS_CSUM_TYPE_CRC32] = { .size = 4, .name = "crc32c" },
  41         [BTRFS_CSUM_TYPE_XXHASH] = { .size = 8, .name = "xxhash64" },
  42         [BTRFS_CSUM_TYPE_SHA256] = { .size = 32, .name = "sha256" },
  43         [BTRFS_CSUM_TYPE_BLAKE2] = { .size = 32, .name = "blake2b",
  44                                      .driver = "blake2b-256" },
  45 };
  46
  47 int btrfs_super_csum_size(const struct btrfs_super_block *s)
  48 {
  49         u16 t = btrfs_super_csum_type(s);
  50         /*
  51          * csum type is validated at mount time
  52          */
  53         return btrfs_csums[t].size;
  54 }
  55
  56 const char *btrfs_super_csum_name(u16 csum_type)
  57 {
  58         /* csum type is validated at mount time */
  59         return btrfs_csums[csum_type].name;
  60 }
  61
  62 /*
  63  * Return driver name if defined, otherwise the name that's also a valid driver
  64  * name
  65  */
  66 const char *btrfs_super_csum_driver(u16 csum_type)
  67 {
  68         /* csum type is validated at mount time */
  69         return btrfs_csums[csum_type].driver[0] ?
  70                 btrfs_csums[csum_type].driver :
  71                 btrfs_csums[csum_type].name;
  72 }
  73
  74 size_t __attribute_const__ btrfs_get_num_csums(void)
  75 {
  76         return ARRAY_SIZE(btrfs_csums);
  77 }
  78
  79 struct btrfs_path *btrfs_alloc_path(void)
  80 {
  81         return kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
  82 }
  83
  84 /* this also releases the path */
  85 void btrfs_free_path(struct btrfs_path *p)
  86 {
  87         if (!p)
  88                 return;
  89         btrfs_release_path(p);
  90         kmem_cache_free(btrfs_path_cachep, p);
  91 }
  92
  93 /*
  94  * path release drops references on the extent buffers in the path
  95  * and it drops any locks held by this path
  96  *
  97  * It is safe to call this on paths that no locks or extent buffers held.
  98  */
  99 noinline void btrfs_release_path(struct btrfs_path *p)
 100 {
 101         int i;
 102
 103         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
 104                 p->slots[i] = 0;
 105                 if (!p->nodes[i])
 106                         continue;
 107                 if (p->locks[i]) {
 108                         btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
 109                         p->locks[i] = 0;
 110                 }
 111                 free_extent_buffer(p->nodes[i]);
 112                 p->nodes[i] = NULL;
 113         }
 114 }
 115
 116 /*
 117  * safely gets a reference on the root node of a tree.  A lock
 118  * is not taken, so a concurrent writer may put a different node
 119  * at the root of the tree.  See btrfs_lock_root_node for the
 120  * looping required.
 121  *
 122  * The extent buffer returned by this has a reference taken, so
 123  * it won't disappear.  It may stop being the root of the tree
 124  * at any time because there are no locks held.
 125  */
 126 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
 127 {
 128         struct extent_buffer *eb;
 129
 130         while (1) {
 131                 rcu_read_lock();
 132                 eb = rcu_dereference(root->node);
 133
 134                 /*
 135                  * RCU really hurts here, we could free up the root node because
 136                  * it was COWed but we may not get the new root node yet so do
 137                  * the inc_not_zero dance and if it doesn't work then
 138                  * synchronize_rcu and try again.
 139                  */
 140                 if (atomic_inc_not_zero(&eb->refs)) {
 141                         rcu_read_unlock();
 142                         break;
 143                 }
 144                 rcu_read_unlock();
 145                 synchronize_rcu();
 146         }
 147         return eb;
 148 }
 149
 150 /*
 151  * Cowonly root (not-shareable trees, everything not subvolume or reloc roots),
 152  * just get put onto a simple dirty list.  Transaction walks this list to make
 153  * sure they get properly updated on disk.
 154  */
 155 static void add_root_to_dirty_list(struct btrfs_root *root)
 156 {
 157         struct btrfs_fs_info *fs_info = root->fs_info;
 158
 159         if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
 160             !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
 161                 return;
 162
 163         spin_lock(&fs_info->trans_lock);
 164         if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
 165                 /* Want the extent tree to be the last on the list */
 166                 if (root->root_key.objectid == BTRFS_EXTENT_TREE_OBJECTID)
 167                         list_move_tail(&root->dirty_list,
 168                                        &fs_info->dirty_cowonly_roots);
 169                 else
 170                         list_move(&root->dirty_list,
 171                                   &fs_info->dirty_cowonly_roots);
 172         }
 173         spin_unlock(&fs_info->trans_lock);
 174 }
 175
 176 /*
 177  * used by snapshot creation to make a copy of a root for a tree with
 178  * a given objectid.  The buffer with the new root node is returned in
 179  * cow_ret, and this func returns zero on success or a negative error code.
 180  */
 181 int btrfs_copy_root(struct btrfs_trans_handle *trans,
 182                       struct btrfs_root *root,
 183                       struct extent_buffer *buf,
 184                       struct extent_buffer **cow_ret, u64 new_root_objectid)
 185 {
 186         struct btrfs_fs_info *fs_info = root->fs_info;
 187         struct extent_buffer *cow;
 188         int ret = 0;
 189         int level;
 190         struct btrfs_disk_key disk_key;
 191
 192         WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 193                 trans->transid != fs_info->running_transaction->transid);
 194         WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 195                 trans->transid != root->last_trans);
 196
 197         level = btrfs_header_level(buf);
 198         if (level == 0)
 199                 btrfs_item_key(buf, &disk_key, 0);
 200         else
 201                 btrfs_node_key(buf, &disk_key, 0);
 202
 203         cow = btrfs_alloc_tree_block(trans, root, 0, new_root_objectid,
 204                                      &disk_key, level, buf->start, 0,
 205                                      BTRFS_NESTING_NEW_ROOT);
 206         if (IS_ERR(cow))
 207                 return PTR_ERR(cow);
 208
 209         copy_extent_buffer_full(cow, buf);
 210         btrfs_set_header_bytenr(cow, cow->start);
 211         btrfs_set_header_generation(cow, trans->transid);
 212         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 213         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 214                                      BTRFS_HEADER_FLAG_RELOC);
 215         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 216                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 217         else
 218                 btrfs_set_header_owner(cow, new_root_objectid);
 219
 220         write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
 221
 222         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 223         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 224                 ret = btrfs_inc_ref(trans, root, cow, 1);
 225         else
 226                 ret = btrfs_inc_ref(trans, root, cow, 0);
 227         if (ret) {
 228                 btrfs_tree_unlock(cow);
 229                 free_extent_buffer(cow);
 230                 btrfs_abort_transaction(trans, ret);
 231                 return ret;
 232         }
 233
 234         btrfs_mark_buffer_dirty(cow);
 235         *cow_ret = cow;
 236         return 0;
 237 }
 238
 239 /*
 240  * check if the tree block can be shared by multiple trees
 241  */
 242 int btrfs_block_can_be_shared(struct btrfs_root *root,
 243                               struct extent_buffer *buf)
 244 {
 245         /*
 246          * Tree blocks not in shareable trees and tree roots are never shared.
 247          * If a block was allocated after the last snapshot and the block was
 248          * not allocated by tree relocation, we know the block is not shared.
 249          */
 250         if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 251             buf != root->node && buf != root->commit_root &&
 252             (btrfs_header_generation(buf) <=
 253              btrfs_root_last_snapshot(&root->root_item) ||
 254              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 255                 return 1;
 256
 257         return 0;
 258 }
 259
 260 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
 261                                        struct btrfs_root *root,
 262                                        struct extent_buffer *buf,
 263                                        struct extent_buffer *cow,
 264                                        int *last_ref)
 265 {
 266         struct btrfs_fs_info *fs_info = root->fs_info;
 267         u64 refs;
 268         u64 owner;
 269         u64 flags;
 270         u64 new_flags = 0;
 271         int ret;
 272
 273         /*
 274          * Backrefs update rules:
 275          *
 276          * Always use full backrefs for extent pointers in tree block
 277          * allocated by tree relocation.
 278          *
 279          * If a shared tree block is no longer referenced by its owner
 280          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
 281          * use full backrefs for extent pointers in tree block.
 282          *
 283          * If a tree block is been relocating
 284          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
 285          * use full backrefs for extent pointers in tree block.
 286          * The reason for this is some operations (such as drop tree)
 287          * are only allowed for blocks use full backrefs.
 288          */
 289
 290         if (btrfs_block_can_be_shared(root, buf)) {
 291                 ret = btrfs_lookup_extent_info(trans, fs_info, buf->start,
 292                                                btrfs_header_level(buf), 1,
 293                                                &refs, &flags);
 294                 if (ret)
 295                         return ret;
 296                 if (refs == 0) {
 297                         ret = -EROFS;
 298                         btrfs_handle_fs_error(fs_info, ret, NULL);
 299                         return ret;
 300                 }
 301         } else {
 302                 refs = 1;
 303                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 304                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 305                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 306                 else
 307                         flags = 0;
 308         }
 309
 310         owner = btrfs_header_owner(buf);
 311         BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
 312                !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
 313
 314         if (refs > 1) {
 315                 if ((owner == root->root_key.objectid ||
 316                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
 317                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
 318                         ret = btrfs_inc_ref(trans, root, buf, 1);
 319                         if (ret)
 320                                 return ret;
 321
 322                         if (root->root_key.objectid ==
 323                             BTRFS_TREE_RELOC_OBJECTID) {
 324                                 ret = btrfs_dec_ref(trans, root, buf, 0);
 325                                 if (ret)
 326                                         return ret;
 327                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 328                                 if (ret)
 329                                         return ret;
 330                         }
 331                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
 332                 } else {
 333
 334                         if (root->root_key.objectid ==
 335                             BTRFS_TREE_RELOC_OBJECTID)
 336                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 337                         else
 338                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 339                         if (ret)
 340                                 return ret;
 341                 }
 342                 if (new_flags != 0) {
 343                         int level = btrfs_header_level(buf);
 344
 345                         ret = btrfs_set_disk_extent_flags(trans, buf,
 346                                                           new_flags, level);
 347                         if (ret)
 348                                 return ret;
 349                 }
 350         } else {
 351                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
 352                         if (root->root_key.objectid ==
 353                             BTRFS_TREE_RELOC_OBJECTID)
 354                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 355                         else
 356                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 357                         if (ret)
 358                                 return ret;
 359                         ret = btrfs_dec_ref(trans, root, buf, 1);
 360                         if (ret)
 361                                 return ret;
 362                 }
 363                 btrfs_clean_tree_block(buf);
 364                 *last_ref = 1;
 365         }
 366         return 0;
 367 }
 368
 369 /*
 370  * does the dirty work in cow of a single block.  The parent block (if
 371  * supplied) is updated to point to the new cow copy.  The new buffer is marked
 372  * dirty and returned locked.  If you modify the block it needs to be marked
 373  * dirty again.
 374  *
 375  * search_start -- an allocation hint for the new block
 376  *
 377  * empty_size -- a hint that you plan on doing more cow.  This is the size in
 378  * bytes the allocator should try to find free next to the block it returns.
 379  * This is just a hint and may be ignored by the allocator.
 380  */
 381 static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 382                              struct btrfs_root *root,
 383                              struct extent_buffer *buf,
 384                              struct extent_buffer *parent, int parent_slot,
 385                              struct extent_buffer **cow_ret,
 386                              u64 search_start, u64 empty_size,
 387                              enum btrfs_lock_nesting nest)
 388 {
 389         struct btrfs_fs_info *fs_info = root->fs_info;
 390         struct btrfs_disk_key disk_key;
 391         struct extent_buffer *cow;
 392         int level, ret;
 393         int last_ref = 0;
 394         int unlock_orig = 0;
 395         u64 parent_start = 0;
 396
 397         if (*cow_ret == buf)
 398                 unlock_orig = 1;
 399
 400         btrfs_assert_tree_write_locked(buf);
 401
 402         WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 403                 trans->transid != fs_info->running_transaction->transid);
 404         WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 405                 trans->transid != root->last_trans);
 406
 407         level = btrfs_header_level(buf);
 408
 409         if (level == 0)
 410                 btrfs_item_key(buf, &disk_key, 0);
 411         else
 412                 btrfs_node_key(buf, &disk_key, 0);
 413
 414         if ((root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) && parent)
 415                 parent_start = parent->start;
 416
 417         cow = btrfs_alloc_tree_block(trans, root, parent_start,
 418                                      root->root_key.objectid, &disk_key, level,
 419                                      search_start, empty_size, nest);
 420         if (IS_ERR(cow))
 421                 return PTR_ERR(cow);
 422
 423         /* cow is set to blocking by btrfs_init_new_buffer */
 424
 425         copy_extent_buffer_full(cow, buf);
 426         btrfs_set_header_bytenr(cow, cow->start);
 427         btrfs_set_header_generation(cow, trans->transid);
 428         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 429         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 430                                      BTRFS_HEADER_FLAG_RELOC);
 431         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 432                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 433         else
 434                 btrfs_set_header_owner(cow, root->root_key.objectid);
 435
 436         write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
 437
 438         ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
 439         if (ret) {
 440                 btrfs_tree_unlock(cow);
 441                 free_extent_buffer(cow);
 442                 btrfs_abort_transaction(trans, ret);
 443                 return ret;
 444         }
 445
 446         if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
 447                 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
 448                 if (ret) {
 449                         btrfs_tree_unlock(cow);
 450                         free_extent_buffer(cow);
 451                         btrfs_abort_transaction(trans, ret);
 452                         return ret;
 453                 }
 454         }
 455
 456         if (buf == root->node) {
 457                 WARN_ON(parent && parent != buf);
 458                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 459                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 460                         parent_start = buf->start;
 461
 462                 atomic_inc(&cow->refs);
 463                 ret = btrfs_tree_mod_log_insert_root(root->node, cow, true);
 464                 BUG_ON(ret < 0);
 465                 rcu_assign_pointer(root->node, cow);
 466
 467                 btrfs_free_tree_block(trans, btrfs_root_id(root), buf,
 468                                       parent_start, last_ref);
 469                 free_extent_buffer(buf);
 470                 add_root_to_dirty_list(root);
 471         } else {
 472                 WARN_ON(trans->transid != btrfs_header_generation(parent));
 473                 btrfs_tree_mod_log_insert_key(parent, parent_slot,
 474                                               BTRFS_MOD_LOG_KEY_REPLACE, GFP_NOFS);
 475                 btrfs_set_node_blockptr(parent, parent_slot,
 476                                         cow->start);
 477                 btrfs_set_node_ptr_generation(parent, parent_slot,
 478                                               trans->transid);
 479                 btrfs_mark_buffer_dirty(parent);
 480                 if (last_ref) {
 481                         ret = btrfs_tree_mod_log_free_eb(buf);
 482                         if (ret) {
 483                                 btrfs_tree_unlock(cow);
 484                                 free_extent_buffer(cow);
 485                                 btrfs_abort_transaction(trans, ret);
 486                                 return ret;
 487                         }
 488                 }
 489                 btrfs_free_tree_block(trans, btrfs_root_id(root), buf,
 490                                       parent_start, last_ref);
 491         }
 492         if (unlock_orig)
 493                 btrfs_tree_unlock(buf);
 494         free_extent_buffer_stale(buf);
 495         btrfs_mark_buffer_dirty(cow);
 496         *cow_ret = cow;
 497         return 0;
 498 }
 499
 500 static inline int should_cow_block(struct btrfs_trans_handle *trans,
 501                                    struct btrfs_root *root,
 502                                    struct extent_buffer *buf)
 503 {
 504         if (btrfs_is_testing(root->fs_info))
 505                 return 0;
 506
 507         /* Ensure we can see the FORCE_COW bit */
 508         smp_mb__before_atomic();
 509
 510         /*
 511          * We do not need to cow a block if
 512          * 1) this block is not created or changed in this transaction;
 513          * 2) this block does not belong to TREE_RELOC tree;
 514          * 3) the root is not forced COW.
 515          *
 516          * What is forced COW:
 517          *    when we create snapshot during committing the transaction,
 518          *    after we've finished copying src root, we must COW the shared
 519          *    block to ensure the metadata consistency.
 520          */
 521         if (btrfs_header_generation(buf) == trans->transid &&
 522             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
 523             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
 524               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
 525             !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
 526                 return 0;
 527         return 1;
 528 }
 529
 530 /*
 531  * cows a single block, see __btrfs_cow_block for the real work.
 532  * This version of it has extra checks so that a block isn't COWed more than
 533  * once per transaction, as long as it hasn't been written yet
 534  */
 535 noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
 536                     struct btrfs_root *root, struct extent_buffer *buf,
 537                     struct extent_buffer *parent, int parent_slot,
 538                     struct extent_buffer **cow_ret,
 539                     enum btrfs_lock_nesting nest)
 540 {
 541         struct btrfs_fs_info *fs_info = root->fs_info;
 542         u64 search_start;
 543         int ret;
 544
 545         if (test_bit(BTRFS_ROOT_DELETING, &root->state))
 546                 btrfs_err(fs_info,
 547                         "COW'ing blocks on a fs root that's being dropped");
 548
 549         if (trans->transaction != fs_info->running_transaction)
 550                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
 551                        trans->transid,
 552                        fs_info->running_transaction->transid);
 553
 554         if (trans->transid != fs_info->generation)
 555                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
 556                        trans->transid, fs_info->generation);
 557
 558         if (!should_cow_block(trans, root, buf)) {
 559                 *cow_ret = buf;
 560                 return 0;
 561         }
 562
 563         search_start = buf->start & ~((u64)SZ_1G - 1);
 564
 565         /*
 566          * Before CoWing this block for later modification, check if it's
 567          * the subtree root and do the delayed subtree trace if needed.
 568          *
 569          * Also We don't care about the error, as it's handled internally.
 570          */
 571         btrfs_qgroup_trace_subtree_after_cow(trans, root, buf);
 572         ret = __btrfs_cow_block(trans, root, buf, parent,
 573                                  parent_slot, cow_ret, search_start, 0, nest);
 574
 575         trace_btrfs_cow_block(root, buf, *cow_ret);
 576
 577         return ret;
 578 }
 579 ALLOW_ERROR_INJECTION(btrfs_cow_block, ERRNO);
 580
 581 /*
 582  * helper function for defrag to decide if two blocks pointed to by a
 583  * node are actually close by
 584  */
 585 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 586 {
 587         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 588                 return 1;
 589         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 590                 return 1;
 591         return 0;
 592 }
 593
 594 #ifdef __LITTLE_ENDIAN
 595
 596 /*
 597  * Compare two keys, on little-endian the disk order is same as CPU order and
 598  * we can avoid the conversion.
 599  */
 600 static int comp_keys(const struct btrfs_disk_key *disk_key,
 601                      const struct btrfs_key *k2)
 602 {
 603         const struct btrfs_key *k1 = (const struct btrfs_key *)disk_key;
 604
 605         return btrfs_comp_cpu_keys(k1, k2);
 606 }
 607
 608 #else
 609
 610 /*
 611  * compare two keys in a memcmp fashion
 612  */
 613 static int comp_keys(const struct btrfs_disk_key *disk,
 614                      const struct btrfs_key *k2)
 615 {
 616         struct btrfs_key k1;
 617
 618         btrfs_disk_key_to_cpu(&k1, disk);
 619
 620         return btrfs_comp_cpu_keys(&k1, k2);
 621 }
 622 #endif
 623
 624 /*
 625  * same as comp_keys only with two btrfs_key's
 626  */
 627 int __pure btrfs_comp_cpu_keys(const struct btrfs_key *k1, const struct btrfs_key *k2)
 628 {
 629         if (k1->objectid > k2->objectid)
 630                 return 1;
 631         if (k1->objectid < k2->objectid)
 632                 return -1;
 633         if (k1->type > k2->type)
 634                 return 1;
 635         if (k1->type < k2->type)
 636                 return -1;
 637         if (k1->offset > k2->offset)
 638                 return 1;
 639         if (k1->offset < k2->offset)
 640                 return -1;
 641         return 0;
 642 }
 643
 644 /*
 645  * this is used by the defrag code to go through all the
 646  * leaves pointed to by a node and reallocate them so that
 647  * disk order is close to key order
 648  */
 649 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 650                        struct btrfs_root *root, struct extent_buffer *parent,
 651                        int start_slot, u64 *last_ret,
 652                        struct btrfs_key *progress)
 653 {
 654         struct btrfs_fs_info *fs_info = root->fs_info;
 655         struct extent_buffer *cur;
 656         u64 blocknr;
 657         u64 search_start = *last_ret;
 658         u64 last_block = 0;
 659         u64 other;
 660         u32 parent_nritems;
 661         int end_slot;
 662         int i;
 663         int err = 0;
 664         u32 blocksize;
 665         int progress_passed = 0;
 666         struct btrfs_disk_key disk_key;
 667
 668         WARN_ON(trans->transaction != fs_info->running_transaction);
 669         WARN_ON(trans->transid != fs_info->generation);
 670
 671         parent_nritems = btrfs_header_nritems(parent);
 672         blocksize = fs_info->nodesize;
 673         end_slot = parent_nritems - 1;
 674
 675         if (parent_nritems <= 1)
 676                 return 0;
 677
 678         for (i = start_slot; i <= end_slot; i++) {
 679                 int close = 1;
 680
 681                 btrfs_node_key(parent, &disk_key, i);
 682                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
 683                         continue;
 684
 685                 progress_passed = 1;
 686                 blocknr = btrfs_node_blockptr(parent, i);
 687                 if (last_block == 0)
 688                         last_block = blocknr;
 689
 690                 if (i > 0) {
 691                         other = btrfs_node_blockptr(parent, i - 1);
 692                         close = close_blocks(blocknr, other, blocksize);
 693                 }
 694                 if (!close && i < end_slot) {
 695                         other = btrfs_node_blockptr(parent, i + 1);
 696                         close = close_blocks(blocknr, other, blocksize);
 697                 }
 698                 if (close) {
 699                         last_block = blocknr;
 700                         continue;
 701                 }
 702
 703                 cur = btrfs_read_node_slot(parent, i);
 704                 if (IS_ERR(cur))
 705                         return PTR_ERR(cur);
 706                 if (search_start == 0)
 707                         search_start = last_block;
 708
 709                 btrfs_tree_lock(cur);
 710                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 711                                         &cur, search_start,
 712                                         min(16 * blocksize,
 713                                             (end_slot - i) * blocksize),
 714                                         BTRFS_NESTING_COW);
 715                 if (err) {
 716                         btrfs_tree_unlock(cur);
 717                         free_extent_buffer(cur);
 718                         break;
 719                 }
 720                 search_start = cur->start;
 721                 last_block = cur->start;
 722                 *last_ret = search_start;
 723                 btrfs_tree_unlock(cur);
 724                 free_extent_buffer(cur);
 725         }
 726         return err;
 727 }
 728
 729 /*
 730  * Search for a key in the given extent_buffer.
 731  *
 732  * The lower boundary for the search is specified by the slot number @low. Use a
 733  * value of 0 to search over the whole extent buffer.
 734  *
 735  * The slot in the extent buffer is returned via @slot. If the key exists in the
 736  * extent buffer, then @slot will point to the slot where the key is, otherwise
 737  * it points to the slot where you would insert the key.
 738  *
 739  * Slot may point to the total number of items (i.e. one position beyond the last
 740  * key) if the key is bigger than the last key in the extent buffer.
 741  */
 742 static noinline int generic_bin_search(struct extent_buffer *eb, int low,
 743                                        const struct btrfs_key *key, int *slot)
 744 {
 745         unsigned long p;
 746         int item_size;
 747         int high = btrfs_header_nritems(eb);
 748         int ret;
 749         const int key_size = sizeof(struct btrfs_disk_key);
 750
 751         if (low > high) {
 752                 btrfs_err(eb->fs_info,
 753                  "%s: low (%d) > high (%d) eb %llu owner %llu level %d",
 754                           __func__, low, high, eb->start,
 755                           btrfs_header_owner(eb), btrfs_header_level(eb));
 756                 return -EINVAL;
 757         }
 758
 759         if (btrfs_header_level(eb) == 0) {
 760                 p = offsetof(struct btrfs_leaf, items);
 761                 item_size = sizeof(struct btrfs_item);
 762         } else {
 763                 p = offsetof(struct btrfs_node, ptrs);
 764                 item_size = sizeof(struct btrfs_key_ptr);
 765         }
 766
 767         while (low < high) {
 768                 unsigned long oip;
 769                 unsigned long offset;
 770                 struct btrfs_disk_key *tmp;
 771                 struct btrfs_disk_key unaligned;
 772                 int mid;
 773
 774                 mid = (low + high) / 2;
 775                 offset = p + mid * item_size;
 776                 oip = offset_in_page(offset);
 777
 778                 if (oip + key_size <= PAGE_SIZE) {
 779                         const unsigned long idx = get_eb_page_index(offset);
 780                         char *kaddr = page_address(eb->pages[idx]);
 781
 782                         oip = get_eb_offset_in_page(eb, offset);
 783                         tmp = (struct btrfs_disk_key *)(kaddr + oip);
 784                 } else {
 785                         read_extent_buffer(eb, &unaligned, offset, key_size);
 786                         tmp = &unaligned;
 787                 }
 788
 789                 ret = comp_keys(tmp, key);
 790
 791                 if (ret < 0)
 792                         low = mid + 1;
 793                 else if (ret > 0)
 794                         high = mid;
 795                 else {
 796                         *slot = mid;
 797                         return 0;
 798                 }
 799         }
 800         *slot = low;
 801         return 1;
 802 }
 803
 804 /*
 805  * Simple binary search on an extent buffer. Works for both leaves and nodes, and
 806  * always searches over the whole range of keys (slot 0 to slot 'nritems - 1').
 807  */
 808 int btrfs_bin_search(struct extent_buffer *eb, const struct btrfs_key *key,
 809                      int *slot)
 810 {
 811         return generic_bin_search(eb, 0, key, slot);
 812 }
 813
 814 static void root_add_used(struct btrfs_root *root, u32 size)
 815 {
 816         spin_lock(&root->accounting_lock);
 817         btrfs_set_root_used(&root->root_item,
 818                             btrfs_root_used(&root->root_item) + size);
 819         spin_unlock(&root->accounting_lock);
 820 }
 821
 822 static void root_sub_used(struct btrfs_root *root, u32 size)
 823 {
 824         spin_lock(&root->accounting_lock);
 825         btrfs_set_root_used(&root->root_item,
 826                             btrfs_root_used(&root->root_item) - size);
 827         spin_unlock(&root->accounting_lock);
 828 }
 829
 830 /* given a node and slot number, this reads the blocks it points to.  The
 831  * extent buffer is returned with a reference taken (but unlocked).
 832  */
 833 struct extent_buffer *btrfs_read_node_slot(struct extent_buffer *parent,
 834                                            int slot)
 835 {
 836         int level = btrfs_header_level(parent);
 837         struct extent_buffer *eb;
 838         struct btrfs_key first_key;
 839
 840         if (slot < 0 || slot >= btrfs_header_nritems(parent))
 841                 return ERR_PTR(-ENOENT);
 842
 843         BUG_ON(level == 0);
 844
 845         btrfs_node_key_to_cpu(parent, &first_key, slot);
 846         eb = read_tree_block(parent->fs_info, btrfs_node_blockptr(parent, slot),
 847                              btrfs_header_owner(parent),
 848                              btrfs_node_ptr_generation(parent, slot),
 849                              level - 1, &first_key);
 850         if (IS_ERR(eb))
 851                 return eb;
 852         if (!extent_buffer_uptodate(eb)) {
 853                 free_extent_buffer(eb);
 854                 return ERR_PTR(-EIO);
 855         }
 856
 857         return eb;
 858 }
 859
 860 /*
 861  * node level balancing, used to make sure nodes are in proper order for
 862  * item deletion.  We balance from the top down, so we have to make sure
 863  * that a deletion won't leave an node completely empty later on.
 864  */
 865 static noinline int balance_level(struct btrfs_trans_handle *trans,
 866                          struct btrfs_root *root,
 867                          struct btrfs_path *path, int level)
 868 {
 869         struct btrfs_fs_info *fs_info = root->fs_info;
 870         struct extent_buffer *right = NULL;
 871         struct extent_buffer *mid;
 872         struct extent_buffer *left = NULL;
 873         struct extent_buffer *parent = NULL;
 874         int ret = 0;
 875         int wret;
 876         int pslot;
 877         int orig_slot = path->slots[level];
 878         u64 orig_ptr;
 879
 880         ASSERT(level > 0);
 881
 882         mid = path->nodes[level];
 883
 884         WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK);
 885         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 886
 887         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 888
 889         if (level < BTRFS_MAX_LEVEL - 1) {
 890                 parent = path->nodes[level + 1];
 891                 pslot = path->slots[level + 1];
 892         }
 893
 894         /*
 895          * deal with the case where there is only one pointer in the root
 896          * by promoting the node below to a root
 897          */
 898         if (!parent) {
 899                 struct extent_buffer *child;
 900
 901                 if (btrfs_header_nritems(mid) != 1)
 902                         return 0;
 903
 904                 /* promote the child to a root */
 905                 child = btrfs_read_node_slot(mid, 0);
 906                 if (IS_ERR(child)) {
 907                         ret = PTR_ERR(child);
 908                         btrfs_handle_fs_error(fs_info, ret, NULL);
 909                         goto enospc;
 910                 }
 911
 912                 btrfs_tree_lock(child);
 913                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child,
 914                                       BTRFS_NESTING_COW);
 915                 if (ret) {
 916                         btrfs_tree_unlock(child);
 917                         free_extent_buffer(child);
 918                         goto enospc;
 919                 }
 920
 921                 ret = btrfs_tree_mod_log_insert_root(root->node, child, true);
 922                 BUG_ON(ret < 0);
 923                 rcu_assign_pointer(root->node, child);
 924
 925                 add_root_to_dirty_list(root);
 926                 btrfs_tree_unlock(child);
 927
 928                 path->locks[level] = 0;
 929                 path->nodes[level] = NULL;
 930                 btrfs_clean_tree_block(mid);
 931                 btrfs_tree_unlock(mid);
 932                 /* once for the path */
 933                 free_extent_buffer(mid);
 934
 935                 root_sub_used(root, mid->len);
 936                 btrfs_free_tree_block(trans, btrfs_root_id(root), mid, 0, 1);
 937                 /* once for the root ptr */
 938                 free_extent_buffer_stale(mid);
 939                 return 0;
 940         }
 941         if (btrfs_header_nritems(mid) >
 942             BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 4)
 943                 return 0;
 944
 945         left = btrfs_read_node_slot(parent, pslot - 1);
 946         if (IS_ERR(left))
 947                 left = NULL;
 948
 949         if (left) {
 950                 __btrfs_tree_lock(left, BTRFS_NESTING_LEFT);
 951                 wret = btrfs_cow_block(trans, root, left,
 952                                        parent, pslot - 1, &left,
 953                                        BTRFS_NESTING_LEFT_COW);
 954                 if (wret) {
 955                         ret = wret;
 956                         goto enospc;
 957                 }
 958         }
 959
 960         right = btrfs_read_node_slot(parent, pslot + 1);
 961         if (IS_ERR(right))
 962                 right = NULL;
 963
 964         if (right) {
 965                 __btrfs_tree_lock(right, BTRFS_NESTING_RIGHT);
 966                 wret = btrfs_cow_block(trans, root, right,
 967                                        parent, pslot + 1, &right,
 968                                        BTRFS_NESTING_RIGHT_COW);
 969                 if (wret) {
 970                         ret = wret;
 971                         goto enospc;
 972                 }
 973         }
 974
 975         /* first, try to make some room in the middle buffer */
 976         if (left) {
 977                 orig_slot += btrfs_header_nritems(left);
 978                 wret = push_node_left(trans, left, mid, 1);
 979                 if (wret < 0)
 980                         ret = wret;
 981         }
 982
 983         /*
 984          * then try to empty the right most buffer into the middle
 985          */
 986         if (right) {
 987                 wret = push_node_left(trans, mid, right, 1);
 988                 if (wret < 0 && wret != -ENOSPC)
 989                         ret = wret;
 990                 if (btrfs_header_nritems(right) == 0) {
 991                         btrfs_clean_tree_block(right);
 992                         btrfs_tree_unlock(right);
 993                         del_ptr(root, path, level + 1, pslot + 1);
 994                         root_sub_used(root, right->len);
 995                         btrfs_free_tree_block(trans, btrfs_root_id(root), right,
 996                                               0, 1);
 997                         free_extent_buffer_stale(right);
 998                         right = NULL;
 999                 } else {
1000                         struct btrfs_disk_key right_key;
1001                         btrfs_node_key(right, &right_key, 0);
1002                         ret = btrfs_tree_mod_log_insert_key(parent, pslot + 1,
1003                                         BTRFS_MOD_LOG_KEY_REPLACE, GFP_NOFS);
1004                         BUG_ON(ret < 0);
1005                         btrfs_set_node_key(parent, &right_key, pslot + 1);
1006                         btrfs_mark_buffer_dirty(parent);
1007                 }
1008         }
1009         if (btrfs_header_nritems(mid) == 1) {
1010                 /*
1011                  * we're not allowed to leave a node with one item in the
1012                  * tree during a delete.  A deletion from lower in the tree
1013                  * could try to delete the only pointer in this node.
1014                  * So, pull some keys from the left.
1015                  * There has to be a left pointer at this point because
1016                  * otherwise we would have pulled some pointers from the
1017                  * right
1018                  */
1019                 if (!left) {
1020                         ret = -EROFS;
1021                         btrfs_handle_fs_error(fs_info, ret, NULL);
1022                         goto enospc;
1023                 }
1024                 wret = balance_node_right(trans, mid, left);
1025                 if (wret < 0) {
1026                         ret = wret;
1027                         goto enospc;
1028                 }
1029                 if (wret == 1) {
1030                         wret = push_node_left(trans, left, mid, 1);
1031                         if (wret < 0)
1032                                 ret = wret;
1033                 }
1034                 BUG_ON(wret == 1);
1035         }
1036         if (btrfs_header_nritems(mid) == 0) {
1037                 btrfs_clean_tree_block(mid);
1038                 btrfs_tree_unlock(mid);
1039                 del_ptr(root, path, level + 1, pslot);
1040                 root_sub_used(root, mid->len);
1041                 btrfs_free_tree_block(trans, btrfs_root_id(root), mid, 0, 1);
1042                 free_extent_buffer_stale(mid);
1043                 mid = NULL;
1044         } else {
1045                 /* update the parent key to reflect our changes */
1046                 struct btrfs_disk_key mid_key;
1047                 btrfs_node_key(mid, &mid_key, 0);
1048                 ret = btrfs_tree_mod_log_insert_key(parent, pslot,
1049                                 BTRFS_MOD_LOG_KEY_REPLACE, GFP_NOFS);
1050                 BUG_ON(ret < 0);
1051                 btrfs_set_node_key(parent, &mid_key, pslot);
1052                 btrfs_mark_buffer_dirty(parent);
1053         }
1054
1055         /* update the path */
1056         if (left) {
1057                 if (btrfs_header_nritems(left) > orig_slot) {
1058                         atomic_inc(&left->refs);
1059                         /* left was locked after cow */
1060                         path->nodes[level] = left;
1061                         path->slots[level + 1] -= 1;
1062                         path->slots[level] = orig_slot;
1063                         if (mid) {
1064                                 btrfs_tree_unlock(mid);
1065                                 free_extent_buffer(mid);
1066                         }
1067                 } else {
1068                         orig_slot -= btrfs_header_nritems(left);
1069                         path->slots[level] = orig_slot;
1070                 }
1071         }
1072         /* double check we haven't messed things up */
1073         if (orig_ptr !=
1074             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
1075                 BUG();
1076 enospc:
1077         if (right) {
1078                 btrfs_tree_unlock(right);
1079                 free_extent_buffer(right);
1080         }
1081         if (left) {
1082                 if (path->nodes[level] != left)
1083                         btrfs_tree_unlock(left);
1084                 free_extent_buffer(left);
1085         }
1086         return ret;
1087 }
1088
1089 /* Node balancing for insertion.  Here we only split or push nodes around
1090  * when they are completely full.  This is also done top down, so we
1091  * have to be pessimistic.
1092  */
1093 static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
1094                                           struct btrfs_root *root,
1095                                           struct btrfs_path *path, int level)
1096 {
1097         struct btrfs_fs_info *fs_info = root->fs_info;
1098         struct extent_buffer *right = NULL;
1099         struct extent_buffer *mid;
1100         struct extent_buffer *left = NULL;
1101         struct extent_buffer *parent = NULL;
1102         int ret = 0;
1103         int wret;
1104         int pslot;
1105         int orig_slot = path->slots[level];
1106
1107         if (level == 0)
1108                 return 1;
1109
1110         mid = path->nodes[level];
1111         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1112
1113         if (level < BTRFS_MAX_LEVEL - 1) {
1114                 parent = path->nodes[level + 1];
1115                 pslot = path->slots[level + 1];
1116         }
1117
1118         if (!parent)
1119                 return 1;
1120
1121         left = btrfs_read_node_slot(parent, pslot - 1);
1122         if (IS_ERR(left))
1123                 left = NULL;
1124
1125         /* first, try to make some room in the middle buffer */
1126         if (left) {
1127                 u32 left_nr;
1128
1129                 __btrfs_tree_lock(left, BTRFS_NESTING_LEFT);
1130
1131                 left_nr = btrfs_header_nritems(left);
1132                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
1133                         wret = 1;
1134                 } else {
1135                         ret = btrfs_cow_block(trans, root, left, parent,
1136                                               pslot - 1, &left,
1137                                               BTRFS_NESTING_LEFT_COW);
1138                         if (ret)
1139                                 wret = 1;
1140                         else {
1141                                 wret = push_node_left(trans, left, mid, 0);
1142                         }
1143                 }
1144                 if (wret < 0)
1145                         ret = wret;
1146                 if (wret == 0) {
1147                         struct btrfs_disk_key disk_key;
1148                         orig_slot += left_nr;
1149                         btrfs_node_key(mid, &disk_key, 0);
1150                         ret = btrfs_tree_mod_log_insert_key(parent, pslot,
1151                                         BTRFS_MOD_LOG_KEY_REPLACE, GFP_NOFS);
1152                         BUG_ON(ret < 0);
1153                         btrfs_set_node_key(parent, &disk_key, pslot);
1154                         btrfs_mark_buffer_dirty(parent);
1155                         if (btrfs_header_nritems(left) > orig_slot) {
1156                                 path->nodes[level] = left;
1157                                 path->slots[level + 1] -= 1;
1158                                 path->slots[level] = orig_slot;
1159                                 btrfs_tree_unlock(mid);
1160                                 free_extent_buffer(mid);
1161                         } else {
1162                                 orig_slot -=
1163                                         btrfs_header_nritems(left);
1164                                 path->slots[level] = orig_slot;
1165                                 btrfs_tree_unlock(left);
1166                                 free_extent_buffer(left);
1167                         }
1168                         return 0;
1169                 }
1170                 btrfs_tree_unlock(left);
1171                 free_extent_buffer(left);
1172         }
1173         right = btrfs_read_node_slot(parent, pslot + 1);
1174         if (IS_ERR(right))
1175                 right = NULL;
1176
1177         /*
1178          * then try to empty the right most buffer into the middle
1179          */
1180         if (right) {
1181                 u32 right_nr;
1182
1183                 __btrfs_tree_lock(right, BTRFS_NESTING_RIGHT);
1184
1185                 right_nr = btrfs_header_nritems(right);
1186                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
1187                         wret = 1;
1188                 } else {
1189                         ret = btrfs_cow_block(trans, root, right,
1190                                               parent, pslot + 1,
1191                                               &right, BTRFS_NESTING_RIGHT_COW);
1192                         if (ret)
1193                                 wret = 1;
1194                         else {
1195                                 wret = balance_node_right(trans, right, mid);
1196                         }
1197                 }
1198                 if (wret < 0)
1199                         ret = wret;
1200                 if (wret == 0) {
1201                         struct btrfs_disk_key disk_key;
1202
1203                         btrfs_node_key(right, &disk_key, 0);
1204                         ret = btrfs_tree_mod_log_insert_key(parent, pslot + 1,
1205                                         BTRFS_MOD_LOG_KEY_REPLACE, GFP_NOFS);
1206                         BUG_ON(ret < 0);
1207                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
1208                         btrfs_mark_buffer_dirty(parent);
1209
1210                         if (btrfs_header_nritems(mid) <= orig_slot) {
1211                                 path->nodes[level] = right;
1212                                 path->slots[level + 1] += 1;
1213                                 path->slots[level] = orig_slot -
1214                                         btrfs_header_nritems(mid);
1215                                 btrfs_tree_unlock(mid);
1216                                 free_extent_buffer(mid);
1217                         } else {
1218                                 btrfs_tree_unlock(right);
1219                                 free_extent_buffer(right);
1220                         }
1221                         return 0;
1222                 }
1223                 btrfs_tree_unlock(right);
1224                 free_extent_buffer(right);
1225         }
1226         return 1;
1227 }
1228
1229 /*
1230  * readahead one full node of leaves, finding things that are close
1231  * to the block in 'slot', and triggering ra on them.
1232  */
1233 static void reada_for_search(struct btrfs_fs_info *fs_info,
1234                              struct btrfs_path *path,
1235                              int level, int slot, u64 objectid)
1236 {
1237         struct extent_buffer *node;
1238         struct btrfs_disk_key disk_key;
1239         u32 nritems;
1240         u64 search;
1241         u64 target;
1242         u64 nread = 0;
1243         u64 nread_max;
1244         u32 nr;
1245         u32 blocksize;
1246         u32 nscan = 0;
1247
1248         if (level != 1 && path->reada != READA_FORWARD_ALWAYS)
1249                 return;
1250
1251         if (!path->nodes[level])
1252                 return;
1253
1254         node = path->nodes[level];
1255
1256         /*
1257          * Since the time between visiting leaves is much shorter than the time
1258          * between visiting nodes, limit read ahead of nodes to 1, to avoid too
1259          * much IO at once (possibly random).
1260          */
1261         if (path->reada == READA_FORWARD_ALWAYS) {
1262                 if (level > 1)
1263                         nread_max = node->fs_info->nodesize;
1264                 else
1265                         nread_max = SZ_128K;
1266         } else {
1267                 nread_max = SZ_64K;
1268         }
1269
1270         search = btrfs_node_blockptr(node, slot);
1271         blocksize = fs_info->nodesize;
1272         if (path->reada != READA_FORWARD_ALWAYS) {
1273                 struct extent_buffer *eb;
1274
1275                 eb = find_extent_buffer(fs_info, search);
1276                 if (eb) {
1277                         free_extent_buffer(eb);
1278                         return;
1279                 }
1280         }
1281
1282         target = search;
1283
1284         nritems = btrfs_header_nritems(node);
1285         nr = slot;
1286
1287         while (1) {
1288                 if (path->reada == READA_BACK) {
1289                         if (nr == 0)
1290                                 break;
1291                         nr--;
1292                 } else if (path->reada == READA_FORWARD ||
1293                            path->reada == READA_FORWARD_ALWAYS) {
1294                         nr++;
1295                         if (nr >= nritems)
1296                                 break;
1297                 }
1298                 if (path->reada == READA_BACK && objectid) {
1299                         btrfs_node_key(node, &disk_key, nr);
1300                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1301                                 break;
1302                 }
1303                 search = btrfs_node_blockptr(node, nr);
1304                 if (path->reada == READA_FORWARD_ALWAYS ||
1305                     (search <= target && target - search <= 65536) ||
1306                     (search > target && search - target <= 65536)) {
1307                         btrfs_readahead_node_child(node, nr);
1308                         nread += blocksize;
1309                 }
1310                 nscan++;
1311                 if (nread > nread_max || nscan > 32)
1312                         break;
1313         }
1314 }
1315
1316 static noinline void reada_for_balance(struct btrfs_path *path, int level)
1317 {
1318         struct extent_buffer *parent;
1319         int slot;
1320         int nritems;
1321
1322         parent = path->nodes[level + 1];
1323         if (!parent)
1324                 return;
1325
1326         nritems = btrfs_header_nritems(parent);
1327         slot = path->slots[level + 1];
1328
1329         if (slot > 0)
1330                 btrfs_readahead_node_child(parent, slot - 1);
1331         if (slot + 1 < nritems)
1332                 btrfs_readahead_node_child(parent, slot + 1);
1333 }
1334
1335
1336 /*
1337  * when we walk down the tree, it is usually safe to unlock the higher layers
1338  * in the tree.  The exceptions are when our path goes through slot 0, because
1339  * operations on the tree might require changing key pointers higher up in the
1340  * tree.
1341  *
1342  * callers might also have set path->keep_locks, which tells this code to keep
1343  * the lock if the path points to the last slot in the block.  This is part of
1344  * walking through the tree, and selecting the next slot in the higher block.
1345  *
1346  * lowest_unlock sets the lowest level in the tree we're allowed to unlock.  so
1347  * if lowest_unlock is 1, level 0 won't be unlocked
1348  */
1349 static noinline void unlock_up(struct btrfs_path *path, int level,
1350                                int lowest_unlock, int min_write_lock_level,
1351                                int *write_lock_level)
1352 {
1353         int i;
1354         int skip_level = level;
1355         bool check_skip = true;
1356
1357         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1358                 if (!path->nodes[i])
1359                         break;
1360                 if (!path->locks[i])
1361                         break;
1362
1363                 if (check_skip) {
1364                         if (path->slots[i] == 0) {
1365                                 skip_level = i + 1;
1366                                 continue;
1367                         }
1368
1369                         if (path->keep_locks) {
1370                                 u32 nritems;
1371
1372                                 nritems = btrfs_header_nritems(path->nodes[i]);
1373                                 if (nritems < 1 || path->slots[i] >= nritems - 1) {
1374                                         skip_level = i + 1;
1375                                         continue;
1376                                 }
1377                         }
1378                 }
1379
1380                 if (i >= lowest_unlock && i > skip_level) {
1381                         check_skip = false;
1382                         btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
1383                         path->locks[i] = 0;
1384                         if (write_lock_level &&
1385                             i > min_write_lock_level &&
1386                             i <= *write_lock_level) {
1387                                 *write_lock_level = i - 1;
1388                         }
1389                 }
1390         }
1391 }
1392
1393 /*
1394  * Helper function for btrfs_search_slot() and other functions that do a search
1395  * on a btree. The goal is to find a tree block in the cache (the radix tree at
1396  * fs_info->buffer_radix), but if we can't find it, or it's not up to date, read
1397  * its pages from disk.
1398  *
1399  * Returns -EAGAIN, with the path unlocked, if the caller needs to repeat the
1400  * whole btree search, starting again from the current root node.
1401  */
1402 static int
1403 read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
1404                       struct extent_buffer **eb_ret, int level, int slot,
1405                       const struct btrfs_key *key)
1406 {
1407         struct btrfs_fs_info *fs_info = root->fs_info;
1408         u64 blocknr;
1409         u64 gen;
1410         struct extent_buffer *tmp;
1411         struct btrfs_key first_key;
1412         int ret;
1413         int parent_level;
1414         bool unlock_up;
1415
1416         unlock_up = ((level + 1 < BTRFS_MAX_LEVEL) && p->locks[level + 1]);
1417         blocknr = btrfs_node_blockptr(*eb_ret, slot);
1418         gen = btrfs_node_ptr_generation(*eb_ret, slot);
1419         parent_level = btrfs_header_level(*eb_ret);
1420         btrfs_node_key_to_cpu(*eb_ret, &first_key, slot);
1421
1422         /*
1423          * If we need to read an extent buffer from disk and we are holding locks
1424          * on upper level nodes, we unlock all the upper nodes before reading the
1425          * extent buffer, and then return -EAGAIN to the caller as it needs to
1426          * restart the search. We don't release the lock on the current level
1427          * because we need to walk this node to figure out which blocks to read.
1428          */
1429         tmp = find_extent_buffer(fs_info, blocknr);
1430         if (tmp) {
1431                 if (p->reada == READA_FORWARD_ALWAYS)
1432                         reada_for_search(fs_info, p, level, slot, key->objectid);
1433
1434                 /* first we do an atomic uptodate check */
1435                 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
1436                         /*
1437                          * Do extra check for first_key, eb can be stale due to
1438                          * being cached, read from scrub, or have multiple
1439                          * parents (shared tree blocks).
1440                          */
1441                         if (btrfs_verify_level_key(tmp,
1442                                         parent_level - 1, &first_key, gen)) {
1443                                 free_extent_buffer(tmp);
1444                                 return -EUCLEAN;
1445                         }
1446                         *eb_ret = tmp;
1447                         return 0;
1448                 }
1449
1450                 if (p->nowait) {
1451                         free_extent_buffer(tmp);
1452                         return -EAGAIN;
1453                 }
1454
1455                 if (unlock_up)
1456                         btrfs_unlock_up_safe(p, level + 1);
1457
1458                 /* now we're allowed to do a blocking uptodate check */
1459                 ret = btrfs_read_extent_buffer(tmp, gen, parent_level - 1, &first_key);
1460                 if (ret) {
1461                         free_extent_buffer(tmp);
1462                         btrfs_release_path(p);
1463                         return -EIO;
1464                 }
1465                 if (btrfs_check_eb_owner(tmp, root->root_key.objectid)) {
1466                         free_extent_buffer(tmp);
1467                         btrfs_release_path(p);
1468                         return -EUCLEAN;
1469                 }
1470
1471                 if (unlock_up)
1472                         ret = -EAGAIN;
1473
1474                 goto out;
1475         } else if (p->nowait) {
1476                 return -EAGAIN;
1477         }
1478
1479         if (unlock_up) {
1480                 btrfs_unlock_up_safe(p, level + 1);
1481                 ret = -EAGAIN;
1482         } else {
1483                 ret = 0;
1484         }
1485
1486         if (p->reada != READA_NONE)
1487                 reada_for_search(fs_info, p, level, slot, key->objectid);
1488
1489         tmp = read_tree_block(fs_info, blocknr, root->root_key.objectid,
1490                               gen, parent_level - 1, &first_key);
1491         if (IS_ERR(tmp)) {
1492                 btrfs_release_path(p);
1493                 return PTR_ERR(tmp);
1494         }
1495         /*
1496          * If the read above didn't mark this buffer up to date,
1497          * it will never end up being up to date.  Set ret to EIO now
1498          * and give up so that our caller doesn't loop forever
1499          * on our EAGAINs.
1500          */
1501         if (!extent_buffer_uptodate(tmp))
1502                 ret = -EIO;
1503
1504 out:
1505         if (ret == 0) {
1506                 *eb_ret = tmp;
1507         } else {
1508                 free_extent_buffer(tmp);
1509                 btrfs_release_path(p);
1510         }
1511
1512         return ret;
1513 }
1514
1515 /*
1516  * helper function for btrfs_search_slot.  This does all of the checks
1517  * for node-level blocks and does any balancing required based on
1518  * the ins_len.
1519  *
1520  * If no extra work was required, zero is returned.  If we had to
1521  * drop the path, -EAGAIN is returned and btrfs_search_slot must
1522  * start over
1523  */
1524 static int
1525 setup_nodes_for_search(struct btrfs_trans_handle *trans,
1526                        struct btrfs_root *root, struct btrfs_path *p,
1527                        struct extent_buffer *b, int level, int ins_len,
1528                        int *write_lock_level)
1529 {
1530         struct btrfs_fs_info *fs_info = root->fs_info;
1531         int ret = 0;
1532
1533         if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
1534             BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3) {
1535
1536                 if (*write_lock_level < level + 1) {
1537                         *write_lock_level = level + 1;
1538                         btrfs_release_path(p);
1539                         return -EAGAIN;
1540                 }
1541
1542                 reada_for_balance(p, level);
1543                 ret = split_node(trans, root, p, level);
1544
1545                 b = p->nodes[level];
1546         } else if (ins_len < 0 && btrfs_header_nritems(b) <
1547                    BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 2) {
1548
1549                 if (*write_lock_level < level + 1) {
1550                         *write_lock_level = level + 1;
1551                         btrfs_release_path(p);
1552                         return -EAGAIN;
1553                 }
1554
1555                 reada_for_balance(p, level);
1556                 ret = balance_level(trans, root, p, level);
1557                 if (ret)
1558                         return ret;
1559
1560                 b = p->nodes[level];
1561                 if (!b) {
1562                         btrfs_release_path(p);
1563                         return -EAGAIN;
1564                 }
1565                 BUG_ON(btrfs_header_nritems(b) == 1);
1566         }
1567         return ret;
1568 }
1569
1570 int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
1571                 u64 iobjectid, u64 ioff, u8 key_type,
1572                 struct btrfs_key *found_key)
1573 {
1574         int ret;
1575         struct btrfs_key key;
1576         struct extent_buffer *eb;
1577
1578         ASSERT(path);
1579         ASSERT(found_key);
1580
1581         key.type = key_type;
1582         key.objectid = iobjectid;
1583         key.offset = ioff;
1584
1585         ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
1586         if (ret < 0)
1587                 return ret;
1588
1589         eb = path->nodes[0];
1590         if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
1591                 ret = btrfs_next_leaf(fs_root, path);
1592                 if (ret)
1593                         return ret;
1594                 eb = path->nodes[0];
1595         }
1596
1597         btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
1598         if (found_key->type != key.type ||
1599                         found_key->objectid != key.objectid)
1600                 return 1;
1601
1602         return 0;
1603 }
1604
1605 static struct extent_buffer *btrfs_search_slot_get_root(struct btrfs_root *root,
1606                                                         struct btrfs_path *p,
1607                                                         int write_lock_level)
1608 {
1609         struct extent_buffer *b;
1610         int root_lock = 0;
1611         int level = 0;
1612
1613         if (p->search_commit_root) {
1614                 b = root->commit_root;
1615                 atomic_inc(&b->refs);
1616                 level = btrfs_header_level(b);
1617                 /*
1618                  * Ensure that all callers have set skip_locking when
1619                  * p->search_commit_root = 1.
1620                  */
1621                 ASSERT(p->skip_locking == 1);
1622
1623                 goto out;
1624         }
1625
1626         if (p->skip_locking) {
1627                 b = btrfs_root_node(root);
1628                 level = btrfs_header_level(b);
1629                 goto out;
1630         }
1631
1632         /* We try very hard to do read locks on the root */
1633         root_lock = BTRFS_READ_LOCK;
1634
1635         /*
1636          * If the level is set to maximum, we can skip trying to get the read
1637          * lock.
1638          */
1639         if (write_lock_level < BTRFS_MAX_LEVEL) {
1640                 /*
1641                  * We don't know the level of the root node until we actually
1642                  * have it read locked
1643                  */
1644                 if (p->nowait) {
1645                         b = btrfs_try_read_lock_root_node(root);
1646                         if (IS_ERR(b))
1647                                 return b;
1648                 } else {
1649                         b = btrfs_read_lock_root_node(root);
1650                 }
1651                 level = btrfs_header_level(b);
1652                 if (level > write_lock_level)
1653                         goto out;
1654
1655                 /* Whoops, must trade for write lock */
1656                 btrfs_tree_read_unlock(b);
1657                 free_extent_buffer(b);
1658         }
1659
1660         b = btrfs_lock_root_node(root);
1661         root_lock = BTRFS_WRITE_LOCK;
1662
1663         /* The level might have changed, check again */
1664         level = btrfs_header_level(b);
1665
1666 out:
1667         /*
1668          * The root may have failed to write out at some point, and thus is no
1669          * longer valid, return an error in this case.
1670          */
1671         if (!extent_buffer_uptodate(b)) {
1672                 if (root_lock)
1673                         btrfs_tree_unlock_rw(b, root_lock);
1674                 free_extent_buffer(b);
1675                 return ERR_PTR(-EIO);
1676         }
1677
1678         p->nodes[level] = b;
1679         if (!p->skip_locking)
1680                 p->locks[level] = root_lock;
1681         /*
1682          * Callers are responsible for dropping b's references.
1683          */
1684         return b;
1685 }
1686
1687 /*
1688  * Replace the extent buffer at the lowest level of the path with a cloned
1689  * version. The purpose is to be able to use it safely, after releasing the
1690  * commit root semaphore, even if relocation is happening in parallel, the
1691  * transaction used for relocation is committed and the extent buffer is
1692  * reallocated in the next transaction.
1693  *
1694  * This is used in a context where the caller does not prevent transaction
1695  * commits from happening, either by holding a transaction handle or holding
1696  * some lock, while it's doing searches through a commit root.
1697  * At the moment it's only used for send operations.
1698  */
1699 static int finish_need_commit_sem_search(struct btrfs_path *path)
1700 {
1701         const int i = path->lowest_level;
1702         const int slot = path->slots[i];
1703         struct extent_buffer *lowest = path->nodes[i];
1704         struct extent_buffer *clone;
1705
1706         ASSERT(path->need_commit_sem);
1707
1708         if (!lowest)
1709                 return 0;
1710
1711         lockdep_assert_held_read(&lowest->fs_info->commit_root_sem);
1712
1713         clone = btrfs_clone_extent_buffer(lowest);
1714         if (!clone)
1715                 return -ENOMEM;
1716
1717         btrfs_release_path(path);
1718         path->nodes[i] = clone;
1719         path->slots[i] = slot;
1720
1721         return 0;
1722 }
1723
1724 static inline int search_for_key_slot(struct extent_buffer *eb,
1725                                       int search_low_slot,
1726                                       const struct btrfs_key *key,
1727                                       int prev_cmp,
1728                                       int *slot)
1729 {
1730         /*
1731          * If a previous call to btrfs_bin_search() on a parent node returned an
1732          * exact match (prev_cmp == 0), we can safely assume the target key will
1733          * always be at slot 0 on lower levels, since each key pointer
1734          * (struct btrfs_key_ptr) refers to the lowest key accessible from the
1735          * subtree it points to. Thus we can skip searching lower levels.
1736          */
1737         if (prev_cmp == 0) {
1738                 *slot = 0;
1739                 return 0;
1740         }
1741
1742         return generic_bin_search(eb, search_low_slot, key, slot);
1743 }
1744
1745 static int search_leaf(struct btrfs_trans_handle *trans,
1746                        struct btrfs_root *root,
1747                        const struct btrfs_key *key,
1748                        struct btrfs_path *path,
1749                        int ins_len,
1750                        int prev_cmp)
1751 {
1752         struct extent_buffer *leaf = path->nodes[0];
1753         int leaf_free_space = -1;
1754         int search_low_slot = 0;
1755         int ret;
1756         bool do_bin_search = true;
1757
1758         /*
1759          * If we are doing an insertion, the leaf has enough free space and the
1760          * destination slot for the key is not slot 0, then we can unlock our
1761          * write lock on the parent, and any other upper nodes, before doing the
1762          * binary search on the leaf (with search_for_key_slot()), allowing other
1763          * tasks to lock the parent and any other upper nodes.
1764          */
1765         if (ins_len > 0) {
1766                 /*
1767                  * Cache the leaf free space, since we will need it later and it
1768                  * will not change until then.
1769                  */
1770                 leaf_free_space = btrfs_leaf_free_space(leaf);
1771
1772                 /*
1773                  * !path->locks[1] means we have a single node tree, the leaf is
1774                  * the root of the tree.
1775                  */
1776                 if (path->locks[1] && leaf_free_space >= ins_len) {
1777                         struct btrfs_disk_key first_key;
1778
1779                         ASSERT(btrfs_header_nritems(leaf) > 0);
1780                         btrfs_item_key(leaf, &first_key, 0);
1781
1782                         /*
1783                          * Doing the extra comparison with the first key is cheap,
1784                          * taking into account that the first key is very likely
1785                          * already in a cache line because it immediately follows
1786                          * the extent buffer's header and we have recently accessed
1787                          * the header's level field.
1788                          */
1789                         ret = comp_keys(&first_key, key);
1790                         if (ret < 0) {
1791                                 /*
1792                                  * The first key is smaller than the key we want
1793                                  * to insert, so we are safe to unlock all upper
1794                                  * nodes and we have to do the binary search.
1795                                  *
1796                                  * We do use btrfs_unlock_up_safe() and not
1797                                  * unlock_up() because the later does not unlock
1798                                  * nodes with a slot of 0 - we can safely unlock
1799                                  * any node even if its slot is 0 since in this
1800                                  * case the key does not end up at slot 0 of the
1801                                  * leaf and there's no need to split the leaf.
1802                                  */
1803                                 btrfs_unlock_up_safe(path, 1);
1804                                 search_low_slot = 1;
1805                         } else {
1806                                 /*
1807                                  * The first key is >= then the key we want to
1808                                  * insert, so we can skip the binary search as
1809                                  * the target key will be at slot 0.
1810                                  *
1811                                  * We can not unlock upper nodes when the key is
1812                                  * less than the first key, because we will need
1813                                  * to update the key at slot 0 of the parent node
1814                                  * and possibly of other upper nodes too.
1815                                  * If the key matches the first key, then we can
1816                                  * unlock all the upper nodes, using
1817                                  * btrfs_unlock_up_safe() instead of unlock_up()
1818                                  * as stated above.
1819                                  */
1820                                 if (ret == 0)
1821                                         btrfs_unlock_up_safe(path, 1);
1822                                 /*
1823                                  * ret is already 0 or 1, matching the result of
1824                                  * a btrfs_bin_search() call, so there is no need
1825                                  * to adjust it.
1826                                  */
1827                                 do_bin_search = false;
1828                                 path->slots[0] = 0;
1829                         }
1830                 }
1831         }
1832
1833         if (do_bin_search) {
1834                 ret = search_for_key_slot(leaf, search_low_slot, key,
1835                                           prev_cmp, &path->slots[0]);
1836                 if (ret < 0)
1837                         return ret;
1838         }
1839
1840         if (ins_len > 0) {
1841                 /*
1842                  * Item key already exists. In this case, if we are allowed to
1843                  * insert the item (for example, in dir_item case, item key
1844                  * collision is allowed), it will be merged with the original
1845                  * item. Only the item size grows, no new btrfs item will be
1846                  * added. If search_for_extension is not set, ins_len already
1847                  * accounts the size btrfs_item, deduct it here so leaf space
1848                  * check will be correct.
1849                  */
1850                 if (ret == 0 && !path->search_for_extension) {
1851                         ASSERT(ins_len >= sizeof(struct btrfs_item));
1852                         ins_len -= sizeof(struct btrfs_item);
1853                 }
1854
1855                 ASSERT(leaf_free_space >= 0);
1856
1857                 if (leaf_free_space < ins_len) {
1858                         int err;
1859
1860                         err = split_leaf(trans, root, key, path, ins_len,
1861                                          (ret == 0));
1862                         ASSERT(err <= 0);
1863                         if (WARN_ON(err > 0))
1864                                 err = -EUCLEAN;
1865                         if (err)
1866                                 ret = err;
1867                 }
1868         }
1869
1870         return ret;
1871 }
1872
1873 /*
1874  * btrfs_search_slot - look for a key in a tree and perform necessary
1875  * modifications to preserve tree invariants.
1876  *
1877  * @trans:      Handle of transaction, used when modifying the tree
1878  * @p:          Holds all btree nodes along the search path
1879  * @root:       The root node of the tree
1880  * @key:        The key we are looking for
1881  * @ins_len:    Indicates purpose of search:
1882  *              >0  for inserts it's size of item inserted (*)
1883  *              <0  for deletions
1884  *               0  for plain searches, not modifying the tree
1885  *
1886  *              (*) If size of item inserted doesn't include
1887  *              sizeof(struct btrfs_item), then p->search_for_extension must
1888  *              be set.
1889  * @cow:        boolean should CoW operations be performed. Must always be 1
1890  *              when modifying the tree.
1891  *
1892  * If @ins_len > 0, nodes and leaves will be split as we walk down the tree.
1893  * If @ins_len < 0, nodes will be merged as we walk down the tree (if possible)
1894  *
1895  * If @key is found, 0 is returned and you can find the item in the leaf level
1896  * of the path (level 0)
1897  *
1898  * If @key isn't found, 1 is returned and the leaf level of the path (level 0)
1899  * points to the slot where it should be inserted
1900  *
1901  * If an error is encountered while searching the tree a negative error number
1902  * is returned
1903  */
1904 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1905                       const struct btrfs_key *key, struct btrfs_path *p,
1906                       int ins_len, int cow)
1907 {
1908         struct btrfs_fs_info *fs_info = root->fs_info;
1909         struct extent_buffer *b;
1910         int slot;
1911         int ret;
1912         int err;
1913         int level;
1914         int lowest_unlock = 1;
1915         /* everything at write_lock_level or lower must be write locked */
1916         int write_lock_level = 0;
1917         u8 lowest_level = 0;
1918         int min_write_lock_level;
1919         int prev_cmp;
1920
1921         lowest_level = p->lowest_level;
1922         WARN_ON(lowest_level && ins_len > 0);
1923         WARN_ON(p->nodes[0] != NULL);
1924         BUG_ON(!cow && ins_len);
1925
1926         /*
1927          * For now only allow nowait for read only operations.  There's no
1928          * strict reason why we can't, we just only need it for reads so it's
1929          * only implemented for reads.
1930          */
1931         ASSERT(!p->nowait || !cow);
1932
1933         if (ins_len < 0) {
1934                 lowest_unlock = 2;
1935
1936                 /* when we are removing items, we might have to go up to level
1937                  * two as we update tree pointers  Make sure we keep write
1938                  * for those levels as well
1939                  */
1940                 write_lock_level = 2;
1941         } else if (ins_len > 0) {
1942                 /*
1943                  * for inserting items, make sure we have a write lock on
1944                  * level 1 so we can update keys
1945                  */
1946                 write_lock_level = 1;
1947         }
1948
1949         if (!cow)
1950                 write_lock_level = -1;
1951
1952         if (cow && (p->keep_locks || p->lowest_level))
1953                 write_lock_level = BTRFS_MAX_LEVEL;
1954
1955         min_write_lock_level = write_lock_level;
1956
1957         if (p->need_commit_sem) {
1958                 ASSERT(p->search_commit_root);
1959                 if (p->nowait) {
1960                         if (!down_read_trylock(&fs_info->commit_root_sem))
1961                                 return -EAGAIN;
1962                 } else {
1963                         down_read(&fs_info->commit_root_sem);
1964                 }
1965         }
1966
1967 again:
1968         prev_cmp = -1;
1969         b = btrfs_search_slot_get_root(root, p, write_lock_level);
1970         if (IS_ERR(b)) {
1971                 ret = PTR_ERR(b);
1972                 goto done;
1973         }
1974
1975         while (b) {
1976                 int dec = 0;
1977
1978                 level = btrfs_header_level(b);
1979
1980                 if (cow) {
1981                         bool last_level = (level == (BTRFS_MAX_LEVEL - 1));
1982
1983                         /*
1984                          * if we don't really need to cow this block
1985                          * then we don't want to set the path blocking,
1986                          * so we test it here
1987                          */
1988                         if (!should_cow_block(trans, root, b))
1989                                 goto cow_done;
1990
1991                         /*
1992                          * must have write locks on this node and the
1993                          * parent
1994                          */
1995                         if (level > write_lock_level ||
1996                             (level + 1 > write_lock_level &&
1997                             level + 1 < BTRFS_MAX_LEVEL &&
1998                             p->nodes[level + 1])) {
1999                                 write_lock_level = level + 1;
2000                                 btrfs_release_path(p);
2001                                 goto again;
2002                         }
2003
2004                         if (last_level)
2005                                 err = btrfs_cow_block(trans, root, b, NULL, 0,
2006                                                       &b,
2007                                                       BTRFS_NESTING_COW);
2008                         else
2009                                 err = btrfs_cow_block(trans, root, b,
2010                                                       p->nodes[level + 1],
2011                                                       p->slots[level + 1], &b,
2012                                                       BTRFS_NESTING_COW);
2013                         if (err) {
2014                                 ret = err;
2015                                 goto done;
2016                         }
2017                 }
2018 cow_done:
2019                 p->nodes[level] = b;
2020
2021                 /*
2022                  * we have a lock on b and as long as we aren't changing
2023                  * the tree, there is no way to for the items in b to change.
2024                  * It is safe to drop the lock on our parent before we
2025                  * go through the expensive btree search on b.
2026                  *
2027                  * If we're inserting or deleting (ins_len != 0), then we might
2028                  * be changing slot zero, which may require changing the parent.
2029                  * So, we can't drop the lock until after we know which slot
2030                  * we're operating on.
2031                  */
2032                 if (!ins_len && !p->keep_locks) {
2033                         int u = level + 1;
2034
2035                         if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
2036                                 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
2037                                 p->locks[u] = 0;
2038                         }
2039                 }
2040
2041                 if (level == 0) {
2042                         if (ins_len > 0)
2043                                 ASSERT(write_lock_level >= 1);
2044
2045                         ret = search_leaf(trans, root, key, p, ins_len, prev_cmp);
2046                         if (!p->search_for_split)
2047                                 unlock_up(p, level, lowest_unlock,
2048                                           min_write_lock_level, NULL);
2049                         goto done;
2050                 }
2051
2052                 ret = search_for_key_slot(b, 0, key, prev_cmp, &slot);
2053                 if (ret < 0)
2054                         goto done;
2055                 prev_cmp = ret;
2056
2057                 if (ret && slot > 0) {
2058                         dec = 1;
2059                         slot--;
2060                 }
2061                 p->slots[level] = slot;
2062                 err = setup_nodes_for_search(trans, root, p, b, level, ins_len,
2063                                              &write_lock_level);
2064                 if (err == -EAGAIN)
2065                         goto again;
2066                 if (err) {
2067                         ret = err;
2068                         goto done;
2069                 }
2070                 b = p->nodes[level];
2071                 slot = p->slots[level];
2072
2073                 /*
2074                  * Slot 0 is special, if we change the key we have to update
2075                  * the parent pointer which means we must have a write lock on
2076                  * the parent
2077                  */
2078                 if (slot == 0 && ins_len && write_lock_level < level + 1) {
2079                         write_lock_level = level + 1;
2080                         btrfs_release_path(p);
2081                         goto again;
2082                 }
2083
2084                 unlock_up(p, level, lowest_unlock, min_write_lock_level,
2085                           &write_lock_level);
2086
2087                 if (level == lowest_level) {
2088                         if (dec)
2089                                 p->slots[level]++;
2090                         goto done;
2091                 }
2092
2093                 err = read_block_for_search(root, p, &b, level, slot, key);
2094                 if (err == -EAGAIN)
2095                         goto again;
2096                 if (err) {
2097                         ret = err;
2098                         goto done;
2099                 }
2100
2101                 if (!p->skip_locking) {
2102                         level = btrfs_header_level(b);
2103
2104                         btrfs_maybe_reset_lockdep_class(root, b);
2105
2106                         if (level <= write_lock_level) {
2107                                 btrfs_tree_lock(b);
2108                                 p->locks[level] = BTRFS_WRITE_LOCK;
2109                         } else {
2110                                 if (p->nowait) {
2111                                         if (!btrfs_try_tree_read_lock(b)) {
2112                                                 free_extent_buffer(b);
2113                                                 ret = -EAGAIN;
2114                                                 goto done;
2115                                         }
2116                                 } else {
2117                                         btrfs_tree_read_lock(b);
2118                                 }
2119                                 p->locks[level] = BTRFS_READ_LOCK;
2120                         }
2121                         p->nodes[level] = b;
2122                 }
2123         }
2124         ret = 1;
2125 done:
2126         if (ret < 0 && !p->skip_release_on_error)
2127                 btrfs_release_path(p);
2128
2129         if (p->need_commit_sem) {
2130                 int ret2;
2131
2132                 ret2 = finish_need_commit_sem_search(p);
2133                 up_read(&fs_info->commit_root_sem);
2134                 if (ret2)
2135                         ret = ret2;
2136         }
2137
2138         return ret;
2139 }
2140 ALLOW_ERROR_INJECTION(btrfs_search_slot, ERRNO);
2141
2142 /*
2143  * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2144  * current state of the tree together with the operations recorded in the tree
2145  * modification log to search for the key in a previous version of this tree, as
2146  * denoted by the time_seq parameter.
2147  *
2148  * Naturally, there is no support for insert, delete or cow operations.
2149  *
2150  * The resulting path and return value will be set up as if we called
2151  * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2152  */
2153 int btrfs_search_old_slot(struct btrfs_root *root, const struct btrfs_key *key,
2154                           struct btrfs_path *p, u64 time_seq)
2155 {
2156         struct btrfs_fs_info *fs_info = root->fs_info;
2157         struct extent_buffer *b;
2158         int slot;
2159         int ret;
2160         int err;
2161         int level;
2162         int lowest_unlock = 1;
2163         u8 lowest_level = 0;
2164
2165         lowest_level = p->lowest_level;
2166         WARN_ON(p->nodes[0] != NULL);
2167         ASSERT(!p->nowait);
2168
2169         if (p->search_commit_root) {
2170                 BUG_ON(time_seq);
2171                 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2172         }
2173
2174 again:
2175         b = btrfs_get_old_root(root, time_seq);
2176         if (!b) {
2177                 ret = -EIO;
2178                 goto done;
2179         }
2180         level = btrfs_header_level(b);
2181         p->locks[level] = BTRFS_READ_LOCK;
2182
2183         while (b) {
2184                 int dec = 0;
2185
2186                 level = btrfs_header_level(b);
2187                 p->nodes[level] = b;
2188
2189                 /*
2190                  * we have a lock on b and as long as we aren't changing
2191                  * the tree, there is no way to for the items in b to change.
2192                  * It is safe to drop the lock on our parent before we
2193                  * go through the expensive btree search on b.
2194                  */
2195                 btrfs_unlock_up_safe(p, level + 1);
2196
2197                 ret = btrfs_bin_search(b, key, &slot);
2198                 if (ret < 0)
2199                         goto done;
2200
2201                 if (level == 0) {
2202                         p->slots[level] = slot;
2203                         unlock_up(p, level, lowest_unlock, 0, NULL);
2204                         goto done;
2205                 }
2206
2207                 if (ret && slot > 0) {
2208                         dec = 1;
2209                         slot--;
2210                 }
2211                 p->slots[level] = slot;
2212                 unlock_up(p, level, lowest_unlock, 0, NULL);
2213
2214                 if (level == lowest_level) {
2215                         if (dec)
2216                                 p->slots[level]++;
2217                         goto done;
2218                 }
2219
2220                 err = read_block_for_search(root, p, &b, level, slot, key);
2221                 if (err == -EAGAIN)
2222                         goto again;
2223                 if (err) {
2224                         ret = err;
2225                         goto done;
2226                 }
2227
2228                 level = btrfs_header_level(b);
2229                 btrfs_tree_read_lock(b);
2230                 b = btrfs_tree_mod_log_rewind(fs_info, p, b, time_seq);
2231                 if (!b) {
2232                         ret = -ENOMEM;
2233                         goto done;
2234                 }
2235                 p->locks[level] = BTRFS_READ_LOCK;
2236                 p->nodes[level] = b;
2237         }
2238         ret = 1;
2239 done:
2240         if (ret < 0)
2241                 btrfs_release_path(p);
2242
2243         return ret;
2244 }
2245
2246 /*
2247  * helper to use instead of search slot if no exact match is needed but
2248  * instead the next or previous item should be returned.
2249  * When find_higher is true, the next higher item is returned, the next lower
2250  * otherwise.
2251  * When return_any and find_higher are both true, and no higher item is found,
2252  * return the next lower instead.
2253  * When return_any is true and find_higher is false, and no lower item is found,
2254  * return the next higher instead.
2255  * It returns 0 if any item is found, 1 if none is found (tree empty), and
2256  * < 0 on error
2257  */
2258 int btrfs_search_slot_for_read(struct btrfs_root *root,
2259                                const struct btrfs_key *key,
2260                                struct btrfs_path *p, int find_higher,
2261                                int return_any)
2262 {
2263         int ret;
2264         struct extent_buffer *leaf;
2265
2266 again:
2267         ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
2268         if (ret <= 0)
2269                 return ret;
2270         /*
2271          * a return value of 1 means the path is at the position where the
2272          * item should be inserted. Normally this is the next bigger item,
2273          * but in case the previous item is the last in a leaf, path points
2274          * to the first free slot in the previous leaf, i.e. at an invalid
2275          * item.
2276          */
2277         leaf = p->nodes[0];
2278
2279         if (find_higher) {
2280                 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
2281                         ret = btrfs_next_leaf(root, p);
2282                         if (ret <= 0)
2283                                 return ret;
2284                         if (!return_any)
2285                                 return 1;
2286                         /*
2287                          * no higher item found, return the next
2288                          * lower instead
2289                          */
2290                         return_any = 0;
2291                         find_higher = 0;
2292                         btrfs_release_path(p);
2293                         goto again;
2294                 }
2295         } else {
2296                 if (p->slots[0] == 0) {
2297                         ret = btrfs_prev_leaf(root, p);
2298                         if (ret < 0)
2299                                 return ret;
2300                         if (!ret) {
2301                                 leaf = p->nodes[0];
2302                                 if (p->slots[0] == btrfs_header_nritems(leaf))
2303                                         p->slots[0]--;
2304                                 return 0;
2305                         }
2306                         if (!return_any)
2307                                 return 1;
2308                         /*
2309                          * no lower item found, return the next
2310                          * higher instead
2311                          */
2312                         return_any = 0;
2313                         find_higher = 1;
2314                         btrfs_release_path(p);
2315                         goto again;
2316                 } else {
2317                         --p->slots[0];
2318                 }
2319         }
2320         return 0;
2321 }
2322
2323 /*
2324  * Execute search and call btrfs_previous_item to traverse backwards if the item
2325  * was not found.
2326  *
2327  * Return 0 if found, 1 if not found and < 0 if error.
2328  */
2329 int btrfs_search_backwards(struct btrfs_root *root, struct btrfs_key *key,
2330                            struct btrfs_path *path)
2331 {
2332         int ret;
2333
2334         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
2335         if (ret > 0)
2336                 ret = btrfs_previous_item(root, path, key->objectid, key->type);
2337
2338         if (ret == 0)
2339                 btrfs_item_key_to_cpu(path->nodes[0], key, path->slots[0]);
2340
2341         return ret;
2342 }
2343
2344 /**
2345  * Search for a valid slot for the given path.
2346  *
2347  * @root:       The root node of the tree.
2348  * @key:        Will contain a valid item if found.
2349  * @path:       The starting point to validate the slot.
2350  *
2351  * Return: 0  if the item is valid
2352  *         1  if not found
2353  *         <0 if error.
2354  */
2355 int btrfs_get_next_valid_item(struct btrfs_root *root, struct btrfs_key *key,
2356                               struct btrfs_path *path)
2357 {
2358         while (1) {
2359                 int ret;
2360                 const int slot = path->slots[0];
2361                 const struct extent_buffer *leaf = path->nodes[0];
2362
2363                 /* This is where we start walking the path. */
2364                 if (slot >= btrfs_header_nritems(leaf)) {
2365                         /*
2366                          * If we've reached the last slot in this leaf we need
2367                          * to go to the next leaf and reset the path.
2368                          */
2369                         ret = btrfs_next_leaf(root, path);
2370                         if (ret)
2371                                 return ret;
2372                         continue;
2373                 }
2374                 /* Store the found, valid item in @key. */
2375                 btrfs_item_key_to_cpu(leaf, key, slot);
2376                 break;
2377         }
2378         return 0;
2379 }
2380
2381 /*
2382  * adjust the pointers going up the tree, starting at level
2383  * making sure the right key of each node is points to 'key'.
2384  * This is used after shifting pointers to the left, so it stops
2385  * fixing up pointers when a given leaf/node is not in slot 0 of the
2386  * higher levels
2387  *
2388  */
2389 static void fixup_low_keys(struct btrfs_path *path,
2390                            struct btrfs_disk_key *key, int level)
2391 {
2392         int i;
2393         struct extent_buffer *t;
2394         int ret;
2395
2396         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2397                 int tslot = path->slots[i];
2398
2399                 if (!path->nodes[i])
2400                         break;
2401                 t = path->nodes[i];
2402                 ret = btrfs_tree_mod_log_insert_key(t, tslot,
2403                                 BTRFS_MOD_LOG_KEY_REPLACE, GFP_ATOMIC);
2404                 BUG_ON(ret < 0);
2405                 btrfs_set_node_key(t, key, tslot);
2406                 btrfs_mark_buffer_dirty(path->nodes[i]);
2407                 if (tslot != 0)
2408                         break;
2409         }
2410 }
2411
2412 /*
2413  * update item key.
2414  *
2415  * This function isn't completely safe. It's the caller's responsibility
2416  * that the new key won't break the order
2417  */
2418 void btrfs_set_item_key_safe(struct btrfs_fs_info *fs_info,
2419                              struct btrfs_path *path,
2420                              const struct btrfs_key *new_key)
2421 {
2422         struct btrfs_disk_key disk_key;
2423         struct extent_buffer *eb;
2424         int slot;
2425
2426         eb = path->nodes[0];
2427         slot = path->slots[0];
2428         if (slot > 0) {
2429                 btrfs_item_key(eb, &disk_key, slot - 1);
2430                 if (unlikely(comp_keys(&disk_key, new_key) >= 0)) {
2431                         btrfs_crit(fs_info,
2432                 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
2433                                    slot, btrfs_disk_key_objectid(&disk_key),
2434                                    btrfs_disk_key_type(&disk_key),
2435                                    btrfs_disk_key_offset(&disk_key),
2436                                    new_key->objectid, new_key->type,
2437                                    new_key->offset);
2438                         btrfs_print_leaf(eb);
2439                         BUG();
2440                 }
2441         }
2442         if (slot < btrfs_header_nritems(eb) - 1) {
2443                 btrfs_item_key(eb, &disk_key, slot + 1);
2444                 if (unlikely(comp_keys(&disk_key, new_key) <= 0)) {
2445                         btrfs_crit(fs_info,
2446                 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
2447                                    slot, btrfs_disk_key_objectid(&disk_key),
2448                                    btrfs_disk_key_type(&disk_key),
2449                                    btrfs_disk_key_offset(&disk_key),
2450                                    new_key->objectid, new_key->type,
2451                                    new_key->offset);
2452                         btrfs_print_leaf(eb);
2453                         BUG();
2454                 }
2455         }
2456
2457         btrfs_cpu_key_to_disk(&disk_key, new_key);
2458         btrfs_set_item_key(eb, &disk_key, slot);
2459         btrfs_mark_buffer_dirty(eb);
2460         if (slot == 0)
2461                 fixup_low_keys(path, &disk_key, 1);
2462 }
2463
2464 /*
2465  * Check key order of two sibling extent buffers.
2466  *
2467  * Return true if something is wrong.
2468  * Return false if everything is fine.
2469  *
2470  * Tree-checker only works inside one tree block, thus the following
2471  * corruption can not be detected by tree-checker:
2472  *
2473  * Leaf @left                   | Leaf @right
2474  * --------------------------------------------------------------
2475  * | 1 | 2 | 3 | 4 | 5 | f6 |   | 7 | 8 |
2476  *
2477  * Key f6 in leaf @left itself is valid, but not valid when the next
2478  * key in leaf @right is 7.
2479  * This can only be checked at tree block merge time.
2480  * And since tree checker has ensured all key order in each tree block
2481  * is correct, we only need to bother the last key of @left and the first
2482  * key of @right.
2483  */
2484 static bool check_sibling_keys(struct extent_buffer *left,
2485                                struct extent_buffer *right)
2486 {
2487         struct btrfs_key left_last;
2488         struct btrfs_key right_first;
2489         int level = btrfs_header_level(left);
2490         int nr_left = btrfs_header_nritems(left);
2491         int nr_right = btrfs_header_nritems(right);
2492
2493         /* No key to check in one of the tree blocks */
2494         if (!nr_left || !nr_right)
2495                 return false;
2496
2497         if (level) {
2498                 btrfs_node_key_to_cpu(left, &left_last, nr_left - 1);
2499                 btrfs_node_key_to_cpu(right, &right_first, 0);
2500         } else {
2501                 btrfs_item_key_to_cpu(left, &left_last, nr_left - 1);
2502                 btrfs_item_key_to_cpu(right, &right_first, 0);
2503         }
2504
2505         if (btrfs_comp_cpu_keys(&left_last, &right_first) >= 0) {
2506                 btrfs_crit(left->fs_info,
2507 "bad key order, sibling blocks, left last (%llu %u %llu) right first (%llu %u %llu)",
2508                            left_last.objectid, left_last.type,
2509                            left_last.offset, right_first.objectid,
2510                            right_first.type, right_first.offset);
2511                 return true;
2512         }
2513         return false;
2514 }
2515
2516 /*
2517  * try to push data from one node into the next node left in the
2518  * tree.
2519  *
2520  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
2521  * error, and > 0 if there was no room in the left hand block.
2522  */
2523 static int push_node_left(struct btrfs_trans_handle *trans,
2524                           struct extent_buffer *dst,
2525                           struct extent_buffer *src, int empty)
2526 {
2527         struct btrfs_fs_info *fs_info = trans->fs_info;
2528         int push_items = 0;
2529         int src_nritems;
2530         int dst_nritems;
2531         int ret = 0;
2532
2533         src_nritems = btrfs_header_nritems(src);
2534         dst_nritems = btrfs_header_nritems(dst);
2535         push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
2536         WARN_ON(btrfs_header_generation(src) != trans->transid);
2537         WARN_ON(btrfs_header_generation(dst) != trans->transid);
2538
2539         if (!empty && src_nritems <= 8)
2540                 return 1;
2541
2542         if (push_items <= 0)
2543                 return 1;
2544
2545         if (empty) {
2546                 push_items = min(src_nritems, push_items);
2547                 if (push_items < src_nritems) {
2548                         /* leave at least 8 pointers in the node if
2549                          * we aren't going to empty it
2550                          */
2551                         if (src_nritems - push_items < 8) {
2552                                 if (push_items <= 8)
2553                                         return 1;
2554                                 push_items -= 8;
2555                         }
2556                 }
2557         } else
2558                 push_items = min(src_nritems - 8, push_items);
2559
2560         /* dst is the left eb, src is the middle eb */
2561         if (check_sibling_keys(dst, src)) {
2562                 ret = -EUCLEAN;
2563                 btrfs_abort_transaction(trans, ret);
2564                 return ret;
2565         }
2566         ret = btrfs_tree_mod_log_eb_copy(dst, src, dst_nritems, 0, push_items);
2567         if (ret) {
2568                 btrfs_abort_transaction(trans, ret);
2569                 return ret;
2570         }
2571         copy_extent_buffer(dst, src,
2572                            btrfs_node_key_ptr_offset(dst_nritems),
2573                            btrfs_node_key_ptr_offset(0),
2574                            push_items * sizeof(struct btrfs_key_ptr));
2575
2576         if (push_items < src_nritems) {
2577                 /*
2578                  * Don't call btrfs_tree_mod_log_insert_move() here, key removal
2579                  * was already fully logged by btrfs_tree_mod_log_eb_copy() above.
2580                  */
2581                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
2582                                       btrfs_node_key_ptr_offset(push_items),
2583                                       (src_nritems - push_items) *
2584                                       sizeof(struct btrfs_key_ptr));
2585         }
2586         btrfs_set_header_nritems(src, src_nritems - push_items);
2587         btrfs_set_header_nritems(dst, dst_nritems + push_items);
2588         btrfs_mark_buffer_dirty(src);
2589         btrfs_mark_buffer_dirty(dst);
2590
2591         return ret;
2592 }
2593
2594 /*
2595  * try to push data from one node into the next node right in the
2596  * tree.
2597  *
2598  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
2599  * error, and > 0 if there was no room in the right hand block.
2600  *
2601  * this will  only push up to 1/2 the contents of the left node over
2602  */
2603 static int balance_node_right(struct btrfs_trans_handle *trans,
2604                               struct extent_buffer *dst,
2605                               struct extent_buffer *src)
2606 {
2607         struct btrfs_fs_info *fs_info = trans->fs_info;
2608         int push_items = 0;
2609         int max_push;
2610         int src_nritems;
2611         int dst_nritems;
2612         int ret = 0;
2613
2614         WARN_ON(btrfs_header_generation(src) != trans->transid);
2615         WARN_ON(btrfs_header_generation(dst) != trans->transid);
2616
2617         src_nritems = btrfs_header_nritems(src);
2618         dst_nritems = btrfs_header_nritems(dst);
2619         push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
2620         if (push_items <= 0)
2621                 return 1;
2622
2623         if (src_nritems < 4)
2624                 return 1;
2625
2626         max_push = src_nritems / 2 + 1;
2627         /* don't try to empty the node */
2628         if (max_push >= src_nritems)
2629                 return 1;
2630
2631         if (max_push < push_items)
2632                 push_items = max_push;
2633
2634         /* dst is the right eb, src is the middle eb */
2635         if (check_sibling_keys(src, dst)) {
2636                 ret = -EUCLEAN;
2637                 btrfs_abort_transaction(trans, ret);
2638                 return ret;
2639         }
2640         ret = btrfs_tree_mod_log_insert_move(dst, push_items, 0, dst_nritems);
2641         BUG_ON(ret < 0);
2642         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
2643                                       btrfs_node_key_ptr_offset(0),
2644                                       (dst_nritems) *
2645                                       sizeof(struct btrfs_key_ptr));
2646
2647         ret = btrfs_tree_mod_log_eb_copy(dst, src, 0, src_nritems - push_items,
2648                                          push_items);
2649         if (ret) {
2650                 btrfs_abort_transaction(trans, ret);
2651                 return ret;
2652         }
2653         copy_extent_buffer(dst, src,
2654                            btrfs_node_key_ptr_offset(0),
2655                            btrfs_node_key_ptr_offset(src_nritems - push_items),
2656                            push_items * sizeof(struct btrfs_key_ptr));
2657
2658         btrfs_set_header_nritems(src, src_nritems - push_items);
2659         btrfs_set_header_nritems(dst, dst_nritems + push_items);
2660
2661         btrfs_mark_buffer_dirty(src);
2662         btrfs_mark_buffer_dirty(dst);
2663
2664         return ret;
2665 }
2666
2667 /*
2668  * helper function to insert a new root level in the tree.
2669  * A new node is allocated, and a single item is inserted to
2670  * point to the existing root
2671  *
2672  * returns zero on success or < 0 on failure.
2673  */
2674 static noinline int insert_new_root(struct btrfs_trans_handle *trans,
2675                            struct btrfs_root *root,
2676                            struct btrfs_path *path, int level)
2677 {
2678         struct btrfs_fs_info *fs_info = root->fs_info;
2679         u64 lower_gen;
2680         struct extent_buffer *lower;
2681         struct extent_buffer *c;
2682         struct extent_buffer *old;
2683         struct btrfs_disk_key lower_key;
2684         int ret;
2685
2686         BUG_ON(path->nodes[level]);
2687         BUG_ON(path->nodes[level-1] != root->node);
2688
2689         lower = path->nodes[level-1];
2690         if (level == 1)
2691                 btrfs_item_key(lower, &lower_key, 0);
2692         else
2693                 btrfs_node_key(lower, &lower_key, 0);
2694
2695         c = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
2696                                    &lower_key, level, root->node->start, 0,
2697                                    BTRFS_NESTING_NEW_ROOT);
2698         if (IS_ERR(c))
2699                 return PTR_ERR(c);
2700
2701         root_add_used(root, fs_info->nodesize);
2702
2703         btrfs_set_header_nritems(c, 1);
2704         btrfs_set_node_key(c, &lower_key, 0);
2705         btrfs_set_node_blockptr(c, 0, lower->start);
2706         lower_gen = btrfs_header_generation(lower);
2707         WARN_ON(lower_gen != trans->transid);
2708
2709         btrfs_set_node_ptr_generation(c, 0, lower_gen);
2710
2711         btrfs_mark_buffer_dirty(c);
2712
2713         old = root->node;
2714         ret = btrfs_tree_mod_log_insert_root(root->node, c, false);
2715         BUG_ON(ret < 0);
2716         rcu_assign_pointer(root->node, c);
2717
2718         /* the super has an extra ref to root->node */
2719         free_extent_buffer(old);
2720
2721         add_root_to_dirty_list(root);
2722         atomic_inc(&c->refs);
2723         path->nodes[level] = c;
2724         path->locks[level] = BTRFS_WRITE_LOCK;
2725         path->slots[level] = 0;
2726         return 0;
2727 }
2728
2729 /*
2730  * worker function to insert a single pointer in a node.
2731  * the node should have enough room for the pointer already
2732  *
2733  * slot and level indicate where you want the key to go, and
2734  * blocknr is the block the key points to.
2735  */
2736 static void insert_ptr(struct btrfs_trans_handle *trans,
2737                        struct btrfs_path *path,
2738                        struct btrfs_disk_key *key, u64 bytenr,
2739                        int slot, int level)
2740 {
2741         struct extent_buffer *lower;
2742         int nritems;
2743         int ret;
2744
2745         BUG_ON(!path->nodes[level]);
2746         btrfs_assert_tree_write_locked(path->nodes[level]);
2747         lower = path->nodes[level];
2748         nritems = btrfs_header_nritems(lower);
2749         BUG_ON(slot > nritems);
2750         BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(trans->fs_info));
2751         if (slot != nritems) {
2752                 if (level) {
2753                         ret = btrfs_tree_mod_log_insert_move(lower, slot + 1,
2754                                         slot, nritems - slot);
2755                         BUG_ON(ret < 0);
2756                 }
2757                 memmove_extent_buffer(lower,
2758                               btrfs_node_key_ptr_offset(slot + 1),
2759                               btrfs_node_key_ptr_offset(slot),
2760                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
2761         }
2762         if (level) {
2763                 ret = btrfs_tree_mod_log_insert_key(lower, slot,
2764                                             BTRFS_MOD_LOG_KEY_ADD, GFP_NOFS);
2765                 BUG_ON(ret < 0);
2766         }
2767         btrfs_set_node_key(lower, key, slot);
2768         btrfs_set_node_blockptr(lower, slot, bytenr);
2769         WARN_ON(trans->transid == 0);
2770         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
2771         btrfs_set_header_nritems(lower, nritems + 1);
2772         btrfs_mark_buffer_dirty(lower);
2773 }
2774
2775 /*
2776  * split the node at the specified level in path in two.
2777  * The path is corrected to point to the appropriate node after the split
2778  *
2779  * Before splitting this tries to make some room in the node by pushing
2780  * left and right, if either one works, it returns right away.
2781  *
2782  * returns 0 on success and < 0 on failure
2783  */
2784 static noinline int split_node(struct btrfs_trans_handle *trans,
2785                                struct btrfs_root *root,
2786                                struct btrfs_path *path, int level)
2787 {
2788         struct btrfs_fs_info *fs_info = root->fs_info;
2789         struct extent_buffer *c;
2790         struct extent_buffer *split;
2791         struct btrfs_disk_key disk_key;
2792         int mid;
2793         int ret;
2794         u32 c_nritems;
2795
2796         c = path->nodes[level];
2797         WARN_ON(btrfs_header_generation(c) != trans->transid);
2798         if (c == root->node) {
2799                 /*
2800                  * trying to split the root, lets make a new one
2801                  *
2802                  * tree mod log: We don't log_removal old root in
2803                  * insert_new_root, because that root buffer will be kept as a
2804                  * normal node. We are going to log removal of half of the
2805                  * elements below with btrfs_tree_mod_log_eb_copy(). We're
2806                  * holding a tree lock on the buffer, which is why we cannot
2807                  * race with other tree_mod_log users.
2808                  */
2809                 ret = insert_new_root(trans, root, path, level + 1);
2810                 if (ret)
2811                         return ret;
2812         } else {
2813                 ret = push_nodes_for_insert(trans, root, path, level);
2814                 c = path->nodes[level];
2815                 if (!ret && btrfs_header_nritems(c) <
2816                     BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3)
2817                         return 0;
2818                 if (ret < 0)
2819                         return ret;
2820         }
2821
2822         c_nritems = btrfs_header_nritems(c);
2823         mid = (c_nritems + 1) / 2;
2824         btrfs_node_key(c, &disk_key, mid);
2825
2826         split = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
2827                                        &disk_key, level, c->start, 0,
2828                                        BTRFS_NESTING_SPLIT);
2829         if (IS_ERR(split))
2830                 return PTR_ERR(split);
2831
2832         root_add_used(root, fs_info->nodesize);
2833         ASSERT(btrfs_header_level(c) == level);
2834
2835         ret = btrfs_tree_mod_log_eb_copy(split, c, 0, mid, c_nritems - mid);
2836         if (ret) {
2837                 btrfs_abort_transaction(trans, ret);
2838                 return ret;
2839         }
2840         copy_extent_buffer(split, c,
2841                            btrfs_node_key_ptr_offset(0),
2842                            btrfs_node_key_ptr_offset(mid),
2843                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
2844         btrfs_set_header_nritems(split, c_nritems - mid);
2845         btrfs_set_header_nritems(c, mid);
2846
2847         btrfs_mark_buffer_dirty(c);
2848         btrfs_mark_buffer_dirty(split);
2849
2850         insert_ptr(trans, path, &disk_key, split->start,
2851                    path->slots[level + 1] + 1, level + 1);
2852
2853         if (path->slots[level] >= mid) {
2854                 path->slots[level] -= mid;
2855                 btrfs_tree_unlock(c);
2856                 free_extent_buffer(c);
2857                 path->nodes[level] = split;
2858                 path->slots[level + 1] += 1;
2859         } else {
2860                 btrfs_tree_unlock(split);
2861                 free_extent_buffer(split);
2862         }
2863         return 0;
2864 }
2865
2866 /*
2867  * how many bytes are required to store the items in a leaf.  start
2868  * and nr indicate which items in the leaf to check.  This totals up the
2869  * space used both by the item structs and the item data
2870  */
2871 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
2872 {
2873         int data_len;
2874         int nritems = btrfs_header_nritems(l);
2875         int end = min(nritems, start + nr) - 1;
2876
2877         if (!nr)
2878                 return 0;
2879         data_len = btrfs_item_offset(l, start) + btrfs_item_size(l, start);
2880         data_len = data_len - btrfs_item_offset(l, end);
2881         data_len += sizeof(struct btrfs_item) * nr;
2882         WARN_ON(data_len < 0);
2883         return data_len;
2884 }
2885
2886 /*
2887  * The space between the end of the leaf items and
2888  * the start of the leaf data.  IOW, how much room
2889  * the leaf has left for both items and data
2890  */
2891 noinline int btrfs_leaf_free_space(struct extent_buffer *leaf)
2892 {
2893         struct btrfs_fs_info *fs_info = leaf->fs_info;
2894         int nritems = btrfs_header_nritems(leaf);
2895         int ret;
2896
2897         ret = BTRFS_LEAF_DATA_SIZE(fs_info) - leaf_space_used(leaf, 0, nritems);
2898         if (ret < 0) {
2899                 btrfs_crit(fs_info,
2900                            "leaf free space ret %d, leaf data size %lu, used %d nritems %d",
2901                            ret,
2902                            (unsigned long) BTRFS_LEAF_DATA_SIZE(fs_info),
2903                            leaf_space_used(leaf, 0, nritems), nritems);
2904         }
2905         return ret;
2906 }
2907
2908 /*
2909  * min slot controls the lowest index we're willing to push to the
2910  * right.  We'll push up to and including min_slot, but no lower
2911  */
2912 static noinline int __push_leaf_right(struct btrfs_path *path,
2913                                       int data_size, int empty,
2914                                       struct extent_buffer *right,
2915                                       int free_space, u32 left_nritems,
2916                                       u32 min_slot)
2917 {
2918         struct btrfs_fs_info *fs_info = right->fs_info;
2919         struct extent_buffer *left = path->nodes[0];
2920         struct extent_buffer *upper = path->nodes[1];
2921         struct btrfs_map_token token;
2922         struct btrfs_disk_key disk_key;
2923         int slot;
2924         u32 i;
2925         int push_space = 0;
2926         int push_items = 0;
2927         u32 nr;
2928         u32 right_nritems;
2929         u32 data_end;
2930         u32 this_item_size;
2931
2932         if (empty)
2933                 nr = 0;
2934         else
2935                 nr = max_t(u32, 1, min_slot);
2936
2937         if (path->slots[0] >= left_nritems)
2938                 push_space += data_size;
2939
2940         slot = path->slots[1];
2941         i = left_nritems - 1;
2942         while (i >= nr) {
2943                 if (!empty && push_items > 0) {
2944                         if (path->slots[0] > i)
2945                                 break;
2946                         if (path->slots[0] == i) {
2947                                 int space = btrfs_leaf_free_space(left);
2948
2949                                 if (space + push_space * 2 > free_space)
2950                                         break;
2951                         }
2952                 }
2953
2954                 if (path->slots[0] == i)
2955                         push_space += data_size;
2956
2957                 this_item_size = btrfs_item_size(left, i);
2958                 if (this_item_size + sizeof(struct btrfs_item) +
2959                     push_space > free_space)
2960                         break;
2961
2962                 push_items++;
2963                 push_space += this_item_size + sizeof(struct btrfs_item);
2964                 if (i == 0)
2965                         break;
2966                 i--;
2967         }
2968
2969         if (push_items == 0)
2970                 goto out_unlock;
2971
2972         WARN_ON(!empty && push_items == left_nritems);
2973
2974         /* push left to right */
2975         right_nritems = btrfs_header_nritems(right);
2976
2977         push_space = btrfs_item_data_end(left, left_nritems - push_items);
2978         push_space -= leaf_data_end(left);
2979
2980         /* make room in the right data area */
2981         data_end = leaf_data_end(right);
2982         memmove_extent_buffer(right,
2983                               BTRFS_LEAF_DATA_OFFSET + data_end - push_space,
2984                               BTRFS_LEAF_DATA_OFFSET + data_end,
2985                               BTRFS_LEAF_DATA_SIZE(fs_info) - data_end);
2986
2987         /* copy from the left data area */
2988         copy_extent_buffer(right, left, BTRFS_LEAF_DATA_OFFSET +
2989                      BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
2990                      BTRFS_LEAF_DATA_OFFSET + leaf_data_end(left),
2991                      push_space);
2992
2993         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
2994                               btrfs_item_nr_offset(0),
2995                               right_nritems * sizeof(struct btrfs_item));
2996
2997         /* copy the items from left to right */
2998         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
2999                    btrfs_item_nr_offset(left_nritems - push_items),
3000                    push_items * sizeof(struct btrfs_item));
3001
3002         /* update the item pointers */
3003         btrfs_init_map_token(&token, right);
3004         right_nritems += push_items;
3005         btrfs_set_header_nritems(right, right_nritems);
3006         push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3007         for (i = 0; i < right_nritems; i++) {
3008                 push_space -= btrfs_token_item_size(&token, i);
3009                 btrfs_set_token_item_offset(&token, i, push_space);
3010         }
3011
3012         left_nritems -= push_items;
3013         btrfs_set_header_nritems(left, left_nritems);
3014
3015         if (left_nritems)
3016                 btrfs_mark_buffer_dirty(left);
3017         else
3018                 btrfs_clean_tree_block(left);
3019
3020         btrfs_mark_buffer_dirty(right);
3021
3022         btrfs_item_key(right, &disk_key, 0);
3023         btrfs_set_node_key(upper, &disk_key, slot + 1);
3024         btrfs_mark_buffer_dirty(upper);
3025
3026         /* then fixup the leaf pointer in the path */
3027         if (path->slots[0] >= left_nritems) {
3028                 path->slots[0] -= left_nritems;
3029                 if (btrfs_header_nritems(path->nodes[0]) == 0)
3030                         btrfs_clean_tree_block(path->nodes[0]);
3031                 btrfs_tree_unlock(path->nodes[0]);
3032                 free_extent_buffer(path->nodes[0]);
3033                 path->nodes[0] = right;
3034                 path->slots[1] += 1;
3035         } else {
3036                 btrfs_tree_unlock(right);
3037                 free_extent_buffer(right);
3038         }
3039         return 0;
3040
3041 out_unlock:
3042         btrfs_tree_unlock(right);
3043         free_extent_buffer(right);
3044         return 1;
3045 }
3046
3047 /*
3048  * push some data in the path leaf to the right, trying to free up at
3049  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
3050  *
3051  * returns 1 if the push failed because the other node didn't have enough
3052  * room, 0 if everything worked out and < 0 if there were major errors.
3053  *
3054  * this will push starting from min_slot to the end of the leaf.  It won't
3055  * push any slot lower than min_slot
3056  */
3057 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
3058                            *root, struct btrfs_path *path,
3059                            int min_data_size, int data_size,
3060                            int empty, u32 min_slot)
3061 {
3062         struct extent_buffer *left = path->nodes[0];
3063         struct extent_buffer *right;
3064         struct extent_buffer *upper;
3065         int slot;
3066         int free_space;
3067         u32 left_nritems;
3068         int ret;
3069
3070         if (!path->nodes[1])
3071                 return 1;
3072
3073         slot = path->slots[1];
3074         upper = path->nodes[1];
3075         if (slot >= btrfs_header_nritems(upper) - 1)
3076                 return 1;
3077
3078         btrfs_assert_tree_write_locked(path->nodes[1]);
3079
3080         right = btrfs_read_node_slot(upper, slot + 1);
3081         /*
3082          * slot + 1 is not valid or we fail to read the right node,
3083          * no big deal, just return.
3084          */
3085         if (IS_ERR(right))
3086                 return 1;
3087
3088         __btrfs_tree_lock(right, BTRFS_NESTING_RIGHT);
3089
3090         free_space = btrfs_leaf_free_space(right);
3091         if (free_space < data_size)
3092                 goto out_unlock;
3093
3094         ret = btrfs_cow_block(trans, root, right, upper,
3095                               slot + 1, &right, BTRFS_NESTING_RIGHT_COW);
3096         if (ret)
3097                 goto out_unlock;
3098
3099         left_nritems = btrfs_header_nritems(left);
3100         if (left_nritems == 0)
3101                 goto out_unlock;
3102
3103         if (check_sibling_keys(left, right)) {
3104                 ret = -EUCLEAN;
3105                 btrfs_tree_unlock(right);
3106                 free_extent_buffer(right);
3107                 return ret;
3108         }
3109         if (path->slots[0] == left_nritems && !empty) {
3110                 /* Key greater than all keys in the leaf, right neighbor has
3111                  * enough room for it and we're not emptying our leaf to delete
3112                  * it, therefore use right neighbor to insert the new item and
3113                  * no need to touch/dirty our left leaf. */
3114                 btrfs_tree_unlock(left);
3115                 free_extent_buffer(left);
3116                 path->nodes[0] = right;
3117                 path->slots[0] = 0;
3118                 path->slots[1]++;
3119                 return 0;
3120         }
3121
3122         return __push_leaf_right(path, min_data_size, empty,
3123                                 right, free_space, left_nritems, min_slot);
3124 out_unlock:
3125         btrfs_tree_unlock(right);
3126         free_extent_buffer(right);
3127         return 1;
3128 }
3129
3130 /*
3131  * push some data in the path leaf to the left, trying to free up at
3132  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
3133  *
3134  * max_slot can put a limit on how far into the leaf we'll push items.  The
3135  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us do all the
3136  * items
3137  */
3138 static noinline int __push_leaf_left(struct btrfs_path *path, int data_size,
3139                                      int empty, struct extent_buffer *left,
3140                                      int free_space, u32 right_nritems,
3141                                      u32 max_slot)
3142 {
3143         struct btrfs_fs_info *fs_info = left->fs_info;
3144         struct btrfs_disk_key disk_key;
3145         struct extent_buffer *right = path->nodes[0];
3146         int i;
3147         int push_space = 0;
3148         int push_items = 0;
3149         u32 old_left_nritems;
3150         u32 nr;
3151         int ret = 0;
3152         u32 this_item_size;
3153         u32 old_left_item_size;
3154         struct btrfs_map_token token;
3155
3156         if (empty)
3157                 nr = min(right_nritems, max_slot);
3158         else
3159                 nr = min(right_nritems - 1, max_slot);
3160
3161         for (i = 0; i < nr; i++) {
3162                 if (!empty && push_items > 0) {
3163                         if (path->slots[0] < i)
3164                                 break;
3165                         if (path->slots[0] == i) {
3166                                 int space = btrfs_leaf_free_space(right);
3167
3168                                 if (space + push_space * 2 > free_space)
3169                                         break;
3170                         }
3171                 }
3172
3173                 if (path->slots[0] == i)
3174                         push_space += data_size;
3175
3176                 this_item_size = btrfs_item_size(right, i);
3177                 if (this_item_size + sizeof(struct btrfs_item) + push_space >
3178                     free_space)
3179                         break;
3180
3181                 push_items++;
3182                 push_space += this_item_size + sizeof(struct btrfs_item);
3183         }
3184
3185         if (push_items == 0) {
3186                 ret = 1;
3187                 goto out;
3188         }
3189         WARN_ON(!empty && push_items == btrfs_header_nritems(right));
3190
3191         /* push data from right to left */
3192         copy_extent_buffer(left, right,
3193                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
3194                            btrfs_item_nr_offset(0),
3195                            push_items * sizeof(struct btrfs_item));
3196
3197         push_space = BTRFS_LEAF_DATA_SIZE(fs_info) -
3198                      btrfs_item_offset(right, push_items - 1);
3199
3200         copy_extent_buffer(left, right, BTRFS_LEAF_DATA_OFFSET +
3201                      leaf_data_end(left) - push_space,
3202                      BTRFS_LEAF_DATA_OFFSET +
3203                      btrfs_item_offset(right, push_items - 1),
3204                      push_space);
3205         old_left_nritems = btrfs_header_nritems(left);
3206         BUG_ON(old_left_nritems <= 0);
3207
3208         btrfs_init_map_token(&token, left);
3209         old_left_item_size = btrfs_item_offset(left, old_left_nritems - 1);
3210         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
3211                 u32 ioff;
3212
3213                 ioff = btrfs_token_item_offset(&token, i);
3214                 btrfs_set_token_item_offset(&token, i,
3215                       ioff - (BTRFS_LEAF_DATA_SIZE(fs_info) - old_left_item_size));
3216         }
3217         btrfs_set_header_nritems(left, old_left_nritems + push_items);
3218
3219         /* fixup right node */
3220         if (push_items > right_nritems)
3221                 WARN(1, KERN_CRIT "push items %d nr %u\n", push_items,
3222                        right_nritems);
3223
3224         if (push_items < right_nritems) {
3225                 push_space = btrfs_item_offset(right, push_items - 1) -
3226                                                   leaf_data_end(right);
3227                 memmove_extent_buffer(right, BTRFS_LEAF_DATA_OFFSET +
3228                                       BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3229                                       BTRFS_LEAF_DATA_OFFSET +
3230                                       leaf_data_end(right), push_space);
3231
3232                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
3233                               btrfs_item_nr_offset(push_items),
3234                              (btrfs_header_nritems(right) - push_items) *
3235                              sizeof(struct btrfs_item));
3236         }
3237
3238         btrfs_init_map_token(&token, right);
3239         right_nritems -= push_items;
3240         btrfs_set_header_nritems(right, right_nritems);
3241         push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3242         for (i = 0; i < right_nritems; i++) {
3243                 push_space = push_space - btrfs_token_item_size(&token, i);
3244                 btrfs_set_token_item_offset(&token, i, push_space);
3245         }
3246
3247         btrfs_mark_buffer_dirty(left);
3248         if (right_nritems)
3249                 btrfs_mark_buffer_dirty(right);
3250         else
3251                 btrfs_clean_tree_block(right);
3252
3253         btrfs_item_key(right, &disk_key, 0);
3254         fixup_low_keys(path, &disk_key, 1);
3255
3256         /* then fixup the leaf pointer in the path */
3257         if (path->slots[0] < push_items) {
3258                 path->slots[0] += old_left_nritems;
3259                 btrfs_tree_unlock(path->nodes[0]);
3260                 free_extent_buffer(path->nodes[0]);
3261                 path->nodes[0] = left;
3262                 path->slots[1] -= 1;
3263         } else {
3264                 btrfs_tree_unlock(left);
3265                 free_extent_buffer(left);
3266                 path->slots[0] -= push_items;
3267         }
3268         BUG_ON(path->slots[0] < 0);
3269         return ret;
3270 out:
3271         btrfs_tree_unlock(left);
3272         free_extent_buffer(left);
3273         return ret;
3274 }
3275
3276 /*
3277  * push some data in the path leaf to the left, trying to free up at
3278  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
3279  *
3280  * max_slot can put a limit on how far into the leaf we'll push items.  The
3281  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us push all the
3282  * items
3283  */
3284 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
3285                           *root, struct btrfs_path *path, int min_data_size,
3286                           int data_size, int empty, u32 max_slot)
3287 {
3288         struct extent_buffer *right = path->nodes[0];
3289         struct extent_buffer *left;
3290         int slot;
3291         int free_space;
3292         u32 right_nritems;
3293         int ret = 0;
3294
3295         slot = path->slots[1];
3296         if (slot == 0)
3297                 return 1;
3298         if (!path->nodes[1])
3299                 return 1;
3300
3301         right_nritems = btrfs_header_nritems(right);
3302         if (right_nritems == 0)
3303                 return 1;
3304
3305         btrfs_assert_tree_write_locked(path->nodes[1]);
3306
3307         left = btrfs_read_node_slot(path->nodes[1], slot - 1);
3308         /*
3309          * slot - 1 is not valid or we fail to read the left node,
3310          * no big deal, just return.
3311          */
3312         if (IS_ERR(left))
3313                 return 1;
3314
3315         __btrfs_tree_lock(left, BTRFS_NESTING_LEFT);
3316
3317         free_space = btrfs_leaf_free_space(left);
3318         if (free_space < data_size) {
3319                 ret = 1;
3320                 goto out;
3321         }
3322
3323         ret = btrfs_cow_block(trans, root, left,
3324                               path->nodes[1], slot - 1, &left,
3325                               BTRFS_NESTING_LEFT_COW);
3326         if (ret) {
3327                 /* we hit -ENOSPC, but it isn't fatal here */
3328                 if (ret == -ENOSPC)
3329                         ret = 1;
3330                 goto out;
3331         }
3332
3333         if (check_sibling_keys(left, right)) {
3334                 ret = -EUCLEAN;
3335                 goto out;
3336         }
3337         return __push_leaf_left(path, min_data_size,
3338                                empty, left, free_space, right_nritems,
3339                                max_slot);
3340 out:
3341         btrfs_tree_unlock(left);
3342         free_extent_buffer(left);
3343         return ret;
3344 }
3345
3346 /*
3347  * split the path's leaf in two, making sure there is at least data_size
3348  * available for the resulting leaf level of the path.
3349  */
3350 static noinline void copy_for_split(struct btrfs_trans_handle *trans,
3351                                     struct btrfs_path *path,
3352                                     struct extent_buffer *l,
3353                                     struct extent_buffer *right,
3354                                     int slot, int mid, int nritems)
3355 {
3356         struct btrfs_fs_info *fs_info = trans->fs_info;
3357         int data_copy_size;
3358         int rt_data_off;
3359         int i;
3360         struct btrfs_disk_key disk_key;
3361         struct btrfs_map_token token;
3362
3363         nritems = nritems - mid;
3364         btrfs_set_header_nritems(right, nritems);
3365         data_copy_size = btrfs_item_data_end(l, mid) - leaf_data_end(l);
3366
3367         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
3368                            btrfs_item_nr_offset(mid),
3369                            nritems * sizeof(struct btrfs_item));
3370
3371         copy_extent_buffer(right, l,
3372                      BTRFS_LEAF_DATA_OFFSET + BTRFS_LEAF_DATA_SIZE(fs_info) -
3373                      data_copy_size, BTRFS_LEAF_DATA_OFFSET +
3374                      leaf_data_end(l), data_copy_size);
3375
3376         rt_data_off = BTRFS_LEAF_DATA_SIZE(fs_info) - btrfs_item_data_end(l, mid);
3377
3378         btrfs_init_map_token(&token, right);
3379         for (i = 0; i < nritems; i++) {
3380                 u32 ioff;
3381
3382                 ioff = btrfs_token_item_offset(&token, i);
3383                 btrfs_set_token_item_offset(&token, i, ioff + rt_data_off);
3384         }
3385
3386         btrfs_set_header_nritems(l, mid);
3387         btrfs_item_key(right, &disk_key, 0);
3388         insert_ptr(trans, path, &disk_key, right->start, path->slots[1] + 1, 1);
3389
3390         btrfs_mark_buffer_dirty(right);
3391         btrfs_mark_buffer_dirty(l);
3392         BUG_ON(path->slots[0] != slot);
3393
3394         if (mid <= slot) {
3395                 btrfs_tree_unlock(path->nodes[0]);
3396                 free_extent_buffer(path->nodes[0]);
3397                 path->nodes[0] = right;
3398                 path->slots[0] -= mid;
3399                 path->slots[1] += 1;
3400         } else {
3401                 btrfs_tree_unlock(right);
3402                 free_extent_buffer(right);
3403         }
3404
3405         BUG_ON(path->slots[0] < 0);
3406 }
3407
3408 /*
3409  * double splits happen when we need to insert a big item in the middle
3410  * of a leaf.  A double split can leave us with 3 mostly empty leaves:
3411  * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
3412  *          A                 B                 C
3413  *
3414  * We avoid this by trying to push the items on either side of our target
3415  * into the adjacent leaves.  If all goes well we can avoid the double split
3416  * completely.
3417  */
3418 static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
3419                                           struct btrfs_root *root,
3420                                           struct btrfs_path *path,
3421                                           int data_size)
3422 {
3423         int ret;
3424         int progress = 0;
3425         int slot;
3426         u32 nritems;
3427         int space_needed = data_size;
3428
3429         slot = path->slots[0];
3430         if (slot < btrfs_header_nritems(path->nodes[0]))
3431                 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
3432
3433         /*
3434          * try to push all the items after our slot into the
3435          * right leaf
3436          */
3437         ret = push_leaf_right(trans, root, path, 1, space_needed, 0, slot);
3438         if (ret < 0)
3439                 return ret;
3440
3441         if (ret == 0)
3442                 progress++;
3443
3444         nritems = btrfs_header_nritems(path->nodes[0]);
3445         /*
3446          * our goal is to get our slot at the start or end of a leaf.  If
3447          * we've done so we're done
3448          */
3449         if (path->slots[0] == 0 || path->slots[0] == nritems)
3450                 return 0;
3451
3452         if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
3453                 return 0;
3454
3455         /* try to push all the items before our slot into the next leaf */
3456         slot = path->slots[0];
3457         space_needed = data_size;
3458         if (slot > 0)
3459                 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
3460         ret = push_leaf_left(trans, root, path, 1, space_needed, 0, slot);
3461         if (ret < 0)
3462                 return ret;
3463
3464         if (ret == 0)
3465                 progress++;
3466
3467         if (progress)
3468                 return 0;
3469         return 1;
3470 }
3471
3472 /*
3473  * split the path's leaf in two, making sure there is at least data_size
3474  * available for the resulting leaf level of the path.
3475  *
3476  * returns 0 if all went well and < 0 on failure.
3477  */
3478 static noinline int split_leaf(struct btrfs_trans_handle *trans,
3479                                struct btrfs_root *root,
3480                                const struct btrfs_key *ins_key,
3481                                struct btrfs_path *path, int data_size,
3482                                int extend)
3483 {
3484         struct btrfs_disk_key disk_key;
3485         struct extent_buffer *l;
3486         u32 nritems;
3487         int mid;
3488         int slot;
3489         struct extent_buffer *right;
3490         struct btrfs_fs_info *fs_info = root->fs_info;
3491         int ret = 0;
3492         int wret;
3493         int split;
3494         int num_doubles = 0;
3495         int tried_avoid_double = 0;
3496
3497         l = path->nodes[0];
3498         slot = path->slots[0];
3499         if (extend && data_size + btrfs_item_size(l, slot) +
3500             sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(fs_info))
3501                 return -EOVERFLOW;
3502
3503         /* first try to make some room by pushing left and right */
3504         if (data_size && path->nodes[1]) {
3505                 int space_needed = data_size;
3506
3507                 if (slot < btrfs_header_nritems(l))
3508                         space_needed -= btrfs_leaf_free_space(l);
3509
3510                 wret = push_leaf_right(trans, root, path, space_needed,
3511                                        space_needed, 0, 0);
3512                 if (wret < 0)
3513                         return wret;
3514                 if (wret) {
3515                         space_needed = data_size;
3516                         if (slot > 0)
3517                                 space_needed -= btrfs_leaf_free_space(l);
3518                         wret = push_leaf_left(trans, root, path, space_needed,
3519                                               space_needed, 0, (u32)-1);
3520                         if (wret < 0)
3521                                 return wret;
3522                 }
3523                 l = path->nodes[0];
3524
3525                 /* did the pushes work? */
3526                 if (btrfs_leaf_free_space(l) >= data_size)
3527                         return 0;
3528         }
3529
3530         if (!path->nodes[1]) {
3531                 ret = insert_new_root(trans, root, path, 1);
3532                 if (ret)
3533                         return ret;
3534         }
3535 again:
3536         split = 1;
3537         l = path->nodes[0];
3538         slot = path->slots[0];
3539         nritems = btrfs_header_nritems(l);
3540         mid = (nritems + 1) / 2;
3541
3542         if (mid <= slot) {
3543                 if (nritems == 1 ||
3544                     leaf_space_used(l, mid, nritems - mid) + data_size >
3545                         BTRFS_LEAF_DATA_SIZE(fs_info)) {
3546                         if (slot >= nritems) {
3547                                 split = 0;
3548                         } else {
3549                                 mid = slot;
3550                                 if (mid != nritems &&
3551                                     leaf_space_used(l, mid, nritems - mid) +
3552                                     data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
3553                                         if (data_size && !tried_avoid_double)
3554                                                 goto push_for_double;
3555                                         split = 2;
3556                                 }
3557                         }
3558                 }
3559         } else {
3560                 if (leaf_space_used(l, 0, mid) + data_size >
3561                         BTRFS_LEAF_DATA_SIZE(fs_info)) {
3562                         if (!extend && data_size && slot == 0) {
3563                                 split = 0;
3564                         } else if ((extend || !data_size) && slot == 0) {
3565                                 mid = 1;
3566                         } else {
3567                                 mid = slot;
3568                                 if (mid != nritems &&
3569                                     leaf_space_used(l, mid, nritems - mid) +
3570                                     data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
3571                                         if (data_size && !tried_avoid_double)
3572                                                 goto push_for_double;
3573                                         split = 2;
3574                                 }
3575                         }
3576                 }
3577         }
3578
3579         if (split == 0)
3580                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
3581         else
3582                 btrfs_item_key(l, &disk_key, mid);
3583
3584         /*
3585          * We have to about BTRFS_NESTING_NEW_ROOT here if we've done a double
3586          * split, because we're only allowed to have MAX_LOCKDEP_SUBCLASSES
3587          * subclasses, which is 8 at the time of this patch, and we've maxed it
3588          * out.  In the future we could add a
3589          * BTRFS_NESTING_SPLIT_THE_SPLITTENING if we need to, but for now just
3590          * use BTRFS_NESTING_NEW_ROOT.
3591          */
3592         right = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
3593                                        &disk_key, 0, l->start, 0,
3594                                        num_doubles ? BTRFS_NESTING_NEW_ROOT :
3595                                        BTRFS_NESTING_SPLIT);
3596         if (IS_ERR(right))
3597                 return PTR_ERR(right);
3598
3599         root_add_used(root, fs_info->nodesize);
3600
3601         if (split == 0) {
3602                 if (mid <= slot) {
3603                         btrfs_set_header_nritems(right, 0);
3604                         insert_ptr(trans, path, &disk_key,
3605                                    right->start, path->slots[1] + 1, 1);
3606                         btrfs_tree_unlock(path->nodes[0]);
3607                         free_extent_buffer(path->nodes[0]);
3608                         path->nodes[0] = right;
3609                         path->slots[0] = 0;
3610                         path->slots[1] += 1;
3611                 } else {
3612                         btrfs_set_header_nritems(right, 0);
3613                         insert_ptr(trans, path, &disk_key,
3614                                    right->start, path->slots[1], 1);
3615                         btrfs_tree_unlock(path->nodes[0]);
3616                         free_extent_buffer(path->nodes[0]);
3617                         path->nodes[0] = right;
3618                         path->slots[0] = 0;
3619                         if (path->slots[1] == 0)
3620                                 fixup_low_keys(path, &disk_key, 1);
3621                 }
3622                 /*
3623                  * We create a new leaf 'right' for the required ins_len and
3624                  * we'll do btrfs_mark_buffer_dirty() on this leaf after copying
3625                  * the content of ins_len to 'right'.
3626                  */
3627                 return ret;
3628         }
3629
3630         copy_for_split(trans, path, l, right, slot, mid, nritems);
3631
3632         if (split == 2) {
3633                 BUG_ON(num_doubles != 0);
3634                 num_doubles++;
3635                 goto again;
3636         }
3637
3638         return 0;
3639
3640 push_for_double:
3641         push_for_double_split(trans, root, path, data_size);
3642         tried_avoid_double = 1;
3643         if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
3644                 return 0;
3645         goto again;
3646 }
3647
3648 static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
3649                                          struct btrfs_root *root,
3650                                          struct btrfs_path *path, int ins_len)
3651 {
3652         struct btrfs_key key;
3653         struct extent_buffer *leaf;
3654         struct btrfs_file_extent_item *fi;
3655         u64 extent_len = 0;
3656         u32 item_size;
3657         int ret;
3658
3659         leaf = path->nodes[0];
3660         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3661
3662         BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
3663                key.type != BTRFS_EXTENT_CSUM_KEY);
3664
3665         if (btrfs_leaf_free_space(leaf) >= ins_len)
3666                 return 0;
3667
3668         item_size = btrfs_item_size(leaf, path->slots[0]);
3669         if (key.type == BTRFS_EXTENT_DATA_KEY) {
3670                 fi = btrfs_item_ptr(leaf, path->slots[0],
3671                                     struct btrfs_file_extent_item);
3672                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
3673         }
3674         btrfs_release_path(path);
3675
3676         path->keep_locks = 1;
3677         path->search_for_split = 1;
3678         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
3679         path->search_for_split = 0;
3680         if (ret > 0)
3681                 ret = -EAGAIN;
3682         if (ret < 0)
3683                 goto err;
3684
3685         ret = -EAGAIN;
3686         leaf = path->nodes[0];
3687         /* if our item isn't there, return now */
3688         if (item_size != btrfs_item_size(leaf, path->slots[0]))
3689                 goto err;
3690
3691         /* the leaf has  changed, it now has room.  return now */
3692         if (btrfs_leaf_free_space(path->nodes[0]) >= ins_len)
3693                 goto err;
3694
3695         if (key.type == BTRFS_EXTENT_DATA_KEY) {
3696                 fi = btrfs_item_ptr(leaf, path->slots[0],
3697                                     struct btrfs_file_extent_item);
3698                 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
3699                         goto err;
3700         }
3701
3702         ret = split_leaf(trans, root, &key, path, ins_len, 1);
3703         if (ret)
3704                 goto err;
3705
3706         path->keep_locks = 0;
3707         btrfs_unlock_up_safe(path, 1);
3708         return 0;
3709 err:
3710         path->keep_locks = 0;
3711         return ret;
3712 }
3713
3714 static noinline int split_item(struct btrfs_path *path,
3715                                const struct btrfs_key *new_key,
3716                                unsigned long split_offset)
3717 {
3718         struct extent_buffer *leaf;
3719         int orig_slot, slot;
3720         char *buf;
3721         u32 nritems;
3722         u32 item_size;
3723         u32 orig_offset;
3724         struct btrfs_disk_key disk_key;
3725
3726         leaf = path->nodes[0];
3727         BUG_ON(btrfs_leaf_free_space(leaf) < sizeof(struct btrfs_item));
3728
3729         orig_slot = path->slots[0];
3730         orig_offset = btrfs_item_offset(leaf, path->slots[0]);
3731         item_size = btrfs_item_size(leaf, path->slots[0]);
3732
3733         buf = kmalloc(item_size, GFP_NOFS);
3734         if (!buf)
3735                 return -ENOMEM;
3736
3737         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
3738                             path->slots[0]), item_size);
3739
3740         slot = path->slots[0] + 1;
3741         nritems = btrfs_header_nritems(leaf);
3742         if (slot != nritems) {
3743                 /* shift the items */
3744                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
3745                                 btrfs_item_nr_offset(slot),
3746                                 (nritems - slot) * sizeof(struct btrfs_item));
3747         }
3748
3749         btrfs_cpu_key_to_disk(&disk_key, new_key);
3750         btrfs_set_item_key(leaf, &disk_key, slot);
3751
3752         btrfs_set_item_offset(leaf, slot, orig_offset);
3753         btrfs_set_item_size(leaf, slot, item_size - split_offset);
3754
3755         btrfs_set_item_offset(leaf, orig_slot,
3756                                  orig_offset + item_size - split_offset);
3757         btrfs_set_item_size(leaf, orig_slot, split_offset);
3758
3759         btrfs_set_header_nritems(leaf, nritems + 1);
3760
3761         /* write the data for the start of the original item */
3762         write_extent_buffer(leaf, buf,
3763                             btrfs_item_ptr_offset(leaf, path->slots[0]),
3764                             split_offset);
3765
3766         /* write the data for the new item */
3767         write_extent_buffer(leaf, buf + split_offset,
3768                             btrfs_item_ptr_offset(leaf, slot),
3769                             item_size - split_offset);
3770         btrfs_mark_buffer_dirty(leaf);
3771
3772         BUG_ON(btrfs_leaf_free_space(leaf) < 0);
3773         kfree(buf);
3774         return 0;
3775 }
3776
3777 /*
3778  * This function splits a single item into two items,
3779  * giving 'new_key' to the new item and splitting the
3780  * old one at split_offset (from the start of the item).
3781  *
3782  * The path may be released by this operation.  After
3783  * the split, the path is pointing to the old item.  The
3784  * new item is going to be in the same node as the old one.
3785  *
3786  * Note, the item being split must be smaller enough to live alone on
3787  * a tree block with room for one extra struct btrfs_item
3788  *
3789  * This allows us to split the item in place, keeping a lock on the
3790  * leaf the entire time.
3791  */
3792 int btrfs_split_item(struct btrfs_trans_handle *trans,
3793                      struct btrfs_root *root,
3794                      struct btrfs_path *path,
3795                      const struct btrfs_key *new_key,
3796                      unsigned long split_offset)
3797 {
3798         int ret;
3799         ret = setup_leaf_for_split(trans, root, path,
3800                                    sizeof(struct btrfs_item));
3801         if (ret)
3802                 return ret;
3803
3804         ret = split_item(path, new_key, split_offset);
3805         return ret;
3806 }
3807
3808 /*
3809  * make the item pointed to by the path smaller.  new_size indicates
3810  * how small to make it, and from_end tells us if we just chop bytes
3811  * off the end of the item or if we shift the item to chop bytes off
3812  * the front.
3813  */
3814 void btrfs_truncate_item(struct btrfs_path *path, u32 new_size, int from_end)
3815 {
3816         int slot;
3817         struct extent_buffer *leaf;
3818         u32 nritems;
3819         unsigned int data_end;
3820         unsigned int old_data_start;
3821         unsigned int old_size;
3822         unsigned int size_diff;
3823         int i;
3824         struct btrfs_map_token token;
3825
3826         leaf = path->nodes[0];
3827         slot = path->slots[0];
3828
3829         old_size = btrfs_item_size(leaf, slot);
3830         if (old_size == new_size)
3831                 return;
3832
3833         nritems = btrfs_header_nritems(leaf);
3834         data_end = leaf_data_end(leaf);
3835
3836         old_data_start = btrfs_item_offset(leaf, slot);
3837
3838         size_diff = old_size - new_size;
3839
3840         BUG_ON(slot < 0);
3841         BUG_ON(slot >= nritems);
3842
3843         /*
3844          * item0..itemN ... dataN.offset..dataN.size .. data0.size
3845          */
3846         /* first correct the data pointers */
3847         btrfs_init_map_token(&token, leaf);
3848         for (i = slot; i < nritems; i++) {
3849                 u32 ioff;
3850
3851                 ioff = btrfs_token_item_offset(&token, i);
3852                 btrfs_set_token_item_offset(&token, i, ioff + size_diff);
3853         }
3854
3855         /* shift the data */
3856         if (from_end) {
3857                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
3858                               data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
3859                               data_end, old_data_start + new_size - data_end);
3860         } else {
3861                 struct btrfs_disk_key disk_key;
3862                 u64 offset;
3863
3864                 btrfs_item_key(leaf, &disk_key, slot);
3865
3866                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
3867                         unsigned long ptr;
3868                         struct btrfs_file_extent_item *fi;
3869
3870                         fi = btrfs_item_ptr(leaf, slot,
3871                                             struct btrfs_file_extent_item);
3872                         fi = (struct btrfs_file_extent_item *)(
3873                              (unsigned long)fi - size_diff);
3874
3875                         if (btrfs_file_extent_type(leaf, fi) ==
3876                             BTRFS_FILE_EXTENT_INLINE) {
3877                                 ptr = btrfs_item_ptr_offset(leaf, slot);
3878                                 memmove_extent_buffer(leaf, ptr,
3879                                       (unsigned long)fi,
3880                                       BTRFS_FILE_EXTENT_INLINE_DATA_START);
3881                         }
3882                 }
3883
3884                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
3885                               data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
3886                               data_end, old_data_start - data_end);
3887
3888                 offset = btrfs_disk_key_offset(&disk_key);
3889                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
3890                 btrfs_set_item_key(leaf, &disk_key, slot);
3891                 if (slot == 0)
3892                         fixup_low_keys(path, &disk_key, 1);
3893         }
3894
3895         btrfs_set_item_size(leaf, slot, new_size);
3896         btrfs_mark_buffer_dirty(leaf);
3897
3898         if (btrfs_leaf_free_space(leaf) < 0) {
3899                 btrfs_print_leaf(leaf);
3900                 BUG();
3901         }
3902 }
3903
3904 /*
3905  * make the item pointed to by the path bigger, data_size is the added size.
3906  */
3907 void btrfs_extend_item(struct btrfs_path *path, u32 data_size)
3908 {
3909         int slot;
3910         struct extent_buffer *leaf;
3911         u32 nritems;
3912         unsigned int data_end;
3913         unsigned int old_data;
3914         unsigned int old_size;
3915         int i;
3916         struct btrfs_map_token token;
3917
3918         leaf = path->nodes[0];
3919
3920         nritems = btrfs_header_nritems(leaf);
3921         data_end = leaf_data_end(leaf);
3922
3923         if (btrfs_leaf_free_space(leaf) < data_size) {
3924                 btrfs_print_leaf(leaf);
3925                 BUG();
3926         }
3927         slot = path->slots[0];
3928         old_data = btrfs_item_data_end(leaf, slot);
3929
3930         BUG_ON(slot < 0);
3931         if (slot >= nritems) {
3932                 btrfs_print_leaf(leaf);
3933                 btrfs_crit(leaf->fs_info, "slot %d too large, nritems %d",
3934                            slot, nritems);
3935                 BUG();
3936         }
3937
3938         /*
3939          * item0..itemN ... dataN.offset..dataN.size .. data0.size
3940          */
3941         /* first correct the data pointers */
3942         btrfs_init_map_token(&token, leaf);
3943         for (i = slot; i < nritems; i++) {
3944                 u32 ioff;
3945
3946                 ioff = btrfs_token_item_offset(&token, i);
3947                 btrfs_set_token_item_offset(&token, i, ioff - data_size);
3948         }
3949
3950         /* shift the data */
3951         memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
3952                       data_end - data_size, BTRFS_LEAF_DATA_OFFSET +
3953                       data_end, old_data - data_end);
3954
3955         data_end = old_data;
3956         old_size = btrfs_item_size(leaf, slot);
3957         btrfs_set_item_size(leaf, slot, old_size + data_size);
3958         btrfs_mark_buffer_dirty(leaf);
3959
3960         if (btrfs_leaf_free_space(leaf) < 0) {
3961                 btrfs_print_leaf(leaf);
3962                 BUG();
3963         }
3964 }
3965
3966 /**
3967  * setup_items_for_insert - Helper called before inserting one or more items
3968  * to a leaf. Main purpose is to save stack depth by doing the bulk of the work
3969  * in a function that doesn't call btrfs_search_slot
3970  *
3971  * @root:       root we are inserting items to
3972  * @path:       points to the leaf/slot where we are going to insert new items
3973  * @batch:      information about the batch of items to insert
3974  */
3975 static void setup_items_for_insert(struct btrfs_root *root, struct btrfs_path *path,
3976                                    const struct btrfs_item_batch *batch)
3977 {
3978         struct btrfs_fs_info *fs_info = root->fs_info;
3979         int i;
3980         u32 nritems;
3981         unsigned int data_end;
3982         struct btrfs_disk_key disk_key;
3983         struct extent_buffer *leaf;
3984         int slot;
3985         struct btrfs_map_token token;
3986         u32 total_size;
3987
3988         /*
3989          * Before anything else, update keys in the parent and other ancestors
3990          * if needed, then release the write locks on them, so that other tasks
3991          * can use them while we modify the leaf.
3992          */
3993         if (path->slots[0] == 0) {
3994                 btrfs_cpu_key_to_disk(&disk_key, &batch->keys[0]);
3995                 fixup_low_keys(path, &disk_key, 1);
3996         }
3997         btrfs_unlock_up_safe(path, 1);
3998
3999         leaf = path->nodes[0];
4000         slot = path->slots[0];
4001
4002         nritems = btrfs_header_nritems(leaf);
4003         data_end = leaf_data_end(leaf);
4004         total_size = batch->total_data_size + (batch->nr * sizeof(struct btrfs_item));
4005
4006         if (btrfs_leaf_free_space(leaf) < total_size) {
4007                 btrfs_print_leaf(leaf);
4008                 btrfs_crit(fs_info, "not enough freespace need %u have %d",
4009                            total_size, btrfs_leaf_free_space(leaf));
4010                 BUG();
4011         }
4012
4013         btrfs_init_map_token(&token, leaf);
4014         if (slot != nritems) {
4015                 unsigned int old_data = btrfs_item_data_end(leaf, slot);
4016
4017                 if (old_data < data_end) {
4018                         btrfs_print_leaf(leaf);
4019                         btrfs_crit(fs_info,
4020                 "item at slot %d with data offset %u beyond data end of leaf %u",
4021                                    slot, old_data, data_end);
4022                         BUG();
4023                 }
4024                 /*
4025                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
4026                  */
4027                 /* first correct the data pointers */
4028                 for (i = slot; i < nritems; i++) {
4029                         u32 ioff;
4030
4031                         ioff = btrfs_token_item_offset(&token, i);
4032                         btrfs_set_token_item_offset(&token, i,
4033                                                        ioff - batch->total_data_size);
4034                 }
4035                 /* shift the items */
4036                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + batch->nr),
4037                               btrfs_item_nr_offset(slot),
4038                               (nritems - slot) * sizeof(struct btrfs_item));
4039
4040                 /* shift the data */
4041                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4042                                       data_end - batch->total_data_size,
4043                                       BTRFS_LEAF_DATA_OFFSET + data_end,
4044                                       old_data - data_end);
4045                 data_end = old_data;
4046         }
4047
4048         /* setup the item for the new data */
4049         for (i = 0; i < batch->nr; i++) {
4050                 btrfs_cpu_key_to_disk(&disk_key, &batch->keys[i]);
4051                 btrfs_set_item_key(leaf, &disk_key, slot + i);
4052                 data_end -= batch->data_sizes[i];
4053                 btrfs_set_token_item_offset(&token, slot + i, data_end);
4054                 btrfs_set_token_item_size(&token, slot + i, batch->data_sizes[i]);
4055         }
4056
4057         btrfs_set_header_nritems(leaf, nritems + batch->nr);
4058         btrfs_mark_buffer_dirty(leaf);
4059
4060         if (btrfs_leaf_free_space(leaf) < 0) {
4061                 btrfs_print_leaf(leaf);
4062                 BUG();
4063         }
4064 }
4065
4066 /*
4067  * Insert a new item into a leaf.
4068  *
4069  * @root:      The root of the btree.
4070  * @path:      A path pointing to the target leaf and slot.
4071  * @key:       The key of the new item.
4072  * @data_size: The size of the data associated with the new key.
4073  */
4074 void btrfs_setup_item_for_insert(struct btrfs_root *root,
4075                                  struct btrfs_path *path,
4076                                  const struct btrfs_key *key,
4077                                  u32 data_size)
4078 {
4079         struct btrfs_item_batch batch;
4080
4081         batch.keys = key;
4082         batch.data_sizes = &data_size;
4083         batch.total_data_size = data_size;
4084         batch.nr = 1;
4085
4086         setup_items_for_insert(root, path, &batch);
4087 }
4088
4089 /*
4090  * Given a key and some data, insert items into the tree.
4091  * This does all the path init required, making room in the tree if needed.
4092  */
4093 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
4094                             struct btrfs_root *root,
4095                             struct btrfs_path *path,
4096                             const struct btrfs_item_batch *batch)
4097 {
4098         int ret = 0;
4099         int slot;
4100         u32 total_size;
4101
4102         total_size = batch->total_data_size + (batch->nr * sizeof(struct btrfs_item));
4103         ret = btrfs_search_slot(trans, root, &batch->keys[0], path, total_size, 1);
4104         if (ret == 0)
4105                 return -EEXIST;
4106         if (ret < 0)
4107                 return ret;
4108
4109         slot = path->slots[0];
4110         BUG_ON(slot < 0);
4111
4112         setup_items_for_insert(root, path, batch);
4113         return 0;
4114 }
4115
4116 /*
4117  * Given a key and some data, insert an item into the tree.
4118  * This does all the path init required, making room in the tree if needed.
4119  */
4120 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4121                       const struct btrfs_key *cpu_key, void *data,
4122                       u32 data_size)
4123 {
4124         int ret = 0;
4125         struct btrfs_path *path;
4126         struct extent_buffer *leaf;
4127         unsigned long ptr;
4128
4129         path = btrfs_alloc_path();
4130         if (!path)
4131                 return -ENOMEM;
4132         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
4133         if (!ret) {
4134                 leaf = path->nodes[0];
4135                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
4136                 write_extent_buffer(leaf, data, ptr, data_size);
4137                 btrfs_mark_buffer_dirty(leaf);
4138         }
4139         btrfs_free_path(path);
4140         return ret;
4141 }
4142
4143 /*
4144  * This function duplicates an item, giving 'new_key' to the new item.
4145  * It guarantees both items live in the same tree leaf and the new item is
4146  * contiguous with the original item.
4147  *
4148  * This allows us to split a file extent in place, keeping a lock on the leaf
4149  * the entire time.
4150  */
4151 int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
4152                          struct btrfs_root *root,
4153                          struct btrfs_path *path,
4154                          const struct btrfs_key *new_key)
4155 {
4156         struct extent_buffer *leaf;
4157         int ret;
4158         u32 item_size;
4159
4160         leaf = path->nodes[0];
4161         item_size = btrfs_item_size(leaf, path->slots[0]);
4162         ret = setup_leaf_for_split(trans, root, path,
4163                                    item_size + sizeof(struct btrfs_item));
4164         if (ret)
4165                 return ret;
4166
4167         path->slots[0]++;
4168         btrfs_setup_item_for_insert(root, path, new_key, item_size);
4169         leaf = path->nodes[0];
4170         memcpy_extent_buffer(leaf,
4171                              btrfs_item_ptr_offset(leaf, path->slots[0]),
4172                              btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
4173                              item_size);
4174         return 0;
4175 }
4176
4177 /*
4178  * delete the pointer from a given node.
4179  *
4180  * the tree should have been previously balanced so the deletion does not
4181  * empty a node.
4182  */
4183 static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
4184                     int level, int slot)
4185 {
4186         struct extent_buffer *parent = path->nodes[level];
4187         u32 nritems;
4188         int ret;
4189
4190         nritems = btrfs_header_nritems(parent);
4191         if (slot != nritems - 1) {
4192                 if (level) {
4193                         ret = btrfs_tree_mod_log_insert_move(parent, slot,
4194                                         slot + 1, nritems - slot - 1);
4195                         BUG_ON(ret < 0);
4196                 }
4197                 memmove_extent_buffer(parent,
4198                               btrfs_node_key_ptr_offset(slot),
4199                               btrfs_node_key_ptr_offset(slot + 1),
4200                               sizeof(struct btrfs_key_ptr) *
4201                               (nritems - slot - 1));
4202         } else if (level) {
4203                 ret = btrfs_tree_mod_log_insert_key(parent, slot,
4204                                 BTRFS_MOD_LOG_KEY_REMOVE, GFP_NOFS);
4205                 BUG_ON(ret < 0);
4206         }
4207
4208         nritems--;
4209         btrfs_set_header_nritems(parent, nritems);
4210         if (nritems == 0 && parent == root->node) {
4211                 BUG_ON(btrfs_header_level(root->node) != 1);
4212                 /* just turn the root into a leaf and break */
4213                 btrfs_set_header_level(root->node, 0);
4214         } else if (slot == 0) {
4215                 struct btrfs_disk_key disk_key;
4216
4217                 btrfs_node_key(parent, &disk_key, 0);
4218                 fixup_low_keys(path, &disk_key, level + 1);
4219         }
4220         btrfs_mark_buffer_dirty(parent);
4221 }
4222
4223 /*
4224  * a helper function to delete the leaf pointed to by path->slots[1] and
4225  * path->nodes[1].
4226  *
4227  * This deletes the pointer in path->nodes[1] and frees the leaf
4228  * block extent.  zero is returned if it all worked out, < 0 otherwise.
4229  *
4230  * The path must have already been setup for deleting the leaf, including
4231  * all the proper balancing.  path->nodes[1] must be locked.
4232  */
4233 static noinline void btrfs_del_leaf(struct btrfs_trans_handle *trans,
4234                                     struct btrfs_root *root,
4235                                     struct btrfs_path *path,
4236                                     struct extent_buffer *leaf)
4237 {
4238         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
4239         del_ptr(root, path, 1, path->slots[1]);
4240
4241         /*
4242          * btrfs_free_extent is expensive, we want to make sure we
4243          * aren't holding any locks when we call it
4244          */
4245         btrfs_unlock_up_safe(path, 0);
4246
4247         root_sub_used(root, leaf->len);
4248
4249         atomic_inc(&leaf->refs);
4250         btrfs_free_tree_block(trans, btrfs_root_id(root), leaf, 0, 1);
4251         free_extent_buffer_stale(leaf);
4252 }
4253 /*
4254  * delete the item at the leaf level in path.  If that empties
4255  * the leaf, remove it from the tree
4256  */
4257 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4258                     struct btrfs_path *path, int slot, int nr)
4259 {
4260         struct btrfs_fs_info *fs_info = root->fs_info;
4261         struct extent_buffer *leaf;
4262         int ret = 0;
4263         int wret;
4264         u32 nritems;
4265
4266         leaf = path->nodes[0];
4267         nritems = btrfs_header_nritems(leaf);
4268
4269         if (slot + nr != nritems) {
4270                 const u32 last_off = btrfs_item_offset(leaf, slot + nr - 1);
4271                 const int data_end = leaf_data_end(leaf);
4272                 struct btrfs_map_token token;
4273                 u32 dsize = 0;
4274                 int i;
4275
4276                 for (i = 0; i < nr; i++)
4277                         dsize += btrfs_item_size(leaf, slot + i);
4278
4279                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4280                               data_end + dsize,
4281                               BTRFS_LEAF_DATA_OFFSET + data_end,
4282                               last_off - data_end);
4283
4284                 btrfs_init_map_token(&token, leaf);
4285                 for (i = slot + nr; i < nritems; i++) {
4286                         u32 ioff;
4287
4288                         ioff = btrfs_token_item_offset(&token, i);
4289                         btrfs_set_token_item_offset(&token, i, ioff + dsize);
4290                 }
4291
4292                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
4293                               btrfs_item_nr_offset(slot + nr),
4294                               sizeof(struct btrfs_item) *
4295                               (nritems - slot - nr));
4296         }
4297         btrfs_set_header_nritems(leaf, nritems - nr);
4298         nritems -= nr;
4299
4300         /* delete the leaf if we've emptied it */
4301         if (nritems == 0) {
4302                 if (leaf == root->node) {
4303                         btrfs_set_header_level(leaf, 0);
4304                 } else {
4305                         btrfs_clean_tree_block(leaf);
4306                         btrfs_del_leaf(trans, root, path, leaf);
4307                 }
4308         } else {
4309                 int used = leaf_space_used(leaf, 0, nritems);
4310                 if (slot == 0) {
4311                         struct btrfs_disk_key disk_key;
4312
4313                         btrfs_item_key(leaf, &disk_key, 0);
4314                         fixup_low_keys(path, &disk_key, 1);
4315                 }
4316
4317                 /*
4318                  * Try to delete the leaf if it is mostly empty. We do this by
4319                  * trying to move all its items into its left and right neighbours.
4320                  * If we can't move all the items, then we don't delete it - it's
4321                  * not ideal, but future insertions might fill the leaf with more
4322                  * items, or items from other leaves might be moved later into our
4323                  * leaf due to deletions on those leaves.
4324                  */
4325                 if (used < BTRFS_LEAF_DATA_SIZE(fs_info) / 3) {
4326                         u32 min_push_space;
4327
4328                         /* push_leaf_left fixes the path.
4329                          * make sure the path still points to our leaf
4330                          * for possible call to del_ptr below
4331                          */
4332                         slot = path->slots[1];
4333                         atomic_inc(&leaf->refs);
4334                         /*
4335                          * We want to be able to at least push one item to the
4336                          * left neighbour leaf, and that's the first item.
4337                          */
4338                         min_push_space = sizeof(struct btrfs_item) +
4339                                 btrfs_item_size(leaf, 0);
4340                         wret = push_leaf_left(trans, root, path, 0,
4341                                               min_push_space, 1, (u32)-1);
4342                         if (wret < 0 && wret != -ENOSPC)
4343                                 ret = wret;
4344
4345                         if (path->nodes[0] == leaf &&
4346                             btrfs_header_nritems(leaf)) {
4347                                 /*
4348                                  * If we were not able to push all items from our
4349                                  * leaf to its left neighbour, then attempt to
4350                                  * either push all the remaining items to the
4351                                  * right neighbour or none. There's no advantage
4352                                  * in pushing only some items, instead of all, as
4353                                  * it's pointless to end up with a leaf having
4354                                  * too few items while the neighbours can be full
4355                                  * or nearly full.
4356                                  */
4357                                 nritems = btrfs_header_nritems(leaf);
4358                                 min_push_space = leaf_space_used(leaf, 0, nritems);
4359                                 wret = push_leaf_right(trans, root, path, 0,
4360                                                        min_push_space, 1, 0);
4361                                 if (wret < 0 && wret != -ENOSPC)
4362                                         ret = wret;
4363                         }
4364
4365                         if (btrfs_header_nritems(leaf) == 0) {
4366                                 path->slots[1] = slot;
4367                                 btrfs_del_leaf(trans, root, path, leaf);
4368                                 free_extent_buffer(leaf);
4369                                 ret = 0;
4370                         } else {
4371                                 /* if we're still in the path, make sure
4372                                  * we're dirty.  Otherwise, one of the
4373                                  * push_leaf functions must have already
4374                                  * dirtied this buffer
4375                                  */
4376                                 if (path->nodes[0] == leaf)
4377                                         btrfs_mark_buffer_dirty(leaf);
4378                                 free_extent_buffer(leaf);
4379                         }
4380                 } else {
4381                         btrfs_mark_buffer_dirty(leaf);
4382                 }
4383         }
4384         return ret;
4385 }
4386
4387 /*
4388  * search the tree again to find a leaf with lesser keys
4389  * returns 0 if it found something or 1 if there are no lesser leaves.
4390  * returns < 0 on io errors.
4391  *
4392  * This may release the path, and so you may lose any locks held at the
4393  * time you call it.
4394  */
4395 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
4396 {
4397         struct btrfs_key key;
4398         struct btrfs_disk_key found_key;
4399         int ret;
4400
4401         btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
4402
4403         if (key.offset > 0) {
4404                 key.offset--;
4405         } else if (key.type > 0) {
4406                 key.type--;
4407                 key.offset = (u64)-1;
4408         } else if (key.objectid > 0) {
4409                 key.objectid--;
4410                 key.type = (u8)-1;
4411                 key.offset = (u64)-1;
4412         } else {
4413                 return 1;
4414         }
4415
4416         btrfs_release_path(path);
4417         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4418         if (ret < 0)
4419                 return ret;
4420         btrfs_item_key(path->nodes[0], &found_key, 0);
4421         ret = comp_keys(&found_key, &key);
4422         /*
4423          * We might have had an item with the previous key in the tree right
4424          * before we released our path. And after we released our path, that
4425          * item might have been pushed to the first slot (0) of the leaf we
4426          * were holding due to a tree balance. Alternatively, an item with the
4427          * previous key can exist as the only element of a leaf (big fat item).
4428          * Therefore account for these 2 cases, so that our callers (like
4429          * btrfs_previous_item) don't miss an existing item with a key matching
4430          * the previous key we computed above.
4431          */
4432         if (ret <= 0)
4433                 return 0;
4434         return 1;
4435 }
4436
4437 /*
4438  * A helper function to walk down the tree starting at min_key, and looking
4439  * for nodes or leaves that are have a minimum transaction id.
4440  * This is used by the btree defrag code, and tree logging
4441  *
4442  * This does not cow, but it does stuff the starting key it finds back
4443  * into min_key, so you can call btrfs_search_slot with cow=1 on the
4444  * key and get a writable path.
4445  *
4446  * This honors path->lowest_level to prevent descent past a given level
4447  * of the tree.
4448  *
4449  * min_trans indicates the oldest transaction that you are interested
4450  * in walking through.  Any nodes or leaves older than min_trans are
4451  * skipped over (without reading them).
4452  *
4453  * returns zero if something useful was found, < 0 on error and 1 if there
4454  * was nothing in the tree that matched the search criteria.
4455  */
4456 int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
4457                          struct btrfs_path *path,
4458                          u64 min_trans)
4459 {
4460         struct extent_buffer *cur;
4461         struct btrfs_key found_key;
4462         int slot;
4463         int sret;
4464         u32 nritems;
4465         int level;
4466         int ret = 1;
4467         int keep_locks = path->keep_locks;
4468
4469         ASSERT(!path->nowait);
4470         path->keep_locks = 1;
4471 again:
4472         cur = btrfs_read_lock_root_node(root);
4473         level = btrfs_header_level(cur);
4474         WARN_ON(path->nodes[level]);
4475         path->nodes[level] = cur;
4476         path->locks[level] = BTRFS_READ_LOCK;
4477
4478         if (btrfs_header_generation(cur) < min_trans) {
4479                 ret = 1;
4480                 goto out;
4481         }
4482         while (1) {
4483                 nritems = btrfs_header_nritems(cur);
4484                 level = btrfs_header_level(cur);
4485                 sret = btrfs_bin_search(cur, min_key, &slot);
4486                 if (sret < 0) {
4487                         ret = sret;
4488                         goto out;
4489                 }
4490
4491                 /* at the lowest level, we're done, setup the path and exit */
4492                 if (level == path->lowest_level) {
4493                         if (slot >= nritems)
4494                                 goto find_next_key;
4495                         ret = 0;
4496                         path->slots[level] = slot;
4497                         btrfs_item_key_to_cpu(cur, &found_key, slot);
4498                         goto out;
4499                 }
4500                 if (sret && slot > 0)
4501                         slot--;
4502                 /*
4503                  * check this node pointer against the min_trans parameters.
4504                  * If it is too old, skip to the next one.
4505                  */
4506                 while (slot < nritems) {
4507                         u64 gen;
4508
4509                         gen = btrfs_node_ptr_generation(cur, slot);
4510                         if (gen < min_trans) {
4511                                 slot++;
4512                                 continue;
4513                         }
4514                         break;
4515                 }
4516 find_next_key:
4517                 /*
4518                  * we didn't find a candidate key in this node, walk forward
4519                  * and find another one
4520                  */
4521                 if (slot >= nritems) {
4522                         path->slots[level] = slot;
4523                         sret = btrfs_find_next_key(root, path, min_key, level,
4524                                                   min_trans);
4525                         if (sret == 0) {
4526                                 btrfs_release_path(path);
4527                                 goto again;
4528                         } else {
4529                                 goto out;
4530                         }
4531                 }
4532                 /* save our key for returning back */
4533                 btrfs_node_key_to_cpu(cur, &found_key, slot);
4534                 path->slots[level] = slot;
4535                 if (level == path->lowest_level) {
4536                         ret = 0;
4537                         goto out;
4538                 }
4539                 cur = btrfs_read_node_slot(cur, slot);
4540                 if (IS_ERR(cur)) {
4541                         ret = PTR_ERR(cur);
4542                         goto out;
4543                 }
4544
4545                 btrfs_tree_read_lock(cur);
4546
4547                 path->locks[level - 1] = BTRFS_READ_LOCK;
4548                 path->nodes[level - 1] = cur;
4549                 unlock_up(path, level, 1, 0, NULL);
4550         }
4551 out:
4552         path->keep_locks = keep_locks;
4553         if (ret == 0) {
4554                 btrfs_unlock_up_safe(path, path->lowest_level + 1);
4555                 memcpy(min_key, &found_key, sizeof(found_key));
4556         }
4557         return ret;
4558 }
4559
4560 /*
4561  * this is similar to btrfs_next_leaf, but does not try to preserve
4562  * and fixup the path.  It looks for and returns the next key in the
4563  * tree based on the current path and the min_trans parameters.
4564  *
4565  * 0 is returned if another key is found, < 0 if there are any errors
4566  * and 1 is returned if there are no higher keys in the tree
4567  *
4568  * path->keep_locks should be set to 1 on the search made before
4569  * calling this function.
4570  */
4571 int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
4572                         struct btrfs_key *key, int level, u64 min_trans)
4573 {
4574         int slot;
4575         struct extent_buffer *c;
4576
4577         WARN_ON(!path->keep_locks && !path->skip_locking);
4578         while (level < BTRFS_MAX_LEVEL) {
4579                 if (!path->nodes[level])
4580                         return 1;
4581
4582                 slot = path->slots[level] + 1;
4583                 c = path->nodes[level];
4584 next:
4585                 if (slot >= btrfs_header_nritems(c)) {
4586                         int ret;
4587                         int orig_lowest;
4588                         struct btrfs_key cur_key;
4589                         if (level + 1 >= BTRFS_MAX_LEVEL ||
4590                             !path->nodes[level + 1])
4591                                 return 1;
4592
4593                         if (path->locks[level + 1] || path->skip_locking) {
4594                                 level++;
4595                                 continue;
4596                         }
4597
4598                         slot = btrfs_header_nritems(c) - 1;
4599                         if (level == 0)
4600                                 btrfs_item_key_to_cpu(c, &cur_key, slot);
4601                         else
4602                                 btrfs_node_key_to_cpu(c, &cur_key, slot);
4603
4604                         orig_lowest = path->lowest_level;
4605                         btrfs_release_path(path);
4606                         path->lowest_level = level;
4607                         ret = btrfs_search_slot(NULL, root, &cur_key, path,
4608                                                 0, 0);
4609                         path->lowest_level = orig_lowest;
4610                         if (ret < 0)
4611                                 return ret;
4612
4613                         c = path->nodes[level];
4614                         slot = path->slots[level];
4615                         if (ret == 0)
4616                                 slot++;
4617                         goto next;
4618                 }
4619
4620                 if (level == 0)
4621                         btrfs_item_key_to_cpu(c, key, slot);
4622                 else {
4623                         u64 gen = btrfs_node_ptr_generation(c, slot);
4624
4625                         if (gen < min_trans) {
4626                                 slot++;
4627                                 goto next;
4628                         }
4629                         btrfs_node_key_to_cpu(c, key, slot);
4630                 }
4631                 return 0;
4632         }
4633         return 1;
4634 }
4635
4636 int btrfs_next_old_leaf(struct btrfs_root *root, struct btrfs_path *path,
4637                         u64 time_seq)
4638 {
4639         int slot;
4640         int level;
4641         struct extent_buffer *c;
4642         struct extent_buffer *next;
4643         struct btrfs_fs_info *fs_info = root->fs_info;
4644         struct btrfs_key key;
4645         bool need_commit_sem = false;
4646         u32 nritems;
4647         int ret;
4648         int i;
4649
4650         ASSERT(!path->nowait);
4651
4652         nritems = btrfs_header_nritems(path->nodes[0]);
4653         if (nritems == 0)
4654                 return 1;
4655
4656         btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
4657 again:
4658         level = 1;
4659         next = NULL;
4660         btrfs_release_path(path);
4661
4662         path->keep_locks = 1;
4663
4664         if (time_seq) {
4665                 ret = btrfs_search_old_slot(root, &key, path, time_seq);
4666         } else {
4667                 if (path->need_commit_sem) {
4668                         path->need_commit_sem = 0;
4669                         need_commit_sem = true;
4670                         down_read(&fs_info->commit_root_sem);
4671                 }
4672                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4673         }
4674         path->keep_locks = 0;
4675
4676         if (ret < 0)
4677                 goto done;
4678
4679         nritems = btrfs_header_nritems(path->nodes[0]);
4680         /*
4681          * by releasing the path above we dropped all our locks.  A balance
4682          * could have added more items next to the key that used to be
4683          * at the very end of the block.  So, check again here and
4684          * advance the path if there are now more items available.
4685          */
4686         if (nritems > 0 && path->slots[0] < nritems - 1) {
4687                 if (ret == 0)
4688                         path->slots[0]++;
4689                 ret = 0;
4690                 goto done;
4691         }
4692         /*
4693          * So the above check misses one case:
4694          * - after releasing the path above, someone has removed the item that
4695          *   used to be at the very end of the block, and balance between leafs
4696          *   gets another one with bigger key.offset to replace it.
4697          *
4698          * This one should be returned as well, or we can get leaf corruption
4699          * later(esp. in __btrfs_drop_extents()).
4700          *
4701          * And a bit more explanation about this check,
4702          * with ret > 0, the key isn't found, the path points to the slot
4703          * where it should be inserted, so the path->slots[0] item must be the
4704          * bigger one.
4705          */
4706         if (nritems > 0 && ret > 0 && path->slots[0] == nritems - 1) {
4707                 ret = 0;
4708                 goto done;
4709         }
4710
4711         while (level < BTRFS_MAX_LEVEL) {
4712                 if (!path->nodes[level]) {
4713                         ret = 1;
4714                         goto done;
4715                 }
4716
4717                 slot = path->slots[level] + 1;
4718                 c = path->nodes[level];
4719                 if (slot >= btrfs_header_nritems(c)) {
4720                         level++;
4721                         if (level == BTRFS_MAX_LEVEL) {
4722                                 ret = 1;
4723                                 goto done;
4724                         }
4725                         continue;
4726                 }
4727
4728
4729                 /*
4730                  * Our current level is where we're going to start from, and to
4731                  * make sure lockdep doesn't complain we need to drop our locks
4732                  * and nodes from 0 to our current level.
4733                  */
4734                 for (i = 0; i < level; i++) {
4735                         if (path->locks[level]) {
4736                                 btrfs_tree_read_unlock(path->nodes[i]);
4737                                 path->locks[i] = 0;
4738                         }
4739                         free_extent_buffer(path->nodes[i]);
4740                         path->nodes[i] = NULL;
4741                 }
4742
4743                 next = c;
4744                 ret = read_block_for_search(root, path, &next, level,
4745                                             slot, &key);
4746                 if (ret == -EAGAIN)
4747                         goto again;
4748
4749                 if (ret < 0) {
4750                         btrfs_release_path(path);
4751                         goto done;
4752                 }
4753
4754                 if (!path->skip_locking) {
4755                         ret = btrfs_try_tree_read_lock(next);
4756                         if (!ret && time_seq) {
4757                                 /*
4758                                  * If we don't get the lock, we may be racing
4759                                  * with push_leaf_left, holding that lock while
4760                                  * itself waiting for the leaf we've currently
4761                                  * locked. To solve this situation, we give up
4762                                  * on our lock and cycle.
4763                                  */
4764                                 free_extent_buffer(next);
4765                                 btrfs_release_path(path);
4766                                 cond_resched();
4767                                 goto again;
4768                         }
4769                         if (!ret)
4770                                 btrfs_tree_read_lock(next);
4771                 }
4772                 break;
4773         }
4774         path->slots[level] = slot;
4775         while (1) {
4776                 level--;
4777                 path->nodes[level] = next;
4778                 path->slots[level] = 0;
4779                 if (!path->skip_locking)
4780                         path->locks[level] = BTRFS_READ_LOCK;
4781                 if (!level)
4782                         break;
4783
4784                 ret = read_block_for_search(root, path, &next, level,
4785                                             0, &key);
4786                 if (ret == -EAGAIN)
4787                         goto again;
4788
4789                 if (ret < 0) {
4790                         btrfs_release_path(path);
4791                         goto done;
4792                 }
4793
4794                 if (!path->skip_locking)
4795                         btrfs_tree_read_lock(next);
4796         }
4797         ret = 0;
4798 done:
4799         unlock_up(path, 0, 1, 0, NULL);
4800         if (need_commit_sem) {
4801                 int ret2;
4802
4803                 path->need_commit_sem = 1;
4804                 ret2 = finish_need_commit_sem_search(path);
4805                 up_read(&fs_info->commit_root_sem);
4806                 if (ret2)
4807                         ret = ret2;
4808         }
4809
4810         return ret;
4811 }
4812
4813 /*
4814  * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
4815  * searching until it gets past min_objectid or finds an item of 'type'
4816  *
4817  * returns 0 if something is found, 1 if nothing was found and < 0 on error
4818  */
4819 int btrfs_previous_item(struct btrfs_root *root,
4820                         struct btrfs_path *path, u64 min_objectid,
4821                         int type)
4822 {
4823         struct btrfs_key found_key;
4824         struct extent_buffer *leaf;
4825         u32 nritems;
4826         int ret;
4827
4828         while (1) {
4829                 if (path->slots[0] == 0) {
4830                         ret = btrfs_prev_leaf(root, path);
4831                         if (ret != 0)
4832                                 return ret;
4833                 } else {
4834                         path->slots[0]--;
4835                 }
4836                 leaf = path->nodes[0];
4837                 nritems = btrfs_header_nritems(leaf);
4838                 if (nritems == 0)
4839                         return 1;
4840                 if (path->slots[0] == nritems)
4841                         path->slots[0]--;
4842
4843                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4844                 if (found_key.objectid < min_objectid)
4845                         break;
4846                 if (found_key.type == type)
4847                         return 0;
4848                 if (found_key.objectid == min_objectid &&
4849                     found_key.type < type)
4850                         break;
4851         }
4852         return 1;
4853 }
4854
4855 /*
4856  * search in extent tree to find a previous Metadata/Data extent item with
4857  * min objecitd.
4858  *
4859  * returns 0 if something is found, 1 if nothing was found and < 0 on error
4860  */
4861 int btrfs_previous_extent_item(struct btrfs_root *root,
4862                         struct btrfs_path *path, u64 min_objectid)
4863 {
4864         struct btrfs_key found_key;
4865         struct extent_buffer *leaf;
4866         u32 nritems;
4867         int ret;
4868
4869         while (1) {
4870                 if (path->slots[0] == 0) {
4871                         ret = btrfs_prev_leaf(root, path);
4872                         if (ret != 0)
4873                                 return ret;
4874                 } else {
4875                         path->slots[0]--;
4876                 }
4877                 leaf = path->nodes[0];
4878                 nritems = btrfs_header_nritems(leaf);
4879                 if (nritems == 0)
4880                         return 1;
4881                 if (path->slots[0] == nritems)
4882                         path->slots[0]--;
4883
4884                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4885                 if (found_key.objectid < min_objectid)
4886                         break;
4887                 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
4888                     found_key.type == BTRFS_METADATA_ITEM_KEY)
4889                         return 0;
4890                 if (found_key.objectid == min_objectid &&
4891                     found_key.type < BTRFS_EXTENT_ITEM_KEY)
4892                         break;
4893         }
4894         return 1;
4895 }