fs/btrfs/ctree.c

   1 /*
   2  * Copyright (C) 2007,2008 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/sched.h>
  20 #include <linux/slab.h>
  21 #include <linux/rbtree.h>
  22 #include <linux/vmalloc.h>
  23 #include "ctree.h"
  24 #include "disk-io.h"
  25 #include "transaction.h"
  26 #include "print-tree.h"
  27 #include "locking.h"
  28
  29 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  30                       *root, struct btrfs_path *path, int level);
  31 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  32                       const struct btrfs_key *ins_key, struct btrfs_path *path,
  33                       int data_size, int extend);
  34 static int push_node_left(struct btrfs_trans_handle *trans,
  35                           struct btrfs_fs_info *fs_info,
  36                           struct extent_buffer *dst,
  37                           struct extent_buffer *src, int empty);
  38 static int balance_node_right(struct btrfs_trans_handle *trans,
  39                               struct btrfs_fs_info *fs_info,
  40                               struct extent_buffer *dst_buf,
  41                               struct extent_buffer *src_buf);
  42 static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
  43                     int level, int slot);
  44 static int tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
  45                                  struct extent_buffer *eb);
  46
  47 struct btrfs_path *btrfs_alloc_path(void)
  48 {
  49         return kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
  50 }
  51
  52 /*
  53  * set all locked nodes in the path to blocking locks.  This should
  54  * be done before scheduling
  55  */
  56 noinline void btrfs_set_path_blocking(struct btrfs_path *p)
  57 {
  58         int i;
  59         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  60                 if (!p->nodes[i] || !p->locks[i])
  61                         continue;
  62                 btrfs_set_lock_blocking_rw(p->nodes[i], p->locks[i]);
  63                 if (p->locks[i] == BTRFS_READ_LOCK)
  64                         p->locks[i] = BTRFS_READ_LOCK_BLOCKING;
  65                 else if (p->locks[i] == BTRFS_WRITE_LOCK)
  66                         p->locks[i] = BTRFS_WRITE_LOCK_BLOCKING;
  67         }
  68 }
  69
  70 /*
  71  * reset all the locked nodes in the patch to spinning locks.
  72  *
  73  * held is used to keep lockdep happy, when lockdep is enabled
  74  * we set held to a blocking lock before we go around and
  75  * retake all the spinlocks in the path.  You can safely use NULL
  76  * for held
  77  */
  78 noinline void btrfs_clear_path_blocking(struct btrfs_path *p,
  79                                         struct extent_buffer *held, int held_rw)
  80 {
  81         int i;
  82
  83         if (held) {
  84                 btrfs_set_lock_blocking_rw(held, held_rw);
  85                 if (held_rw == BTRFS_WRITE_LOCK)
  86                         held_rw = BTRFS_WRITE_LOCK_BLOCKING;
  87                 else if (held_rw == BTRFS_READ_LOCK)
  88                         held_rw = BTRFS_READ_LOCK_BLOCKING;
  89         }
  90         btrfs_set_path_blocking(p);
  91
  92         for (i = BTRFS_MAX_LEVEL - 1; i >= 0; i--) {
  93                 if (p->nodes[i] && p->locks[i]) {
  94                         btrfs_clear_lock_blocking_rw(p->nodes[i], p->locks[i]);
  95                         if (p->locks[i] == BTRFS_WRITE_LOCK_BLOCKING)
  96                                 p->locks[i] = BTRFS_WRITE_LOCK;
  97                         else if (p->locks[i] == BTRFS_READ_LOCK_BLOCKING)
  98                                 p->locks[i] = BTRFS_READ_LOCK;
  99                 }
 100         }
 101
 102         if (held)
 103                 btrfs_clear_lock_blocking_rw(held, held_rw);
 104 }
 105
 106 /* this also releases the path */
 107 void btrfs_free_path(struct btrfs_path *p)
 108 {
 109         if (!p)
 110                 return;
 111         btrfs_release_path(p);
 112         kmem_cache_free(btrfs_path_cachep, p);
 113 }
 114
 115 /*
 116  * path release drops references on the extent buffers in the path
 117  * and it drops any locks held by this path
 118  *
 119  * It is safe to call this on paths that no locks or extent buffers held.
 120  */
 121 noinline void btrfs_release_path(struct btrfs_path *p)
 122 {
 123         int i;
 124
 125         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
 126                 p->slots[i] = 0;
 127                 if (!p->nodes[i])
 128                         continue;
 129                 if (p->locks[i]) {
 130                         btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
 131                         p->locks[i] = 0;
 132                 }
 133                 free_extent_buffer(p->nodes[i]);
 134                 p->nodes[i] = NULL;
 135         }
 136 }
 137
 138 /*
 139  * safely gets a reference on the root node of a tree.  A lock
 140  * is not taken, so a concurrent writer may put a different node
 141  * at the root of the tree.  See btrfs_lock_root_node for the
 142  * looping required.
 143  *
 144  * The extent buffer returned by this has a reference taken, so
 145  * it won't disappear.  It may stop being the root of the tree
 146  * at any time because there are no locks held.
 147  */
 148 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
 149 {
 150         struct extent_buffer *eb;
 151
 152         while (1) {
 153                 rcu_read_lock();
 154                 eb = rcu_dereference(root->node);
 155
 156                 /*
 157                  * RCU really hurts here, we could free up the root node because
 158                  * it was COWed but we may not get the new root node yet so do
 159                  * the inc_not_zero dance and if it doesn't work then
 160                  * synchronize_rcu and try again.
 161                  */
 162                 if (atomic_inc_not_zero(&eb->refs)) {
 163                         rcu_read_unlock();
 164                         break;
 165                 }
 166                 rcu_read_unlock();
 167                 synchronize_rcu();
 168         }
 169         return eb;
 170 }
 171
 172 /* loop around taking references on and locking the root node of the
 173  * tree until you end up with a lock on the root.  A locked buffer
 174  * is returned, with a reference held.
 175  */
 176 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
 177 {
 178         struct extent_buffer *eb;
 179
 180         while (1) {
 181                 eb = btrfs_root_node(root);
 182                 btrfs_tree_lock(eb);
 183                 if (eb == root->node)
 184                         break;
 185                 btrfs_tree_unlock(eb);
 186                 free_extent_buffer(eb);
 187         }
 188         return eb;
 189 }
 190
 191 /* loop around taking references on and locking the root node of the
 192  * tree until you end up with a lock on the root.  A locked buffer
 193  * is returned, with a reference held.
 194  */
 195 static struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
 196 {
 197         struct extent_buffer *eb;
 198
 199         while (1) {
 200                 eb = btrfs_root_node(root);
 201                 btrfs_tree_read_lock(eb);
 202                 if (eb == root->node)
 203                         break;
 204                 btrfs_tree_read_unlock(eb);
 205                 free_extent_buffer(eb);
 206         }
 207         return eb;
 208 }
 209
 210 /* cowonly root (everything not a reference counted cow subvolume), just get
 211  * put onto a simple dirty list.  transaction.c walks this to make sure they
 212  * get properly updated on disk.
 213  */
 214 static void add_root_to_dirty_list(struct btrfs_root *root)
 215 {
 216         struct btrfs_fs_info *fs_info = root->fs_info;
 217
 218         if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
 219             !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
 220                 return;
 221
 222         spin_lock(&fs_info->trans_lock);
 223         if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
 224                 /* Want the extent tree to be the last on the list */
 225                 if (root->objectid == BTRFS_EXTENT_TREE_OBJECTID)
 226                         list_move_tail(&root->dirty_list,
 227                                        &fs_info->dirty_cowonly_roots);
 228                 else
 229                         list_move(&root->dirty_list,
 230                                   &fs_info->dirty_cowonly_roots);
 231         }
 232         spin_unlock(&fs_info->trans_lock);
 233 }
 234
 235 /*
 236  * used by snapshot creation to make a copy of a root for a tree with
 237  * a given objectid.  The buffer with the new root node is returned in
 238  * cow_ret, and this func returns zero on success or a negative error code.
 239  */
 240 int btrfs_copy_root(struct btrfs_trans_handle *trans,
 241                       struct btrfs_root *root,
 242                       struct extent_buffer *buf,
 243                       struct extent_buffer **cow_ret, u64 new_root_objectid)
 244 {
 245         struct btrfs_fs_info *fs_info = root->fs_info;
 246         struct extent_buffer *cow;
 247         int ret = 0;
 248         int level;
 249         struct btrfs_disk_key disk_key;
 250
 251         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 252                 trans->transid != fs_info->running_transaction->transid);
 253         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 254                 trans->transid != root->last_trans);
 255
 256         level = btrfs_header_level(buf);
 257         if (level == 0)
 258                 btrfs_item_key(buf, &disk_key, 0);
 259         else
 260                 btrfs_node_key(buf, &disk_key, 0);
 261
 262         cow = btrfs_alloc_tree_block(trans, root, 0, new_root_objectid,
 263                         &disk_key, level, buf->start, 0);
 264         if (IS_ERR(cow))
 265                 return PTR_ERR(cow);
 266
 267         copy_extent_buffer_full(cow, buf);
 268         btrfs_set_header_bytenr(cow, cow->start);
 269         btrfs_set_header_generation(cow, trans->transid);
 270         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 271         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 272                                      BTRFS_HEADER_FLAG_RELOC);
 273         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 274                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 275         else
 276                 btrfs_set_header_owner(cow, new_root_objectid);
 277
 278         write_extent_buffer_fsid(cow, fs_info->fsid);
 279
 280         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 281         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 282                 ret = btrfs_inc_ref(trans, root, cow, 1);
 283         else
 284                 ret = btrfs_inc_ref(trans, root, cow, 0);
 285
 286         if (ret)
 287                 return ret;
 288
 289         btrfs_mark_buffer_dirty(cow);
 290         *cow_ret = cow;
 291         return 0;
 292 }
 293
 294 enum mod_log_op {
 295         MOD_LOG_KEY_REPLACE,
 296         MOD_LOG_KEY_ADD,
 297         MOD_LOG_KEY_REMOVE,
 298         MOD_LOG_KEY_REMOVE_WHILE_FREEING,
 299         MOD_LOG_KEY_REMOVE_WHILE_MOVING,
 300         MOD_LOG_MOVE_KEYS,
 301         MOD_LOG_ROOT_REPLACE,
 302 };
 303
 304 struct tree_mod_move {
 305         int dst_slot;
 306         int nr_items;
 307 };
 308
 309 struct tree_mod_root {
 310         u64 logical;
 311         u8 level;
 312 };
 313
 314 struct tree_mod_elem {
 315         struct rb_node node;
 316         u64 logical;
 317         u64 seq;
 318         enum mod_log_op op;
 319
 320         /* this is used for MOD_LOG_KEY_* and MOD_LOG_MOVE_KEYS operations */
 321         int slot;
 322
 323         /* this is used for MOD_LOG_KEY* and MOD_LOG_ROOT_REPLACE */
 324         u64 generation;
 325
 326         /* those are used for op == MOD_LOG_KEY_{REPLACE,REMOVE} */
 327         struct btrfs_disk_key key;
 328         u64 blockptr;
 329
 330         /* this is used for op == MOD_LOG_MOVE_KEYS */
 331         struct tree_mod_move move;
 332
 333         /* this is used for op == MOD_LOG_ROOT_REPLACE */
 334         struct tree_mod_root old_root;
 335 };
 336
 337 static inline void tree_mod_log_read_lock(struct btrfs_fs_info *fs_info)
 338 {
 339         read_lock(&fs_info->tree_mod_log_lock);
 340 }
 341
 342 static inline void tree_mod_log_read_unlock(struct btrfs_fs_info *fs_info)
 343 {
 344         read_unlock(&fs_info->tree_mod_log_lock);
 345 }
 346
 347 static inline void tree_mod_log_write_lock(struct btrfs_fs_info *fs_info)
 348 {
 349         write_lock(&fs_info->tree_mod_log_lock);
 350 }
 351
 352 static inline void tree_mod_log_write_unlock(struct btrfs_fs_info *fs_info)
 353 {
 354         write_unlock(&fs_info->tree_mod_log_lock);
 355 }
 356
 357 /*
 358  * Pull a new tree mod seq number for our operation.
 359  */
 360 static inline u64 btrfs_inc_tree_mod_seq(struct btrfs_fs_info *fs_info)
 361 {
 362         return atomic64_inc_return(&fs_info->tree_mod_seq);
 363 }
 364
 365 /*
 366  * This adds a new blocker to the tree mod log's blocker list if the @elem
 367  * passed does not already have a sequence number set. So when a caller expects
 368  * to record tree modifications, it should ensure to set elem->seq to zero
 369  * before calling btrfs_get_tree_mod_seq.
 370  * Returns a fresh, unused tree log modification sequence number, even if no new
 371  * blocker was added.
 372  */
 373 u64 btrfs_get_tree_mod_seq(struct btrfs_fs_info *fs_info,
 374                            struct seq_list *elem)
 375 {
 376         tree_mod_log_write_lock(fs_info);
 377         spin_lock(&fs_info->tree_mod_seq_lock);
 378         if (!elem->seq) {
 379                 elem->seq = btrfs_inc_tree_mod_seq(fs_info);
 380                 list_add_tail(&elem->list, &fs_info->tree_mod_seq_list);
 381         }
 382         spin_unlock(&fs_info->tree_mod_seq_lock);
 383         tree_mod_log_write_unlock(fs_info);
 384
 385         return elem->seq;
 386 }
 387
 388 void btrfs_put_tree_mod_seq(struct btrfs_fs_info *fs_info,
 389                             struct seq_list *elem)
 390 {
 391         struct rb_root *tm_root;
 392         struct rb_node *node;
 393         struct rb_node *next;
 394         struct seq_list *cur_elem;
 395         struct tree_mod_elem *tm;
 396         u64 min_seq = (u64)-1;
 397         u64 seq_putting = elem->seq;
 398
 399         if (!seq_putting)
 400                 return;
 401
 402         spin_lock(&fs_info->tree_mod_seq_lock);
 403         list_del(&elem->list);
 404         elem->seq = 0;
 405
 406         list_for_each_entry(cur_elem, &fs_info->tree_mod_seq_list, list) {
 407                 if (cur_elem->seq < min_seq) {
 408                         if (seq_putting > cur_elem->seq) {
 409                                 /*
 410                                  * blocker with lower sequence number exists, we
 411                                  * cannot remove anything from the log
 412                                  */
 413                                 spin_unlock(&fs_info->tree_mod_seq_lock);
 414                                 return;
 415                         }
 416                         min_seq = cur_elem->seq;
 417                 }
 418         }
 419         spin_unlock(&fs_info->tree_mod_seq_lock);
 420
 421         /*
 422          * anything that's lower than the lowest existing (read: blocked)
 423          * sequence number can be removed from the tree.
 424          */
 425         tree_mod_log_write_lock(fs_info);
 426         tm_root = &fs_info->tree_mod_log;
 427         for (node = rb_first(tm_root); node; node = next) {
 428                 next = rb_next(node);
 429                 tm = rb_entry(node, struct tree_mod_elem, node);
 430                 if (tm->seq > min_seq)
 431                         continue;
 432                 rb_erase(node, tm_root);
 433                 kfree(tm);
 434         }
 435         tree_mod_log_write_unlock(fs_info);
 436 }
 437
 438 /*
 439  * key order of the log:
 440  *       node/leaf start address -> sequence
 441  *
 442  * The 'start address' is the logical address of the *new* root node
 443  * for root replace operations, or the logical address of the affected
 444  * block for all other operations.
 445  *
 446  * Note: must be called with write lock (tree_mod_log_write_lock).
 447  */
 448 static noinline int
 449 __tree_mod_log_insert(struct btrfs_fs_info *fs_info, struct tree_mod_elem *tm)
 450 {
 451         struct rb_root *tm_root;
 452         struct rb_node **new;
 453         struct rb_node *parent = NULL;
 454         struct tree_mod_elem *cur;
 455
 456         BUG_ON(!tm);
 457
 458         tm->seq = btrfs_inc_tree_mod_seq(fs_info);
 459
 460         tm_root = &fs_info->tree_mod_log;
 461         new = &tm_root->rb_node;
 462         while (*new) {
 463                 cur = rb_entry(*new, struct tree_mod_elem, node);
 464                 parent = *new;
 465                 if (cur->logical < tm->logical)
 466                         new = &((*new)->rb_left);
 467                 else if (cur->logical > tm->logical)
 468                         new = &((*new)->rb_right);
 469                 else if (cur->seq < tm->seq)
 470                         new = &((*new)->rb_left);
 471                 else if (cur->seq > tm->seq)
 472                         new = &((*new)->rb_right);
 473                 else
 474                         return -EEXIST;
 475         }
 476
 477         rb_link_node(&tm->node, parent, new);
 478         rb_insert_color(&tm->node, tm_root);
 479         return 0;
 480 }
 481
 482 /*
 483  * Determines if logging can be omitted. Returns 1 if it can. Otherwise, it
 484  * returns zero with the tree_mod_log_lock acquired. The caller must hold
 485  * this until all tree mod log insertions are recorded in the rb tree and then
 486  * call tree_mod_log_write_unlock() to release.
 487  */
 488 static inline int tree_mod_dont_log(struct btrfs_fs_info *fs_info,
 489                                     struct extent_buffer *eb) {
 490         smp_mb();
 491         if (list_empty(&(fs_info)->tree_mod_seq_list))
 492                 return 1;
 493         if (eb && btrfs_header_level(eb) == 0)
 494                 return 1;
 495
 496         tree_mod_log_write_lock(fs_info);
 497         if (list_empty(&(fs_info)->tree_mod_seq_list)) {
 498                 tree_mod_log_write_unlock(fs_info);
 499                 return 1;
 500         }
 501
 502         return 0;
 503 }
 504
 505 /* Similar to tree_mod_dont_log, but doesn't acquire any locks. */
 506 static inline int tree_mod_need_log(const struct btrfs_fs_info *fs_info,
 507                                     struct extent_buffer *eb)
 508 {
 509         smp_mb();
 510         if (list_empty(&(fs_info)->tree_mod_seq_list))
 511                 return 0;
 512         if (eb && btrfs_header_level(eb) == 0)
 513                 return 0;
 514
 515         return 1;
 516 }
 517
 518 static struct tree_mod_elem *
 519 alloc_tree_mod_elem(struct extent_buffer *eb, int slot,
 520                     enum mod_log_op op, gfp_t flags)
 521 {
 522         struct tree_mod_elem *tm;
 523
 524         tm = kzalloc(sizeof(*tm), flags);
 525         if (!tm)
 526                 return NULL;
 527
 528         tm->logical = eb->start;
 529         if (op != MOD_LOG_KEY_ADD) {
 530                 btrfs_node_key(eb, &tm->key, slot);
 531                 tm->blockptr = btrfs_node_blockptr(eb, slot);
 532         }
 533         tm->op = op;
 534         tm->slot = slot;
 535         tm->generation = btrfs_node_ptr_generation(eb, slot);
 536         RB_CLEAR_NODE(&tm->node);
 537
 538         return tm;
 539 }
 540
 541 static noinline int
 542 tree_mod_log_insert_key(struct btrfs_fs_info *fs_info,
 543                         struct extent_buffer *eb, int slot,
 544                         enum mod_log_op op, gfp_t flags)
 545 {
 546         struct tree_mod_elem *tm;
 547         int ret;
 548
 549         if (!tree_mod_need_log(fs_info, eb))
 550                 return 0;
 551
 552         tm = alloc_tree_mod_elem(eb, slot, op, flags);
 553         if (!tm)
 554                 return -ENOMEM;
 555
 556         if (tree_mod_dont_log(fs_info, eb)) {
 557                 kfree(tm);
 558                 return 0;
 559         }
 560
 561         ret = __tree_mod_log_insert(fs_info, tm);
 562         tree_mod_log_write_unlock(fs_info);
 563         if (ret)
 564                 kfree(tm);
 565
 566         return ret;
 567 }
 568
 569 static noinline int
 570 tree_mod_log_insert_move(struct btrfs_fs_info *fs_info,
 571                          struct extent_buffer *eb, int dst_slot, int src_slot,
 572                          int nr_items, gfp_t flags)
 573 {
 574         struct tree_mod_elem *tm = NULL;
 575         struct tree_mod_elem **tm_list = NULL;
 576         int ret = 0;
 577         int i;
 578         int locked = 0;
 579
 580         if (!tree_mod_need_log(fs_info, eb))
 581                 return 0;
 582
 583         tm_list = kcalloc(nr_items, sizeof(struct tree_mod_elem *), flags);
 584         if (!tm_list)
 585                 return -ENOMEM;
 586
 587         tm = kzalloc(sizeof(*tm), flags);
 588         if (!tm) {
 589                 ret = -ENOMEM;
 590                 goto free_tms;
 591         }
 592
 593         tm->logical = eb->start;
 594         tm->slot = src_slot;
 595         tm->move.dst_slot = dst_slot;
 596         tm->move.nr_items = nr_items;
 597         tm->op = MOD_LOG_MOVE_KEYS;
 598
 599         for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
 600                 tm_list[i] = alloc_tree_mod_elem(eb, i + dst_slot,
 601                     MOD_LOG_KEY_REMOVE_WHILE_MOVING, flags);
 602                 if (!tm_list[i]) {
 603                         ret = -ENOMEM;
 604                         goto free_tms;
 605                 }
 606         }
 607
 608         if (tree_mod_dont_log(fs_info, eb))
 609                 goto free_tms;
 610         locked = 1;
 611
 612         /*
 613          * When we override something during the move, we log these removals.
 614          * This can only happen when we move towards the beginning of the
 615          * buffer, i.e. dst_slot < src_slot.
 616          */
 617         for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
 618                 ret = __tree_mod_log_insert(fs_info, tm_list[i]);
 619                 if (ret)
 620                         goto free_tms;
 621         }
 622
 623         ret = __tree_mod_log_insert(fs_info, tm);
 624         if (ret)
 625                 goto free_tms;
 626         tree_mod_log_write_unlock(fs_info);
 627         kfree(tm_list);
 628
 629         return 0;
 630 free_tms:
 631         for (i = 0; i < nr_items; i++) {
 632                 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
 633                         rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
 634                 kfree(tm_list[i]);
 635         }
 636         if (locked)
 637                 tree_mod_log_write_unlock(fs_info);
 638         kfree(tm_list);
 639         kfree(tm);
 640
 641         return ret;
 642 }
 643
 644 static inline int
 645 __tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
 646                        struct tree_mod_elem **tm_list,
 647                        int nritems)
 648 {
 649         int i, j;
 650         int ret;
 651
 652         for (i = nritems - 1; i >= 0; i--) {
 653                 ret = __tree_mod_log_insert(fs_info, tm_list[i]);
 654                 if (ret) {
 655                         for (j = nritems - 1; j > i; j--)
 656                                 rb_erase(&tm_list[j]->node,
 657                                          &fs_info->tree_mod_log);
 658                         return ret;
 659                 }
 660         }
 661
 662         return 0;
 663 }
 664
 665 static noinline int
 666 tree_mod_log_insert_root(struct btrfs_fs_info *fs_info,
 667                          struct extent_buffer *old_root,
 668                          struct extent_buffer *new_root, gfp_t flags,
 669                          int log_removal)
 670 {
 671         struct tree_mod_elem *tm = NULL;
 672         struct tree_mod_elem **tm_list = NULL;
 673         int nritems = 0;
 674         int ret = 0;
 675         int i;
 676
 677         if (!tree_mod_need_log(fs_info, NULL))
 678                 return 0;
 679
 680         if (log_removal && btrfs_header_level(old_root) > 0) {
 681                 nritems = btrfs_header_nritems(old_root);
 682                 tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *),
 683                                   flags);
 684                 if (!tm_list) {
 685                         ret = -ENOMEM;
 686                         goto free_tms;
 687                 }
 688                 for (i = 0; i < nritems; i++) {
 689                         tm_list[i] = alloc_tree_mod_elem(old_root, i,
 690                             MOD_LOG_KEY_REMOVE_WHILE_FREEING, flags);
 691                         if (!tm_list[i]) {
 692                                 ret = -ENOMEM;
 693                                 goto free_tms;
 694                         }
 695                 }
 696         }
 697
 698         tm = kzalloc(sizeof(*tm), flags);
 699         if (!tm) {
 700                 ret = -ENOMEM;
 701                 goto free_tms;
 702         }
 703
 704         tm->logical = new_root->start;
 705         tm->old_root.logical = old_root->start;
 706         tm->old_root.level = btrfs_header_level(old_root);
 707         tm->generation = btrfs_header_generation(old_root);
 708         tm->op = MOD_LOG_ROOT_REPLACE;
 709
 710         if (tree_mod_dont_log(fs_info, NULL))
 711                 goto free_tms;
 712
 713         if (tm_list)
 714                 ret = __tree_mod_log_free_eb(fs_info, tm_list, nritems);
 715         if (!ret)
 716                 ret = __tree_mod_log_insert(fs_info, tm);
 717
 718         tree_mod_log_write_unlock(fs_info);
 719         if (ret)
 720                 goto free_tms;
 721         kfree(tm_list);
 722
 723         return ret;
 724
 725 free_tms:
 726         if (tm_list) {
 727                 for (i = 0; i < nritems; i++)
 728                         kfree(tm_list[i]);
 729                 kfree(tm_list);
 730         }
 731         kfree(tm);
 732
 733         return ret;
 734 }
 735
 736 static struct tree_mod_elem *
 737 __tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq,
 738                       int smallest)
 739 {
 740         struct rb_root *tm_root;
 741         struct rb_node *node;
 742         struct tree_mod_elem *cur = NULL;
 743         struct tree_mod_elem *found = NULL;
 744
 745         tree_mod_log_read_lock(fs_info);
 746         tm_root = &fs_info->tree_mod_log;
 747         node = tm_root->rb_node;
 748         while (node) {
 749                 cur = rb_entry(node, struct tree_mod_elem, node);
 750                 if (cur->logical < start) {
 751                         node = node->rb_left;
 752                 } else if (cur->logical > start) {
 753                         node = node->rb_right;
 754                 } else if (cur->seq < min_seq) {
 755                         node = node->rb_left;
 756                 } else if (!smallest) {
 757                         /* we want the node with the highest seq */
 758                         if (found)
 759                                 BUG_ON(found->seq > cur->seq);
 760                         found = cur;
 761                         node = node->rb_left;
 762                 } else if (cur->seq > min_seq) {
 763                         /* we want the node with the smallest seq */
 764                         if (found)
 765                                 BUG_ON(found->seq < cur->seq);
 766                         found = cur;
 767                         node = node->rb_right;
 768                 } else {
 769                         found = cur;
 770                         break;
 771                 }
 772         }
 773         tree_mod_log_read_unlock(fs_info);
 774
 775         return found;
 776 }
 777
 778 /*
 779  * this returns the element from the log with the smallest time sequence
 780  * value that's in the log (the oldest log item). any element with a time
 781  * sequence lower than min_seq will be ignored.
 782  */
 783 static struct tree_mod_elem *
 784 tree_mod_log_search_oldest(struct btrfs_fs_info *fs_info, u64 start,
 785                            u64 min_seq)
 786 {
 787         return __tree_mod_log_search(fs_info, start, min_seq, 1);
 788 }
 789
 790 /*
 791  * this returns the element from the log with the largest time sequence
 792  * value that's in the log (the most recent log item). any element with
 793  * a time sequence lower than min_seq will be ignored.
 794  */
 795 static struct tree_mod_elem *
 796 tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq)
 797 {
 798         return __tree_mod_log_search(fs_info, start, min_seq, 0);
 799 }
 800
 801 static noinline int
 802 tree_mod_log_eb_copy(struct btrfs_fs_info *fs_info, struct extent_buffer *dst,
 803                      struct extent_buffer *src, unsigned long dst_offset,
 804                      unsigned long src_offset, int nr_items)
 805 {
 806         int ret = 0;
 807         struct tree_mod_elem **tm_list = NULL;
 808         struct tree_mod_elem **tm_list_add, **tm_list_rem;
 809         int i;
 810         int locked = 0;
 811
 812         if (!tree_mod_need_log(fs_info, NULL))
 813                 return 0;
 814
 815         if (btrfs_header_level(dst) == 0 && btrfs_header_level(src) == 0)
 816                 return 0;
 817
 818         tm_list = kcalloc(nr_items * 2, sizeof(struct tree_mod_elem *),
 819                           GFP_NOFS);
 820         if (!tm_list)
 821                 return -ENOMEM;
 822
 823         tm_list_add = tm_list;
 824         tm_list_rem = tm_list + nr_items;
 825         for (i = 0; i < nr_items; i++) {
 826                 tm_list_rem[i] = alloc_tree_mod_elem(src, i + src_offset,
 827                     MOD_LOG_KEY_REMOVE, GFP_NOFS);
 828                 if (!tm_list_rem[i]) {
 829                         ret = -ENOMEM;
 830                         goto free_tms;
 831                 }
 832
 833                 tm_list_add[i] = alloc_tree_mod_elem(dst, i + dst_offset,
 834                     MOD_LOG_KEY_ADD, GFP_NOFS);
 835                 if (!tm_list_add[i]) {
 836                         ret = -ENOMEM;
 837                         goto free_tms;
 838                 }
 839         }
 840
 841         if (tree_mod_dont_log(fs_info, NULL))
 842                 goto free_tms;
 843         locked = 1;
 844
 845         for (i = 0; i < nr_items; i++) {
 846                 ret = __tree_mod_log_insert(fs_info, tm_list_rem[i]);
 847                 if (ret)
 848                         goto free_tms;
 849                 ret = __tree_mod_log_insert(fs_info, tm_list_add[i]);
 850                 if (ret)
 851                         goto free_tms;
 852         }
 853
 854         tree_mod_log_write_unlock(fs_info);
 855         kfree(tm_list);
 856
 857         return 0;
 858
 859 free_tms:
 860         for (i = 0; i < nr_items * 2; i++) {
 861                 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
 862                         rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
 863                 kfree(tm_list[i]);
 864         }
 865         if (locked)
 866                 tree_mod_log_write_unlock(fs_info);
 867         kfree(tm_list);
 868
 869         return ret;
 870 }
 871
 872 static inline void
 873 tree_mod_log_eb_move(struct btrfs_fs_info *fs_info, struct extent_buffer *dst,
 874                      int dst_offset, int src_offset, int nr_items)
 875 {
 876         int ret;
 877         ret = tree_mod_log_insert_move(fs_info, dst, dst_offset, src_offset,
 878                                        nr_items, GFP_NOFS);
 879         BUG_ON(ret < 0);
 880 }
 881
 882 static noinline void
 883 tree_mod_log_set_node_key(struct btrfs_fs_info *fs_info,
 884                           struct extent_buffer *eb, int slot, int atomic)
 885 {
 886         int ret;
 887
 888         ret = tree_mod_log_insert_key(fs_info, eb, slot,
 889                                         MOD_LOG_KEY_REPLACE,
 890                                         atomic ? GFP_ATOMIC : GFP_NOFS);
 891         BUG_ON(ret < 0);
 892 }
 893
 894 static noinline int
 895 tree_mod_log_free_eb(struct btrfs_fs_info *fs_info, struct extent_buffer *eb)
 896 {
 897         struct tree_mod_elem **tm_list = NULL;
 898         int nritems = 0;
 899         int i;
 900         int ret = 0;
 901
 902         if (btrfs_header_level(eb) == 0)
 903                 return 0;
 904
 905         if (!tree_mod_need_log(fs_info, NULL))
 906                 return 0;
 907
 908         nritems = btrfs_header_nritems(eb);
 909         tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *), GFP_NOFS);
 910         if (!tm_list)
 911                 return -ENOMEM;
 912
 913         for (i = 0; i < nritems; i++) {
 914                 tm_list[i] = alloc_tree_mod_elem(eb, i,
 915                     MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
 916                 if (!tm_list[i]) {
 917                         ret = -ENOMEM;
 918                         goto free_tms;
 919                 }
 920         }
 921
 922         if (tree_mod_dont_log(fs_info, eb))
 923                 goto free_tms;
 924
 925         ret = __tree_mod_log_free_eb(fs_info, tm_list, nritems);
 926         tree_mod_log_write_unlock(fs_info);
 927         if (ret)
 928                 goto free_tms;
 929         kfree(tm_list);
 930
 931         return 0;
 932
 933 free_tms:
 934         for (i = 0; i < nritems; i++)
 935                 kfree(tm_list[i]);
 936         kfree(tm_list);
 937
 938         return ret;
 939 }
 940
 941 static noinline void
 942 tree_mod_log_set_root_pointer(struct btrfs_root *root,
 943                               struct extent_buffer *new_root_node,
 944                               int log_removal)
 945 {
 946         int ret;
 947         ret = tree_mod_log_insert_root(root->fs_info, root->node,
 948                                        new_root_node, GFP_NOFS, log_removal);
 949         BUG_ON(ret < 0);
 950 }
 951
 952 /*
 953  * check if the tree block can be shared by multiple trees
 954  */
 955 int btrfs_block_can_be_shared(struct btrfs_root *root,
 956                               struct extent_buffer *buf)
 957 {
 958         /*
 959          * Tree blocks not in reference counted trees and tree roots
 960          * are never shared. If a block was allocated after the last
 961          * snapshot and the block was not allocated by tree relocation,
 962          * we know the block is not shared.
 963          */
 964         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 965             buf != root->node && buf != root->commit_root &&
 966             (btrfs_header_generation(buf) <=
 967              btrfs_root_last_snapshot(&root->root_item) ||
 968              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 969                 return 1;
 970 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 971         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 972             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 973                 return 1;
 974 #endif
 975         return 0;
 976 }
 977
 978 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
 979                                        struct btrfs_root *root,
 980                                        struct extent_buffer *buf,
 981                                        struct extent_buffer *cow,
 982                                        int *last_ref)
 983 {
 984         struct btrfs_fs_info *fs_info = root->fs_info;
 985         u64 refs;
 986         u64 owner;
 987         u64 flags;
 988         u64 new_flags = 0;
 989         int ret;
 990
 991         /*
 992          * Backrefs update rules:
 993          *
 994          * Always use full backrefs for extent pointers in tree block
 995          * allocated by tree relocation.
 996          *
 997          * If a shared tree block is no longer referenced by its owner
 998          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
 999          * use full backrefs for extent pointers in tree block.
1000          *
1001          * If a tree block is been relocating
1002          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
1003          * use full backrefs for extent pointers in tree block.
1004          * The reason for this is some operations (such as drop tree)
1005          * are only allowed for blocks use full backrefs.
1006          */
1007
1008         if (btrfs_block_can_be_shared(root, buf)) {
1009                 ret = btrfs_lookup_extent_info(trans, fs_info, buf->start,
1010                                                btrfs_header_level(buf), 1,
1011                                                &refs, &flags);
1012                 if (ret)
1013                         return ret;
1014                 if (refs == 0) {
1015                         ret = -EROFS;
1016                         btrfs_handle_fs_error(fs_info, ret, NULL);
1017                         return ret;
1018                 }
1019         } else {
1020                 refs = 1;
1021                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1022                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1023                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
1024                 else
1025                         flags = 0;
1026         }
1027
1028         owner = btrfs_header_owner(buf);
1029         BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
1030                !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
1031
1032         if (refs > 1) {
1033                 if ((owner == root->root_key.objectid ||
1034                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
1035                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
1036                         ret = btrfs_inc_ref(trans, root, buf, 1);
1037                         BUG_ON(ret); /* -ENOMEM */
1038
1039                         if (root->root_key.objectid ==
1040                             BTRFS_TREE_RELOC_OBJECTID) {
1041                                 ret = btrfs_dec_ref(trans, root, buf, 0);
1042                                 BUG_ON(ret); /* -ENOMEM */
1043                                 ret = btrfs_inc_ref(trans, root, cow, 1);
1044                                 BUG_ON(ret); /* -ENOMEM */
1045                         }
1046                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
1047                 } else {
1048
1049                         if (root->root_key.objectid ==
1050                             BTRFS_TREE_RELOC_OBJECTID)
1051                                 ret = btrfs_inc_ref(trans, root, cow, 1);
1052                         else
1053                                 ret = btrfs_inc_ref(trans, root, cow, 0);
1054                         BUG_ON(ret); /* -ENOMEM */
1055                 }
1056                 if (new_flags != 0) {
1057                         int level = btrfs_header_level(buf);
1058
1059                         ret = btrfs_set_disk_extent_flags(trans, fs_info,
1060                                                           buf->start,
1061                                                           buf->len,
1062                                                           new_flags, level, 0);
1063                         if (ret)
1064                                 return ret;
1065                 }
1066         } else {
1067                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
1068                         if (root->root_key.objectid ==
1069                             BTRFS_TREE_RELOC_OBJECTID)
1070                                 ret = btrfs_inc_ref(trans, root, cow, 1);
1071                         else
1072                                 ret = btrfs_inc_ref(trans, root, cow, 0);
1073                         BUG_ON(ret); /* -ENOMEM */
1074                         ret = btrfs_dec_ref(trans, root, buf, 1);
1075                         BUG_ON(ret); /* -ENOMEM */
1076                 }
1077                 clean_tree_block(fs_info, buf);
1078                 *last_ref = 1;
1079         }
1080         return 0;
1081 }
1082
1083 /*
1084  * does the dirty work in cow of a single block.  The parent block (if
1085  * supplied) is updated to point to the new cow copy.  The new buffer is marked
1086  * dirty and returned locked.  If you modify the block it needs to be marked
1087  * dirty again.
1088  *
1089  * search_start -- an allocation hint for the new block
1090  *
1091  * empty_size -- a hint that you plan on doing more cow.  This is the size in
1092  * bytes the allocator should try to find free next to the block it returns.
1093  * This is just a hint and may be ignored by the allocator.
1094  */
1095 static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
1096                              struct btrfs_root *root,
1097                              struct extent_buffer *buf,
1098                              struct extent_buffer *parent, int parent_slot,
1099                              struct extent_buffer **cow_ret,
1100                              u64 search_start, u64 empty_size)
1101 {
1102         struct btrfs_fs_info *fs_info = root->fs_info;
1103         struct btrfs_disk_key disk_key;
1104         struct extent_buffer *cow;
1105         int level, ret;
1106         int last_ref = 0;
1107         int unlock_orig = 0;
1108         u64 parent_start = 0;
1109
1110         if (*cow_ret == buf)
1111                 unlock_orig = 1;
1112
1113         btrfs_assert_tree_locked(buf);
1114
1115         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1116                 trans->transid != fs_info->running_transaction->transid);
1117         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1118                 trans->transid != root->last_trans);
1119
1120         level = btrfs_header_level(buf);
1121
1122         if (level == 0)
1123                 btrfs_item_key(buf, &disk_key, 0);
1124         else
1125                 btrfs_node_key(buf, &disk_key, 0);
1126
1127         if ((root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) && parent)
1128                 parent_start = parent->start;
1129
1130         cow = btrfs_alloc_tree_block(trans, root, parent_start,
1131                         root->root_key.objectid, &disk_key, level,
1132                         search_start, empty_size);
1133         if (IS_ERR(cow))
1134                 return PTR_ERR(cow);
1135
1136         /* cow is set to blocking by btrfs_init_new_buffer */
1137
1138         copy_extent_buffer_full(cow, buf);
1139         btrfs_set_header_bytenr(cow, cow->start);
1140         btrfs_set_header_generation(cow, trans->transid);
1141         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
1142         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
1143                                      BTRFS_HEADER_FLAG_RELOC);
1144         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1145                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
1146         else
1147                 btrfs_set_header_owner(cow, root->root_key.objectid);
1148
1149         write_extent_buffer_fsid(cow, fs_info->fsid);
1150
1151         ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
1152         if (ret) {
1153                 btrfs_abort_transaction(trans, ret);
1154                 return ret;
1155         }
1156
1157         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
1158                 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
1159                 if (ret) {
1160                         btrfs_abort_transaction(trans, ret);
1161                         return ret;
1162                 }
1163         }
1164
1165         if (buf == root->node) {
1166                 WARN_ON(parent && parent != buf);
1167                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1168                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1169                         parent_start = buf->start;
1170
1171                 extent_buffer_get(cow);
1172                 tree_mod_log_set_root_pointer(root, cow, 1);
1173                 rcu_assign_pointer(root->node, cow);
1174
1175                 btrfs_free_tree_block(trans, root, buf, parent_start,
1176                                       last_ref);
1177                 free_extent_buffer(buf);
1178                 add_root_to_dirty_list(root);
1179         } else {
1180                 WARN_ON(trans->transid != btrfs_header_generation(parent));
1181                 tree_mod_log_insert_key(fs_info, parent, parent_slot,
1182                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
1183                 btrfs_set_node_blockptr(parent, parent_slot,
1184                                         cow->start);
1185                 btrfs_set_node_ptr_generation(parent, parent_slot,
1186                                               trans->transid);
1187                 btrfs_mark_buffer_dirty(parent);
1188                 if (last_ref) {
1189                         ret = tree_mod_log_free_eb(fs_info, buf);
1190                         if (ret) {
1191                                 btrfs_abort_transaction(trans, ret);
1192                                 return ret;
1193                         }
1194                 }
1195                 btrfs_free_tree_block(trans, root, buf, parent_start,
1196                                       last_ref);
1197         }
1198         if (unlock_orig)
1199                 btrfs_tree_unlock(buf);
1200         free_extent_buffer_stale(buf);
1201         btrfs_mark_buffer_dirty(cow);
1202         *cow_ret = cow;
1203         return 0;
1204 }
1205
1206 /*
1207  * returns the logical address of the oldest predecessor of the given root.
1208  * entries older than time_seq are ignored.
1209  */
1210 static struct tree_mod_elem *
1211 __tree_mod_log_oldest_root(struct btrfs_fs_info *fs_info,
1212                            struct extent_buffer *eb_root, u64 time_seq)
1213 {
1214         struct tree_mod_elem *tm;
1215         struct tree_mod_elem *found = NULL;
1216         u64 root_logical = eb_root->start;
1217         int looped = 0;
1218
1219         if (!time_seq)
1220                 return NULL;
1221
1222         /*
1223          * the very last operation that's logged for a root is the
1224          * replacement operation (if it is replaced at all). this has
1225          * the logical address of the *new* root, making it the very
1226          * first operation that's logged for this root.
1227          */
1228         while (1) {
1229                 tm = tree_mod_log_search_oldest(fs_info, root_logical,
1230                                                 time_seq);
1231                 if (!looped && !tm)
1232                         return NULL;
1233                 /*
1234                  * if there are no tree operation for the oldest root, we simply
1235                  * return it. this should only happen if that (old) root is at
1236                  * level 0.
1237                  */
1238                 if (!tm)
1239                         break;
1240
1241                 /*
1242                  * if there's an operation that's not a root replacement, we
1243                  * found the oldest version of our root. normally, we'll find a
1244                  * MOD_LOG_KEY_REMOVE_WHILE_FREEING operation here.
1245                  */
1246                 if (tm->op != MOD_LOG_ROOT_REPLACE)
1247                         break;
1248
1249                 found = tm;
1250                 root_logical = tm->old_root.logical;
1251                 looped = 1;
1252         }
1253
1254         /* if there's no old root to return, return what we found instead */
1255         if (!found)
1256                 found = tm;
1257
1258         return found;
1259 }
1260
1261 /*
1262  * tm is a pointer to the first operation to rewind within eb. then, all
1263  * previous operations will be rewound (until we reach something older than
1264  * time_seq).
1265  */
1266 static void
1267 __tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct extent_buffer *eb,
1268                       u64 time_seq, struct tree_mod_elem *first_tm)
1269 {
1270         u32 n;
1271         struct rb_node *next;
1272         struct tree_mod_elem *tm = first_tm;
1273         unsigned long o_dst;
1274         unsigned long o_src;
1275         unsigned long p_size = sizeof(struct btrfs_key_ptr);
1276
1277         n = btrfs_header_nritems(eb);
1278         tree_mod_log_read_lock(fs_info);
1279         while (tm && tm->seq >= time_seq) {
1280                 /*
1281                  * all the operations are recorded with the operator used for
1282                  * the modification. as we're going backwards, we do the
1283                  * opposite of each operation here.
1284                  */
1285                 switch (tm->op) {
1286                 case MOD_LOG_KEY_REMOVE_WHILE_FREEING:
1287                         BUG_ON(tm->slot < n);
1288                         /* Fallthrough */
1289                 case MOD_LOG_KEY_REMOVE_WHILE_MOVING:
1290                 case MOD_LOG_KEY_REMOVE:
1291                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1292                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1293                         btrfs_set_node_ptr_generation(eb, tm->slot,
1294                                                       tm->generation);
1295                         n++;
1296                         break;
1297                 case MOD_LOG_KEY_REPLACE:
1298                         BUG_ON(tm->slot >= n);
1299                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1300                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1301                         btrfs_set_node_ptr_generation(eb, tm->slot,
1302                                                       tm->generation);
1303                         break;
1304                 case MOD_LOG_KEY_ADD:
1305                         /* if a move operation is needed it's in the log */
1306                         n--;
1307                         break;
1308                 case MOD_LOG_MOVE_KEYS:
1309                         o_dst = btrfs_node_key_ptr_offset(tm->slot);
1310                         o_src = btrfs_node_key_ptr_offset(tm->move.dst_slot);
1311                         memmove_extent_buffer(eb, o_dst, o_src,
1312                                               tm->move.nr_items * p_size);
1313                         break;
1314                 case MOD_LOG_ROOT_REPLACE:
1315                         /*
1316                          * this operation is special. for roots, this must be
1317                          * handled explicitly before rewinding.
1318                          * for non-roots, this operation may exist if the node
1319                          * was a root: root A -> child B; then A gets empty and
1320                          * B is promoted to the new root. in the mod log, we'll
1321                          * have a root-replace operation for B, a tree block
1322                          * that is no root. we simply ignore that operation.
1323                          */
1324                         break;
1325                 }
1326                 next = rb_next(&tm->node);
1327                 if (!next)
1328                         break;
1329                 tm = rb_entry(next, struct tree_mod_elem, node);
1330                 if (tm->logical != first_tm->logical)
1331                         break;
1332         }
1333         tree_mod_log_read_unlock(fs_info);
1334         btrfs_set_header_nritems(eb, n);
1335 }
1336
1337 /*
1338  * Called with eb read locked. If the buffer cannot be rewound, the same buffer
1339  * is returned. If rewind operations happen, a fresh buffer is returned. The
1340  * returned buffer is always read-locked. If the returned buffer is not the
1341  * input buffer, the lock on the input buffer is released and the input buffer
1342  * is freed (its refcount is decremented).
1343  */
1344 static struct extent_buffer *
1345 tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
1346                     struct extent_buffer *eb, u64 time_seq)
1347 {
1348         struct extent_buffer *eb_rewin;
1349         struct tree_mod_elem *tm;
1350
1351         if (!time_seq)
1352                 return eb;
1353
1354         if (btrfs_header_level(eb) == 0)
1355                 return eb;
1356
1357         tm = tree_mod_log_search(fs_info, eb->start, time_seq);
1358         if (!tm)
1359                 return eb;
1360
1361         btrfs_set_path_blocking(path);
1362         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1363
1364         if (tm->op == MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1365                 BUG_ON(tm->slot != 0);
1366                 eb_rewin = alloc_dummy_extent_buffer(fs_info, eb->start);
1367                 if (!eb_rewin) {
1368                         btrfs_tree_read_unlock_blocking(eb);
1369                         free_extent_buffer(eb);
1370                         return NULL;
1371                 }
1372                 btrfs_set_header_bytenr(eb_rewin, eb->start);
1373                 btrfs_set_header_backref_rev(eb_rewin,
1374                                              btrfs_header_backref_rev(eb));
1375                 btrfs_set_header_owner(eb_rewin, btrfs_header_owner(eb));
1376                 btrfs_set_header_level(eb_rewin, btrfs_header_level(eb));
1377         } else {
1378                 eb_rewin = btrfs_clone_extent_buffer(eb);
1379                 if (!eb_rewin) {
1380                         btrfs_tree_read_unlock_blocking(eb);
1381                         free_extent_buffer(eb);
1382                         return NULL;
1383                 }
1384         }
1385
1386         btrfs_clear_path_blocking(path, NULL, BTRFS_READ_LOCK);
1387         btrfs_tree_read_unlock_blocking(eb);
1388         free_extent_buffer(eb);
1389
1390         extent_buffer_get(eb_rewin);
1391         btrfs_tree_read_lock(eb_rewin);
1392         __tree_mod_log_rewind(fs_info, eb_rewin, time_seq, tm);
1393         WARN_ON(btrfs_header_nritems(eb_rewin) >
1394                 BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1395
1396         return eb_rewin;
1397 }
1398
1399 /*
1400  * get_old_root() rewinds the state of @root's root node to the given @time_seq
1401  * value. If there are no changes, the current root->root_node is returned. If
1402  * anything changed in between, there's a fresh buffer allocated on which the
1403  * rewind operations are done. In any case, the returned buffer is read locked.
1404  * Returns NULL on error (with no locks held).
1405  */
1406 static inline struct extent_buffer *
1407 get_old_root(struct btrfs_root *root, u64 time_seq)
1408 {
1409         struct btrfs_fs_info *fs_info = root->fs_info;
1410         struct tree_mod_elem *tm;
1411         struct extent_buffer *eb = NULL;
1412         struct extent_buffer *eb_root;
1413         struct extent_buffer *old;
1414         struct tree_mod_root *old_root = NULL;
1415         u64 old_generation = 0;
1416         u64 logical;
1417
1418         eb_root = btrfs_read_lock_root_node(root);
1419         tm = __tree_mod_log_oldest_root(fs_info, eb_root, time_seq);
1420         if (!tm)
1421                 return eb_root;
1422
1423         if (tm->op == MOD_LOG_ROOT_REPLACE) {
1424                 old_root = &tm->old_root;
1425                 old_generation = tm->generation;
1426                 logical = old_root->logical;
1427         } else {
1428                 logical = eb_root->start;
1429         }
1430
1431         tm = tree_mod_log_search(fs_info, logical, time_seq);
1432         if (old_root && tm && tm->op != MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1433                 btrfs_tree_read_unlock(eb_root);
1434                 free_extent_buffer(eb_root);
1435                 old = read_tree_block(fs_info, logical, 0);
1436                 if (WARN_ON(IS_ERR(old) || !extent_buffer_uptodate(old))) {
1437                         if (!IS_ERR(old))
1438                                 free_extent_buffer(old);
1439                         btrfs_warn(fs_info,
1440                                    "failed to read tree block %llu from get_old_root",
1441                                    logical);
1442                 } else {
1443                         eb = btrfs_clone_extent_buffer(old);
1444                         free_extent_buffer(old);
1445                 }
1446         } else if (old_root) {
1447                 btrfs_tree_read_unlock(eb_root);
1448                 free_extent_buffer(eb_root);
1449                 eb = alloc_dummy_extent_buffer(fs_info, logical);
1450         } else {
1451                 btrfs_set_lock_blocking_rw(eb_root, BTRFS_READ_LOCK);
1452                 eb = btrfs_clone_extent_buffer(eb_root);
1453                 btrfs_tree_read_unlock_blocking(eb_root);
1454                 free_extent_buffer(eb_root);
1455         }
1456
1457         if (!eb)
1458                 return NULL;
1459         extent_buffer_get(eb);
1460         btrfs_tree_read_lock(eb);
1461         if (old_root) {
1462                 btrfs_set_header_bytenr(eb, eb->start);
1463                 btrfs_set_header_backref_rev(eb, BTRFS_MIXED_BACKREF_REV);
1464                 btrfs_set_header_owner(eb, btrfs_header_owner(eb_root));
1465                 btrfs_set_header_level(eb, old_root->level);
1466                 btrfs_set_header_generation(eb, old_generation);
1467         }
1468         if (tm)
1469                 __tree_mod_log_rewind(fs_info, eb, time_seq, tm);
1470         else
1471                 WARN_ON(btrfs_header_level(eb) != 0);
1472         WARN_ON(btrfs_header_nritems(eb) > BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1473
1474         return eb;
1475 }
1476
1477 int btrfs_old_root_level(struct btrfs_root *root, u64 time_seq)
1478 {
1479         struct tree_mod_elem *tm;
1480         int level;
1481         struct extent_buffer *eb_root = btrfs_root_node(root);
1482
1483         tm = __tree_mod_log_oldest_root(root->fs_info, eb_root, time_seq);
1484         if (tm && tm->op == MOD_LOG_ROOT_REPLACE) {
1485                 level = tm->old_root.level;
1486         } else {
1487                 level = btrfs_header_level(eb_root);
1488         }
1489         free_extent_buffer(eb_root);
1490
1491         return level;
1492 }
1493
1494 static inline int should_cow_block(struct btrfs_trans_handle *trans,
1495                                    struct btrfs_root *root,
1496                                    struct extent_buffer *buf)
1497 {
1498         if (btrfs_is_testing(root->fs_info))
1499                 return 0;
1500
1501         /* ensure we can see the force_cow */
1502         smp_rmb();
1503
1504         /*
1505          * We do not need to cow a block if
1506          * 1) this block is not created or changed in this transaction;
1507          * 2) this block does not belong to TREE_RELOC tree;
1508          * 3) the root is not forced COW.
1509          *
1510          * What is forced COW:
1511          *    when we create snapshot during committing the transaction,
1512          *    after we've finished coping src root, we must COW the shared
1513          *    block to ensure the metadata consistency.
1514          */
1515         if (btrfs_header_generation(buf) == trans->transid &&
1516             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
1517             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
1518               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
1519             !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
1520                 return 0;
1521         return 1;
1522 }
1523
1524 /*
1525  * cows a single block, see __btrfs_cow_block for the real work.
1526  * This version of it has extra checks so that a block isn't COWed more than
1527  * once per transaction, as long as it hasn't been written yet
1528  */
1529 noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
1530                     struct btrfs_root *root, struct extent_buffer *buf,
1531                     struct extent_buffer *parent, int parent_slot,
1532                     struct extent_buffer **cow_ret)
1533 {
1534         struct btrfs_fs_info *fs_info = root->fs_info;
1535         u64 search_start;
1536         int ret;
1537
1538         if (trans->transaction != fs_info->running_transaction)
1539                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1540                        trans->transid,
1541                        fs_info->running_transaction->transid);
1542
1543         if (trans->transid != fs_info->generation)
1544                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1545                        trans->transid, fs_info->generation);
1546
1547         if (!should_cow_block(trans, root, buf)) {
1548                 trans->dirty = true;
1549                 *cow_ret = buf;
1550                 return 0;
1551         }
1552
1553         search_start = buf->start & ~((u64)SZ_1G - 1);
1554
1555         if (parent)
1556                 btrfs_set_lock_blocking(parent);
1557         btrfs_set_lock_blocking(buf);
1558
1559         ret = __btrfs_cow_block(trans, root, buf, parent,
1560                                  parent_slot, cow_ret, search_start, 0);
1561
1562         trace_btrfs_cow_block(root, buf, *cow_ret);
1563
1564         return ret;
1565 }
1566
1567 /*
1568  * helper function for defrag to decide if two blocks pointed to by a
1569  * node are actually close by
1570  */
1571 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
1572 {
1573         if (blocknr < other && other - (blocknr + blocksize) < 32768)
1574                 return 1;
1575         if (blocknr > other && blocknr - (other + blocksize) < 32768)
1576                 return 1;
1577         return 0;
1578 }
1579
1580 /*
1581  * compare two keys in a memcmp fashion
1582  */
1583 static int comp_keys(const struct btrfs_disk_key *disk,
1584                      const struct btrfs_key *k2)
1585 {
1586         struct btrfs_key k1;
1587
1588         btrfs_disk_key_to_cpu(&k1, disk);
1589
1590         return btrfs_comp_cpu_keys(&k1, k2);
1591 }
1592
1593 /*
1594  * same as comp_keys only with two btrfs_key's
1595  */
1596 int btrfs_comp_cpu_keys(const struct btrfs_key *k1, const struct btrfs_key *k2)
1597 {
1598         if (k1->objectid > k2->objectid)
1599                 return 1;
1600         if (k1->objectid < k2->objectid)
1601                 return -1;
1602         if (k1->type > k2->type)
1603                 return 1;
1604         if (k1->type < k2->type)
1605                 return -1;
1606         if (k1->offset > k2->offset)
1607                 return 1;
1608         if (k1->offset < k2->offset)
1609                 return -1;
1610         return 0;
1611 }
1612
1613 /*
1614  * this is used by the defrag code to go through all the
1615  * leaves pointed to by a node and reallocate them so that
1616  * disk order is close to key order
1617  */
1618 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
1619                        struct btrfs_root *root, struct extent_buffer *parent,
1620                        int start_slot, u64 *last_ret,
1621                        struct btrfs_key *progress)
1622 {
1623         struct btrfs_fs_info *fs_info = root->fs_info;
1624         struct extent_buffer *cur;
1625         u64 blocknr;
1626         u64 gen;
1627         u64 search_start = *last_ret;
1628         u64 last_block = 0;
1629         u64 other;
1630         u32 parent_nritems;
1631         int end_slot;
1632         int i;
1633         int err = 0;
1634         int parent_level;
1635         int uptodate;
1636         u32 blocksize;
1637         int progress_passed = 0;
1638         struct btrfs_disk_key disk_key;
1639
1640         parent_level = btrfs_header_level(parent);
1641
1642         WARN_ON(trans->transaction != fs_info->running_transaction);
1643         WARN_ON(trans->transid != fs_info->generation);
1644
1645         parent_nritems = btrfs_header_nritems(parent);
1646         blocksize = fs_info->nodesize;
1647         end_slot = parent_nritems - 1;
1648
1649         if (parent_nritems <= 1)
1650                 return 0;
1651
1652         btrfs_set_lock_blocking(parent);
1653
1654         for (i = start_slot; i <= end_slot; i++) {
1655                 int close = 1;
1656
1657                 btrfs_node_key(parent, &disk_key, i);
1658                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
1659                         continue;
1660
1661                 progress_passed = 1;
1662                 blocknr = btrfs_node_blockptr(parent, i);
1663                 gen = btrfs_node_ptr_generation(parent, i);
1664                 if (last_block == 0)
1665                         last_block = blocknr;
1666
1667                 if (i > 0) {
1668                         other = btrfs_node_blockptr(parent, i - 1);
1669                         close = close_blocks(blocknr, other, blocksize);
1670                 }
1671                 if (!close && i < end_slot) {
1672                         other = btrfs_node_blockptr(parent, i + 1);
1673                         close = close_blocks(blocknr, other, blocksize);
1674                 }
1675                 if (close) {
1676                         last_block = blocknr;
1677                         continue;
1678                 }
1679
1680                 cur = find_extent_buffer(fs_info, blocknr);
1681                 if (cur)
1682                         uptodate = btrfs_buffer_uptodate(cur, gen, 0);
1683                 else
1684                         uptodate = 0;
1685                 if (!cur || !uptodate) {
1686                         if (!cur) {
1687                                 cur = read_tree_block(fs_info, blocknr, gen);
1688                                 if (IS_ERR(cur)) {
1689                                         return PTR_ERR(cur);
1690                                 } else if (!extent_buffer_uptodate(cur)) {
1691                                         free_extent_buffer(cur);
1692                                         return -EIO;
1693                                 }
1694                         } else if (!uptodate) {
1695                                 err = btrfs_read_buffer(cur, gen);
1696                                 if (err) {
1697                                         free_extent_buffer(cur);
1698                                         return err;
1699                                 }
1700                         }
1701                 }
1702                 if (search_start == 0)
1703                         search_start = last_block;
1704
1705                 btrfs_tree_lock(cur);
1706                 btrfs_set_lock_blocking(cur);
1707                 err = __btrfs_cow_block(trans, root, cur, parent, i,
1708                                         &cur, search_start,
1709                                         min(16 * blocksize,
1710                                             (end_slot - i) * blocksize));
1711                 if (err) {
1712                         btrfs_tree_unlock(cur);
1713                         free_extent_buffer(cur);
1714                         break;
1715                 }
1716                 search_start = cur->start;
1717                 last_block = cur->start;
1718                 *last_ret = search_start;
1719                 btrfs_tree_unlock(cur);
1720                 free_extent_buffer(cur);
1721         }
1722         return err;
1723 }
1724
1725 /*
1726  * search for key in the extent_buffer.  The items start at offset p,
1727  * and they are item_size apart.  There are 'max' items in p.
1728  *
1729  * the slot in the array is returned via slot, and it points to
1730  * the place where you would insert key if it is not found in
1731  * the array.
1732  *
1733  * slot may point to max if the key is bigger than all of the keys
1734  */
1735 static noinline int generic_bin_search(struct extent_buffer *eb,
1736                                        unsigned long p, int item_size,
1737                                        const struct btrfs_key *key,
1738                                        int max, int *slot)
1739 {
1740         int low = 0;
1741         int high = max;
1742         int mid;
1743         int ret;
1744         struct btrfs_disk_key *tmp = NULL;
1745         struct btrfs_disk_key unaligned;
1746         unsigned long offset;
1747         char *kaddr = NULL;
1748         unsigned long map_start = 0;
1749         unsigned long map_len = 0;
1750         int err;
1751
1752         if (low > high) {
1753                 btrfs_err(eb->fs_info,
1754                  "%s: low (%d) > high (%d) eb %llu owner %llu level %d",
1755                           __func__, low, high, eb->start,
1756                           btrfs_header_owner(eb), btrfs_header_level(eb));
1757                 return -EINVAL;
1758         }
1759
1760         while (low < high) {
1761                 mid = (low + high) / 2;
1762                 offset = p + mid * item_size;
1763
1764                 if (!kaddr || offset < map_start ||
1765                     (offset + sizeof(struct btrfs_disk_key)) >
1766                     map_start + map_len) {
1767
1768                         err = map_private_extent_buffer(eb, offset,
1769                                                 sizeof(struct btrfs_disk_key),
1770                                                 &kaddr, &map_start, &map_len);
1771
1772                         if (!err) {
1773                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1774                                                         map_start);
1775                         } else if (err == 1) {
1776                                 read_extent_buffer(eb, &unaligned,
1777                                                    offset, sizeof(unaligned));
1778                                 tmp = &unaligned;
1779                         } else {
1780                                 return err;
1781                         }
1782
1783                 } else {
1784                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
1785                                                         map_start);
1786                 }
1787                 ret = comp_keys(tmp, key);
1788
1789                 if (ret < 0)
1790                         low = mid + 1;
1791                 else if (ret > 0)
1792                         high = mid;
1793                 else {
1794                         *slot = mid;
1795                         return 0;
1796                 }
1797         }
1798         *slot = low;
1799         return 1;
1800 }
1801
1802 /*
1803  * simple bin_search frontend that does the right thing for
1804  * leaves vs nodes
1805  */
1806 static int bin_search(struct extent_buffer *eb, const struct btrfs_key *key,
1807                       int level, int *slot)
1808 {
1809         if (level == 0)
1810                 return generic_bin_search(eb,
1811                                           offsetof(struct btrfs_leaf, items),
1812                                           sizeof(struct btrfs_item),
1813                                           key, btrfs_header_nritems(eb),
1814                                           slot);
1815         else
1816                 return generic_bin_search(eb,
1817                                           offsetof(struct btrfs_node, ptrs),
1818                                           sizeof(struct btrfs_key_ptr),
1819                                           key, btrfs_header_nritems(eb),
1820                                           slot);
1821 }
1822
1823 int btrfs_bin_search(struct extent_buffer *eb, const struct btrfs_key *key,
1824                      int level, int *slot)
1825 {
1826         return bin_search(eb, key, level, slot);
1827 }
1828
1829 static void root_add_used(struct btrfs_root *root, u32 size)
1830 {
1831         spin_lock(&root->accounting_lock);
1832         btrfs_set_root_used(&root->root_item,
1833                             btrfs_root_used(&root->root_item) + size);
1834         spin_unlock(&root->accounting_lock);
1835 }
1836
1837 static void root_sub_used(struct btrfs_root *root, u32 size)
1838 {
1839         spin_lock(&root->accounting_lock);
1840         btrfs_set_root_used(&root->root_item,
1841                             btrfs_root_used(&root->root_item) - size);
1842         spin_unlock(&root->accounting_lock);
1843 }
1844
1845 /* given a node and slot number, this reads the blocks it points to.  The
1846  * extent buffer is returned with a reference taken (but unlocked).
1847  */
1848 static noinline struct extent_buffer *
1849 read_node_slot(struct btrfs_fs_info *fs_info, struct extent_buffer *parent,
1850                int slot)
1851 {
1852         int level = btrfs_header_level(parent);
1853         struct extent_buffer *eb;
1854
1855         if (slot < 0 || slot >= btrfs_header_nritems(parent))
1856                 return ERR_PTR(-ENOENT);
1857
1858         BUG_ON(level == 0);
1859
1860         eb = read_tree_block(fs_info, btrfs_node_blockptr(parent, slot),
1861                              btrfs_node_ptr_generation(parent, slot));
1862         if (!IS_ERR(eb) && !extent_buffer_uptodate(eb)) {
1863                 free_extent_buffer(eb);
1864                 eb = ERR_PTR(-EIO);
1865         }
1866
1867         return eb;
1868 }
1869
1870 /*
1871  * node level balancing, used to make sure nodes are in proper order for
1872  * item deletion.  We balance from the top down, so we have to make sure
1873  * that a deletion won't leave an node completely empty later on.
1874  */
1875 static noinline int balance_level(struct btrfs_trans_handle *trans,
1876                          struct btrfs_root *root,
1877                          struct btrfs_path *path, int level)
1878 {
1879         struct btrfs_fs_info *fs_info = root->fs_info;
1880         struct extent_buffer *right = NULL;
1881         struct extent_buffer *mid;
1882         struct extent_buffer *left = NULL;
1883         struct extent_buffer *parent = NULL;
1884         int ret = 0;
1885         int wret;
1886         int pslot;
1887         int orig_slot = path->slots[level];
1888         u64 orig_ptr;
1889
1890         if (level == 0)
1891                 return 0;
1892
1893         mid = path->nodes[level];
1894
1895         WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK &&
1896                 path->locks[level] != BTRFS_WRITE_LOCK_BLOCKING);
1897         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1898
1899         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1900
1901         if (level < BTRFS_MAX_LEVEL - 1) {
1902                 parent = path->nodes[level + 1];
1903                 pslot = path->slots[level + 1];
1904         }
1905
1906         /*
1907          * deal with the case where there is only one pointer in the root
1908          * by promoting the node below to a root
1909          */
1910         if (!parent) {
1911                 struct extent_buffer *child;
1912
1913                 if (btrfs_header_nritems(mid) != 1)
1914                         return 0;
1915
1916                 /* promote the child to a root */
1917                 child = read_node_slot(fs_info, mid, 0);
1918                 if (IS_ERR(child)) {
1919                         ret = PTR_ERR(child);
1920                         btrfs_handle_fs_error(fs_info, ret, NULL);
1921                         goto enospc;
1922                 }
1923
1924                 btrfs_tree_lock(child);
1925                 btrfs_set_lock_blocking(child);
1926                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
1927                 if (ret) {
1928                         btrfs_tree_unlock(child);
1929                         free_extent_buffer(child);
1930                         goto enospc;
1931                 }
1932
1933                 tree_mod_log_set_root_pointer(root, child, 1);
1934                 rcu_assign_pointer(root->node, child);
1935
1936                 add_root_to_dirty_list(root);
1937                 btrfs_tree_unlock(child);
1938
1939                 path->locks[level] = 0;
1940                 path->nodes[level] = NULL;
1941                 clean_tree_block(fs_info, mid);
1942                 btrfs_tree_unlock(mid);
1943                 /* once for the path */
1944                 free_extent_buffer(mid);
1945
1946                 root_sub_used(root, mid->len);
1947                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1948                 /* once for the root ptr */
1949                 free_extent_buffer_stale(mid);
1950                 return 0;
1951         }
1952         if (btrfs_header_nritems(mid) >
1953             BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 4)
1954                 return 0;
1955
1956         left = read_node_slot(fs_info, parent, pslot - 1);
1957         if (IS_ERR(left))
1958                 left = NULL;
1959
1960         if (left) {
1961                 btrfs_tree_lock(left);
1962                 btrfs_set_lock_blocking(left);
1963                 wret = btrfs_cow_block(trans, root, left,
1964                                        parent, pslot - 1, &left);
1965                 if (wret) {
1966                         ret = wret;
1967                         goto enospc;
1968                 }
1969         }
1970
1971         right = read_node_slot(fs_info, parent, pslot + 1);
1972         if (IS_ERR(right))
1973                 right = NULL;
1974
1975         if (right) {
1976                 btrfs_tree_lock(right);
1977                 btrfs_set_lock_blocking(right);
1978                 wret = btrfs_cow_block(trans, root, right,
1979                                        parent, pslot + 1, &right);
1980                 if (wret) {
1981                         ret = wret;
1982                         goto enospc;
1983                 }
1984         }
1985
1986         /* first, try to make some room in the middle buffer */
1987         if (left) {
1988                 orig_slot += btrfs_header_nritems(left);
1989                 wret = push_node_left(trans, fs_info, left, mid, 1);
1990                 if (wret < 0)
1991                         ret = wret;
1992         }
1993
1994         /*
1995          * then try to empty the right most buffer into the middle
1996          */
1997         if (right) {
1998                 wret = push_node_left(trans, fs_info, mid, right, 1);
1999                 if (wret < 0 && wret != -ENOSPC)
2000                         ret = wret;
2001                 if (btrfs_header_nritems(right) == 0) {
2002                         clean_tree_block(fs_info, right);
2003                         btrfs_tree_unlock(right);
2004                         del_ptr(root, path, level + 1, pslot + 1);
2005                         root_sub_used(root, right->len);
2006                         btrfs_free_tree_block(trans, root, right, 0, 1);
2007                         free_extent_buffer_stale(right);
2008                         right = NULL;
2009                 } else {
2010                         struct btrfs_disk_key right_key;
2011                         btrfs_node_key(right, &right_key, 0);
2012                         tree_mod_log_set_node_key(fs_info, parent,
2013                                                   pslot + 1, 0);
2014                         btrfs_set_node_key(parent, &right_key, pslot + 1);
2015                         btrfs_mark_buffer_dirty(parent);
2016                 }
2017         }
2018         if (btrfs_header_nritems(mid) == 1) {
2019                 /*
2020                  * we're not allowed to leave a node with one item in the
2021                  * tree during a delete.  A deletion from lower in the tree
2022                  * could try to delete the only pointer in this node.
2023                  * So, pull some keys from the left.
2024                  * There has to be a left pointer at this point because
2025                  * otherwise we would have pulled some pointers from the
2026                  * right
2027                  */
2028                 if (!left) {
2029                         ret = -EROFS;
2030                         btrfs_handle_fs_error(fs_info, ret, NULL);
2031                         goto enospc;
2032                 }
2033                 wret = balance_node_right(trans, fs_info, mid, left);
2034                 if (wret < 0) {
2035                         ret = wret;
2036                         goto enospc;
2037                 }
2038                 if (wret == 1) {
2039                         wret = push_node_left(trans, fs_info, left, mid, 1);
2040                         if (wret < 0)
2041                                 ret = wret;
2042                 }
2043                 BUG_ON(wret == 1);
2044         }
2045         if (btrfs_header_nritems(mid) == 0) {
2046                 clean_tree_block(fs_info, mid);
2047                 btrfs_tree_unlock(mid);
2048                 del_ptr(root, path, level + 1, pslot);
2049                 root_sub_used(root, mid->len);
2050                 btrfs_free_tree_block(trans, root, mid, 0, 1);
2051                 free_extent_buffer_stale(mid);
2052                 mid = NULL;
2053         } else {
2054                 /* update the parent key to reflect our changes */
2055                 struct btrfs_disk_key mid_key;
2056                 btrfs_node_key(mid, &mid_key, 0);
2057                 tree_mod_log_set_node_key(fs_info, parent, pslot, 0);
2058                 btrfs_set_node_key(parent, &mid_key, pslot);
2059                 btrfs_mark_buffer_dirty(parent);
2060         }
2061
2062         /* update the path */
2063         if (left) {
2064                 if (btrfs_header_nritems(left) > orig_slot) {
2065                         extent_buffer_get(left);
2066                         /* left was locked after cow */
2067                         path->nodes[level] = left;
2068                         path->slots[level + 1] -= 1;
2069                         path->slots[level] = orig_slot;
2070                         if (mid) {
2071                                 btrfs_tree_unlock(mid);
2072                                 free_extent_buffer(mid);
2073                         }
2074                 } else {
2075                         orig_slot -= btrfs_header_nritems(left);
2076                         path->slots[level] = orig_slot;
2077                 }
2078         }
2079         /* double check we haven't messed things up */
2080         if (orig_ptr !=
2081             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
2082                 BUG();
2083 enospc:
2084         if (right) {
2085                 btrfs_tree_unlock(right);
2086                 free_extent_buffer(right);
2087         }
2088         if (left) {
2089                 if (path->nodes[level] != left)
2090                         btrfs_tree_unlock(left);
2091                 free_extent_buffer(left);
2092         }
2093         return ret;
2094 }
2095
2096 /* Node balancing for insertion.  Here we only split or push nodes around
2097  * when they are completely full.  This is also done top down, so we
2098  * have to be pessimistic.
2099  */
2100 static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
2101                                           struct btrfs_root *root,
2102                                           struct btrfs_path *path, int level)
2103 {
2104         struct btrfs_fs_info *fs_info = root->fs_info;
2105         struct extent_buffer *right = NULL;
2106         struct extent_buffer *mid;
2107         struct extent_buffer *left = NULL;
2108         struct extent_buffer *parent = NULL;
2109         int ret = 0;
2110         int wret;
2111         int pslot;
2112         int orig_slot = path->slots[level];
2113
2114         if (level == 0)
2115                 return 1;
2116
2117         mid = path->nodes[level];
2118         WARN_ON(btrfs_header_generation(mid) != trans->transid);
2119
2120         if (level < BTRFS_MAX_LEVEL - 1) {
2121                 parent = path->nodes[level + 1];
2122                 pslot = path->slots[level + 1];
2123         }
2124
2125         if (!parent)
2126                 return 1;
2127
2128         left = read_node_slot(fs_info, parent, pslot - 1);
2129         if (IS_ERR(left))
2130                 left = NULL;
2131
2132         /* first, try to make some room in the middle buffer */
2133         if (left) {
2134                 u32 left_nr;
2135
2136                 btrfs_tree_lock(left);
2137                 btrfs_set_lock_blocking(left);
2138
2139                 left_nr = btrfs_header_nritems(left);
2140                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2141                         wret = 1;
2142                 } else {
2143                         ret = btrfs_cow_block(trans, root, left, parent,
2144                                               pslot - 1, &left);
2145                         if (ret)
2146                                 wret = 1;
2147                         else {
2148                                 wret = push_node_left(trans, fs_info,
2149                                                       left, mid, 0);
2150                         }
2151                 }
2152                 if (wret < 0)
2153                         ret = wret;
2154                 if (wret == 0) {
2155                         struct btrfs_disk_key disk_key;
2156                         orig_slot += left_nr;
2157                         btrfs_node_key(mid, &disk_key, 0);
2158                         tree_mod_log_set_node_key(fs_info, parent, pslot, 0);
2159                         btrfs_set_node_key(parent, &disk_key, pslot);
2160                         btrfs_mark_buffer_dirty(parent);
2161                         if (btrfs_header_nritems(left) > orig_slot) {
2162                                 path->nodes[level] = left;
2163                                 path->slots[level + 1] -= 1;
2164                                 path->slots[level] = orig_slot;
2165                                 btrfs_tree_unlock(mid);
2166                                 free_extent_buffer(mid);
2167                         } else {
2168                                 orig_slot -=
2169                                         btrfs_header_nritems(left);
2170                                 path->slots[level] = orig_slot;
2171                                 btrfs_tree_unlock(left);
2172                                 free_extent_buffer(left);
2173                         }
2174                         return 0;
2175                 }
2176                 btrfs_tree_unlock(left);
2177                 free_extent_buffer(left);
2178         }
2179         right = read_node_slot(fs_info, parent, pslot + 1);
2180         if (IS_ERR(right))
2181                 right = NULL;
2182
2183         /*
2184          * then try to empty the right most buffer into the middle
2185          */
2186         if (right) {
2187                 u32 right_nr;
2188
2189                 btrfs_tree_lock(right);
2190                 btrfs_set_lock_blocking(right);
2191
2192                 right_nr = btrfs_header_nritems(right);
2193                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2194                         wret = 1;
2195                 } else {
2196                         ret = btrfs_cow_block(trans, root, right,
2197                                               parent, pslot + 1,
2198                                               &right);
2199                         if (ret)
2200                                 wret = 1;
2201                         else {
2202                                 wret = balance_node_right(trans, fs_info,
2203                                                           right, mid);
2204                         }
2205                 }
2206                 if (wret < 0)
2207                         ret = wret;
2208                 if (wret == 0) {
2209                         struct btrfs_disk_key disk_key;
2210
2211                         btrfs_node_key(right, &disk_key, 0);
2212                         tree_mod_log_set_node_key(fs_info, parent,
2213                                                   pslot + 1, 0);
2214                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
2215                         btrfs_mark_buffer_dirty(parent);
2216
2217                         if (btrfs_header_nritems(mid) <= orig_slot) {
2218                                 path->nodes[level] = right;
2219                                 path->slots[level + 1] += 1;
2220                                 path->slots[level] = orig_slot -
2221                                         btrfs_header_nritems(mid);
2222                                 btrfs_tree_unlock(mid);
2223                                 free_extent_buffer(mid);
2224                         } else {
2225                                 btrfs_tree_unlock(right);
2226                                 free_extent_buffer(right);
2227                         }
2228                         return 0;
2229                 }
2230                 btrfs_tree_unlock(right);
2231                 free_extent_buffer(right);
2232         }
2233         return 1;
2234 }
2235
2236 /*
2237  * readahead one full node of leaves, finding things that are close
2238  * to the block in 'slot', and triggering ra on them.
2239  */
2240 static void reada_for_search(struct btrfs_fs_info *fs_info,
2241                              struct btrfs_path *path,
2242                              int level, int slot, u64 objectid)
2243 {
2244         struct extent_buffer *node;
2245         struct btrfs_disk_key disk_key;
2246         u32 nritems;
2247         u64 search;
2248         u64 target;
2249         u64 nread = 0;
2250         struct extent_buffer *eb;
2251         u32 nr;
2252         u32 blocksize;
2253         u32 nscan = 0;
2254
2255         if (level != 1)
2256                 return;
2257
2258         if (!path->nodes[level])
2259                 return;
2260
2261         node = path->nodes[level];
2262
2263         search = btrfs_node_blockptr(node, slot);
2264         blocksize = fs_info->nodesize;
2265         eb = find_extent_buffer(fs_info, search);
2266         if (eb) {
2267                 free_extent_buffer(eb);
2268                 return;
2269         }
2270
2271         target = search;
2272
2273         nritems = btrfs_header_nritems(node);
2274         nr = slot;
2275
2276         while (1) {
2277                 if (path->reada == READA_BACK) {
2278                         if (nr == 0)
2279                                 break;
2280                         nr--;
2281                 } else if (path->reada == READA_FORWARD) {
2282                         nr++;
2283                         if (nr >= nritems)
2284                                 break;
2285                 }
2286                 if (path->reada == READA_BACK && objectid) {
2287                         btrfs_node_key(node, &disk_key, nr);
2288                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
2289                                 break;
2290                 }
2291                 search = btrfs_node_blockptr(node, nr);
2292                 if ((search <= target && target - search <= 65536) ||
2293                     (search > target && search - target <= 65536)) {
2294                         readahead_tree_block(fs_info, search);
2295                         nread += blocksize;
2296                 }
2297                 nscan++;
2298                 if ((nread > 65536 || nscan > 32))
2299                         break;
2300         }
2301 }
2302
2303 static noinline void reada_for_balance(struct btrfs_fs_info *fs_info,
2304                                        struct btrfs_path *path, int level)
2305 {
2306         int slot;
2307         int nritems;
2308         struct extent_buffer *parent;
2309         struct extent_buffer *eb;
2310         u64 gen;
2311         u64 block1 = 0;
2312         u64 block2 = 0;
2313
2314         parent = path->nodes[level + 1];
2315         if (!parent)
2316                 return;
2317
2318         nritems = btrfs_header_nritems(parent);
2319         slot = path->slots[level + 1];
2320
2321         if (slot > 0) {
2322                 block1 = btrfs_node_blockptr(parent, slot - 1);
2323                 gen = btrfs_node_ptr_generation(parent, slot - 1);
2324                 eb = find_extent_buffer(fs_info, block1);
2325                 /*
2326                  * if we get -eagain from btrfs_buffer_uptodate, we
2327                  * don't want to return eagain here.  That will loop
2328                  * forever
2329                  */
2330                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2331                         block1 = 0;
2332                 free_extent_buffer(eb);
2333         }
2334         if (slot + 1 < nritems) {
2335                 block2 = btrfs_node_blockptr(parent, slot + 1);
2336                 gen = btrfs_node_ptr_generation(parent, slot + 1);
2337                 eb = find_extent_buffer(fs_info, block2);
2338                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2339                         block2 = 0;
2340                 free_extent_buffer(eb);
2341         }
2342
2343         if (block1)
2344                 readahead_tree_block(fs_info, block1);
2345         if (block2)
2346                 readahead_tree_block(fs_info, block2);
2347 }
2348
2349
2350 /*
2351  * when we walk down the tree, it is usually safe to unlock the higher layers
2352  * in the tree.  The exceptions are when our path goes through slot 0, because
2353  * operations on the tree might require changing key pointers higher up in the
2354  * tree.
2355  *
2356  * callers might also have set path->keep_locks, which tells this code to keep
2357  * the lock if the path points to the last slot in the block.  This is part of
2358  * walking through the tree, and selecting the next slot in the higher block.
2359  *
2360  * lowest_unlock sets the lowest level in the tree we're allowed to unlock.  so
2361  * if lowest_unlock is 1, level 0 won't be unlocked
2362  */
2363 static noinline void unlock_up(struct btrfs_path *path, int level,
2364                                int lowest_unlock, int min_write_lock_level,
2365                                int *write_lock_level)
2366 {
2367         int i;
2368         int skip_level = level;
2369         int no_skips = 0;
2370         struct extent_buffer *t;
2371
2372         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2373                 if (!path->nodes[i])
2374                         break;
2375                 if (!path->locks[i])
2376                         break;
2377                 if (!no_skips && path->slots[i] == 0) {
2378                         skip_level = i + 1;
2379                         continue;
2380                 }
2381                 if (!no_skips && path->keep_locks) {
2382                         u32 nritems;
2383                         t = path->nodes[i];
2384                         nritems = btrfs_header_nritems(t);
2385                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
2386                                 skip_level = i + 1;
2387                                 continue;
2388                         }
2389                 }
2390                 if (skip_level < i && i >= lowest_unlock)
2391                         no_skips = 1;
2392
2393                 t = path->nodes[i];
2394                 if (i >= lowest_unlock && i > skip_level && path->locks[i]) {
2395                         btrfs_tree_unlock_rw(t, path->locks[i]);
2396                         path->locks[i] = 0;
2397                         if (write_lock_level &&
2398                             i > min_write_lock_level &&
2399                             i <= *write_lock_level) {
2400                                 *write_lock_level = i - 1;
2401                         }
2402                 }
2403         }
2404 }
2405
2406 /*
2407  * This releases any locks held in the path starting at level and
2408  * going all the way up to the root.
2409  *
2410  * btrfs_search_slot will keep the lock held on higher nodes in a few
2411  * corner cases, such as COW of the block at slot zero in the node.  This
2412  * ignores those rules, and it should only be called when there are no
2413  * more updates to be done higher up in the tree.
2414  */
2415 noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
2416 {
2417         int i;
2418
2419         if (path->keep_locks)
2420                 return;
2421
2422         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2423                 if (!path->nodes[i])
2424                         continue;
2425                 if (!path->locks[i])
2426                         continue;
2427                 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
2428                 path->locks[i] = 0;
2429         }
2430 }
2431
2432 /*
2433  * helper function for btrfs_search_slot.  The goal is to find a block
2434  * in cache without setting the path to blocking.  If we find the block
2435  * we return zero and the path is unchanged.
2436  *
2437  * If we can't find the block, we set the path blocking and do some
2438  * reada.  -EAGAIN is returned and the search must be repeated.
2439  */
2440 static int
2441 read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
2442                       struct extent_buffer **eb_ret, int level, int slot,
2443                       const struct btrfs_key *key)
2444 {
2445         struct btrfs_fs_info *fs_info = root->fs_info;
2446         u64 blocknr;
2447         u64 gen;
2448         struct extent_buffer *b = *eb_ret;
2449         struct extent_buffer *tmp;
2450         int ret;
2451
2452         blocknr = btrfs_node_blockptr(b, slot);
2453         gen = btrfs_node_ptr_generation(b, slot);
2454
2455         tmp = find_extent_buffer(fs_info, blocknr);
2456         if (tmp) {
2457                 /* first we do an atomic uptodate check */
2458                 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
2459                         *eb_ret = tmp;
2460                         return 0;
2461                 }
2462
2463                 /* the pages were up to date, but we failed
2464                  * the generation number check.  Do a full
2465                  * read for the generation number that is correct.
2466                  * We must do this without dropping locks so
2467                  * we can trust our generation number
2468                  */
2469                 btrfs_set_path_blocking(p);
2470
2471                 /* now we're allowed to do a blocking uptodate check */
2472                 ret = btrfs_read_buffer(tmp, gen);
2473                 if (!ret) {
2474                         *eb_ret = tmp;
2475                         return 0;
2476                 }
2477                 free_extent_buffer(tmp);
2478                 btrfs_release_path(p);
2479                 return -EIO;
2480         }
2481
2482         /*
2483          * reduce lock contention at high levels
2484          * of the btree by dropping locks before
2485          * we read.  Don't release the lock on the current
2486          * level because we need to walk this node to figure
2487          * out which blocks to read.
2488          */
2489         btrfs_unlock_up_safe(p, level + 1);
2490         btrfs_set_path_blocking(p);
2491
2492         free_extent_buffer(tmp);
2493         if (p->reada != READA_NONE)
2494                 reada_for_search(fs_info, p, level, slot, key->objectid);
2495
2496         btrfs_release_path(p);
2497
2498         ret = -EAGAIN;
2499         tmp = read_tree_block(fs_info, blocknr, 0);
2500         if (!IS_ERR(tmp)) {
2501                 /*
2502                  * If the read above didn't mark this buffer up to date,
2503                  * it will never end up being up to date.  Set ret to EIO now
2504                  * and give up so that our caller doesn't loop forever
2505                  * on our EAGAINs.
2506                  */
2507                 if (!btrfs_buffer_uptodate(tmp, 0, 0))
2508                         ret = -EIO;
2509                 free_extent_buffer(tmp);
2510         } else {
2511                 ret = PTR_ERR(tmp);
2512         }
2513         return ret;
2514 }
2515
2516 /*
2517  * helper function for btrfs_search_slot.  This does all of the checks
2518  * for node-level blocks and does any balancing required based on
2519  * the ins_len.
2520  *
2521  * If no extra work was required, zero is returned.  If we had to
2522  * drop the path, -EAGAIN is returned and btrfs_search_slot must
2523  * start over
2524  */
2525 static int
2526 setup_nodes_for_search(struct btrfs_trans_handle *trans,
2527                        struct btrfs_root *root, struct btrfs_path *p,
2528                        struct extent_buffer *b, int level, int ins_len,
2529                        int *write_lock_level)
2530 {
2531         struct btrfs_fs_info *fs_info = root->fs_info;
2532         int ret;
2533
2534         if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
2535             BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3) {
2536                 int sret;
2537
2538                 if (*write_lock_level < level + 1) {
2539                         *write_lock_level = level + 1;
2540                         btrfs_release_path(p);
2541                         goto again;
2542                 }
2543
2544                 btrfs_set_path_blocking(p);
2545                 reada_for_balance(fs_info, p, level);
2546                 sret = split_node(trans, root, p, level);
2547                 btrfs_clear_path_blocking(p, NULL, 0);
2548
2549                 BUG_ON(sret > 0);
2550                 if (sret) {
2551                         ret = sret;
2552                         goto done;
2553                 }
2554                 b = p->nodes[level];
2555         } else if (ins_len < 0 && btrfs_header_nritems(b) <
2556                    BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 2) {
2557                 int sret;
2558
2559                 if (*write_lock_level < level + 1) {
2560                         *write_lock_level = level + 1;
2561                         btrfs_release_path(p);
2562                         goto again;
2563                 }
2564
2565                 btrfs_set_path_blocking(p);
2566                 reada_for_balance(fs_info, p, level);
2567                 sret = balance_level(trans, root, p, level);
2568                 btrfs_clear_path_blocking(p, NULL, 0);
2569
2570                 if (sret) {
2571                         ret = sret;
2572                         goto done;
2573                 }
2574                 b = p->nodes[level];
2575                 if (!b) {
2576                         btrfs_release_path(p);
2577                         goto again;
2578                 }
2579                 BUG_ON(btrfs_header_nritems(b) == 1);
2580         }
2581         return 0;
2582
2583 again:
2584         ret = -EAGAIN;
2585 done:
2586         return ret;
2587 }
2588
2589 static void key_search_validate(struct extent_buffer *b,
2590                                 const struct btrfs_key *key,
2591                                 int level)
2592 {
2593 #ifdef CONFIG_BTRFS_ASSERT
2594         struct btrfs_disk_key disk_key;
2595
2596         btrfs_cpu_key_to_disk(&disk_key, key);
2597
2598         if (level == 0)
2599                 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2600                     offsetof(struct btrfs_leaf, items[0].key),
2601                     sizeof(disk_key)));
2602         else
2603                 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2604                     offsetof(struct btrfs_node, ptrs[0].key),
2605                     sizeof(disk_key)));
2606 #endif
2607 }
2608
2609 static int key_search(struct extent_buffer *b, const struct btrfs_key *key,
2610                       int level, int *prev_cmp, int *slot)
2611 {
2612         if (*prev_cmp != 0) {
2613                 *prev_cmp = bin_search(b, key, level, slot);
2614                 return *prev_cmp;
2615         }
2616
2617         key_search_validate(b, key, level);
2618         *slot = 0;
2619
2620         return 0;
2621 }
2622
2623 int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
2624                 u64 iobjectid, u64 ioff, u8 key_type,
2625                 struct btrfs_key *found_key)
2626 {
2627         int ret;
2628         struct btrfs_key key;
2629         struct extent_buffer *eb;
2630
2631         ASSERT(path);
2632         ASSERT(found_key);
2633
2634         key.type = key_type;
2635         key.objectid = iobjectid;
2636         key.offset = ioff;
2637
2638         ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
2639         if (ret < 0)
2640                 return ret;
2641
2642         eb = path->nodes[0];
2643         if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
2644                 ret = btrfs_next_leaf(fs_root, path);
2645                 if (ret)
2646                         return ret;
2647                 eb = path->nodes[0];
2648         }
2649
2650         btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
2651         if (found_key->type != key.type ||
2652                         found_key->objectid != key.objectid)
2653                 return 1;
2654
2655         return 0;
2656 }
2657
2658 /*
2659  * look for key in the tree.  path is filled in with nodes along the way
2660  * if key is found, we return zero and you can find the item in the leaf
2661  * level of the path (level 0)
2662  *
2663  * If the key isn't found, the path points to the slot where it should
2664  * be inserted, and 1 is returned.  If there are other errors during the
2665  * search a negative error number is returned.
2666  *
2667  * if ins_len > 0, nodes and leaves will be split as we walk down the
2668  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
2669  * possible)
2670  */
2671 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2672                       const struct btrfs_key *key, struct btrfs_path *p,
2673                       int ins_len, int cow)
2674 {
2675         struct btrfs_fs_info *fs_info = root->fs_info;
2676         struct extent_buffer *b;
2677         int slot;
2678         int ret;
2679         int err;
2680         int level;
2681         int lowest_unlock = 1;
2682         int root_lock;
2683         /* everything at write_lock_level or lower must be write locked */
2684         int write_lock_level = 0;
2685         u8 lowest_level = 0;
2686         int min_write_lock_level;
2687         int prev_cmp;
2688
2689         lowest_level = p->lowest_level;
2690         WARN_ON(lowest_level && ins_len > 0);
2691         WARN_ON(p->nodes[0] != NULL);
2692         BUG_ON(!cow && ins_len);
2693
2694         if (ins_len < 0) {
2695                 lowest_unlock = 2;
2696
2697                 /* when we are removing items, we might have to go up to level
2698                  * two as we update tree pointers  Make sure we keep write
2699                  * for those levels as well
2700                  */
2701                 write_lock_level = 2;
2702         } else if (ins_len > 0) {
2703                 /*
2704                  * for inserting items, make sure we have a write lock on
2705                  * level 1 so we can update keys
2706                  */
2707                 write_lock_level = 1;
2708         }
2709
2710         if (!cow)
2711                 write_lock_level = -1;
2712
2713         if (cow && (p->keep_locks || p->lowest_level))
2714                 write_lock_level = BTRFS_MAX_LEVEL;
2715
2716         min_write_lock_level = write_lock_level;
2717
2718 again:
2719         prev_cmp = -1;
2720         /*
2721          * we try very hard to do read locks on the root
2722          */
2723         root_lock = BTRFS_READ_LOCK;
2724         level = 0;
2725         if (p->search_commit_root) {
2726                 /*
2727                  * the commit roots are read only
2728                  * so we always do read locks
2729                  */
2730                 if (p->need_commit_sem)
2731                         down_read(&fs_info->commit_root_sem);
2732                 b = root->commit_root;
2733                 extent_buffer_get(b);
2734                 level = btrfs_header_level(b);
2735                 if (p->need_commit_sem)
2736                         up_read(&fs_info->commit_root_sem);
2737                 if (!p->skip_locking)
2738                         btrfs_tree_read_lock(b);
2739         } else {
2740                 if (p->skip_locking) {
2741                         b = btrfs_root_node(root);
2742                         level = btrfs_header_level(b);
2743                 } else {
2744                         /* we don't know the level of the root node
2745                          * until we actually have it read locked
2746                          */
2747                         b = btrfs_read_lock_root_node(root);
2748                         level = btrfs_header_level(b);
2749                         if (level <= write_lock_level) {
2750                                 /* whoops, must trade for write lock */
2751                                 btrfs_tree_read_unlock(b);
2752                                 free_extent_buffer(b);
2753                                 b = btrfs_lock_root_node(root);
2754                                 root_lock = BTRFS_WRITE_LOCK;
2755
2756                                 /* the level might have changed, check again */
2757                                 level = btrfs_header_level(b);
2758                         }
2759                 }
2760         }
2761         p->nodes[level] = b;
2762         if (!p->skip_locking)
2763                 p->locks[level] = root_lock;
2764
2765         while (b) {
2766                 level = btrfs_header_level(b);
2767
2768                 /*
2769                  * setup the path here so we can release it under lock
2770                  * contention with the cow code
2771                  */
2772                 if (cow) {
2773                         /*
2774                          * if we don't really need to cow this block
2775                          * then we don't want to set the path blocking,
2776                          * so we test it here
2777                          */
2778                         if (!should_cow_block(trans, root, b)) {
2779                                 trans->dirty = true;
2780                                 goto cow_done;
2781                         }
2782
2783                         /*
2784                          * must have write locks on this node and the
2785                          * parent
2786                          */
2787                         if (level > write_lock_level ||
2788                             (level + 1 > write_lock_level &&
2789                             level + 1 < BTRFS_MAX_LEVEL &&
2790                             p->nodes[level + 1])) {
2791                                 write_lock_level = level + 1;
2792                                 btrfs_release_path(p);
2793                                 goto again;
2794                         }
2795
2796                         btrfs_set_path_blocking(p);
2797                         err = btrfs_cow_block(trans, root, b,
2798                                               p->nodes[level + 1],
2799                                               p->slots[level + 1], &b);
2800                         if (err) {
2801                                 ret = err;
2802                                 goto done;
2803                         }
2804                 }
2805 cow_done:
2806                 p->nodes[level] = b;
2807                 btrfs_clear_path_blocking(p, NULL, 0);
2808
2809                 /*
2810                  * we have a lock on b and as long as we aren't changing
2811                  * the tree, there is no way to for the items in b to change.
2812                  * It is safe to drop the lock on our parent before we
2813                  * go through the expensive btree search on b.
2814                  *
2815                  * If we're inserting or deleting (ins_len != 0), then we might
2816                  * be changing slot zero, which may require changing the parent.
2817                  * So, we can't drop the lock until after we know which slot
2818                  * we're operating on.
2819                  */
2820                 if (!ins_len && !p->keep_locks) {
2821                         int u = level + 1;
2822
2823                         if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
2824                                 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
2825                                 p->locks[u] = 0;
2826                         }
2827                 }
2828
2829                 ret = key_search(b, key, level, &prev_cmp, &slot);
2830                 if (ret < 0)
2831                         goto done;
2832
2833                 if (level != 0) {
2834                         int dec = 0;
2835                         if (ret && slot > 0) {
2836                                 dec = 1;
2837                                 slot -= 1;
2838                         }
2839                         p->slots[level] = slot;
2840                         err = setup_nodes_for_search(trans, root, p, b, level,
2841                                              ins_len, &write_lock_level);
2842                         if (err == -EAGAIN)
2843                                 goto again;
2844                         if (err) {
2845                                 ret = err;
2846                                 goto done;
2847                         }
2848                         b = p->nodes[level];
2849                         slot = p->slots[level];
2850
2851                         /*
2852                          * slot 0 is special, if we change the key
2853                          * we have to update the parent pointer
2854                          * which means we must have a write lock
2855                          * on the parent
2856                          */
2857                         if (slot == 0 && ins_len &&
2858                             write_lock_level < level + 1) {
2859                                 write_lock_level = level + 1;
2860                                 btrfs_release_path(p);
2861                                 goto again;
2862                         }
2863
2864                         unlock_up(p, level, lowest_unlock,
2865                                   min_write_lock_level, &write_lock_level);
2866
2867                         if (level == lowest_level) {
2868                                 if (dec)
2869                                         p->slots[level]++;
2870                                 goto done;
2871                         }
2872
2873                         err = read_block_for_search(root, p, &b, level,
2874                                                     slot, key);
2875                         if (err == -EAGAIN)
2876                                 goto again;
2877                         if (err) {
2878                                 ret = err;
2879                                 goto done;
2880                         }
2881
2882                         if (!p->skip_locking) {
2883                                 level = btrfs_header_level(b);
2884                                 if (level <= write_lock_level) {
2885                                         err = btrfs_try_tree_write_lock(b);
2886                                         if (!err) {
2887                                                 btrfs_set_path_blocking(p);
2888                                                 btrfs_tree_lock(b);
2889                                                 btrfs_clear_path_blocking(p, b,
2890                                                                   BTRFS_WRITE_LOCK);
2891                                         }
2892                                         p->locks[level] = BTRFS_WRITE_LOCK;
2893                                 } else {
2894                                         err = btrfs_tree_read_lock_atomic(b);
2895                                         if (!err) {
2896                                                 btrfs_set_path_blocking(p);
2897                                                 btrfs_tree_read_lock(b);
2898                                                 btrfs_clear_path_blocking(p, b,
2899                                                                   BTRFS_READ_LOCK);
2900                                         }
2901                                         p->locks[level] = BTRFS_READ_LOCK;
2902                                 }
2903                                 p->nodes[level] = b;
2904                         }
2905                 } else {
2906                         p->slots[level] = slot;
2907                         if (ins_len > 0 &&
2908                             btrfs_leaf_free_space(fs_info, b) < ins_len) {
2909                                 if (write_lock_level < 1) {
2910                                         write_lock_level = 1;
2911                                         btrfs_release_path(p);
2912                                         goto again;
2913                                 }
2914
2915                                 btrfs_set_path_blocking(p);
2916                                 err = split_leaf(trans, root, key,
2917                                                  p, ins_len, ret == 0);
2918                                 btrfs_clear_path_blocking(p, NULL, 0);
2919
2920                                 BUG_ON(err > 0);
2921                                 if (err) {
2922                                         ret = err;
2923                                         goto done;
2924                                 }
2925                         }
2926                         if (!p->search_for_split)
2927                                 unlock_up(p, level, lowest_unlock,
2928                                           min_write_lock_level, &write_lock_level);
2929                         goto done;
2930                 }
2931         }
2932         ret = 1;
2933 done:
2934         /*
2935          * we don't really know what they plan on doing with the path
2936          * from here on, so for now just mark it as blocking
2937          */
2938         if (!p->leave_spinning)
2939                 btrfs_set_path_blocking(p);
2940         if (ret < 0 && !p->skip_release_on_error)
2941                 btrfs_release_path(p);
2942         return ret;
2943 }
2944
2945 /*
2946  * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2947  * current state of the tree together with the operations recorded in the tree
2948  * modification log to search for the key in a previous version of this tree, as
2949  * denoted by the time_seq parameter.
2950  *
2951  * Naturally, there is no support for insert, delete or cow operations.
2952  *
2953  * The resulting path and return value will be set up as if we called
2954  * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2955  */
2956 int btrfs_search_old_slot(struct btrfs_root *root, const struct btrfs_key *key,
2957                           struct btrfs_path *p, u64 time_seq)
2958 {
2959         struct btrfs_fs_info *fs_info = root->fs_info;
2960         struct extent_buffer *b;
2961         int slot;
2962         int ret;
2963         int err;
2964         int level;
2965         int lowest_unlock = 1;
2966         u8 lowest_level = 0;
2967         int prev_cmp = -1;
2968
2969         lowest_level = p->lowest_level;
2970         WARN_ON(p->nodes[0] != NULL);
2971
2972         if (p->search_commit_root) {
2973                 BUG_ON(time_seq);
2974                 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2975         }
2976
2977 again:
2978         b = get_old_root(root, time_seq);
2979         level = btrfs_header_level(b);
2980         p->locks[level] = BTRFS_READ_LOCK;
2981
2982         while (b) {
2983                 level = btrfs_header_level(b);
2984                 p->nodes[level] = b;
2985                 btrfs_clear_path_blocking(p, NULL, 0);
2986
2987                 /*
2988                  * we have a lock on b and as long as we aren't changing
2989                  * the tree, there is no way to for the items in b to change.
2990                  * It is safe to drop the lock on our parent before we
2991                  * go through the expensive btree search on b.
2992                  */
2993                 btrfs_unlock_up_safe(p, level + 1);
2994
2995                 /*
2996                  * Since we can unwind ebs we want to do a real search every
2997                  * time.
2998                  */
2999                 prev_cmp = -1;
3000                 ret = key_search(b, key, level, &prev_cmp, &slot);
3001
3002                 if (level != 0) {
3003                         int dec = 0;
3004                         if (ret && slot > 0) {
3005                                 dec = 1;
3006                                 slot -= 1;
3007                         }
3008                         p->slots[level] = slot;
3009                         unlock_up(p, level, lowest_unlock, 0, NULL);
3010
3011                         if (level == lowest_level) {
3012                                 if (dec)
3013                                         p->slots[level]++;
3014                                 goto done;
3015                         }
3016
3017                         err = read_block_for_search(root, p, &b, level,
3018                                                     slot, key);
3019                         if (err == -EAGAIN)
3020                                 goto again;
3021                         if (err) {
3022                                 ret = err;
3023                                 goto done;
3024                         }
3025
3026                         level = btrfs_header_level(b);
3027                         err = btrfs_tree_read_lock_atomic(b);
3028                         if (!err) {
3029                                 btrfs_set_path_blocking(p);
3030                                 btrfs_tree_read_lock(b);
3031                                 btrfs_clear_path_blocking(p, b,
3032                                                           BTRFS_READ_LOCK);
3033                         }
3034                         b = tree_mod_log_rewind(fs_info, p, b, time_seq);
3035                         if (!b) {
3036                                 ret = -ENOMEM;
3037                                 goto done;
3038                         }
3039                         p->locks[level] = BTRFS_READ_LOCK;
3040                         p->nodes[level] = b;
3041                 } else {
3042                         p->slots[level] = slot;
3043                         unlock_up(p, level, lowest_unlock, 0, NULL);
3044                         goto done;
3045                 }
3046         }
3047         ret = 1;
3048 done:
3049         if (!p->leave_spinning)
3050                 btrfs_set_path_blocking(p);
3051         if (ret < 0)
3052                 btrfs_release_path(p);
3053
3054         return ret;
3055 }
3056
3057 /*
3058  * helper to use instead of search slot if no exact match is needed but
3059  * instead the next or previous item should be returned.
3060  * When find_higher is true, the next higher item is returned, the next lower
3061  * otherwise.
3062  * When return_any and find_higher are both true, and no higher item is found,
3063  * return the next lower instead.
3064  * When return_any is true and find_higher is false, and no lower item is found,
3065  * return the next higher instead.
3066  * It returns 0 if any item is found, 1 if none is found (tree empty), and
3067  * < 0 on error
3068  */
3069 int btrfs_search_slot_for_read(struct btrfs_root *root,
3070                                const struct btrfs_key *key,
3071                                struct btrfs_path *p, int find_higher,
3072                                int return_any)
3073 {
3074         int ret;
3075         struct extent_buffer *leaf;
3076
3077 again:
3078         ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
3079         if (ret <= 0)
3080                 return ret;
3081         /*
3082          * a return value of 1 means the path is at the position where the
3083          * item should be inserted. Normally this is the next bigger item,
3084          * but in case the previous item is the last in a leaf, path points
3085          * to the first free slot in the previous leaf, i.e. at an invalid
3086          * item.
3087          */
3088         leaf = p->nodes[0];
3089
3090         if (find_higher) {
3091                 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
3092                         ret = btrfs_next_leaf(root, p);
3093                         if (ret <= 0)
3094                                 return ret;
3095                         if (!return_any)
3096                                 return 1;
3097                         /*
3098                          * no higher item found, return the next
3099                          * lower instead
3100                          */
3101                         return_any = 0;
3102                         find_higher = 0;
3103                         btrfs_release_path(p);
3104                         goto again;
3105                 }
3106         } else {
3107                 if (p->slots[0] == 0) {
3108                         ret = btrfs_prev_leaf(root, p);
3109                         if (ret < 0)
3110                                 return ret;
3111                         if (!ret) {
3112                                 leaf = p->nodes[0];
3113                                 if (p->slots[0] == btrfs_header_nritems(leaf))
3114                                         p->slots[0]--;
3115                                 return 0;
3116                         }
3117                         if (!return_any)
3118                                 return 1;
3119                         /*
3120                          * no lower item found, return the next
3121                          * higher instead
3122                          */
3123                         return_any = 0;
3124                         find_higher = 1;
3125                         btrfs_release_path(p);
3126                         goto again;
3127                 } else {
3128                         --p->slots[0];
3129                 }
3130         }
3131         return 0;
3132 }
3133
3134 /*
3135  * adjust the pointers going up the tree, starting at level
3136  * making sure the right key of each node is points to 'key'.
3137  * This is used after shifting pointers to the left, so it stops
3138  * fixing up pointers when a given leaf/node is not in slot 0 of the
3139  * higher levels
3140  *
3141  */
3142 static void fixup_low_keys(struct btrfs_fs_info *fs_info,
3143                            struct btrfs_path *path,
3144                            struct btrfs_disk_key *key, int level)
3145 {
3146         int i;
3147         struct extent_buffer *t;
3148
3149         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
3150                 int tslot = path->slots[i];
3151                 if (!path->nodes[i])
3152                         break;
3153                 t = path->nodes[i];
3154                 tree_mod_log_set_node_key(fs_info, t, tslot, 1);
3155                 btrfs_set_node_key(t, key, tslot);
3156                 btrfs_mark_buffer_dirty(path->nodes[i]);
3157                 if (tslot != 0)
3158                         break;
3159         }
3160 }
3161
3162 /*
3163  * update item key.
3164  *
3165  * This function isn't completely safe. It's the caller's responsibility
3166  * that the new key won't break the order
3167  */
3168 void btrfs_set_item_key_safe(struct btrfs_fs_info *fs_info,
3169                              struct btrfs_path *path,
3170                              const struct btrfs_key *new_key)
3171 {
3172         struct btrfs_disk_key disk_key;
3173         struct extent_buffer *eb;
3174         int slot;
3175
3176         eb = path->nodes[0];
3177         slot = path->slots[0];
3178         if (slot > 0) {
3179                 btrfs_item_key(eb, &disk_key, slot - 1);
3180                 BUG_ON(comp_keys(&disk_key, new_key) >= 0);
3181         }
3182         if (slot < btrfs_header_nritems(eb) - 1) {
3183                 btrfs_item_key(eb, &disk_key, slot + 1);
3184                 BUG_ON(comp_keys(&disk_key, new_key) <= 0);
3185         }
3186
3187         btrfs_cpu_key_to_disk(&disk_key, new_key);
3188         btrfs_set_item_key(eb, &disk_key, slot);
3189         btrfs_mark_buffer_dirty(eb);
3190         if (slot == 0)
3191                 fixup_low_keys(fs_info, path, &disk_key, 1);
3192 }
3193
3194 /*
3195  * try to push data from one node into the next node left in the
3196  * tree.
3197  *
3198  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
3199  * error, and > 0 if there was no room in the left hand block.
3200  */
3201 static int push_node_left(struct btrfs_trans_handle *trans,
3202                           struct btrfs_fs_info *fs_info,
3203                           struct extent_buffer *dst,
3204                           struct extent_buffer *src, int empty)
3205 {
3206         int push_items = 0;
3207         int src_nritems;
3208         int dst_nritems;
3209         int ret = 0;
3210
3211         src_nritems = btrfs_header_nritems(src);
3212         dst_nritems = btrfs_header_nritems(dst);
3213         push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3214         WARN_ON(btrfs_header_generation(src) != trans->transid);
3215         WARN_ON(btrfs_header_generation(dst) != trans->transid);
3216
3217         if (!empty && src_nritems <= 8)
3218                 return 1;
3219
3220         if (push_items <= 0)
3221                 return 1;
3222
3223         if (empty) {
3224                 push_items = min(src_nritems, push_items);
3225                 if (push_items < src_nritems) {
3226                         /* leave at least 8 pointers in the node if
3227                          * we aren't going to empty it
3228                          */
3229                         if (src_nritems - push_items < 8) {
3230                                 if (push_items <= 8)
3231                                         return 1;
3232                                 push_items -= 8;
3233                         }
3234                 }
3235         } else
3236                 push_items = min(src_nritems - 8, push_items);
3237
3238         ret = tree_mod_log_eb_copy(fs_info, dst, src, dst_nritems, 0,
3239                                    push_items);
3240         if (ret) {
3241                 btrfs_abort_transaction(trans, ret);
3242                 return ret;
3243         }
3244         copy_extent_buffer(dst, src,
3245                            btrfs_node_key_ptr_offset(dst_nritems),
3246                            btrfs_node_key_ptr_offset(0),
3247                            push_items * sizeof(struct btrfs_key_ptr));
3248
3249         if (push_items < src_nritems) {
3250                 /*
3251                  * don't call tree_mod_log_eb_move here, key removal was already
3252                  * fully logged by tree_mod_log_eb_copy above.
3253                  */
3254                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
3255                                       btrfs_node_key_ptr_offset(push_items),
3256                                       (src_nritems - push_items) *
3257                                       sizeof(struct btrfs_key_ptr));
3258         }
3259         btrfs_set_header_nritems(src, src_nritems - push_items);
3260         btrfs_set_header_nritems(dst, dst_nritems + push_items);
3261         btrfs_mark_buffer_dirty(src);
3262         btrfs_mark_buffer_dirty(dst);
3263
3264         return ret;
3265 }
3266
3267 /*
3268  * try to push data from one node into the next node right in the
3269  * tree.
3270  *
3271  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
3272  * error, and > 0 if there was no room in the right hand block.
3273  *
3274  * this will  only push up to 1/2 the contents of the left node over
3275  */
3276 static int balance_node_right(struct btrfs_trans_handle *trans,
3277                               struct btrfs_fs_info *fs_info,
3278                               struct extent_buffer *dst,
3279                               struct extent_buffer *src)
3280 {
3281         int push_items = 0;
3282         int max_push;
3283         int src_nritems;
3284         int dst_nritems;
3285         int ret = 0;
3286
3287         WARN_ON(btrfs_header_generation(src) != trans->transid);
3288         WARN_ON(btrfs_header_generation(dst) != trans->transid);
3289
3290         src_nritems = btrfs_header_nritems(src);
3291         dst_nritems = btrfs_header_nritems(dst);
3292         push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3293         if (push_items <= 0)
3294                 return 1;
3295
3296         if (src_nritems < 4)
3297                 return 1;
3298
3299         max_push = src_nritems / 2 + 1;
3300         /* don't try to empty the node */
3301         if (max_push >= src_nritems)
3302                 return 1;
3303
3304         if (max_push < push_items)
3305                 push_items = max_push;
3306
3307         tree_mod_log_eb_move(fs_info, dst, push_items, 0, dst_nritems);
3308         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
3309                                       btrfs_node_key_ptr_offset(0),
3310                                       (dst_nritems) *
3311                                       sizeof(struct btrfs_key_ptr));
3312
3313         ret = tree_mod_log_eb_copy(fs_info, dst, src, 0,
3314                                    src_nritems - push_items, push_items);
3315         if (ret) {
3316                 btrfs_abort_transaction(trans, ret);
3317                 return ret;
3318         }
3319         copy_extent_buffer(dst, src,
3320                            btrfs_node_key_ptr_offset(0),
3321                            btrfs_node_key_ptr_offset(src_nritems - push_items),
3322                            push_items * sizeof(struct btrfs_key_ptr));
3323
3324         btrfs_set_header_nritems(src, src_nritems - push_items);
3325         btrfs_set_header_nritems(dst, dst_nritems + push_items);
3326
3327         btrfs_mark_buffer_dirty(src);
3328         btrfs_mark_buffer_dirty(dst);
3329
3330         return ret;
3331 }
3332
3333 /*
3334  * helper function to insert a new root level in the tree.
3335  * A new node is allocated, and a single item is inserted to
3336  * point to the existing root
3337  *
3338  * returns zero on success or < 0 on failure.
3339  */
3340 static noinline int insert_new_root(struct btrfs_trans_handle *trans,
3341                            struct btrfs_root *root,
3342                            struct btrfs_path *path, int level)
3343 {
3344         struct btrfs_fs_info *fs_info = root->fs_info;
3345         u64 lower_gen;
3346         struct extent_buffer *lower;
3347         struct extent_buffer *c;
3348         struct extent_buffer *old;
3349         struct btrfs_disk_key lower_key;
3350
3351         BUG_ON(path->nodes[level]);
3352         BUG_ON(path->nodes[level-1] != root->node);
3353
3354         lower = path->nodes[level-1];
3355         if (level == 1)
3356                 btrfs_item_key(lower, &lower_key, 0);
3357         else
3358                 btrfs_node_key(lower, &lower_key, 0);
3359
3360         c = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
3361                                    &lower_key, level, root->node->start, 0);
3362         if (IS_ERR(c))
3363                 return PTR_ERR(c);
3364
3365         root_add_used(root, fs_info->nodesize);
3366
3367         memzero_extent_buffer(c, 0, sizeof(struct btrfs_header));
3368         btrfs_set_header_nritems(c, 1);
3369         btrfs_set_header_level(c, level);
3370         btrfs_set_header_bytenr(c, c->start);
3371         btrfs_set_header_generation(c, trans->transid);
3372         btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
3373         btrfs_set_header_owner(c, root->root_key.objectid);
3374
3375         write_extent_buffer_fsid(c, fs_info->fsid);
3376         write_extent_buffer_chunk_tree_uuid(c, fs_info->chunk_tree_uuid);
3377
3378         btrfs_set_node_key(c, &lower_key, 0);
3379         btrfs_set_node_blockptr(c, 0, lower->start);
3380         lower_gen = btrfs_header_generation(lower);
3381         WARN_ON(lower_gen != trans->transid);
3382
3383         btrfs_set_node_ptr_generation(c, 0, lower_gen);
3384
3385         btrfs_mark_buffer_dirty(c);
3386
3387         old = root->node;
3388         tree_mod_log_set_root_pointer(root, c, 0);
3389         rcu_assign_pointer(root->node, c);
3390
3391         /* the super has an extra ref to root->node */
3392         free_extent_buffer(old);
3393
3394         add_root_to_dirty_list(root);
3395         extent_buffer_get(c);
3396         path->nodes[level] = c;
3397         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
3398         path->slots[level] = 0;
3399         return 0;
3400 }
3401
3402 /*
3403  * worker function to insert a single pointer in a node.
3404  * the node should have enough room for the pointer already
3405  *
3406  * slot and level indicate where you want the key to go, and
3407  * blocknr is the block the key points to.
3408  */
3409 static void insert_ptr(struct btrfs_trans_handle *trans,
3410                        struct btrfs_fs_info *fs_info, struct btrfs_path *path,
3411                        struct btrfs_disk_key *key, u64 bytenr,
3412                        int slot, int level)
3413 {
3414         struct extent_buffer *lower;
3415         int nritems;
3416         int ret;
3417
3418         BUG_ON(!path->nodes[level]);
3419         btrfs_assert_tree_locked(path->nodes[level]);
3420         lower = path->nodes[level];
3421         nritems = btrfs_header_nritems(lower);
3422         BUG_ON(slot > nritems);
3423         BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(fs_info));
3424         if (slot != nritems) {
3425                 if (level)
3426                         tree_mod_log_eb_move(fs_info, lower, slot + 1,
3427                                              slot, nritems - slot);
3428                 memmove_extent_buffer(lower,
3429                               btrfs_node_key_ptr_offset(slot + 1),
3430                               btrfs_node_key_ptr_offset(slot),
3431                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
3432         }
3433         if (level) {
3434                 ret = tree_mod_log_insert_key(fs_info, lower, slot,
3435                                               MOD_LOG_KEY_ADD, GFP_NOFS);
3436                 BUG_ON(ret < 0);
3437         }
3438         btrfs_set_node_key(lower, key, slot);
3439         btrfs_set_node_blockptr(lower, slot, bytenr);
3440         WARN_ON(trans->transid == 0);
3441         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
3442         btrfs_set_header_nritems(lower, nritems + 1);
3443         btrfs_mark_buffer_dirty(lower);
3444 }
3445
3446 /*
3447  * split the node at the specified level in path in two.
3448  * The path is corrected to point to the appropriate node after the split
3449  *
3450  * Before splitting this tries to make some room in the node by pushing
3451  * left and right, if either one works, it returns right away.
3452  *
3453  * returns 0 on success and < 0 on failure
3454  */
3455 static noinline int split_node(struct btrfs_trans_handle *trans,
3456                                struct btrfs_root *root,
3457                                struct btrfs_path *path, int level)
3458 {
3459         struct btrfs_fs_info *fs_info = root->fs_info;
3460         struct extent_buffer *c;
3461         struct extent_buffer *split;
3462         struct btrfs_disk_key disk_key;
3463         int mid;
3464         int ret;
3465         u32 c_nritems;
3466
3467         c = path->nodes[level];
3468         WARN_ON(btrfs_header_generation(c) != trans->transid);
3469         if (c == root->node) {
3470                 /*
3471                  * trying to split the root, lets make a new one
3472                  *
3473                  * tree mod log: We don't log_removal old root in
3474                  * insert_new_root, because that root buffer will be kept as a
3475                  * normal node. We are going to log removal of half of the
3476                  * elements below with tree_mod_log_eb_copy. We're holding a
3477                  * tree lock on the buffer, which is why we cannot race with
3478                  * other tree_mod_log users.
3479                  */
3480                 ret = insert_new_root(trans, root, path, level + 1);
3481                 if (ret)
3482                         return ret;
3483         } else {
3484                 ret = push_nodes_for_insert(trans, root, path, level);
3485                 c = path->nodes[level];
3486                 if (!ret && btrfs_header_nritems(c) <
3487                     BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3)
3488                         return 0;
3489                 if (ret < 0)
3490                         return ret;
3491         }
3492
3493         c_nritems = btrfs_header_nritems(c);
3494         mid = (c_nritems + 1) / 2;
3495         btrfs_node_key(c, &disk_key, mid);
3496
3497         split = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
3498                         &disk_key, level, c->start, 0);
3499         if (IS_ERR(split))
3500                 return PTR_ERR(split);
3501
3502         root_add_used(root, fs_info->nodesize);
3503
3504         memzero_extent_buffer(split, 0, sizeof(struct btrfs_header));
3505         btrfs_set_header_level(split, btrfs_header_level(c));
3506         btrfs_set_header_bytenr(split, split->start);
3507         btrfs_set_header_generation(split, trans->transid);
3508         btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
3509         btrfs_set_header_owner(split, root->root_key.objectid);
3510         write_extent_buffer_fsid(split, fs_info->fsid);
3511         write_extent_buffer_chunk_tree_uuid(split, fs_info->chunk_tree_uuid);
3512
3513         ret = tree_mod_log_eb_copy(fs_info, split, c, 0, mid, c_nritems - mid);
3514         if (ret) {
3515                 btrfs_abort_transaction(trans, ret);
3516                 return ret;
3517         }
3518         copy_extent_buffer(split, c,
3519                            btrfs_node_key_ptr_offset(0),
3520                            btrfs_node_key_ptr_offset(mid),
3521                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
3522         btrfs_set_header_nritems(split, c_nritems - mid);
3523         btrfs_set_header_nritems(c, mid);
3524         ret = 0;
3525
3526         btrfs_mark_buffer_dirty(c);
3527         btrfs_mark_buffer_dirty(split);
3528
3529         insert_ptr(trans, fs_info, path, &disk_key, split->start,
3530                    path->slots[level + 1] + 1, level + 1);
3531
3532         if (path->slots[level] >= mid) {
3533                 path->slots[level] -= mid;
3534                 btrfs_tree_unlock(c);
3535                 free_extent_buffer(c);
3536                 path->nodes[level] = split;
3537                 path->slots[level + 1] += 1;
3538         } else {
3539                 btrfs_tree_unlock(split);
3540                 free_extent_buffer(split);
3541         }
3542         return ret;
3543 }
3544
3545 /*
3546  * how many bytes are required to store the items in a leaf.  start
3547  * and nr indicate which items in the leaf to check.  This totals up the
3548  * space used both by the item structs and the item data
3549  */
3550 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
3551 {
3552         struct btrfs_item *start_item;
3553         struct btrfs_item *end_item;
3554         struct btrfs_map_token token;
3555         int data_len;
3556         int nritems = btrfs_header_nritems(l);
3557         int end = min(nritems, start + nr) - 1;
3558
3559         if (!nr)
3560                 return 0;
3561         btrfs_init_map_token(&token);
3562         start_item = btrfs_item_nr(start);
3563         end_item = btrfs_item_nr(end);
3564         data_len = btrfs_token_item_offset(l, start_item, &token) +
3565                 btrfs_token_item_size(l, start_item, &token);
3566         data_len = data_len - btrfs_token_item_offset(l, end_item, &token);
3567         data_len += sizeof(struct btrfs_item) * nr;
3568         WARN_ON(data_len < 0);
3569         return data_len;
3570 }
3571
3572 /*
3573  * The space between the end of the leaf items and
3574  * the start of the leaf data.  IOW, how much room
3575  * the leaf has left for both items and data
3576  */
3577 noinline int btrfs_leaf_free_space(struct btrfs_fs_info *fs_info,
3578                                    struct extent_buffer *leaf)
3579 {
3580         int nritems = btrfs_header_nritems(leaf);
3581         int ret;
3582
3583         ret = BTRFS_LEAF_DATA_SIZE(fs_info) - leaf_space_used(leaf, 0, nritems);
3584         if (ret < 0) {
3585                 btrfs_crit(fs_info,
3586                            "leaf free space ret %d, leaf data size %lu, used %d nritems %d",
3587                            ret,
3588                            (unsigned long) BTRFS_LEAF_DATA_SIZE(fs_info),
3589                            leaf_space_used(leaf, 0, nritems), nritems);
3590         }
3591         return ret;
3592 }
3593
3594 /*
3595  * min slot controls the lowest index we're willing to push to the
3596  * right.  We'll push up to and including min_slot, but no lower
3597  */
3598 static noinline int __push_leaf_right(struct btrfs_fs_info *fs_info,
3599                                       struct btrfs_path *path,
3600                                       int data_size, int empty,
3601                                       struct extent_buffer *right,
3602                                       int free_space, u32 left_nritems,
3603                                       u32 min_slot)
3604 {
3605         struct extent_buffer *left = path->nodes[0];
3606         struct extent_buffer *upper = path->nodes[1];
3607         struct btrfs_map_token token;
3608         struct btrfs_disk_key disk_key;
3609         int slot;
3610         u32 i;
3611         int push_space = 0;
3612         int push_items = 0;
3613         struct btrfs_item *item;
3614         u32 nr;
3615         u32 right_nritems;
3616         u32 data_end;
3617         u32 this_item_size;
3618
3619         btrfs_init_map_token(&token);
3620
3621         if (empty)
3622                 nr = 0;
3623         else
3624                 nr = max_t(u32, 1, min_slot);
3625
3626         if (path->slots[0] >= left_nritems)
3627                 push_space += data_size;
3628
3629         slot = path->slots[1];
3630         i = left_nritems - 1;
3631         while (i >= nr) {
3632                 item = btrfs_item_nr(i);
3633
3634                 if (!empty && push_items > 0) {
3635                         if (path->slots[0] > i)
3636                                 break;
3637                         if (path->slots[0] == i) {
3638                                 int space = btrfs_leaf_free_space(fs_info, left);
3639                                 if (space + push_space * 2 > free_space)
3640                                         break;
3641                         }
3642                 }
3643
3644                 if (path->slots[0] == i)
3645                         push_space += data_size;
3646
3647                 this_item_size = btrfs_item_size(left, item);
3648                 if (this_item_size + sizeof(*item) + push_space > free_space)
3649                         break;
3650
3651                 push_items++;
3652                 push_space += this_item_size + sizeof(*item);
3653                 if (i == 0)
3654                         break;
3655                 i--;
3656         }
3657
3658         if (push_items == 0)
3659                 goto out_unlock;
3660
3661         WARN_ON(!empty && push_items == left_nritems);
3662
3663         /* push left to right */
3664         right_nritems = btrfs_header_nritems(right);
3665
3666         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
3667         push_space -= leaf_data_end(fs_info, left);
3668
3669         /* make room in the right data area */
3670         data_end = leaf_data_end(fs_info, right);
3671         memmove_extent_buffer(right,
3672                               btrfs_leaf_data(right) + data_end - push_space,
3673                               btrfs_leaf_data(right) + data_end,
3674                               BTRFS_LEAF_DATA_SIZE(fs_info) - data_end);
3675
3676         /* copy from the left data area */
3677         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
3678                      BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3679                      btrfs_leaf_data(left) + leaf_data_end(fs_info, left),
3680                      push_space);
3681
3682         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
3683                               btrfs_item_nr_offset(0),
3684                               right_nritems * sizeof(struct btrfs_item));
3685
3686         /* copy the items from left to right */
3687         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
3688                    btrfs_item_nr_offset(left_nritems - push_items),
3689                    push_items * sizeof(struct btrfs_item));
3690
3691         /* update the item pointers */
3692         right_nritems += push_items;
3693         btrfs_set_header_nritems(right, right_nritems);
3694         push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3695         for (i = 0; i < right_nritems; i++) {
3696                 item = btrfs_item_nr(i);
3697                 push_space -= btrfs_token_item_size(right, item, &token);
3698                 btrfs_set_token_item_offset(right, item, push_space, &token);
3699         }
3700
3701         left_nritems -= push_items;
3702         btrfs_set_header_nritems(left, left_nritems);
3703
3704         if (left_nritems)
3705                 btrfs_mark_buffer_dirty(left);
3706         else
3707                 clean_tree_block(fs_info, left);
3708
3709         btrfs_mark_buffer_dirty(right);
3710
3711         btrfs_item_key(right, &disk_key, 0);
3712         btrfs_set_node_key(upper, &disk_key, slot + 1);
3713         btrfs_mark_buffer_dirty(upper);
3714
3715         /* then fixup the leaf pointer in the path */
3716         if (path->slots[0] >= left_nritems) {
3717                 path->slots[0] -= left_nritems;
3718                 if (btrfs_header_nritems(path->nodes[0]) == 0)
3719                         clean_tree_block(fs_info, path->nodes[0]);
3720                 btrfs_tree_unlock(path->nodes[0]);
3721                 free_extent_buffer(path->nodes[0]);
3722                 path->nodes[0] = right;
3723                 path->slots[1] += 1;
3724         } else {
3725                 btrfs_tree_unlock(right);
3726                 free_extent_buffer(right);
3727         }
3728         return 0;
3729
3730 out_unlock:
3731         btrfs_tree_unlock(right);
3732         free_extent_buffer(right);
3733         return 1;
3734 }
3735
3736 /*
3737  * push some data in the path leaf to the right, trying to free up at
3738  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
3739  *
3740  * returns 1 if the push failed because the other node didn't have enough
3741  * room, 0 if everything worked out and < 0 if there were major errors.
3742  *
3743  * this will push starting from min_slot to the end of the leaf.  It won't
3744  * push any slot lower than min_slot
3745  */
3746 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
3747                            *root, struct btrfs_path *path,
3748                            int min_data_size, int data_size,
3749                            int empty, u32 min_slot)
3750 {
3751         struct btrfs_fs_info *fs_info = root->fs_info;
3752         struct extent_buffer *left = path->nodes[0];
3753         struct extent_buffer *right;
3754         struct extent_buffer *upper;
3755         int slot;
3756         int free_space;
3757         u32 left_nritems;
3758         int ret;
3759
3760         if (!path->nodes[1])
3761                 return 1;
3762
3763         slot = path->slots[1];
3764         upper = path->nodes[1];
3765         if (slot >= btrfs_header_nritems(upper) - 1)
3766                 return 1;
3767
3768         btrfs_assert_tree_locked(path->nodes[1]);
3769
3770         right = read_node_slot(fs_info, upper, slot + 1);
3771         /*
3772          * slot + 1 is not valid or we fail to read the right node,
3773          * no big deal, just return.
3774          */
3775         if (IS_ERR(right))
3776                 return 1;
3777
3778         btrfs_tree_lock(right);
3779         btrfs_set_lock_blocking(right);
3780
3781         free_space = btrfs_leaf_free_space(fs_info, right);
3782         if (free_space < data_size)
3783                 goto out_unlock;
3784
3785         /* cow and double check */
3786         ret = btrfs_cow_block(trans, root, right, upper,
3787                               slot + 1, &right);
3788         if (ret)
3789                 goto out_unlock;
3790
3791         free_space = btrfs_leaf_free_space(fs_info, right);
3792         if (free_space < data_size)
3793                 goto out_unlock;
3794
3795         left_nritems = btrfs_header_nritems(left);
3796         if (left_nritems == 0)
3797                 goto out_unlock;
3798
3799         if (path->slots[0] == left_nritems && !empty) {
3800                 /* Key greater than all keys in the leaf, right neighbor has
3801                  * enough room for it and we're not emptying our leaf to delete
3802                  * it, therefore use right neighbor to insert the new item and
3803                  * no need to touch/dirty our left leaft. */
3804                 btrfs_tree_unlock(left);
3805                 free_extent_buffer(left);
3806                 path->nodes[0] = right;
3807                 path->slots[0] = 0;
3808                 path->slots[1]++;
3809                 return 0;
3810         }
3811
3812         return __push_leaf_right(fs_info, path, min_data_size, empty,
3813                                 right, free_space, left_nritems, min_slot);
3814 out_unlock:
3815         btrfs_tree_unlock(right);
3816         free_extent_buffer(right);
3817         return 1;
3818 }
3819
3820 /*
3821  * push some data in the path leaf to the left, trying to free up at
3822  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
3823  *
3824  * max_slot can put a limit on how far into the leaf we'll push items.  The
3825  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us do all the
3826  * items
3827  */
3828 static noinline int __push_leaf_left(struct btrfs_fs_info *fs_info,
3829                                      struct btrfs_path *path, int data_size,
3830                                      int empty, struct extent_buffer *left,
3831                                      int free_space, u32 right_nritems,
3832                                      u32 max_slot)
3833 {
3834         struct btrfs_disk_key disk_key;
3835         struct extent_buffer *right = path->nodes[0];
3836         int i;
3837         int push_space = 0;
3838         int push_items = 0;
3839         struct btrfs_item *item;
3840         u32 old_left_nritems;
3841         u32 nr;
3842         int ret = 0;
3843         u32 this_item_size;
3844         u32 old_left_item_size;
3845         struct btrfs_map_token token;
3846
3847         btrfs_init_map_token(&token);
3848
3849         if (empty)
3850                 nr = min(right_nritems, max_slot);
3851         else
3852                 nr = min(right_nritems - 1, max_slot);
3853
3854         for (i = 0; i < nr; i++) {
3855                 item = btrfs_item_nr(i);
3856
3857                 if (!empty && push_items > 0) {
3858                         if (path->slots[0] < i)
3859                                 break;
3860                         if (path->slots[0] == i) {
3861                                 int space = btrfs_leaf_free_space(fs_info, right);
3862                                 if (space + push_space * 2 > free_space)
3863                                         break;
3864                         }
3865                 }
3866
3867                 if (path->slots[0] == i)
3868                         push_space += data_size;
3869
3870                 this_item_size = btrfs_item_size(right, item);
3871                 if (this_item_size + sizeof(*item) + push_space > free_space)
3872                         break;
3873
3874                 push_items++;
3875                 push_space += this_item_size + sizeof(*item);
3876         }
3877
3878         if (push_items == 0) {
3879                 ret = 1;
3880                 goto out;
3881         }
3882         WARN_ON(!empty && push_items == btrfs_header_nritems(right));
3883
3884         /* push data from right to left */
3885         copy_extent_buffer(left, right,
3886                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
3887                            btrfs_item_nr_offset(0),
3888                            push_items * sizeof(struct btrfs_item));
3889
3890         push_space = BTRFS_LEAF_DATA_SIZE(fs_info) -
3891                      btrfs_item_offset_nr(right, push_items - 1);
3892
3893         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
3894                      leaf_data_end(fs_info, left) - push_space,
3895                      btrfs_leaf_data(right) +
3896                      btrfs_item_offset_nr(right, push_items - 1),
3897                      push_space);
3898         old_left_nritems = btrfs_header_nritems(left);
3899         BUG_ON(old_left_nritems <= 0);
3900
3901         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
3902         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
3903                 u32 ioff;
3904
3905                 item = btrfs_item_nr(i);
3906
3907                 ioff = btrfs_token_item_offset(left, item, &token);
3908                 btrfs_set_token_item_offset(left, item,
3909                       ioff - (BTRFS_LEAF_DATA_SIZE(fs_info) - old_left_item_size),
3910                       &token);
3911         }
3912         btrfs_set_header_nritems(left, old_left_nritems + push_items);
3913
3914         /* fixup right node */
3915         if (push_items > right_nritems)
3916                 WARN(1, KERN_CRIT "push items %d nr %u\n", push_items,
3917                        right_nritems);
3918
3919         if (push_items < right_nritems) {
3920                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
3921                                                   leaf_data_end(fs_info, right);
3922                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
3923                                       BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3924                                       btrfs_leaf_data(right) +
3925                                       leaf_data_end(fs_info, right), push_space);
3926
3927                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
3928                               btrfs_item_nr_offset(push_items),
3929                              (btrfs_header_nritems(right) - push_items) *
3930                              sizeof(struct btrfs_item));
3931         }
3932         right_nritems -= push_items;
3933         btrfs_set_header_nritems(right, right_nritems);
3934         push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3935         for (i = 0; i < right_nritems; i++) {
3936                 item = btrfs_item_nr(i);
3937
3938                 push_space = push_space - btrfs_token_item_size(right,
3939                                                                 item, &token);
3940                 btrfs_set_token_item_offset(right, item, push_space, &token);
3941         }
3942
3943         btrfs_mark_buffer_dirty(left);
3944         if (right_nritems)
3945                 btrfs_mark_buffer_dirty(right);
3946         else
3947                 clean_tree_block(fs_info, right);
3948
3949         btrfs_item_key(right, &disk_key, 0);
3950         fixup_low_keys(fs_info, path, &disk_key, 1);
3951
3952         /* then fixup the leaf pointer in the path */
3953         if (path->slots[0] < push_items) {
3954                 path->slots[0] += old_left_nritems;
3955                 btrfs_tree_unlock(path->nodes[0]);
3956                 free_extent_buffer(path->nodes[0]);
3957                 path->nodes[0] = left;
3958                 path->slots[1] -= 1;
3959         } else {
3960                 btrfs_tree_unlock(left);
3961                 free_extent_buffer(left);
3962                 path->slots[0] -= push_items;
3963         }
3964         BUG_ON(path->slots[0] < 0);
3965         return ret;
3966 out:
3967         btrfs_tree_unlock(left);
3968         free_extent_buffer(left);
3969         return ret;
3970 }
3971
3972 /*
3973  * push some data in the path leaf to the left, trying to free up at
3974  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
3975  *
3976  * max_slot can put a limit on how far into the leaf we'll push items.  The
3977  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us push all the
3978  * items
3979  */
3980 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
3981                           *root, struct btrfs_path *path, int min_data_size,
3982                           int data_size, int empty, u32 max_slot)
3983 {
3984         struct btrfs_fs_info *fs_info = root->fs_info;
3985         struct extent_buffer *right = path->nodes[0];
3986         struct extent_buffer *left;
3987         int slot;
3988         int free_space;
3989         u32 right_nritems;
3990         int ret = 0;
3991
3992         slot = path->slots[1];
3993         if (slot == 0)
3994                 return 1;
3995         if (!path->nodes[1])
3996                 return 1;
3997
3998         right_nritems = btrfs_header_nritems(right);
3999         if (right_nritems == 0)
4000                 return 1;
4001
4002         btrfs_assert_tree_locked(path->nodes[1]);
4003
4004         left = read_node_slot(fs_info, path->nodes[1], slot - 1);
4005         /*
4006          * slot - 1 is not valid or we fail to read the left node,
4007          * no big deal, just return.
4008          */
4009         if (IS_ERR(left))
4010                 return 1;
4011
4012         btrfs_tree_lock(left);
4013         btrfs_set_lock_blocking(left);
4014
4015         free_space = btrfs_leaf_free_space(fs_info, left);
4016         if (free_space < data_size) {
4017                 ret = 1;
4018                 goto out;
4019         }
4020
4021         /* cow and double check */
4022         ret = btrfs_cow_block(trans, root, left,
4023                               path->nodes[1], slot - 1, &left);
4024         if (ret) {
4025                 /* we hit -ENOSPC, but it isn't fatal here */
4026                 if (ret == -ENOSPC)
4027                         ret = 1;
4028                 goto out;
4029         }
4030
4031         free_space = btrfs_leaf_free_space(fs_info, left);
4032         if (free_space < data_size) {
4033                 ret = 1;
4034                 goto out;
4035         }
4036
4037         return __push_leaf_left(fs_info, path, min_data_size,
4038                                empty, left, free_space, right_nritems,
4039                                max_slot);
4040 out:
4041         btrfs_tree_unlock(left);
4042         free_extent_buffer(left);
4043         return ret;
4044 }
4045
4046 /*
4047  * split the path's leaf in two, making sure there is at least data_size
4048  * available for the resulting leaf level of the path.
4049  */
4050 static noinline void copy_for_split(struct btrfs_trans_handle *trans,
4051                                     struct btrfs_fs_info *fs_info,
4052                                     struct btrfs_path *path,
4053                                     struct extent_buffer *l,
4054                                     struct extent_buffer *right,
4055                                     int slot, int mid, int nritems)
4056 {
4057         int data_copy_size;
4058         int rt_data_off;
4059         int i;
4060         struct btrfs_disk_key disk_key;
4061         struct btrfs_map_token token;
4062
4063         btrfs_init_map_token(&token);
4064
4065         nritems = nritems - mid;
4066         btrfs_set_header_nritems(right, nritems);
4067         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(fs_info, l);
4068
4069         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
4070                            btrfs_item_nr_offset(mid),
4071                            nritems * sizeof(struct btrfs_item));
4072
4073         copy_extent_buffer(right, l,
4074                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(fs_info) -
4075                      data_copy_size, btrfs_leaf_data(l) +
4076                      leaf_data_end(fs_info, l), data_copy_size);
4077
4078         rt_data_off = BTRFS_LEAF_DATA_SIZE(fs_info) - btrfs_item_end_nr(l, mid);
4079
4080         for (i = 0; i < nritems; i++) {
4081                 struct btrfs_item *item = btrfs_item_nr(i);
4082                 u32 ioff;
4083
4084                 ioff = btrfs_token_item_offset(right, item, &token);
4085                 btrfs_set_token_item_offset(right, item,
4086                                             ioff + rt_data_off, &token);
4087         }
4088
4089         btrfs_set_header_nritems(l, mid);
4090         btrfs_item_key(right, &disk_key, 0);
4091         insert_ptr(trans, fs_info, path, &disk_key, right->start,
4092                    path->slots[1] + 1, 1);
4093
4094         btrfs_mark_buffer_dirty(right);
4095         btrfs_mark_buffer_dirty(l);
4096         BUG_ON(path->slots[0] != slot);
4097
4098         if (mid <= slot) {
4099                 btrfs_tree_unlock(path->nodes[0]);
4100                 free_extent_buffer(path->nodes[0]);
4101                 path->nodes[0] = right;
4102                 path->slots[0] -= mid;
4103                 path->slots[1] += 1;
4104         } else {
4105                 btrfs_tree_unlock(right);
4106                 free_extent_buffer(right);
4107         }
4108
4109         BUG_ON(path->slots[0] < 0);
4110 }
4111
4112 /*
4113  * double splits happen when we need to insert a big item in the middle
4114  * of a leaf.  A double split can leave us with 3 mostly empty leaves:
4115  * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
4116  *          A                 B                 C
4117  *
4118  * We avoid this by trying to push the items on either side of our target
4119  * into the adjacent leaves.  If all goes well we can avoid the double split
4120  * completely.
4121  */
4122 static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
4123                                           struct btrfs_root *root,
4124                                           struct btrfs_path *path,
4125                                           int data_size)
4126 {
4127         struct btrfs_fs_info *fs_info = root->fs_info;
4128         int ret;
4129         int progress = 0;
4130         int slot;
4131         u32 nritems;
4132         int space_needed = data_size;
4133
4134         slot = path->slots[0];
4135         if (slot < btrfs_header_nritems(path->nodes[0]))
4136                 space_needed -= btrfs_leaf_free_space(fs_info, path->nodes[0]);
4137
4138         /*
4139          * try to push all the items after our slot into the
4140          * right leaf
4141          */
4142         ret = push_leaf_right(trans, root, path, 1, space_needed, 0, slot);
4143         if (ret < 0)
4144                 return ret;
4145
4146         if (ret == 0)
4147                 progress++;
4148
4149         nritems = btrfs_header_nritems(path->nodes[0]);
4150         /*
4151          * our goal is to get our slot at the start or end of a leaf.  If
4152          * we've done so we're done
4153          */
4154         if (path->slots[0] == 0 || path->slots[0] == nritems)
4155                 return 0;
4156
4157         if (btrfs_leaf_free_space(fs_info, path->nodes[0]) >= data_size)
4158                 return 0;
4159
4160         /* try to push all the items before our slot into the next leaf */
4161         slot = path->slots[0];
4162         ret = push_leaf_left(trans, root, path, 1, space_needed, 0, slot);
4163         if (ret < 0)
4164                 return ret;
4165
4166         if (ret == 0)
4167                 progress++;
4168
4169         if (progress)
4170                 return 0;
4171         return 1;
4172 }
4173
4174 /*
4175  * split the path's leaf in two, making sure there is at least data_size
4176  * available for the resulting leaf level of the path.
4177  *
4178  * returns 0 if all went well and < 0 on failure.
4179  */
4180 static noinline int split_leaf(struct btrfs_trans_handle *trans,
4181                                struct btrfs_root *root,
4182                                const struct btrfs_key *ins_key,
4183                                struct btrfs_path *path, int data_size,
4184                                int extend)
4185 {
4186         struct btrfs_disk_key disk_key;
4187         struct extent_buffer *l;
4188         u32 nritems;
4189         int mid;
4190         int slot;
4191         struct extent_buffer *right;
4192         struct btrfs_fs_info *fs_info = root->fs_info;
4193         int ret = 0;
4194         int wret;
4195         int split;
4196         int num_doubles = 0;
4197         int tried_avoid_double = 0;
4198
4199         l = path->nodes[0];
4200         slot = path->slots[0];
4201         if (extend && data_size + btrfs_item_size_nr(l, slot) +
4202             sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(fs_info))
4203                 return -EOVERFLOW;
4204
4205         /* first try to make some room by pushing left and right */
4206         if (data_size && path->nodes[1]) {
4207                 int space_needed = data_size;
4208
4209                 if (slot < btrfs_header_nritems(l))
4210                         space_needed -= btrfs_leaf_free_space(fs_info, l);
4211
4212                 wret = push_leaf_right(trans, root, path, space_needed,
4213                                        space_needed, 0, 0);
4214                 if (wret < 0)
4215                         return wret;
4216                 if (wret) {
4217                         wret = push_leaf_left(trans, root, path, space_needed,
4218                                               space_needed, 0, (u32)-1);
4219                         if (wret < 0)
4220                                 return wret;
4221                 }
4222                 l = path->nodes[0];
4223
4224                 /* did the pushes work? */
4225                 if (btrfs_leaf_free_space(fs_info, l) >= data_size)
4226                         return 0;
4227         }
4228
4229         if (!path->nodes[1]) {
4230                 ret = insert_new_root(trans, root, path, 1);
4231                 if (ret)
4232                         return ret;
4233         }
4234 again:
4235         split = 1;
4236         l = path->nodes[0];
4237         slot = path->slots[0];
4238         nritems = btrfs_header_nritems(l);
4239         mid = (nritems + 1) / 2;
4240
4241         if (mid <= slot) {
4242                 if (nritems == 1 ||
4243                     leaf_space_used(l, mid, nritems - mid) + data_size >
4244                         BTRFS_LEAF_DATA_SIZE(fs_info)) {
4245                         if (slot >= nritems) {
4246                                 split = 0;
4247                         } else {
4248                                 mid = slot;
4249                                 if (mid != nritems &&
4250                                     leaf_space_used(l, mid, nritems - mid) +
4251                                     data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
4252                                         if (data_size && !tried_avoid_double)
4253                                                 goto push_for_double;
4254                                         split = 2;
4255                                 }
4256                         }
4257                 }
4258         } else {
4259                 if (leaf_space_used(l, 0, mid) + data_size >
4260                         BTRFS_LEAF_DATA_SIZE(fs_info)) {
4261                         if (!extend && data_size && slot == 0) {
4262                                 split = 0;
4263                         } else if ((extend || !data_size) && slot == 0) {
4264                                 mid = 1;
4265                         } else {
4266                                 mid = slot;
4267                                 if (mid != nritems &&
4268                                     leaf_space_used(l, mid, nritems - mid) +
4269                                     data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
4270                                         if (data_size && !tried_avoid_double)
4271                                                 goto push_for_double;
4272                                         split = 2;
4273                                 }
4274                         }
4275                 }
4276         }
4277
4278         if (split == 0)
4279                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
4280         else
4281                 btrfs_item_key(l, &disk_key, mid);
4282
4283         right = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
4284                         &disk_key, 0, l->start, 0);
4285         if (IS_ERR(right))
4286                 return PTR_ERR(right);
4287
4288         root_add_used(root, fs_info->nodesize);
4289
4290         memzero_extent_buffer(right, 0, sizeof(struct btrfs_header));
4291         btrfs_set_header_bytenr(right, right->start);
4292         btrfs_set_header_generation(right, trans->transid);
4293         btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
4294         btrfs_set_header_owner(right, root->root_key.objectid);
4295         btrfs_set_header_level(right, 0);
4296         write_extent_buffer_fsid(right, fs_info->fsid);
4297         write_extent_buffer_chunk_tree_uuid(right, fs_info->chunk_tree_uuid);
4298
4299         if (split == 0) {
4300                 if (mid <= slot) {
4301                         btrfs_set_header_nritems(right, 0);
4302                         insert_ptr(trans, fs_info, path, &disk_key,
4303                                    right->start, path->slots[1] + 1, 1);
4304                         btrfs_tree_unlock(path->nodes[0]);
4305                         free_extent_buffer(path->nodes[0]);
4306                         path->nodes[0] = right;
4307                         path->slots[0] = 0;
4308                         path->slots[1] += 1;
4309                 } else {
4310                         btrfs_set_header_nritems(right, 0);
4311                         insert_ptr(trans, fs_info, path, &disk_key,
4312                                    right->start, path->slots[1], 1);
4313                         btrfs_tree_unlock(path->nodes[0]);
4314                         free_extent_buffer(path->nodes[0]);
4315                         path->nodes[0] = right;
4316                         path->slots[0] = 0;
4317                         if (path->slots[1] == 0)
4318                                 fixup_low_keys(fs_info, path, &disk_key, 1);
4319                 }
4320                 /*
4321                  * We create a new leaf 'right' for the required ins_len and
4322                  * we'll do btrfs_mark_buffer_dirty() on this leaf after copying
4323                  * the content of ins_len to 'right'.
4324                  */
4325                 return ret;
4326         }
4327
4328         copy_for_split(trans, fs_info, path, l, right, slot, mid, nritems);
4329
4330         if (split == 2) {
4331                 BUG_ON(num_doubles != 0);
4332                 num_doubles++;
4333                 goto again;
4334         }
4335
4336         return 0;
4337
4338 push_for_double:
4339         push_for_double_split(trans, root, path, data_size);
4340         tried_avoid_double = 1;
4341         if (btrfs_leaf_free_space(fs_info, path->nodes[0]) >= data_size)
4342                 return 0;
4343         goto again;
4344 }
4345
4346 static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
4347                                          struct btrfs_root *root,
4348                                          struct btrfs_path *path, int ins_len)
4349 {
4350         struct btrfs_fs_info *fs_info = root->fs_info;
4351         struct btrfs_key key;
4352         struct extent_buffer *leaf;
4353         struct btrfs_file_extent_item *fi;
4354         u64 extent_len = 0;
4355         u32 item_size;
4356         int ret;
4357
4358         leaf = path->nodes[0];
4359         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4360
4361         BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
4362                key.type != BTRFS_EXTENT_CSUM_KEY);
4363
4364         if (btrfs_leaf_free_space(fs_info, leaf) >= ins_len)
4365                 return 0;
4366
4367         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4368         if (key.type == BTRFS_EXTENT_DATA_KEY) {
4369                 fi = btrfs_item_ptr(leaf, path->slots[0],
4370                                     struct btrfs_file_extent_item);
4371                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
4372         }
4373         btrfs_release_path(path);
4374
4375         path->keep_locks = 1;
4376         path->search_for_split = 1;
4377         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
4378         path->search_for_split = 0;
4379         if (ret > 0)
4380                 ret = -EAGAIN;
4381         if (ret < 0)
4382                 goto err;
4383
4384         ret = -EAGAIN;
4385         leaf = path->nodes[0];
4386         /* if our item isn't there, return now */
4387         if (item_size != btrfs_item_size_nr(leaf, path->slots[0]))
4388                 goto err;
4389
4390         /* the leaf has  changed, it now has room.  return now */
4391         if (btrfs_leaf_free_space(fs_info, path->nodes[0]) >= ins_len)
4392                 goto err;
4393
4394         if (key.type == BTRFS_EXTENT_DATA_KEY) {
4395                 fi = btrfs_item_ptr(leaf, path->slots[0],
4396                                     struct btrfs_file_extent_item);
4397                 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
4398                         goto err;
4399         }
4400
4401         btrfs_set_path_blocking(path);
4402         ret = split_leaf(trans, root, &key, path, ins_len, 1);
4403         if (ret)
4404                 goto err;
4405
4406         path->keep_locks = 0;
4407         btrfs_unlock_up_safe(path, 1);
4408         return 0;
4409 err:
4410         path->keep_locks = 0;
4411         return ret;
4412 }
4413
4414 static noinline int split_item(struct btrfs_fs_info *fs_info,
4415                                struct btrfs_path *path,
4416                                const struct btrfs_key *new_key,
4417                                unsigned long split_offset)
4418 {
4419         struct extent_buffer *leaf;
4420         struct btrfs_item *item;
4421         struct btrfs_item *new_item;
4422         int slot;
4423         char *buf;
4424         u32 nritems;
4425         u32 item_size;
4426         u32 orig_offset;
4427         struct btrfs_disk_key disk_key;
4428
4429         leaf = path->nodes[0];
4430         BUG_ON(btrfs_leaf_free_space(fs_info, leaf) < sizeof(struct btrfs_item));
4431
4432         btrfs_set_path_blocking(path);
4433
4434         item = btrfs_item_nr(path->slots[0]);
4435         orig_offset = btrfs_item_offset(leaf, item);
4436         item_size = btrfs_item_size(leaf, item);
4437
4438         buf = kmalloc(item_size, GFP_NOFS);
4439         if (!buf)
4440                 return -ENOMEM;
4441
4442         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
4443                             path->slots[0]), item_size);
4444
4445         slot = path->slots[0] + 1;
4446         nritems = btrfs_header_nritems(leaf);
4447         if (slot != nritems) {
4448                 /* shift the items */
4449                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
4450                                 btrfs_item_nr_offset(slot),
4451                                 (nritems - slot) * sizeof(struct btrfs_item));
4452         }
4453
4454         btrfs_cpu_key_to_disk(&disk_key, new_key);
4455         btrfs_set_item_key(leaf, &disk_key, slot);
4456
4457         new_item = btrfs_item_nr(slot);
4458
4459         btrfs_set_item_offset(leaf, new_item, orig_offset);
4460         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
4461
4462         btrfs_set_item_offset(leaf, item,
4463                               orig_offset + item_size - split_offset);
4464         btrfs_set_item_size(leaf, item, split_offset);
4465
4466         btrfs_set_header_nritems(leaf, nritems + 1);
4467
4468         /* write the data for the start of the original item */
4469         write_extent_buffer(leaf, buf,
4470                             btrfs_item_ptr_offset(leaf, path->slots[0]),
4471                             split_offset);
4472
4473         /* write the data for the new item */
4474         write_extent_buffer(leaf, buf + split_offset,
4475                             btrfs_item_ptr_offset(leaf, slot),
4476                             item_size - split_offset);
4477         btrfs_mark_buffer_dirty(leaf);
4478
4479         BUG_ON(btrfs_leaf_free_space(fs_info, leaf) < 0);
4480         kfree(buf);
4481         return 0;
4482 }
4483
4484 /*
4485  * This function splits a single item into two items,
4486  * giving 'new_key' to the new item and splitting the
4487  * old one at split_offset (from the start of the item).
4488  *
4489  * The path may be released by this operation.  After
4490  * the split, the path is pointing to the old item.  The
4491  * new item is going to be in the same node as the old one.
4492  *
4493  * Note, the item being split must be smaller enough to live alone on
4494  * a tree block with room for one extra struct btrfs_item
4495  *
4496  * This allows us to split the item in place, keeping a lock on the
4497  * leaf the entire time.
4498  */
4499 int btrfs_split_item(struct btrfs_trans_handle *trans,
4500                      struct btrfs_root *root,
4501                      struct btrfs_path *path,
4502                      const struct btrfs_key *new_key,
4503                      unsigned long split_offset)
4504 {
4505         int ret;
4506         ret = setup_leaf_for_split(trans, root, path,
4507                                    sizeof(struct btrfs_item));
4508         if (ret)
4509                 return ret;
4510
4511         ret = split_item(root->fs_info, path, new_key, split_offset);
4512         return ret;
4513 }
4514
4515 /*
4516  * This function duplicate a item, giving 'new_key' to the new item.
4517  * It guarantees both items live in the same tree leaf and the new item
4518  * is contiguous with the original item.
4519  *
4520  * This allows us to split file extent in place, keeping a lock on the
4521  * leaf the entire time.
4522  */
4523 int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
4524                          struct btrfs_root *root,
4525                          struct btrfs_path *path,
4526                          const struct btrfs_key *new_key)
4527 {
4528         struct extent_buffer *leaf;
4529         int ret;
4530         u32 item_size;
4531
4532         leaf = path->nodes[0];
4533         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4534         ret = setup_leaf_for_split(trans, root, path,
4535                                    item_size + sizeof(struct btrfs_item));
4536         if (ret)
4537                 return ret;
4538
4539         path->slots[0]++;
4540         setup_items_for_insert(root, path, new_key, &item_size,
4541                                item_size, item_size +
4542                                sizeof(struct btrfs_item), 1);
4543         leaf = path->nodes[0];
4544         memcpy_extent_buffer(leaf,
4545                              btrfs_item_ptr_offset(leaf, path->slots[0]),
4546                              btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
4547                              item_size);
4548         return 0;
4549 }
4550
4551 /*
4552  * make the item pointed to by the path smaller.  new_size indicates
4553  * how small to make it, and from_end tells us if we just chop bytes
4554  * off the end of the item or if we shift the item to chop bytes off
4555  * the front.
4556  */
4557 void btrfs_truncate_item(struct btrfs_fs_info *fs_info,
4558                          struct btrfs_path *path, u32 new_size, int from_end)
4559 {
4560         int slot;
4561         struct extent_buffer *leaf;
4562         struct btrfs_item *item;
4563         u32 nritems;
4564         unsigned int data_end;
4565         unsigned int old_data_start;
4566         unsigned int old_size;
4567         unsigned int size_diff;
4568         int i;
4569         struct btrfs_map_token token;
4570
4571         btrfs_init_map_token(&token);
4572
4573         leaf = path->nodes[0];
4574         slot = path->slots[0];
4575
4576         old_size = btrfs_item_size_nr(leaf, slot);
4577         if (old_size == new_size)
4578                 return;
4579
4580         nritems = btrfs_header_nritems(leaf);
4581         data_end = leaf_data_end(fs_info, leaf);
4582
4583         old_data_start = btrfs_item_offset_nr(leaf, slot);
4584
4585         size_diff = old_size - new_size;
4586
4587         BUG_ON(slot < 0);
4588         BUG_ON(slot >= nritems);
4589
4590         /*
4591          * item0..itemN ... dataN.offset..dataN.size .. data0.size
4592          */
4593         /* first correct the data pointers */
4594         for (i = slot; i < nritems; i++) {
4595                 u32 ioff;
4596                 item = btrfs_item_nr(i);
4597
4598                 ioff = btrfs_token_item_offset(leaf, item, &token);
4599                 btrfs_set_token_item_offset(leaf, item,
4600                                             ioff + size_diff, &token);
4601         }
4602
4603         /* shift the data */
4604         if (from_end) {
4605                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4606                               data_end + size_diff, btrfs_leaf_data(leaf) +
4607                               data_end, old_data_start + new_size - data_end);
4608         } else {
4609                 struct btrfs_disk_key disk_key;
4610                 u64 offset;
4611
4612                 btrfs_item_key(leaf, &disk_key, slot);
4613
4614                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
4615                         unsigned long ptr;
4616                         struct btrfs_file_extent_item *fi;
4617
4618                         fi = btrfs_item_ptr(leaf, slot,
4619                                             struct btrfs_file_extent_item);
4620                         fi = (struct btrfs_file_extent_item *)(
4621                              (unsigned long)fi - size_diff);
4622
4623                         if (btrfs_file_extent_type(leaf, fi) ==
4624                             BTRFS_FILE_EXTENT_INLINE) {
4625                                 ptr = btrfs_item_ptr_offset(leaf, slot);
4626                                 memmove_extent_buffer(leaf, ptr,
4627                                       (unsigned long)fi,
4628                                       BTRFS_FILE_EXTENT_INLINE_DATA_START);
4629                         }
4630                 }
4631
4632                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4633                               data_end + size_diff, btrfs_leaf_data(leaf) +
4634                               data_end, old_data_start - data_end);
4635
4636                 offset = btrfs_disk_key_offset(&disk_key);
4637                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
4638                 btrfs_set_item_key(leaf, &disk_key, slot);
4639                 if (slot == 0)
4640                         fixup_low_keys(fs_info, path, &disk_key, 1);
4641         }
4642
4643         item = btrfs_item_nr(slot);
4644         btrfs_set_item_size(leaf, item, new_size);
4645         btrfs_mark_buffer_dirty(leaf);
4646
4647         if (btrfs_leaf_free_space(fs_info, leaf) < 0) {
4648                 btrfs_print_leaf(fs_info, leaf);
4649                 BUG();
4650         }
4651 }
4652
4653 /*
4654  * make the item pointed to by the path bigger, data_size is the added size.
4655  */
4656 void btrfs_extend_item(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
4657                        u32 data_size)
4658 {
4659         int slot;
4660         struct extent_buffer *leaf;
4661         struct btrfs_item *item;
4662         u32 nritems;
4663         unsigned int data_end;
4664         unsigned int old_data;
4665         unsigned int old_size;
4666         int i;
4667         struct btrfs_map_token token;
4668
4669         btrfs_init_map_token(&token);
4670
4671         leaf = path->nodes[0];
4672
4673         nritems = btrfs_header_nritems(leaf);
4674         data_end = leaf_data_end(fs_info, leaf);
4675
4676         if (btrfs_leaf_free_space(fs_info, leaf) < data_size) {
4677                 btrfs_print_leaf(fs_info, leaf);
4678                 BUG();
4679         }
4680         slot = path->slots[0];
4681         old_data = btrfs_item_end_nr(leaf, slot);
4682
4683         BUG_ON(slot < 0);
4684         if (slot >= nritems) {
4685                 btrfs_print_leaf(fs_info, leaf);
4686                 btrfs_crit(fs_info, "slot %d too large, nritems %d",
4687                            slot, nritems);
4688                 BUG_ON(1);
4689         }
4690
4691         /*
4692          * item0..itemN ... dataN.offset..dataN.size .. data0.size
4693          */
4694         /* first correct the data pointers */
4695         for (i = slot; i < nritems; i++) {
4696                 u32 ioff;
4697                 item = btrfs_item_nr(i);
4698
4699                 ioff = btrfs_token_item_offset(leaf, item, &token);
4700                 btrfs_set_token_item_offset(leaf, item,
4701                                             ioff - data_size, &token);
4702         }
4703
4704         /* shift the data */
4705         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4706                       data_end - data_size, btrfs_leaf_data(leaf) +
4707                       data_end, old_data - data_end);
4708
4709         data_end = old_data;
4710         old_size = btrfs_item_size_nr(leaf, slot);
4711         item = btrfs_item_nr(slot);
4712         btrfs_set_item_size(leaf, item, old_size + data_size);
4713         btrfs_mark_buffer_dirty(leaf);
4714
4715         if (btrfs_leaf_free_space(fs_info, leaf) < 0) {
4716                 btrfs_print_leaf(fs_info, leaf);
4717                 BUG();
4718         }
4719 }
4720
4721 /*
4722  * this is a helper for btrfs_insert_empty_items, the main goal here is
4723  * to save stack depth by doing the bulk of the work in a function
4724  * that doesn't call btrfs_search_slot
4725  */
4726 void setup_items_for_insert(struct btrfs_root *root, struct btrfs_path *path,
4727                             const struct btrfs_key *cpu_key, u32 *data_size,
4728                             u32 total_data, u32 total_size, int nr)
4729 {
4730         struct btrfs_fs_info *fs_info = root->fs_info;
4731         struct btrfs_item *item;
4732         int i;
4733         u32 nritems;
4734         unsigned int data_end;
4735         struct btrfs_disk_key disk_key;
4736         struct extent_buffer *leaf;
4737         int slot;
4738         struct btrfs_map_token token;
4739
4740         if (path->slots[0] == 0) {
4741                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
4742                 fixup_low_keys(fs_info, path, &disk_key, 1);
4743         }
4744         btrfs_unlock_up_safe(path, 1);
4745
4746         btrfs_init_map_token(&token);
4747
4748         leaf = path->nodes[0];
4749         slot = path->slots[0];
4750
4751         nritems = btrfs_header_nritems(leaf);
4752         data_end = leaf_data_end(fs_info, leaf);
4753
4754         if (btrfs_leaf_free_space(fs_info, leaf) < total_size) {
4755                 btrfs_print_leaf(fs_info, leaf);
4756                 btrfs_crit(fs_info, "not enough freespace need %u have %d",
4757                            total_size, btrfs_leaf_free_space(fs_info, leaf));
4758                 BUG();
4759         }
4760
4761         if (slot != nritems) {
4762                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
4763
4764                 if (old_data < data_end) {
4765                         btrfs_print_leaf(fs_info, leaf);
4766                         btrfs_crit(fs_info, "slot %d old_data %d data_end %d",
4767                                    slot, old_data, data_end);
4768                         BUG_ON(1);
4769                 }
4770                 /*
4771                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
4772                  */
4773                 /* first correct the data pointers */
4774                 for (i = slot; i < nritems; i++) {
4775                         u32 ioff;
4776
4777                         item = btrfs_item_nr(i);
4778                         ioff = btrfs_token_item_offset(leaf, item, &token);
4779                         btrfs_set_token_item_offset(leaf, item,
4780                                                     ioff - total_data, &token);
4781                 }
4782                 /* shift the items */
4783                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
4784                               btrfs_item_nr_offset(slot),
4785                               (nritems - slot) * sizeof(struct btrfs_item));
4786
4787                 /* shift the data */
4788                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4789                               data_end - total_data, btrfs_leaf_data(leaf) +
4790                               data_end, old_data - data_end);
4791                 data_end = old_data;
4792         }
4793
4794         /* setup the item for the new data */
4795         for (i = 0; i < nr; i++) {
4796                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
4797                 btrfs_set_item_key(leaf, &disk_key, slot + i);
4798                 item = btrfs_item_nr(slot + i);
4799                 btrfs_set_token_item_offset(leaf, item,
4800                                             data_end - data_size[i], &token);
4801                 data_end -= data_size[i];
4802                 btrfs_set_token_item_size(leaf, item, data_size[i], &token);
4803         }
4804
4805         btrfs_set_header_nritems(leaf, nritems + nr);
4806         btrfs_mark_buffer_dirty(leaf);
4807
4808         if (btrfs_leaf_free_space(fs_info, leaf) < 0) {
4809                 btrfs_print_leaf(fs_info, leaf);
4810                 BUG();
4811         }
4812 }
4813
4814 /*
4815  * Given a key and some data, insert items into the tree.
4816  * This does all the path init required, making room in the tree if needed.
4817  */
4818 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
4819                             struct btrfs_root *root,
4820                             struct btrfs_path *path,
4821                             const struct btrfs_key *cpu_key, u32 *data_size,
4822                             int nr)
4823 {
4824         int ret = 0;
4825         int slot;
4826         int i;
4827         u32 total_size = 0;
4828         u32 total_data = 0;
4829
4830         for (i = 0; i < nr; i++)
4831                 total_data += data_size[i];
4832
4833         total_size = total_data + (nr * sizeof(struct btrfs_item));
4834         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
4835         if (ret == 0)
4836                 return -EEXIST;
4837         if (ret < 0)
4838                 return ret;
4839
4840         slot = path->slots[0];
4841         BUG_ON(slot < 0);
4842
4843         setup_items_for_insert(root, path, cpu_key, data_size,
4844                                total_data, total_size, nr);
4845         return 0;
4846 }
4847
4848 /*
4849  * Given a key and some data, insert an item into the tree.
4850  * This does all the path init required, making room in the tree if needed.
4851  */
4852 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4853                       const struct btrfs_key *cpu_key, void *data,
4854                       u32 data_size)
4855 {
4856         int ret = 0;
4857         struct btrfs_path *path;
4858         struct extent_buffer *leaf;
4859         unsigned long ptr;
4860
4861         path = btrfs_alloc_path();
4862         if (!path)
4863                 return -ENOMEM;
4864         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
4865         if (!ret) {
4866                 leaf = path->nodes[0];
4867                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
4868                 write_extent_buffer(leaf, data, ptr, data_size);
4869                 btrfs_mark_buffer_dirty(leaf);
4870         }
4871         btrfs_free_path(path);
4872         return ret;
4873 }
4874
4875 /*
4876  * delete the pointer from a given node.
4877  *
4878  * the tree should have been previously balanced so the deletion does not
4879  * empty a node.
4880  */
4881 static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
4882                     int level, int slot)
4883 {
4884         struct btrfs_fs_info *fs_info = root->fs_info;
4885         struct extent_buffer *parent = path->nodes[level];
4886         u32 nritems;
4887         int ret;
4888
4889         nritems = btrfs_header_nritems(parent);
4890         if (slot != nritems - 1) {
4891                 if (level)
4892                         tree_mod_log_eb_move(fs_info, parent, slot,
4893                                              slot + 1, nritems - slot - 1);
4894                 memmove_extent_buffer(parent,
4895                               btrfs_node_key_ptr_offset(slot),
4896                               btrfs_node_key_ptr_offset(slot + 1),
4897                               sizeof(struct btrfs_key_ptr) *
4898                               (nritems - slot - 1));
4899         } else if (level) {
4900                 ret = tree_mod_log_insert_key(fs_info, parent, slot,
4901                                               MOD_LOG_KEY_REMOVE, GFP_NOFS);
4902                 BUG_ON(ret < 0);
4903         }
4904
4905         nritems--;
4906         btrfs_set_header_nritems(parent, nritems);
4907         if (nritems == 0 && parent == root->node) {
4908                 BUG_ON(btrfs_header_level(root->node) != 1);
4909                 /* just turn the root into a leaf and break */
4910                 btrfs_set_header_level(root->node, 0);
4911         } else if (slot == 0) {
4912                 struct btrfs_disk_key disk_key;
4913
4914                 btrfs_node_key(parent, &disk_key, 0);
4915                 fixup_low_keys(fs_info, path, &disk_key, level + 1);
4916         }
4917         btrfs_mark_buffer_dirty(parent);
4918 }
4919
4920 /*
4921  * a helper function to delete the leaf pointed to by path->slots[1] and
4922  * path->nodes[1].
4923  *
4924  * This deletes the pointer in path->nodes[1] and frees the leaf
4925  * block extent.  zero is returned if it all worked out, < 0 otherwise.
4926  *
4927  * The path must have already been setup for deleting the leaf, including
4928  * all the proper balancing.  path->nodes[1] must be locked.
4929  */
4930 static noinline void btrfs_del_leaf(struct btrfs_trans_handle *trans,
4931                                     struct btrfs_root *root,
4932                                     struct btrfs_path *path,
4933                                     struct extent_buffer *leaf)
4934 {
4935         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
4936         del_ptr(root, path, 1, path->slots[1]);
4937
4938         /*
4939          * btrfs_free_extent is expensive, we want to make sure we
4940          * aren't holding any locks when we call it
4941          */
4942         btrfs_unlock_up_safe(path, 0);
4943
4944         root_sub_used(root, leaf->len);
4945
4946         extent_buffer_get(leaf);
4947         btrfs_free_tree_block(trans, root, leaf, 0, 1);
4948         free_extent_buffer_stale(leaf);
4949 }
4950 /*
4951  * delete the item at the leaf level in path.  If that empties
4952  * the leaf, remove it from the tree
4953  */
4954 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4955                     struct btrfs_path *path, int slot, int nr)
4956 {
4957         struct btrfs_fs_info *fs_info = root->fs_info;
4958         struct extent_buffer *leaf;
4959         struct btrfs_item *item;
4960         u32 last_off;
4961         u32 dsize = 0;
4962         int ret = 0;
4963         int wret;
4964         int i;
4965         u32 nritems;
4966         struct btrfs_map_token token;
4967
4968         btrfs_init_map_token(&token);
4969
4970         leaf = path->nodes[0];
4971         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
4972
4973         for (i = 0; i < nr; i++)
4974                 dsize += btrfs_item_size_nr(leaf, slot + i);
4975
4976         nritems = btrfs_header_nritems(leaf);
4977
4978         if (slot + nr != nritems) {
4979                 int data_end = leaf_data_end(fs_info, leaf);
4980
4981                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
4982                               data_end + dsize,
4983                               btrfs_leaf_data(leaf) + data_end,
4984                               last_off - data_end);
4985
4986                 for (i = slot + nr; i < nritems; i++) {
4987                         u32 ioff;
4988
4989                         item = btrfs_item_nr(i);
4990                         ioff = btrfs_token_item_offset(leaf, item, &token);
4991                         btrfs_set_token_item_offset(leaf, item,
4992                                                     ioff + dsize, &token);
4993                 }
4994
4995                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
4996                               btrfs_item_nr_offset(slot + nr),
4997                               sizeof(struct btrfs_item) *
4998                               (nritems - slot - nr));
4999         }
5000         btrfs_set_header_nritems(leaf, nritems - nr);
5001         nritems -= nr;
5002
5003         /* delete the leaf if we've emptied it */
5004         if (nritems == 0) {
5005                 if (leaf == root->node) {
5006                         btrfs_set_header_level(leaf, 0);
5007                 } else {
5008                         btrfs_set_path_blocking(path);
5009                         clean_tree_block(fs_info, leaf);
5010                         btrfs_del_leaf(trans, root, path, leaf);
5011                 }
5012         } else {
5013                 int used = leaf_space_used(leaf, 0, nritems);
5014                 if (slot == 0) {
5015                         struct btrfs_disk_key disk_key;
5016
5017                         btrfs_item_key(leaf, &disk_key, 0);
5018                         fixup_low_keys(fs_info, path, &disk_key, 1);
5019                 }
5020
5021                 /* delete the leaf if it is mostly empty */
5022                 if (used < BTRFS_LEAF_DATA_SIZE(fs_info) / 3) {
5023                         /* push_leaf_left fixes the path.
5024                          * make sure the path still points to our leaf
5025                          * for possible call to del_ptr below
5026                          */
5027                         slot = path->slots[1];
5028                         extent_buffer_get(leaf);
5029
5030                         btrfs_set_path_blocking(path);
5031                         wret = push_leaf_left(trans, root, path, 1, 1,
5032                                               1, (u32)-1);
5033                         if (wret < 0 && wret != -ENOSPC)
5034                                 ret = wret;
5035
5036                         if (path->nodes[0] == leaf &&
5037                             btrfs_header_nritems(leaf)) {
5038                                 wret = push_leaf_right(trans, root, path, 1,
5039                                                        1, 1, 0);
5040                                 if (wret < 0 && wret != -ENOSPC)
5041                                         ret = wret;
5042                         }
5043
5044                         if (btrfs_header_nritems(leaf) == 0) {
5045                                 path->slots[1] = slot;
5046                                 btrfs_del_leaf(trans, root, path, leaf);
5047                                 free_extent_buffer(leaf);
5048                                 ret = 0;
5049                         } else {
5050                                 /* if we're still in the path, make sure
5051                                  * we're dirty.  Otherwise, one of the
5052                                  * push_leaf functions must have already
5053                                  * dirtied this buffer
5054                                  */
5055                                 if (path->nodes[0] == leaf)
5056                                         btrfs_mark_buffer_dirty(leaf);
5057                                 free_extent_buffer(leaf);
5058                         }
5059                 } else {
5060                         btrfs_mark_buffer_dirty(leaf);
5061                 }
5062         }
5063         return ret;
5064 }
5065
5066 /*
5067  * search the tree again to find a leaf with lesser keys
5068  * returns 0 if it found something or 1 if there are no lesser leaves.
5069  * returns < 0 on io errors.
5070  *
5071  * This may release the path, and so you may lose any locks held at the
5072  * time you call it.
5073  */
5074 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
5075 {
5076         struct btrfs_key key;
5077         struct btrfs_disk_key found_key;
5078         int ret;
5079
5080         btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
5081
5082         if (key.offset > 0) {
5083                 key.offset--;
5084         } else if (key.type > 0) {
5085                 key.type--;
5086                 key.offset = (u64)-1;
5087         } else if (key.objectid > 0) {
5088                 key.objectid--;
5089                 key.type = (u8)-1;
5090                 key.offset = (u64)-1;
5091         } else {
5092                 return 1;
5093         }
5094
5095         btrfs_release_path(path);
5096         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5097         if (ret < 0)
5098                 return ret;
5099         btrfs_item_key(path->nodes[0], &found_key, 0);
5100         ret = comp_keys(&found_key, &key);
5101         /*
5102          * We might have had an item with the previous key in the tree right
5103          * before we released our path. And after we released our path, that
5104          * item might have been pushed to the first slot (0) of the leaf we
5105          * were holding due to a tree balance. Alternatively, an item with the
5106          * previous key can exist as the only element of a leaf (big fat item).
5107          * Therefore account for these 2 cases, so that our callers (like
5108          * btrfs_previous_item) don't miss an existing item with a key matching
5109          * the previous key we computed above.
5110          */
5111         if (ret <= 0)
5112                 return 0;
5113         return 1;
5114 }
5115
5116 /*
5117  * A helper function to walk down the tree starting at min_key, and looking
5118  * for nodes or leaves that are have a minimum transaction id.
5119  * This is used by the btree defrag code, and tree logging
5120  *
5121  * This does not cow, but it does stuff the starting key it finds back
5122  * into min_key, so you can call btrfs_search_slot with cow=1 on the
5123  * key and get a writable path.
5124  *
5125  * This does lock as it descends, and path->keep_locks should be set
5126  * to 1 by the caller.
5127  *
5128  * This honors path->lowest_level to prevent descent past a given level
5129  * of the tree.
5130  *
5131  * min_trans indicates the oldest transaction that you are interested
5132  * in walking through.  Any nodes or leaves older than min_trans are
5133  * skipped over (without reading them).
5134  *
5135  * returns zero if something useful was found, < 0 on error and 1 if there
5136  * was nothing in the tree that matched the search criteria.
5137  */
5138 int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
5139                          struct btrfs_path *path,
5140                          u64 min_trans)
5141 {
5142         struct btrfs_fs_info *fs_info = root->fs_info;
5143         struct extent_buffer *cur;
5144         struct btrfs_key found_key;
5145         int slot;
5146         int sret;
5147         u32 nritems;
5148         int level;
5149         int ret = 1;
5150         int keep_locks = path->keep_locks;
5151
5152         path->keep_locks = 1;
5153 again:
5154         cur = btrfs_read_lock_root_node(root);
5155         level = btrfs_header_level(cur);
5156         WARN_ON(path->nodes[level]);
5157         path->nodes[level] = cur;
5158         path->locks[level] = BTRFS_READ_LOCK;
5159
5160         if (btrfs_header_generation(cur) < min_trans) {
5161                 ret = 1;
5162                 goto out;
5163         }
5164         while (1) {
5165                 nritems = btrfs_header_nritems(cur);
5166                 level = btrfs_header_level(cur);
5167                 sret = bin_search(cur, min_key, level, &slot);
5168
5169                 /* at the lowest level, we're done, setup the path and exit */
5170                 if (level == path->lowest_level) {
5171                         if (slot >= nritems)
5172                                 goto find_next_key;
5173                         ret = 0;
5174                         path->slots[level] = slot;
5175                         btrfs_item_key_to_cpu(cur, &found_key, slot);
5176                         goto out;
5177                 }
5178                 if (sret && slot > 0)
5179                         slot--;
5180                 /*
5181                  * check this node pointer against the min_trans parameters.
5182                  * If it is too old, old, skip to the next one.
5183                  */
5184                 while (slot < nritems) {
5185                         u64 gen;
5186
5187                         gen = btrfs_node_ptr_generation(cur, slot);
5188                         if (gen < min_trans) {
5189                                 slot++;
5190                                 continue;
5191                         }
5192                         break;
5193                 }
5194 find_next_key:
5195                 /*
5196                  * we didn't find a candidate key in this node, walk forward
5197                  * and find another one
5198                  */
5199                 if (slot >= nritems) {
5200                         path->slots[level] = slot;
5201                         btrfs_set_path_blocking(path);
5202                         sret = btrfs_find_next_key(root, path, min_key, level,
5203                                                   min_trans);
5204                         if (sret == 0) {
5205                                 btrfs_release_path(path);
5206                                 goto again;
5207                         } else {
5208                                 goto out;
5209                         }
5210                 }
5211                 /* save our key for returning back */
5212                 btrfs_node_key_to_cpu(cur, &found_key, slot);
5213                 path->slots[level] = slot;
5214                 if (level == path->lowest_level) {
5215                         ret = 0;
5216                         goto out;
5217                 }
5218                 btrfs_set_path_blocking(path);
5219                 cur = read_node_slot(fs_info, cur, slot);
5220                 if (IS_ERR(cur)) {
5221                         ret = PTR_ERR(cur);
5222                         goto out;
5223                 }
5224
5225                 btrfs_tree_read_lock(cur);
5226
5227                 path->locks[level - 1] = BTRFS_READ_LOCK;
5228                 path->nodes[level - 1] = cur;
5229                 unlock_up(path, level, 1, 0, NULL);
5230                 btrfs_clear_path_blocking(path, NULL, 0);
5231         }
5232 out:
5233         path->keep_locks = keep_locks;
5234         if (ret == 0) {
5235                 btrfs_unlock_up_safe(path, path->lowest_level + 1);
5236                 btrfs_set_path_blocking(path);
5237                 memcpy(min_key, &found_key, sizeof(found_key));
5238         }
5239         return ret;
5240 }
5241
5242 static int tree_move_down(struct btrfs_fs_info *fs_info,
5243                            struct btrfs_path *path,
5244                            int *level)
5245 {
5246         struct extent_buffer *eb;
5247
5248         BUG_ON(*level == 0);
5249         eb = read_node_slot(fs_info, path->nodes[*level], path->slots[*level]);
5250         if (IS_ERR(eb))
5251                 return PTR_ERR(eb);
5252
5253         path->nodes[*level - 1] = eb;
5254         path->slots[*level - 1] = 0;
5255         (*level)--;
5256         return 0;
5257 }
5258
5259 static int tree_move_next_or_upnext(struct btrfs_path *path,
5260                                     int *level, int root_level)
5261 {
5262         int ret = 0;
5263         int nritems;
5264         nritems = btrfs_header_nritems(path->nodes[*level]);
5265
5266         path->slots[*level]++;
5267
5268         while (path->slots[*level] >= nritems) {
5269                 if (*level == root_level)
5270                         return -1;
5271
5272                 /* move upnext */
5273                 path->slots[*level] = 0;
5274                 free_extent_buffer(path->nodes[*level]);
5275                 path->nodes[*level] = NULL;
5276                 (*level)++;
5277                 path->slots[*level]++;
5278
5279                 nritems = btrfs_header_nritems(path->nodes[*level]);
5280                 ret = 1;
5281         }
5282         return ret;
5283 }
5284
5285 /*
5286  * Returns 1 if it had to move up and next. 0 is returned if it moved only next
5287  * or down.
5288  */
5289 static int tree_advance(struct btrfs_fs_info *fs_info,
5290                         struct btrfs_path *path,
5291                         int *level, int root_level,
5292                         int allow_down,
5293                         struct btrfs_key *key)
5294 {
5295         int ret;
5296
5297         if (*level == 0 || !allow_down) {
5298                 ret = tree_move_next_or_upnext(path, level, root_level);
5299         } else {
5300                 ret = tree_move_down(fs_info, path, level);
5301         }
5302         if (ret >= 0) {
5303                 if (*level == 0)
5304                         btrfs_item_key_to_cpu(path->nodes[*level], key,
5305                                         path->slots[*level]);
5306                 else
5307                         btrfs_node_key_to_cpu(path->nodes[*level], key,
5308                                         path->slots[*level]);
5309         }
5310         return ret;
5311 }
5312
5313 static int tree_compare_item(struct btrfs_path *left_path,
5314                              struct btrfs_path *right_path,
5315                              char *tmp_buf)
5316 {
5317         int cmp;
5318         int len1, len2;
5319         unsigned long off1, off2;
5320
5321         len1 = btrfs_item_size_nr(left_path->nodes[0], left_path->slots[0]);
5322         len2 = btrfs_item_size_nr(right_path->nodes[0], right_path->slots[0]);
5323         if (len1 != len2)
5324                 return 1;
5325
5326         off1 = btrfs_item_ptr_offset(left_path->nodes[0], left_path->slots[0]);
5327         off2 = btrfs_item_ptr_offset(right_path->nodes[0],
5328                                 right_path->slots[0]);
5329
5330         read_extent_buffer(left_path->nodes[0], tmp_buf, off1, len1);
5331
5332         cmp = memcmp_extent_buffer(right_path->nodes[0], tmp_buf, off2, len1);
5333         if (cmp)
5334                 return 1;
5335         return 0;
5336 }
5337
5338 #define ADVANCE 1
5339 #define ADVANCE_ONLY_NEXT -1
5340
5341 /*
5342  * This function compares two trees and calls the provided callback for
5343  * every changed/new/deleted item it finds.
5344  * If shared tree blocks are encountered, whole subtrees are skipped, making
5345  * the compare pretty fast on snapshotted subvolumes.
5346  *
5347  * This currently works on commit roots only. As commit roots are read only,
5348  * we don't do any locking. The commit roots are protected with transactions.
5349  * Transactions are ended and rejoined when a commit is tried in between.
5350  *
5351  * This function checks for modifications done to the trees while comparing.
5352  * If it detects a change, it aborts immediately.
5353  */
5354 int btrfs_compare_trees(struct btrfs_root *left_root,
5355                         struct btrfs_root *right_root,
5356                         btrfs_changed_cb_t changed_cb, void *ctx)
5357 {
5358         struct btrfs_fs_info *fs_info = left_root->fs_info;
5359         int ret;
5360         int cmp;
5361         struct btrfs_path *left_path = NULL;
5362         struct btrfs_path *right_path = NULL;
5363         struct btrfs_key left_key;
5364         struct btrfs_key right_key;
5365         char *tmp_buf = NULL;
5366         int left_root_level;
5367         int right_root_level;
5368         int left_level;
5369         int right_level;
5370         int left_end_reached;
5371         int right_end_reached;
5372         int advance_left;
5373         int advance_right;
5374         u64 left_blockptr;
5375         u64 right_blockptr;
5376         u64 left_gen;
5377         u64 right_gen;
5378
5379         left_path = btrfs_alloc_path();
5380         if (!left_path) {
5381                 ret = -ENOMEM;
5382                 goto out;
5383         }
5384         right_path = btrfs_alloc_path();
5385         if (!right_path) {
5386                 ret = -ENOMEM;
5387                 goto out;
5388         }
5389
5390         tmp_buf = kmalloc(fs_info->nodesize, GFP_KERNEL | __GFP_NOWARN);
5391         if (!tmp_buf) {
5392                 tmp_buf = vmalloc(fs_info->nodesize);
5393                 if (!tmp_buf) {
5394                         ret = -ENOMEM;
5395                         goto out;
5396                 }
5397         }
5398
5399         left_path->search_commit_root = 1;
5400         left_path->skip_locking = 1;
5401         right_path->search_commit_root = 1;
5402         right_path->skip_locking = 1;
5403
5404         /*
5405          * Strategy: Go to the first items of both trees. Then do
5406          *
5407          * If both trees are at level 0
5408          *   Compare keys of current items
5409          *     If left < right treat left item as new, advance left tree
5410          *       and repeat
5411          *     If left > right treat right item as deleted, advance right tree
5412          *       and repeat
5413          *     If left == right do deep compare of items, treat as changed if
5414          *       needed, advance both trees and repeat
5415          * If both trees are at the same level but not at level 0
5416          *   Compare keys of current nodes/leafs
5417          *     If left < right advance left tree and repeat
5418          *     If left > right advance right tree and repeat
5419          *     If left == right compare blockptrs of the next nodes/leafs
5420          *       If they match advance both trees but stay at the same level
5421          *         and repeat
5422          *       If they don't match advance both trees while allowing to go
5423          *         deeper and repeat
5424          * If tree levels are different
5425          *   Advance the tree that needs it and repeat
5426          *
5427          * Advancing a tree means:
5428          *   If we are at level 0, try to go to the next slot. If that's not
5429          *   possible, go one level up and repeat. Stop when we found a level
5430          *   where we could go to the next slot. We may at this point be on a
5431          *   node or a leaf.
5432          *
5433          *   If we are not at level 0 and not on shared tree blocks, go one
5434          *   level deeper.
5435          *
5436          *   If we are not at level 0 and on shared tree blocks, go one slot to
5437          *   the right if possible or go up and right.
5438          */
5439
5440         down_read(&fs_info->commit_root_sem);
5441         left_level = btrfs_header_level(left_root->commit_root);
5442         left_root_level = left_level;
5443         left_path->nodes[left_level] = left_root->commit_root;
5444         extent_buffer_get(left_path->nodes[left_level]);
5445
5446         right_level = btrfs_header_level(right_root->commit_root);
5447         right_root_level = right_level;
5448         right_path->nodes[right_level] = right_root->commit_root;
5449         extent_buffer_get(right_path->nodes[right_level]);
5450         up_read(&fs_info->commit_root_sem);
5451
5452         if (left_level == 0)
5453                 btrfs_item_key_to_cpu(left_path->nodes[left_level],
5454                                 &left_key, left_path->slots[left_level]);
5455         else
5456                 btrfs_node_key_to_cpu(left_path->nodes[left_level],
5457                                 &left_key, left_path->slots[left_level]);
5458         if (right_level == 0)
5459                 btrfs_item_key_to_cpu(right_path->nodes[right_level],
5460                                 &right_key, right_path->slots[right_level]);
5461         else
5462                 btrfs_node_key_to_cpu(right_path->nodes[right_level],
5463                                 &right_key, right_path->slots[right_level]);
5464
5465         left_end_reached = right_end_reached = 0;
5466         advance_left = advance_right = 0;
5467
5468         while (1) {
5469                 if (advance_left && !left_end_reached) {
5470                         ret = tree_advance(fs_info, left_path, &left_level,
5471                                         left_root_level,
5472                                         advance_left != ADVANCE_ONLY_NEXT,
5473                                         &left_key);
5474                         if (ret == -1)
5475                                 left_end_reached = ADVANCE;
5476                         else if (ret < 0)
5477                                 goto out;
5478                         advance_left = 0;
5479                 }
5480                 if (advance_right && !right_end_reached) {
5481                         ret = tree_advance(fs_info, right_path, &right_level,
5482                                         right_root_level,
5483                                         advance_right != ADVANCE_ONLY_NEXT,
5484                                         &right_key);
5485                         if (ret == -1)
5486                                 right_end_reached = ADVANCE;
5487                         else if (ret < 0)
5488                                 goto out;
5489                         advance_right = 0;
5490                 }
5491
5492                 if (left_end_reached && right_end_reached) {
5493                         ret = 0;
5494                         goto out;
5495                 } else if (left_end_reached) {
5496                         if (right_level == 0) {
5497                                 ret = changed_cb(left_root, right_root,
5498                                                 left_path, right_path,
5499                                                 &right_key,
5500                                                 BTRFS_COMPARE_TREE_DELETED,
5501                                                 ctx);
5502                                 if (ret < 0)
5503                                         goto out;
5504                         }
5505                         advance_right = ADVANCE;
5506                         continue;
5507                 } else if (right_end_reached) {
5508                         if (left_level == 0) {
5509                                 ret = changed_cb(left_root, right_root,
5510                                                 left_path, right_path,
5511                                                 &left_key,
5512                                                 BTRFS_COMPARE_TREE_NEW,
5513                                                 ctx);
5514                                 if (ret < 0)
5515                                         goto out;
5516                         }
5517                         advance_left = ADVANCE;
5518                         continue;
5519                 }
5520
5521                 if (left_level == 0 && right_level == 0) {
5522                         cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
5523                         if (cmp < 0) {
5524                                 ret = changed_cb(left_root, right_root,
5525                                                 left_path, right_path,
5526                                                 &left_key,
5527                                                 BTRFS_COMPARE_TREE_NEW,
5528                                                 ctx);
5529                                 if (ret < 0)
5530                                         goto out;
5531                                 advance_left = ADVANCE;
5532                         } else if (cmp > 0) {
5533                                 ret = changed_cb(left_root, right_root,
5534                                                 left_path, right_path,
5535                                                 &right_key,
5536                                                 BTRFS_COMPARE_TREE_DELETED,
5537                                                 ctx);
5538                                 if (ret < 0)
5539                                         goto out;
5540                                 advance_right = ADVANCE;
5541                         } else {
5542                                 enum btrfs_compare_tree_result result;
5543
5544                                 WARN_ON(!extent_buffer_uptodate(left_path->nodes[0]));
5545                                 ret = tree_compare_item(left_path, right_path,
5546                                                         tmp_buf);
5547                                 if (ret)
5548                                         result = BTRFS_COMPARE_TREE_CHANGED;
5549                                 else
5550                                         result = BTRFS_COMPARE_TREE_SAME;
5551                                 ret = changed_cb(left_root, right_root,
5552                                                  left_path, right_path,
5553                                                  &left_key, result, ctx);
5554                                 if (ret < 0)
5555                                         goto out;
5556                                 advance_left = ADVANCE;
5557                                 advance_right = ADVANCE;
5558                         }
5559                 } else if (left_level == right_level) {
5560                         cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
5561                         if (cmp < 0) {
5562                                 advance_left = ADVANCE;
5563                         } else if (cmp > 0) {
5564                                 advance_right = ADVANCE;
5565                         } else {
5566                                 left_blockptr = btrfs_node_blockptr(
5567                                                 left_path->nodes[left_level],
5568                                                 left_path->slots[left_level]);
5569                                 right_blockptr = btrfs_node_blockptr(
5570                                                 right_path->nodes[right_level],
5571                                                 right_path->slots[right_level]);
5572                                 left_gen = btrfs_node_ptr_generation(
5573                                                 left_path->nodes[left_level],
5574                                                 left_path->slots[left_level]);
5575                                 right_gen = btrfs_node_ptr_generation(
5576                                                 right_path->nodes[right_level],
5577                                                 right_path->slots[right_level]);
5578                                 if (left_blockptr == right_blockptr &&
5579                                     left_gen == right_gen) {
5580                                         /*
5581                                          * As we're on a shared block, don't
5582                                          * allow to go deeper.
5583                                          */
5584                                         advance_left = ADVANCE_ONLY_NEXT;
5585                                         advance_right = ADVANCE_ONLY_NEXT;
5586                                 } else {
5587                                         advance_left = ADVANCE;
5588                                         advance_right = ADVANCE;
5589                                 }
5590                         }
5591                 } else if (left_level < right_level) {
5592                         advance_right = ADVANCE;
5593                 } else {
5594                         advance_left = ADVANCE;
5595                 }
5596         }
5597
5598 out:
5599         btrfs_free_path(left_path);
5600         btrfs_free_path(right_path);
5601         kvfree(tmp_buf);
5602         return ret;
5603 }
5604
5605 /*
5606  * this is similar to btrfs_next_leaf, but does not try to preserve
5607  * and fixup the path.  It looks for and returns the next key in the
5608  * tree based on the current path and the min_trans parameters.
5609  *
5610  * 0 is returned if another key is found, < 0 if there are any errors
5611  * and 1 is returned if there are no higher keys in the tree
5612  *
5613  * path->keep_locks should be set to 1 on the search made before
5614  * calling this function.
5615  */
5616 int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
5617                         struct btrfs_key *key, int level, u64 min_trans)
5618 {
5619         int slot;
5620         struct extent_buffer *c;
5621
5622         WARN_ON(!path->keep_locks);
5623         while (level < BTRFS_MAX_LEVEL) {
5624                 if (!path->nodes[level])
5625                         return 1;
5626
5627                 slot = path->slots[level] + 1;
5628                 c = path->nodes[level];
5629 next:
5630                 if (slot >= btrfs_header_nritems(c)) {
5631                         int ret;
5632                         int orig_lowest;
5633                         struct btrfs_key cur_key;
5634                         if (level + 1 >= BTRFS_MAX_LEVEL ||
5635                             !path->nodes[level + 1])
5636                                 return 1;
5637
5638                         if (path->locks[level + 1]) {
5639                                 level++;
5640                                 continue;
5641                         }
5642
5643                         slot = btrfs_header_nritems(c) - 1;
5644                         if (level == 0)
5645                                 btrfs_item_key_to_cpu(c, &cur_key, slot);
5646                         else
5647                                 btrfs_node_key_to_cpu(c, &cur_key, slot);
5648
5649                         orig_lowest = path->lowest_level;
5650                         btrfs_release_path(path);
5651                         path->lowest_level = level;
5652                         ret = btrfs_search_slot(NULL, root, &cur_key, path,
5653                                                 0, 0);
5654                         path->lowest_level = orig_lowest;
5655                         if (ret < 0)
5656                                 return ret;
5657
5658                         c = path->nodes[level];
5659                         slot = path->slots[level];
5660                         if (ret == 0)
5661                                 slot++;
5662                         goto next;
5663                 }
5664
5665                 if (level == 0)
5666                         btrfs_item_key_to_cpu(c, key, slot);
5667                 else {
5668                         u64 gen = btrfs_node_ptr_generation(c, slot);
5669
5670                         if (gen < min_trans) {
5671                                 slot++;
5672                                 goto next;
5673                         }
5674                         btrfs_node_key_to_cpu(c, key, slot);
5675                 }
5676                 return 0;
5677         }
5678         return 1;
5679 }
5680
5681 /*
5682  * search the tree again to find a leaf with greater keys
5683  * returns 0 if it found something or 1 if there are no greater leaves.
5684  * returns < 0 on io errors.
5685  */
5686 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
5687 {
5688         return btrfs_next_old_leaf(root, path, 0);
5689 }
5690
5691 int btrfs_next_old_leaf(struct btrfs_root *root, struct btrfs_path *path,
5692                         u64 time_seq)
5693 {
5694         int slot;
5695         int level;
5696         struct extent_buffer *c;
5697         struct extent_buffer *next;
5698         struct btrfs_key key;
5699         u32 nritems;
5700         int ret;
5701         int old_spinning = path->leave_spinning;
5702         int next_rw_lock = 0;
5703
5704         nritems = btrfs_header_nritems(path->nodes[0]);
5705         if (nritems == 0)
5706                 return 1;
5707
5708         btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
5709 again:
5710         level = 1;
5711         next = NULL;
5712         next_rw_lock = 0;
5713         btrfs_release_path(path);
5714
5715         path->keep_locks = 1;
5716         path->leave_spinning = 1;
5717
5718         if (time_seq)
5719                 ret = btrfs_search_old_slot(root, &key, path, time_seq);
5720         else
5721                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5722         path->keep_locks = 0;
5723
5724         if (ret < 0)
5725                 return ret;
5726
5727         nritems = btrfs_header_nritems(path->nodes[0]);
5728         /*
5729          * by releasing the path above we dropped all our locks.  A balance
5730          * could have added more items next to the key that used to be
5731          * at the very end of the block.  So, check again here and
5732          * advance the path if there are now more items available.
5733          */
5734         if (nritems > 0 && path->slots[0] < nritems - 1) {
5735                 if (ret == 0)
5736                         path->slots[0]++;
5737                 ret = 0;
5738                 goto done;
5739         }
5740         /*
5741          * So the above check misses one case:
5742          * - after releasing the path above, someone has removed the item that
5743          *   used to be at the very end of the block, and balance between leafs
5744          *   gets another one with bigger key.offset to replace it.
5745          *
5746          * This one should be returned as well, or we can get leaf corruption
5747          * later(esp. in __btrfs_drop_extents()).
5748          *
5749          * And a bit more explanation about this check,
5750          * with ret > 0, the key isn't found, the path points to the slot
5751          * where it should be inserted, so the path->slots[0] item must be the
5752          * bigger one.
5753          */
5754         if (nritems > 0 && ret > 0 && path->slots[0] == nritems - 1) {
5755                 ret = 0;
5756                 goto done;
5757         }
5758
5759         while (level < BTRFS_MAX_LEVEL) {
5760                 if (!path->nodes[level]) {
5761                         ret = 1;
5762                         goto done;
5763                 }
5764
5765                 slot = path->slots[level] + 1;
5766                 c = path->nodes[level];
5767                 if (slot >= btrfs_header_nritems(c)) {
5768                         level++;
5769                         if (level == BTRFS_MAX_LEVEL) {
5770                                 ret = 1;
5771                                 goto done;
5772                         }
5773                         continue;
5774                 }
5775
5776                 if (next) {
5777                         btrfs_tree_unlock_rw(next, next_rw_lock);
5778                         free_extent_buffer(next);
5779                 }
5780
5781                 next = c;
5782                 next_rw_lock = path->locks[level];
5783                 ret = read_block_for_search(root, path, &next, level,
5784                                             slot, &key);
5785                 if (ret == -EAGAIN)
5786                         goto again;
5787
5788                 if (ret < 0) {
5789                         btrfs_release_path(path);
5790                         goto done;
5791                 }
5792
5793                 if (!path->skip_locking) {
5794                         ret = btrfs_try_tree_read_lock(next);
5795                         if (!ret && time_seq) {
5796                                 /*
5797                                  * If we don't get the lock, we may be racing
5798                                  * with push_leaf_left, holding that lock while
5799                                  * itself waiting for the leaf we've currently
5800                                  * locked. To solve this situation, we give up
5801                                  * on our lock and cycle.
5802                                  */
5803                                 free_extent_buffer(next);
5804                                 btrfs_release_path(path);
5805                                 cond_resched();
5806                                 goto again;
5807                         }
5808                         if (!ret) {
5809                                 btrfs_set_path_blocking(path);
5810                                 btrfs_tree_read_lock(next);
5811                                 btrfs_clear_path_blocking(path, next,
5812                                                           BTRFS_READ_LOCK);
5813                         }
5814                         next_rw_lock = BTRFS_READ_LOCK;
5815                 }
5816                 break;
5817         }
5818         path->slots[level] = slot;
5819         while (1) {
5820                 level--;
5821                 c = path->nodes[level];
5822                 if (path->locks[level])
5823                         btrfs_tree_unlock_rw(c, path->locks[level]);
5824
5825                 free_extent_buffer(c);
5826                 path->nodes[level] = next;
5827                 path->slots[level] = 0;
5828                 if (!path->skip_locking)
5829                         path->locks[level] = next_rw_lock;
5830                 if (!level)
5831                         break;
5832
5833                 ret = read_block_for_search(root, path, &next, level,
5834                                             0, &key);
5835                 if (ret == -EAGAIN)
5836                         goto again;
5837
5838                 if (ret < 0) {
5839                         btrfs_release_path(path);
5840                         goto done;
5841                 }
5842
5843                 if (!path->skip_locking) {
5844                         ret = btrfs_try_tree_read_lock(next);
5845                         if (!ret) {
5846                                 btrfs_set_path_blocking(path);
5847                                 btrfs_tree_read_lock(next);
5848                                 btrfs_clear_path_blocking(path, next,
5849                                                           BTRFS_READ_LOCK);
5850                         }
5851                         next_rw_lock = BTRFS_READ_LOCK;
5852                 }
5853         }
5854         ret = 0;
5855 done:
5856         unlock_up(path, 0, 1, 0, NULL);
5857         path->leave_spinning = old_spinning;
5858         if (!old_spinning)
5859                 btrfs_set_path_blocking(path);
5860
5861         return ret;
5862 }
5863
5864 /*
5865  * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
5866  * searching until it gets past min_objectid or finds an item of 'type'
5867  *
5868  * returns 0 if something is found, 1 if nothing was found and < 0 on error
5869  */
5870 int btrfs_previous_item(struct btrfs_root *root,
5871                         struct btrfs_path *path, u64 min_objectid,
5872                         int type)
5873 {
5874         struct btrfs_key found_key;
5875         struct extent_buffer *leaf;
5876         u32 nritems;
5877         int ret;
5878
5879         while (1) {
5880                 if (path->slots[0] == 0) {
5881                         btrfs_set_path_blocking(path);
5882                         ret = btrfs_prev_leaf(root, path);
5883                         if (ret != 0)
5884                                 return ret;
5885                 } else {
5886                         path->slots[0]--;
5887                 }
5888                 leaf = path->nodes[0];
5889                 nritems = btrfs_header_nritems(leaf);
5890                 if (nritems == 0)
5891                         return 1;
5892                 if (path->slots[0] == nritems)
5893                         path->slots[0]--;
5894
5895                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5896                 if (found_key.objectid < min_objectid)
5897                         break;
5898                 if (found_key.type == type)
5899                         return 0;
5900                 if (found_key.objectid == min_objectid &&
5901                     found_key.type < type)
5902                         break;
5903         }
5904         return 1;
5905 }
5906
5907 /*
5908  * search in extent tree to find a previous Metadata/Data extent item with
5909  * min objecitd.
5910  *
5911  * returns 0 if something is found, 1 if nothing was found and < 0 on error
5912  */
5913 int btrfs_previous_extent_item(struct btrfs_root *root,
5914                         struct btrfs_path *path, u64 min_objectid)
5915 {
5916         struct btrfs_key found_key;
5917         struct extent_buffer *leaf;
5918         u32 nritems;
5919         int ret;
5920
5921         while (1) {
5922                 if (path->slots[0] == 0) {
5923                         btrfs_set_path_blocking(path);
5924                         ret = btrfs_prev_leaf(root, path);
5925                         if (ret != 0)
5926                                 return ret;
5927                 } else {
5928                         path->slots[0]--;
5929                 }
5930                 leaf = path->nodes[0];
5931                 nritems = btrfs_header_nritems(leaf);
5932                 if (nritems == 0)
5933                         return 1;
5934                 if (path->slots[0] == nritems)
5935                         path->slots[0]--;
5936
5937                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5938                 if (found_key.objectid < min_objectid)
5939                         break;
5940                 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
5941                     found_key.type == BTRFS_METADATA_ITEM_KEY)
5942                         return 0;
5943                 if (found_key.objectid == min_objectid &&
5944                     found_key.type < BTRFS_EXTENT_ITEM_KEY)
5945                         break;
5946         }
5947         return 1;
5948 }