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