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