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