fs/btrfs/ctree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007,2008 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/slab.h>
   8 #include <linux/rbtree.h>
   9 #include <linux/mm.h>
  10 #include "ctree.h"
  11 #include "disk-io.h"
  12 #include "transaction.h"
  13 #include "print-tree.h"
  14 #include "locking.h"
  15
  16 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  17                       *root, struct btrfs_path *path, int level);
  18 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  19                       const struct btrfs_key *ins_key, struct btrfs_path *path,
  20                       int data_size, int extend);
  21 static int push_node_left(struct btrfs_trans_handle *trans,
  22                           struct btrfs_fs_info *fs_info,
  23                           struct extent_buffer *dst,
  24                           struct extent_buffer *src, int empty);
  25 static int balance_node_right(struct btrfs_trans_handle *trans,
  26                               struct btrfs_fs_info *fs_info,
  27                               struct extent_buffer *dst_buf,
  28                               struct extent_buffer *src_buf);
  29 static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
  30                     int level, int slot);
  31
  32 struct btrfs_path *btrfs_alloc_path(void)
  33 {
  34         return kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
  35 }
  36
  37 /*
  38  * set all locked nodes in the path to blocking locks.  This should
  39  * be done before scheduling
  40  */
  41 noinline void btrfs_set_path_blocking(struct btrfs_path *p)
  42 {
  43         int i;
  44         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  45                 if (!p->nodes[i] || !p->locks[i])
  46                         continue;
  47                 btrfs_set_lock_blocking_rw(p->nodes[i], p->locks[i]);
  48                 if (p->locks[i] == BTRFS_READ_LOCK)
  49                         p->locks[i] = BTRFS_READ_LOCK_BLOCKING;
  50                 else if (p->locks[i] == BTRFS_WRITE_LOCK)
  51                         p->locks[i] = BTRFS_WRITE_LOCK_BLOCKING;
  52         }
  53 }
  54
  55 /* this also releases the path */
  56 void btrfs_free_path(struct btrfs_path *p)
  57 {
  58         if (!p)
  59                 return;
  60         btrfs_release_path(p);
  61         kmem_cache_free(btrfs_path_cachep, p);
  62 }
  63
  64 /*
  65  * path release drops references on the extent buffers in the path
  66  * and it drops any locks held by this path
  67  *
  68  * It is safe to call this on paths that no locks or extent buffers held.
  69  */
  70 noinline void btrfs_release_path(struct btrfs_path *p)
  71 {
  72         int i;
  73
  74         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  75                 p->slots[i] = 0;
  76                 if (!p->nodes[i])
  77                         continue;
  78                 if (p->locks[i]) {
  79                         btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
  80                         p->locks[i] = 0;
  81                 }
  82                 free_extent_buffer(p->nodes[i]);
  83                 p->nodes[i] = NULL;
  84         }
  85 }
  86
  87 /*
  88  * safely gets a reference on the root node of a tree.  A lock
  89  * is not taken, so a concurrent writer may put a different node
  90  * at the root of the tree.  See btrfs_lock_root_node for the
  91  * looping required.
  92  *
  93  * The extent buffer returned by this has a reference taken, so
  94  * it won't disappear.  It may stop being the root of the tree
  95  * at any time because there are no locks held.
  96  */
  97 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
  98 {
  99         struct extent_buffer *eb;
 100
 101         while (1) {
 102                 rcu_read_lock();
 103                 eb = rcu_dereference(root->node);
 104
 105                 /*
 106                  * RCU really hurts here, we could free up the root node because
 107                  * it was COWed but we may not get the new root node yet so do
 108                  * the inc_not_zero dance and if it doesn't work then
 109                  * synchronize_rcu and try again.
 110                  */
 111                 if (atomic_inc_not_zero(&eb->refs)) {
 112                         rcu_read_unlock();
 113                         break;
 114                 }
 115                 rcu_read_unlock();
 116                 synchronize_rcu();
 117         }
 118         return eb;
 119 }
 120
 121 /* loop around taking references on and locking the root node of the
 122  * tree until you end up with a lock on the root.  A locked buffer
 123  * is returned, with a reference held.
 124  */
 125 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
 126 {
 127         struct extent_buffer *eb;
 128
 129         while (1) {
 130                 eb = btrfs_root_node(root);
 131                 btrfs_tree_lock(eb);
 132                 if (eb == root->node)
 133                         break;
 134                 btrfs_tree_unlock(eb);
 135                 free_extent_buffer(eb);
 136         }
 137         return eb;
 138 }
 139
 140 /* loop around taking references on and locking the root node of the
 141  * tree until you end up with a lock on the root.  A locked buffer
 142  * is returned, with a reference held.
 143  */
 144 struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
 145 {
 146         struct extent_buffer *eb;
 147
 148         while (1) {
 149                 eb = btrfs_root_node(root);
 150                 btrfs_tree_read_lock(eb);
 151                 if (eb == root->node)
 152                         break;
 153                 btrfs_tree_read_unlock(eb);
 154                 free_extent_buffer(eb);
 155         }
 156         return eb;
 157 }
 158
 159 /* cowonly root (everything not a reference counted cow subvolume), just get
 160  * put onto a simple dirty list.  transaction.c walks this to make sure they
 161  * get properly updated on disk.
 162  */
 163 static void add_root_to_dirty_list(struct btrfs_root *root)
 164 {
 165         struct btrfs_fs_info *fs_info = root->fs_info;
 166
 167         if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
 168             !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
 169                 return;
 170
 171         spin_lock(&fs_info->trans_lock);
 172         if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
 173                 /* Want the extent tree to be the last on the list */
 174                 if (root->root_key.objectid == BTRFS_EXTENT_TREE_OBJECTID)
 175                         list_move_tail(&root->dirty_list,
 176                                        &fs_info->dirty_cowonly_roots);
 177                 else
 178                         list_move(&root->dirty_list,
 179                                   &fs_info->dirty_cowonly_roots);
 180         }
 181         spin_unlock(&fs_info->trans_lock);
 182 }
 183
 184 /*
 185  * used by snapshot creation to make a copy of a root for a tree with
 186  * a given objectid.  The buffer with the new root node is returned in
 187  * cow_ret, and this func returns zero on success or a negative error code.
 188  */
 189 int btrfs_copy_root(struct btrfs_trans_handle *trans,
 190                       struct btrfs_root *root,
 191                       struct extent_buffer *buf,
 192                       struct extent_buffer **cow_ret, u64 new_root_objectid)
 193 {
 194         struct btrfs_fs_info *fs_info = root->fs_info;
 195         struct extent_buffer *cow;
 196         int ret = 0;
 197         int level;
 198         struct btrfs_disk_key disk_key;
 199
 200         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 201                 trans->transid != fs_info->running_transaction->transid);
 202         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 203                 trans->transid != root->last_trans);
 204
 205         level = btrfs_header_level(buf);
 206         if (level == 0)
 207                 btrfs_item_key(buf, &disk_key, 0);
 208         else
 209                 btrfs_node_key(buf, &disk_key, 0);
 210
 211         cow = btrfs_alloc_tree_block(trans, root, 0, new_root_objectid,
 212                         &disk_key, level, buf->start, 0);
 213         if (IS_ERR(cow))
 214                 return PTR_ERR(cow);
 215
 216         copy_extent_buffer_full(cow, buf);
 217         btrfs_set_header_bytenr(cow, cow->start);
 218         btrfs_set_header_generation(cow, trans->transid);
 219         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 220         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 221                                      BTRFS_HEADER_FLAG_RELOC);
 222         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 223                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 224         else
 225                 btrfs_set_header_owner(cow, new_root_objectid);
 226
 227         write_extent_buffer_fsid(cow, fs_info->fsid);
 228
 229         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 230         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 231                 ret = btrfs_inc_ref(trans, root, cow, 1);
 232         else
 233                 ret = btrfs_inc_ref(trans, root, cow, 0);
 234
 235         if (ret)
 236                 return ret;
 237
 238         btrfs_mark_buffer_dirty(cow);
 239         *cow_ret = cow;
 240         return 0;
 241 }
 242
 243 enum mod_log_op {
 244         MOD_LOG_KEY_REPLACE,
 245         MOD_LOG_KEY_ADD,
 246         MOD_LOG_KEY_REMOVE,
 247         MOD_LOG_KEY_REMOVE_WHILE_FREEING,
 248         MOD_LOG_KEY_REMOVE_WHILE_MOVING,
 249         MOD_LOG_MOVE_KEYS,
 250         MOD_LOG_ROOT_REPLACE,
 251 };
 252
 253 struct tree_mod_root {
 254         u64 logical;
 255         u8 level;
 256 };
 257
 258 struct tree_mod_elem {
 259         struct rb_node node;
 260         u64 logical;
 261         u64 seq;
 262         enum mod_log_op op;
 263
 264         /* this is used for MOD_LOG_KEY_* and MOD_LOG_MOVE_KEYS operations */
 265         int slot;
 266
 267         /* this is used for MOD_LOG_KEY* and MOD_LOG_ROOT_REPLACE */
 268         u64 generation;
 269
 270         /* those are used for op == MOD_LOG_KEY_{REPLACE,REMOVE} */
 271         struct btrfs_disk_key key;
 272         u64 blockptr;
 273
 274         /* this is used for op == MOD_LOG_MOVE_KEYS */
 275         struct {
 276                 int dst_slot;
 277                 int nr_items;
 278         } move;
 279
 280         /* this is used for op == MOD_LOG_ROOT_REPLACE */
 281         struct tree_mod_root old_root;
 282 };
 283
 284 /*
 285  * Pull a new tree mod seq number for our operation.
 286  */
 287 static inline u64 btrfs_inc_tree_mod_seq(struct btrfs_fs_info *fs_info)
 288 {
 289         return atomic64_inc_return(&fs_info->tree_mod_seq);
 290 }
 291
 292 /*
 293  * This adds a new blocker to the tree mod log's blocker list if the @elem
 294  * passed does not already have a sequence number set. So when a caller expects
 295  * to record tree modifications, it should ensure to set elem->seq to zero
 296  * before calling btrfs_get_tree_mod_seq.
 297  * Returns a fresh, unused tree log modification sequence number, even if no new
 298  * blocker was added.
 299  */
 300 u64 btrfs_get_tree_mod_seq(struct btrfs_fs_info *fs_info,
 301                            struct seq_list *elem)
 302 {
 303         write_lock(&fs_info->tree_mod_log_lock);
 304         spin_lock(&fs_info->tree_mod_seq_lock);
 305         if (!elem->seq) {
 306                 elem->seq = btrfs_inc_tree_mod_seq(fs_info);
 307                 list_add_tail(&elem->list, &fs_info->tree_mod_seq_list);
 308         }
 309         spin_unlock(&fs_info->tree_mod_seq_lock);
 310         write_unlock(&fs_info->tree_mod_log_lock);
 311
 312         return elem->seq;
 313 }
 314
 315 void btrfs_put_tree_mod_seq(struct btrfs_fs_info *fs_info,
 316                             struct seq_list *elem)
 317 {
 318         struct rb_root *tm_root;
 319         struct rb_node *node;
 320         struct rb_node *next;
 321         struct seq_list *cur_elem;
 322         struct tree_mod_elem *tm;
 323         u64 min_seq = (u64)-1;
 324         u64 seq_putting = elem->seq;
 325
 326         if (!seq_putting)
 327                 return;
 328
 329         spin_lock(&fs_info->tree_mod_seq_lock);
 330         list_del(&elem->list);
 331         elem->seq = 0;
 332
 333         list_for_each_entry(cur_elem, &fs_info->tree_mod_seq_list, list) {
 334                 if (cur_elem->seq < min_seq) {
 335                         if (seq_putting > cur_elem->seq) {
 336                                 /*
 337                                  * blocker with lower sequence number exists, we
 338                                  * cannot remove anything from the log
 339                                  */
 340                                 spin_unlock(&fs_info->tree_mod_seq_lock);
 341                                 return;
 342                         }
 343                         min_seq = cur_elem->seq;
 344                 }
 345         }
 346         spin_unlock(&fs_info->tree_mod_seq_lock);
 347
 348         /*
 349          * anything that's lower than the lowest existing (read: blocked)
 350          * sequence number can be removed from the tree.
 351          */
 352         write_lock(&fs_info->tree_mod_log_lock);
 353         tm_root = &fs_info->tree_mod_log;
 354         for (node = rb_first(tm_root); node; node = next) {
 355                 next = rb_next(node);
 356                 tm = rb_entry(node, struct tree_mod_elem, node);
 357                 if (tm->seq > min_seq)
 358                         continue;
 359                 rb_erase(node, tm_root);
 360                 kfree(tm);
 361         }
 362         write_unlock(&fs_info->tree_mod_log_lock);
 363 }
 364
 365 /*
 366  * key order of the log:
 367  *       node/leaf start address -> sequence
 368  *
 369  * The 'start address' is the logical address of the *new* root node
 370  * for root replace operations, or the logical address of the affected
 371  * block for all other operations.
 372  *
 373  * Note: must be called with write lock for fs_info::tree_mod_log_lock.
 374  */
 375 static noinline int
 376 __tree_mod_log_insert(struct btrfs_fs_info *fs_info, struct tree_mod_elem *tm)
 377 {
 378         struct rb_root *tm_root;
 379         struct rb_node **new;
 380         struct rb_node *parent = NULL;
 381         struct tree_mod_elem *cur;
 382
 383         tm->seq = btrfs_inc_tree_mod_seq(fs_info);
 384
 385         tm_root = &fs_info->tree_mod_log;
 386         new = &tm_root->rb_node;
 387         while (*new) {
 388                 cur = rb_entry(*new, struct tree_mod_elem, node);
 389                 parent = *new;
 390                 if (cur->logical < tm->logical)
 391                         new = &((*new)->rb_left);
 392                 else if (cur->logical > tm->logical)
 393                         new = &((*new)->rb_right);
 394                 else if (cur->seq < tm->seq)
 395                         new = &((*new)->rb_left);
 396                 else if (cur->seq > tm->seq)
 397                         new = &((*new)->rb_right);
 398                 else
 399                         return -EEXIST;
 400         }
 401
 402         rb_link_node(&tm->node, parent, new);
 403         rb_insert_color(&tm->node, tm_root);
 404         return 0;
 405 }
 406
 407 /*
 408  * Determines if logging can be omitted. Returns 1 if it can. Otherwise, it
 409  * returns zero with the tree_mod_log_lock acquired. The caller must hold
 410  * this until all tree mod log insertions are recorded in the rb tree and then
 411  * write unlock fs_info::tree_mod_log_lock.
 412  */
 413 static inline int tree_mod_dont_log(struct btrfs_fs_info *fs_info,
 414                                     struct extent_buffer *eb) {
 415         smp_mb();
 416         if (list_empty(&(fs_info)->tree_mod_seq_list))
 417                 return 1;
 418         if (eb && btrfs_header_level(eb) == 0)
 419                 return 1;
 420
 421         write_lock(&fs_info->tree_mod_log_lock);
 422         if (list_empty(&(fs_info)->tree_mod_seq_list)) {
 423                 write_unlock(&fs_info->tree_mod_log_lock);
 424                 return 1;
 425         }
 426
 427         return 0;
 428 }
 429
 430 /* Similar to tree_mod_dont_log, but doesn't acquire any locks. */
 431 static inline int tree_mod_need_log(const struct btrfs_fs_info *fs_info,
 432                                     struct extent_buffer *eb)
 433 {
 434         smp_mb();
 435         if (list_empty(&(fs_info)->tree_mod_seq_list))
 436                 return 0;
 437         if (eb && btrfs_header_level(eb) == 0)
 438                 return 0;
 439
 440         return 1;
 441 }
 442
 443 static struct tree_mod_elem *
 444 alloc_tree_mod_elem(struct extent_buffer *eb, int slot,
 445                     enum mod_log_op op, gfp_t flags)
 446 {
 447         struct tree_mod_elem *tm;
 448
 449         tm = kzalloc(sizeof(*tm), flags);
 450         if (!tm)
 451                 return NULL;
 452
 453         tm->logical = eb->start;
 454         if (op != MOD_LOG_KEY_ADD) {
 455                 btrfs_node_key(eb, &tm->key, slot);
 456                 tm->blockptr = btrfs_node_blockptr(eb, slot);
 457         }
 458         tm->op = op;
 459         tm->slot = slot;
 460         tm->generation = btrfs_node_ptr_generation(eb, slot);
 461         RB_CLEAR_NODE(&tm->node);
 462
 463         return tm;
 464 }
 465
 466 static noinline int tree_mod_log_insert_key(struct extent_buffer *eb, int slot,
 467                 enum mod_log_op op, gfp_t flags)
 468 {
 469         struct tree_mod_elem *tm;
 470         int ret;
 471
 472         if (!tree_mod_need_log(eb->fs_info, eb))
 473                 return 0;
 474
 475         tm = alloc_tree_mod_elem(eb, slot, op, flags);
 476         if (!tm)
 477                 return -ENOMEM;
 478
 479         if (tree_mod_dont_log(eb->fs_info, eb)) {
 480                 kfree(tm);
 481                 return 0;
 482         }
 483
 484         ret = __tree_mod_log_insert(eb->fs_info, tm);
 485         write_unlock(&eb->fs_info->tree_mod_log_lock);
 486         if (ret)
 487                 kfree(tm);
 488
 489         return ret;
 490 }
 491
 492 static noinline int tree_mod_log_insert_move(struct extent_buffer *eb,
 493                 int dst_slot, int src_slot, int nr_items)
 494 {
 495         struct tree_mod_elem *tm = NULL;
 496         struct tree_mod_elem **tm_list = NULL;
 497         int ret = 0;
 498         int i;
 499         int locked = 0;
 500
 501         if (!tree_mod_need_log(eb->fs_info, eb))
 502                 return 0;
 503
 504         tm_list = kcalloc(nr_items, sizeof(struct tree_mod_elem *), GFP_NOFS);
 505         if (!tm_list)
 506                 return -ENOMEM;
 507
 508         tm = kzalloc(sizeof(*tm), GFP_NOFS);
 509         if (!tm) {
 510                 ret = -ENOMEM;
 511                 goto free_tms;
 512         }
 513
 514         tm->logical = eb->start;
 515         tm->slot = src_slot;
 516         tm->move.dst_slot = dst_slot;
 517         tm->move.nr_items = nr_items;
 518         tm->op = MOD_LOG_MOVE_KEYS;
 519
 520         for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
 521                 tm_list[i] = alloc_tree_mod_elem(eb, i + dst_slot,
 522                     MOD_LOG_KEY_REMOVE_WHILE_MOVING, GFP_NOFS);
 523                 if (!tm_list[i]) {
 524                         ret = -ENOMEM;
 525                         goto free_tms;
 526                 }
 527         }
 528
 529         if (tree_mod_dont_log(eb->fs_info, eb))
 530                 goto free_tms;
 531         locked = 1;
 532
 533         /*
 534          * When we override something during the move, we log these removals.
 535          * This can only happen when we move towards the beginning of the
 536          * buffer, i.e. dst_slot < src_slot.
 537          */
 538         for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
 539                 ret = __tree_mod_log_insert(eb->fs_info, tm_list[i]);
 540                 if (ret)
 541                         goto free_tms;
 542         }
 543
 544         ret = __tree_mod_log_insert(eb->fs_info, tm);
 545         if (ret)
 546                 goto free_tms;
 547         write_unlock(&eb->fs_info->tree_mod_log_lock);
 548         kfree(tm_list);
 549
 550         return 0;
 551 free_tms:
 552         for (i = 0; i < nr_items; i++) {
 553                 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
 554                         rb_erase(&tm_list[i]->node, &eb->fs_info->tree_mod_log);
 555                 kfree(tm_list[i]);
 556         }
 557         if (locked)
 558                 write_unlock(&eb->fs_info->tree_mod_log_lock);
 559         kfree(tm_list);
 560         kfree(tm);
 561
 562         return ret;
 563 }
 564
 565 static inline int
 566 __tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
 567                        struct tree_mod_elem **tm_list,
 568                        int nritems)
 569 {
 570         int i, j;
 571         int ret;
 572
 573         for (i = nritems - 1; i >= 0; i--) {
 574                 ret = __tree_mod_log_insert(fs_info, tm_list[i]);
 575                 if (ret) {
 576                         for (j = nritems - 1; j > i; j--)
 577                                 rb_erase(&tm_list[j]->node,
 578                                          &fs_info->tree_mod_log);
 579                         return ret;
 580                 }
 581         }
 582
 583         return 0;
 584 }
 585
 586 static noinline int tree_mod_log_insert_root(struct extent_buffer *old_root,
 587                          struct extent_buffer *new_root, int log_removal)
 588 {
 589         struct btrfs_fs_info *fs_info = old_root->fs_info;
 590         struct tree_mod_elem *tm = NULL;
 591         struct tree_mod_elem **tm_list = NULL;
 592         int nritems = 0;
 593         int ret = 0;
 594         int i;
 595
 596         if (!tree_mod_need_log(fs_info, NULL))
 597                 return 0;
 598
 599         if (log_removal && btrfs_header_level(old_root) > 0) {
 600                 nritems = btrfs_header_nritems(old_root);
 601                 tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *),
 602                                   GFP_NOFS);
 603                 if (!tm_list) {
 604                         ret = -ENOMEM;
 605                         goto free_tms;
 606                 }
 607                 for (i = 0; i < nritems; i++) {
 608                         tm_list[i] = alloc_tree_mod_elem(old_root, i,
 609                             MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
 610                         if (!tm_list[i]) {
 611                                 ret = -ENOMEM;
 612                                 goto free_tms;
 613                         }
 614                 }
 615         }
 616
 617         tm = kzalloc(sizeof(*tm), GFP_NOFS);
 618         if (!tm) {
 619                 ret = -ENOMEM;
 620                 goto free_tms;
 621         }
 622
 623         tm->logical = new_root->start;
 624         tm->old_root.logical = old_root->start;
 625         tm->old_root.level = btrfs_header_level(old_root);
 626         tm->generation = btrfs_header_generation(old_root);
 627         tm->op = MOD_LOG_ROOT_REPLACE;
 628
 629         if (tree_mod_dont_log(fs_info, NULL))
 630                 goto free_tms;
 631
 632         if (tm_list)
 633                 ret = __tree_mod_log_free_eb(fs_info, tm_list, nritems);
 634         if (!ret)
 635                 ret = __tree_mod_log_insert(fs_info, tm);
 636
 637         write_unlock(&fs_info->tree_mod_log_lock);
 638         if (ret)
 639                 goto free_tms;
 640         kfree(tm_list);
 641
 642         return ret;
 643
 644 free_tms:
 645         if (tm_list) {
 646                 for (i = 0; i < nritems; i++)
 647                         kfree(tm_list[i]);
 648                 kfree(tm_list);
 649         }
 650         kfree(tm);
 651
 652         return ret;
 653 }
 654
 655 static struct tree_mod_elem *
 656 __tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq,
 657                       int smallest)
 658 {
 659         struct rb_root *tm_root;
 660         struct rb_node *node;
 661         struct tree_mod_elem *cur = NULL;
 662         struct tree_mod_elem *found = NULL;
 663
 664         read_lock(&fs_info->tree_mod_log_lock);
 665         tm_root = &fs_info->tree_mod_log;
 666         node = tm_root->rb_node;
 667         while (node) {
 668                 cur = rb_entry(node, struct tree_mod_elem, node);
 669                 if (cur->logical < start) {
 670                         node = node->rb_left;
 671                 } else if (cur->logical > start) {
 672                         node = node->rb_right;
 673                 } else if (cur->seq < min_seq) {
 674                         node = node->rb_left;
 675                 } else if (!smallest) {
 676                         /* we want the node with the highest seq */
 677                         if (found)
 678                                 BUG_ON(found->seq > cur->seq);
 679                         found = cur;
 680                         node = node->rb_left;
 681                 } else if (cur->seq > min_seq) {
 682                         /* we want the node with the smallest seq */
 683                         if (found)
 684                                 BUG_ON(found->seq < cur->seq);
 685                         found = cur;
 686                         node = node->rb_right;
 687                 } else {
 688                         found = cur;
 689                         break;
 690                 }
 691         }
 692         read_unlock(&fs_info->tree_mod_log_lock);
 693
 694         return found;
 695 }
 696
 697 /*
 698  * this returns the element from the log with the smallest time sequence
 699  * value that's in the log (the oldest log item). any element with a time
 700  * sequence lower than min_seq will be ignored.
 701  */
 702 static struct tree_mod_elem *
 703 tree_mod_log_search_oldest(struct btrfs_fs_info *fs_info, u64 start,
 704                            u64 min_seq)
 705 {
 706         return __tree_mod_log_search(fs_info, start, min_seq, 1);
 707 }
 708
 709 /*
 710  * this returns the element from the log with the largest time sequence
 711  * value that's in the log (the most recent log item). any element with
 712  * a time sequence lower than min_seq will be ignored.
 713  */
 714 static struct tree_mod_elem *
 715 tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq)
 716 {
 717         return __tree_mod_log_search(fs_info, start, min_seq, 0);
 718 }
 719
 720 static noinline int
 721 tree_mod_log_eb_copy(struct btrfs_fs_info *fs_info, struct extent_buffer *dst,
 722                      struct extent_buffer *src, unsigned long dst_offset,
 723                      unsigned long src_offset, int nr_items)
 724 {
 725         int ret = 0;
 726         struct tree_mod_elem **tm_list = NULL;
 727         struct tree_mod_elem **tm_list_add, **tm_list_rem;
 728         int i;
 729         int locked = 0;
 730
 731         if (!tree_mod_need_log(fs_info, NULL))
 732                 return 0;
 733
 734         if (btrfs_header_level(dst) == 0 && btrfs_header_level(src) == 0)
 735                 return 0;
 736
 737         tm_list = kcalloc(nr_items * 2, sizeof(struct tree_mod_elem *),
 738                           GFP_NOFS);
 739         if (!tm_list)
 740                 return -ENOMEM;
 741
 742         tm_list_add = tm_list;
 743         tm_list_rem = tm_list + nr_items;
 744         for (i = 0; i < nr_items; i++) {
 745                 tm_list_rem[i] = alloc_tree_mod_elem(src, i + src_offset,
 746                     MOD_LOG_KEY_REMOVE, GFP_NOFS);
 747                 if (!tm_list_rem[i]) {
 748                         ret = -ENOMEM;
 749                         goto free_tms;
 750                 }
 751
 752                 tm_list_add[i] = alloc_tree_mod_elem(dst, i + dst_offset,
 753                     MOD_LOG_KEY_ADD, GFP_NOFS);
 754                 if (!tm_list_add[i]) {
 755                         ret = -ENOMEM;
 756                         goto free_tms;
 757                 }
 758         }
 759
 760         if (tree_mod_dont_log(fs_info, NULL))
 761                 goto free_tms;
 762         locked = 1;
 763
 764         for (i = 0; i < nr_items; i++) {
 765                 ret = __tree_mod_log_insert(fs_info, tm_list_rem[i]);
 766                 if (ret)
 767                         goto free_tms;
 768                 ret = __tree_mod_log_insert(fs_info, tm_list_add[i]);
 769                 if (ret)
 770                         goto free_tms;
 771         }
 772
 773         write_unlock(&fs_info->tree_mod_log_lock);
 774         kfree(tm_list);
 775
 776         return 0;
 777
 778 free_tms:
 779         for (i = 0; i < nr_items * 2; i++) {
 780                 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
 781                         rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
 782                 kfree(tm_list[i]);
 783         }
 784         if (locked)
 785                 write_unlock(&fs_info->tree_mod_log_lock);
 786         kfree(tm_list);
 787
 788         return ret;
 789 }
 790
 791 static noinline int tree_mod_log_free_eb(struct extent_buffer *eb)
 792 {
 793         struct tree_mod_elem **tm_list = NULL;
 794         int nritems = 0;
 795         int i;
 796         int ret = 0;
 797
 798         if (btrfs_header_level(eb) == 0)
 799                 return 0;
 800
 801         if (!tree_mod_need_log(eb->fs_info, NULL))
 802                 return 0;
 803
 804         nritems = btrfs_header_nritems(eb);
 805         tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *), GFP_NOFS);
 806         if (!tm_list)
 807                 return -ENOMEM;
 808
 809         for (i = 0; i < nritems; i++) {
 810                 tm_list[i] = alloc_tree_mod_elem(eb, i,
 811                     MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
 812                 if (!tm_list[i]) {
 813                         ret = -ENOMEM;
 814                         goto free_tms;
 815                 }
 816         }
 817
 818         if (tree_mod_dont_log(eb->fs_info, eb))
 819                 goto free_tms;
 820
 821         ret = __tree_mod_log_free_eb(eb->fs_info, tm_list, nritems);
 822         write_unlock(&eb->fs_info->tree_mod_log_lock);
 823         if (ret)
 824                 goto free_tms;
 825         kfree(tm_list);
 826
 827         return 0;
 828
 829 free_tms:
 830         for (i = 0; i < nritems; i++)
 831                 kfree(tm_list[i]);
 832         kfree(tm_list);
 833
 834         return ret;
 835 }
 836
 837 /*
 838  * check if the tree block can be shared by multiple trees
 839  */
 840 int btrfs_block_can_be_shared(struct btrfs_root *root,
 841                               struct extent_buffer *buf)
 842 {
 843         /*
 844          * Tree blocks not in reference counted trees and tree roots
 845          * are never shared. If a block was allocated after the last
 846          * snapshot and the block was not allocated by tree relocation,
 847          * we know the block is not shared.
 848          */
 849         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 850             buf != root->node && buf != root->commit_root &&
 851             (btrfs_header_generation(buf) <=
 852              btrfs_root_last_snapshot(&root->root_item) ||
 853              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 854                 return 1;
 855
 856         return 0;
 857 }
 858
 859 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
 860                                        struct btrfs_root *root,
 861                                        struct extent_buffer *buf,
 862                                        struct extent_buffer *cow,
 863                                        int *last_ref)
 864 {
 865         struct btrfs_fs_info *fs_info = root->fs_info;
 866         u64 refs;
 867         u64 owner;
 868         u64 flags;
 869         u64 new_flags = 0;
 870         int ret;
 871
 872         /*
 873          * Backrefs update rules:
 874          *
 875          * Always use full backrefs for extent pointers in tree block
 876          * allocated by tree relocation.
 877          *
 878          * If a shared tree block is no longer referenced by its owner
 879          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
 880          * use full backrefs for extent pointers in tree block.
 881          *
 882          * If a tree block is been relocating
 883          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
 884          * use full backrefs for extent pointers in tree block.
 885          * The reason for this is some operations (such as drop tree)
 886          * are only allowed for blocks use full backrefs.
 887          */
 888
 889         if (btrfs_block_can_be_shared(root, buf)) {
 890                 ret = btrfs_lookup_extent_info(trans, fs_info, buf->start,
 891                                                btrfs_header_level(buf), 1,
 892                                                &refs, &flags);
 893                 if (ret)
 894                         return ret;
 895                 if (refs == 0) {
 896                         ret = -EROFS;
 897                         btrfs_handle_fs_error(fs_info, ret, NULL);
 898                         return ret;
 899                 }
 900         } else {
 901                 refs = 1;
 902                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 903                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 904                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 905                 else
 906                         flags = 0;
 907         }
 908
 909         owner = btrfs_header_owner(buf);
 910         BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
 911                !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
 912
 913         if (refs > 1) {
 914                 if ((owner == root->root_key.objectid ||
 915                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
 916                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
 917                         ret = btrfs_inc_ref(trans, root, buf, 1);
 918                         if (ret)
 919                                 return ret;
 920
 921                         if (root->root_key.objectid ==
 922                             BTRFS_TREE_RELOC_OBJECTID) {
 923                                 ret = btrfs_dec_ref(trans, root, buf, 0);
 924                                 if (ret)
 925                                         return ret;
 926                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 927                                 if (ret)
 928                                         return ret;
 929                         }
 930                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
 931                 } else {
 932
 933                         if (root->root_key.objectid ==
 934                             BTRFS_TREE_RELOC_OBJECTID)
 935                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 936                         else
 937                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 938                         if (ret)
 939                                 return ret;
 940                 }
 941                 if (new_flags != 0) {
 942                         int level = btrfs_header_level(buf);
 943
 944                         ret = btrfs_set_disk_extent_flags(trans, fs_info,
 945                                                           buf->start,
 946                                                           buf->len,
 947                                                           new_flags, level, 0);
 948                         if (ret)
 949                                 return ret;
 950                 }
 951         } else {
 952                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
 953                         if (root->root_key.objectid ==
 954                             BTRFS_TREE_RELOC_OBJECTID)
 955                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 956                         else
 957                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 958                         if (ret)
 959                                 return ret;
 960                         ret = btrfs_dec_ref(trans, root, buf, 1);
 961                         if (ret)
 962                                 return ret;
 963                 }
 964                 clean_tree_block(fs_info, buf);
 965                 *last_ref = 1;
 966         }
 967         return 0;
 968 }
 969
 970 /*
 971  * does the dirty work in cow of a single block.  The parent block (if
 972  * supplied) is updated to point to the new cow copy.  The new buffer is marked
 973  * dirty and returned locked.  If you modify the block it needs to be marked
 974  * dirty again.
 975  *
 976  * search_start -- an allocation hint for the new block
 977  *
 978  * empty_size -- a hint that you plan on doing more cow.  This is the size in
 979  * bytes the allocator should try to find free next to the block it returns.
 980  * This is just a hint and may be ignored by the allocator.
 981  */
 982 static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 983                              struct btrfs_root *root,
 984                              struct extent_buffer *buf,
 985                              struct extent_buffer *parent, int parent_slot,
 986                              struct extent_buffer **cow_ret,
 987                              u64 search_start, u64 empty_size)
 988 {
 989         struct btrfs_fs_info *fs_info = root->fs_info;
 990         struct btrfs_disk_key disk_key;
 991         struct extent_buffer *cow;
 992         int level, ret;
 993         int last_ref = 0;
 994         int unlock_orig = 0;
 995         u64 parent_start = 0;
 996
 997         if (*cow_ret == buf)
 998                 unlock_orig = 1;
 999
1000         btrfs_assert_tree_locked(buf);
1001
1002         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1003                 trans->transid != fs_info->running_transaction->transid);
1004         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1005                 trans->transid != root->last_trans);
1006
1007         level = btrfs_header_level(buf);
1008
1009         if (level == 0)
1010                 btrfs_item_key(buf, &disk_key, 0);
1011         else
1012                 btrfs_node_key(buf, &disk_key, 0);
1013
1014         if ((root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) && parent)
1015                 parent_start = parent->start;
1016
1017         cow = btrfs_alloc_tree_block(trans, root, parent_start,
1018                         root->root_key.objectid, &disk_key, level,
1019                         search_start, empty_size);
1020         if (IS_ERR(cow))
1021                 return PTR_ERR(cow);
1022
1023         /* cow is set to blocking by btrfs_init_new_buffer */
1024
1025         copy_extent_buffer_full(cow, buf);
1026         btrfs_set_header_bytenr(cow, cow->start);
1027         btrfs_set_header_generation(cow, trans->transid);
1028         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
1029         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
1030                                      BTRFS_HEADER_FLAG_RELOC);
1031         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1032                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
1033         else
1034                 btrfs_set_header_owner(cow, root->root_key.objectid);
1035
1036         write_extent_buffer_fsid(cow, fs_info->fsid);
1037
1038         ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
1039         if (ret) {
1040                 btrfs_abort_transaction(trans, ret);
1041                 return ret;
1042         }
1043
1044         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
1045                 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
1046                 if (ret) {
1047                         btrfs_abort_transaction(trans, ret);
1048                         return ret;
1049                 }
1050         }
1051
1052         if (buf == root->node) {
1053                 WARN_ON(parent && parent != buf);
1054                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1055                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1056                         parent_start = buf->start;
1057
1058                 extent_buffer_get(cow);
1059                 ret = tree_mod_log_insert_root(root->node, cow, 1);
1060                 BUG_ON(ret < 0);
1061                 rcu_assign_pointer(root->node, cow);
1062
1063                 btrfs_free_tree_block(trans, root, buf, parent_start,
1064                                       last_ref);
1065                 free_extent_buffer(buf);
1066                 add_root_to_dirty_list(root);
1067         } else {
1068                 WARN_ON(trans->transid != btrfs_header_generation(parent));
1069                 tree_mod_log_insert_key(parent, parent_slot,
1070                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
1071                 btrfs_set_node_blockptr(parent, parent_slot,
1072                                         cow->start);
1073                 btrfs_set_node_ptr_generation(parent, parent_slot,
1074                                               trans->transid);
1075                 btrfs_mark_buffer_dirty(parent);
1076                 if (last_ref) {
1077                         ret = tree_mod_log_free_eb(buf);
1078                         if (ret) {
1079                                 btrfs_abort_transaction(trans, ret);
1080                                 return ret;
1081                         }
1082                 }
1083                 btrfs_free_tree_block(trans, root, buf, parent_start,
1084                                       last_ref);
1085         }
1086         if (unlock_orig)
1087                 btrfs_tree_unlock(buf);
1088         free_extent_buffer_stale(buf);
1089         btrfs_mark_buffer_dirty(cow);
1090         *cow_ret = cow;
1091         return 0;
1092 }
1093
1094 /*
1095  * returns the logical address of the oldest predecessor of the given root.
1096  * entries older than time_seq are ignored.
1097  */
1098 static struct tree_mod_elem *__tree_mod_log_oldest_root(
1099                 struct extent_buffer *eb_root, u64 time_seq)
1100 {
1101         struct tree_mod_elem *tm;
1102         struct tree_mod_elem *found = NULL;
1103         u64 root_logical = eb_root->start;
1104         int looped = 0;
1105
1106         if (!time_seq)
1107                 return NULL;
1108
1109         /*
1110          * the very last operation that's logged for a root is the
1111          * replacement operation (if it is replaced at all). this has
1112          * the logical address of the *new* root, making it the very
1113          * first operation that's logged for this root.
1114          */
1115         while (1) {
1116                 tm = tree_mod_log_search_oldest(eb_root->fs_info, root_logical,
1117                                                 time_seq);
1118                 if (!looped && !tm)
1119                         return NULL;
1120                 /*
1121                  * if there are no tree operation for the oldest root, we simply
1122                  * return it. this should only happen if that (old) root is at
1123                  * level 0.
1124                  */
1125                 if (!tm)
1126                         break;
1127
1128                 /*
1129                  * if there's an operation that's not a root replacement, we
1130                  * found the oldest version of our root. normally, we'll find a
1131                  * MOD_LOG_KEY_REMOVE_WHILE_FREEING operation here.
1132                  */
1133                 if (tm->op != MOD_LOG_ROOT_REPLACE)
1134                         break;
1135
1136                 found = tm;
1137                 root_logical = tm->old_root.logical;
1138                 looped = 1;
1139         }
1140
1141         /* if there's no old root to return, return what we found instead */
1142         if (!found)
1143                 found = tm;
1144
1145         return found;
1146 }
1147
1148 /*
1149  * tm is a pointer to the first operation to rewind within eb. then, all
1150  * previous operations will be rewound (until we reach something older than
1151  * time_seq).
1152  */
1153 static void
1154 __tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct extent_buffer *eb,
1155                       u64 time_seq, struct tree_mod_elem *first_tm)
1156 {
1157         u32 n;
1158         struct rb_node *next;
1159         struct tree_mod_elem *tm = first_tm;
1160         unsigned long o_dst;
1161         unsigned long o_src;
1162         unsigned long p_size = sizeof(struct btrfs_key_ptr);
1163
1164         n = btrfs_header_nritems(eb);
1165         read_lock(&fs_info->tree_mod_log_lock);
1166         while (tm && tm->seq >= time_seq) {
1167                 /*
1168                  * all the operations are recorded with the operator used for
1169                  * the modification. as we're going backwards, we do the
1170                  * opposite of each operation here.
1171                  */
1172                 switch (tm->op) {
1173                 case MOD_LOG_KEY_REMOVE_WHILE_FREEING:
1174                         BUG_ON(tm->slot < n);
1175                         /* Fallthrough */
1176                 case MOD_LOG_KEY_REMOVE_WHILE_MOVING:
1177                 case MOD_LOG_KEY_REMOVE:
1178                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1179                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1180                         btrfs_set_node_ptr_generation(eb, tm->slot,
1181                                                       tm->generation);
1182                         n++;
1183                         break;
1184                 case MOD_LOG_KEY_REPLACE:
1185                         BUG_ON(tm->slot >= n);
1186                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1187                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1188                         btrfs_set_node_ptr_generation(eb, tm->slot,
1189                                                       tm->generation);
1190                         break;
1191                 case MOD_LOG_KEY_ADD:
1192                         /* if a move operation is needed it's in the log */
1193                         n--;
1194                         break;
1195                 case MOD_LOG_MOVE_KEYS:
1196                         o_dst = btrfs_node_key_ptr_offset(tm->slot);
1197                         o_src = btrfs_node_key_ptr_offset(tm->move.dst_slot);
1198                         memmove_extent_buffer(eb, o_dst, o_src,
1199                                               tm->move.nr_items * p_size);
1200                         break;
1201                 case MOD_LOG_ROOT_REPLACE:
1202                         /*
1203                          * this operation is special. for roots, this must be
1204                          * handled explicitly before rewinding.
1205                          * for non-roots, this operation may exist if the node
1206                          * was a root: root A -> child B; then A gets empty and
1207                          * B is promoted to the new root. in the mod log, we'll
1208                          * have a root-replace operation for B, a tree block
1209                          * that is no root. we simply ignore that operation.
1210                          */
1211                         break;
1212                 }
1213                 next = rb_next(&tm->node);
1214                 if (!next)
1215                         break;
1216                 tm = rb_entry(next, struct tree_mod_elem, node);
1217                 if (tm->logical != first_tm->logical)
1218                         break;
1219         }
1220         read_unlock(&fs_info->tree_mod_log_lock);
1221         btrfs_set_header_nritems(eb, n);
1222 }
1223
1224 /*
1225  * Called with eb read locked. If the buffer cannot be rewound, the same buffer
1226  * is returned. If rewind operations happen, a fresh buffer is returned. The
1227  * returned buffer is always read-locked. If the returned buffer is not the
1228  * input buffer, the lock on the input buffer is released and the input buffer
1229  * is freed (its refcount is decremented).
1230  */
1231 static struct extent_buffer *
1232 tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
1233                     struct extent_buffer *eb, u64 time_seq)
1234 {
1235         struct extent_buffer *eb_rewin;
1236         struct tree_mod_elem *tm;
1237
1238         if (!time_seq)
1239                 return eb;
1240
1241         if (btrfs_header_level(eb) == 0)
1242                 return eb;
1243
1244         tm = tree_mod_log_search(fs_info, eb->start, time_seq);
1245         if (!tm)
1246                 return eb;
1247
1248         btrfs_set_path_blocking(path);
1249         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1250
1251         if (tm->op == MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1252                 BUG_ON(tm->slot != 0);
1253                 eb_rewin = alloc_dummy_extent_buffer(fs_info, eb->start);
1254                 if (!eb_rewin) {
1255                         btrfs_tree_read_unlock_blocking(eb);
1256                         free_extent_buffer(eb);
1257                         return NULL;
1258                 }
1259                 btrfs_set_header_bytenr(eb_rewin, eb->start);
1260                 btrfs_set_header_backref_rev(eb_rewin,
1261                                              btrfs_header_backref_rev(eb));
1262                 btrfs_set_header_owner(eb_rewin, btrfs_header_owner(eb));
1263                 btrfs_set_header_level(eb_rewin, btrfs_header_level(eb));
1264         } else {
1265                 eb_rewin = btrfs_clone_extent_buffer(eb);
1266                 if (!eb_rewin) {
1267                         btrfs_tree_read_unlock_blocking(eb);
1268                         free_extent_buffer(eb);
1269                         return NULL;
1270                 }
1271         }
1272
1273         btrfs_tree_read_unlock_blocking(eb);
1274         free_extent_buffer(eb);
1275
1276         extent_buffer_get(eb_rewin);
1277         btrfs_tree_read_lock(eb_rewin);
1278         __tree_mod_log_rewind(fs_info, eb_rewin, time_seq, tm);
1279         WARN_ON(btrfs_header_nritems(eb_rewin) >
1280                 BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1281
1282         return eb_rewin;
1283 }
1284
1285 /*
1286  * get_old_root() rewinds the state of @root's root node to the given @time_seq
1287  * value. If there are no changes, the current root->root_node is returned. If
1288  * anything changed in between, there's a fresh buffer allocated on which the
1289  * rewind operations are done. In any case, the returned buffer is read locked.
1290  * Returns NULL on error (with no locks held).
1291  */
1292 static inline struct extent_buffer *
1293 get_old_root(struct btrfs_root *root, u64 time_seq)
1294 {
1295         struct btrfs_fs_info *fs_info = root->fs_info;
1296         struct tree_mod_elem *tm;
1297         struct extent_buffer *eb = NULL;
1298         struct extent_buffer *eb_root;
1299         struct extent_buffer *old;
1300         struct tree_mod_root *old_root = NULL;
1301         u64 old_generation = 0;
1302         u64 logical;
1303         int level;
1304
1305         eb_root = btrfs_read_lock_root_node(root);
1306         tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1307         if (!tm)
1308                 return eb_root;
1309
1310         if (tm->op == MOD_LOG_ROOT_REPLACE) {
1311                 old_root = &tm->old_root;
1312                 old_generation = tm->generation;
1313                 logical = old_root->logical;
1314                 level = old_root->level;
1315         } else {
1316                 logical = eb_root->start;
1317                 level = btrfs_header_level(eb_root);
1318         }
1319
1320         tm = tree_mod_log_search(fs_info, logical, time_seq);
1321         if (old_root && tm && tm->op != MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1322                 btrfs_tree_read_unlock(eb_root);
1323                 free_extent_buffer(eb_root);
1324                 old = read_tree_block(fs_info, logical, 0, level, NULL);
1325                 if (WARN_ON(IS_ERR(old) || !extent_buffer_uptodate(old))) {
1326                         if (!IS_ERR(old))
1327                                 free_extent_buffer(old);
1328                         btrfs_warn(fs_info,
1329                                    "failed to read tree block %llu from get_old_root",
1330                                    logical);
1331                 } else {
1332                         eb = btrfs_clone_extent_buffer(old);
1333                         free_extent_buffer(old);
1334                 }
1335         } else if (old_root) {
1336                 btrfs_tree_read_unlock(eb_root);
1337                 free_extent_buffer(eb_root);
1338                 eb = alloc_dummy_extent_buffer(fs_info, logical);
1339         } else {
1340                 btrfs_set_lock_blocking_rw(eb_root, BTRFS_READ_LOCK);
1341                 eb = btrfs_clone_extent_buffer(eb_root);
1342                 btrfs_tree_read_unlock_blocking(eb_root);
1343                 free_extent_buffer(eb_root);
1344         }
1345
1346         if (!eb)
1347                 return NULL;
1348         extent_buffer_get(eb);
1349         btrfs_tree_read_lock(eb);
1350         if (old_root) {
1351                 btrfs_set_header_bytenr(eb, eb->start);
1352                 btrfs_set_header_backref_rev(eb, BTRFS_MIXED_BACKREF_REV);
1353                 btrfs_set_header_owner(eb, btrfs_header_owner(eb_root));
1354                 btrfs_set_header_level(eb, old_root->level);
1355                 btrfs_set_header_generation(eb, old_generation);
1356         }
1357         if (tm)
1358                 __tree_mod_log_rewind(fs_info, eb, time_seq, tm);
1359         else
1360                 WARN_ON(btrfs_header_level(eb) != 0);
1361         WARN_ON(btrfs_header_nritems(eb) > BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1362
1363         return eb;
1364 }
1365
1366 int btrfs_old_root_level(struct btrfs_root *root, u64 time_seq)
1367 {
1368         struct tree_mod_elem *tm;
1369         int level;
1370         struct extent_buffer *eb_root = btrfs_root_node(root);
1371
1372         tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1373         if (tm && tm->op == MOD_LOG_ROOT_REPLACE) {
1374                 level = tm->old_root.level;
1375         } else {
1376                 level = btrfs_header_level(eb_root);
1377         }
1378         free_extent_buffer(eb_root);
1379
1380         return level;
1381 }
1382
1383 static inline int should_cow_block(struct btrfs_trans_handle *trans,
1384                                    struct btrfs_root *root,
1385                                    struct extent_buffer *buf)
1386 {
1387         if (btrfs_is_testing(root->fs_info))
1388                 return 0;
1389
1390         /* Ensure we can see the FORCE_COW bit */
1391         smp_mb__before_atomic();
1392
1393         /*
1394          * We do not need to cow a block if
1395          * 1) this block is not created or changed in this transaction;
1396          * 2) this block does not belong to TREE_RELOC tree;
1397          * 3) the root is not forced COW.
1398          *
1399          * What is forced COW:
1400          *    when we create snapshot during committing the transaction,
1401          *    after we've finished coping src root, we must COW the shared
1402          *    block to ensure the metadata consistency.
1403          */
1404         if (btrfs_header_generation(buf) == trans->transid &&
1405             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
1406             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
1407               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
1408             !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
1409                 return 0;
1410         return 1;
1411 }
1412
1413 /*
1414  * cows a single block, see __btrfs_cow_block for the real work.
1415  * This version of it has extra checks so that a block isn't COWed more than
1416  * once per transaction, as long as it hasn't been written yet
1417  */
1418 noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
1419                     struct btrfs_root *root, struct extent_buffer *buf,
1420                     struct extent_buffer *parent, int parent_slot,
1421                     struct extent_buffer **cow_ret)
1422 {
1423         struct btrfs_fs_info *fs_info = root->fs_info;
1424         u64 search_start;
1425         int ret;
1426
1427         if (trans->transaction != fs_info->running_transaction)
1428                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1429                        trans->transid,
1430                        fs_info->running_transaction->transid);
1431
1432         if (trans->transid != fs_info->generation)
1433                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1434                        trans->transid, fs_info->generation);
1435
1436         if (!should_cow_block(trans, root, buf)) {
1437                 trans->dirty = true;
1438                 *cow_ret = buf;
1439                 return 0;
1440         }
1441
1442         search_start = buf->start & ~((u64)SZ_1G - 1);
1443
1444         if (parent)
1445                 btrfs_set_lock_blocking(parent);
1446         btrfs_set_lock_blocking(buf);
1447
1448         ret = __btrfs_cow_block(trans, root, buf, parent,
1449                                  parent_slot, cow_ret, search_start, 0);
1450
1451         trace_btrfs_cow_block(root, buf, *cow_ret);
1452
1453         return ret;
1454 }
1455
1456 /*
1457  * helper function for defrag to decide if two blocks pointed to by a
1458  * node are actually close by
1459  */
1460 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
1461 {
1462         if (blocknr < other && other - (blocknr + blocksize) < 32768)
1463                 return 1;
1464         if (blocknr > other && blocknr - (other + blocksize) < 32768)
1465                 return 1;
1466         return 0;
1467 }
1468
1469 /*
1470  * compare two keys in a memcmp fashion
1471  */
1472 static int comp_keys(const struct btrfs_disk_key *disk,
1473                      const struct btrfs_key *k2)
1474 {
1475         struct btrfs_key k1;
1476
1477         btrfs_disk_key_to_cpu(&k1, disk);
1478
1479         return btrfs_comp_cpu_keys(&k1, k2);
1480 }
1481
1482 /*
1483  * same as comp_keys only with two btrfs_key's
1484  */
1485 int btrfs_comp_cpu_keys(const struct btrfs_key *k1, const struct btrfs_key *k2)
1486 {
1487         if (k1->objectid > k2->objectid)
1488                 return 1;
1489         if (k1->objectid < k2->objectid)
1490                 return -1;
1491         if (k1->type > k2->type)
1492                 return 1;
1493         if (k1->type < k2->type)
1494                 return -1;
1495         if (k1->offset > k2->offset)
1496                 return 1;
1497         if (k1->offset < k2->offset)
1498                 return -1;
1499         return 0;
1500 }
1501
1502 /*
1503  * this is used by the defrag code to go through all the
1504  * leaves pointed to by a node and reallocate them so that
1505  * disk order is close to key order
1506  */
1507 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
1508                        struct btrfs_root *root, struct extent_buffer *parent,
1509                        int start_slot, u64 *last_ret,
1510                        struct btrfs_key *progress)
1511 {
1512         struct btrfs_fs_info *fs_info = root->fs_info;
1513         struct extent_buffer *cur;
1514         u64 blocknr;
1515         u64 gen;
1516         u64 search_start = *last_ret;
1517         u64 last_block = 0;
1518         u64 other;
1519         u32 parent_nritems;
1520         int end_slot;
1521         int i;
1522         int err = 0;
1523         int parent_level;
1524         int uptodate;
1525         u32 blocksize;
1526         int progress_passed = 0;
1527         struct btrfs_disk_key disk_key;
1528
1529         parent_level = btrfs_header_level(parent);
1530
1531         WARN_ON(trans->transaction != fs_info->running_transaction);
1532         WARN_ON(trans->transid != fs_info->generation);
1533
1534         parent_nritems = btrfs_header_nritems(parent);
1535         blocksize = fs_info->nodesize;
1536         end_slot = parent_nritems - 1;
1537
1538         if (parent_nritems <= 1)
1539                 return 0;
1540
1541         btrfs_set_lock_blocking(parent);
1542
1543         for (i = start_slot; i <= end_slot; i++) {
1544                 struct btrfs_key first_key;
1545                 int close = 1;
1546
1547                 btrfs_node_key(parent, &disk_key, i);
1548                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
1549                         continue;
1550
1551                 progress_passed = 1;
1552                 blocknr = btrfs_node_blockptr(parent, i);
1553                 gen = btrfs_node_ptr_generation(parent, i);
1554                 btrfs_node_key_to_cpu(parent, &first_key, i);
1555                 if (last_block == 0)
1556                         last_block = blocknr;
1557
1558                 if (i > 0) {
1559                         other = btrfs_node_blockptr(parent, i - 1);
1560                         close = close_blocks(blocknr, other, blocksize);
1561                 }
1562                 if (!close && i < end_slot) {
1563                         other = btrfs_node_blockptr(parent, i + 1);
1564                         close = close_blocks(blocknr, other, blocksize);
1565                 }
1566                 if (close) {
1567                         last_block = blocknr;
1568                         continue;
1569                 }
1570
1571                 cur = find_extent_buffer(fs_info, blocknr);
1572                 if (cur)
1573                         uptodate = btrfs_buffer_uptodate(cur, gen, 0);
1574                 else
1575                         uptodate = 0;
1576                 if (!cur || !uptodate) {
1577                         if (!cur) {
1578                                 cur = read_tree_block(fs_info, blocknr, gen,
1579                                                       parent_level - 1,
1580                                                       &first_key);
1581                                 if (IS_ERR(cur)) {
1582                                         return PTR_ERR(cur);
1583                                 } else if (!extent_buffer_uptodate(cur)) {
1584                                         free_extent_buffer(cur);
1585                                         return -EIO;
1586                                 }
1587                         } else if (!uptodate) {
1588                                 err = btrfs_read_buffer(cur, gen,
1589                                                 parent_level - 1,&first_key);
1590                                 if (err) {
1591                                         free_extent_buffer(cur);
1592                                         return err;
1593                                 }
1594                         }
1595                 }
1596                 if (search_start == 0)
1597                         search_start = last_block;
1598
1599                 btrfs_tree_lock(cur);
1600                 btrfs_set_lock_blocking(cur);
1601                 err = __btrfs_cow_block(trans, root, cur, parent, i,
1602                                         &cur, search_start,
1603                                         min(16 * blocksize,
1604                                             (end_slot - i) * blocksize));
1605                 if (err) {
1606                         btrfs_tree_unlock(cur);
1607                         free_extent_buffer(cur);
1608                         break;
1609                 }
1610                 search_start = cur->start;
1611                 last_block = cur->start;
1612                 *last_ret = search_start;
1613                 btrfs_tree_unlock(cur);
1614                 free_extent_buffer(cur);
1615         }
1616         return err;
1617 }
1618
1619 /*
1620  * search for key in the extent_buffer.  The items start at offset p,
1621  * and they are item_size apart.  There are 'max' items in p.
1622  *
1623  * the slot in the array is returned via slot, and it points to
1624  * the place where you would insert key if it is not found in
1625  * the array.
1626  *
1627  * slot may point to max if the key is bigger than all of the keys
1628  */
1629 static noinline int generic_bin_search(struct extent_buffer *eb,
1630                                        unsigned long p, int item_size,
1631                                        const struct btrfs_key *key,
1632                                        int max, int *slot)
1633 {
1634         int low = 0;
1635         int high = max;
1636         int mid;
1637         int ret;
1638         struct btrfs_disk_key *tmp = NULL;
1639         struct btrfs_disk_key unaligned;
1640         unsigned long offset;
1641         char *kaddr = NULL;
1642         unsigned long map_start = 0;
1643         unsigned long map_len = 0;
1644         int err;
1645
1646         if (low > high) {
1647                 btrfs_err(eb->fs_info,
1648                  "%s: low (%d) > high (%d) eb %llu owner %llu level %d",
1649                           __func__, low, high, eb->start,
1650                           btrfs_header_owner(eb), btrfs_header_level(eb));
1651                 return -EINVAL;
1652         }
1653
1654         while (low < high) {
1655                 mid = (low + high) / 2;
1656                 offset = p + mid * item_size;
1657
1658                 if (!kaddr || offset < map_start ||
1659                     (offset + sizeof(struct btrfs_disk_key)) >
1660                     map_start + map_len) {
1661
1662                         err = map_private_extent_buffer(eb, offset,
1663                                                 sizeof(struct btrfs_disk_key),
1664                                                 &kaddr, &map_start, &map_len);
1665
1666                         if (!err) {
1667                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1668                                                         map_start);
1669                         } else if (err == 1) {
1670                                 read_extent_buffer(eb, &unaligned,
1671                                                    offset, sizeof(unaligned));
1672                                 tmp = &unaligned;
1673                         } else {
1674                                 return err;
1675                         }
1676
1677                 } else {
1678                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
1679                                                         map_start);
1680                 }
1681                 ret = comp_keys(tmp, key);
1682
1683                 if (ret < 0)
1684                         low = mid + 1;
1685                 else if (ret > 0)
1686                         high = mid;
1687                 else {
1688                         *slot = mid;
1689                         return 0;
1690                 }
1691         }
1692         *slot = low;
1693         return 1;
1694 }
1695
1696 /*
1697  * simple bin_search frontend that does the right thing for
1698  * leaves vs nodes
1699  */
1700 int btrfs_bin_search(struct extent_buffer *eb, const struct btrfs_key *key,
1701                      int level, int *slot)
1702 {
1703         if (level == 0)
1704                 return generic_bin_search(eb,
1705                                           offsetof(struct btrfs_leaf, items),
1706                                           sizeof(struct btrfs_item),
1707                                           key, btrfs_header_nritems(eb),
1708                                           slot);
1709         else
1710                 return generic_bin_search(eb,
1711                                           offsetof(struct btrfs_node, ptrs),
1712                                           sizeof(struct btrfs_key_ptr),
1713                                           key, btrfs_header_nritems(eb),
1714                                           slot);
1715 }
1716
1717 static void root_add_used(struct btrfs_root *root, u32 size)
1718 {
1719         spin_lock(&root->accounting_lock);
1720         btrfs_set_root_used(&root->root_item,
1721                             btrfs_root_used(&root->root_item) + size);
1722         spin_unlock(&root->accounting_lock);
1723 }
1724
1725 static void root_sub_used(struct btrfs_root *root, u32 size)
1726 {
1727         spin_lock(&root->accounting_lock);
1728         btrfs_set_root_used(&root->root_item,
1729                             btrfs_root_used(&root->root_item) - size);
1730         spin_unlock(&root->accounting_lock);
1731 }
1732
1733 /* given a node and slot number, this reads the blocks it points to.  The
1734  * extent buffer is returned with a reference taken (but unlocked).
1735  */
1736 static noinline struct extent_buffer *
1737 read_node_slot(struct btrfs_fs_info *fs_info, struct extent_buffer *parent,
1738                int slot)
1739 {
1740         int level = btrfs_header_level(parent);
1741         struct extent_buffer *eb;
1742         struct btrfs_key first_key;
1743
1744         if (slot < 0 || slot >= btrfs_header_nritems(parent))
1745                 return ERR_PTR(-ENOENT);
1746
1747         BUG_ON(level == 0);
1748
1749         btrfs_node_key_to_cpu(parent, &first_key, slot);
1750         eb = read_tree_block(fs_info, btrfs_node_blockptr(parent, slot),
1751                              btrfs_node_ptr_generation(parent, slot),
1752                              level - 1, &first_key);
1753         if (!IS_ERR(eb) && !extent_buffer_uptodate(eb)) {
1754                 free_extent_buffer(eb);
1755                 eb = ERR_PTR(-EIO);
1756         }
1757
1758         return eb;
1759 }
1760
1761 /*
1762  * node level balancing, used to make sure nodes are in proper order for
1763  * item deletion.  We balance from the top down, so we have to make sure
1764  * that a deletion won't leave an node completely empty later on.
1765  */
1766 static noinline int balance_level(struct btrfs_trans_handle *trans,
1767                          struct btrfs_root *root,
1768                          struct btrfs_path *path, int level)
1769 {
1770         struct btrfs_fs_info *fs_info = root->fs_info;
1771         struct extent_buffer *right = NULL;
1772         struct extent_buffer *mid;
1773         struct extent_buffer *left = NULL;
1774         struct extent_buffer *parent = NULL;
1775         int ret = 0;
1776         int wret;
1777         int pslot;
1778         int orig_slot = path->slots[level];
1779         u64 orig_ptr;
1780
1781         ASSERT(level > 0);
1782
1783         mid = path->nodes[level];
1784
1785         WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK &&
1786                 path->locks[level] != BTRFS_WRITE_LOCK_BLOCKING);
1787         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1788
1789         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1790
1791         if (level < BTRFS_MAX_LEVEL - 1) {
1792                 parent = path->nodes[level + 1];
1793                 pslot = path->slots[level + 1];
1794         }
1795
1796         /*
1797          * deal with the case where there is only one pointer in the root
1798          * by promoting the node below to a root
1799          */
1800         if (!parent) {
1801                 struct extent_buffer *child;
1802
1803                 if (btrfs_header_nritems(mid) != 1)
1804                         return 0;
1805
1806                 /* promote the child to a root */
1807                 child = read_node_slot(fs_info, mid, 0);
1808                 if (IS_ERR(child)) {
1809                         ret = PTR_ERR(child);
1810                         btrfs_handle_fs_error(fs_info, ret, NULL);
1811                         goto enospc;
1812                 }
1813
1814                 btrfs_tree_lock(child);
1815                 btrfs_set_lock_blocking(child);
1816                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
1817                 if (ret) {
1818                         btrfs_tree_unlock(child);
1819                         free_extent_buffer(child);
1820                         goto enospc;
1821                 }
1822
1823                 ret = tree_mod_log_insert_root(root->node, child, 1);
1824                 BUG_ON(ret < 0);
1825                 rcu_assign_pointer(root->node, child);
1826
1827                 add_root_to_dirty_list(root);
1828                 btrfs_tree_unlock(child);
1829
1830                 path->locks[level] = 0;
1831                 path->nodes[level] = NULL;
1832                 clean_tree_block(fs_info, mid);
1833                 btrfs_tree_unlock(mid);
1834                 /* once for the path */
1835                 free_extent_buffer(mid);
1836
1837                 root_sub_used(root, mid->len);
1838                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1839                 /* once for the root ptr */
1840                 free_extent_buffer_stale(mid);
1841                 return 0;
1842         }
1843         if (btrfs_header_nritems(mid) >
1844             BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 4)
1845                 return 0;
1846
1847         left = read_node_slot(fs_info, parent, pslot - 1);
1848         if (IS_ERR(left))
1849                 left = NULL;
1850
1851         if (left) {
1852                 btrfs_tree_lock(left);
1853                 btrfs_set_lock_blocking(left);
1854                 wret = btrfs_cow_block(trans, root, left,
1855                                        parent, pslot - 1, &left);
1856                 if (wret) {
1857                         ret = wret;
1858                         goto enospc;
1859                 }
1860         }
1861
1862         right = read_node_slot(fs_info, parent, pslot + 1);
1863         if (IS_ERR(right))
1864                 right = NULL;
1865
1866         if (right) {
1867                 btrfs_tree_lock(right);
1868                 btrfs_set_lock_blocking(right);
1869                 wret = btrfs_cow_block(trans, root, right,
1870                                        parent, pslot + 1, &right);
1871                 if (wret) {
1872                         ret = wret;
1873                         goto enospc;
1874                 }
1875         }
1876
1877         /* first, try to make some room in the middle buffer */
1878         if (left) {
1879                 orig_slot += btrfs_header_nritems(left);
1880                 wret = push_node_left(trans, fs_info, left, mid, 1);
1881                 if (wret < 0)
1882                         ret = wret;
1883         }
1884
1885         /*
1886          * then try to empty the right most buffer into the middle
1887          */
1888         if (right) {
1889                 wret = push_node_left(trans, fs_info, mid, right, 1);
1890                 if (wret < 0 && wret != -ENOSPC)
1891                         ret = wret;
1892                 if (btrfs_header_nritems(right) == 0) {
1893                         clean_tree_block(fs_info, right);
1894                         btrfs_tree_unlock(right);
1895                         del_ptr(root, path, level + 1, pslot + 1);
1896                         root_sub_used(root, right->len);
1897                         btrfs_free_tree_block(trans, root, right, 0, 1);
1898                         free_extent_buffer_stale(right);
1899                         right = NULL;
1900                 } else {
1901                         struct btrfs_disk_key right_key;
1902                         btrfs_node_key(right, &right_key, 0);
1903                         ret = tree_mod_log_insert_key(parent, pslot + 1,
1904                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
1905                         BUG_ON(ret < 0);
1906                         btrfs_set_node_key(parent, &right_key, pslot + 1);
1907                         btrfs_mark_buffer_dirty(parent);
1908                 }
1909         }
1910         if (btrfs_header_nritems(mid) == 1) {
1911                 /*
1912                  * we're not allowed to leave a node with one item in the
1913                  * tree during a delete.  A deletion from lower in the tree
1914                  * could try to delete the only pointer in this node.
1915                  * So, pull some keys from the left.
1916                  * There has to be a left pointer at this point because
1917                  * otherwise we would have pulled some pointers from the
1918                  * right
1919                  */
1920                 if (!left) {
1921                         ret = -EROFS;
1922                         btrfs_handle_fs_error(fs_info, ret, NULL);
1923                         goto enospc;
1924                 }
1925                 wret = balance_node_right(trans, fs_info, mid, left);
1926                 if (wret < 0) {
1927                         ret = wret;
1928                         goto enospc;
1929                 }
1930                 if (wret == 1) {
1931                         wret = push_node_left(trans, fs_info, left, mid, 1);
1932                         if (wret < 0)
1933                                 ret = wret;
1934                 }
1935                 BUG_ON(wret == 1);
1936         }
1937         if (btrfs_header_nritems(mid) == 0) {
1938                 clean_tree_block(fs_info, mid);
1939                 btrfs_tree_unlock(mid);
1940                 del_ptr(root, path, level + 1, pslot);
1941                 root_sub_used(root, mid->len);
1942                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1943                 free_extent_buffer_stale(mid);
1944                 mid = NULL;
1945         } else {
1946                 /* update the parent key to reflect our changes */
1947                 struct btrfs_disk_key mid_key;
1948                 btrfs_node_key(mid, &mid_key, 0);
1949                 ret = tree_mod_log_insert_key(parent, pslot,
1950                                 MOD_LOG_KEY_REPLACE, GFP_NOFS);
1951                 BUG_ON(ret < 0);
1952                 btrfs_set_node_key(parent, &mid_key, pslot);
1953                 btrfs_mark_buffer_dirty(parent);
1954         }
1955
1956         /* update the path */
1957         if (left) {
1958                 if (btrfs_header_nritems(left) > orig_slot) {
1959                         extent_buffer_get(left);
1960                         /* left was locked after cow */
1961                         path->nodes[level] = left;
1962                         path->slots[level + 1] -= 1;
1963                         path->slots[level] = orig_slot;
1964                         if (mid) {
1965                                 btrfs_tree_unlock(mid);
1966                                 free_extent_buffer(mid);
1967                         }
1968                 } else {
1969                         orig_slot -= btrfs_header_nritems(left);
1970                         path->slots[level] = orig_slot;
1971                 }
1972         }
1973         /* double check we haven't messed things up */
1974         if (orig_ptr !=
1975             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
1976                 BUG();
1977 enospc:
1978         if (right) {
1979                 btrfs_tree_unlock(right);
1980                 free_extent_buffer(right);
1981         }
1982         if (left) {
1983                 if (path->nodes[level] != left)
1984                         btrfs_tree_unlock(left);
1985                 free_extent_buffer(left);
1986         }
1987         return ret;
1988 }
1989
1990 /* Node balancing for insertion.  Here we only split or push nodes around
1991  * when they are completely full.  This is also done top down, so we
1992  * have to be pessimistic.
1993  */
1994 static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
1995                                           struct btrfs_root *root,
1996                                           struct btrfs_path *path, int level)
1997 {
1998         struct btrfs_fs_info *fs_info = root->fs_info;
1999         struct extent_buffer *right = NULL;
2000         struct extent_buffer *mid;
2001         struct extent_buffer *left = NULL;
2002         struct extent_buffer *parent = NULL;
2003         int ret = 0;
2004         int wret;
2005         int pslot;
2006         int orig_slot = path->slots[level];
2007
2008         if (level == 0)
2009                 return 1;
2010
2011         mid = path->nodes[level];
2012         WARN_ON(btrfs_header_generation(mid) != trans->transid);
2013
2014         if (level < BTRFS_MAX_LEVEL - 1) {
2015                 parent = path->nodes[level + 1];
2016                 pslot = path->slots[level + 1];
2017         }
2018
2019         if (!parent)
2020                 return 1;
2021
2022         left = read_node_slot(fs_info, parent, pslot - 1);
2023         if (IS_ERR(left))
2024                 left = NULL;
2025
2026         /* first, try to make some room in the middle buffer */
2027         if (left) {
2028                 u32 left_nr;
2029
2030                 btrfs_tree_lock(left);
2031                 btrfs_set_lock_blocking(left);
2032
2033                 left_nr = btrfs_header_nritems(left);
2034                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2035                         wret = 1;
2036                 } else {
2037                         ret = btrfs_cow_block(trans, root, left, parent,
2038                                               pslot - 1, &left);
2039                         if (ret)
2040                                 wret = 1;
2041                         else {
2042                                 wret = push_node_left(trans, fs_info,
2043                                                       left, mid, 0);
2044                         }
2045                 }
2046                 if (wret < 0)
2047                         ret = wret;
2048                 if (wret == 0) {
2049                         struct btrfs_disk_key disk_key;
2050                         orig_slot += left_nr;
2051                         btrfs_node_key(mid, &disk_key, 0);
2052                         ret = tree_mod_log_insert_key(parent, pslot,
2053                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
2054                         BUG_ON(ret < 0);
2055                         btrfs_set_node_key(parent, &disk_key, pslot);
2056                         btrfs_mark_buffer_dirty(parent);
2057                         if (btrfs_header_nritems(left) > orig_slot) {
2058                                 path->nodes[level] = left;
2059                                 path->slots[level + 1] -= 1;
2060                                 path->slots[level] = orig_slot;
2061                                 btrfs_tree_unlock(mid);
2062                                 free_extent_buffer(mid);
2063                         } else {
2064                                 orig_slot -=
2065                                         btrfs_header_nritems(left);
2066                                 path->slots[level] = orig_slot;
2067                                 btrfs_tree_unlock(left);
2068                                 free_extent_buffer(left);
2069                         }
2070                         return 0;
2071                 }
2072                 btrfs_tree_unlock(left);
2073                 free_extent_buffer(left);
2074         }
2075         right = read_node_slot(fs_info, parent, pslot + 1);
2076         if (IS_ERR(right))
2077                 right = NULL;
2078
2079         /*
2080          * then try to empty the right most buffer into the middle
2081          */
2082         if (right) {
2083                 u32 right_nr;
2084
2085                 btrfs_tree_lock(right);
2086                 btrfs_set_lock_blocking(right);
2087
2088                 right_nr = btrfs_header_nritems(right);
2089                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2090                         wret = 1;
2091                 } else {
2092                         ret = btrfs_cow_block(trans, root, right,
2093                                               parent, pslot + 1,
2094                                               &right);
2095                         if (ret)
2096                                 wret = 1;
2097                         else {
2098                                 wret = balance_node_right(trans, fs_info,
2099                                                           right, mid);
2100                         }
2101                 }
2102                 if (wret < 0)
2103                         ret = wret;
2104                 if (wret == 0) {
2105                         struct btrfs_disk_key disk_key;
2106
2107                         btrfs_node_key(right, &disk_key, 0);
2108                         ret = tree_mod_log_insert_key(parent, pslot + 1,
2109                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
2110                         BUG_ON(ret < 0);
2111                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
2112                         btrfs_mark_buffer_dirty(parent);
2113
2114                         if (btrfs_header_nritems(mid) <= orig_slot) {
2115                                 path->nodes[level] = right;
2116                                 path->slots[level + 1] += 1;
2117                                 path->slots[level] = orig_slot -
2118                                         btrfs_header_nritems(mid);
2119                                 btrfs_tree_unlock(mid);
2120                                 free_extent_buffer(mid);
2121                         } else {
2122                                 btrfs_tree_unlock(right);
2123                                 free_extent_buffer(right);
2124                         }
2125                         return 0;
2126                 }
2127                 btrfs_tree_unlock(right);
2128                 free_extent_buffer(right);
2129         }
2130         return 1;
2131 }
2132
2133 /*
2134  * readahead one full node of leaves, finding things that are close
2135  * to the block in 'slot', and triggering ra on them.
2136  */
2137 static void reada_for_search(struct btrfs_fs_info *fs_info,
2138                              struct btrfs_path *path,
2139                              int level, int slot, u64 objectid)
2140 {
2141         struct extent_buffer *node;
2142         struct btrfs_disk_key disk_key;
2143         u32 nritems;
2144         u64 search;
2145         u64 target;
2146         u64 nread = 0;
2147         struct extent_buffer *eb;
2148         u32 nr;
2149         u32 blocksize;
2150         u32 nscan = 0;
2151
2152         if (level != 1)
2153                 return;
2154
2155         if (!path->nodes[level])
2156                 return;
2157
2158         node = path->nodes[level];
2159
2160         search = btrfs_node_blockptr(node, slot);
2161         blocksize = fs_info->nodesize;
2162         eb = find_extent_buffer(fs_info, search);
2163         if (eb) {
2164                 free_extent_buffer(eb);
2165                 return;
2166         }
2167
2168         target = search;
2169
2170         nritems = btrfs_header_nritems(node);
2171         nr = slot;
2172
2173         while (1) {
2174                 if (path->reada == READA_BACK) {
2175                         if (nr == 0)
2176                                 break;
2177                         nr--;
2178                 } else if (path->reada == READA_FORWARD) {
2179                         nr++;
2180                         if (nr >= nritems)
2181                                 break;
2182                 }
2183                 if (path->reada == READA_BACK && objectid) {
2184                         btrfs_node_key(node, &disk_key, nr);
2185                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
2186                                 break;
2187                 }
2188                 search = btrfs_node_blockptr(node, nr);
2189                 if ((search <= target && target - search <= 65536) ||
2190                     (search > target && search - target <= 65536)) {
2191                         readahead_tree_block(fs_info, search);
2192                         nread += blocksize;
2193                 }
2194                 nscan++;
2195                 if ((nread > 65536 || nscan > 32))
2196                         break;
2197         }
2198 }
2199
2200 static noinline void reada_for_balance(struct btrfs_fs_info *fs_info,
2201                                        struct btrfs_path *path, int level)
2202 {
2203         int slot;
2204         int nritems;
2205         struct extent_buffer *parent;
2206         struct extent_buffer *eb;
2207         u64 gen;
2208         u64 block1 = 0;
2209         u64 block2 = 0;
2210
2211         parent = path->nodes[level + 1];
2212         if (!parent)
2213                 return;
2214
2215         nritems = btrfs_header_nritems(parent);
2216         slot = path->slots[level + 1];
2217
2218         if (slot > 0) {
2219                 block1 = btrfs_node_blockptr(parent, slot - 1);
2220                 gen = btrfs_node_ptr_generation(parent, slot - 1);
2221                 eb = find_extent_buffer(fs_info, block1);
2222                 /*
2223                  * if we get -eagain from btrfs_buffer_uptodate, we
2224                  * don't want to return eagain here.  That will loop
2225                  * forever
2226                  */
2227                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2228                         block1 = 0;
2229                 free_extent_buffer(eb);
2230         }
2231         if (slot + 1 < nritems) {
2232                 block2 = btrfs_node_blockptr(parent, slot + 1);
2233                 gen = btrfs_node_ptr_generation(parent, slot + 1);
2234                 eb = find_extent_buffer(fs_info, block2);
2235                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2236                         block2 = 0;
2237                 free_extent_buffer(eb);
2238         }
2239
2240         if (block1)
2241                 readahead_tree_block(fs_info, block1);
2242         if (block2)
2243                 readahead_tree_block(fs_info, block2);
2244 }
2245
2246
2247 /*
2248  * when we walk down the tree, it is usually safe to unlock the higher layers
2249  * in the tree.  The exceptions are when our path goes through slot 0, because
2250  * operations on the tree might require changing key pointers higher up in the
2251  * tree.
2252  *
2253  * callers might also have set path->keep_locks, which tells this code to keep
2254  * the lock if the path points to the last slot in the block.  This is part of
2255  * walking through the tree, and selecting the next slot in the higher block.
2256  *
2257  * lowest_unlock sets the lowest level in the tree we're allowed to unlock.  so
2258  * if lowest_unlock is 1, level 0 won't be unlocked
2259  */
2260 static noinline void unlock_up(struct btrfs_path *path, int level,
2261                                int lowest_unlock, int min_write_lock_level,
2262                                int *write_lock_level)
2263 {
2264         int i;
2265         int skip_level = level;
2266         int no_skips = 0;
2267         struct extent_buffer *t;
2268
2269         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2270                 if (!path->nodes[i])
2271                         break;
2272                 if (!path->locks[i])
2273                         break;
2274                 if (!no_skips && path->slots[i] == 0) {
2275                         skip_level = i + 1;
2276                         continue;
2277                 }
2278                 if (!no_skips && path->keep_locks) {
2279                         u32 nritems;
2280                         t = path->nodes[i];
2281                         nritems = btrfs_header_nritems(t);
2282                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
2283                                 skip_level = i + 1;
2284                                 continue;
2285                         }
2286                 }
2287                 if (skip_level < i && i >= lowest_unlock)
2288                         no_skips = 1;
2289
2290                 t = path->nodes[i];
2291                 if (i >= lowest_unlock && i > skip_level) {
2292                         btrfs_tree_unlock_rw(t, path->locks[i]);
2293                         path->locks[i] = 0;
2294                         if (write_lock_level &&
2295                             i > min_write_lock_level &&
2296                             i <= *write_lock_level) {
2297                                 *write_lock_level = i - 1;
2298                         }
2299                 }
2300         }
2301 }
2302
2303 /*
2304  * This releases any locks held in the path starting at level and
2305  * going all the way up to the root.
2306  *
2307  * btrfs_search_slot will keep the lock held on higher nodes in a few
2308  * corner cases, such as COW of the block at slot zero in the node.  This
2309  * ignores those rules, and it should only be called when there are no
2310  * more updates to be done higher up in the tree.
2311  */
2312 noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
2313 {
2314         int i;
2315
2316         if (path->keep_locks)
2317                 return;
2318
2319         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2320                 if (!path->nodes[i])
2321                         continue;
2322                 if (!path->locks[i])
2323                         continue;
2324                 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
2325                 path->locks[i] = 0;
2326         }
2327 }
2328
2329 /*
2330  * helper function for btrfs_search_slot.  The goal is to find a block
2331  * in cache without setting the path to blocking.  If we find the block
2332  * we return zero and the path is unchanged.
2333  *
2334  * If we can't find the block, we set the path blocking and do some
2335  * reada.  -EAGAIN is returned and the search must be repeated.
2336  */
2337 static int
2338 read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
2339                       struct extent_buffer **eb_ret, int level, int slot,
2340                       const struct btrfs_key *key)
2341 {
2342         struct btrfs_fs_info *fs_info = root->fs_info;
2343         u64 blocknr;
2344         u64 gen;
2345         struct extent_buffer *b = *eb_ret;
2346         struct extent_buffer *tmp;
2347         struct btrfs_key first_key;
2348         int ret;
2349         int parent_level;
2350
2351         blocknr = btrfs_node_blockptr(b, slot);
2352         gen = btrfs_node_ptr_generation(b, slot);
2353         parent_level = btrfs_header_level(b);
2354         btrfs_node_key_to_cpu(b, &first_key, slot);
2355
2356         tmp = find_extent_buffer(fs_info, blocknr);
2357         if (tmp) {
2358                 /* first we do an atomic uptodate check */
2359                 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
2360                         *eb_ret = tmp;
2361                         return 0;
2362                 }
2363
2364                 /* the pages were up to date, but we failed
2365                  * the generation number check.  Do a full
2366                  * read for the generation number that is correct.
2367                  * We must do this without dropping locks so
2368                  * we can trust our generation number
2369                  */
2370                 btrfs_set_path_blocking(p);
2371
2372                 /* now we're allowed to do a blocking uptodate check */
2373                 ret = btrfs_read_buffer(tmp, gen, parent_level - 1, &first_key);
2374                 if (!ret) {
2375                         *eb_ret = tmp;
2376                         return 0;
2377                 }
2378                 free_extent_buffer(tmp);
2379                 btrfs_release_path(p);
2380                 return -EIO;
2381         }
2382
2383         /*
2384          * reduce lock contention at high levels
2385          * of the btree by dropping locks before
2386          * we read.  Don't release the lock on the current
2387          * level because we need to walk this node to figure
2388          * out which blocks to read.
2389          */
2390         btrfs_unlock_up_safe(p, level + 1);
2391         btrfs_set_path_blocking(p);
2392
2393         if (p->reada != READA_NONE)
2394                 reada_for_search(fs_info, p, level, slot, key->objectid);
2395
2396         ret = -EAGAIN;
2397         tmp = read_tree_block(fs_info, blocknr, gen, parent_level - 1,
2398                               &first_key);
2399         if (!IS_ERR(tmp)) {
2400                 /*
2401                  * If the read above didn't mark this buffer up to date,
2402                  * it will never end up being up to date.  Set ret to EIO now
2403                  * and give up so that our caller doesn't loop forever
2404                  * on our EAGAINs.
2405                  */
2406                 if (!extent_buffer_uptodate(tmp))
2407                         ret = -EIO;
2408                 free_extent_buffer(tmp);
2409         } else {
2410                 ret = PTR_ERR(tmp);
2411         }
2412
2413         btrfs_release_path(p);
2414         return ret;
2415 }
2416
2417 /*
2418  * helper function for btrfs_search_slot.  This does all of the checks
2419  * for node-level blocks and does any balancing required based on
2420  * the ins_len.
2421  *
2422  * If no extra work was required, zero is returned.  If we had to
2423  * drop the path, -EAGAIN is returned and btrfs_search_slot must
2424  * start over
2425  */
2426 static int
2427 setup_nodes_for_search(struct btrfs_trans_handle *trans,
2428                        struct btrfs_root *root, struct btrfs_path *p,
2429                        struct extent_buffer *b, int level, int ins_len,
2430                        int *write_lock_level)
2431 {
2432         struct btrfs_fs_info *fs_info = root->fs_info;
2433         int ret;
2434
2435         if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
2436             BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3) {
2437                 int sret;
2438
2439                 if (*write_lock_level < level + 1) {
2440                         *write_lock_level = level + 1;
2441                         btrfs_release_path(p);
2442                         goto again;
2443                 }
2444
2445                 btrfs_set_path_blocking(p);
2446                 reada_for_balance(fs_info, p, level);
2447                 sret = split_node(trans, root, p, level);
2448
2449                 BUG_ON(sret > 0);
2450                 if (sret) {
2451                         ret = sret;
2452                         goto done;
2453                 }
2454                 b = p->nodes[level];
2455         } else if (ins_len < 0 && btrfs_header_nritems(b) <
2456                    BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 2) {
2457                 int sret;
2458
2459                 if (*write_lock_level < level + 1) {
2460                         *write_lock_level = level + 1;
2461                         btrfs_release_path(p);
2462                         goto again;
2463                 }
2464
2465                 btrfs_set_path_blocking(p);
2466                 reada_for_balance(fs_info, p, level);
2467                 sret = balance_level(trans, root, p, level);
2468
2469                 if (sret) {
2470                         ret = sret;
2471                         goto done;
2472                 }
2473                 b = p->nodes[level];
2474                 if (!b) {
2475                         btrfs_release_path(p);
2476                         goto again;
2477                 }
2478                 BUG_ON(btrfs_header_nritems(b) == 1);
2479         }
2480         return 0;
2481
2482 again:
2483         ret = -EAGAIN;
2484 done:
2485         return ret;
2486 }
2487
2488 static void key_search_validate(struct extent_buffer *b,
2489                                 const struct btrfs_key *key,
2490                                 int level)
2491 {
2492 #ifdef CONFIG_BTRFS_ASSERT
2493         struct btrfs_disk_key disk_key;
2494
2495         btrfs_cpu_key_to_disk(&disk_key, key);
2496
2497         if (level == 0)
2498                 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2499                     offsetof(struct btrfs_leaf, items[0].key),
2500                     sizeof(disk_key)));
2501         else
2502                 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2503                     offsetof(struct btrfs_node, ptrs[0].key),
2504                     sizeof(disk_key)));
2505 #endif
2506 }
2507
2508 static int key_search(struct extent_buffer *b, const struct btrfs_key *key,
2509                       int level, int *prev_cmp, int *slot)
2510 {
2511         if (*prev_cmp != 0) {
2512                 *prev_cmp = btrfs_bin_search(b, key, level, slot);
2513                 return *prev_cmp;
2514         }
2515
2516         key_search_validate(b, key, level);
2517         *slot = 0;
2518
2519         return 0;
2520 }
2521
2522 int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
2523                 u64 iobjectid, u64 ioff, u8 key_type,
2524                 struct btrfs_key *found_key)
2525 {
2526         int ret;
2527         struct btrfs_key key;
2528         struct extent_buffer *eb;
2529
2530         ASSERT(path);
2531         ASSERT(found_key);
2532
2533         key.type = key_type;
2534         key.objectid = iobjectid;
2535         key.offset = ioff;
2536
2537         ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
2538         if (ret < 0)
2539                 return ret;
2540
2541         eb = path->nodes[0];
2542         if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
2543                 ret = btrfs_next_leaf(fs_root, path);
2544                 if (ret)
2545                         return ret;
2546                 eb = path->nodes[0];
2547         }
2548
2549         btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
2550         if (found_key->type != key.type ||
2551                         found_key->objectid != key.objectid)
2552                 return 1;
2553
2554         return 0;
2555 }
2556
2557 static struct extent_buffer *btrfs_search_slot_get_root(struct btrfs_root *root,
2558                                                         struct btrfs_path *p,
2559                                                         int write_lock_level)
2560 {
2561         struct btrfs_fs_info *fs_info = root->fs_info;
2562         struct extent_buffer *b;
2563         int root_lock;
2564         int level = 0;
2565
2566         /* We try very hard to do read locks on the root */
2567         root_lock = BTRFS_READ_LOCK;
2568
2569         if (p->search_commit_root) {
2570                 /* The commit roots are read only so we always do read locks */
2571                 if (p->need_commit_sem)
2572                         down_read(&fs_info->commit_root_sem);
2573                 b = root->commit_root;
2574                 extent_buffer_get(b);
2575                 level = btrfs_header_level(b);
2576                 if (p->need_commit_sem)
2577                         up_read(&fs_info->commit_root_sem);
2578                 /*
2579                  * Ensure that all callers have set skip_locking when
2580                  * p->search_commit_root = 1.
2581                  */
2582                 ASSERT(p->skip_locking == 1);
2583
2584                 goto out;
2585         }
2586
2587         if (p->skip_locking) {
2588                 b = btrfs_root_node(root);
2589                 level = btrfs_header_level(b);
2590                 goto out;
2591         }
2592
2593         /*
2594          * If the level is set to maximum, we can skip trying to get the read
2595          * lock.
2596          */
2597         if (write_lock_level < BTRFS_MAX_LEVEL) {
2598                 /*
2599                  * We don't know the level of the root node until we actually
2600                  * have it read locked
2601                  */
2602                 b = btrfs_read_lock_root_node(root);
2603                 level = btrfs_header_level(b);
2604                 if (level > write_lock_level)
2605                         goto out;
2606
2607                 /* Whoops, must trade for write lock */
2608                 btrfs_tree_read_unlock(b);
2609                 free_extent_buffer(b);
2610         }
2611
2612         b = btrfs_lock_root_node(root);
2613         root_lock = BTRFS_WRITE_LOCK;
2614
2615         /* The level might have changed, check again */
2616         level = btrfs_header_level(b);
2617
2618 out:
2619         p->nodes[level] = b;
2620         if (!p->skip_locking)
2621                 p->locks[level] = root_lock;
2622         /*
2623          * Callers are responsible for dropping b's references.
2624          */
2625         return b;
2626 }
2627
2628
2629 /*
2630  * btrfs_search_slot - look for a key in a tree and perform necessary
2631  * modifications to preserve tree invariants.
2632  *
2633  * @trans:      Handle of transaction, used when modifying the tree
2634  * @p:          Holds all btree nodes along the search path
2635  * @root:       The root node of the tree
2636  * @key:        The key we are looking for
2637  * @ins_len:    Indicates purpose of search, for inserts it is 1, for
2638  *              deletions it's -1. 0 for plain searches
2639  * @cow:        boolean should CoW operations be performed. Must always be 1
2640  *              when modifying the tree.
2641  *
2642  * If @ins_len > 0, nodes and leaves will be split as we walk down the tree.
2643  * If @ins_len < 0, nodes will be merged as we walk down the tree (if possible)
2644  *
2645  * If @key is found, 0 is returned and you can find the item in the leaf level
2646  * of the path (level 0)
2647  *
2648  * If @key isn't found, 1 is returned and the leaf level of the path (level 0)
2649  * points to the slot where it should be inserted
2650  *
2651  * If an error is encountered while searching the tree a negative error number
2652  * is returned
2653  */
2654 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2655                       const struct btrfs_key *key, struct btrfs_path *p,
2656                       int ins_len, int cow)
2657 {
2658         struct btrfs_fs_info *fs_info = root->fs_info;
2659         struct extent_buffer *b;
2660         int slot;
2661         int ret;
2662         int err;
2663         int level;
2664         int lowest_unlock = 1;
2665         /* everything at write_lock_level or lower must be write locked */
2666         int write_lock_level = 0;
2667         u8 lowest_level = 0;
2668         int min_write_lock_level;
2669         int prev_cmp;
2670
2671         lowest_level = p->lowest_level;
2672         WARN_ON(lowest_level && ins_len > 0);
2673         WARN_ON(p->nodes[0] != NULL);
2674         BUG_ON(!cow && ins_len);
2675
2676         if (ins_len < 0) {
2677                 lowest_unlock = 2;
2678
2679                 /* when we are removing items, we might have to go up to level
2680                  * two as we update tree pointers  Make sure we keep write
2681                  * for those levels as well
2682                  */
2683                 write_lock_level = 2;
2684         } else if (ins_len > 0) {
2685                 /*
2686                  * for inserting items, make sure we have a write lock on
2687                  * level 1 so we can update keys
2688                  */
2689                 write_lock_level = 1;
2690         }
2691
2692         if (!cow)
2693                 write_lock_level = -1;
2694
2695         if (cow && (p->keep_locks || p->lowest_level))
2696                 write_lock_level = BTRFS_MAX_LEVEL;
2697
2698         min_write_lock_level = write_lock_level;
2699
2700 again:
2701         prev_cmp = -1;
2702         b = btrfs_search_slot_get_root(root, p, write_lock_level);
2703
2704         while (b) {
2705                 level = btrfs_header_level(b);
2706
2707                 /*
2708                  * setup the path here so we can release it under lock
2709                  * contention with the cow code
2710                  */
2711                 if (cow) {
2712                         bool last_level = (level == (BTRFS_MAX_LEVEL - 1));
2713
2714                         /*
2715                          * if we don't really need to cow this block
2716                          * then we don't want to set the path blocking,
2717                          * so we test it here
2718                          */
2719                         if (!should_cow_block(trans, root, b)) {
2720                                 trans->dirty = true;
2721                                 goto cow_done;
2722                         }
2723
2724                         /*
2725                          * must have write locks on this node and the
2726                          * parent
2727                          */
2728                         if (level > write_lock_level ||
2729                             (level + 1 > write_lock_level &&
2730                             level + 1 < BTRFS_MAX_LEVEL &&
2731                             p->nodes[level + 1])) {
2732                                 write_lock_level = level + 1;
2733                                 btrfs_release_path(p);
2734                                 goto again;
2735                         }
2736
2737                         btrfs_set_path_blocking(p);
2738                         if (last_level)
2739                                 err = btrfs_cow_block(trans, root, b, NULL, 0,
2740                                                       &b);
2741                         else
2742                                 err = btrfs_cow_block(trans, root, b,
2743                                                       p->nodes[level + 1],
2744                                                       p->slots[level + 1], &b);
2745                         if (err) {
2746                                 ret = err;
2747                                 goto done;
2748                         }
2749                 }
2750 cow_done:
2751                 p->nodes[level] = b;
2752                 /*
2753                  * Leave path with blocking locks to avoid massive
2754                  * lock context switch, this is made on purpose.
2755                  */
2756
2757                 /*
2758                  * we have a lock on b and as long as we aren't changing
2759                  * the tree, there is no way to for the items in b to change.
2760                  * It is safe to drop the lock on our parent before we
2761                  * go through the expensive btree search on b.
2762                  *
2763                  * If we're inserting or deleting (ins_len != 0), then we might
2764                  * be changing slot zero, which may require changing the parent.
2765                  * So, we can't drop the lock until after we know which slot
2766                  * we're operating on.
2767                  */
2768                 if (!ins_len && !p->keep_locks) {
2769                         int u = level + 1;
2770
2771                         if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
2772                                 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
2773                                 p->locks[u] = 0;
2774                         }
2775                 }
2776
2777                 ret = key_search(b, key, level, &prev_cmp, &slot);
2778                 if (ret < 0)
2779                         goto done;
2780
2781                 if (level != 0) {
2782                         int dec = 0;
2783                         if (ret && slot > 0) {
2784                                 dec = 1;
2785                                 slot -= 1;
2786                         }
2787                         p->slots[level] = slot;
2788                         err = setup_nodes_for_search(trans, root, p, b, level,
2789                                              ins_len, &write_lock_level);
2790                         if (err == -EAGAIN)
2791                                 goto again;
2792                         if (err) {
2793                                 ret = err;
2794                                 goto done;
2795                         }
2796                         b = p->nodes[level];
2797                         slot = p->slots[level];
2798
2799                         /*
2800                          * slot 0 is special, if we change the key
2801                          * we have to update the parent pointer
2802                          * which means we must have a write lock
2803                          * on the parent
2804                          */
2805                         if (slot == 0 && ins_len &&
2806                             write_lock_level < level + 1) {
2807                                 write_lock_level = level + 1;
2808                                 btrfs_release_path(p);
2809                                 goto again;
2810                         }
2811
2812                         unlock_up(p, level, lowest_unlock,
2813                                   min_write_lock_level, &write_lock_level);
2814
2815                         if (level == lowest_level) {
2816                                 if (dec)
2817                                         p->slots[level]++;
2818                                 goto done;
2819                         }
2820
2821                         err = read_block_for_search(root, p, &b, level,
2822                                                     slot, key);
2823                         if (err == -EAGAIN)
2824                                 goto again;
2825                         if (err) {
2826                                 ret = err;
2827                                 goto done;
2828                         }
2829
2830                         if (!p->skip_locking) {
2831                                 level = btrfs_header_level(b);
2832                                 if (level <= write_lock_level) {
2833                                         err = btrfs_try_tree_write_lock(b);
2834                                         if (!err) {
2835                                                 btrfs_set_path_blocking(p);
2836                                                 btrfs_tree_lock(b);
2837                                         }
2838                                         p->locks[level] = BTRFS_WRITE_LOCK;
2839                                 } else {
2840                                         err = btrfs_tree_read_lock_atomic(b);
2841                                         if (!err) {
2842                                                 btrfs_set_path_blocking(p);
2843                                                 btrfs_tree_read_lock(b);
2844                                         }
2845                                         p->locks[level] = BTRFS_READ_LOCK;
2846                                 }
2847                                 p->nodes[level] = b;
2848                         }
2849                 } else {
2850                         p->slots[level] = slot;
2851                         if (ins_len > 0 &&
2852                             btrfs_leaf_free_space(fs_info, b) < ins_len) {
2853                                 if (write_lock_level < 1) {
2854                                         write_lock_level = 1;
2855                                         btrfs_release_path(p);
2856                                         goto again;
2857                                 }
2858
2859                                 btrfs_set_path_blocking(p);
2860                                 err = split_leaf(trans, root, key,
2861                                                  p, ins_len, ret == 0);
2862
2863                                 BUG_ON(err > 0);
2864                                 if (err) {
2865                                         ret = err;
2866                                         goto done;
2867                                 }
2868                         }
2869                         if (!p->search_for_split)
2870                                 unlock_up(p, level, lowest_unlock,
2871                                           min_write_lock_level, NULL);
2872                         goto done;
2873                 }
2874         }
2875         ret = 1;
2876 done:
2877         /*
2878          * we don't really know what they plan on doing with the path
2879          * from here on, so for now just mark it as blocking
2880          */
2881         if (!p->leave_spinning)
2882                 btrfs_set_path_blocking(p);
2883         if (ret < 0 && !p->skip_release_on_error)
2884                 btrfs_release_path(p);
2885         return ret;
2886 }
2887
2888 /*
2889  * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2890  * current state of the tree together with the operations recorded in the tree
2891  * modification log to search for the key in a previous version of this tree, as
2892  * denoted by the time_seq parameter.
2893  *
2894  * Naturally, there is no support for insert, delete or cow operations.
2895  *
2896  * The resulting path and return value will be set up as if we called
2897  * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2898  */
2899 int btrfs_search_old_slot(struct btrfs_root *root, const struct btrfs_key *key,
2900                           struct btrfs_path *p, u64 time_seq)
2901 {
2902         struct btrfs_fs_info *fs_info = root->fs_info;
2903         struct extent_buffer *b;
2904         int slot;
2905         int ret;
2906         int err;
2907         int level;
2908         int lowest_unlock = 1;
2909         u8 lowest_level = 0;
2910         int prev_cmp = -1;
2911
2912         lowest_level = p->lowest_level;
2913         WARN_ON(p->nodes[0] != NULL);
2914
2915         if (p->search_commit_root) {
2916                 BUG_ON(time_seq);
2917                 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2918         }
2919
2920 again:
2921         b = get_old_root(root, time_seq);
2922         if (!b) {
2923                 ret = -EIO;
2924                 goto done;
2925         }
2926         level = btrfs_header_level(b);
2927         p->locks[level] = BTRFS_READ_LOCK;
2928
2929         while (b) {
2930                 level = btrfs_header_level(b);
2931                 p->nodes[level] = b;
2932
2933                 /*
2934                  * we have a lock on b and as long as we aren't changing
2935                  * the tree, there is no way to for the items in b to change.
2936                  * It is safe to drop the lock on our parent before we
2937                  * go through the expensive btree search on b.
2938                  */
2939                 btrfs_unlock_up_safe(p, level + 1);
2940
2941                 /*
2942                  * Since we can unwind ebs we want to do a real search every
2943                  * time.
2944                  */
2945                 prev_cmp = -1;
2946                 ret = key_search(b, key, level, &prev_cmp, &slot);
2947
2948                 if (level != 0) {
2949                         int dec = 0;
2950                         if (ret && slot > 0) {
2951                                 dec = 1;
2952                                 slot -= 1;
2953                         }
2954                         p->slots[level] = slot;
2955                         unlock_up(p, level, lowest_unlock, 0, NULL);
2956
2957                         if (level == lowest_level) {
2958                                 if (dec)
2959                                         p->slots[level]++;
2960                                 goto done;
2961                         }
2962
2963                         err = read_block_for_search(root, p, &b, level,
2964                                                     slot, key);
2965                         if (err == -EAGAIN)
2966                                 goto again;
2967                         if (err) {
2968                                 ret = err;
2969                                 goto done;
2970                         }
2971
2972                         level = btrfs_header_level(b);
2973                         err = btrfs_tree_read_lock_atomic(b);
2974                         if (!err) {
2975                                 btrfs_set_path_blocking(p);
2976                                 btrfs_tree_read_lock(b);
2977                         }
2978                         b = tree_mod_log_rewind(fs_info, p, b, time_seq);
2979                         if (!b) {
2980                                 ret = -ENOMEM;
2981                                 goto done;
2982                         }
2983                         p->locks[level] = BTRFS_READ_LOCK;
2984                         p->nodes[level] = b;
2985                 } else {
2986                         p->slots[level] = slot;
2987                         unlock_up(p, level, lowest_unlock, 0, NULL);
2988                         goto done;
2989                 }
2990         }
2991         ret = 1;
2992 done:
2993         if (!p->leave_spinning)
2994                 btrfs_set_path_blocking(p);
2995         if (ret < 0)
2996                 btrfs_release_path(p);
2997
2998         return ret;
2999 }
3000
3001 /*
3002  * helper to use instead of search slot if no exact match is needed but
3003  * instead the next or previous item should be returned.
3004  * When find_higher is true, the next higher item is returned, the next lower
3005  * otherwise.
3006  * When return_any and find_higher are both true, and no higher item is found,
3007  * return the next lower instead.
3008  * When return_any is true and find_higher is false, and no lower item is found,
3009  * return the next higher instead.
3010  * It returns 0 if any item is found, 1 if none is found (tree empty), and
3011  * < 0 on error
3012  */
3013 int btrfs_search_slot_for_read(struct btrfs_root *root,
3014                                const struct btrfs_key *key,
3015                                struct btrfs_path *p, int find_higher,
3016                                int return_any)
3017 {
3018         int ret;
3019         struct extent_buffer *leaf;
3020
3021 again:
3022         ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
3023         if (ret <= 0)
3024                 return ret;
3025         /*
3026          * a return value of 1 means the path is at the position where the
3027          * item should be inserted. Normally this is the next bigger item,
3028          * but in case the previous item is the last in a leaf, path points
3029          * to the first free slot in the previous leaf, i.e. at an invalid
3030          * item.
3031          */
3032         leaf = p->nodes[0];
3033
3034         if (find_higher) {
3035                 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
3036                         ret = btrfs_next_leaf(root, p);
3037                         if (ret <= 0)
3038                                 return ret;
3039                         if (!return_any)
3040                                 return 1;
3041                         /*
3042                          * no higher item found, return the next
3043                          * lower instead
3044                          */
3045                         return_any = 0;
3046                         find_higher = 0;
3047                         btrfs_release_path(p);
3048                         goto again;
3049                 }
3050         } else {
3051                 if (p->slots[0] == 0) {
3052                         ret = btrfs_prev_leaf(root, p);
3053                         if (ret < 0)
3054                                 return ret;
3055                         if (!ret) {
3056                                 leaf = p->nodes[0];
3057                                 if (p->slots[0] == btrfs_header_nritems(leaf))
3058                                         p->slots[0]--;
3059                                 return 0;
3060                         }
3061                         if (!return_any)
3062                                 return 1;
3063                         /*
3064                          * no lower item found, return the next
3065                          * higher instead
3066                          */
3067                         return_any = 0;
3068                         find_higher = 1;
3069                         btrfs_release_path(p);
3070                         goto again;
3071                 } else {
3072                         --p->slots[0];
3073                 }
3074         }
3075         return 0;
3076 }
3077
3078 /*
3079  * adjust the pointers going up the tree, starting at level
3080  * making sure the right key of each node is points to 'key'.
3081  * This is used after shifting pointers to the left, so it stops
3082  * fixing up pointers when a given leaf/node is not in slot 0 of the
3083  * higher levels
3084  *
3085  */
3086 static void fixup_low_keys(struct btrfs_path *path,
3087                            struct btrfs_disk_key *key, int level)
3088 {
3089         int i;
3090         struct extent_buffer *t;
3091         int ret;
3092
3093         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
3094                 int tslot = path->slots[i];
3095
3096                 if (!path->nodes[i])
3097                         break;
3098                 t = path->nodes[i];
3099                 ret = tree_mod_log_insert_key(t, tslot, MOD_LOG_KEY_REPLACE,
3100                                 GFP_ATOMIC);
3101                 BUG_ON(ret < 0);
3102                 btrfs_set_node_key(t, key, tslot);
3103                 btrfs_mark_buffer_dirty(path->nodes[i]);
3104                 if (tslot != 0)
3105                         break;
3106         }
3107 }
3108
3109 /*
3110  * update item key.
3111  *
3112  * This function isn't completely safe. It's the caller's responsibility
3113  * that the new key won't break the order
3114  */
3115 void btrfs_set_item_key_safe(struct btrfs_fs_info *fs_info,
3116                              struct btrfs_path *path,
3117                              const struct btrfs_key *new_key)
3118 {
3119         struct btrfs_disk_key disk_key;
3120         struct extent_buffer *eb;
3121         int slot;
3122
3123         eb = path->nodes[0];
3124         slot = path->slots[0];
3125         if (slot > 0) {
3126                 btrfs_item_key(eb, &disk_key, slot - 1);
3127                 BUG_ON(comp_keys(&disk_key, new_key) >= 0);
3128         }
3129         if (slot < btrfs_header_nritems(eb) - 1) {
3130                 btrfs_item_key(eb, &disk_key, slot + 1);
3131                 BUG_ON(comp_keys(&disk_key, new_key) <= 0);
3132         }
3133
3134         btrfs_cpu_key_to_disk(&disk_key, new_key);
3135         btrfs_set_item_key(eb, &disk_key, slot);
3136         btrfs_mark_buffer_dirty(eb);
3137         if (slot == 0)
3138                 fixup_low_keys(path, &disk_key, 1);
3139 }
3140
3141 /*
3142  * try to push data from one node into the next node left in the
3143  * tree.
3144  *
3145  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
3146  * error, and > 0 if there was no room in the left hand block.
3147  */
3148 static int push_node_left(struct btrfs_trans_handle *trans,
3149                           struct btrfs_fs_info *fs_info,
3150                           struct extent_buffer *dst,
3151                           struct extent_buffer *src, int empty)
3152 {
3153         int push_items = 0;
3154         int src_nritems;
3155         int dst_nritems;
3156         int ret = 0;
3157
3158         src_nritems = btrfs_header_nritems(src);
3159         dst_nritems = btrfs_header_nritems(dst);
3160         push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3161         WARN_ON(btrfs_header_generation(src) != trans->transid);
3162         WARN_ON(btrfs_header_generation(dst) != trans->transid);
3163
3164         if (!empty && src_nritems <= 8)
3165                 return 1;
3166
3167         if (push_items <= 0)
3168                 return 1;
3169
3170         if (empty) {
3171                 push_items = min(src_nritems, push_items);
3172                 if (push_items < src_nritems) {
3173                         /* leave at least 8 pointers in the node if
3174                          * we aren't going to empty it
3175                          */
3176                         if (src_nritems - push_items < 8) {
3177                                 if (push_items <= 8)
3178                                         return 1;
3179                                 push_items -= 8;
3180                         }
3181                 }
3182         } else
3183                 push_items = min(src_nritems - 8, push_items);
3184
3185         ret = tree_mod_log_eb_copy(fs_info, dst, src, dst_nritems, 0,
3186                                    push_items);
3187         if (ret) {
3188                 btrfs_abort_transaction(trans, ret);
3189                 return ret;
3190         }
3191         copy_extent_buffer(dst, src,
3192                            btrfs_node_key_ptr_offset(dst_nritems),
3193                            btrfs_node_key_ptr_offset(0),
3194                            push_items * sizeof(struct btrfs_key_ptr));
3195
3196         if (push_items < src_nritems) {
3197                 /*
3198                  * Don't call tree_mod_log_insert_move here, key removal was
3199                  * already fully logged by tree_mod_log_eb_copy above.
3200                  */
3201                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
3202                                       btrfs_node_key_ptr_offset(push_items),
3203                                       (src_nritems - push_items) *
3204                                       sizeof(struct btrfs_key_ptr));
3205         }
3206         btrfs_set_header_nritems(src, src_nritems - push_items);
3207         btrfs_set_header_nritems(dst, dst_nritems + push_items);
3208         btrfs_mark_buffer_dirty(src);
3209         btrfs_mark_buffer_dirty(dst);
3210
3211         return ret;
3212 }
3213
3214 /*
3215  * try to push data from one node into the next node right in the
3216  * tree.
3217  *
3218  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
3219  * error, and > 0 if there was no room in the right hand block.
3220  *
3221  * this will  only push up to 1/2 the contents of the left node over
3222  */
3223 static int balance_node_right(struct btrfs_trans_handle *trans,
3224                               struct btrfs_fs_info *fs_info,
3225                               struct extent_buffer *dst,
3226                               struct extent_buffer *src)
3227 {
3228         int push_items = 0;
3229         int max_push;
3230         int src_nritems;
3231         int dst_nritems;
3232         int ret = 0;
3233
3234         WARN_ON(btrfs_header_generation(src) != trans->transid);
3235         WARN_ON(btrfs_header_generation(dst) != trans->transid);
3236
3237         src_nritems = btrfs_header_nritems(src);
3238         dst_nritems = btrfs_header_nritems(dst);
3239         push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3240         if (push_items <= 0)
3241                 return 1;
3242
3243         if (src_nritems < 4)
3244                 return 1;
3245
3246         max_push = src_nritems / 2 + 1;
3247         /* don't try to empty the node */
3248         if (max_push >= src_nritems)
3249                 return 1;
3250
3251         if (max_push < push_items)
3252                 push_items = max_push;
3253
3254         ret = tree_mod_log_insert_move(dst, push_items, 0, dst_nritems);
3255         BUG_ON(ret < 0);
3256         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
3257                                       btrfs_node_key_ptr_offset(0),
3258                                       (dst_nritems) *
3259                                       sizeof(struct btrfs_key_ptr));
3260
3261         ret = tree_mod_log_eb_copy(fs_info, dst, src, 0,
3262                                    src_nritems - push_items, push_items);
3263         if (ret) {
3264                 btrfs_abort_transaction(trans, ret);
3265                 return ret;
3266         }
3267         copy_extent_buffer(dst, src,
3268                            btrfs_node_key_ptr_offset(0),
3269                            btrfs_node_key_ptr_offset(src_nritems - push_items),
3270                            push_items * sizeof(struct btrfs_key_ptr));
3271
3272         btrfs_set_header_nritems(src, src_nritems - push_items);
3273         btrfs_set_header_nritems(dst, dst_nritems + push_items);
3274
3275         btrfs_mark_buffer_dirty(src);
3276         btrfs_mark_buffer_dirty(dst);
3277
3278         return ret;
3279 }
3280
3281 /*
3282  * helper function to insert a new root level in the tree.
3283  * A new node is allocated, and a single item is inserted to
3284  * point to the existing root
3285  *
3286  * returns zero on success or < 0 on failure.
3287  */
3288 static noinline int insert_new_root(struct btrfs_trans_handle *trans,
3289                            struct btrfs_root *root,
3290                            struct btrfs_path *path, int level)
3291 {
3292         struct btrfs_fs_info *fs_info = root->fs_info;
3293         u64 lower_gen;
3294         struct extent_buffer *lower;
3295         struct extent_buffer *c;
3296         struct extent_buffer *old;
3297         struct btrfs_disk_key lower_key;
3298         int ret;
3299
3300         BUG_ON(path->nodes[level]);
3301         BUG_ON(path->nodes[level-1] != root->node);
3302
3303         lower = path->nodes[level-1];
3304         if (level == 1)
3305                 btrfs_item_key(lower, &lower_key, 0);
3306         else
3307                 btrfs_node_key(lower, &lower_key, 0);
3308
3309         c = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
3310                                    &lower_key, level, root->node->start, 0);
3311         if (IS_ERR(c))
3312                 return PTR_ERR(c);
3313
3314         root_add_used(root, fs_info->nodesize);
3315
3316         btrfs_set_header_nritems(c, 1);
3317         btrfs_set_node_key(c, &lower_key, 0);
3318         btrfs_set_node_blockptr(c, 0, lower->start);
3319         lower_gen = btrfs_header_generation(lower);
3320         WARN_ON(lower_gen != trans->transid);
3321
3322         btrfs_set_node_ptr_generation(c, 0, lower_gen);
3323
3324         btrfs_mark_buffer_dirty(c);
3325
3326         old = root->node;
3327         ret = tree_mod_log_insert_root(root->node, c, 0);
3328         BUG_ON(ret < 0);
3329         rcu_assign_pointer(root->node, c);
3330
3331         /* the super has an extra ref to root->node */
3332         free_extent_buffer(old);
3333
3334         add_root_to_dirty_list(root);
3335         extent_buffer_get(c);
3336         path->nodes[level] = c;
3337         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
3338         path->slots[level] = 0;
3339         return 0;
3340 }
3341
3342 /*
3343  * worker function to insert a single pointer in a node.
3344  * the node should have enough room for the pointer already
3345  *
3346  * slot and level indicate where you want the key to go, and
3347  * blocknr is the block the key points to.
3348  */
3349 static void insert_ptr(struct btrfs_trans_handle *trans,
3350                        struct btrfs_fs_info *fs_info, struct btrfs_path *path,
3351                        struct btrfs_disk_key *key, u64 bytenr,
3352                        int slot, int level)
3353 {
3354         struct extent_buffer *lower;
3355         int nritems;
3356         int ret;
3357
3358         BUG_ON(!path->nodes[level]);
3359         btrfs_assert_tree_locked(path->nodes[level]);
3360         lower = path->nodes[level];
3361         nritems = btrfs_header_nritems(lower);
3362         BUG_ON(slot > nritems);
3363         BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(fs_info));
3364         if (slot != nritems) {
3365                 if (level) {
3366                         ret = tree_mod_log_insert_move(lower, slot + 1, slot,
3367                                         nritems - slot);
3368                         BUG_ON(ret < 0);
3369                 }
3370                 memmove_extent_buffer(lower,
3371                               btrfs_node_key_ptr_offset(slot + 1),
3372                               btrfs_node_key_ptr_offset(slot),
3373                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
3374         }
3375         if (level) {
3376                 ret = tree_mod_log_insert_key(lower, slot, MOD_LOG_KEY_ADD,
3377                                 GFP_NOFS);
3378                 BUG_ON(ret < 0);
3379         }
3380         btrfs_set_node_key(lower, key, slot);
3381         btrfs_set_node_blockptr(lower, slot, bytenr);
3382         WARN_ON(trans->transid == 0);
3383         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
3384         btrfs_set_header_nritems(lower, nritems + 1);
3385         btrfs_mark_buffer_dirty(lower);
3386 }
3387
3388 /*
3389  * split the node at the specified level in path in two.
3390  * The path is corrected to point to the appropriate node after the split
3391  *
3392  * Before splitting this tries to make some room in the node by pushing
3393  * left and right, if either one works, it returns right away.
3394  *
3395  * returns 0 on success and < 0 on failure
3396  */
3397 static noinline int split_node(struct btrfs_trans_handle *trans,
3398                                struct btrfs_root *root,
3399                                struct btrfs_path *path, int level)
3400 {
3401         struct btrfs_fs_info *fs_info = root->fs_info;
3402         struct extent_buffer *c;
3403         struct extent_buffer *split;
3404         struct btrfs_disk_key disk_key;
3405         int mid;
3406         int ret;
3407         u32 c_nritems;
3408
3409         c = path->nodes[level];
3410         WARN_ON(btrfs_header_generation(c) != trans->transid);
3411         if (c == root->node) {
3412                 /*
3413                  * trying to split the root, lets make a new one
3414                  *
3415                  * tree mod log: We don't log_removal old root in
3416                  * insert_new_root, because that root buffer will be kept as a
3417                  * normal node. We are going to log removal of half of the
3418                  * elements below with tree_mod_log_eb_copy. We're holding a
3419                  * tree lock on the buffer, which is why we cannot race with
3420                  * other tree_mod_log users.
3421                  */
3422                 ret = insert_new_root(trans, root, path, level + 1);
3423                 if (ret)
3424                         return ret;
3425         } else {
3426                 ret = push_nodes_for_insert(trans, root, path, level);
3427                 c = path->nodes[level];
3428                 if (!ret && btrfs_header_nritems(c) <
3429                     BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3)
3430                         return 0;
3431                 if (ret < 0)
3432                         return ret;
3433         }
3434
3435         c_nritems = btrfs_header_nritems(c);
3436         mid = (c_nritems + 1) / 2;
3437         btrfs_node_key(c, &disk_key, mid);
3438
3439         split = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
3440                         &disk_key, level, c->start, 0);
3441         if (IS_ERR(split))
3442                 return PTR_ERR(split);
3443
3444         root_add_used(root, fs_info->nodesize);
3445         ASSERT(btrfs_header_level(c) == level);
3446
3447         ret = tree_mod_log_eb_copy(fs_info, split, c, 0, mid, c_nritems - mid);
3448         if (ret) {
3449                 btrfs_abort_transaction(trans, ret);
3450                 return ret;
3451         }
3452         copy_extent_buffer(split, c,
3453                            btrfs_node_key_ptr_offset(0),
3454                            btrfs_node_key_ptr_offset(mid),
3455                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
3456         btrfs_set_header_nritems(split, c_nritems - mid);
3457         btrfs_set_header_nritems(c, mid);
3458         ret = 0;
3459
3460         btrfs_mark_buffer_dirty(c);
3461         btrfs_mark_buffer_dirty(split);
3462
3463         insert_ptr(trans, fs_info, path, &disk_key, split->start,
3464                    path->slots[level + 1] + 1, level + 1);
3465
3466         if (path->slots[level] >= mid) {
3467                 path->slots[level] -= mid;
3468                 btrfs_tree_unlock(c);
3469                 free_extent_buffer(c);
3470                 path->nodes[level] = split;
3471                 path->slots[level + 1] += 1;
3472         } else {
3473                 btrfs_tree_unlock(split);
3474                 free_extent_buffer(split);
3475         }
3476         return ret;
3477 }
3478
3479 /*
3480  * how many bytes are required to store the items in a leaf.  start
3481  * and nr indicate which items in the leaf to check.  This totals up the
3482  * space used both by the item structs and the item data
3483  */
3484 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
3485 {
3486         struct btrfs_item *start_item;
3487         struct btrfs_item *end_item;
3488         struct btrfs_map_token token;
3489         int data_len;
3490         int nritems = btrfs_header_nritems(l);
3491         int end = min(nritems, start + nr) - 1;
3492
3493         if (!nr)
3494                 return 0;
3495         btrfs_init_map_token(&token);
3496         start_item = btrfs_item_nr(start);
3497         end_item = btrfs_item_nr(end);
3498         data_len = btrfs_token_item_offset(l, start_item, &token) +
3499                 btrfs_token_item_size(l, start_item, &token);
3500         data_len = data_len - btrfs_token_item_offset(l, end_item, &token);
3501         data_len += sizeof(struct btrfs_item) * nr;
3502         WARN_ON(data_len < 0);
3503         return data_len;
3504 }
3505
3506 /*
3507  * The space between the end of the leaf items and
3508  * the start of the leaf data.  IOW, how much room
3509  * the leaf has left for both items and data
3510  */
3511 noinline int btrfs_leaf_free_space(struct btrfs_fs_info *fs_info,
3512                                    struct extent_buffer *leaf)
3513 {
3514         int nritems = btrfs_header_nritems(leaf);
3515         int ret;
3516
3517         ret = BTRFS_LEAF_DATA_SIZE(fs_info) - leaf_space_used(leaf, 0, nritems);
3518         if (ret < 0) {
3519                 btrfs_crit(fs_info,
3520                            "leaf free space ret %d, leaf data size %lu, used %d nritems %d",
3521                            ret,
3522                            (unsigned long) BTRFS_LEAF_DATA_SIZE(fs_info),
3523                            leaf_space_used(leaf, 0, nritems), nritems);
3524         }
3525         return ret;
3526 }
3527
3528 /*
3529  * min slot controls the lowest index we're willing to push to the
3530  * right.  We'll push up to and including min_slot, but no lower
3531  */
3532 static noinline int __push_leaf_right(struct btrfs_fs_info *fs_info,
3533                                       struct btrfs_path *path,
3534                                       int data_size, int empty,
3535                                       struct extent_buffer *right,
3536                                       int free_space, u32 left_nritems,
3537                                       u32 min_slot)
3538 {
3539         struct extent_buffer *left = path->nodes[0];
3540         struct extent_buffer *upper = path->nodes[1];
3541         struct btrfs_map_token token;
3542         struct btrfs_disk_key disk_key;
3543         int slot;
3544         u32 i;
3545         int push_space = 0;
3546         int push_items = 0;
3547         struct btrfs_item *item;
3548         u32 nr;
3549         u32 right_nritems;
3550         u32 data_end;
3551         u32 this_item_size;
3552
3553         btrfs_init_map_token(&token);
3554
3555         if (empty)
3556                 nr = 0;
3557         else
3558                 nr = max_t(u32, 1, min_slot);
3559
3560         if (path->slots[0] >= left_nritems)
3561                 push_space += data_size;
3562
3563         slot = path->slots[1];
3564         i = left_nritems - 1;
3565         while (i >= nr) {
3566                 item = btrfs_item_nr(i);
3567
3568                 if (!empty && push_items > 0) {
3569                         if (path->slots[0] > i)
3570                                 break;
3571                         if (path->slots[0] == i) {
3572                                 int space = btrfs_leaf_free_space(fs_info, left);
3573                                 if (space + push_space * 2 > free_space)
3574                                         break;
3575                         }
3576                 }
3577
3578                 if (path->slots[0] == i)
3579                         push_space += data_size;
3580
3581                 this_item_size = btrfs_item_size(left, item);
3582                 if (this_item_size + sizeof(*item) + push_space > free_space)
3583                         break;
3584
3585                 push_items++;
3586                 push_space += this_item_size + sizeof(*item);
3587                 if (i == 0)
3588                         break;
3589                 i--;
3590         }
3591
3592         if (push_items == 0)
3593                 goto out_unlock;
3594
3595         WARN_ON(!empty && push_items == left_nritems);
3596
3597         /* push left to right */
3598         right_nritems = btrfs_header_nritems(right);
3599
3600         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
3601         push_space -= leaf_data_end(fs_info, left);
3602
3603         /* make room in the right data area */
3604         data_end = leaf_data_end(fs_info, right);
3605         memmove_extent_buffer(right,
3606                               BTRFS_LEAF_DATA_OFFSET + data_end - push_space,
3607                               BTRFS_LEAF_DATA_OFFSET + data_end,
3608                               BTRFS_LEAF_DATA_SIZE(fs_info) - data_end);
3609
3610         /* copy from the left data area */
3611         copy_extent_buffer(right, left, BTRFS_LEAF_DATA_OFFSET +
3612                      BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3613                      BTRFS_LEAF_DATA_OFFSET + leaf_data_end(fs_info, left),
3614                      push_space);
3615
3616         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
3617                               btrfs_item_nr_offset(0),
3618                               right_nritems * sizeof(struct btrfs_item));
3619
3620         /* copy the items from left to right */
3621         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
3622                    btrfs_item_nr_offset(left_nritems - push_items),
3623                    push_items * sizeof(struct btrfs_item));
3624
3625         /* update the item pointers */
3626         right_nritems += push_items;
3627         btrfs_set_header_nritems(right, right_nritems);
3628         push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3629         for (i = 0; i < right_nritems; i++) {
3630                 item = btrfs_item_nr(i);
3631                 push_space -= btrfs_token_item_size(right, item, &token);
3632                 btrfs_set_token_item_offset(right, item, push_space, &token);
3633         }
3634
3635         left_nritems -= push_items;
3636         btrfs_set_header_nritems(left, left_nritems);
3637
3638         if (left_nritems)
3639                 btrfs_mark_buffer_dirty(left);
3640         else
3641                 clean_tree_block(fs_info, left);
3642
3643         btrfs_mark_buffer_dirty(right);
3644
3645         btrfs_item_key(right, &disk_key, 0);
3646         btrfs_set_node_key(upper, &disk_key, slot + 1);
3647         btrfs_mark_buffer_dirty(upper);
3648
3649         /* then fixup the leaf pointer in the path */
3650         if (path->slots[0] >= left_nritems) {
3651                 path->slots[0] -= left_nritems;
3652                 if (btrfs_header_nritems(path->nodes[0]) == 0)
3653                         clean_tree_block(fs_info, path->nodes[0]);
3654                 btrfs_tree_unlock(path->nodes[0]);
3655                 free_extent_buffer(path->nodes[0]);
3656                 path->nodes[0] = right;
3657                 path->slots[1] += 1;
3658         } else {
3659                 btrfs_tree_unlock(right);
3660                 free_extent_buffer(right);
3661         }
3662         return 0;
3663
3664 out_unlock:
3665         btrfs_tree_unlock(right);
3666         free_extent_buffer(right);
3667         return 1;
3668 }
3669
3670 /*
3671  * push some data in the path leaf to the right, trying to free up at
3672  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
3673  *
3674  * returns 1 if the push failed because the other node didn't have enough
3675  * room, 0 if everything worked out and < 0 if there were major errors.
3676  *
3677  * this will push starting from min_slot to the end of the leaf.  It won't
3678  * push any slot lower than min_slot
3679  */
3680 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
3681                            *root, struct btrfs_path *path,
3682                            int min_data_size, int data_size,
3683                            int empty, u32 min_slot)
3684 {
3685         struct btrfs_fs_info *fs_info = root->fs_info;
3686         struct extent_buffer *left = path->nodes[0];
3687         struct extent_buffer *right;
3688         struct extent_buffer *upper;
3689         int slot;
3690         int free_space;
3691         u32 left_nritems;
3692         int ret;
3693
3694         if (!path->nodes[1])
3695                 return 1;
3696
3697         slot = path->slots[1];
3698         upper = path->nodes[1];
3699         if (slot >= btrfs_header_nritems(upper) - 1)
3700                 return 1;
3701
3702         btrfs_assert_tree_locked(path->nodes[1]);
3703
3704         right = read_node_slot(fs_info, upper, slot + 1);
3705         /*
3706          * slot + 1 is not valid or we fail to read the right node,
3707          * no big deal, just return.
3708          */
3709         if (IS_ERR(right))
3710                 return 1;
3711
3712         btrfs_tree_lock(right);
3713         btrfs_set_lock_blocking(right);
3714
3715         free_space = btrfs_leaf_free_space(fs_info, right);
3716         if (free_space < data_size)
3717                 goto out_unlock;
3718
3719         /* cow and double check */
3720         ret = btrfs_cow_block(trans, root, right, upper,
3721                               slot + 1, &right);
3722         if (ret)
3723                 goto out_unlock;
3724
3725         free_space = btrfs_leaf_free_space(fs_info, right);
3726         if (free_space < data_size)
3727                 goto out_unlock;
3728
3729         left_nritems = btrfs_header_nritems(left);
3730         if (left_nritems == 0)
3731                 goto out_unlock;
3732
3733         if (path->slots[0] == left_nritems && !empty) {
3734                 /* Key greater than all keys in the leaf, right neighbor has
3735                  * enough room for it and we're not emptying our leaf to delete
3736                  * it, therefore use right neighbor to insert the new item and
3737                  * no need to touch/dirty our left leaft. */
3738                 btrfs_tree_unlock(left);
3739                 free_extent_buffer(left);
3740                 path->nodes[0] = right;
3741                 path->slots[0] = 0;
3742                 path->slots[1]++;
3743                 return 0;
3744         }
3745
3746         return __push_leaf_right(fs_info, path, min_data_size, empty,
3747                                 right, free_space, left_nritems, min_slot);
3748 out_unlock:
3749         btrfs_tree_unlock(right);
3750         free_extent_buffer(right);
3751         return 1;
3752 }
3753
3754 /*
3755  * push some data in the path leaf to the left, trying to free up at
3756  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
3757  *
3758  * max_slot can put a limit on how far into the leaf we'll push items.  The
3759  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us do all the
3760  * items
3761  */
3762 static noinline int __push_leaf_left(struct btrfs_fs_info *fs_info,
3763                                      struct btrfs_path *path, int data_size,
3764                                      int empty, struct extent_buffer *left,
3765                                      int free_space, u32 right_nritems,
3766                                      u32 max_slot)
3767 {
3768         struct btrfs_disk_key disk_key;
3769         struct extent_buffer *right = path->nodes[0];
3770         int i;
3771         int push_space = 0;
3772         int push_items = 0;
3773         struct btrfs_item *item;
3774         u32 old_left_nritems;
3775         u32 nr;
3776         int ret = 0;
3777         u32 this_item_size;
3778         u32 old_left_item_size;
3779         struct btrfs_map_token token;
3780
3781         btrfs_init_map_token(&token);
3782
3783         if (empty)
3784                 nr = min(right_nritems, max_slot);
3785         else
3786                 nr = min(right_nritems - 1, max_slot);
3787
3788         for (i = 0; i < nr; i++) {
3789                 item = btrfs_item_nr(i);
3790
3791                 if (!empty && push_items > 0) {
3792                         if (path->slots[0] < i)
3793                                 break;
3794                         if (path->slots[0] == i) {
3795                                 int space = btrfs_leaf_free_space(fs_info, right);
3796                                 if (space + push_space * 2 > free_space)
3797                                         break;
3798                         }
3799                 }
3800
3801                 if (path->slots[0] == i)
3802                         push_space += data_size;
3803
3804                 this_item_size = btrfs_item_size(right, item);
3805                 if (this_item_size + sizeof(*item) + push_space > free_space)
3806                         break;
3807
3808                 push_items++;
3809                 push_space += this_item_size + sizeof(*item);
3810         }
3811
3812         if (push_items == 0) {
3813                 ret = 1;
3814                 goto out;
3815         }
3816         WARN_ON(!empty && push_items == btrfs_header_nritems(right));
3817
3818         /* push data from right to left */
3819         copy_extent_buffer(left, right,
3820                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
3821                            btrfs_item_nr_offset(0),
3822                            push_items * sizeof(struct btrfs_item));
3823
3824         push_space = BTRFS_LEAF_DATA_SIZE(fs_info) -
3825                      btrfs_item_offset_nr(right, push_items - 1);
3826
3827         copy_extent_buffer(left, right, BTRFS_LEAF_DATA_OFFSET +
3828                      leaf_data_end(fs_info, left) - push_space,
3829                      BTRFS_LEAF_DATA_OFFSET +
3830                      btrfs_item_offset_nr(right, push_items - 1),
3831                      push_space);
3832         old_left_nritems = btrfs_header_nritems(left);
3833         BUG_ON(old_left_nritems <= 0);
3834
3835         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
3836         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
3837                 u32 ioff;
3838
3839                 item = btrfs_item_nr(i);
3840
3841                 ioff = btrfs_token_item_offset(left, item, &token);
3842                 btrfs_set_token_item_offset(left, item,
3843                       ioff - (BTRFS_LEAF_DATA_SIZE(fs_info) - old_left_item_size),
3844                       &token);
3845         }
3846         btrfs_set_header_nritems(left, old_left_nritems + push_items);
3847
3848         /* fixup right node */
3849         if (push_items > right_nritems)
3850                 WARN(1, KERN_CRIT "push items %d nr %u\n", push_items,
3851                        right_nritems);
3852
3853         if (push_items < right_nritems) {
3854                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
3855                                                   leaf_data_end(fs_info, right);
3856                 memmove_extent_buffer(right, BTRFS_LEAF_DATA_OFFSET +
3857                                       BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3858                                       BTRFS_LEAF_DATA_OFFSET +
3859                                       leaf_data_end(fs_info, right), push_space);
3860
3861                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
3862                               btrfs_item_nr_offset(push_items),
3863                              (btrfs_header_nritems(right) - push_items) *
3864                              sizeof(struct btrfs_item));
3865         }
3866         right_nritems -= push_items;
3867         btrfs_set_header_nritems(right, right_nritems);
3868         push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3869         for (i = 0; i < right_nritems; i++) {
3870                 item = btrfs_item_nr(i);
3871
3872                 push_space = push_space - btrfs_token_item_size(right,
3873                                                                 item, &token);
3874                 btrfs_set_token_item_offset(right, item, push_space, &token);
3875         }
3876
3877         btrfs_mark_buffer_dirty(left);
3878         if (right_nritems)
3879                 btrfs_mark_buffer_dirty(right);
3880         else
3881                 clean_tree_block(fs_info, right);
3882
3883         btrfs_item_key(right, &disk_key, 0);
3884         fixup_low_keys(path, &disk_key, 1);
3885
3886         /* then fixup the leaf pointer in the path */
3887         if (path->slots[0] < push_items) {
3888                 path->slots[0] += old_left_nritems;
3889                 btrfs_tree_unlock(path->nodes[0]);
3890                 free_extent_buffer(path->nodes[0]);
3891                 path->nodes[0] = left;
3892                 path->slots[1] -= 1;
3893         } else {
3894                 btrfs_tree_unlock(left);
3895                 free_extent_buffer(left);
3896                 path->slots[0] -= push_items;
3897         }
3898         BUG_ON(path->slots[0] < 0);
3899         return ret;
3900 out:
3901         btrfs_tree_unlock(left);
3902         free_extent_buffer(left);
3903         return ret;
3904 }
3905
3906 /*
3907  * push some data in the path leaf to the left, trying to free up at
3908  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
3909  *
3910  * max_slot can put a limit on how far into the leaf we'll push items.  The
3911  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us push all the
3912  * items
3913  */
3914 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
3915                           *root, struct btrfs_path *path, int min_data_size,
3916                           int data_size, int empty, u32 max_slot)
3917 {
3918         struct btrfs_fs_info *fs_info = root->fs_info;
3919         struct extent_buffer *right = path->nodes[0];
3920         struct extent_buffer *left;
3921         int slot;
3922         int free_space;
3923         u32 right_nritems;
3924         int ret = 0;
3925
3926         slot = path->slots[1];
3927         if (slot == 0)
3928                 return 1;
3929         if (!path->nodes[1])
3930                 return 1;
3931
3932         right_nritems = btrfs_header_nritems(right);
3933         if (right_nritems == 0)
3934                 return 1;
3935
3936         btrfs_assert_tree_locked(path->nodes[1]);
3937
3938         left = read_node_slot(fs_info, path->nodes[1], slot - 1);
3939         /*
3940          * slot - 1 is not valid or we fail to read the left node,
3941          * no big deal, just return.
3942          */
3943         if (IS_ERR(left))
3944                 return 1;
3945
3946         btrfs_tree_lock(left);
3947         btrfs_set_lock_blocking(left);
3948
3949         free_space = btrfs_leaf_free_space(fs_info, left);
3950         if (free_space < data_size) {
3951                 ret = 1;
3952                 goto out;
3953         }
3954
3955         /* cow and double check */
3956         ret = btrfs_cow_block(trans, root, left,
3957                               path->nodes[1], slot - 1, &left);
3958         if (ret) {
3959                 /* we hit -ENOSPC, but it isn't fatal here */
3960                 if (ret == -ENOSPC)
3961                         ret = 1;
3962                 goto out;
3963         }
3964
3965         free_space = btrfs_leaf_free_space(fs_info, left);
3966         if (free_space < data_size) {
3967                 ret = 1;
3968                 goto out;
3969         }
3970
3971         return __push_leaf_left(fs_info, path, min_data_size,
3972                                empty, left, free_space, right_nritems,
3973                                max_slot);
3974 out:
3975         btrfs_tree_unlock(left);
3976         free_extent_buffer(left);
3977         return ret;
3978 }
3979
3980 /*
3981  * split the path's leaf in two, making sure there is at least data_size
3982  * available for the resulting leaf level of the path.
3983  */
3984 static noinline void copy_for_split(struct btrfs_trans_handle *trans,
3985                                     struct btrfs_fs_info *fs_info,
3986                                     struct btrfs_path *path,
3987                                     struct extent_buffer *l,
3988                                     struct extent_buffer *right,
3989                                     int slot, int mid, int nritems)
3990 {
3991         int data_copy_size;
3992         int rt_data_off;
3993         int i;
3994         struct btrfs_disk_key disk_key;
3995         struct btrfs_map_token token;
3996
3997         btrfs_init_map_token(&token);
3998
3999         nritems = nritems - mid;
4000         btrfs_set_header_nritems(right, nritems);
4001         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(fs_info, l);
4002
4003         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
4004                            btrfs_item_nr_offset(mid),
4005                            nritems * sizeof(struct btrfs_item));
4006
4007         copy_extent_buffer(right, l,
4008                      BTRFS_LEAF_DATA_OFFSET + BTRFS_LEAF_DATA_SIZE(fs_info) -
4009                      data_copy_size, BTRFS_LEAF_DATA_OFFSET +
4010                      leaf_data_end(fs_info, l), data_copy_size);
4011
4012         rt_data_off = BTRFS_LEAF_DATA_SIZE(fs_info) - btrfs_item_end_nr(l, mid);
4013
4014         for (i = 0; i < nritems; i++) {
4015                 struct btrfs_item *item = btrfs_item_nr(i);
4016                 u32 ioff;
4017
4018                 ioff = btrfs_token_item_offset(right, item, &token);
4019                 btrfs_set_token_item_offset(right, item,
4020                                             ioff + rt_data_off, &token);
4021         }
4022
4023         btrfs_set_header_nritems(l, mid);
4024         btrfs_item_key(right, &disk_key, 0);
4025         insert_ptr(trans, fs_info, path, &disk_key, right->start,
4026                    path->slots[1] + 1, 1);
4027
4028         btrfs_mark_buffer_dirty(right);
4029         btrfs_mark_buffer_dirty(l);
4030         BUG_ON(path->slots[0] != slot);
4031
4032         if (mid <= slot) {
4033                 btrfs_tree_unlock(path->nodes[0]);
4034                 free_extent_buffer(path->nodes[0]);
4035                 path->nodes[0] = right;
4036                 path->slots[0] -= mid;
4037                 path->slots[1] += 1;
4038         } else {
4039                 btrfs_tree_unlock(right);
4040                 free_extent_buffer(right);
4041         }
4042
4043         BUG_ON(path->slots[0] < 0);
4044 }
4045
4046 /*
4047  * double splits happen when we need to insert a big item in the middle
4048  * of a leaf.  A double split can leave us with 3 mostly empty leaves:
4049  * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
4050  *          A                 B                 C
4051  *
4052  * We avoid this by trying to push the items on either side of our target
4053  * into the adjacent leaves.  If all goes well we can avoid the double split
4054  * completely.
4055  */
4056 static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
4057                                           struct btrfs_root *root,
4058                                           struct btrfs_path *path,
4059                                           int data_size)
4060 {
4061         struct btrfs_fs_info *fs_info = root->fs_info;
4062         int ret;
4063         int progress = 0;
4064         int slot;
4065         u32 nritems;
4066         int space_needed = data_size;
4067
4068         slot = path->slots[0];
4069         if (slot < btrfs_header_nritems(path->nodes[0]))
4070                 space_needed -= btrfs_leaf_free_space(fs_info, path->nodes[0]);
4071
4072         /*
4073          * try to push all the items after our slot into the
4074          * right leaf
4075          */
4076         ret = push_leaf_right(trans, root, path, 1, space_needed, 0, slot);
4077         if (ret < 0)
4078                 return ret;
4079
4080         if (ret == 0)
4081                 progress++;
4082
4083         nritems = btrfs_header_nritems(path->nodes[0]);
4084         /*
4085          * our goal is to get our slot at the start or end of a leaf.  If
4086          * we've done so we're done
4087          */
4088         if (path->slots[0] == 0 || path->slots[0] == nritems)
4089                 return 0;
4090
4091         if (btrfs_leaf_free_space(fs_info, path->nodes[0]) >= data_size)
4092                 return 0;
4093
4094         /* try to push all the items before our slot into the next leaf */
4095         slot = path->slots[0];
4096         space_needed = data_size;
4097         if (slot > 0)
4098                 space_needed -= btrfs_leaf_free_space(fs_info, path->nodes[0]);
4099         ret = push_leaf_left(trans, root, path, 1, space_needed, 0, slot);
4100         if (ret < 0)
4101                 return ret;
4102
4103         if (ret == 0)
4104                 progress++;
4105
4106         if (progress)
4107                 return 0;
4108         return 1;
4109 }
4110
4111 /*
4112  * split the path's leaf in two, making sure there is at least data_size
4113  * available for the resulting leaf level of the path.
4114  *
4115  * returns 0 if all went well and < 0 on failure.
4116  */
4117 static noinline int split_leaf(struct btrfs_trans_handle *trans,
4118                                struct btrfs_root *root,
4119                                const struct btrfs_key *ins_key,
4120                                struct btrfs_path *path, int data_size,
4121                                int extend)
4122 {
4123         struct btrfs_disk_key disk_key;
4124         struct extent_buffer *l;
4125         u32 nritems;
4126         int mid;
4127         int slot;
4128         struct extent_buffer *right;
4129         struct btrfs_fs_info *fs_info = root->fs_info;
4130         int ret = 0;
4131         int wret;
4132         int split;
4133         int num_doubles = 0;
4134         int tried_avoid_double = 0;
4135
4136         l = path->nodes[0];
4137         slot = path->slots[0];
4138         if (extend && data_size + btrfs_item_size_nr(l, slot) +
4139             sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(fs_info))
4140                 return -EOVERFLOW;
4141
4142         /* first try to make some room by pushing left and right */
4143         if (data_size && path->nodes[1]) {
4144                 int space_needed = data_size;
4145
4146                 if (slot < btrfs_header_nritems(l))
4147                         space_needed -= btrfs_leaf_free_space(fs_info, l);
4148
4149                 wret = push_leaf_right(trans, root, path, space_needed,
4150                                        space_needed, 0, 0);
4151                 if (wret < 0)
4152                         return wret;
4153                 if (wret) {
4154                         space_needed = data_size;
4155                         if (slot > 0)
4156                                 space_needed -= btrfs_leaf_free_space(fs_info,
4157                                                                       l);
4158                         wret = push_leaf_left(trans, root, path, space_needed,
4159                                               space_needed, 0, (u32)-1);
4160                         if (wret < 0)
4161                                 return wret;
4162                 }
4163                 l = path->nodes[0];
4164
4165                 /* did the pushes work? */
4166                 if (btrfs_leaf_free_space(fs_info, l) >= data_size)
4167                         return 0;
4168         }
4169
4170         if (!path->nodes[1]) {
4171                 ret = insert_new_root(trans, root, path, 1);
4172                 if (ret)
4173                         return ret;
4174         }
4175 again:
4176         split = 1;
4177         l = path->nodes[0];
4178         slot = path->slots[0];
4179         nritems = btrfs_header_nritems(l);
4180         mid = (nritems + 1) / 2;
4181
4182         if (mid <= slot) {
4183                 if (nritems == 1 ||
4184                     leaf_space_used(l, mid, nritems - mid) + data_size >
4185                         BTRFS_LEAF_DATA_SIZE(fs_info)) {
4186                         if (slot >= nritems) {
4187                                 split = 0;
4188                         } else {
4189                                 mid = slot;
4190                                 if (mid != nritems &&
4191                                     leaf_space_used(l, mid, nritems - mid) +
4192                                     data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
4193                                         if (data_size && !tried_avoid_double)
4194                                                 goto push_for_double;
4195                                         split = 2;
4196                                 }
4197                         }
4198                 }
4199         } else {
4200                 if (leaf_space_used(l, 0, mid) + data_size >
4201                         BTRFS_LEAF_DATA_SIZE(fs_info)) {
4202                         if (!extend && data_size && slot == 0) {
4203                                 split = 0;
4204                         } else if ((extend || !data_size) && slot == 0) {
4205                                 mid = 1;
4206                         } else {
4207                                 mid = slot;
4208                                 if (mid != nritems &&
4209                                     leaf_space_used(l, mid, nritems - mid) +
4210                                     data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
4211                                         if (data_size && !tried_avoid_double)
4212                                                 goto push_for_double;
4213                                         split = 2;
4214                                 }
4215                         }
4216                 }
4217         }
4218
4219         if (split == 0)
4220                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
4221         else
4222                 btrfs_item_key(l, &disk_key, mid);
4223
4224         right = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
4225                         &disk_key, 0, l->start, 0);
4226         if (IS_ERR(right))
4227                 return PTR_ERR(right);
4228
4229         root_add_used(root, fs_info->nodesize);
4230
4231         if (split == 0) {
4232                 if (mid <= slot) {
4233                         btrfs_set_header_nritems(right, 0);
4234                         insert_ptr(trans, fs_info, path, &disk_key,
4235                                    right->start, path->slots[1] + 1, 1);
4236                         btrfs_tree_unlock(path->nodes[0]);
4237                         free_extent_buffer(path->nodes[0]);
4238                         path->nodes[0] = right;
4239                         path->slots[0] = 0;
4240                         path->slots[1] += 1;
4241                 } else {
4242                         btrfs_set_header_nritems(right, 0);
4243                         insert_ptr(trans, fs_info, path, &disk_key,
4244                                    right->start, path->slots[1], 1);
4245                         btrfs_tree_unlock(path->nodes[0]);
4246                         free_extent_buffer(path->nodes[0]);
4247                         path->nodes[0] = right;
4248                         path->slots[0] = 0;
4249                         if (path->slots[1] == 0)
4250                                 fixup_low_keys(path, &disk_key, 1);
4251                 }
4252                 /*
4253                  * We create a new leaf 'right' for the required ins_len and
4254                  * we'll do btrfs_mark_buffer_dirty() on this leaf after copying
4255                  * the content of ins_len to 'right'.
4256                  */
4257                 return ret;
4258         }
4259
4260         copy_for_split(trans, fs_info, path, l, right, slot, mid, nritems);
4261
4262         if (split == 2) {
4263                 BUG_ON(num_doubles != 0);
4264                 num_doubles++;
4265                 goto again;
4266         }
4267
4268         return 0;
4269
4270 push_for_double:
4271         push_for_double_split(trans, root, path, data_size);
4272         tried_avoid_double = 1;
4273         if (btrfs_leaf_free_space(fs_info, path->nodes[0]) >= data_size)
4274                 return 0;
4275         goto again;
4276 }
4277
4278 static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
4279                                          struct btrfs_root *root,
4280                                          struct btrfs_path *path, int ins_len)
4281 {
4282         struct btrfs_fs_info *fs_info = root->fs_info;
4283         struct btrfs_key key;
4284         struct extent_buffer *leaf;
4285         struct btrfs_file_extent_item *fi;
4286         u64 extent_len = 0;
4287         u32 item_size;
4288         int ret;
4289
4290         leaf = path->nodes[0];
4291         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4292
4293         BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
4294                key.type != BTRFS_EXTENT_CSUM_KEY);
4295
4296         if (btrfs_leaf_free_space(fs_info, leaf) >= ins_len)
4297                 return 0;
4298
4299         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4300         if (key.type == BTRFS_EXTENT_DATA_KEY) {
4301                 fi = btrfs_item_ptr(leaf, path->slots[0],
4302                                     struct btrfs_file_extent_item);
4303                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
4304         }
4305         btrfs_release_path(path);
4306
4307         path->keep_locks = 1;
4308         path->search_for_split = 1;
4309         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
4310         path->search_for_split = 0;
4311         if (ret > 0)
4312                 ret = -EAGAIN;
4313         if (ret < 0)
4314                 goto err;
4315
4316         ret = -EAGAIN;
4317         leaf = path->nodes[0];
4318         /* if our item isn't there, return now */
4319         if (item_size != btrfs_item_size_nr(leaf, path->slots[0]))
4320                 goto err;
4321
4322         /* the leaf has  changed, it now has room.  return now */
4323         if (btrfs_leaf_free_space(fs_info, path->nodes[0]) >= ins_len)
4324                 goto err;
4325
4326         if (key.type == BTRFS_EXTENT_DATA_KEY) {
4327                 fi = btrfs_item_ptr(leaf, path->slots[0],
4328                                     struct btrfs_file_extent_item);
4329                 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
4330                         goto err;
4331         }
4332
4333         btrfs_set_path_blocking(path);
4334         ret = split_leaf(trans, root, &key, path, ins_len, 1);
4335         if (ret)
4336                 goto err;
4337
4338         path->keep_locks = 0;
4339         btrfs_unlock_up_safe(path, 1);
4340         return 0;
4341 err:
4342         path->keep_locks = 0;
4343         return ret;
4344 }
4345
4346 static noinline int split_item(struct btrfs_fs_info *fs_info,
4347                                struct btrfs_path *path,
4348                                const struct btrfs_key *new_key,
4349                                unsigned long split_offset)
4350 {
4351         struct extent_buffer *leaf;
4352         struct btrfs_item *item;
4353         struct btrfs_item *new_item;
4354         int slot;
4355         char *buf;
4356         u32 nritems;
4357         u32 item_size;
4358         u32 orig_offset;
4359         struct btrfs_disk_key disk_key;
4360
4361         leaf = path->nodes[0];
4362         BUG_ON(btrfs_leaf_free_space(fs_info, leaf) < sizeof(struct btrfs_item));
4363
4364         btrfs_set_path_blocking(path);
4365
4366         item = btrfs_item_nr(path->slots[0]);
4367         orig_offset = btrfs_item_offset(leaf, item);
4368         item_size = btrfs_item_size(leaf, item);
4369
4370         buf = kmalloc(item_size, GFP_NOFS);
4371         if (!buf)
4372                 return -ENOMEM;
4373
4374         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
4375                             path->slots[0]), item_size);
4376
4377         slot = path->slots[0] + 1;
4378         nritems = btrfs_header_nritems(leaf);
4379         if (slot != nritems) {
4380                 /* shift the items */
4381                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
4382                                 btrfs_item_nr_offset(slot),
4383                                 (nritems - slot) * sizeof(struct btrfs_item));
4384         }
4385
4386         btrfs_cpu_key_to_disk(&disk_key, new_key);
4387         btrfs_set_item_key(leaf, &disk_key, slot);
4388
4389         new_item = btrfs_item_nr(slot);
4390
4391         btrfs_set_item_offset(leaf, new_item, orig_offset);
4392         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
4393
4394         btrfs_set_item_offset(leaf, item,
4395                               orig_offset + item_size - split_offset);
4396         btrfs_set_item_size(leaf, item, split_offset);
4397
4398         btrfs_set_header_nritems(leaf, nritems + 1);
4399
4400         /* write the data for the start of the original item */
4401         write_extent_buffer(leaf, buf,
4402                             btrfs_item_ptr_offset(leaf, path->slots[0]),
4403                             split_offset);
4404
4405         /* write the data for the new item */
4406         write_extent_buffer(leaf, buf + split_offset,
4407                             btrfs_item_ptr_offset(leaf, slot),
4408                             item_size - split_offset);
4409         btrfs_mark_buffer_dirty(leaf);
4410
4411         BUG_ON(btrfs_leaf_free_space(fs_info, leaf) < 0);
4412         kfree(buf);
4413         return 0;
4414 }
4415
4416 /*
4417  * This function splits a single item into two items,
4418  * giving 'new_key' to the new item and splitting the
4419  * old one at split_offset (from the start of the item).
4420  *
4421  * The path may be released by this operation.  After
4422  * the split, the path is pointing to the old item.  The
4423  * new item is going to be in the same node as the old one.
4424  *
4425  * Note, the item being split must be smaller enough to live alone on
4426  * a tree block with room for one extra struct btrfs_item
4427  *
4428  * This allows us to split the item in place, keeping a lock on the
4429  * leaf the entire time.
4430  */
4431 int btrfs_split_item(struct btrfs_trans_handle *trans,
4432                      struct btrfs_root *root,
4433                      struct btrfs_path *path,
4434                      const struct btrfs_key *new_key,
4435                      unsigned long split_offset)
4436 {
4437         int ret;
4438         ret = setup_leaf_for_split(trans, root, path,
4439                                    sizeof(struct btrfs_item));
4440         if (ret)
4441                 return ret;
4442
4443         ret = split_item(root->fs_info, path, new_key, split_offset);
4444         return ret;
4445 }
4446
4447 /*
4448  * This function duplicate a item, giving 'new_key' to the new item.
4449  * It guarantees both items live in the same tree leaf and the new item
4450  * is contiguous with the original item.
4451  *
4452  * This allows us to split file extent in place, keeping a lock on the
4453  * leaf the entire time.
4454  */
4455 int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
4456                          struct btrfs_root *root,
4457                          struct btrfs_path *path,
4458                          const struct btrfs_key *new_key)
4459 {
4460         struct extent_buffer *leaf;
4461         int ret;
4462         u32 item_size;
4463
4464         leaf = path->nodes[0];
4465         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4466         ret = setup_leaf_for_split(trans, root, path,
4467                                    item_size + sizeof(struct btrfs_item));
4468         if (ret)
4469                 return ret;
4470
4471         path->slots[0]++;
4472         setup_items_for_insert(root, path, new_key, &item_size,
4473                                item_size, item_size +
4474                                sizeof(struct btrfs_item), 1);
4475         leaf = path->nodes[0];
4476         memcpy_extent_buffer(leaf,
4477                              btrfs_item_ptr_offset(leaf, path->slots[0]),
4478                              btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
4479                              item_size);
4480         return 0;
4481 }
4482
4483 /*
4484  * make the item pointed to by the path smaller.  new_size indicates
4485  * how small to make it, and from_end tells us if we just chop bytes
4486  * off the end of the item or if we shift the item to chop bytes off
4487  * the front.
4488  */
4489 void btrfs_truncate_item(struct btrfs_fs_info *fs_info,
4490                          struct btrfs_path *path, u32 new_size, int from_end)
4491 {
4492         int slot;
4493         struct extent_buffer *leaf;
4494         struct btrfs_item *item;
4495         u32 nritems;
4496         unsigned int data_end;
4497         unsigned int old_data_start;
4498         unsigned int old_size;
4499         unsigned int size_diff;
4500         int i;
4501         struct btrfs_map_token token;
4502
4503         btrfs_init_map_token(&token);
4504
4505         leaf = path->nodes[0];
4506         slot = path->slots[0];
4507
4508         old_size = btrfs_item_size_nr(leaf, slot);
4509         if (old_size == new_size)
4510                 return;
4511
4512         nritems = btrfs_header_nritems(leaf);
4513         data_end = leaf_data_end(fs_info, leaf);
4514
4515         old_data_start = btrfs_item_offset_nr(leaf, slot);
4516
4517         size_diff = old_size - new_size;
4518
4519         BUG_ON(slot < 0);
4520         BUG_ON(slot >= nritems);
4521
4522         /*
4523          * item0..itemN ... dataN.offset..dataN.size .. data0.size
4524          */
4525         /* first correct the data pointers */
4526         for (i = slot; i < nritems; i++) {
4527                 u32 ioff;
4528                 item = btrfs_item_nr(i);
4529
4530                 ioff = btrfs_token_item_offset(leaf, item, &token);
4531                 btrfs_set_token_item_offset(leaf, item,
4532                                             ioff + size_diff, &token);
4533         }
4534
4535         /* shift the data */
4536         if (from_end) {
4537                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4538                               data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
4539                               data_end, old_data_start + new_size - data_end);
4540         } else {
4541                 struct btrfs_disk_key disk_key;
4542                 u64 offset;
4543
4544                 btrfs_item_key(leaf, &disk_key, slot);
4545
4546                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
4547                         unsigned long ptr;
4548                         struct btrfs_file_extent_item *fi;
4549
4550                         fi = btrfs_item_ptr(leaf, slot,
4551                                             struct btrfs_file_extent_item);
4552                         fi = (struct btrfs_file_extent_item *)(
4553                              (unsigned long)fi - size_diff);
4554
4555                         if (btrfs_file_extent_type(leaf, fi) ==
4556                             BTRFS_FILE_EXTENT_INLINE) {
4557                                 ptr = btrfs_item_ptr_offset(leaf, slot);
4558                                 memmove_extent_buffer(leaf, ptr,
4559                                       (unsigned long)fi,
4560                                       BTRFS_FILE_EXTENT_INLINE_DATA_START);
4561                         }
4562                 }
4563
4564                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4565                               data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
4566                               data_end, old_data_start - data_end);
4567
4568                 offset = btrfs_disk_key_offset(&disk_key);
4569                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
4570                 btrfs_set_item_key(leaf, &disk_key, slot);
4571                 if (slot == 0)
4572                         fixup_low_keys(path, &disk_key, 1);
4573         }
4574
4575         item = btrfs_item_nr(slot);
4576         btrfs_set_item_size(leaf, item, new_size);
4577         btrfs_mark_buffer_dirty(leaf);
4578
4579         if (btrfs_leaf_free_space(fs_info, leaf) < 0) {
4580                 btrfs_print_leaf(leaf);
4581                 BUG();
4582         }
4583 }
4584
4585 /*
4586  * make the item pointed to by the path bigger, data_size is the added size.
4587  */
4588 void btrfs_extend_item(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
4589                        u32 data_size)
4590 {
4591         int slot;
4592         struct extent_buffer *leaf;
4593         struct btrfs_item *item;
4594         u32 nritems;
4595         unsigned int data_end;
4596         unsigned int old_data;
4597         unsigned int old_size;
4598         int i;
4599         struct btrfs_map_token token;
4600
4601         btrfs_init_map_token(&token);
4602
4603         leaf = path->nodes[0];
4604
4605         nritems = btrfs_header_nritems(leaf);
4606         data_end = leaf_data_end(fs_info, leaf);
4607
4608         if (btrfs_leaf_free_space(fs_info, leaf) < data_size) {
4609                 btrfs_print_leaf(leaf);
4610                 BUG();
4611         }
4612         slot = path->slots[0];
4613         old_data = btrfs_item_end_nr(leaf, slot);
4614
4615         BUG_ON(slot < 0);
4616         if (slot >= nritems) {
4617                 btrfs_print_leaf(leaf);
4618                 btrfs_crit(fs_info, "slot %d too large, nritems %d",
4619                            slot, nritems);
4620                 BUG_ON(1);
4621         }
4622
4623         /*
4624          * item0..itemN ... dataN.offset..dataN.size .. data0.size
4625          */
4626         /* first correct the data pointers */
4627         for (i = slot; i < nritems; i++) {
4628                 u32 ioff;
4629                 item = btrfs_item_nr(i);
4630
4631                 ioff = btrfs_token_item_offset(leaf, item, &token);
4632                 btrfs_set_token_item_offset(leaf, item,
4633                                             ioff - data_size, &token);
4634         }
4635
4636         /* shift the data */
4637         memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4638                       data_end - data_size, BTRFS_LEAF_DATA_OFFSET +
4639                       data_end, old_data - data_end);
4640
4641         data_end = old_data;
4642         old_size = btrfs_item_size_nr(leaf, slot);
4643         item = btrfs_item_nr(slot);
4644         btrfs_set_item_size(leaf, item, old_size + data_size);
4645         btrfs_mark_buffer_dirty(leaf);
4646
4647         if (btrfs_leaf_free_space(fs_info, leaf) < 0) {
4648                 btrfs_print_leaf(leaf);
4649                 BUG();
4650         }
4651 }
4652
4653 /*
4654  * this is a helper for btrfs_insert_empty_items, the main goal here is
4655  * to save stack depth by doing the bulk of the work in a function
4656  * that doesn't call btrfs_search_slot
4657  */
4658 void setup_items_for_insert(struct btrfs_root *root, struct btrfs_path *path,
4659                             const struct btrfs_key *cpu_key, u32 *data_size,
4660                             u32 total_data, u32 total_size, int nr)
4661 {
4662         struct btrfs_fs_info *fs_info = root->fs_info;
4663         struct btrfs_item *item;
4664         int i;
4665         u32 nritems;
4666         unsigned int data_end;
4667         struct btrfs_disk_key disk_key;
4668         struct extent_buffer *leaf;
4669         int slot;
4670         struct btrfs_map_token token;
4671
4672         if (path->slots[0] == 0) {
4673                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
4674                 fixup_low_keys(path, &disk_key, 1);
4675         }
4676         btrfs_unlock_up_safe(path, 1);
4677
4678         btrfs_init_map_token(&token);
4679
4680         leaf = path->nodes[0];
4681         slot = path->slots[0];
4682
4683         nritems = btrfs_header_nritems(leaf);
4684         data_end = leaf_data_end(fs_info, leaf);
4685
4686         if (btrfs_leaf_free_space(fs_info, leaf) < total_size) {
4687                 btrfs_print_leaf(leaf);
4688                 btrfs_crit(fs_info, "not enough freespace need %u have %d",
4689                            total_size, btrfs_leaf_free_space(fs_info, leaf));
4690                 BUG();
4691         }
4692
4693         if (slot != nritems) {
4694                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
4695
4696                 if (old_data < data_end) {
4697                         btrfs_print_leaf(leaf);
4698                         btrfs_crit(fs_info, "slot %d old_data %d data_end %d",
4699                                    slot, old_data, data_end);
4700                         BUG_ON(1);
4701                 }
4702                 /*
4703                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
4704                  */
4705                 /* first correct the data pointers */
4706                 for (i = slot; i < nritems; i++) {
4707                         u32 ioff;
4708
4709                         item = btrfs_item_nr(i);
4710                         ioff = btrfs_token_item_offset(leaf, item, &token);
4711                         btrfs_set_token_item_offset(leaf, item,
4712                                                     ioff - total_data, &token);
4713                 }
4714                 /* shift the items */
4715                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
4716                               btrfs_item_nr_offset(slot),
4717                               (nritems - slot) * sizeof(struct btrfs_item));
4718
4719                 /* shift the data */
4720                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4721                               data_end - total_data, BTRFS_LEAF_DATA_OFFSET +
4722                               data_end, old_data - data_end);
4723                 data_end = old_data;
4724         }
4725
4726         /* setup the item for the new data */
4727         for (i = 0; i < nr; i++) {
4728                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
4729                 btrfs_set_item_key(leaf, &disk_key, slot + i);
4730                 item = btrfs_item_nr(slot + i);
4731                 btrfs_set_token_item_offset(leaf, item,
4732                                             data_end - data_size[i], &token);
4733                 data_end -= data_size[i];
4734                 btrfs_set_token_item_size(leaf, item, data_size[i], &token);
4735         }
4736
4737         btrfs_set_header_nritems(leaf, nritems + nr);
4738         btrfs_mark_buffer_dirty(leaf);
4739
4740         if (btrfs_leaf_free_space(fs_info, leaf) < 0) {
4741                 btrfs_print_leaf(leaf);
4742                 BUG();
4743         }
4744 }
4745
4746 /*
4747  * Given a key and some data, insert items into the tree.
4748  * This does all the path init required, making room in the tree if needed.
4749  */
4750 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
4751                             struct btrfs_root *root,
4752                             struct btrfs_path *path,
4753                             const struct btrfs_key *cpu_key, u32 *data_size,
4754                             int nr)
4755 {
4756         int ret = 0;
4757         int slot;
4758         int i;
4759         u32 total_size = 0;
4760         u32 total_data = 0;
4761
4762         for (i = 0; i < nr; i++)
4763                 total_data += data_size[i];
4764
4765         total_size = total_data + (nr * sizeof(struct btrfs_item));
4766         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
4767         if (ret == 0)
4768                 return -EEXIST;
4769         if (ret < 0)
4770                 return ret;
4771
4772         slot = path->slots[0];
4773         BUG_ON(slot < 0);
4774
4775         setup_items_for_insert(root, path, cpu_key, data_size,
4776                                total_data, total_size, nr);
4777         return 0;
4778 }
4779
4780 /*
4781  * Given a key and some data, insert an item into the tree.
4782  * This does all the path init required, making room in the tree if needed.
4783  */
4784 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4785                       const struct btrfs_key *cpu_key, void *data,
4786                       u32 data_size)
4787 {
4788         int ret = 0;
4789         struct btrfs_path *path;
4790         struct extent_buffer *leaf;
4791         unsigned long ptr;
4792
4793         path = btrfs_alloc_path();
4794         if (!path)
4795                 return -ENOMEM;
4796         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
4797         if (!ret) {
4798                 leaf = path->nodes[0];
4799                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
4800                 write_extent_buffer(leaf, data, ptr, data_size);
4801                 btrfs_mark_buffer_dirty(leaf);
4802         }
4803         btrfs_free_path(path);
4804         return ret;
4805 }
4806
4807 /*
4808  * delete the pointer from a given node.
4809  *
4810  * the tree should have been previously balanced so the deletion does not
4811  * empty a node.
4812  */
4813 static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
4814                     int level, int slot)
4815 {
4816         struct extent_buffer *parent = path->nodes[level];
4817         u32 nritems;
4818         int ret;
4819
4820         nritems = btrfs_header_nritems(parent);
4821         if (slot != nritems - 1) {
4822                 if (level) {
4823                         ret = tree_mod_log_insert_move(parent, slot, slot + 1,
4824                                         nritems - slot - 1);
4825                         BUG_ON(ret < 0);
4826                 }
4827                 memmove_extent_buffer(parent,
4828                               btrfs_node_key_ptr_offset(slot),
4829                               btrfs_node_key_ptr_offset(slot + 1),
4830                               sizeof(struct btrfs_key_ptr) *
4831                               (nritems - slot - 1));
4832         } else if (level) {
4833                 ret = tree_mod_log_insert_key(parent, slot, MOD_LOG_KEY_REMOVE,
4834                                 GFP_NOFS);
4835                 BUG_ON(ret < 0);
4836         }
4837
4838         nritems--;
4839         btrfs_set_header_nritems(parent, nritems);
4840         if (nritems == 0 && parent == root->node) {
4841                 BUG_ON(btrfs_header_level(root->node) != 1);
4842                 /* just turn the root into a leaf and break */
4843                 btrfs_set_header_level(root->node, 0);
4844         } else if (slot == 0) {
4845                 struct btrfs_disk_key disk_key;
4846
4847                 btrfs_node_key(parent, &disk_key, 0);
4848                 fixup_low_keys(path, &disk_key, level + 1);
4849         }
4850         btrfs_mark_buffer_dirty(parent);
4851 }
4852
4853 /*
4854  * a helper function to delete the leaf pointed to by path->slots[1] and
4855  * path->nodes[1].
4856  *
4857  * This deletes the pointer in path->nodes[1] and frees the leaf
4858  * block extent.  zero is returned if it all worked out, < 0 otherwise.
4859  *
4860  * The path must have already been setup for deleting the leaf, including
4861  * all the proper balancing.  path->nodes[1] must be locked.
4862  */
4863 static noinline void btrfs_del_leaf(struct btrfs_trans_handle *trans,
4864                                     struct btrfs_root *root,
4865                                     struct btrfs_path *path,
4866                                     struct extent_buffer *leaf)
4867 {
4868         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
4869         del_ptr(root, path, 1, path->slots[1]);
4870
4871         /*
4872          * btrfs_free_extent is expensive, we want to make sure we
4873          * aren't holding any locks when we call it
4874          */
4875         btrfs_unlock_up_safe(path, 0);
4876
4877         root_sub_used(root, leaf->len);
4878
4879         extent_buffer_get(leaf);
4880         btrfs_free_tree_block(trans, root, leaf, 0, 1);
4881         free_extent_buffer_stale(leaf);
4882 }
4883 /*
4884  * delete the item at the leaf level in path.  If that empties
4885  * the leaf, remove it from the tree
4886  */
4887 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4888                     struct btrfs_path *path, int slot, int nr)
4889 {
4890         struct btrfs_fs_info *fs_info = root->fs_info;
4891         struct extent_buffer *leaf;
4892         struct btrfs_item *item;
4893         u32 last_off;
4894         u32 dsize = 0;
4895         int ret = 0;
4896         int wret;
4897         int i;
4898         u32 nritems;
4899         struct btrfs_map_token token;
4900
4901         btrfs_init_map_token(&token);
4902
4903         leaf = path->nodes[0];
4904         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
4905
4906         for (i = 0; i < nr; i++)
4907                 dsize += btrfs_item_size_nr(leaf, slot + i);
4908
4909         nritems = btrfs_header_nritems(leaf);
4910
4911         if (slot + nr != nritems) {
4912                 int data_end = leaf_data_end(fs_info, leaf);
4913
4914                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4915                               data_end + dsize,
4916                               BTRFS_LEAF_DATA_OFFSET + data_end,
4917                               last_off - data_end);
4918
4919                 for (i = slot + nr; i < nritems; i++) {
4920                         u32 ioff;
4921
4922                         item = btrfs_item_nr(i);
4923                         ioff = btrfs_token_item_offset(leaf, item, &token);
4924                         btrfs_set_token_item_offset(leaf, item,
4925                                                     ioff + dsize, &token);
4926                 }
4927
4928                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
4929                               btrfs_item_nr_offset(slot + nr),
4930                               sizeof(struct btrfs_item) *
4931                               (nritems - slot - nr));
4932         }
4933         btrfs_set_header_nritems(leaf, nritems - nr);
4934         nritems -= nr;
4935
4936         /* delete the leaf if we've emptied it */
4937         if (nritems == 0) {
4938                 if (leaf == root->node) {
4939                         btrfs_set_header_level(leaf, 0);
4940                 } else {
4941                         btrfs_set_path_blocking(path);
4942                         clean_tree_block(fs_info, leaf);
4943                         btrfs_del_leaf(trans, root, path, leaf);
4944                 }
4945         } else {
4946                 int used = leaf_space_used(leaf, 0, nritems);
4947                 if (slot == 0) {
4948                         struct btrfs_disk_key disk_key;
4949
4950                         btrfs_item_key(leaf, &disk_key, 0);
4951                         fixup_low_keys(path, &disk_key, 1);
4952                 }
4953
4954                 /* delete the leaf if it is mostly empty */
4955                 if (used < BTRFS_LEAF_DATA_SIZE(fs_info) / 3) {
4956                         /* push_leaf_left fixes the path.
4957                          * make sure the path still points to our leaf
4958                          * for possible call to del_ptr below
4959                          */
4960                         slot = path->slots[1];
4961                         extent_buffer_get(leaf);
4962
4963                         btrfs_set_path_blocking(path);
4964                         wret = push_leaf_left(trans, root, path, 1, 1,
4965                                               1, (u32)-1);
4966                         if (wret < 0 && wret != -ENOSPC)
4967                                 ret = wret;
4968
4969                         if (path->nodes[0] == leaf &&
4970                             btrfs_header_nritems(leaf)) {
4971                                 wret = push_leaf_right(trans, root, path, 1,
4972                                                        1, 1, 0);
4973                                 if (wret < 0 && wret != -ENOSPC)
4974                                         ret = wret;
4975                         }
4976
4977                         if (btrfs_header_nritems(leaf) == 0) {
4978                                 path->slots[1] = slot;
4979                                 btrfs_del_leaf(trans, root, path, leaf);
4980                                 free_extent_buffer(leaf);
4981                                 ret = 0;
4982                         } else {
4983                                 /* if we're still in the path, make sure
4984                                  * we're dirty.  Otherwise, one of the
4985                                  * push_leaf functions must have already
4986                                  * dirtied this buffer
4987                                  */
4988                                 if (path->nodes[0] == leaf)
4989                                         btrfs_mark_buffer_dirty(leaf);
4990                                 free_extent_buffer(leaf);
4991                         }
4992                 } else {
4993                         btrfs_mark_buffer_dirty(leaf);
4994                 }
4995         }
4996         return ret;
4997 }
4998
4999 /*
5000  * search the tree again to find a leaf with lesser keys
5001  * returns 0 if it found something or 1 if there are no lesser leaves.
5002  * returns < 0 on io errors.
5003  *
5004  * This may release the path, and so you may lose any locks held at the
5005  * time you call it.
5006  */
5007 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
5008 {
5009         struct btrfs_key key;
5010         struct btrfs_disk_key found_key;
5011         int ret;
5012
5013         btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
5014
5015         if (key.offset > 0) {
5016                 key.offset--;
5017         } else if (key.type > 0) {
5018                 key.type--;
5019                 key.offset = (u64)-1;
5020         } else if (key.objectid > 0) {
5021                 key.objectid--;
5022                 key.type = (u8)-1;
5023                 key.offset = (u64)-1;
5024         } else {
5025                 return 1;
5026         }
5027
5028         btrfs_release_path(path);
5029         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5030         if (ret < 0)
5031                 return ret;
5032         btrfs_item_key(path->nodes[0], &found_key, 0);
5033         ret = comp_keys(&found_key, &key);
5034         /*
5035          * We might have had an item with the previous key in the tree right
5036          * before we released our path. And after we released our path, that
5037          * item might have been pushed to the first slot (0) of the leaf we
5038          * were holding due to a tree balance. Alternatively, an item with the
5039          * previous key can exist as the only element of a leaf (big fat item).
5040          * Therefore account for these 2 cases, so that our callers (like
5041          * btrfs_previous_item) don't miss an existing item with a key matching
5042          * the previous key we computed above.
5043          */
5044         if (ret <= 0)
5045                 return 0;
5046         return 1;
5047 }
5048
5049 /*
5050  * A helper function to walk down the tree starting at min_key, and looking
5051  * for nodes or leaves that are have a minimum transaction id.
5052  * This is used by the btree defrag code, and tree logging
5053  *
5054  * This does not cow, but it does stuff the starting key it finds back
5055  * into min_key, so you can call btrfs_search_slot with cow=1 on the
5056  * key and get a writable path.
5057  *
5058  * This honors path->lowest_level to prevent descent past a given level
5059  * of the tree.
5060  *
5061  * min_trans indicates the oldest transaction that you are interested
5062  * in walking through.  Any nodes or leaves older than min_trans are
5063  * skipped over (without reading them).
5064  *
5065  * returns zero if something useful was found, < 0 on error and 1 if there
5066  * was nothing in the tree that matched the search criteria.
5067  */
5068 int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
5069                          struct btrfs_path *path,
5070                          u64 min_trans)
5071 {
5072         struct btrfs_fs_info *fs_info = root->fs_info;
5073         struct extent_buffer *cur;
5074         struct btrfs_key found_key;
5075         int slot;
5076         int sret;
5077         u32 nritems;
5078         int level;
5079         int ret = 1;
5080         int keep_locks = path->keep_locks;
5081
5082         path->keep_locks = 1;
5083 again:
5084         cur = btrfs_read_lock_root_node(root);
5085         level = btrfs_header_level(cur);
5086         WARN_ON(path->nodes[level]);
5087         path->nodes[level] = cur;
5088         path->locks[level] = BTRFS_READ_LOCK;
5089
5090         if (btrfs_header_generation(cur) < min_trans) {
5091                 ret = 1;
5092                 goto out;
5093         }
5094         while (1) {
5095                 nritems = btrfs_header_nritems(cur);
5096                 level = btrfs_header_level(cur);
5097                 sret = btrfs_bin_search(cur, min_key, level, &slot);
5098
5099                 /* at the lowest level, we're done, setup the path and exit */
5100                 if (level == path->lowest_level) {
5101                         if (slot >= nritems)
5102                                 goto find_next_key;
5103                         ret = 0;
5104                         path->slots[level] = slot;
5105                         btrfs_item_key_to_cpu(cur, &found_key, slot);
5106                         goto out;
5107                 }
5108                 if (sret && slot > 0)
5109                         slot--;
5110                 /*
5111                  * check this node pointer against the min_trans parameters.
5112                  * If it is too old, old, skip to the next one.
5113                  */
5114                 while (slot < nritems) {
5115                         u64 gen;
5116
5117                         gen = btrfs_node_ptr_generation(cur, slot);
5118                         if (gen < min_trans) {
5119                                 slot++;
5120                                 continue;
5121                         }
5122                         break;
5123                 }
5124 find_next_key:
5125                 /*
5126                  * we didn't find a candidate key in this node, walk forward
5127                  * and find another one
5128                  */
5129                 if (slot >= nritems) {
5130                         path->slots[level] = slot;
5131                         btrfs_set_path_blocking(path);
5132                         sret = btrfs_find_next_key(root, path, min_key, level,
5133                                                   min_trans);
5134                         if (sret == 0) {
5135                                 btrfs_release_path(path);
5136                                 goto again;
5137                         } else {
5138                                 goto out;
5139                         }
5140                 }
5141                 /* save our key for returning back */
5142                 btrfs_node_key_to_cpu(cur, &found_key, slot);
5143                 path->slots[level] = slot;
5144                 if (level == path->lowest_level) {
5145                         ret = 0;
5146                         goto out;
5147                 }
5148                 btrfs_set_path_blocking(path);
5149                 cur = read_node_slot(fs_info, cur, slot);
5150                 if (IS_ERR(cur)) {
5151                         ret = PTR_ERR(cur);
5152                         goto out;
5153                 }
5154
5155                 btrfs_tree_read_lock(cur);
5156
5157                 path->locks[level - 1] = BTRFS_READ_LOCK;
5158                 path->nodes[level - 1] = cur;
5159                 unlock_up(path, level, 1, 0, NULL);
5160         }
5161 out:
5162         path->keep_locks = keep_locks;
5163         if (ret == 0) {
5164                 btrfs_unlock_up_safe(path, path->lowest_level + 1);
5165                 btrfs_set_path_blocking(path);
5166                 memcpy(min_key, &found_key, sizeof(found_key));
5167         }
5168         return ret;
5169 }
5170
5171 static int tree_move_down(struct btrfs_fs_info *fs_info,
5172                            struct btrfs_path *path,
5173                            int *level)
5174 {
5175         struct extent_buffer *eb;
5176
5177         BUG_ON(*level == 0);
5178         eb = read_node_slot(fs_info, path->nodes[*level], path->slots[*level]);
5179         if (IS_ERR(eb))
5180                 return PTR_ERR(eb);
5181
5182         path->nodes[*level - 1] = eb;
5183         path->slots[*level - 1] = 0;
5184         (*level)--;
5185         return 0;
5186 }
5187
5188 static int tree_move_next_or_upnext(struct btrfs_path *path,
5189                                     int *level, int root_level)
5190 {
5191         int ret = 0;
5192         int nritems;
5193         nritems = btrfs_header_nritems(path->nodes[*level]);
5194
5195         path->slots[*level]++;
5196
5197         while (path->slots[*level] >= nritems) {
5198                 if (*level == root_level)
5199                         return -1;
5200
5201                 /* move upnext */
5202                 path->slots[*level] = 0;
5203                 free_extent_buffer(path->nodes[*level]);
5204                 path->nodes[*level] = NULL;
5205                 (*level)++;
5206                 path->slots[*level]++;
5207
5208                 nritems = btrfs_header_nritems(path->nodes[*level]);
5209                 ret = 1;
5210         }
5211         return ret;
5212 }
5213
5214 /*
5215  * Returns 1 if it had to move up and next. 0 is returned if it moved only next
5216  * or down.
5217  */
5218 static int tree_advance(struct btrfs_fs_info *fs_info,
5219                         struct btrfs_path *path,
5220                         int *level, int root_level,
5221                         int allow_down,
5222                         struct btrfs_key *key)
5223 {
5224         int ret;
5225
5226         if (*level == 0 || !allow_down) {
5227                 ret = tree_move_next_or_upnext(path, level, root_level);
5228         } else {
5229                 ret = tree_move_down(fs_info, path, level);
5230         }
5231         if (ret >= 0) {
5232                 if (*level == 0)
5233                         btrfs_item_key_to_cpu(path->nodes[*level], key,
5234                                         path->slots[*level]);
5235                 else
5236                         btrfs_node_key_to_cpu(path->nodes[*level], key,
5237                                         path->slots[*level]);
5238         }
5239         return ret;
5240 }
5241
5242 static int tree_compare_item(struct btrfs_path *left_path,
5243                              struct btrfs_path *right_path,
5244                              char *tmp_buf)
5245 {
5246         int cmp;
5247         int len1, len2;
5248         unsigned long off1, off2;
5249
5250         len1 = btrfs_item_size_nr(left_path->nodes[0], left_path->slots[0]);
5251         len2 = btrfs_item_size_nr(right_path->nodes[0], right_path->slots[0]);
5252         if (len1 != len2)
5253                 return 1;
5254
5255         off1 = btrfs_item_ptr_offset(left_path->nodes[0], left_path->slots[0]);
5256         off2 = btrfs_item_ptr_offset(right_path->nodes[0],
5257                                 right_path->slots[0]);
5258
5259         read_extent_buffer(left_path->nodes[0], tmp_buf, off1, len1);
5260
5261         cmp = memcmp_extent_buffer(right_path->nodes[0], tmp_buf, off2, len1);
5262         if (cmp)
5263                 return 1;
5264         return 0;
5265 }
5266
5267 #define ADVANCE 1
5268 #define ADVANCE_ONLY_NEXT -1
5269
5270 /*
5271  * This function compares two trees and calls the provided callback for
5272  * every changed/new/deleted item it finds.
5273  * If shared tree blocks are encountered, whole subtrees are skipped, making
5274  * the compare pretty fast on snapshotted subvolumes.
5275  *
5276  * This currently works on commit roots only. As commit roots are read only,
5277  * we don't do any locking. The commit roots are protected with transactions.
5278  * Transactions are ended and rejoined when a commit is tried in between.
5279  *
5280  * This function checks for modifications done to the trees while comparing.
5281  * If it detects a change, it aborts immediately.
5282  */
5283 int btrfs_compare_trees(struct btrfs_root *left_root,
5284                         struct btrfs_root *right_root,
5285                         btrfs_changed_cb_t changed_cb, void *ctx)
5286 {
5287         struct btrfs_fs_info *fs_info = left_root->fs_info;
5288         int ret;
5289         int cmp;
5290         struct btrfs_path *left_path = NULL;
5291         struct btrfs_path *right_path = NULL;
5292         struct btrfs_key left_key;
5293         struct btrfs_key right_key;
5294         char *tmp_buf = NULL;
5295         int left_root_level;
5296         int right_root_level;
5297         int left_level;
5298         int right_level;
5299         int left_end_reached;
5300         int right_end_reached;
5301         int advance_left;
5302         int advance_right;
5303         u64 left_blockptr;
5304         u64 right_blockptr;
5305         u64 left_gen;
5306         u64 right_gen;
5307
5308         left_path = btrfs_alloc_path();
5309         if (!left_path) {
5310                 ret = -ENOMEM;
5311                 goto out;
5312         }
5313         right_path = btrfs_alloc_path();
5314         if (!right_path) {
5315                 ret = -ENOMEM;
5316                 goto out;
5317         }
5318
5319         tmp_buf = kvmalloc(fs_info->nodesize, GFP_KERNEL);
5320         if (!tmp_buf) {
5321                 ret = -ENOMEM;
5322                 goto out;
5323         }
5324
5325         left_path->search_commit_root = 1;
5326         left_path->skip_locking = 1;
5327         right_path->search_commit_root = 1;
5328         right_path->skip_locking = 1;
5329
5330         /*
5331          * Strategy: Go to the first items of both trees. Then do
5332          *
5333          * If both trees are at level 0
5334          *   Compare keys of current items
5335          *     If left < right treat left item as new, advance left tree
5336          *       and repeat
5337          *     If left > right treat right item as deleted, advance right tree
5338          *       and repeat
5339          *     If left == right do deep compare of items, treat as changed if
5340          *       needed, advance both trees and repeat
5341          * If both trees are at the same level but not at level 0
5342          *   Compare keys of current nodes/leafs
5343          *     If left < right advance left tree and repeat
5344          *     If left > right advance right tree and repeat
5345          *     If left == right compare blockptrs of the next nodes/leafs
5346          *       If they match advance both trees but stay at the same level
5347          *         and repeat
5348          *       If they don't match advance both trees while allowing to go
5349          *         deeper and repeat
5350          * If tree levels are different
5351          *   Advance the tree that needs it and repeat
5352          *
5353          * Advancing a tree means:
5354          *   If we are at level 0, try to go to the next slot. If that's not
5355          *   possible, go one level up and repeat. Stop when we found a level
5356          *   where we could go to the next slot. We may at this point be on a
5357          *   node or a leaf.
5358          *
5359          *   If we are not at level 0 and not on shared tree blocks, go one
5360          *   level deeper.
5361          *
5362          *   If we are not at level 0 and on shared tree blocks, go one slot to
5363          *   the right if possible or go up and right.
5364          */
5365
5366         down_read(&fs_info->commit_root_sem);
5367         left_level = btrfs_header_level(left_root->commit_root);
5368         left_root_level = left_level;
5369         left_path->nodes[left_level] =
5370                         btrfs_clone_extent_buffer(left_root->commit_root);
5371         if (!left_path->nodes[left_level]) {
5372                 up_read(&fs_info->commit_root_sem);
5373                 ret = -ENOMEM;
5374                 goto out;
5375         }
5376         extent_buffer_get(left_path->nodes[left_level]);
5377
5378         right_level = btrfs_header_level(right_root->commit_root);
5379         right_root_level = right_level;
5380         right_path->nodes[right_level] =
5381                         btrfs_clone_extent_buffer(right_root->commit_root);
5382         if (!right_path->nodes[right_level]) {
5383                 up_read(&fs_info->commit_root_sem);
5384                 ret = -ENOMEM;
5385                 goto out;
5386         }
5387         extent_buffer_get(right_path->nodes[right_level]);
5388         up_read(&fs_info->commit_root_sem);
5389
5390         if (left_level == 0)
5391                 btrfs_item_key_to_cpu(left_path->nodes[left_level],
5392                                 &left_key, left_path->slots[left_level]);
5393         else
5394                 btrfs_node_key_to_cpu(left_path->nodes[left_level],
5395                                 &left_key, left_path->slots[left_level]);
5396         if (right_level == 0)
5397                 btrfs_item_key_to_cpu(right_path->nodes[right_level],
5398                                 &right_key, right_path->slots[right_level]);
5399         else
5400                 btrfs_node_key_to_cpu(right_path->nodes[right_level],
5401                                 &right_key, right_path->slots[right_level]);
5402
5403         left_end_reached = right_end_reached = 0;
5404         advance_left = advance_right = 0;
5405
5406         while (1) {
5407                 if (advance_left && !left_end_reached) {
5408                         ret = tree_advance(fs_info, left_path, &left_level,
5409                                         left_root_level,
5410                                         advance_left != ADVANCE_ONLY_NEXT,
5411                                         &left_key);
5412                         if (ret == -1)
5413                                 left_end_reached = ADVANCE;
5414                         else if (ret < 0)
5415                                 goto out;
5416                         advance_left = 0;
5417                 }
5418                 if (advance_right && !right_end_reached) {
5419                         ret = tree_advance(fs_info, right_path, &right_level,
5420                                         right_root_level,
5421                                         advance_right != ADVANCE_ONLY_NEXT,
5422                                         &right_key);
5423                         if (ret == -1)
5424                                 right_end_reached = ADVANCE;
5425                         else if (ret < 0)
5426                                 goto out;
5427                         advance_right = 0;
5428                 }
5429
5430                 if (left_end_reached && right_end_reached) {
5431                         ret = 0;
5432                         goto out;
5433                 } else if (left_end_reached) {
5434                         if (right_level == 0) {
5435                                 ret = changed_cb(left_path, right_path,
5436                                                 &right_key,
5437                                                 BTRFS_COMPARE_TREE_DELETED,
5438                                                 ctx);
5439                                 if (ret < 0)
5440                                         goto out;
5441                         }
5442                         advance_right = ADVANCE;
5443                         continue;
5444                 } else if (right_end_reached) {
5445                         if (left_level == 0) {
5446                                 ret = changed_cb(left_path, right_path,
5447                                                 &left_key,
5448                                                 BTRFS_COMPARE_TREE_NEW,
5449                                                 ctx);
5450                                 if (ret < 0)
5451                                         goto out;
5452                         }
5453                         advance_left = ADVANCE;
5454                         continue;
5455                 }
5456
5457                 if (left_level == 0 && right_level == 0) {
5458                         cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
5459                         if (cmp < 0) {
5460                                 ret = changed_cb(left_path, right_path,
5461                                                 &left_key,
5462                                                 BTRFS_COMPARE_TREE_NEW,
5463                                                 ctx);
5464                                 if (ret < 0)
5465                                         goto out;
5466                                 advance_left = ADVANCE;
5467                         } else if (cmp > 0) {
5468                                 ret = changed_cb(left_path, right_path,
5469                                                 &right_key,
5470                                                 BTRFS_COMPARE_TREE_DELETED,
5471                                                 ctx);
5472                                 if (ret < 0)
5473                                         goto out;
5474                                 advance_right = ADVANCE;
5475                         } else {
5476                                 enum btrfs_compare_tree_result result;
5477
5478                                 WARN_ON(!extent_buffer_uptodate(left_path->nodes[0]));
5479                                 ret = tree_compare_item(left_path, right_path,
5480                                                         tmp_buf);
5481                                 if (ret)
5482                                         result = BTRFS_COMPARE_TREE_CHANGED;
5483                                 else
5484                                         result = BTRFS_COMPARE_TREE_SAME;
5485                                 ret = changed_cb(left_path, right_path,
5486                                                  &left_key, result, ctx);
5487                                 if (ret < 0)
5488                                         goto out;
5489                                 advance_left = ADVANCE;
5490                                 advance_right = ADVANCE;
5491                         }
5492                 } else if (left_level == right_level) {
5493                         cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
5494                         if (cmp < 0) {
5495                                 advance_left = ADVANCE;
5496                         } else if (cmp > 0) {
5497                                 advance_right = ADVANCE;
5498                         } else {
5499                                 left_blockptr = btrfs_node_blockptr(
5500                                                 left_path->nodes[left_level],
5501                                                 left_path->slots[left_level]);
5502                                 right_blockptr = btrfs_node_blockptr(
5503                                                 right_path->nodes[right_level],
5504                                                 right_path->slots[right_level]);
5505                                 left_gen = btrfs_node_ptr_generation(
5506                                                 left_path->nodes[left_level],
5507                                                 left_path->slots[left_level]);
5508                                 right_gen = btrfs_node_ptr_generation(
5509                                                 right_path->nodes[right_level],
5510                                                 right_path->slots[right_level]);
5511                                 if (left_blockptr == right_blockptr &&
5512                                     left_gen == right_gen) {
5513                                         /*
5514                                          * As we're on a shared block, don't
5515                                          * allow to go deeper.
5516                                          */
5517                                         advance_left = ADVANCE_ONLY_NEXT;
5518                                         advance_right = ADVANCE_ONLY_NEXT;
5519                                 } else {
5520                                         advance_left = ADVANCE;
5521                                         advance_right = ADVANCE;
5522                                 }
5523                         }
5524                 } else if (left_level < right_level) {
5525                         advance_right = ADVANCE;
5526                 } else {
5527                         advance_left = ADVANCE;
5528                 }
5529         }
5530
5531 out:
5532         btrfs_free_path(left_path);
5533         btrfs_free_path(right_path);
5534         kvfree(tmp_buf);
5535         return ret;
5536 }
5537
5538 /*
5539  * this is similar to btrfs_next_leaf, but does not try to preserve
5540  * and fixup the path.  It looks for and returns the next key in the
5541  * tree based on the current path and the min_trans parameters.
5542  *
5543  * 0 is returned if another key is found, < 0 if there are any errors
5544  * and 1 is returned if there are no higher keys in the tree
5545  *
5546  * path->keep_locks should be set to 1 on the search made before
5547  * calling this function.
5548  */
5549 int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
5550                         struct btrfs_key *key, int level, u64 min_trans)
5551 {
5552         int slot;
5553         struct extent_buffer *c;
5554
5555         WARN_ON(!path->keep_locks);
5556         while (level < BTRFS_MAX_LEVEL) {
5557                 if (!path->nodes[level])
5558                         return 1;
5559
5560                 slot = path->slots[level] + 1;
5561                 c = path->nodes[level];
5562 next:
5563                 if (slot >= btrfs_header_nritems(c)) {
5564                         int ret;
5565                         int orig_lowest;
5566                         struct btrfs_key cur_key;
5567                         if (level + 1 >= BTRFS_MAX_LEVEL ||
5568                             !path->nodes[level + 1])
5569                                 return 1;
5570
5571                         if (path->locks[level + 1]) {
5572                                 level++;
5573                                 continue;
5574                         }
5575
5576                         slot = btrfs_header_nritems(c) - 1;
5577                         if (level == 0)
5578                                 btrfs_item_key_to_cpu(c, &cur_key, slot);
5579                         else
5580                                 btrfs_node_key_to_cpu(c, &cur_key, slot);
5581
5582                         orig_lowest = path->lowest_level;
5583                         btrfs_release_path(path);
5584                         path->lowest_level = level;
5585                         ret = btrfs_search_slot(NULL, root, &cur_key, path,
5586                                                 0, 0);
5587                         path->lowest_level = orig_lowest;
5588                         if (ret < 0)
5589                                 return ret;
5590
5591                         c = path->nodes[level];
5592                         slot = path->slots[level];
5593                         if (ret == 0)
5594                                 slot++;
5595                         goto next;
5596                 }
5597
5598                 if (level == 0)
5599                         btrfs_item_key_to_cpu(c, key, slot);
5600                 else {
5601                         u64 gen = btrfs_node_ptr_generation(c, slot);
5602
5603                         if (gen < min_trans) {
5604                                 slot++;
5605                                 goto next;
5606                         }
5607                         btrfs_node_key_to_cpu(c, key, slot);
5608                 }
5609                 return 0;
5610         }
5611         return 1;
5612 }
5613
5614 /*
5615  * search the tree again to find a leaf with greater keys
5616  * returns 0 if it found something or 1 if there are no greater leaves.
5617  * returns < 0 on io errors.
5618  */
5619 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
5620 {
5621         return btrfs_next_old_leaf(root, path, 0);
5622 }
5623
5624 int btrfs_next_old_leaf(struct btrfs_root *root, struct btrfs_path *path,
5625                         u64 time_seq)
5626 {
5627         int slot;
5628         int level;
5629         struct extent_buffer *c;
5630         struct extent_buffer *next;
5631         struct btrfs_key key;
5632         u32 nritems;
5633         int ret;
5634         int old_spinning = path->leave_spinning;
5635         int next_rw_lock = 0;
5636
5637         nritems = btrfs_header_nritems(path->nodes[0]);
5638         if (nritems == 0)
5639                 return 1;
5640
5641         btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
5642 again:
5643         level = 1;
5644         next = NULL;
5645         next_rw_lock = 0;
5646         btrfs_release_path(path);
5647
5648         path->keep_locks = 1;
5649         path->leave_spinning = 1;
5650
5651         if (time_seq)
5652                 ret = btrfs_search_old_slot(root, &key, path, time_seq);
5653         else
5654                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5655         path->keep_locks = 0;
5656
5657         if (ret < 0)
5658                 return ret;
5659
5660         nritems = btrfs_header_nritems(path->nodes[0]);
5661         /*
5662          * by releasing the path above we dropped all our locks.  A balance
5663          * could have added more items next to the key that used to be
5664          * at the very end of the block.  So, check again here and
5665          * advance the path if there are now more items available.
5666          */
5667         if (nritems > 0 && path->slots[0] < nritems - 1) {
5668                 if (ret == 0)
5669                         path->slots[0]++;
5670                 ret = 0;
5671                 goto done;
5672         }
5673         /*
5674          * So the above check misses one case:
5675          * - after releasing the path above, someone has removed the item that
5676          *   used to be at the very end of the block, and balance between leafs
5677          *   gets another one with bigger key.offset to replace it.
5678          *
5679          * This one should be returned as well, or we can get leaf corruption
5680          * later(esp. in __btrfs_drop_extents()).
5681          *
5682          * And a bit more explanation about this check,
5683          * with ret > 0, the key isn't found, the path points to the slot
5684          * where it should be inserted, so the path->slots[0] item must be the
5685          * bigger one.
5686          */
5687         if (nritems > 0 && ret > 0 && path->slots[0] == nritems - 1) {
5688                 ret = 0;
5689                 goto done;
5690         }
5691
5692         while (level < BTRFS_MAX_LEVEL) {
5693                 if (!path->nodes[level]) {
5694                         ret = 1;
5695                         goto done;
5696                 }
5697
5698                 slot = path->slots[level] + 1;
5699                 c = path->nodes[level];
5700                 if (slot >= btrfs_header_nritems(c)) {
5701                         level++;
5702                         if (level == BTRFS_MAX_LEVEL) {
5703                                 ret = 1;
5704                                 goto done;
5705                         }
5706                         continue;
5707                 }
5708
5709                 if (next) {
5710                         btrfs_tree_unlock_rw(next, next_rw_lock);
5711                         free_extent_buffer(next);
5712                 }
5713
5714                 next = c;
5715                 next_rw_lock = path->locks[level];
5716                 ret = read_block_for_search(root, path, &next, level,
5717                                             slot, &key);
5718                 if (ret == -EAGAIN)
5719                         goto again;
5720
5721                 if (ret < 0) {
5722                         btrfs_release_path(path);
5723                         goto done;
5724                 }
5725
5726                 if (!path->skip_locking) {
5727                         ret = btrfs_try_tree_read_lock(next);
5728                         if (!ret && time_seq) {
5729                                 /*
5730                                  * If we don't get the lock, we may be racing
5731                                  * with push_leaf_left, holding that lock while
5732                                  * itself waiting for the leaf we've currently
5733                                  * locked. To solve this situation, we give up
5734                                  * on our lock and cycle.
5735                                  */
5736                                 free_extent_buffer(next);
5737                                 btrfs_release_path(path);
5738                                 cond_resched();
5739                                 goto again;
5740                         }
5741                         if (!ret) {
5742                                 btrfs_set_path_blocking(path);
5743                                 btrfs_tree_read_lock(next);
5744                         }
5745                         next_rw_lock = BTRFS_READ_LOCK;
5746                 }
5747                 break;
5748         }
5749         path->slots[level] = slot;
5750         while (1) {
5751                 level--;
5752                 c = path->nodes[level];
5753                 if (path->locks[level])
5754                         btrfs_tree_unlock_rw(c, path->locks[level]);
5755
5756                 free_extent_buffer(c);
5757                 path->nodes[level] = next;
5758                 path->slots[level] = 0;
5759                 if (!path->skip_locking)
5760                         path->locks[level] = next_rw_lock;
5761                 if (!level)
5762                         break;
5763
5764                 ret = read_block_for_search(root, path, &next, level,
5765                                             0, &key);
5766                 if (ret == -EAGAIN)
5767                         goto again;
5768
5769                 if (ret < 0) {
5770                         btrfs_release_path(path);
5771                         goto done;
5772                 }
5773
5774                 if (!path->skip_locking) {
5775                         ret = btrfs_try_tree_read_lock(next);
5776                         if (!ret) {
5777                                 btrfs_set_path_blocking(path);
5778                                 btrfs_tree_read_lock(next);
5779                         }
5780                         next_rw_lock = BTRFS_READ_LOCK;
5781                 }
5782         }
5783         ret = 0;
5784 done:
5785         unlock_up(path, 0, 1, 0, NULL);
5786         path->leave_spinning = old_spinning;
5787         if (!old_spinning)
5788                 btrfs_set_path_blocking(path);
5789
5790         return ret;
5791 }
5792
5793 /*
5794  * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
5795  * searching until it gets past min_objectid or finds an item of 'type'
5796  *
5797  * returns 0 if something is found, 1 if nothing was found and < 0 on error
5798  */
5799 int btrfs_previous_item(struct btrfs_root *root,
5800                         struct btrfs_path *path, u64 min_objectid,
5801                         int type)
5802 {
5803         struct btrfs_key found_key;
5804         struct extent_buffer *leaf;
5805         u32 nritems;
5806         int ret;
5807
5808         while (1) {
5809                 if (path->slots[0] == 0) {
5810                         btrfs_set_path_blocking(path);
5811                         ret = btrfs_prev_leaf(root, path);
5812                         if (ret != 0)
5813                                 return ret;
5814                 } else {
5815                         path->slots[0]--;
5816                 }
5817                 leaf = path->nodes[0];
5818                 nritems = btrfs_header_nritems(leaf);
5819                 if (nritems == 0)
5820                         return 1;
5821                 if (path->slots[0] == nritems)
5822                         path->slots[0]--;
5823
5824                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5825                 if (found_key.objectid < min_objectid)
5826                         break;
5827                 if (found_key.type == type)
5828                         return 0;
5829                 if (found_key.objectid == min_objectid &&
5830                     found_key.type < type)
5831                         break;
5832         }
5833         return 1;
5834 }
5835
5836 /*
5837  * search in extent tree to find a previous Metadata/Data extent item with
5838  * min objecitd.
5839  *
5840  * returns 0 if something is found, 1 if nothing was found and < 0 on error
5841  */
5842 int btrfs_previous_extent_item(struct btrfs_root *root,
5843                         struct btrfs_path *path, u64 min_objectid)
5844 {
5845         struct btrfs_key found_key;
5846         struct extent_buffer *leaf;
5847         u32 nritems;
5848         int ret;
5849
5850         while (1) {
5851                 if (path->slots[0] == 0) {
5852                         btrfs_set_path_blocking(path);
5853                         ret = btrfs_prev_leaf(root, path);
5854                         if (ret != 0)
5855                                 return ret;
5856                 } else {
5857                         path->slots[0]--;
5858                 }
5859                 leaf = path->nodes[0];
5860                 nritems = btrfs_header_nritems(leaf);
5861                 if (nritems == 0)
5862                         return 1;
5863                 if (path->slots[0] == nritems)
5864                         path->slots[0]--;
5865
5866                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5867                 if (found_key.objectid < min_objectid)
5868                         break;
5869                 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
5870                     found_key.type == BTRFS_METADATA_ITEM_KEY)
5871                         return 0;
5872                 if (found_key.objectid == min_objectid &&
5873                     found_key.type < BTRFS_EXTENT_ITEM_KEY)
5874                         break;
5875         }
5876         return 1;
5877 }