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