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 static const struct btrfs_csums {
  33         u16             size;
  34         const char      *name;
  35 } btrfs_csums[] = {
  36         [BTRFS_CSUM_TYPE_CRC32] = { .size = 4, .name = "crc32c" },
  37 };
  38
  39 int btrfs_super_csum_size(const struct btrfs_super_block *s)
  40 {
  41         u16 t = btrfs_super_csum_type(s);
  42         /*
  43          * csum type is validated at mount time
  44          */
  45         return btrfs_csums[t].size;
  46 }
  47
  48 const char *btrfs_super_csum_name(u16 csum_type)
  49 {
  50         /* csum type is validated at mount time */
  51         return btrfs_csums[csum_type].name;
  52 }
  53
  54 struct btrfs_path *btrfs_alloc_path(void)
  55 {
  56         return kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
  57 }
  58
  59 /*
  60  * set all locked nodes in the path to blocking locks.  This should
  61  * be done before scheduling
  62  */
  63 noinline void btrfs_set_path_blocking(struct btrfs_path *p)
  64 {
  65         int i;
  66         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  67                 if (!p->nodes[i] || !p->locks[i])
  68                         continue;
  69                 /*
  70                  * If we currently have a spinning reader or writer lock this
  71                  * will bump the count of blocking holders and drop the
  72                  * spinlock.
  73                  */
  74                 if (p->locks[i] == BTRFS_READ_LOCK) {
  75                         btrfs_set_lock_blocking_read(p->nodes[i]);
  76                         p->locks[i] = BTRFS_READ_LOCK_BLOCKING;
  77                 } else if (p->locks[i] == BTRFS_WRITE_LOCK) {
  78                         btrfs_set_lock_blocking_write(p->nodes[i]);
  79                         p->locks[i] = BTRFS_WRITE_LOCK_BLOCKING;
  80                 }
  81         }
  82 }
  83
  84 /* this also releases the path */
  85 void btrfs_free_path(struct btrfs_path *p)
  86 {
  87         if (!p)
  88                 return;
  89         btrfs_release_path(p);
  90         kmem_cache_free(btrfs_path_cachep, p);
  91 }
  92
  93 /*
  94  * path release drops references on the extent buffers in the path
  95  * and it drops any locks held by this path
  96  *
  97  * It is safe to call this on paths that no locks or extent buffers held.
  98  */
  99 noinline void btrfs_release_path(struct btrfs_path *p)
 100 {
 101         int i;
 102
 103         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
 104                 p->slots[i] = 0;
 105                 if (!p->nodes[i])
 106                         continue;
 107                 if (p->locks[i]) {
 108                         btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
 109                         p->locks[i] = 0;
 110                 }
 111                 free_extent_buffer(p->nodes[i]);
 112                 p->nodes[i] = NULL;
 113         }
 114 }
 115
 116 /*
 117  * safely gets a reference on the root node of a tree.  A lock
 118  * is not taken, so a concurrent writer may put a different node
 119  * at the root of the tree.  See btrfs_lock_root_node for the
 120  * looping required.
 121  *
 122  * The extent buffer returned by this has a reference taken, so
 123  * it won't disappear.  It may stop being the root of the tree
 124  * at any time because there are no locks held.
 125  */
 126 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
 127 {
 128         struct extent_buffer *eb;
 129
 130         while (1) {
 131                 rcu_read_lock();
 132                 eb = rcu_dereference(root->node);
 133
 134                 /*
 135                  * RCU really hurts here, we could free up the root node because
 136                  * it was COWed but we may not get the new root node yet so do
 137                  * the inc_not_zero dance and if it doesn't work then
 138                  * synchronize_rcu and try again.
 139                  */
 140                 if (atomic_inc_not_zero(&eb->refs)) {
 141                         rcu_read_unlock();
 142                         break;
 143                 }
 144                 rcu_read_unlock();
 145                 synchronize_rcu();
 146         }
 147         return eb;
 148 }
 149
 150 /* loop around taking references on and locking the root node of the
 151  * tree until you end up with a lock on the root.  A locked buffer
 152  * is returned, with a reference held.
 153  */
 154 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
 155 {
 156         struct extent_buffer *eb;
 157
 158         while (1) {
 159                 eb = btrfs_root_node(root);
 160                 btrfs_tree_lock(eb);
 161                 if (eb == root->node)
 162                         break;
 163                 btrfs_tree_unlock(eb);
 164                 free_extent_buffer(eb);
 165         }
 166         return eb;
 167 }
 168
 169 /* loop around taking references on and locking the root node of the
 170  * tree until you end up with a lock on the root.  A locked buffer
 171  * is returned, with a reference held.
 172  */
 173 struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
 174 {
 175         struct extent_buffer *eb;
 176
 177         while (1) {
 178                 eb = btrfs_root_node(root);
 179                 btrfs_tree_read_lock(eb);
 180                 if (eb == root->node)
 181                         break;
 182                 btrfs_tree_read_unlock(eb);
 183                 free_extent_buffer(eb);
 184         }
 185         return eb;
 186 }
 187
 188 /* cowonly root (everything not a reference counted cow subvolume), just get
 189  * put onto a simple dirty list.  transaction.c walks this to make sure they
 190  * get properly updated on disk.
 191  */
 192 static void add_root_to_dirty_list(struct btrfs_root *root)
 193 {
 194         struct btrfs_fs_info *fs_info = root->fs_info;
 195
 196         if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
 197             !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
 198                 return;
 199
 200         spin_lock(&fs_info->trans_lock);
 201         if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
 202                 /* Want the extent tree to be the last on the list */
 203                 if (root->root_key.objectid == BTRFS_EXTENT_TREE_OBJECTID)
 204                         list_move_tail(&root->dirty_list,
 205                                        &fs_info->dirty_cowonly_roots);
 206                 else
 207                         list_move(&root->dirty_list,
 208                                   &fs_info->dirty_cowonly_roots);
 209         }
 210         spin_unlock(&fs_info->trans_lock);
 211 }
 212
 213 /*
 214  * used by snapshot creation to make a copy of a root for a tree with
 215  * a given objectid.  The buffer with the new root node is returned in
 216  * cow_ret, and this func returns zero on success or a negative error code.
 217  */
 218 int btrfs_copy_root(struct btrfs_trans_handle *trans,
 219                       struct btrfs_root *root,
 220                       struct extent_buffer *buf,
 221                       struct extent_buffer **cow_ret, u64 new_root_objectid)
 222 {
 223         struct btrfs_fs_info *fs_info = root->fs_info;
 224         struct extent_buffer *cow;
 225         int ret = 0;
 226         int level;
 227         struct btrfs_disk_key disk_key;
 228
 229         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 230                 trans->transid != fs_info->running_transaction->transid);
 231         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 232                 trans->transid != root->last_trans);
 233
 234         level = btrfs_header_level(buf);
 235         if (level == 0)
 236                 btrfs_item_key(buf, &disk_key, 0);
 237         else
 238                 btrfs_node_key(buf, &disk_key, 0);
 239
 240         cow = btrfs_alloc_tree_block(trans, root, 0, new_root_objectid,
 241                         &disk_key, level, buf->start, 0);
 242         if (IS_ERR(cow))
 243                 return PTR_ERR(cow);
 244
 245         copy_extent_buffer_full(cow, buf);
 246         btrfs_set_header_bytenr(cow, cow->start);
 247         btrfs_set_header_generation(cow, trans->transid);
 248         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 249         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 250                                      BTRFS_HEADER_FLAG_RELOC);
 251         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 252                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 253         else
 254                 btrfs_set_header_owner(cow, new_root_objectid);
 255
 256         write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
 257
 258         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 259         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 260                 ret = btrfs_inc_ref(trans, root, cow, 1);
 261         else
 262                 ret = btrfs_inc_ref(trans, root, cow, 0);
 263
 264         if (ret)
 265                 return ret;
 266
 267         btrfs_mark_buffer_dirty(cow);
 268         *cow_ret = cow;
 269         return 0;
 270 }
 271
 272 enum mod_log_op {
 273         MOD_LOG_KEY_REPLACE,
 274         MOD_LOG_KEY_ADD,
 275         MOD_LOG_KEY_REMOVE,
 276         MOD_LOG_KEY_REMOVE_WHILE_FREEING,
 277         MOD_LOG_KEY_REMOVE_WHILE_MOVING,
 278         MOD_LOG_MOVE_KEYS,
 279         MOD_LOG_ROOT_REPLACE,
 280 };
 281
 282 struct tree_mod_root {
 283         u64 logical;
 284         u8 level;
 285 };
 286
 287 struct tree_mod_elem {
 288         struct rb_node node;
 289         u64 logical;
 290         u64 seq;
 291         enum mod_log_op op;
 292
 293         /* this is used for MOD_LOG_KEY_* and MOD_LOG_MOVE_KEYS operations */
 294         int slot;
 295
 296         /* this is used for MOD_LOG_KEY* and MOD_LOG_ROOT_REPLACE */
 297         u64 generation;
 298
 299         /* those are used for op == MOD_LOG_KEY_{REPLACE,REMOVE} */
 300         struct btrfs_disk_key key;
 301         u64 blockptr;
 302
 303         /* this is used for op == MOD_LOG_MOVE_KEYS */
 304         struct {
 305                 int dst_slot;
 306                 int nr_items;
 307         } move;
 308
 309         /* this is used for op == MOD_LOG_ROOT_REPLACE */
 310         struct tree_mod_root old_root;
 311 };
 312
 313 /*
 314  * Pull a new tree mod seq number for our operation.
 315  */
 316 static inline u64 btrfs_inc_tree_mod_seq(struct btrfs_fs_info *fs_info)
 317 {
 318         return atomic64_inc_return(&fs_info->tree_mod_seq);
 319 }
 320
 321 /*
 322  * This adds a new blocker to the tree mod log's blocker list if the @elem
 323  * passed does not already have a sequence number set. So when a caller expects
 324  * to record tree modifications, it should ensure to set elem->seq to zero
 325  * before calling btrfs_get_tree_mod_seq.
 326  * Returns a fresh, unused tree log modification sequence number, even if no new
 327  * blocker was added.
 328  */
 329 u64 btrfs_get_tree_mod_seq(struct btrfs_fs_info *fs_info,
 330                            struct seq_list *elem)
 331 {
 332         write_lock(&fs_info->tree_mod_log_lock);
 333         spin_lock(&fs_info->tree_mod_seq_lock);
 334         if (!elem->seq) {
 335                 elem->seq = btrfs_inc_tree_mod_seq(fs_info);
 336                 list_add_tail(&elem->list, &fs_info->tree_mod_seq_list);
 337         }
 338         spin_unlock(&fs_info->tree_mod_seq_lock);
 339         write_unlock(&fs_info->tree_mod_log_lock);
 340
 341         return elem->seq;
 342 }
 343
 344 void btrfs_put_tree_mod_seq(struct btrfs_fs_info *fs_info,
 345                             struct seq_list *elem)
 346 {
 347         struct rb_root *tm_root;
 348         struct rb_node *node;
 349         struct rb_node *next;
 350         struct seq_list *cur_elem;
 351         struct tree_mod_elem *tm;
 352         u64 min_seq = (u64)-1;
 353         u64 seq_putting = elem->seq;
 354
 355         if (!seq_putting)
 356                 return;
 357
 358         spin_lock(&fs_info->tree_mod_seq_lock);
 359         list_del(&elem->list);
 360         elem->seq = 0;
 361
 362         list_for_each_entry(cur_elem, &fs_info->tree_mod_seq_list, list) {
 363                 if (cur_elem->seq < min_seq) {
 364                         if (seq_putting > cur_elem->seq) {
 365                                 /*
 366                                  * blocker with lower sequence number exists, we
 367                                  * cannot remove anything from the log
 368                                  */
 369                                 spin_unlock(&fs_info->tree_mod_seq_lock);
 370                                 return;
 371                         }
 372                         min_seq = cur_elem->seq;
 373                 }
 374         }
 375         spin_unlock(&fs_info->tree_mod_seq_lock);
 376
 377         /*
 378          * anything that's lower than the lowest existing (read: blocked)
 379          * sequence number can be removed from the tree.
 380          */
 381         write_lock(&fs_info->tree_mod_log_lock);
 382         tm_root = &fs_info->tree_mod_log;
 383         for (node = rb_first(tm_root); node; node = next) {
 384                 next = rb_next(node);
 385                 tm = rb_entry(node, struct tree_mod_elem, node);
 386                 if (tm->seq > min_seq)
 387                         continue;
 388                 rb_erase(node, tm_root);
 389                 kfree(tm);
 390         }
 391         write_unlock(&fs_info->tree_mod_log_lock);
 392 }
 393
 394 /*
 395  * key order of the log:
 396  *       node/leaf start address -> sequence
 397  *
 398  * The 'start address' is the logical address of the *new* root node
 399  * for root replace operations, or the logical address of the affected
 400  * block for all other operations.
 401  */
 402 static noinline int
 403 __tree_mod_log_insert(struct btrfs_fs_info *fs_info, struct tree_mod_elem *tm)
 404 {
 405         struct rb_root *tm_root;
 406         struct rb_node **new;
 407         struct rb_node *parent = NULL;
 408         struct tree_mod_elem *cur;
 409
 410         lockdep_assert_held_write(&fs_info->tree_mod_log_lock);
 411
 412         tm->seq = btrfs_inc_tree_mod_seq(fs_info);
 413
 414         tm_root = &fs_info->tree_mod_log;
 415         new = &tm_root->rb_node;
 416         while (*new) {
 417                 cur = rb_entry(*new, struct tree_mod_elem, node);
 418                 parent = *new;
 419                 if (cur->logical < tm->logical)
 420                         new = &((*new)->rb_left);
 421                 else if (cur->logical > tm->logical)
 422                         new = &((*new)->rb_right);
 423                 else if (cur->seq < tm->seq)
 424                         new = &((*new)->rb_left);
 425                 else if (cur->seq > tm->seq)
 426                         new = &((*new)->rb_right);
 427                 else
 428                         return -EEXIST;
 429         }
 430
 431         rb_link_node(&tm->node, parent, new);
 432         rb_insert_color(&tm->node, tm_root);
 433         return 0;
 434 }
 435
 436 /*
 437  * Determines if logging can be omitted. Returns 1 if it can. Otherwise, it
 438  * returns zero with the tree_mod_log_lock acquired. The caller must hold
 439  * this until all tree mod log insertions are recorded in the rb tree and then
 440  * write unlock fs_info::tree_mod_log_lock.
 441  */
 442 static inline int tree_mod_dont_log(struct btrfs_fs_info *fs_info,
 443                                     struct extent_buffer *eb) {
 444         smp_mb();
 445         if (list_empty(&(fs_info)->tree_mod_seq_list))
 446                 return 1;
 447         if (eb && btrfs_header_level(eb) == 0)
 448                 return 1;
 449
 450         write_lock(&fs_info->tree_mod_log_lock);
 451         if (list_empty(&(fs_info)->tree_mod_seq_list)) {
 452                 write_unlock(&fs_info->tree_mod_log_lock);
 453                 return 1;
 454         }
 455
 456         return 0;
 457 }
 458
 459 /* Similar to tree_mod_dont_log, but doesn't acquire any locks. */
 460 static inline int tree_mod_need_log(const struct btrfs_fs_info *fs_info,
 461                                     struct extent_buffer *eb)
 462 {
 463         smp_mb();
 464         if (list_empty(&(fs_info)->tree_mod_seq_list))
 465                 return 0;
 466         if (eb && btrfs_header_level(eb) == 0)
 467                 return 0;
 468
 469         return 1;
 470 }
 471
 472 static struct tree_mod_elem *
 473 alloc_tree_mod_elem(struct extent_buffer *eb, int slot,
 474                     enum mod_log_op op, gfp_t flags)
 475 {
 476         struct tree_mod_elem *tm;
 477
 478         tm = kzalloc(sizeof(*tm), flags);
 479         if (!tm)
 480                 return NULL;
 481
 482         tm->logical = eb->start;
 483         if (op != MOD_LOG_KEY_ADD) {
 484                 btrfs_node_key(eb, &tm->key, slot);
 485                 tm->blockptr = btrfs_node_blockptr(eb, slot);
 486         }
 487         tm->op = op;
 488         tm->slot = slot;
 489         tm->generation = btrfs_node_ptr_generation(eb, slot);
 490         RB_CLEAR_NODE(&tm->node);
 491
 492         return tm;
 493 }
 494
 495 static noinline int tree_mod_log_insert_key(struct extent_buffer *eb, int slot,
 496                 enum mod_log_op op, gfp_t flags)
 497 {
 498         struct tree_mod_elem *tm;
 499         int ret;
 500
 501         if (!tree_mod_need_log(eb->fs_info, eb))
 502                 return 0;
 503
 504         tm = alloc_tree_mod_elem(eb, slot, op, flags);
 505         if (!tm)
 506                 return -ENOMEM;
 507
 508         if (tree_mod_dont_log(eb->fs_info, eb)) {
 509                 kfree(tm);
 510                 return 0;
 511         }
 512
 513         ret = __tree_mod_log_insert(eb->fs_info, tm);
 514         write_unlock(&eb->fs_info->tree_mod_log_lock);
 515         if (ret)
 516                 kfree(tm);
 517
 518         return ret;
 519 }
 520
 521 static noinline int tree_mod_log_insert_move(struct extent_buffer *eb,
 522                 int dst_slot, int src_slot, int nr_items)
 523 {
 524         struct tree_mod_elem *tm = NULL;
 525         struct tree_mod_elem **tm_list = NULL;
 526         int ret = 0;
 527         int i;
 528         int locked = 0;
 529
 530         if (!tree_mod_need_log(eb->fs_info, eb))
 531                 return 0;
 532
 533         tm_list = kcalloc(nr_items, sizeof(struct tree_mod_elem *), GFP_NOFS);
 534         if (!tm_list)
 535                 return -ENOMEM;
 536
 537         tm = kzalloc(sizeof(*tm), GFP_NOFS);
 538         if (!tm) {
 539                 ret = -ENOMEM;
 540                 goto free_tms;
 541         }
 542
 543         tm->logical = eb->start;
 544         tm->slot = src_slot;
 545         tm->move.dst_slot = dst_slot;
 546         tm->move.nr_items = nr_items;
 547         tm->op = MOD_LOG_MOVE_KEYS;
 548
 549         for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
 550                 tm_list[i] = alloc_tree_mod_elem(eb, i + dst_slot,
 551                     MOD_LOG_KEY_REMOVE_WHILE_MOVING, GFP_NOFS);
 552                 if (!tm_list[i]) {
 553                         ret = -ENOMEM;
 554                         goto free_tms;
 555                 }
 556         }
 557
 558         if (tree_mod_dont_log(eb->fs_info, eb))
 559                 goto free_tms;
 560         locked = 1;
 561
 562         /*
 563          * When we override something during the move, we log these removals.
 564          * This can only happen when we move towards the beginning of the
 565          * buffer, i.e. dst_slot < src_slot.
 566          */
 567         for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
 568                 ret = __tree_mod_log_insert(eb->fs_info, tm_list[i]);
 569                 if (ret)
 570                         goto free_tms;
 571         }
 572
 573         ret = __tree_mod_log_insert(eb->fs_info, tm);
 574         if (ret)
 575                 goto free_tms;
 576         write_unlock(&eb->fs_info->tree_mod_log_lock);
 577         kfree(tm_list);
 578
 579         return 0;
 580 free_tms:
 581         for (i = 0; i < nr_items; i++) {
 582                 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
 583                         rb_erase(&tm_list[i]->node, &eb->fs_info->tree_mod_log);
 584                 kfree(tm_list[i]);
 585         }
 586         if (locked)
 587                 write_unlock(&eb->fs_info->tree_mod_log_lock);
 588         kfree(tm_list);
 589         kfree(tm);
 590
 591         return ret;
 592 }
 593
 594 static inline int
 595 __tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
 596                        struct tree_mod_elem **tm_list,
 597                        int nritems)
 598 {
 599         int i, j;
 600         int ret;
 601
 602         for (i = nritems - 1; i >= 0; i--) {
 603                 ret = __tree_mod_log_insert(fs_info, tm_list[i]);
 604                 if (ret) {
 605                         for (j = nritems - 1; j > i; j--)
 606                                 rb_erase(&tm_list[j]->node,
 607                                          &fs_info->tree_mod_log);
 608                         return ret;
 609                 }
 610         }
 611
 612         return 0;
 613 }
 614
 615 static noinline int tree_mod_log_insert_root(struct extent_buffer *old_root,
 616                          struct extent_buffer *new_root, int log_removal)
 617 {
 618         struct btrfs_fs_info *fs_info = old_root->fs_info;
 619         struct tree_mod_elem *tm = NULL;
 620         struct tree_mod_elem **tm_list = NULL;
 621         int nritems = 0;
 622         int ret = 0;
 623         int i;
 624
 625         if (!tree_mod_need_log(fs_info, NULL))
 626                 return 0;
 627
 628         if (log_removal && btrfs_header_level(old_root) > 0) {
 629                 nritems = btrfs_header_nritems(old_root);
 630                 tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *),
 631                                   GFP_NOFS);
 632                 if (!tm_list) {
 633                         ret = -ENOMEM;
 634                         goto free_tms;
 635                 }
 636                 for (i = 0; i < nritems; i++) {
 637                         tm_list[i] = alloc_tree_mod_elem(old_root, i,
 638                             MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
 639                         if (!tm_list[i]) {
 640                                 ret = -ENOMEM;
 641                                 goto free_tms;
 642                         }
 643                 }
 644         }
 645
 646         tm = kzalloc(sizeof(*tm), GFP_NOFS);
 647         if (!tm) {
 648                 ret = -ENOMEM;
 649                 goto free_tms;
 650         }
 651
 652         tm->logical = new_root->start;
 653         tm->old_root.logical = old_root->start;
 654         tm->old_root.level = btrfs_header_level(old_root);
 655         tm->generation = btrfs_header_generation(old_root);
 656         tm->op = MOD_LOG_ROOT_REPLACE;
 657
 658         if (tree_mod_dont_log(fs_info, NULL))
 659                 goto free_tms;
 660
 661         if (tm_list)
 662                 ret = __tree_mod_log_free_eb(fs_info, tm_list, nritems);
 663         if (!ret)
 664                 ret = __tree_mod_log_insert(fs_info, tm);
 665
 666         write_unlock(&fs_info->tree_mod_log_lock);
 667         if (ret)
 668                 goto free_tms;
 669         kfree(tm_list);
 670
 671         return ret;
 672
 673 free_tms:
 674         if (tm_list) {
 675                 for (i = 0; i < nritems; i++)
 676                         kfree(tm_list[i]);
 677                 kfree(tm_list);
 678         }
 679         kfree(tm);
 680
 681         return ret;
 682 }
 683
 684 static struct tree_mod_elem *
 685 __tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq,
 686                       int smallest)
 687 {
 688         struct rb_root *tm_root;
 689         struct rb_node *node;
 690         struct tree_mod_elem *cur = NULL;
 691         struct tree_mod_elem *found = NULL;
 692
 693         read_lock(&fs_info->tree_mod_log_lock);
 694         tm_root = &fs_info->tree_mod_log;
 695         node = tm_root->rb_node;
 696         while (node) {
 697                 cur = rb_entry(node, struct tree_mod_elem, node);
 698                 if (cur->logical < start) {
 699                         node = node->rb_left;
 700                 } else if (cur->logical > start) {
 701                         node = node->rb_right;
 702                 } else if (cur->seq < min_seq) {
 703                         node = node->rb_left;
 704                 } else if (!smallest) {
 705                         /* we want the node with the highest seq */
 706                         if (found)
 707                                 BUG_ON(found->seq > cur->seq);
 708                         found = cur;
 709                         node = node->rb_left;
 710                 } else if (cur->seq > min_seq) {
 711                         /* we want the node with the smallest seq */
 712                         if (found)
 713                                 BUG_ON(found->seq < cur->seq);
 714                         found = cur;
 715                         node = node->rb_right;
 716                 } else {
 717                         found = cur;
 718                         break;
 719                 }
 720         }
 721         read_unlock(&fs_info->tree_mod_log_lock);
 722
 723         return found;
 724 }
 725
 726 /*
 727  * this returns the element from the log with the smallest time sequence
 728  * value that's in the log (the oldest log item). any element with a time
 729  * sequence lower than min_seq will be ignored.
 730  */
 731 static struct tree_mod_elem *
 732 tree_mod_log_search_oldest(struct btrfs_fs_info *fs_info, u64 start,
 733                            u64 min_seq)
 734 {
 735         return __tree_mod_log_search(fs_info, start, min_seq, 1);
 736 }
 737
 738 /*
 739  * this returns the element from the log with the largest time sequence
 740  * value that's in the log (the most recent log item). any element with
 741  * a time sequence lower than min_seq will be ignored.
 742  */
 743 static struct tree_mod_elem *
 744 tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq)
 745 {
 746         return __tree_mod_log_search(fs_info, start, min_seq, 0);
 747 }
 748
 749 static noinline int tree_mod_log_eb_copy(struct extent_buffer *dst,
 750                      struct extent_buffer *src, unsigned long dst_offset,
 751                      unsigned long src_offset, int nr_items)
 752 {
 753         struct btrfs_fs_info *fs_info = dst->fs_info;
 754         int ret = 0;
 755         struct tree_mod_elem **tm_list = NULL;
 756         struct tree_mod_elem **tm_list_add, **tm_list_rem;
 757         int i;
 758         int locked = 0;
 759
 760         if (!tree_mod_need_log(fs_info, NULL))
 761                 return 0;
 762
 763         if (btrfs_header_level(dst) == 0 && btrfs_header_level(src) == 0)
 764                 return 0;
 765
 766         tm_list = kcalloc(nr_items * 2, sizeof(struct tree_mod_elem *),
 767                           GFP_NOFS);
 768         if (!tm_list)
 769                 return -ENOMEM;
 770
 771         tm_list_add = tm_list;
 772         tm_list_rem = tm_list + nr_items;
 773         for (i = 0; i < nr_items; i++) {
 774                 tm_list_rem[i] = alloc_tree_mod_elem(src, i + src_offset,
 775                     MOD_LOG_KEY_REMOVE, GFP_NOFS);
 776                 if (!tm_list_rem[i]) {
 777                         ret = -ENOMEM;
 778                         goto free_tms;
 779                 }
 780
 781                 tm_list_add[i] = alloc_tree_mod_elem(dst, i + dst_offset,
 782                     MOD_LOG_KEY_ADD, GFP_NOFS);
 783                 if (!tm_list_add[i]) {
 784                         ret = -ENOMEM;
 785                         goto free_tms;
 786                 }
 787         }
 788
 789         if (tree_mod_dont_log(fs_info, NULL))
 790                 goto free_tms;
 791         locked = 1;
 792
 793         for (i = 0; i < nr_items; i++) {
 794                 ret = __tree_mod_log_insert(fs_info, tm_list_rem[i]);
 795                 if (ret)
 796                         goto free_tms;
 797                 ret = __tree_mod_log_insert(fs_info, tm_list_add[i]);
 798                 if (ret)
 799                         goto free_tms;
 800         }
 801
 802         write_unlock(&fs_info->tree_mod_log_lock);
 803         kfree(tm_list);
 804
 805         return 0;
 806
 807 free_tms:
 808         for (i = 0; i < nr_items * 2; i++) {
 809                 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
 810                         rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
 811                 kfree(tm_list[i]);
 812         }
 813         if (locked)
 814                 write_unlock(&fs_info->tree_mod_log_lock);
 815         kfree(tm_list);
 816
 817         return ret;
 818 }
 819
 820 static noinline int tree_mod_log_free_eb(struct extent_buffer *eb)
 821 {
 822         struct tree_mod_elem **tm_list = NULL;
 823         int nritems = 0;
 824         int i;
 825         int ret = 0;
 826
 827         if (btrfs_header_level(eb) == 0)
 828                 return 0;
 829
 830         if (!tree_mod_need_log(eb->fs_info, NULL))
 831                 return 0;
 832
 833         nritems = btrfs_header_nritems(eb);
 834         tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *), GFP_NOFS);
 835         if (!tm_list)
 836                 return -ENOMEM;
 837
 838         for (i = 0; i < nritems; i++) {
 839                 tm_list[i] = alloc_tree_mod_elem(eb, i,
 840                     MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
 841                 if (!tm_list[i]) {
 842                         ret = -ENOMEM;
 843                         goto free_tms;
 844                 }
 845         }
 846
 847         if (tree_mod_dont_log(eb->fs_info, eb))
 848                 goto free_tms;
 849
 850         ret = __tree_mod_log_free_eb(eb->fs_info, tm_list, nritems);
 851         write_unlock(&eb->fs_info->tree_mod_log_lock);
 852         if (ret)
 853                 goto free_tms;
 854         kfree(tm_list);
 855
 856         return 0;
 857
 858 free_tms:
 859         for (i = 0; i < nritems; i++)
 860                 kfree(tm_list[i]);
 861         kfree(tm_list);
 862
 863         return ret;
 864 }
 865
 866 /*
 867  * check if the tree block can be shared by multiple trees
 868  */
 869 int btrfs_block_can_be_shared(struct btrfs_root *root,
 870                               struct extent_buffer *buf)
 871 {
 872         /*
 873          * Tree blocks not in reference counted trees and tree roots
 874          * are never shared. If a block was allocated after the last
 875          * snapshot and the block was not allocated by tree relocation,
 876          * we know the block is not shared.
 877          */
 878         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 879             buf != root->node && buf != root->commit_root &&
 880             (btrfs_header_generation(buf) <=
 881              btrfs_root_last_snapshot(&root->root_item) ||
 882              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 883                 return 1;
 884
 885         return 0;
 886 }
 887
 888 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
 889                                        struct btrfs_root *root,
 890                                        struct extent_buffer *buf,
 891                                        struct extent_buffer *cow,
 892                                        int *last_ref)
 893 {
 894         struct btrfs_fs_info *fs_info = root->fs_info;
 895         u64 refs;
 896         u64 owner;
 897         u64 flags;
 898         u64 new_flags = 0;
 899         int ret;
 900
 901         /*
 902          * Backrefs update rules:
 903          *
 904          * Always use full backrefs for extent pointers in tree block
 905          * allocated by tree relocation.
 906          *
 907          * If a shared tree block is no longer referenced by its owner
 908          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
 909          * use full backrefs for extent pointers in tree block.
 910          *
 911          * If a tree block is been relocating
 912          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
 913          * use full backrefs for extent pointers in tree block.
 914          * The reason for this is some operations (such as drop tree)
 915          * are only allowed for blocks use full backrefs.
 916          */
 917
 918         if (btrfs_block_can_be_shared(root, buf)) {
 919                 ret = btrfs_lookup_extent_info(trans, fs_info, buf->start,
 920                                                btrfs_header_level(buf), 1,
 921                                                &refs, &flags);
 922                 if (ret)
 923                         return ret;
 924                 if (refs == 0) {
 925                         ret = -EROFS;
 926                         btrfs_handle_fs_error(fs_info, ret, NULL);
 927                         return ret;
 928                 }
 929         } else {
 930                 refs = 1;
 931                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 932                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 933                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 934                 else
 935                         flags = 0;
 936         }
 937
 938         owner = btrfs_header_owner(buf);
 939         BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
 940                !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
 941
 942         if (refs > 1) {
 943                 if ((owner == root->root_key.objectid ||
 944                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
 945                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
 946                         ret = btrfs_inc_ref(trans, root, buf, 1);
 947                         if (ret)
 948                                 return ret;
 949
 950                         if (root->root_key.objectid ==
 951                             BTRFS_TREE_RELOC_OBJECTID) {
 952                                 ret = btrfs_dec_ref(trans, root, buf, 0);
 953                                 if (ret)
 954                                         return ret;
 955                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 956                                 if (ret)
 957                                         return ret;
 958                         }
 959                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
 960                 } else {
 961
 962                         if (root->root_key.objectid ==
 963                             BTRFS_TREE_RELOC_OBJECTID)
 964                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 965                         else
 966                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 967                         if (ret)
 968                                 return ret;
 969                 }
 970                 if (new_flags != 0) {
 971                         int level = btrfs_header_level(buf);
 972
 973                         ret = btrfs_set_disk_extent_flags(trans,
 974                                                           buf->start,
 975                                                           buf->len,
 976                                                           new_flags, level, 0);
 977                         if (ret)
 978                                 return ret;
 979                 }
 980         } else {
 981                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
 982                         if (root->root_key.objectid ==
 983                             BTRFS_TREE_RELOC_OBJECTID)
 984                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 985                         else
 986                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 987                         if (ret)
 988                                 return ret;
 989                         ret = btrfs_dec_ref(trans, root, buf, 1);
 990                         if (ret)
 991                                 return ret;
 992                 }
 993                 btrfs_clean_tree_block(buf);
 994                 *last_ref = 1;
 995         }
 996         return 0;
 997 }
 998
 999 static struct extent_buffer *alloc_tree_block_no_bg_flush(
1000                                           struct btrfs_trans_handle *trans,
1001                                           struct btrfs_root *root,
1002                                           u64 parent_start,
1003                                           const struct btrfs_disk_key *disk_key,
1004                                           int level,
1005                                           u64 hint,
1006                                           u64 empty_size)
1007 {
1008         struct btrfs_fs_info *fs_info = root->fs_info;
1009         struct extent_buffer *ret;
1010
1011         /*
1012          * If we are COWing a node/leaf from the extent, chunk, device or free
1013          * space trees, make sure that we do not finish block group creation of
1014          * pending block groups. We do this to avoid a deadlock.
1015          * COWing can result in allocation of a new chunk, and flushing pending
1016          * block groups (btrfs_create_pending_block_groups()) can be triggered
1017          * when finishing allocation of a new chunk. Creation of a pending block
1018          * group modifies the extent, chunk, device and free space trees,
1019          * therefore we could deadlock with ourselves since we are holding a
1020          * lock on an extent buffer that btrfs_create_pending_block_groups() may
1021          * try to COW later.
1022          * For similar reasons, we also need to delay flushing pending block
1023          * groups when splitting a leaf or node, from one of those trees, since
1024          * we are holding a write lock on it and its parent or when inserting a
1025          * new root node for one of those trees.
1026          */
1027         if (root == fs_info->extent_root ||
1028             root == fs_info->chunk_root ||
1029             root == fs_info->dev_root ||
1030             root == fs_info->free_space_root)
1031                 trans->can_flush_pending_bgs = false;
1032
1033         ret = btrfs_alloc_tree_block(trans, root, parent_start,
1034                                      root->root_key.objectid, disk_key, level,
1035                                      hint, empty_size);
1036         trans->can_flush_pending_bgs = true;
1037
1038         return ret;
1039 }
1040
1041 /*
1042  * does the dirty work in cow of a single block.  The parent block (if
1043  * supplied) is updated to point to the new cow copy.  The new buffer is marked
1044  * dirty and returned locked.  If you modify the block it needs to be marked
1045  * dirty again.
1046  *
1047  * search_start -- an allocation hint for the new block
1048  *
1049  * empty_size -- a hint that you plan on doing more cow.  This is the size in
1050  * bytes the allocator should try to find free next to the block it returns.
1051  * This is just a hint and may be ignored by the allocator.
1052  */
1053 static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
1054                              struct btrfs_root *root,
1055                              struct extent_buffer *buf,
1056                              struct extent_buffer *parent, int parent_slot,
1057                              struct extent_buffer **cow_ret,
1058                              u64 search_start, u64 empty_size)
1059 {
1060         struct btrfs_fs_info *fs_info = root->fs_info;
1061         struct btrfs_disk_key disk_key;
1062         struct extent_buffer *cow;
1063         int level, ret;
1064         int last_ref = 0;
1065         int unlock_orig = 0;
1066         u64 parent_start = 0;
1067
1068         if (*cow_ret == buf)
1069                 unlock_orig = 1;
1070
1071         btrfs_assert_tree_locked(buf);
1072
1073         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1074                 trans->transid != fs_info->running_transaction->transid);
1075         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1076                 trans->transid != root->last_trans);
1077
1078         level = btrfs_header_level(buf);
1079
1080         if (level == 0)
1081                 btrfs_item_key(buf, &disk_key, 0);
1082         else
1083                 btrfs_node_key(buf, &disk_key, 0);
1084
1085         if ((root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) && parent)
1086                 parent_start = parent->start;
1087
1088         cow = alloc_tree_block_no_bg_flush(trans, root, parent_start, &disk_key,
1089                                            level, search_start, empty_size);
1090         if (IS_ERR(cow))
1091                 return PTR_ERR(cow);
1092
1093         /* cow is set to blocking by btrfs_init_new_buffer */
1094
1095         copy_extent_buffer_full(cow, buf);
1096         btrfs_set_header_bytenr(cow, cow->start);
1097         btrfs_set_header_generation(cow, trans->transid);
1098         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
1099         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
1100                                      BTRFS_HEADER_FLAG_RELOC);
1101         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1102                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
1103         else
1104                 btrfs_set_header_owner(cow, root->root_key.objectid);
1105
1106         write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
1107
1108         ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
1109         if (ret) {
1110                 btrfs_abort_transaction(trans, ret);
1111                 return ret;
1112         }
1113
1114         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
1115                 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
1116                 if (ret) {
1117                         btrfs_abort_transaction(trans, ret);
1118                         return ret;
1119                 }
1120         }
1121
1122         if (buf == root->node) {
1123                 WARN_ON(parent && parent != buf);
1124                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1125                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1126                         parent_start = buf->start;
1127
1128                 extent_buffer_get(cow);
1129                 ret = tree_mod_log_insert_root(root->node, cow, 1);
1130                 BUG_ON(ret < 0);
1131                 rcu_assign_pointer(root->node, cow);
1132
1133                 btrfs_free_tree_block(trans, root, buf, parent_start,
1134                                       last_ref);
1135                 free_extent_buffer(buf);
1136                 add_root_to_dirty_list(root);
1137         } else {
1138                 WARN_ON(trans->transid != btrfs_header_generation(parent));
1139                 tree_mod_log_insert_key(parent, parent_slot,
1140                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
1141                 btrfs_set_node_blockptr(parent, parent_slot,
1142                                         cow->start);
1143                 btrfs_set_node_ptr_generation(parent, parent_slot,
1144                                               trans->transid);
1145                 btrfs_mark_buffer_dirty(parent);
1146                 if (last_ref) {
1147                         ret = tree_mod_log_free_eb(buf);
1148                         if (ret) {
1149                                 btrfs_abort_transaction(trans, ret);
1150                                 return ret;
1151                         }
1152                 }
1153                 btrfs_free_tree_block(trans, root, buf, parent_start,
1154                                       last_ref);
1155         }
1156         if (unlock_orig)
1157                 btrfs_tree_unlock(buf);
1158         free_extent_buffer_stale(buf);
1159         btrfs_mark_buffer_dirty(cow);
1160         *cow_ret = cow;
1161         return 0;
1162 }
1163
1164 /*
1165  * returns the logical address of the oldest predecessor of the given root.
1166  * entries older than time_seq are ignored.
1167  */
1168 static struct tree_mod_elem *__tree_mod_log_oldest_root(
1169                 struct extent_buffer *eb_root, u64 time_seq)
1170 {
1171         struct tree_mod_elem *tm;
1172         struct tree_mod_elem *found = NULL;
1173         u64 root_logical = eb_root->start;
1174         int looped = 0;
1175
1176         if (!time_seq)
1177                 return NULL;
1178
1179         /*
1180          * the very last operation that's logged for a root is the
1181          * replacement operation (if it is replaced at all). this has
1182          * the logical address of the *new* root, making it the very
1183          * first operation that's logged for this root.
1184          */
1185         while (1) {
1186                 tm = tree_mod_log_search_oldest(eb_root->fs_info, root_logical,
1187                                                 time_seq);
1188                 if (!looped && !tm)
1189                         return NULL;
1190                 /*
1191                  * if there are no tree operation for the oldest root, we simply
1192                  * return it. this should only happen if that (old) root is at
1193                  * level 0.
1194                  */
1195                 if (!tm)
1196                         break;
1197
1198                 /*
1199                  * if there's an operation that's not a root replacement, we
1200                  * found the oldest version of our root. normally, we'll find a
1201                  * MOD_LOG_KEY_REMOVE_WHILE_FREEING operation here.
1202                  */
1203                 if (tm->op != MOD_LOG_ROOT_REPLACE)
1204                         break;
1205
1206                 found = tm;
1207                 root_logical = tm->old_root.logical;
1208                 looped = 1;
1209         }
1210
1211         /* if there's no old root to return, return what we found instead */
1212         if (!found)
1213                 found = tm;
1214
1215         return found;
1216 }
1217
1218 /*
1219  * tm is a pointer to the first operation to rewind within eb. then, all
1220  * previous operations will be rewound (until we reach something older than
1221  * time_seq).
1222  */
1223 static void
1224 __tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct extent_buffer *eb,
1225                       u64 time_seq, struct tree_mod_elem *first_tm)
1226 {
1227         u32 n;
1228         struct rb_node *next;
1229         struct tree_mod_elem *tm = first_tm;
1230         unsigned long o_dst;
1231         unsigned long o_src;
1232         unsigned long p_size = sizeof(struct btrfs_key_ptr);
1233
1234         n = btrfs_header_nritems(eb);
1235         read_lock(&fs_info->tree_mod_log_lock);
1236         while (tm && tm->seq >= time_seq) {
1237                 /*
1238                  * all the operations are recorded with the operator used for
1239                  * the modification. as we're going backwards, we do the
1240                  * opposite of each operation here.
1241                  */
1242                 switch (tm->op) {
1243                 case MOD_LOG_KEY_REMOVE_WHILE_FREEING:
1244                         BUG_ON(tm->slot < n);
1245                         /* Fallthrough */
1246                 case MOD_LOG_KEY_REMOVE_WHILE_MOVING:
1247                 case MOD_LOG_KEY_REMOVE:
1248                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1249                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1250                         btrfs_set_node_ptr_generation(eb, tm->slot,
1251                                                       tm->generation);
1252                         n++;
1253                         break;
1254                 case MOD_LOG_KEY_REPLACE:
1255                         BUG_ON(tm->slot >= n);
1256                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1257                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1258                         btrfs_set_node_ptr_generation(eb, tm->slot,
1259                                                       tm->generation);
1260                         break;
1261                 case MOD_LOG_KEY_ADD:
1262                         /* if a move operation is needed it's in the log */
1263                         n--;
1264                         break;
1265                 case MOD_LOG_MOVE_KEYS:
1266                         o_dst = btrfs_node_key_ptr_offset(tm->slot);
1267                         o_src = btrfs_node_key_ptr_offset(tm->move.dst_slot);
1268                         memmove_extent_buffer(eb, o_dst, o_src,
1269                                               tm->move.nr_items * p_size);
1270                         break;
1271                 case MOD_LOG_ROOT_REPLACE:
1272                         /*
1273                          * this operation is special. for roots, this must be
1274                          * handled explicitly before rewinding.
1275                          * for non-roots, this operation may exist if the node
1276                          * was a root: root A -> child B; then A gets empty and
1277                          * B is promoted to the new root. in the mod log, we'll
1278                          * have a root-replace operation for B, a tree block
1279                          * that is no root. we simply ignore that operation.
1280                          */
1281                         break;
1282                 }
1283                 next = rb_next(&tm->node);
1284                 if (!next)
1285                         break;
1286                 tm = rb_entry(next, struct tree_mod_elem, node);
1287                 if (tm->logical != first_tm->logical)
1288                         break;
1289         }
1290         read_unlock(&fs_info->tree_mod_log_lock);
1291         btrfs_set_header_nritems(eb, n);
1292 }
1293
1294 /*
1295  * Called with eb read locked. If the buffer cannot be rewound, the same buffer
1296  * is returned. If rewind operations happen, a fresh buffer is returned. The
1297  * returned buffer is always read-locked. If the returned buffer is not the
1298  * input buffer, the lock on the input buffer is released and the input buffer
1299  * is freed (its refcount is decremented).
1300  */
1301 static struct extent_buffer *
1302 tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
1303                     struct extent_buffer *eb, u64 time_seq)
1304 {
1305         struct extent_buffer *eb_rewin;
1306         struct tree_mod_elem *tm;
1307
1308         if (!time_seq)
1309                 return eb;
1310
1311         if (btrfs_header_level(eb) == 0)
1312                 return eb;
1313
1314         tm = tree_mod_log_search(fs_info, eb->start, time_seq);
1315         if (!tm)
1316                 return eb;
1317
1318         btrfs_set_path_blocking(path);
1319         btrfs_set_lock_blocking_read(eb);
1320
1321         if (tm->op == MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1322                 BUG_ON(tm->slot != 0);
1323                 eb_rewin = alloc_dummy_extent_buffer(fs_info, eb->start);
1324                 if (!eb_rewin) {
1325                         btrfs_tree_read_unlock_blocking(eb);
1326                         free_extent_buffer(eb);
1327                         return NULL;
1328                 }
1329                 btrfs_set_header_bytenr(eb_rewin, eb->start);
1330                 btrfs_set_header_backref_rev(eb_rewin,
1331                                              btrfs_header_backref_rev(eb));
1332                 btrfs_set_header_owner(eb_rewin, btrfs_header_owner(eb));
1333                 btrfs_set_header_level(eb_rewin, btrfs_header_level(eb));
1334         } else {
1335                 eb_rewin = btrfs_clone_extent_buffer(eb);
1336                 if (!eb_rewin) {
1337                         btrfs_tree_read_unlock_blocking(eb);
1338                         free_extent_buffer(eb);
1339                         return NULL;
1340                 }
1341         }
1342
1343         btrfs_tree_read_unlock_blocking(eb);
1344         free_extent_buffer(eb);
1345
1346         btrfs_tree_read_lock(eb_rewin);
1347         __tree_mod_log_rewind(fs_info, eb_rewin, time_seq, tm);
1348         WARN_ON(btrfs_header_nritems(eb_rewin) >
1349                 BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1350
1351         return eb_rewin;
1352 }
1353
1354 /*
1355  * get_old_root() rewinds the state of @root's root node to the given @time_seq
1356  * value. If there are no changes, the current root->root_node is returned. If
1357  * anything changed in between, there's a fresh buffer allocated on which the
1358  * rewind operations are done. In any case, the returned buffer is read locked.
1359  * Returns NULL on error (with no locks held).
1360  */
1361 static inline struct extent_buffer *
1362 get_old_root(struct btrfs_root *root, u64 time_seq)
1363 {
1364         struct btrfs_fs_info *fs_info = root->fs_info;
1365         struct tree_mod_elem *tm;
1366         struct extent_buffer *eb = NULL;
1367         struct extent_buffer *eb_root;
1368         u64 eb_root_owner = 0;
1369         struct extent_buffer *old;
1370         struct tree_mod_root *old_root = NULL;
1371         u64 old_generation = 0;
1372         u64 logical;
1373         int level;
1374
1375         eb_root = btrfs_read_lock_root_node(root);
1376         tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1377         if (!tm)
1378                 return eb_root;
1379
1380         if (tm->op == MOD_LOG_ROOT_REPLACE) {
1381                 old_root = &tm->old_root;
1382                 old_generation = tm->generation;
1383                 logical = old_root->logical;
1384                 level = old_root->level;
1385         } else {
1386                 logical = eb_root->start;
1387                 level = btrfs_header_level(eb_root);
1388         }
1389
1390         tm = tree_mod_log_search(fs_info, logical, time_seq);
1391         if (old_root && tm && tm->op != MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1392                 btrfs_tree_read_unlock(eb_root);
1393                 free_extent_buffer(eb_root);
1394                 old = read_tree_block(fs_info, logical, 0, level, NULL);
1395                 if (WARN_ON(IS_ERR(old) || !extent_buffer_uptodate(old))) {
1396                         if (!IS_ERR(old))
1397                                 free_extent_buffer(old);
1398                         btrfs_warn(fs_info,
1399                                    "failed to read tree block %llu from get_old_root",
1400                                    logical);
1401                 } else {
1402                         eb = btrfs_clone_extent_buffer(old);
1403                         free_extent_buffer(old);
1404                 }
1405         } else if (old_root) {
1406                 eb_root_owner = btrfs_header_owner(eb_root);
1407                 btrfs_tree_read_unlock(eb_root);
1408                 free_extent_buffer(eb_root);
1409                 eb = alloc_dummy_extent_buffer(fs_info, logical);
1410         } else {
1411                 btrfs_set_lock_blocking_read(eb_root);
1412                 eb = btrfs_clone_extent_buffer(eb_root);
1413                 btrfs_tree_read_unlock_blocking(eb_root);
1414                 free_extent_buffer(eb_root);
1415         }
1416
1417         if (!eb)
1418                 return NULL;
1419         btrfs_tree_read_lock(eb);
1420         if (old_root) {
1421                 btrfs_set_header_bytenr(eb, eb->start);
1422                 btrfs_set_header_backref_rev(eb, BTRFS_MIXED_BACKREF_REV);
1423                 btrfs_set_header_owner(eb, eb_root_owner);
1424                 btrfs_set_header_level(eb, old_root->level);
1425                 btrfs_set_header_generation(eb, old_generation);
1426         }
1427         if (tm)
1428                 __tree_mod_log_rewind(fs_info, eb, time_seq, tm);
1429         else
1430                 WARN_ON(btrfs_header_level(eb) != 0);
1431         WARN_ON(btrfs_header_nritems(eb) > BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1432
1433         return eb;
1434 }
1435
1436 int btrfs_old_root_level(struct btrfs_root *root, u64 time_seq)
1437 {
1438         struct tree_mod_elem *tm;
1439         int level;
1440         struct extent_buffer *eb_root = btrfs_root_node(root);
1441
1442         tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1443         if (tm && tm->op == MOD_LOG_ROOT_REPLACE) {
1444                 level = tm->old_root.level;
1445         } else {
1446                 level = btrfs_header_level(eb_root);
1447         }
1448         free_extent_buffer(eb_root);
1449
1450         return level;
1451 }
1452
1453 static inline int should_cow_block(struct btrfs_trans_handle *trans,
1454                                    struct btrfs_root *root,
1455                                    struct extent_buffer *buf)
1456 {
1457         if (btrfs_is_testing(root->fs_info))
1458                 return 0;
1459
1460         /* Ensure we can see the FORCE_COW bit */
1461         smp_mb__before_atomic();
1462
1463         /*
1464          * We do not need to cow a block if
1465          * 1) this block is not created or changed in this transaction;
1466          * 2) this block does not belong to TREE_RELOC tree;
1467          * 3) the root is not forced COW.
1468          *
1469          * What is forced COW:
1470          *    when we create snapshot during committing the transaction,
1471          *    after we've finished copying src root, we must COW the shared
1472          *    block to ensure the metadata consistency.
1473          */
1474         if (btrfs_header_generation(buf) == trans->transid &&
1475             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
1476             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
1477               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
1478             !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
1479                 return 0;
1480         return 1;
1481 }
1482
1483 /*
1484  * cows a single block, see __btrfs_cow_block for the real work.
1485  * This version of it has extra checks so that a block isn't COWed more than
1486  * once per transaction, as long as it hasn't been written yet
1487  */
1488 noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
1489                     struct btrfs_root *root, struct extent_buffer *buf,
1490                     struct extent_buffer *parent, int parent_slot,
1491                     struct extent_buffer **cow_ret)
1492 {
1493         struct btrfs_fs_info *fs_info = root->fs_info;
1494         u64 search_start;
1495         int ret;
1496
1497         if (test_bit(BTRFS_ROOT_DELETING, &root->state))
1498                 btrfs_err(fs_info,
1499                         "COW'ing blocks on a fs root that's being dropped");
1500
1501         if (trans->transaction != fs_info->running_transaction)
1502                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1503                        trans->transid,
1504                        fs_info->running_transaction->transid);
1505
1506         if (trans->transid != fs_info->generation)
1507                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1508                        trans->transid, fs_info->generation);
1509
1510         if (!should_cow_block(trans, root, buf)) {
1511                 trans->dirty = true;
1512                 *cow_ret = buf;
1513                 return 0;
1514         }
1515
1516         search_start = buf->start & ~((u64)SZ_1G - 1);
1517
1518         if (parent)
1519                 btrfs_set_lock_blocking_write(parent);
1520         btrfs_set_lock_blocking_write(buf);
1521
1522         /*
1523          * Before CoWing this block for later modification, check if it's
1524          * the subtree root and do the delayed subtree trace if needed.
1525          *
1526          * Also We don't care about the error, as it's handled internally.
1527          */
1528         btrfs_qgroup_trace_subtree_after_cow(trans, root, buf);
1529         ret = __btrfs_cow_block(trans, root, buf, parent,
1530                                  parent_slot, cow_ret, search_start, 0);
1531
1532         trace_btrfs_cow_block(root, buf, *cow_ret);
1533
1534         return ret;
1535 }
1536
1537 /*
1538  * helper function for defrag to decide if two blocks pointed to by a
1539  * node are actually close by
1540  */
1541 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
1542 {
1543         if (blocknr < other && other - (blocknr + blocksize) < 32768)
1544                 return 1;
1545         if (blocknr > other && blocknr - (other + blocksize) < 32768)
1546                 return 1;
1547         return 0;
1548 }
1549
1550 /*
1551  * compare two keys in a memcmp fashion
1552  */
1553 static int comp_keys(const struct btrfs_disk_key *disk,
1554                      const struct btrfs_key *k2)
1555 {
1556         struct btrfs_key k1;
1557
1558         btrfs_disk_key_to_cpu(&k1, disk);
1559
1560         return btrfs_comp_cpu_keys(&k1, k2);
1561 }
1562
1563 /*
1564  * same as comp_keys only with two btrfs_key's
1565  */
1566 int btrfs_comp_cpu_keys(const struct btrfs_key *k1, const struct btrfs_key *k2)
1567 {
1568         if (k1->objectid > k2->objectid)
1569                 return 1;
1570         if (k1->objectid < k2->objectid)
1571                 return -1;
1572         if (k1->type > k2->type)
1573                 return 1;
1574         if (k1->type < k2->type)
1575                 return -1;
1576         if (k1->offset > k2->offset)
1577                 return 1;
1578         if (k1->offset < k2->offset)
1579                 return -1;
1580         return 0;
1581 }
1582
1583 /*
1584  * this is used by the defrag code to go through all the
1585  * leaves pointed to by a node and reallocate them so that
1586  * disk order is close to key order
1587  */
1588 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
1589                        struct btrfs_root *root, struct extent_buffer *parent,
1590                        int start_slot, u64 *last_ret,
1591                        struct btrfs_key *progress)
1592 {
1593         struct btrfs_fs_info *fs_info = root->fs_info;
1594         struct extent_buffer *cur;
1595         u64 blocknr;
1596         u64 gen;
1597         u64 search_start = *last_ret;
1598         u64 last_block = 0;
1599         u64 other;
1600         u32 parent_nritems;
1601         int end_slot;
1602         int i;
1603         int err = 0;
1604         int parent_level;
1605         int uptodate;
1606         u32 blocksize;
1607         int progress_passed = 0;
1608         struct btrfs_disk_key disk_key;
1609
1610         parent_level = btrfs_header_level(parent);
1611
1612         WARN_ON(trans->transaction != fs_info->running_transaction);
1613         WARN_ON(trans->transid != fs_info->generation);
1614
1615         parent_nritems = btrfs_header_nritems(parent);
1616         blocksize = fs_info->nodesize;
1617         end_slot = parent_nritems - 1;
1618
1619         if (parent_nritems <= 1)
1620                 return 0;
1621
1622         btrfs_set_lock_blocking_write(parent);
1623
1624         for (i = start_slot; i <= end_slot; i++) {
1625                 struct btrfs_key first_key;
1626                 int close = 1;
1627
1628                 btrfs_node_key(parent, &disk_key, i);
1629                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
1630                         continue;
1631
1632                 progress_passed = 1;
1633                 blocknr = btrfs_node_blockptr(parent, i);
1634                 gen = btrfs_node_ptr_generation(parent, i);
1635                 btrfs_node_key_to_cpu(parent, &first_key, i);
1636                 if (last_block == 0)
1637                         last_block = blocknr;
1638
1639                 if (i > 0) {
1640                         other = btrfs_node_blockptr(parent, i - 1);
1641                         close = close_blocks(blocknr, other, blocksize);
1642                 }
1643                 if (!close && i < end_slot) {
1644                         other = btrfs_node_blockptr(parent, i + 1);
1645                         close = close_blocks(blocknr, other, blocksize);
1646                 }
1647                 if (close) {
1648                         last_block = blocknr;
1649                         continue;
1650                 }
1651
1652                 cur = find_extent_buffer(fs_info, blocknr);
1653                 if (cur)
1654                         uptodate = btrfs_buffer_uptodate(cur, gen, 0);
1655                 else
1656                         uptodate = 0;
1657                 if (!cur || !uptodate) {
1658                         if (!cur) {
1659                                 cur = read_tree_block(fs_info, blocknr, gen,
1660                                                       parent_level - 1,
1661                                                       &first_key);
1662                                 if (IS_ERR(cur)) {
1663                                         return PTR_ERR(cur);
1664                                 } else if (!extent_buffer_uptodate(cur)) {
1665                                         free_extent_buffer(cur);
1666                                         return -EIO;
1667                                 }
1668                         } else if (!uptodate) {
1669                                 err = btrfs_read_buffer(cur, gen,
1670                                                 parent_level - 1,&first_key);
1671                                 if (err) {
1672                                         free_extent_buffer(cur);
1673                                         return err;
1674                                 }
1675                         }
1676                 }
1677                 if (search_start == 0)
1678                         search_start = last_block;
1679
1680                 btrfs_tree_lock(cur);
1681                 btrfs_set_lock_blocking_write(cur);
1682                 err = __btrfs_cow_block(trans, root, cur, parent, i,
1683                                         &cur, search_start,
1684                                         min(16 * blocksize,
1685                                             (end_slot - i) * blocksize));
1686                 if (err) {
1687                         btrfs_tree_unlock(cur);
1688                         free_extent_buffer(cur);
1689                         break;
1690                 }
1691                 search_start = cur->start;
1692                 last_block = cur->start;
1693                 *last_ret = search_start;
1694                 btrfs_tree_unlock(cur);
1695                 free_extent_buffer(cur);
1696         }
1697         return err;
1698 }
1699
1700 /*
1701  * search for key in the extent_buffer.  The items start at offset p,
1702  * and they are item_size apart.  There are 'max' items in p.
1703  *
1704  * the slot in the array is returned via slot, and it points to
1705  * the place where you would insert key if it is not found in
1706  * the array.
1707  *
1708  * slot may point to max if the key is bigger than all of the keys
1709  */
1710 static noinline int generic_bin_search(struct extent_buffer *eb,
1711                                        unsigned long p, int item_size,
1712                                        const struct btrfs_key *key,
1713                                        int max, int *slot)
1714 {
1715         int low = 0;
1716         int high = max;
1717         int mid;
1718         int ret;
1719         struct btrfs_disk_key *tmp = NULL;
1720         struct btrfs_disk_key unaligned;
1721         unsigned long offset;
1722         char *kaddr = NULL;
1723         unsigned long map_start = 0;
1724         unsigned long map_len = 0;
1725         int err;
1726
1727         if (low > high) {
1728                 btrfs_err(eb->fs_info,
1729                  "%s: low (%d) > high (%d) eb %llu owner %llu level %d",
1730                           __func__, low, high, eb->start,
1731                           btrfs_header_owner(eb), btrfs_header_level(eb));
1732                 return -EINVAL;
1733         }
1734
1735         while (low < high) {
1736                 mid = (low + high) / 2;
1737                 offset = p + mid * item_size;
1738
1739                 if (!kaddr || offset < map_start ||
1740                     (offset + sizeof(struct btrfs_disk_key)) >
1741                     map_start + map_len) {
1742
1743                         err = map_private_extent_buffer(eb, offset,
1744                                                 sizeof(struct btrfs_disk_key),
1745                                                 &kaddr, &map_start, &map_len);
1746
1747                         if (!err) {
1748                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1749                                                         map_start);
1750                         } else if (err == 1) {
1751                                 read_extent_buffer(eb, &unaligned,
1752                                                    offset, sizeof(unaligned));
1753                                 tmp = &unaligned;
1754                         } else {
1755                                 return err;
1756                         }
1757
1758                 } else {
1759                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
1760                                                         map_start);
1761                 }
1762                 ret = comp_keys(tmp, key);
1763
1764                 if (ret < 0)
1765                         low = mid + 1;
1766                 else if (ret > 0)
1767                         high = mid;
1768                 else {
1769                         *slot = mid;
1770                         return 0;
1771                 }
1772         }
1773         *slot = low;
1774         return 1;
1775 }
1776
1777 /*
1778  * simple bin_search frontend that does the right thing for
1779  * leaves vs nodes
1780  */
1781 int btrfs_bin_search(struct extent_buffer *eb, const struct btrfs_key *key,
1782                      int level, int *slot)
1783 {
1784         if (level == 0)
1785                 return generic_bin_search(eb,
1786                                           offsetof(struct btrfs_leaf, items),
1787                                           sizeof(struct btrfs_item),
1788                                           key, btrfs_header_nritems(eb),
1789                                           slot);
1790         else
1791                 return generic_bin_search(eb,
1792                                           offsetof(struct btrfs_node, ptrs),
1793                                           sizeof(struct btrfs_key_ptr),
1794                                           key, btrfs_header_nritems(eb),
1795                                           slot);
1796 }
1797
1798 static void root_add_used(struct btrfs_root *root, u32 size)
1799 {
1800         spin_lock(&root->accounting_lock);
1801         btrfs_set_root_used(&root->root_item,
1802                             btrfs_root_used(&root->root_item) + size);
1803         spin_unlock(&root->accounting_lock);
1804 }
1805
1806 static void root_sub_used(struct btrfs_root *root, u32 size)
1807 {
1808         spin_lock(&root->accounting_lock);
1809         btrfs_set_root_used(&root->root_item,
1810                             btrfs_root_used(&root->root_item) - size);
1811         spin_unlock(&root->accounting_lock);
1812 }
1813
1814 /* given a node and slot number, this reads the blocks it points to.  The
1815  * extent buffer is returned with a reference taken (but unlocked).
1816  */
1817 struct extent_buffer *btrfs_read_node_slot(struct extent_buffer *parent,
1818                                            int slot)
1819 {
1820         int level = btrfs_header_level(parent);
1821         struct extent_buffer *eb;
1822         struct btrfs_key first_key;
1823
1824         if (slot < 0 || slot >= btrfs_header_nritems(parent))
1825                 return ERR_PTR(-ENOENT);
1826
1827         BUG_ON(level == 0);
1828
1829         btrfs_node_key_to_cpu(parent, &first_key, slot);
1830         eb = read_tree_block(parent->fs_info, btrfs_node_blockptr(parent, slot),
1831                              btrfs_node_ptr_generation(parent, slot),
1832                              level - 1, &first_key);
1833         if (!IS_ERR(eb) && !extent_buffer_uptodate(eb)) {
1834                 free_extent_buffer(eb);
1835                 eb = ERR_PTR(-EIO);
1836         }
1837
1838         return eb;
1839 }
1840
1841 /*
1842  * node level balancing, used to make sure nodes are in proper order for
1843  * item deletion.  We balance from the top down, so we have to make sure
1844  * that a deletion won't leave an node completely empty later on.
1845  */
1846 static noinline int balance_level(struct btrfs_trans_handle *trans,
1847                          struct btrfs_root *root,
1848                          struct btrfs_path *path, int level)
1849 {
1850         struct btrfs_fs_info *fs_info = root->fs_info;
1851         struct extent_buffer *right = NULL;
1852         struct extent_buffer *mid;
1853         struct extent_buffer *left = NULL;
1854         struct extent_buffer *parent = NULL;
1855         int ret = 0;
1856         int wret;
1857         int pslot;
1858         int orig_slot = path->slots[level];
1859         u64 orig_ptr;
1860
1861         ASSERT(level > 0);
1862
1863         mid = path->nodes[level];
1864
1865         WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK &&
1866                 path->locks[level] != BTRFS_WRITE_LOCK_BLOCKING);
1867         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1868
1869         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1870
1871         if (level < BTRFS_MAX_LEVEL - 1) {
1872                 parent = path->nodes[level + 1];
1873                 pslot = path->slots[level + 1];
1874         }
1875
1876         /*
1877          * deal with the case where there is only one pointer in the root
1878          * by promoting the node below to a root
1879          */
1880         if (!parent) {
1881                 struct extent_buffer *child;
1882
1883                 if (btrfs_header_nritems(mid) != 1)
1884                         return 0;
1885
1886                 /* promote the child to a root */
1887                 child = btrfs_read_node_slot(mid, 0);
1888                 if (IS_ERR(child)) {
1889                         ret = PTR_ERR(child);
1890                         btrfs_handle_fs_error(fs_info, ret, NULL);
1891                         goto enospc;
1892                 }
1893
1894                 btrfs_tree_lock(child);
1895                 btrfs_set_lock_blocking_write(child);
1896                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
1897                 if (ret) {
1898                         btrfs_tree_unlock(child);
1899                         free_extent_buffer(child);
1900                         goto enospc;
1901                 }
1902
1903                 ret = tree_mod_log_insert_root(root->node, child, 1);
1904                 BUG_ON(ret < 0);
1905                 rcu_assign_pointer(root->node, child);
1906
1907                 add_root_to_dirty_list(root);
1908                 btrfs_tree_unlock(child);
1909
1910                 path->locks[level] = 0;
1911                 path->nodes[level] = NULL;
1912                 btrfs_clean_tree_block(mid);
1913                 btrfs_tree_unlock(mid);
1914                 /* once for the path */
1915                 free_extent_buffer(mid);
1916
1917                 root_sub_used(root, mid->len);
1918                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1919                 /* once for the root ptr */
1920                 free_extent_buffer_stale(mid);
1921                 return 0;
1922         }
1923         if (btrfs_header_nritems(mid) >
1924             BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 4)
1925                 return 0;
1926
1927         left = btrfs_read_node_slot(parent, pslot - 1);
1928         if (IS_ERR(left))
1929                 left = NULL;
1930
1931         if (left) {
1932                 btrfs_tree_lock(left);
1933                 btrfs_set_lock_blocking_write(left);
1934                 wret = btrfs_cow_block(trans, root, left,
1935                                        parent, pslot - 1, &left);
1936                 if (wret) {
1937                         ret = wret;
1938                         goto enospc;
1939                 }
1940         }
1941
1942         right = btrfs_read_node_slot(parent, pslot + 1);
1943         if (IS_ERR(right))
1944                 right = NULL;
1945
1946         if (right) {
1947                 btrfs_tree_lock(right);
1948                 btrfs_set_lock_blocking_write(right);
1949                 wret = btrfs_cow_block(trans, root, right,
1950                                        parent, pslot + 1, &right);
1951                 if (wret) {
1952                         ret = wret;
1953                         goto enospc;
1954                 }
1955         }
1956
1957         /* first, try to make some room in the middle buffer */
1958         if (left) {
1959                 orig_slot += btrfs_header_nritems(left);
1960                 wret = push_node_left(trans, left, mid, 1);
1961                 if (wret < 0)
1962                         ret = wret;
1963         }
1964
1965         /*
1966          * then try to empty the right most buffer into the middle
1967          */
1968         if (right) {
1969                 wret = push_node_left(trans, mid, right, 1);
1970                 if (wret < 0 && wret != -ENOSPC)
1971                         ret = wret;
1972                 if (btrfs_header_nritems(right) == 0) {
1973                         btrfs_clean_tree_block(right);
1974                         btrfs_tree_unlock(right);
1975                         del_ptr(root, path, level + 1, pslot + 1);
1976                         root_sub_used(root, right->len);
1977                         btrfs_free_tree_block(trans, root, right, 0, 1);
1978                         free_extent_buffer_stale(right);
1979                         right = NULL;
1980                 } else {
1981                         struct btrfs_disk_key right_key;
1982                         btrfs_node_key(right, &right_key, 0);
1983                         ret = tree_mod_log_insert_key(parent, pslot + 1,
1984                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
1985                         BUG_ON(ret < 0);
1986                         btrfs_set_node_key(parent, &right_key, pslot + 1);
1987                         btrfs_mark_buffer_dirty(parent);
1988                 }
1989         }
1990         if (btrfs_header_nritems(mid) == 1) {
1991                 /*
1992                  * we're not allowed to leave a node with one item in the
1993                  * tree during a delete.  A deletion from lower in the tree
1994                  * could try to delete the only pointer in this node.
1995                  * So, pull some keys from the left.
1996                  * There has to be a left pointer at this point because
1997                  * otherwise we would have pulled some pointers from the
1998                  * right
1999                  */
2000                 if (!left) {
2001                         ret = -EROFS;
2002                         btrfs_handle_fs_error(fs_info, ret, NULL);
2003                         goto enospc;
2004                 }
2005                 wret = balance_node_right(trans, mid, left);
2006                 if (wret < 0) {
2007                         ret = wret;
2008                         goto enospc;
2009                 }
2010                 if (wret == 1) {
2011                         wret = push_node_left(trans, left, mid, 1);
2012                         if (wret < 0)
2013                                 ret = wret;
2014                 }
2015                 BUG_ON(wret == 1);
2016         }
2017         if (btrfs_header_nritems(mid) == 0) {
2018                 btrfs_clean_tree_block(mid);
2019                 btrfs_tree_unlock(mid);
2020                 del_ptr(root, path, level + 1, pslot);
2021                 root_sub_used(root, mid->len);
2022                 btrfs_free_tree_block(trans, root, mid, 0, 1);
2023                 free_extent_buffer_stale(mid);
2024                 mid = NULL;
2025         } else {
2026                 /* update the parent key to reflect our changes */
2027                 struct btrfs_disk_key mid_key;
2028                 btrfs_node_key(mid, &mid_key, 0);
2029                 ret = tree_mod_log_insert_key(parent, pslot,
2030                                 MOD_LOG_KEY_REPLACE, GFP_NOFS);
2031                 BUG_ON(ret < 0);
2032                 btrfs_set_node_key(parent, &mid_key, pslot);
2033                 btrfs_mark_buffer_dirty(parent);
2034         }
2035
2036         /* update the path */
2037         if (left) {
2038                 if (btrfs_header_nritems(left) > orig_slot) {
2039                         extent_buffer_get(left);
2040                         /* left was locked after cow */
2041                         path->nodes[level] = left;
2042                         path->slots[level + 1] -= 1;
2043                         path->slots[level] = orig_slot;
2044                         if (mid) {
2045                                 btrfs_tree_unlock(mid);
2046                                 free_extent_buffer(mid);
2047                         }
2048                 } else {
2049                         orig_slot -= btrfs_header_nritems(left);
2050                         path->slots[level] = orig_slot;
2051                 }
2052         }
2053         /* double check we haven't messed things up */
2054         if (orig_ptr !=
2055             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
2056                 BUG();
2057 enospc:
2058         if (right) {
2059                 btrfs_tree_unlock(right);
2060                 free_extent_buffer(right);
2061         }
2062         if (left) {
2063                 if (path->nodes[level] != left)
2064                         btrfs_tree_unlock(left);
2065                 free_extent_buffer(left);
2066         }
2067         return ret;
2068 }
2069
2070 /* Node balancing for insertion.  Here we only split or push nodes around
2071  * when they are completely full.  This is also done top down, so we
2072  * have to be pessimistic.
2073  */
2074 static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
2075                                           struct btrfs_root *root,
2076                                           struct btrfs_path *path, int level)
2077 {
2078         struct btrfs_fs_info *fs_info = root->fs_info;
2079         struct extent_buffer *right = NULL;
2080         struct extent_buffer *mid;
2081         struct extent_buffer *left = NULL;
2082         struct extent_buffer *parent = NULL;
2083         int ret = 0;
2084         int wret;
2085         int pslot;
2086         int orig_slot = path->slots[level];
2087
2088         if (level == 0)
2089                 return 1;
2090
2091         mid = path->nodes[level];
2092         WARN_ON(btrfs_header_generation(mid) != trans->transid);
2093
2094         if (level < BTRFS_MAX_LEVEL - 1) {
2095                 parent = path->nodes[level + 1];
2096                 pslot = path->slots[level + 1];
2097         }
2098
2099         if (!parent)
2100                 return 1;
2101
2102         left = btrfs_read_node_slot(parent, pslot - 1);
2103         if (IS_ERR(left))
2104                 left = NULL;
2105
2106         /* first, try to make some room in the middle buffer */
2107         if (left) {
2108                 u32 left_nr;
2109
2110                 btrfs_tree_lock(left);
2111                 btrfs_set_lock_blocking_write(left);
2112
2113                 left_nr = btrfs_header_nritems(left);
2114                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2115                         wret = 1;
2116                 } else {
2117                         ret = btrfs_cow_block(trans, root, left, parent,
2118                                               pslot - 1, &left);
2119                         if (ret)
2120                                 wret = 1;
2121                         else {
2122                                 wret = push_node_left(trans, left, mid, 0);
2123                         }
2124                 }
2125                 if (wret < 0)
2126                         ret = wret;
2127                 if (wret == 0) {
2128                         struct btrfs_disk_key disk_key;
2129                         orig_slot += left_nr;
2130                         btrfs_node_key(mid, &disk_key, 0);
2131                         ret = tree_mod_log_insert_key(parent, pslot,
2132                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
2133                         BUG_ON(ret < 0);
2134                         btrfs_set_node_key(parent, &disk_key, pslot);
2135                         btrfs_mark_buffer_dirty(parent);
2136                         if (btrfs_header_nritems(left) > orig_slot) {
2137                                 path->nodes[level] = left;
2138                                 path->slots[level + 1] -= 1;
2139                                 path->slots[level] = orig_slot;
2140                                 btrfs_tree_unlock(mid);
2141                                 free_extent_buffer(mid);
2142                         } else {
2143                                 orig_slot -=
2144                                         btrfs_header_nritems(left);
2145                                 path->slots[level] = orig_slot;
2146                                 btrfs_tree_unlock(left);
2147                                 free_extent_buffer(left);
2148                         }
2149                         return 0;
2150                 }
2151                 btrfs_tree_unlock(left);
2152                 free_extent_buffer(left);
2153         }
2154         right = btrfs_read_node_slot(parent, pslot + 1);
2155         if (IS_ERR(right))
2156                 right = NULL;
2157
2158         /*
2159          * then try to empty the right most buffer into the middle
2160          */
2161         if (right) {
2162                 u32 right_nr;
2163
2164                 btrfs_tree_lock(right);
2165                 btrfs_set_lock_blocking_write(right);
2166
2167                 right_nr = btrfs_header_nritems(right);
2168                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2169                         wret = 1;
2170                 } else {
2171                         ret = btrfs_cow_block(trans, root, right,
2172                                               parent, pslot + 1,
2173                                               &right);
2174                         if (ret)
2175                                 wret = 1;
2176                         else {
2177                                 wret = balance_node_right(trans, right, mid);
2178                         }
2179                 }
2180                 if (wret < 0)
2181                         ret = wret;
2182                 if (wret == 0) {
2183                         struct btrfs_disk_key disk_key;
2184
2185                         btrfs_node_key(right, &disk_key, 0);
2186                         ret = tree_mod_log_insert_key(parent, pslot + 1,
2187                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
2188                         BUG_ON(ret < 0);
2189                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
2190                         btrfs_mark_buffer_dirty(parent);
2191
2192                         if (btrfs_header_nritems(mid) <= orig_slot) {
2193                                 path->nodes[level] = right;
2194                                 path->slots[level + 1] += 1;
2195                                 path->slots[level] = orig_slot -
2196                                         btrfs_header_nritems(mid);
2197                                 btrfs_tree_unlock(mid);
2198                                 free_extent_buffer(mid);
2199                         } else {
2200                                 btrfs_tree_unlock(right);
2201                                 free_extent_buffer(right);
2202                         }
2203                         return 0;
2204                 }
2205                 btrfs_tree_unlock(right);
2206                 free_extent_buffer(right);
2207         }
2208         return 1;
2209 }
2210
2211 /*
2212  * readahead one full node of leaves, finding things that are close
2213  * to the block in 'slot', and triggering ra on them.
2214  */
2215 static void reada_for_search(struct btrfs_fs_info *fs_info,
2216                              struct btrfs_path *path,
2217                              int level, int slot, u64 objectid)
2218 {
2219         struct extent_buffer *node;
2220         struct btrfs_disk_key disk_key;
2221         u32 nritems;
2222         u64 search;
2223         u64 target;
2224         u64 nread = 0;
2225         struct extent_buffer *eb;
2226         u32 nr;
2227         u32 blocksize;
2228         u32 nscan = 0;
2229
2230         if (level != 1)
2231                 return;
2232
2233         if (!path->nodes[level])
2234                 return;
2235
2236         node = path->nodes[level];
2237
2238         search = btrfs_node_blockptr(node, slot);
2239         blocksize = fs_info->nodesize;
2240         eb = find_extent_buffer(fs_info, search);
2241         if (eb) {
2242                 free_extent_buffer(eb);
2243                 return;
2244         }
2245
2246         target = search;
2247
2248         nritems = btrfs_header_nritems(node);
2249         nr = slot;
2250
2251         while (1) {
2252                 if (path->reada == READA_BACK) {
2253                         if (nr == 0)
2254                                 break;
2255                         nr--;
2256                 } else if (path->reada == READA_FORWARD) {
2257                         nr++;
2258                         if (nr >= nritems)
2259                                 break;
2260                 }
2261                 if (path->reada == READA_BACK && objectid) {
2262                         btrfs_node_key(node, &disk_key, nr);
2263                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
2264                                 break;
2265                 }
2266                 search = btrfs_node_blockptr(node, nr);
2267                 if ((search <= target && target - search <= 65536) ||
2268                     (search > target && search - target <= 65536)) {
2269                         readahead_tree_block(fs_info, search);
2270                         nread += blocksize;
2271                 }
2272                 nscan++;
2273                 if ((nread > 65536 || nscan > 32))
2274                         break;
2275         }
2276 }
2277
2278 static noinline void reada_for_balance(struct btrfs_fs_info *fs_info,
2279                                        struct btrfs_path *path, int level)
2280 {
2281         int slot;
2282         int nritems;
2283         struct extent_buffer *parent;
2284         struct extent_buffer *eb;
2285         u64 gen;
2286         u64 block1 = 0;
2287         u64 block2 = 0;
2288
2289         parent = path->nodes[level + 1];
2290         if (!parent)
2291                 return;
2292
2293         nritems = btrfs_header_nritems(parent);
2294         slot = path->slots[level + 1];
2295
2296         if (slot > 0) {
2297                 block1 = btrfs_node_blockptr(parent, slot - 1);
2298                 gen = btrfs_node_ptr_generation(parent, slot - 1);
2299                 eb = find_extent_buffer(fs_info, block1);
2300                 /*
2301                  * if we get -eagain from btrfs_buffer_uptodate, we
2302                  * don't want to return eagain here.  That will loop
2303                  * forever
2304                  */
2305                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2306                         block1 = 0;
2307                 free_extent_buffer(eb);
2308         }
2309         if (slot + 1 < nritems) {
2310                 block2 = btrfs_node_blockptr(parent, slot + 1);
2311                 gen = btrfs_node_ptr_generation(parent, slot + 1);
2312                 eb = find_extent_buffer(fs_info, block2);
2313                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2314                         block2 = 0;
2315                 free_extent_buffer(eb);
2316         }
2317
2318         if (block1)
2319                 readahead_tree_block(fs_info, block1);
2320         if (block2)
2321                 readahead_tree_block(fs_info, block2);
2322 }
2323
2324
2325 /*
2326  * when we walk down the tree, it is usually safe to unlock the higher layers
2327  * in the tree.  The exceptions are when our path goes through slot 0, because
2328  * operations on the tree might require changing key pointers higher up in the
2329  * tree.
2330  *
2331  * callers might also have set path->keep_locks, which tells this code to keep
2332  * the lock if the path points to the last slot in the block.  This is part of
2333  * walking through the tree, and selecting the next slot in the higher block.
2334  *
2335  * lowest_unlock sets the lowest level in the tree we're allowed to unlock.  so
2336  * if lowest_unlock is 1, level 0 won't be unlocked
2337  */
2338 static noinline void unlock_up(struct btrfs_path *path, int level,
2339                                int lowest_unlock, int min_write_lock_level,
2340                                int *write_lock_level)
2341 {
2342         int i;
2343         int skip_level = level;
2344         int no_skips = 0;
2345         struct extent_buffer *t;
2346
2347         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2348                 if (!path->nodes[i])
2349                         break;
2350                 if (!path->locks[i])
2351                         break;
2352                 if (!no_skips && path->slots[i] == 0) {
2353                         skip_level = i + 1;
2354                         continue;
2355                 }
2356                 if (!no_skips && path->keep_locks) {
2357                         u32 nritems;
2358                         t = path->nodes[i];
2359                         nritems = btrfs_header_nritems(t);
2360                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
2361                                 skip_level = i + 1;
2362                                 continue;
2363                         }
2364                 }
2365                 if (skip_level < i && i >= lowest_unlock)
2366                         no_skips = 1;
2367
2368                 t = path->nodes[i];
2369                 if (i >= lowest_unlock && i > skip_level) {
2370                         btrfs_tree_unlock_rw(t, path->locks[i]);
2371                         path->locks[i] = 0;
2372                         if (write_lock_level &&
2373                             i > min_write_lock_level &&
2374                             i <= *write_lock_level) {
2375                                 *write_lock_level = i - 1;
2376                         }
2377                 }
2378         }
2379 }
2380
2381 /*
2382  * This releases any locks held in the path starting at level and
2383  * going all the way up to the root.
2384  *
2385  * btrfs_search_slot will keep the lock held on higher nodes in a few
2386  * corner cases, such as COW of the block at slot zero in the node.  This
2387  * ignores those rules, and it should only be called when there are no
2388  * more updates to be done higher up in the tree.
2389  */
2390 noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
2391 {
2392         int i;
2393
2394         if (path->keep_locks)
2395                 return;
2396
2397         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2398                 if (!path->nodes[i])
2399                         continue;
2400                 if (!path->locks[i])
2401                         continue;
2402                 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
2403                 path->locks[i] = 0;
2404         }
2405 }
2406
2407 /*
2408  * helper function for btrfs_search_slot.  The goal is to find a block
2409  * in cache without setting the path to blocking.  If we find the block
2410  * we return zero and the path is unchanged.
2411  *
2412  * If we can't find the block, we set the path blocking and do some
2413  * reada.  -EAGAIN is returned and the search must be repeated.
2414  */
2415 static int
2416 read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
2417                       struct extent_buffer **eb_ret, int level, int slot,
2418                       const struct btrfs_key *key)
2419 {
2420         struct btrfs_fs_info *fs_info = root->fs_info;
2421         u64 blocknr;
2422         u64 gen;
2423         struct extent_buffer *b = *eb_ret;
2424         struct extent_buffer *tmp;
2425         struct btrfs_key first_key;
2426         int ret;
2427         int parent_level;
2428
2429         blocknr = btrfs_node_blockptr(b, slot);
2430         gen = btrfs_node_ptr_generation(b, slot);
2431         parent_level = btrfs_header_level(b);
2432         btrfs_node_key_to_cpu(b, &first_key, slot);
2433
2434         tmp = find_extent_buffer(fs_info, blocknr);
2435         if (tmp) {
2436                 /* first we do an atomic uptodate check */
2437                 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
2438                         /*
2439                          * Do extra check for first_key, eb can be stale due to
2440                          * being cached, read from scrub, or have multiple
2441                          * parents (shared tree blocks).
2442                          */
2443                         if (btrfs_verify_level_key(tmp,
2444                                         parent_level - 1, &first_key, gen)) {
2445                                 free_extent_buffer(tmp);
2446                                 return -EUCLEAN;
2447                         }
2448                         *eb_ret = tmp;
2449                         return 0;
2450                 }
2451
2452                 /* the pages were up to date, but we failed
2453                  * the generation number check.  Do a full
2454                  * read for the generation number that is correct.
2455                  * We must do this without dropping locks so
2456                  * we can trust our generation number
2457                  */
2458                 btrfs_set_path_blocking(p);
2459
2460                 /* now we're allowed to do a blocking uptodate check */
2461                 ret = btrfs_read_buffer(tmp, gen, parent_level - 1, &first_key);
2462                 if (!ret) {
2463                         *eb_ret = tmp;
2464                         return 0;
2465                 }
2466                 free_extent_buffer(tmp);
2467                 btrfs_release_path(p);
2468                 return -EIO;
2469         }
2470
2471         /*
2472          * reduce lock contention at high levels
2473          * of the btree by dropping locks before
2474          * we read.  Don't release the lock on the current
2475          * level because we need to walk this node to figure
2476          * out which blocks to read.
2477          */
2478         btrfs_unlock_up_safe(p, level + 1);
2479         btrfs_set_path_blocking(p);
2480
2481         if (p->reada != READA_NONE)
2482                 reada_for_search(fs_info, p, level, slot, key->objectid);
2483
2484         ret = -EAGAIN;
2485         tmp = read_tree_block(fs_info, blocknr, gen, parent_level - 1,
2486                               &first_key);
2487         if (!IS_ERR(tmp)) {
2488                 /*
2489                  * If the read above didn't mark this buffer up to date,
2490                  * it will never end up being up to date.  Set ret to EIO now
2491                  * and give up so that our caller doesn't loop forever
2492                  * on our EAGAINs.
2493                  */
2494                 if (!extent_buffer_uptodate(tmp))
2495                         ret = -EIO;
2496                 free_extent_buffer(tmp);
2497         } else {
2498                 ret = PTR_ERR(tmp);
2499         }
2500
2501         btrfs_release_path(p);
2502         return ret;
2503 }
2504
2505 /*
2506  * helper function for btrfs_search_slot.  This does all of the checks
2507  * for node-level blocks and does any balancing required based on
2508  * the ins_len.
2509  *
2510  * If no extra work was required, zero is returned.  If we had to
2511  * drop the path, -EAGAIN is returned and btrfs_search_slot must
2512  * start over
2513  */
2514 static int
2515 setup_nodes_for_search(struct btrfs_trans_handle *trans,
2516                        struct btrfs_root *root, struct btrfs_path *p,
2517                        struct extent_buffer *b, int level, int ins_len,
2518                        int *write_lock_level)
2519 {
2520         struct btrfs_fs_info *fs_info = root->fs_info;
2521         int ret;
2522
2523         if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
2524             BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3) {
2525                 int sret;
2526
2527                 if (*write_lock_level < level + 1) {
2528                         *write_lock_level = level + 1;
2529                         btrfs_release_path(p);
2530                         goto again;
2531                 }
2532
2533                 btrfs_set_path_blocking(p);
2534                 reada_for_balance(fs_info, p, level);
2535                 sret = split_node(trans, root, p, level);
2536
2537                 BUG_ON(sret > 0);
2538                 if (sret) {
2539                         ret = sret;
2540                         goto done;
2541                 }
2542                 b = p->nodes[level];
2543         } else if (ins_len < 0 && btrfs_header_nritems(b) <
2544                    BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 2) {
2545                 int sret;
2546
2547                 if (*write_lock_level < level + 1) {
2548                         *write_lock_level = level + 1;
2549                         btrfs_release_path(p);
2550                         goto again;
2551                 }
2552
2553                 btrfs_set_path_blocking(p);
2554                 reada_for_balance(fs_info, p, level);
2555                 sret = balance_level(trans, root, p, level);
2556
2557                 if (sret) {
2558                         ret = sret;
2559                         goto done;
2560                 }
2561                 b = p->nodes[level];
2562                 if (!b) {
2563                         btrfs_release_path(p);
2564                         goto again;
2565                 }
2566                 BUG_ON(btrfs_header_nritems(b) == 1);
2567         }
2568         return 0;
2569
2570 again:
2571         ret = -EAGAIN;
2572 done:
2573         return ret;
2574 }
2575
2576 static int key_search(struct extent_buffer *b, const struct btrfs_key *key,
2577                       int level, int *prev_cmp, int *slot)
2578 {
2579         if (*prev_cmp != 0) {
2580                 *prev_cmp = btrfs_bin_search(b, key, level, slot);
2581                 return *prev_cmp;
2582         }
2583
2584         *slot = 0;
2585
2586         return 0;
2587 }
2588
2589 int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
2590                 u64 iobjectid, u64 ioff, u8 key_type,
2591                 struct btrfs_key *found_key)
2592 {
2593         int ret;
2594         struct btrfs_key key;
2595         struct extent_buffer *eb;
2596
2597         ASSERT(path);
2598         ASSERT(found_key);
2599
2600         key.type = key_type;
2601         key.objectid = iobjectid;
2602         key.offset = ioff;
2603
2604         ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
2605         if (ret < 0)
2606                 return ret;
2607
2608         eb = path->nodes[0];
2609         if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
2610                 ret = btrfs_next_leaf(fs_root, path);
2611                 if (ret)
2612                         return ret;
2613                 eb = path->nodes[0];
2614         }
2615
2616         btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
2617         if (found_key->type != key.type ||
2618                         found_key->objectid != key.objectid)
2619                 return 1;
2620
2621         return 0;
2622 }
2623
2624 static struct extent_buffer *btrfs_search_slot_get_root(struct btrfs_root *root,
2625                                                         struct btrfs_path *p,
2626                                                         int write_lock_level)
2627 {
2628         struct btrfs_fs_info *fs_info = root->fs_info;
2629         struct extent_buffer *b;
2630         int root_lock;
2631         int level = 0;
2632
2633         /* We try very hard to do read locks on the root */
2634         root_lock = BTRFS_READ_LOCK;
2635
2636         if (p->search_commit_root) {
2637                 /*
2638                  * The commit roots are read only so we always do read locks,
2639                  * and we always must hold the commit_root_sem when doing
2640                  * searches on them, the only exception is send where we don't
2641                  * want to block transaction commits for a long time, so
2642                  * we need to clone the commit root in order to avoid races
2643                  * with transaction commits that create a snapshot of one of
2644                  * the roots used by a send operation.
2645                  */
2646                 if (p->need_commit_sem) {
2647                         down_read(&fs_info->commit_root_sem);
2648                         b = btrfs_clone_extent_buffer(root->commit_root);
2649                         up_read(&fs_info->commit_root_sem);
2650                         if (!b)
2651                                 return ERR_PTR(-ENOMEM);
2652
2653                 } else {
2654                         b = root->commit_root;
2655                         extent_buffer_get(b);
2656                 }
2657                 level = btrfs_header_level(b);
2658                 /*
2659                  * Ensure that all callers have set skip_locking when
2660                  * p->search_commit_root = 1.
2661                  */
2662                 ASSERT(p->skip_locking == 1);
2663
2664                 goto out;
2665         }
2666
2667         if (p->skip_locking) {
2668                 b = btrfs_root_node(root);
2669                 level = btrfs_header_level(b);
2670                 goto out;
2671         }
2672
2673         /*
2674          * If the level is set to maximum, we can skip trying to get the read
2675          * lock.
2676          */
2677         if (write_lock_level < BTRFS_MAX_LEVEL) {
2678                 /*
2679                  * We don't know the level of the root node until we actually
2680                  * have it read locked
2681                  */
2682                 b = btrfs_read_lock_root_node(root);
2683                 level = btrfs_header_level(b);
2684                 if (level > write_lock_level)
2685                         goto out;
2686
2687                 /* Whoops, must trade for write lock */
2688                 btrfs_tree_read_unlock(b);
2689                 free_extent_buffer(b);
2690         }
2691
2692         b = btrfs_lock_root_node(root);
2693         root_lock = BTRFS_WRITE_LOCK;
2694
2695         /* The level might have changed, check again */
2696         level = btrfs_header_level(b);
2697
2698 out:
2699         p->nodes[level] = b;
2700         if (!p->skip_locking)
2701                 p->locks[level] = root_lock;
2702         /*
2703          * Callers are responsible for dropping b's references.
2704          */
2705         return b;
2706 }
2707
2708
2709 /*
2710  * btrfs_search_slot - look for a key in a tree and perform necessary
2711  * modifications to preserve tree invariants.
2712  *
2713  * @trans:      Handle of transaction, used when modifying the tree
2714  * @p:          Holds all btree nodes along the search path
2715  * @root:       The root node of the tree
2716  * @key:        The key we are looking for
2717  * @ins_len:    Indicates purpose of search, for inserts it is 1, for
2718  *              deletions it's -1. 0 for plain searches
2719  * @cow:        boolean should CoW operations be performed. Must always be 1
2720  *              when modifying the tree.
2721  *
2722  * If @ins_len > 0, nodes and leaves will be split as we walk down the tree.
2723  * If @ins_len < 0, nodes will be merged as we walk down the tree (if possible)
2724  *
2725  * If @key is found, 0 is returned and you can find the item in the leaf level
2726  * of the path (level 0)
2727  *
2728  * If @key isn't found, 1 is returned and the leaf level of the path (level 0)
2729  * points to the slot where it should be inserted
2730  *
2731  * If an error is encountered while searching the tree a negative error number
2732  * is returned
2733  */
2734 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2735                       const struct btrfs_key *key, struct btrfs_path *p,
2736                       int ins_len, int cow)
2737 {
2738         struct extent_buffer *b;
2739         int slot;
2740         int ret;
2741         int err;
2742         int level;
2743         int lowest_unlock = 1;
2744         /* everything at write_lock_level or lower must be write locked */
2745         int write_lock_level = 0;
2746         u8 lowest_level = 0;
2747         int min_write_lock_level;
2748         int prev_cmp;
2749
2750         lowest_level = p->lowest_level;
2751         WARN_ON(lowest_level && ins_len > 0);
2752         WARN_ON(p->nodes[0] != NULL);
2753         BUG_ON(!cow && ins_len);
2754
2755         if (ins_len < 0) {
2756                 lowest_unlock = 2;
2757
2758                 /* when we are removing items, we might have to go up to level
2759                  * two as we update tree pointers  Make sure we keep write
2760                  * for those levels as well
2761                  */
2762                 write_lock_level = 2;
2763         } else if (ins_len > 0) {
2764                 /*
2765                  * for inserting items, make sure we have a write lock on
2766                  * level 1 so we can update keys
2767                  */
2768                 write_lock_level = 1;
2769         }
2770
2771         if (!cow)
2772                 write_lock_level = -1;
2773
2774         if (cow && (p->keep_locks || p->lowest_level))
2775                 write_lock_level = BTRFS_MAX_LEVEL;
2776
2777         min_write_lock_level = write_lock_level;
2778
2779 again:
2780         prev_cmp = -1;
2781         b = btrfs_search_slot_get_root(root, p, write_lock_level);
2782         if (IS_ERR(b)) {
2783                 ret = PTR_ERR(b);
2784                 goto done;
2785         }
2786
2787         while (b) {
2788                 level = btrfs_header_level(b);
2789
2790                 /*
2791                  * setup the path here so we can release it under lock
2792                  * contention with the cow code
2793                  */
2794                 if (cow) {
2795                         bool last_level = (level == (BTRFS_MAX_LEVEL - 1));
2796
2797                         /*
2798                          * if we don't really need to cow this block
2799                          * then we don't want to set the path blocking,
2800                          * so we test it here
2801                          */
2802                         if (!should_cow_block(trans, root, b)) {
2803                                 trans->dirty = true;
2804                                 goto cow_done;
2805                         }
2806
2807                         /*
2808                          * must have write locks on this node and the
2809                          * parent
2810                          */
2811                         if (level > write_lock_level ||
2812                             (level + 1 > write_lock_level &&
2813                             level + 1 < BTRFS_MAX_LEVEL &&
2814                             p->nodes[level + 1])) {
2815                                 write_lock_level = level + 1;
2816                                 btrfs_release_path(p);
2817                                 goto again;
2818                         }
2819
2820                         btrfs_set_path_blocking(p);
2821                         if (last_level)
2822                                 err = btrfs_cow_block(trans, root, b, NULL, 0,
2823                                                       &b);
2824                         else
2825                                 err = btrfs_cow_block(trans, root, b,
2826                                                       p->nodes[level + 1],
2827                                                       p->slots[level + 1], &b);
2828                         if (err) {
2829                                 ret = err;
2830                                 goto done;
2831                         }
2832                 }
2833 cow_done:
2834                 p->nodes[level] = b;
2835                 /*
2836                  * Leave path with blocking locks to avoid massive
2837                  * lock context switch, this is made on purpose.
2838                  */
2839
2840                 /*
2841                  * we have a lock on b and as long as we aren't changing
2842                  * the tree, there is no way to for the items in b to change.
2843                  * It is safe to drop the lock on our parent before we
2844                  * go through the expensive btree search on b.
2845                  *
2846                  * If we're inserting or deleting (ins_len != 0), then we might
2847                  * be changing slot zero, which may require changing the parent.
2848                  * So, we can't drop the lock until after we know which slot
2849                  * we're operating on.
2850                  */
2851                 if (!ins_len && !p->keep_locks) {
2852                         int u = level + 1;
2853
2854                         if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
2855                                 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
2856                                 p->locks[u] = 0;
2857                         }
2858                 }
2859
2860                 ret = key_search(b, key, level, &prev_cmp, &slot);
2861                 if (ret < 0)
2862                         goto done;
2863
2864                 if (level != 0) {
2865                         int dec = 0;
2866                         if (ret && slot > 0) {
2867                                 dec = 1;
2868                                 slot -= 1;
2869                         }
2870                         p->slots[level] = slot;
2871                         err = setup_nodes_for_search(trans, root, p, b, level,
2872                                              ins_len, &write_lock_level);
2873                         if (err == -EAGAIN)
2874                                 goto again;
2875                         if (err) {
2876                                 ret = err;
2877                                 goto done;
2878                         }
2879                         b = p->nodes[level];
2880                         slot = p->slots[level];
2881
2882                         /*
2883                          * slot 0 is special, if we change the key
2884                          * we have to update the parent pointer
2885                          * which means we must have a write lock
2886                          * on the parent
2887                          */
2888                         if (slot == 0 && ins_len &&
2889                             write_lock_level < level + 1) {
2890                                 write_lock_level = level + 1;
2891                                 btrfs_release_path(p);
2892                                 goto again;
2893                         }
2894
2895                         unlock_up(p, level, lowest_unlock,
2896                                   min_write_lock_level, &write_lock_level);
2897
2898                         if (level == lowest_level) {
2899                                 if (dec)
2900                                         p->slots[level]++;
2901                                 goto done;
2902                         }
2903
2904                         err = read_block_for_search(root, p, &b, level,
2905                                                     slot, key);
2906                         if (err == -EAGAIN)
2907                                 goto again;
2908                         if (err) {
2909                                 ret = err;
2910                                 goto done;
2911                         }
2912
2913                         if (!p->skip_locking) {
2914                                 level = btrfs_header_level(b);
2915                                 if (level <= write_lock_level) {
2916                                         if (!btrfs_try_tree_write_lock(b)) {
2917                                                 btrfs_set_path_blocking(p);
2918                                                 btrfs_tree_lock(b);
2919                                         }
2920                                         p->locks[level] = BTRFS_WRITE_LOCK;
2921                                 } else {
2922                                         if (!btrfs_tree_read_lock_atomic(b)) {
2923                                                 btrfs_set_path_blocking(p);
2924                                                 btrfs_tree_read_lock(b);
2925                                         }
2926                                         p->locks[level] = BTRFS_READ_LOCK;
2927                                 }
2928                                 p->nodes[level] = b;
2929                         }
2930                 } else {
2931                         p->slots[level] = slot;
2932                         if (ins_len > 0 &&
2933                             btrfs_leaf_free_space(b) < ins_len) {
2934                                 if (write_lock_level < 1) {
2935                                         write_lock_level = 1;
2936                                         btrfs_release_path(p);
2937                                         goto again;
2938                                 }
2939
2940                                 btrfs_set_path_blocking(p);
2941                                 err = split_leaf(trans, root, key,
2942                                                  p, ins_len, ret == 0);
2943
2944                                 BUG_ON(err > 0);
2945                                 if (err) {
2946                                         ret = err;
2947                                         goto done;
2948                                 }
2949                         }
2950                         if (!p->search_for_split)
2951                                 unlock_up(p, level, lowest_unlock,
2952                                           min_write_lock_level, NULL);
2953                         goto done;
2954                 }
2955         }
2956         ret = 1;
2957 done:
2958         /*
2959          * we don't really know what they plan on doing with the path
2960          * from here on, so for now just mark it as blocking
2961          */
2962         if (!p->leave_spinning)
2963                 btrfs_set_path_blocking(p);
2964         if (ret < 0 && !p->skip_release_on_error)
2965                 btrfs_release_path(p);
2966         return ret;
2967 }
2968
2969 /*
2970  * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2971  * current state of the tree together with the operations recorded in the tree
2972  * modification log to search for the key in a previous version of this tree, as
2973  * denoted by the time_seq parameter.
2974  *
2975  * Naturally, there is no support for insert, delete or cow operations.
2976  *
2977  * The resulting path and return value will be set up as if we called
2978  * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2979  */
2980 int btrfs_search_old_slot(struct btrfs_root *root, const struct btrfs_key *key,
2981                           struct btrfs_path *p, u64 time_seq)
2982 {
2983         struct btrfs_fs_info *fs_info = root->fs_info;
2984         struct extent_buffer *b;
2985         int slot;
2986         int ret;
2987         int err;
2988         int level;
2989         int lowest_unlock = 1;
2990         u8 lowest_level = 0;
2991         int prev_cmp = -1;
2992
2993         lowest_level = p->lowest_level;
2994         WARN_ON(p->nodes[0] != NULL);
2995
2996         if (p->search_commit_root) {
2997                 BUG_ON(time_seq);
2998                 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2999         }
3000
3001 again:
3002         b = get_old_root(root, time_seq);
3003         if (!b) {
3004                 ret = -EIO;
3005                 goto done;
3006         }
3007         level = btrfs_header_level(b);
3008         p->locks[level] = BTRFS_READ_LOCK;
3009
3010         while (b) {
3011                 level = btrfs_header_level(b);
3012                 p->nodes[level] = b;
3013
3014                 /*
3015                  * we have a lock on b and as long as we aren't changing
3016                  * the tree, there is no way to for the items in b to change.
3017                  * It is safe to drop the lock on our parent before we
3018                  * go through the expensive btree search on b.
3019                  */
3020                 btrfs_unlock_up_safe(p, level + 1);
3021
3022                 /*
3023                  * Since we can unwind ebs we want to do a real search every
3024                  * time.
3025                  */
3026                 prev_cmp = -1;
3027                 ret = key_search(b, key, level, &prev_cmp, &slot);
3028                 if (ret < 0)
3029                         goto done;
3030
3031                 if (level != 0) {
3032                         int dec = 0;
3033                         if (ret && slot > 0) {
3034                                 dec = 1;
3035                                 slot -= 1;
3036                         }
3037                         p->slots[level] = slot;
3038                         unlock_up(p, level, lowest_unlock, 0, NULL);
3039
3040                         if (level == lowest_level) {
3041                                 if (dec)
3042                                         p->slots[level]++;
3043                                 goto done;
3044                         }
3045
3046                         err = read_block_for_search(root, p, &b, level,
3047                                                     slot, key);
3048                         if (err == -EAGAIN)
3049                                 goto again;
3050                         if (err) {
3051                                 ret = err;
3052                                 goto done;
3053                         }
3054
3055                         level = btrfs_header_level(b);
3056                         if (!btrfs_tree_read_lock_atomic(b)) {
3057                                 btrfs_set_path_blocking(p);
3058                                 btrfs_tree_read_lock(b);
3059                         }
3060                         b = tree_mod_log_rewind(fs_info, p, b, time_seq);
3061                         if (!b) {
3062                                 ret = -ENOMEM;
3063                                 goto done;
3064                         }
3065                         p->locks[level] = BTRFS_READ_LOCK;
3066                         p->nodes[level] = b;
3067                 } else {
3068                         p->slots[level] = slot;
3069                         unlock_up(p, level, lowest_unlock, 0, NULL);
3070                         goto done;
3071                 }
3072         }
3073         ret = 1;
3074 done:
3075         if (!p->leave_spinning)
3076                 btrfs_set_path_blocking(p);
3077         if (ret < 0)
3078                 btrfs_release_path(p);
3079
3080         return ret;
3081 }
3082
3083 /*
3084  * helper to use instead of search slot if no exact match is needed but
3085  * instead the next or previous item should be returned.
3086  * When find_higher is true, the next higher item is returned, the next lower
3087  * otherwise.
3088  * When return_any and find_higher are both true, and no higher item is found,
3089  * return the next lower instead.
3090  * When return_any is true and find_higher is false, and no lower item is found,
3091  * return the next higher instead.
3092  * It returns 0 if any item is found, 1 if none is found (tree empty), and
3093  * < 0 on error
3094  */
3095 int btrfs_search_slot_for_read(struct btrfs_root *root,
3096                                const struct btrfs_key *key,
3097                                struct btrfs_path *p, int find_higher,
3098                                int return_any)
3099 {
3100         int ret;
3101         struct extent_buffer *leaf;
3102
3103 again:
3104         ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
3105         if (ret <= 0)
3106                 return ret;
3107         /*
3108          * a return value of 1 means the path is at the position where the
3109          * item should be inserted. Normally this is the next bigger item,
3110          * but in case the previous item is the last in a leaf, path points
3111          * to the first free slot in the previous leaf, i.e. at an invalid
3112          * item.
3113          */
3114         leaf = p->nodes[0];
3115
3116         if (find_higher) {
3117                 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
3118                         ret = btrfs_next_leaf(root, p);
3119                         if (ret <= 0)
3120                                 return ret;
3121                         if (!return_any)
3122                                 return 1;
3123                         /*
3124                          * no higher item found, return the next
3125                          * lower instead
3126                          */
3127                         return_any = 0;
3128                         find_higher = 0;
3129                         btrfs_release_path(p);
3130                         goto again;
3131                 }
3132         } else {
3133                 if (p->slots[0] == 0) {
3134                         ret = btrfs_prev_leaf(root, p);
3135                         if (ret < 0)
3136                                 return ret;
3137                         if (!ret) {
3138                                 leaf = p->nodes[0];
3139                                 if (p->slots[0] == btrfs_header_nritems(leaf))
3140                                         p->slots[0]--;
3141                                 return 0;
3142                         }
3143                         if (!return_any)
3144                                 return 1;
3145                         /*
3146                          * no lower item found, return the next
3147                          * higher instead
3148                          */
3149                         return_any = 0;
3150                         find_higher = 1;
3151                         btrfs_release_path(p);
3152                         goto again;
3153                 } else {
3154                         --p->slots[0];
3155                 }
3156         }
3157         return 0;
3158 }
3159
3160 /*
3161  * adjust the pointers going up the tree, starting at level
3162  * making sure the right key of each node is points to 'key'.
3163  * This is used after shifting pointers to the left, so it stops
3164  * fixing up pointers when a given leaf/node is not in slot 0 of the
3165  * higher levels
3166  *
3167  */
3168 static void fixup_low_keys(struct btrfs_path *path,
3169                            struct btrfs_disk_key *key, int level)
3170 {
3171         int i;
3172         struct extent_buffer *t;
3173         int ret;
3174
3175         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
3176                 int tslot = path->slots[i];
3177
3178                 if (!path->nodes[i])
3179                         break;
3180                 t = path->nodes[i];
3181                 ret = tree_mod_log_insert_key(t, tslot, MOD_LOG_KEY_REPLACE,
3182                                 GFP_ATOMIC);
3183                 BUG_ON(ret < 0);
3184                 btrfs_set_node_key(t, key, tslot);
3185                 btrfs_mark_buffer_dirty(path->nodes[i]);
3186                 if (tslot != 0)
3187                         break;
3188         }
3189 }
3190
3191 /*
3192  * update item key.
3193  *
3194  * This function isn't completely safe. It's the caller's responsibility
3195  * that the new key won't break the order
3196  */
3197 void btrfs_set_item_key_safe(struct btrfs_fs_info *fs_info,
3198                              struct btrfs_path *path,
3199                              const struct btrfs_key *new_key)
3200 {
3201         struct btrfs_disk_key disk_key;
3202         struct extent_buffer *eb;
3203         int slot;
3204
3205         eb = path->nodes[0];
3206         slot = path->slots[0];
3207         if (slot > 0) {
3208                 btrfs_item_key(eb, &disk_key, slot - 1);
3209                 if (unlikely(comp_keys(&disk_key, new_key) >= 0)) {
3210                         btrfs_crit(fs_info,
3211                 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
3212                                    slot, btrfs_disk_key_objectid(&disk_key),
3213                                    btrfs_disk_key_type(&disk_key),
3214                                    btrfs_disk_key_offset(&disk_key),
3215                                    new_key->objectid, new_key->type,
3216                                    new_key->offset);
3217                         btrfs_print_leaf(eb);
3218                         BUG();
3219                 }
3220         }
3221         if (slot < btrfs_header_nritems(eb) - 1) {
3222                 btrfs_item_key(eb, &disk_key, slot + 1);
3223                 if (unlikely(comp_keys(&disk_key, new_key) <= 0)) {
3224                         btrfs_crit(fs_info,
3225                 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
3226                                    slot, btrfs_disk_key_objectid(&disk_key),
3227                                    btrfs_disk_key_type(&disk_key),
3228                                    btrfs_disk_key_offset(&disk_key),
3229                                    new_key->objectid, new_key->type,
3230                                    new_key->offset);
3231                         btrfs_print_leaf(eb);
3232                         BUG();
3233                 }
3234         }
3235
3236         btrfs_cpu_key_to_disk(&disk_key, new_key);
3237         btrfs_set_item_key(eb, &disk_key, slot);
3238         btrfs_mark_buffer_dirty(eb);
3239         if (slot == 0)
3240                 fixup_low_keys(path, &disk_key, 1);
3241 }
3242
3243 /*
3244  * try to push data from one node into the next node left in the
3245  * tree.
3246  *
3247  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
3248  * error, and > 0 if there was no room in the left hand block.
3249  */
3250 static int push_node_left(struct btrfs_trans_handle *trans,
3251                           struct extent_buffer *dst,
3252                           struct extent_buffer *src, int empty)
3253 {
3254         struct btrfs_fs_info *fs_info = trans->fs_info;
3255         int push_items = 0;
3256         int src_nritems;
3257         int dst_nritems;
3258         int ret = 0;
3259
3260         src_nritems = btrfs_header_nritems(src);
3261         dst_nritems = btrfs_header_nritems(dst);
3262         push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3263         WARN_ON(btrfs_header_generation(src) != trans->transid);
3264         WARN_ON(btrfs_header_generation(dst) != trans->transid);
3265
3266         if (!empty && src_nritems <= 8)
3267                 return 1;
3268
3269         if (push_items <= 0)
3270                 return 1;
3271
3272         if (empty) {
3273                 push_items = min(src_nritems, push_items);
3274                 if (push_items < src_nritems) {
3275                         /* leave at least 8 pointers in the node if
3276                          * we aren't going to empty it
3277                          */
3278                         if (src_nritems - push_items < 8) {
3279                                 if (push_items <= 8)
3280                                         return 1;
3281                                 push_items -= 8;
3282                         }
3283                 }
3284         } else
3285                 push_items = min(src_nritems - 8, push_items);
3286
3287         ret = tree_mod_log_eb_copy(dst, src, dst_nritems, 0, push_items);
3288         if (ret) {
3289                 btrfs_abort_transaction(trans, ret);
3290                 return ret;
3291         }
3292         copy_extent_buffer(dst, src,
3293                            btrfs_node_key_ptr_offset(dst_nritems),
3294                            btrfs_node_key_ptr_offset(0),
3295                            push_items * sizeof(struct btrfs_key_ptr));
3296
3297         if (push_items < src_nritems) {
3298                 /*
3299                  * Don't call tree_mod_log_insert_move here, key removal was
3300                  * already fully logged by tree_mod_log_eb_copy above.
3301                  */
3302                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
3303                                       btrfs_node_key_ptr_offset(push_items),
3304                                       (src_nritems - push_items) *
3305                                       sizeof(struct btrfs_key_ptr));
3306         }
3307         btrfs_set_header_nritems(src, src_nritems - push_items);
3308         btrfs_set_header_nritems(dst, dst_nritems + push_items);
3309         btrfs_mark_buffer_dirty(src);
3310         btrfs_mark_buffer_dirty(dst);
3311
3312         return ret;
3313 }
3314
3315 /*
3316  * try to push data from one node into the next node right in the
3317  * tree.
3318  *
3319  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
3320  * error, and > 0 if there was no room in the right hand block.
3321  *
3322  * this will  only push up to 1/2 the contents of the left node over
3323  */
3324 static int balance_node_right(struct btrfs_trans_handle *trans,
3325                               struct extent_buffer *dst,
3326                               struct extent_buffer *src)
3327 {
3328         struct btrfs_fs_info *fs_info = trans->fs_info;
3329         int push_items = 0;
3330         int max_push;
3331         int src_nritems;
3332         int dst_nritems;
3333         int ret = 0;
3334
3335         WARN_ON(btrfs_header_generation(src) != trans->transid);
3336         WARN_ON(btrfs_header_generation(dst) != trans->transid);
3337
3338         src_nritems = btrfs_header_nritems(src);
3339         dst_nritems = btrfs_header_nritems(dst);
3340         push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3341         if (push_items <= 0)
3342                 return 1;
3343
3344         if (src_nritems < 4)
3345                 return 1;
3346
3347         max_push = src_nritems / 2 + 1;
3348         /* don't try to empty the node */
3349         if (max_push >= src_nritems)
3350                 return 1;
3351
3352         if (max_push < push_items)
3353                 push_items = max_push;
3354
3355         ret = tree_mod_log_insert_move(dst, push_items, 0, dst_nritems);
3356         BUG_ON(ret < 0);
3357         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
3358                                       btrfs_node_key_ptr_offset(0),
3359                                       (dst_nritems) *
3360                                       sizeof(struct btrfs_key_ptr));
3361
3362         ret = tree_mod_log_eb_copy(dst, src, 0, src_nritems - push_items,
3363                                    push_items);
3364         if (ret) {
3365                 btrfs_abort_transaction(trans, ret);
3366                 return ret;
3367         }
3368         copy_extent_buffer(dst, src,
3369                            btrfs_node_key_ptr_offset(0),
3370                            btrfs_node_key_ptr_offset(src_nritems - push_items),
3371                            push_items * sizeof(struct btrfs_key_ptr));
3372
3373         btrfs_set_header_nritems(src, src_nritems - push_items);
3374         btrfs_set_header_nritems(dst, dst_nritems + push_items);
3375
3376         btrfs_mark_buffer_dirty(src);
3377         btrfs_mark_buffer_dirty(dst);
3378
3379         return ret;
3380 }
3381
3382 /*
3383  * helper function to insert a new root level in the tree.
3384  * A new node is allocated, and a single item is inserted to
3385  * point to the existing root
3386  *
3387  * returns zero on success or < 0 on failure.
3388  */
3389 static noinline int insert_new_root(struct btrfs_trans_handle *trans,
3390                            struct btrfs_root *root,
3391                            struct btrfs_path *path, int level)
3392 {
3393         struct btrfs_fs_info *fs_info = root->fs_info;
3394         u64 lower_gen;
3395         struct extent_buffer *lower;
3396         struct extent_buffer *c;
3397         struct extent_buffer *old;
3398         struct btrfs_disk_key lower_key;
3399         int ret;
3400
3401         BUG_ON(path->nodes[level]);
3402         BUG_ON(path->nodes[level-1] != root->node);
3403
3404         lower = path->nodes[level-1];
3405         if (level == 1)
3406                 btrfs_item_key(lower, &lower_key, 0);
3407         else
3408                 btrfs_node_key(lower, &lower_key, 0);
3409
3410         c = alloc_tree_block_no_bg_flush(trans, root, 0, &lower_key, level,
3411                                          root->node->start, 0);
3412         if (IS_ERR(c))
3413                 return PTR_ERR(c);
3414
3415         root_add_used(root, fs_info->nodesize);
3416
3417         btrfs_set_header_nritems(c, 1);
3418         btrfs_set_node_key(c, &lower_key, 0);
3419         btrfs_set_node_blockptr(c, 0, lower->start);
3420         lower_gen = btrfs_header_generation(lower);
3421         WARN_ON(lower_gen != trans->transid);
3422
3423         btrfs_set_node_ptr_generation(c, 0, lower_gen);
3424
3425         btrfs_mark_buffer_dirty(c);
3426
3427         old = root->node;
3428         ret = tree_mod_log_insert_root(root->node, c, 0);
3429         BUG_ON(ret < 0);
3430         rcu_assign_pointer(root->node, c);
3431
3432         /* the super has an extra ref to root->node */
3433         free_extent_buffer(old);
3434
3435         add_root_to_dirty_list(root);
3436         extent_buffer_get(c);
3437         path->nodes[level] = c;
3438         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
3439         path->slots[level] = 0;
3440         return 0;
3441 }
3442
3443 /*
3444  * worker function to insert a single pointer in a node.
3445  * the node should have enough room for the pointer already
3446  *
3447  * slot and level indicate where you want the key to go, and
3448  * blocknr is the block the key points to.
3449  */
3450 static void insert_ptr(struct btrfs_trans_handle *trans,
3451                        struct btrfs_path *path,
3452                        struct btrfs_disk_key *key, u64 bytenr,
3453                        int slot, int level)
3454 {
3455         struct extent_buffer *lower;
3456         int nritems;
3457         int ret;
3458
3459         BUG_ON(!path->nodes[level]);
3460         btrfs_assert_tree_locked(path->nodes[level]);
3461         lower = path->nodes[level];
3462         nritems = btrfs_header_nritems(lower);
3463         BUG_ON(slot > nritems);
3464         BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(trans->fs_info));
3465         if (slot != nritems) {
3466                 if (level) {
3467                         ret = tree_mod_log_insert_move(lower, slot + 1, slot,
3468                                         nritems - slot);
3469                         BUG_ON(ret < 0);
3470                 }
3471                 memmove_extent_buffer(lower,
3472                               btrfs_node_key_ptr_offset(slot + 1),
3473                               btrfs_node_key_ptr_offset(slot),
3474                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
3475         }
3476         if (level) {
3477                 ret = tree_mod_log_insert_key(lower, slot, MOD_LOG_KEY_ADD,
3478                                 GFP_NOFS);
3479                 BUG_ON(ret < 0);
3480         }
3481         btrfs_set_node_key(lower, key, slot);
3482         btrfs_set_node_blockptr(lower, slot, bytenr);
3483         WARN_ON(trans->transid == 0);
3484         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
3485         btrfs_set_header_nritems(lower, nritems + 1);
3486         btrfs_mark_buffer_dirty(lower);
3487 }
3488
3489 /*
3490  * split the node at the specified level in path in two.
3491  * The path is corrected to point to the appropriate node after the split
3492  *
3493  * Before splitting this tries to make some room in the node by pushing
3494  * left and right, if either one works, it returns right away.
3495  *
3496  * returns 0 on success and < 0 on failure
3497  */
3498 static noinline int split_node(struct btrfs_trans_handle *trans,
3499                                struct btrfs_root *root,
3500                                struct btrfs_path *path, int level)
3501 {
3502         struct btrfs_fs_info *fs_info = root->fs_info;
3503         struct extent_buffer *c;
3504         struct extent_buffer *split;
3505         struct btrfs_disk_key disk_key;
3506         int mid;
3507         int ret;
3508         u32 c_nritems;
3509
3510         c = path->nodes[level];
3511         WARN_ON(btrfs_header_generation(c) != trans->transid);
3512         if (c == root->node) {
3513                 /*
3514                  * trying to split the root, lets make a new one
3515                  *
3516                  * tree mod log: We don't log_removal old root in
3517                  * insert_new_root, because that root buffer will be kept as a
3518                  * normal node. We are going to log removal of half of the
3519                  * elements below with tree_mod_log_eb_copy. We're holding a
3520                  * tree lock on the buffer, which is why we cannot race with
3521                  * other tree_mod_log users.
3522                  */
3523                 ret = insert_new_root(trans, root, path, level + 1);
3524                 if (ret)
3525                         return ret;
3526         } else {
3527                 ret = push_nodes_for_insert(trans, root, path, level);
3528                 c = path->nodes[level];
3529                 if (!ret && btrfs_header_nritems(c) <
3530                     BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3)
3531                         return 0;
3532                 if (ret < 0)
3533                         return ret;
3534         }
3535
3536         c_nritems = btrfs_header_nritems(c);
3537         mid = (c_nritems + 1) / 2;
3538         btrfs_node_key(c, &disk_key, mid);
3539
3540         split = alloc_tree_block_no_bg_flush(trans, root, 0, &disk_key, level,
3541                                              c->start, 0);
3542         if (IS_ERR(split))
3543                 return PTR_ERR(split);
3544
3545         root_add_used(root, fs_info->nodesize);
3546         ASSERT(btrfs_header_level(c) == level);
3547
3548         ret = tree_mod_log_eb_copy(split, c, 0, mid, c_nritems - mid);
3549         if (ret) {
3550                 btrfs_abort_transaction(trans, ret);
3551                 return ret;
3552         }
3553         copy_extent_buffer(split, c,
3554                            btrfs_node_key_ptr_offset(0),
3555                            btrfs_node_key_ptr_offset(mid),
3556                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
3557         btrfs_set_header_nritems(split, c_nritems - mid);
3558         btrfs_set_header_nritems(c, mid);
3559         ret = 0;
3560
3561         btrfs_mark_buffer_dirty(c);
3562         btrfs_mark_buffer_dirty(split);
3563
3564         insert_ptr(trans, path, &disk_key, split->start,
3565                    path->slots[level + 1] + 1, level + 1);
3566
3567         if (path->slots[level] >= mid) {
3568                 path->slots[level] -= mid;
3569                 btrfs_tree_unlock(c);
3570                 free_extent_buffer(c);
3571                 path->nodes[level] = split;
3572                 path->slots[level + 1] += 1;
3573         } else {
3574                 btrfs_tree_unlock(split);
3575                 free_extent_buffer(split);
3576         }
3577         return ret;
3578 }
3579
3580 /*
3581  * how many bytes are required to store the items in a leaf.  start
3582  * and nr indicate which items in the leaf to check.  This totals up the
3583  * space used both by the item structs and the item data
3584  */
3585 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
3586 {
3587         struct btrfs_item *start_item;
3588         struct btrfs_item *end_item;
3589         struct btrfs_map_token token;
3590         int data_len;
3591         int nritems = btrfs_header_nritems(l);
3592         int end = min(nritems, start + nr) - 1;
3593
3594         if (!nr)
3595                 return 0;
3596         btrfs_init_map_token(&token, l);
3597         start_item = btrfs_item_nr(start);
3598         end_item = btrfs_item_nr(end);
3599         data_len = btrfs_token_item_offset(l, start_item, &token) +
3600                 btrfs_token_item_size(l, start_item, &token);
3601         data_len = data_len - btrfs_token_item_offset(l, end_item, &token);
3602         data_len += sizeof(struct btrfs_item) * nr;
3603         WARN_ON(data_len < 0);
3604         return data_len;
3605 }
3606
3607 /*
3608  * The space between the end of the leaf items and
3609  * the start of the leaf data.  IOW, how much room
3610  * the leaf has left for both items and data
3611  */
3612 noinline int btrfs_leaf_free_space(struct extent_buffer *leaf)
3613 {
3614         struct btrfs_fs_info *fs_info = leaf->fs_info;
3615         int nritems = btrfs_header_nritems(leaf);
3616         int ret;
3617
3618         ret = BTRFS_LEAF_DATA_SIZE(fs_info) - leaf_space_used(leaf, 0, nritems);
3619         if (ret < 0) {
3620                 btrfs_crit(fs_info,
3621                            "leaf free space ret %d, leaf data size %lu, used %d nritems %d",
3622                            ret,
3623                            (unsigned long) BTRFS_LEAF_DATA_SIZE(fs_info),
3624                            leaf_space_used(leaf, 0, nritems), nritems);
3625         }
3626         return ret;
3627 }
3628
3629 /*
3630  * min slot controls the lowest index we're willing to push to the
3631  * right.  We'll push up to and including min_slot, but no lower
3632  */
3633 static noinline int __push_leaf_right(struct btrfs_path *path,
3634                                       int data_size, int empty,
3635                                       struct extent_buffer *right,
3636                                       int free_space, u32 left_nritems,
3637                                       u32 min_slot)
3638 {
3639         struct btrfs_fs_info *fs_info = right->fs_info;
3640         struct extent_buffer *left = path->nodes[0];
3641         struct extent_buffer *upper = path->nodes[1];
3642         struct btrfs_map_token token;
3643         struct btrfs_disk_key disk_key;
3644         int slot;
3645         u32 i;
3646         int push_space = 0;
3647         int push_items = 0;
3648         struct btrfs_item *item;
3649         u32 nr;
3650         u32 right_nritems;
3651         u32 data_end;
3652         u32 this_item_size;
3653
3654         if (empty)
3655                 nr = 0;
3656         else
3657                 nr = max_t(u32, 1, min_slot);
3658
3659         if (path->slots[0] >= left_nritems)
3660                 push_space += data_size;
3661
3662         slot = path->slots[1];
3663         i = left_nritems - 1;
3664         while (i >= nr) {
3665                 item = btrfs_item_nr(i);
3666
3667                 if (!empty && push_items > 0) {
3668                         if (path->slots[0] > i)
3669                                 break;
3670                         if (path->slots[0] == i) {
3671                                 int space = btrfs_leaf_free_space(left);
3672
3673                                 if (space + push_space * 2 > free_space)
3674                                         break;
3675                         }
3676                 }
3677
3678                 if (path->slots[0] == i)
3679                         push_space += data_size;
3680
3681                 this_item_size = btrfs_item_size(left, item);
3682                 if (this_item_size + sizeof(*item) + push_space > free_space)
3683                         break;
3684
3685                 push_items++;
3686                 push_space += this_item_size + sizeof(*item);
3687                 if (i == 0)
3688                         break;
3689                 i--;
3690         }
3691
3692         if (push_items == 0)
3693                 goto out_unlock;
3694
3695         WARN_ON(!empty && push_items == left_nritems);
3696
3697         /* push left to right */
3698         right_nritems = btrfs_header_nritems(right);
3699
3700         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
3701         push_space -= leaf_data_end(left);
3702
3703         /* make room in the right data area */
3704         data_end = leaf_data_end(right);
3705         memmove_extent_buffer(right,
3706                               BTRFS_LEAF_DATA_OFFSET + data_end - push_space,
3707                               BTRFS_LEAF_DATA_OFFSET + data_end,
3708                               BTRFS_LEAF_DATA_SIZE(fs_info) - data_end);
3709
3710         /* copy from the left data area */
3711         copy_extent_buffer(right, left, BTRFS_LEAF_DATA_OFFSET +
3712                      BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3713                      BTRFS_LEAF_DATA_OFFSET + leaf_data_end(left),
3714                      push_space);
3715
3716         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
3717                               btrfs_item_nr_offset(0),
3718                               right_nritems * sizeof(struct btrfs_item));
3719
3720         /* copy the items from left to right */
3721         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
3722                    btrfs_item_nr_offset(left_nritems - push_items),
3723                    push_items * sizeof(struct btrfs_item));
3724
3725         /* update the item pointers */
3726         btrfs_init_map_token(&token, right);
3727         right_nritems += push_items;
3728         btrfs_set_header_nritems(right, right_nritems);
3729         push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3730         for (i = 0; i < right_nritems; i++) {
3731                 item = btrfs_item_nr(i);
3732                 push_space -= btrfs_token_item_size(right, item, &token);
3733                 btrfs_set_token_item_offset(right, item, push_space, &token);
3734         }
3735
3736         left_nritems -= push_items;
3737         btrfs_set_header_nritems(left, left_nritems);
3738
3739         if (left_nritems)
3740                 btrfs_mark_buffer_dirty(left);
3741         else
3742                 btrfs_clean_tree_block(left);
3743
3744         btrfs_mark_buffer_dirty(right);
3745
3746         btrfs_item_key(right, &disk_key, 0);
3747         btrfs_set_node_key(upper, &disk_key, slot + 1);
3748         btrfs_mark_buffer_dirty(upper);
3749
3750         /* then fixup the leaf pointer in the path */
3751         if (path->slots[0] >= left_nritems) {
3752                 path->slots[0] -= left_nritems;
3753                 if (btrfs_header_nritems(path->nodes[0]) == 0)
3754                         btrfs_clean_tree_block(path->nodes[0]);
3755                 btrfs_tree_unlock(path->nodes[0]);
3756                 free_extent_buffer(path->nodes[0]);
3757                 path->nodes[0] = right;
3758                 path->slots[1] += 1;
3759         } else {
3760                 btrfs_tree_unlock(right);
3761                 free_extent_buffer(right);
3762         }
3763         return 0;
3764
3765 out_unlock:
3766         btrfs_tree_unlock(right);
3767         free_extent_buffer(right);
3768         return 1;
3769 }
3770
3771 /*
3772  * push some data in the path leaf to the right, trying to free up at
3773  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
3774  *
3775  * returns 1 if the push failed because the other node didn't have enough
3776  * room, 0 if everything worked out and < 0 if there were major errors.
3777  *
3778  * this will push starting from min_slot to the end of the leaf.  It won't
3779  * push any slot lower than min_slot
3780  */
3781 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
3782                            *root, struct btrfs_path *path,
3783                            int min_data_size, int data_size,
3784                            int empty, u32 min_slot)
3785 {
3786         struct extent_buffer *left = path->nodes[0];
3787         struct extent_buffer *right;
3788         struct extent_buffer *upper;
3789         int slot;
3790         int free_space;
3791         u32 left_nritems;
3792         int ret;
3793
3794         if (!path->nodes[1])
3795                 return 1;
3796
3797         slot = path->slots[1];
3798         upper = path->nodes[1];
3799         if (slot >= btrfs_header_nritems(upper) - 1)
3800                 return 1;
3801
3802         btrfs_assert_tree_locked(path->nodes[1]);
3803
3804         right = btrfs_read_node_slot(upper, slot + 1);
3805         /*
3806          * slot + 1 is not valid or we fail to read the right node,
3807          * no big deal, just return.
3808          */
3809         if (IS_ERR(right))
3810                 return 1;
3811
3812         btrfs_tree_lock(right);
3813         btrfs_set_lock_blocking_write(right);
3814
3815         free_space = btrfs_leaf_free_space(right);
3816         if (free_space < data_size)
3817                 goto out_unlock;
3818
3819         /* cow and double check */
3820         ret = btrfs_cow_block(trans, root, right, upper,
3821                               slot + 1, &right);
3822         if (ret)
3823                 goto out_unlock;
3824
3825         free_space = btrfs_leaf_free_space(right);
3826         if (free_space < data_size)
3827                 goto out_unlock;
3828
3829         left_nritems = btrfs_header_nritems(left);
3830         if (left_nritems == 0)
3831                 goto out_unlock;
3832
3833         if (path->slots[0] == left_nritems && !empty) {
3834                 /* Key greater than all keys in the leaf, right neighbor has
3835                  * enough room for it and we're not emptying our leaf to delete
3836                  * it, therefore use right neighbor to insert the new item and
3837                  * no need to touch/dirty our left leaf. */
3838                 btrfs_tree_unlock(left);
3839                 free_extent_buffer(left);
3840                 path->nodes[0] = right;
3841                 path->slots[0] = 0;
3842                 path->slots[1]++;
3843                 return 0;
3844         }
3845
3846         return __push_leaf_right(path, min_data_size, empty,
3847                                 right, free_space, left_nritems, min_slot);
3848 out_unlock:
3849         btrfs_tree_unlock(right);
3850         free_extent_buffer(right);
3851         return 1;
3852 }
3853
3854 /*
3855  * push some data in the path leaf to the left, trying to free up at
3856  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
3857  *
3858  * max_slot can put a limit on how far into the leaf we'll push items.  The
3859  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us do all the
3860  * items
3861  */
3862 static noinline int __push_leaf_left(struct btrfs_path *path, int data_size,
3863                                      int empty, struct extent_buffer *left,
3864                                      int free_space, u32 right_nritems,
3865                                      u32 max_slot)
3866 {
3867         struct btrfs_fs_info *fs_info = left->fs_info;
3868         struct btrfs_disk_key disk_key;
3869         struct extent_buffer *right = path->nodes[0];
3870         int i;
3871         int push_space = 0;
3872         int push_items = 0;
3873         struct btrfs_item *item;
3874         u32 old_left_nritems;
3875         u32 nr;
3876         int ret = 0;
3877         u32 this_item_size;
3878         u32 old_left_item_size;
3879         struct btrfs_map_token token;
3880
3881         if (empty)
3882                 nr = min(right_nritems, max_slot);
3883         else
3884                 nr = min(right_nritems - 1, max_slot);
3885
3886         for (i = 0; i < nr; i++) {
3887                 item = btrfs_item_nr(i);
3888
3889                 if (!empty && push_items > 0) {
3890                         if (path->slots[0] < i)
3891                                 break;
3892                         if (path->slots[0] == i) {
3893                                 int space = btrfs_leaf_free_space(right);
3894
3895                                 if (space + push_space * 2 > free_space)
3896                                         break;
3897                         }
3898                 }
3899
3900                 if (path->slots[0] == i)
3901                         push_space += data_size;
3902
3903                 this_item_size = btrfs_item_size(right, item);
3904                 if (this_item_size + sizeof(*item) + push_space > free_space)
3905                         break;
3906
3907                 push_items++;
3908                 push_space += this_item_size + sizeof(*item);
3909         }
3910
3911         if (push_items == 0) {
3912                 ret = 1;
3913                 goto out;
3914         }
3915         WARN_ON(!empty && push_items == btrfs_header_nritems(right));
3916
3917         /* push data from right to left */
3918         copy_extent_buffer(left, right,
3919                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
3920                            btrfs_item_nr_offset(0),
3921                            push_items * sizeof(struct btrfs_item));
3922
3923         push_space = BTRFS_LEAF_DATA_SIZE(fs_info) -
3924                      btrfs_item_offset_nr(right, push_items - 1);
3925
3926         copy_extent_buffer(left, right, BTRFS_LEAF_DATA_OFFSET +
3927                      leaf_data_end(left) - push_space,
3928                      BTRFS_LEAF_DATA_OFFSET +
3929                      btrfs_item_offset_nr(right, push_items - 1),
3930                      push_space);
3931         old_left_nritems = btrfs_header_nritems(left);
3932         BUG_ON(old_left_nritems <= 0);
3933
3934         btrfs_init_map_token(&token, left);
3935         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
3936         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
3937                 u32 ioff;
3938
3939                 item = btrfs_item_nr(i);
3940
3941                 ioff = btrfs_token_item_offset(left, item, &token);
3942                 btrfs_set_token_item_offset(left, item,
3943                       ioff - (BTRFS_LEAF_DATA_SIZE(fs_info) - old_left_item_size),
3944                       &token);
3945         }
3946         btrfs_set_header_nritems(left, old_left_nritems + push_items);
3947
3948         /* fixup right node */
3949         if (push_items > right_nritems)
3950                 WARN(1, KERN_CRIT "push items %d nr %u\n", push_items,
3951                        right_nritems);
3952
3953         if (push_items < right_nritems) {
3954                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
3955                                                   leaf_data_end(right);
3956                 memmove_extent_buffer(right, BTRFS_LEAF_DATA_OFFSET +
3957                                       BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3958                                       BTRFS_LEAF_DATA_OFFSET +
3959                                       leaf_data_end(right), push_space);
3960
3961                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
3962                               btrfs_item_nr_offset(push_items),
3963                              (btrfs_header_nritems(right) - push_items) *
3964                              sizeof(struct btrfs_item));
3965         }
3966
3967         btrfs_init_map_token(&token, right);
3968         right_nritems -= push_items;
3969         btrfs_set_header_nritems(right, right_nritems);
3970         push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3971         for (i = 0; i < right_nritems; i++) {
3972                 item = btrfs_item_nr(i);
3973
3974                 push_space = push_space - btrfs_token_item_size(right,
3975                                                                 item, &token);
3976                 btrfs_set_token_item_offset(right, item, push_space, &token);
3977         }
3978
3979         btrfs_mark_buffer_dirty(left);
3980         if (right_nritems)
3981                 btrfs_mark_buffer_dirty(right);
3982         else
3983                 btrfs_clean_tree_block(right);
3984
3985         btrfs_item_key(right, &disk_key, 0);
3986         fixup_low_keys(path, &disk_key, 1);
3987
3988         /* then fixup the leaf pointer in the path */
3989         if (path->slots[0] < push_items) {
3990                 path->slots[0] += old_left_nritems;
3991                 btrfs_tree_unlock(path->nodes[0]);
3992                 free_extent_buffer(path->nodes[0]);
3993                 path->nodes[0] = left;
3994                 path->slots[1] -= 1;
3995         } else {
3996                 btrfs_tree_unlock(left);
3997                 free_extent_buffer(left);
3998                 path->slots[0] -= push_items;
3999         }
4000         BUG_ON(path->slots[0] < 0);
4001         return ret;
4002 out:
4003         btrfs_tree_unlock(left);
4004         free_extent_buffer(left);
4005         return ret;
4006 }
4007
4008 /*
4009  * push some data in the path leaf to the left, trying to free up at
4010  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
4011  *
4012  * max_slot can put a limit on how far into the leaf we'll push items.  The
4013  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us push all the
4014  * items
4015  */
4016 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
4017                           *root, struct btrfs_path *path, int min_data_size,
4018                           int data_size, int empty, u32 max_slot)
4019 {
4020         struct extent_buffer *right = path->nodes[0];
4021         struct extent_buffer *left;
4022         int slot;
4023         int free_space;
4024         u32 right_nritems;
4025         int ret = 0;
4026
4027         slot = path->slots[1];
4028         if (slot == 0)
4029                 return 1;
4030         if (!path->nodes[1])
4031                 return 1;
4032
4033         right_nritems = btrfs_header_nritems(right);
4034         if (right_nritems == 0)
4035                 return 1;
4036
4037         btrfs_assert_tree_locked(path->nodes[1]);
4038
4039         left = btrfs_read_node_slot(path->nodes[1], slot - 1);
4040         /*
4041          * slot - 1 is not valid or we fail to read the left node,
4042          * no big deal, just return.
4043          */
4044         if (IS_ERR(left))
4045                 return 1;
4046
4047         btrfs_tree_lock(left);
4048         btrfs_set_lock_blocking_write(left);
4049
4050         free_space = btrfs_leaf_free_space(left);
4051         if (free_space < data_size) {
4052                 ret = 1;
4053                 goto out;
4054         }
4055
4056         /* cow and double check */
4057         ret = btrfs_cow_block(trans, root, left,
4058                               path->nodes[1], slot - 1, &left);
4059         if (ret) {
4060                 /* we hit -ENOSPC, but it isn't fatal here */
4061                 if (ret == -ENOSPC)
4062                         ret = 1;
4063                 goto out;
4064         }
4065
4066         free_space = btrfs_leaf_free_space(left);
4067         if (free_space < data_size) {
4068                 ret = 1;
4069                 goto out;
4070         }
4071
4072         return __push_leaf_left(path, min_data_size,
4073                                empty, left, free_space, right_nritems,
4074                                max_slot);
4075 out:
4076         btrfs_tree_unlock(left);
4077         free_extent_buffer(left);
4078         return ret;
4079 }
4080
4081 /*
4082  * split the path's leaf in two, making sure there is at least data_size
4083  * available for the resulting leaf level of the path.
4084  */
4085 static noinline void copy_for_split(struct btrfs_trans_handle *trans,
4086                                     struct btrfs_path *path,
4087                                     struct extent_buffer *l,
4088                                     struct extent_buffer *right,
4089                                     int slot, int mid, int nritems)
4090 {
4091         struct btrfs_fs_info *fs_info = trans->fs_info;
4092         int data_copy_size;
4093         int rt_data_off;
4094         int i;
4095         struct btrfs_disk_key disk_key;
4096         struct btrfs_map_token token;
4097
4098         nritems = nritems - mid;
4099         btrfs_set_header_nritems(right, nritems);
4100         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(l);
4101
4102         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
4103                            btrfs_item_nr_offset(mid),
4104                            nritems * sizeof(struct btrfs_item));
4105
4106         copy_extent_buffer(right, l,
4107                      BTRFS_LEAF_DATA_OFFSET + BTRFS_LEAF_DATA_SIZE(fs_info) -
4108                      data_copy_size, BTRFS_LEAF_DATA_OFFSET +
4109                      leaf_data_end(l), data_copy_size);
4110
4111         rt_data_off = BTRFS_LEAF_DATA_SIZE(fs_info) - btrfs_item_end_nr(l, mid);
4112
4113         btrfs_init_map_token(&token, right);
4114         for (i = 0; i < nritems; i++) {
4115                 struct btrfs_item *item = btrfs_item_nr(i);
4116                 u32 ioff;
4117
4118                 ioff = btrfs_token_item_offset(right, item, &token);
4119                 btrfs_set_token_item_offset(right, item,
4120                                             ioff + rt_data_off, &token);
4121         }
4122
4123         btrfs_set_header_nritems(l, mid);
4124         btrfs_item_key(right, &disk_key, 0);
4125         insert_ptr(trans, path, &disk_key, right->start, path->slots[1] + 1, 1);
4126
4127         btrfs_mark_buffer_dirty(right);
4128         btrfs_mark_buffer_dirty(l);
4129         BUG_ON(path->slots[0] != slot);
4130
4131         if (mid <= slot) {
4132                 btrfs_tree_unlock(path->nodes[0]);
4133                 free_extent_buffer(path->nodes[0]);
4134                 path->nodes[0] = right;
4135                 path->slots[0] -= mid;
4136                 path->slots[1] += 1;
4137         } else {
4138                 btrfs_tree_unlock(right);
4139                 free_extent_buffer(right);
4140         }
4141
4142         BUG_ON(path->slots[0] < 0);
4143 }
4144
4145 /*
4146  * double splits happen when we need to insert a big item in the middle
4147  * of a leaf.  A double split can leave us with 3 mostly empty leaves:
4148  * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
4149  *          A                 B                 C
4150  *
4151  * We avoid this by trying to push the items on either side of our target
4152  * into the adjacent leaves.  If all goes well we can avoid the double split
4153  * completely.
4154  */
4155 static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
4156                                           struct btrfs_root *root,
4157                                           struct btrfs_path *path,
4158                                           int data_size)
4159 {
4160         int ret;
4161         int progress = 0;
4162         int slot;
4163         u32 nritems;
4164         int space_needed = data_size;
4165
4166         slot = path->slots[0];
4167         if (slot < btrfs_header_nritems(path->nodes[0]))
4168                 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
4169
4170         /*
4171          * try to push all the items after our slot into the
4172          * right leaf
4173          */
4174         ret = push_leaf_right(trans, root, path, 1, space_needed, 0, slot);
4175         if (ret < 0)
4176                 return ret;
4177
4178         if (ret == 0)
4179                 progress++;
4180
4181         nritems = btrfs_header_nritems(path->nodes[0]);
4182         /*
4183          * our goal is to get our slot at the start or end of a leaf.  If
4184          * we've done so we're done
4185          */
4186         if (path->slots[0] == 0 || path->slots[0] == nritems)
4187                 return 0;
4188
4189         if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
4190                 return 0;
4191
4192         /* try to push all the items before our slot into the next leaf */
4193         slot = path->slots[0];
4194         space_needed = data_size;
4195         if (slot > 0)
4196                 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
4197         ret = push_leaf_left(trans, root, path, 1, space_needed, 0, slot);
4198         if (ret < 0)
4199                 return ret;
4200
4201         if (ret == 0)
4202                 progress++;
4203
4204         if (progress)
4205                 return 0;
4206         return 1;
4207 }
4208
4209 /*
4210  * split the path's leaf in two, making sure there is at least data_size
4211  * available for the resulting leaf level of the path.
4212  *
4213  * returns 0 if all went well and < 0 on failure.
4214  */
4215 static noinline int split_leaf(struct btrfs_trans_handle *trans,
4216                                struct btrfs_root *root,
4217                                const struct btrfs_key *ins_key,
4218                                struct btrfs_path *path, int data_size,
4219                                int extend)
4220 {
4221         struct btrfs_disk_key disk_key;
4222         struct extent_buffer *l;
4223         u32 nritems;
4224         int mid;
4225         int slot;
4226         struct extent_buffer *right;
4227         struct btrfs_fs_info *fs_info = root->fs_info;
4228         int ret = 0;
4229         int wret;
4230         int split;
4231         int num_doubles = 0;
4232         int tried_avoid_double = 0;
4233
4234         l = path->nodes[0];
4235         slot = path->slots[0];
4236         if (extend && data_size + btrfs_item_size_nr(l, slot) +
4237             sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(fs_info))
4238                 return -EOVERFLOW;
4239
4240         /* first try to make some room by pushing left and right */
4241         if (data_size && path->nodes[1]) {
4242                 int space_needed = data_size;
4243
4244                 if (slot < btrfs_header_nritems(l))
4245                         space_needed -= btrfs_leaf_free_space(l);
4246
4247                 wret = push_leaf_right(trans, root, path, space_needed,
4248                                        space_needed, 0, 0);
4249                 if (wret < 0)
4250                         return wret;
4251                 if (wret) {
4252                         space_needed = data_size;
4253                         if (slot > 0)
4254                                 space_needed -= btrfs_leaf_free_space(l);
4255                         wret = push_leaf_left(trans, root, path, space_needed,
4256                                               space_needed, 0, (u32)-1);
4257                         if (wret < 0)
4258                                 return wret;
4259                 }
4260                 l = path->nodes[0];
4261
4262                 /* did the pushes work? */
4263                 if (btrfs_leaf_free_space(l) >= data_size)
4264                         return 0;
4265         }
4266
4267         if (!path->nodes[1]) {
4268                 ret = insert_new_root(trans, root, path, 1);
4269                 if (ret)
4270                         return ret;
4271         }
4272 again:
4273         split = 1;
4274         l = path->nodes[0];
4275         slot = path->slots[0];
4276         nritems = btrfs_header_nritems(l);
4277         mid = (nritems + 1) / 2;
4278
4279         if (mid <= slot) {
4280                 if (nritems == 1 ||
4281                     leaf_space_used(l, mid, nritems - mid) + data_size >
4282                         BTRFS_LEAF_DATA_SIZE(fs_info)) {
4283                         if (slot >= nritems) {
4284                                 split = 0;
4285                         } else {
4286                                 mid = slot;
4287                                 if (mid != nritems &&
4288                                     leaf_space_used(l, mid, nritems - mid) +
4289                                     data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
4290                                         if (data_size && !tried_avoid_double)
4291                                                 goto push_for_double;
4292                                         split = 2;
4293                                 }
4294                         }
4295                 }
4296         } else {
4297                 if (leaf_space_used(l, 0, mid) + data_size >
4298                         BTRFS_LEAF_DATA_SIZE(fs_info)) {
4299                         if (!extend && data_size && slot == 0) {
4300                                 split = 0;
4301                         } else if ((extend || !data_size) && slot == 0) {
4302                                 mid = 1;
4303                         } else {
4304                                 mid = slot;
4305                                 if (mid != nritems &&
4306                                     leaf_space_used(l, mid, nritems - mid) +
4307                                     data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
4308                                         if (data_size && !tried_avoid_double)
4309                                                 goto push_for_double;
4310                                         split = 2;
4311                                 }
4312                         }
4313                 }
4314         }
4315
4316         if (split == 0)
4317                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
4318         else
4319                 btrfs_item_key(l, &disk_key, mid);
4320
4321         right = alloc_tree_block_no_bg_flush(trans, root, 0, &disk_key, 0,
4322                                              l->start, 0);
4323         if (IS_ERR(right))
4324                 return PTR_ERR(right);
4325
4326         root_add_used(root, fs_info->nodesize);
4327
4328         if (split == 0) {
4329                 if (mid <= slot) {
4330                         btrfs_set_header_nritems(right, 0);
4331                         insert_ptr(trans, path, &disk_key,
4332                                    right->start, path->slots[1] + 1, 1);
4333                         btrfs_tree_unlock(path->nodes[0]);
4334                         free_extent_buffer(path->nodes[0]);
4335                         path->nodes[0] = right;
4336                         path->slots[0] = 0;
4337                         path->slots[1] += 1;
4338                 } else {
4339                         btrfs_set_header_nritems(right, 0);
4340                         insert_ptr(trans, path, &disk_key,
4341                                    right->start, path->slots[1], 1);
4342                         btrfs_tree_unlock(path->nodes[0]);
4343                         free_extent_buffer(path->nodes[0]);
4344                         path->nodes[0] = right;
4345                         path->slots[0] = 0;
4346                         if (path->slots[1] == 0)
4347                                 fixup_low_keys(path, &disk_key, 1);
4348                 }
4349                 /*
4350                  * We create a new leaf 'right' for the required ins_len and
4351                  * we'll do btrfs_mark_buffer_dirty() on this leaf after copying
4352                  * the content of ins_len to 'right'.
4353                  */
4354                 return ret;
4355         }
4356
4357         copy_for_split(trans, path, l, right, slot, mid, nritems);
4358
4359         if (split == 2) {
4360                 BUG_ON(num_doubles != 0);
4361                 num_doubles++;
4362                 goto again;
4363         }
4364
4365         return 0;
4366
4367 push_for_double:
4368         push_for_double_split(trans, root, path, data_size);
4369         tried_avoid_double = 1;
4370         if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
4371                 return 0;
4372         goto again;
4373 }
4374
4375 static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
4376                                          struct btrfs_root *root,
4377                                          struct btrfs_path *path, int ins_len)
4378 {
4379         struct btrfs_key key;
4380         struct extent_buffer *leaf;
4381         struct btrfs_file_extent_item *fi;
4382         u64 extent_len = 0;
4383         u32 item_size;
4384         int ret;
4385
4386         leaf = path->nodes[0];
4387         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4388
4389         BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
4390                key.type != BTRFS_EXTENT_CSUM_KEY);
4391
4392         if (btrfs_leaf_free_space(leaf) >= ins_len)
4393                 return 0;
4394
4395         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4396         if (key.type == BTRFS_EXTENT_DATA_KEY) {
4397                 fi = btrfs_item_ptr(leaf, path->slots[0],
4398                                     struct btrfs_file_extent_item);
4399                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
4400         }
4401         btrfs_release_path(path);
4402
4403         path->keep_locks = 1;
4404         path->search_for_split = 1;
4405         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
4406         path->search_for_split = 0;
4407         if (ret > 0)
4408                 ret = -EAGAIN;
4409         if (ret < 0)
4410                 goto err;
4411
4412         ret = -EAGAIN;
4413         leaf = path->nodes[0];
4414         /* if our item isn't there, return now */
4415         if (item_size != btrfs_item_size_nr(leaf, path->slots[0]))
4416                 goto err;
4417
4418         /* the leaf has  changed, it now has room.  return now */
4419         if (btrfs_leaf_free_space(path->nodes[0]) >= ins_len)
4420                 goto err;
4421
4422         if (key.type == BTRFS_EXTENT_DATA_KEY) {
4423                 fi = btrfs_item_ptr(leaf, path->slots[0],
4424                                     struct btrfs_file_extent_item);
4425                 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
4426                         goto err;
4427         }
4428
4429         btrfs_set_path_blocking(path);
4430         ret = split_leaf(trans, root, &key, path, ins_len, 1);
4431         if (ret)
4432                 goto err;
4433
4434         path->keep_locks = 0;
4435         btrfs_unlock_up_safe(path, 1);
4436         return 0;
4437 err:
4438         path->keep_locks = 0;
4439         return ret;
4440 }
4441
4442 static noinline int split_item(struct btrfs_path *path,
4443                                const struct btrfs_key *new_key,
4444                                unsigned long split_offset)
4445 {
4446         struct extent_buffer *leaf;
4447         struct btrfs_item *item;
4448         struct btrfs_item *new_item;
4449         int slot;
4450         char *buf;
4451         u32 nritems;
4452         u32 item_size;
4453         u32 orig_offset;
4454         struct btrfs_disk_key disk_key;
4455
4456         leaf = path->nodes[0];
4457         BUG_ON(btrfs_leaf_free_space(leaf) < sizeof(struct btrfs_item));
4458
4459         btrfs_set_path_blocking(path);
4460
4461         item = btrfs_item_nr(path->slots[0]);
4462         orig_offset = btrfs_item_offset(leaf, item);
4463         item_size = btrfs_item_size(leaf, item);
4464
4465         buf = kmalloc(item_size, GFP_NOFS);
4466         if (!buf)
4467                 return -ENOMEM;
4468
4469         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
4470                             path->slots[0]), item_size);
4471
4472         slot = path->slots[0] + 1;
4473         nritems = btrfs_header_nritems(leaf);
4474         if (slot != nritems) {
4475                 /* shift the items */
4476                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
4477                                 btrfs_item_nr_offset(slot),
4478                                 (nritems - slot) * sizeof(struct btrfs_item));
4479         }
4480
4481         btrfs_cpu_key_to_disk(&disk_key, new_key);
4482         btrfs_set_item_key(leaf, &disk_key, slot);
4483
4484         new_item = btrfs_item_nr(slot);
4485
4486         btrfs_set_item_offset(leaf, new_item, orig_offset);
4487         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
4488
4489         btrfs_set_item_offset(leaf, item,
4490                               orig_offset + item_size - split_offset);
4491         btrfs_set_item_size(leaf, item, split_offset);
4492
4493         btrfs_set_header_nritems(leaf, nritems + 1);
4494
4495         /* write the data for the start of the original item */
4496         write_extent_buffer(leaf, buf,
4497                             btrfs_item_ptr_offset(leaf, path->slots[0]),
4498                             split_offset);
4499
4500         /* write the data for the new item */
4501         write_extent_buffer(leaf, buf + split_offset,
4502                             btrfs_item_ptr_offset(leaf, slot),
4503                             item_size - split_offset);
4504         btrfs_mark_buffer_dirty(leaf);
4505
4506         BUG_ON(btrfs_leaf_free_space(leaf) < 0);
4507         kfree(buf);
4508         return 0;
4509 }
4510
4511 /*
4512  * This function splits a single item into two items,
4513  * giving 'new_key' to the new item and splitting the
4514  * old one at split_offset (from the start of the item).
4515  *
4516  * The path may be released by this operation.  After
4517  * the split, the path is pointing to the old item.  The
4518  * new item is going to be in the same node as the old one.
4519  *
4520  * Note, the item being split must be smaller enough to live alone on
4521  * a tree block with room for one extra struct btrfs_item
4522  *
4523  * This allows us to split the item in place, keeping a lock on the
4524  * leaf the entire time.
4525  */
4526 int btrfs_split_item(struct btrfs_trans_handle *trans,
4527                      struct btrfs_root *root,
4528                      struct btrfs_path *path,
4529                      const struct btrfs_key *new_key,
4530                      unsigned long split_offset)
4531 {
4532         int ret;
4533         ret = setup_leaf_for_split(trans, root, path,
4534                                    sizeof(struct btrfs_item));
4535         if (ret)
4536                 return ret;
4537
4538         ret = split_item(path, new_key, split_offset);
4539         return ret;
4540 }
4541
4542 /*
4543  * This function duplicate a item, giving 'new_key' to the new item.
4544  * It guarantees both items live in the same tree leaf and the new item
4545  * is contiguous with the original item.
4546  *
4547  * This allows us to split file extent in place, keeping a lock on the
4548  * leaf the entire time.
4549  */
4550 int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
4551                          struct btrfs_root *root,
4552                          struct btrfs_path *path,
4553                          const struct btrfs_key *new_key)
4554 {
4555         struct extent_buffer *leaf;
4556         int ret;
4557         u32 item_size;
4558
4559         leaf = path->nodes[0];
4560         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4561         ret = setup_leaf_for_split(trans, root, path,
4562                                    item_size + sizeof(struct btrfs_item));
4563         if (ret)
4564                 return ret;
4565
4566         path->slots[0]++;
4567         setup_items_for_insert(root, path, new_key, &item_size,
4568                                item_size, item_size +
4569                                sizeof(struct btrfs_item), 1);
4570         leaf = path->nodes[0];
4571         memcpy_extent_buffer(leaf,
4572                              btrfs_item_ptr_offset(leaf, path->slots[0]),
4573                              btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
4574                              item_size);
4575         return 0;
4576 }
4577
4578 /*
4579  * make the item pointed to by the path smaller.  new_size indicates
4580  * how small to make it, and from_end tells us if we just chop bytes
4581  * off the end of the item or if we shift the item to chop bytes off
4582  * the front.
4583  */
4584 void btrfs_truncate_item(struct btrfs_path *path, u32 new_size, int from_end)
4585 {
4586         int slot;
4587         struct extent_buffer *leaf;
4588         struct btrfs_item *item;
4589         u32 nritems;
4590         unsigned int data_end;
4591         unsigned int old_data_start;
4592         unsigned int old_size;
4593         unsigned int size_diff;
4594         int i;
4595         struct btrfs_map_token token;
4596
4597         leaf = path->nodes[0];
4598         slot = path->slots[0];
4599
4600         old_size = btrfs_item_size_nr(leaf, slot);
4601         if (old_size == new_size)
4602                 return;
4603
4604         nritems = btrfs_header_nritems(leaf);
4605         data_end = leaf_data_end(leaf);
4606
4607         old_data_start = btrfs_item_offset_nr(leaf, slot);
4608
4609         size_diff = old_size - new_size;
4610
4611         BUG_ON(slot < 0);
4612         BUG_ON(slot >= nritems);
4613
4614         /*
4615          * item0..itemN ... dataN.offset..dataN.size .. data0.size
4616          */
4617         /* first correct the data pointers */
4618         btrfs_init_map_token(&token, leaf);
4619         for (i = slot; i < nritems; i++) {
4620                 u32 ioff;
4621                 item = btrfs_item_nr(i);
4622
4623                 ioff = btrfs_token_item_offset(leaf, item, &token);
4624                 btrfs_set_token_item_offset(leaf, item,
4625                                             ioff + size_diff, &token);
4626         }
4627
4628         /* shift the data */
4629         if (from_end) {
4630                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4631                               data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
4632                               data_end, old_data_start + new_size - data_end);
4633         } else {
4634                 struct btrfs_disk_key disk_key;
4635                 u64 offset;
4636
4637                 btrfs_item_key(leaf, &disk_key, slot);
4638
4639                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
4640                         unsigned long ptr;
4641                         struct btrfs_file_extent_item *fi;
4642
4643                         fi = btrfs_item_ptr(leaf, slot,
4644                                             struct btrfs_file_extent_item);
4645                         fi = (struct btrfs_file_extent_item *)(
4646                              (unsigned long)fi - size_diff);
4647
4648                         if (btrfs_file_extent_type(leaf, fi) ==
4649                             BTRFS_FILE_EXTENT_INLINE) {
4650                                 ptr = btrfs_item_ptr_offset(leaf, slot);
4651                                 memmove_extent_buffer(leaf, ptr,
4652                                       (unsigned long)fi,
4653                                       BTRFS_FILE_EXTENT_INLINE_DATA_START);
4654                         }
4655                 }
4656
4657                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4658                               data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
4659                               data_end, old_data_start - data_end);
4660
4661                 offset = btrfs_disk_key_offset(&disk_key);
4662                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
4663                 btrfs_set_item_key(leaf, &disk_key, slot);
4664                 if (slot == 0)
4665                         fixup_low_keys(path, &disk_key, 1);
4666         }
4667
4668         item = btrfs_item_nr(slot);
4669         btrfs_set_item_size(leaf, item, new_size);
4670         btrfs_mark_buffer_dirty(leaf);
4671
4672         if (btrfs_leaf_free_space(leaf) < 0) {
4673                 btrfs_print_leaf(leaf);
4674                 BUG();
4675         }
4676 }
4677
4678 /*
4679  * make the item pointed to by the path bigger, data_size is the added size.
4680  */
4681 void btrfs_extend_item(struct btrfs_path *path, u32 data_size)
4682 {
4683         int slot;
4684         struct extent_buffer *leaf;
4685         struct btrfs_item *item;
4686         u32 nritems;
4687         unsigned int data_end;
4688         unsigned int old_data;
4689         unsigned int old_size;
4690         int i;
4691         struct btrfs_map_token token;
4692
4693         leaf = path->nodes[0];
4694
4695         nritems = btrfs_header_nritems(leaf);
4696         data_end = leaf_data_end(leaf);
4697
4698         if (btrfs_leaf_free_space(leaf) < data_size) {
4699                 btrfs_print_leaf(leaf);
4700                 BUG();
4701         }
4702         slot = path->slots[0];
4703         old_data = btrfs_item_end_nr(leaf, slot);
4704
4705         BUG_ON(slot < 0);
4706         if (slot >= nritems) {
4707                 btrfs_print_leaf(leaf);
4708                 btrfs_crit(leaf->fs_info, "slot %d too large, nritems %d",
4709                            slot, nritems);
4710                 BUG();
4711         }
4712
4713         /*
4714          * item0..itemN ... dataN.offset..dataN.size .. data0.size
4715          */
4716         /* first correct the data pointers */
4717         btrfs_init_map_token(&token, leaf);
4718         for (i = slot; i < nritems; i++) {
4719                 u32 ioff;
4720                 item = btrfs_item_nr(i);
4721
4722                 ioff = btrfs_token_item_offset(leaf, item, &token);
4723                 btrfs_set_token_item_offset(leaf, item,
4724                                             ioff - data_size, &token);
4725         }
4726
4727         /* shift the data */
4728         memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4729                       data_end - data_size, BTRFS_LEAF_DATA_OFFSET +
4730                       data_end, old_data - data_end);
4731
4732         data_end = old_data;
4733         old_size = btrfs_item_size_nr(leaf, slot);
4734         item = btrfs_item_nr(slot);
4735         btrfs_set_item_size(leaf, item, old_size + data_size);
4736         btrfs_mark_buffer_dirty(leaf);
4737
4738         if (btrfs_leaf_free_space(leaf) < 0) {
4739                 btrfs_print_leaf(leaf);
4740                 BUG();
4741         }
4742 }
4743
4744 /*
4745  * this is a helper for btrfs_insert_empty_items, the main goal here is
4746  * to save stack depth by doing the bulk of the work in a function
4747  * that doesn't call btrfs_search_slot
4748  */
4749 void setup_items_for_insert(struct btrfs_root *root, struct btrfs_path *path,
4750                             const struct btrfs_key *cpu_key, u32 *data_size,
4751                             u32 total_data, u32 total_size, int nr)
4752 {
4753         struct btrfs_fs_info *fs_info = root->fs_info;
4754         struct btrfs_item *item;
4755         int i;
4756         u32 nritems;
4757         unsigned int data_end;
4758         struct btrfs_disk_key disk_key;
4759         struct extent_buffer *leaf;
4760         int slot;
4761         struct btrfs_map_token token;
4762
4763         if (path->slots[0] == 0) {
4764                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
4765                 fixup_low_keys(path, &disk_key, 1);
4766         }
4767         btrfs_unlock_up_safe(path, 1);
4768
4769         leaf = path->nodes[0];
4770         slot = path->slots[0];
4771
4772         nritems = btrfs_header_nritems(leaf);
4773         data_end = leaf_data_end(leaf);
4774
4775         if (btrfs_leaf_free_space(leaf) < total_size) {
4776                 btrfs_print_leaf(leaf);
4777                 btrfs_crit(fs_info, "not enough freespace need %u have %d",
4778                            total_size, btrfs_leaf_free_space(leaf));
4779                 BUG();
4780         }
4781
4782         btrfs_init_map_token(&token, leaf);
4783         if (slot != nritems) {
4784                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
4785
4786                 if (old_data < data_end) {
4787                         btrfs_print_leaf(leaf);
4788                         btrfs_crit(fs_info, "slot %d old_data %d data_end %d",
4789                                    slot, old_data, data_end);
4790                         BUG();
4791                 }
4792                 /*
4793                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
4794                  */
4795                 /* first correct the data pointers */
4796                 for (i = slot; i < nritems; i++) {
4797                         u32 ioff;
4798
4799                         item = btrfs_item_nr(i);
4800                         ioff = btrfs_token_item_offset(leaf, item, &token);
4801                         btrfs_set_token_item_offset(leaf, item,
4802                                                     ioff - total_data, &token);
4803                 }
4804                 /* shift the items */
4805                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
4806                               btrfs_item_nr_offset(slot),
4807                               (nritems - slot) * sizeof(struct btrfs_item));
4808
4809                 /* shift the data */
4810                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4811                               data_end - total_data, BTRFS_LEAF_DATA_OFFSET +
4812                               data_end, old_data - data_end);
4813                 data_end = old_data;
4814         }
4815
4816         /* setup the item for the new data */
4817         for (i = 0; i < nr; i++) {
4818                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
4819                 btrfs_set_item_key(leaf, &disk_key, slot + i);
4820                 item = btrfs_item_nr(slot + i);
4821                 btrfs_set_token_item_offset(leaf, item,
4822                                             data_end - data_size[i], &token);
4823                 data_end -= data_size[i];
4824                 btrfs_set_token_item_size(leaf, item, data_size[i], &token);
4825         }
4826
4827         btrfs_set_header_nritems(leaf, nritems + nr);
4828         btrfs_mark_buffer_dirty(leaf);
4829
4830         if (btrfs_leaf_free_space(leaf) < 0) {
4831                 btrfs_print_leaf(leaf);
4832                 BUG();
4833         }
4834 }
4835
4836 /*
4837  * Given a key and some data, insert items into the tree.
4838  * This does all the path init required, making room in the tree if needed.
4839  */
4840 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
4841                             struct btrfs_root *root,
4842                             struct btrfs_path *path,
4843                             const struct btrfs_key *cpu_key, u32 *data_size,
4844                             int nr)
4845 {
4846         int ret = 0;
4847         int slot;
4848         int i;
4849         u32 total_size = 0;
4850         u32 total_data = 0;
4851
4852         for (i = 0; i < nr; i++)
4853                 total_data += data_size[i];
4854
4855         total_size = total_data + (nr * sizeof(struct btrfs_item));
4856         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
4857         if (ret == 0)
4858                 return -EEXIST;
4859         if (ret < 0)
4860                 return ret;
4861
4862         slot = path->slots[0];
4863         BUG_ON(slot < 0);
4864
4865         setup_items_for_insert(root, path, cpu_key, data_size,
4866                                total_data, total_size, nr);
4867         return 0;
4868 }
4869
4870 /*
4871  * Given a key and some data, insert an item into the tree.
4872  * This does all the path init required, making room in the tree if needed.
4873  */
4874 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4875                       const struct btrfs_key *cpu_key, void *data,
4876                       u32 data_size)
4877 {
4878         int ret = 0;
4879         struct btrfs_path *path;
4880         struct extent_buffer *leaf;
4881         unsigned long ptr;
4882
4883         path = btrfs_alloc_path();
4884         if (!path)
4885                 return -ENOMEM;
4886         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
4887         if (!ret) {
4888                 leaf = path->nodes[0];
4889                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
4890                 write_extent_buffer(leaf, data, ptr, data_size);
4891                 btrfs_mark_buffer_dirty(leaf);
4892         }
4893         btrfs_free_path(path);
4894         return ret;
4895 }
4896
4897 /*
4898  * delete the pointer from a given node.
4899  *
4900  * the tree should have been previously balanced so the deletion does not
4901  * empty a node.
4902  */
4903 static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
4904                     int level, int slot)
4905 {
4906         struct extent_buffer *parent = path->nodes[level];
4907         u32 nritems;
4908         int ret;
4909
4910         nritems = btrfs_header_nritems(parent);
4911         if (slot != nritems - 1) {
4912                 if (level) {
4913                         ret = tree_mod_log_insert_move(parent, slot, slot + 1,
4914                                         nritems - slot - 1);
4915                         BUG_ON(ret < 0);
4916                 }
4917                 memmove_extent_buffer(parent,
4918                               btrfs_node_key_ptr_offset(slot),
4919                               btrfs_node_key_ptr_offset(slot + 1),
4920                               sizeof(struct btrfs_key_ptr) *
4921                               (nritems - slot - 1));
4922         } else if (level) {
4923                 ret = tree_mod_log_insert_key(parent, slot, MOD_LOG_KEY_REMOVE,
4924                                 GFP_NOFS);
4925                 BUG_ON(ret < 0);
4926         }
4927
4928         nritems--;
4929         btrfs_set_header_nritems(parent, nritems);
4930         if (nritems == 0 && parent == root->node) {
4931                 BUG_ON(btrfs_header_level(root->node) != 1);
4932                 /* just turn the root into a leaf and break */
4933                 btrfs_set_header_level(root->node, 0);
4934         } else if (slot == 0) {
4935                 struct btrfs_disk_key disk_key;
4936
4937                 btrfs_node_key(parent, &disk_key, 0);
4938                 fixup_low_keys(path, &disk_key, level + 1);
4939         }
4940         btrfs_mark_buffer_dirty(parent);
4941 }
4942
4943 /*
4944  * a helper function to delete the leaf pointed to by path->slots[1] and
4945  * path->nodes[1].
4946  *
4947  * This deletes the pointer in path->nodes[1] and frees the leaf
4948  * block extent.  zero is returned if it all worked out, < 0 otherwise.
4949  *
4950  * The path must have already been setup for deleting the leaf, including
4951  * all the proper balancing.  path->nodes[1] must be locked.
4952  */
4953 static noinline void btrfs_del_leaf(struct btrfs_trans_handle *trans,
4954                                     struct btrfs_root *root,
4955                                     struct btrfs_path *path,
4956                                     struct extent_buffer *leaf)
4957 {
4958         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
4959         del_ptr(root, path, 1, path->slots[1]);
4960
4961         /*
4962          * btrfs_free_extent is expensive, we want to make sure we
4963          * aren't holding any locks when we call it
4964          */
4965         btrfs_unlock_up_safe(path, 0);
4966
4967         root_sub_used(root, leaf->len);
4968
4969         extent_buffer_get(leaf);
4970         btrfs_free_tree_block(trans, root, leaf, 0, 1);
4971         free_extent_buffer_stale(leaf);
4972 }
4973 /*
4974  * delete the item at the leaf level in path.  If that empties
4975  * the leaf, remove it from the tree
4976  */
4977 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4978                     struct btrfs_path *path, int slot, int nr)
4979 {
4980         struct btrfs_fs_info *fs_info = root->fs_info;
4981         struct extent_buffer *leaf;
4982         struct btrfs_item *item;
4983         u32 last_off;
4984         u32 dsize = 0;
4985         int ret = 0;
4986         int wret;
4987         int i;
4988         u32 nritems;
4989
4990         leaf = path->nodes[0];
4991         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
4992
4993         for (i = 0; i < nr; i++)
4994                 dsize += btrfs_item_size_nr(leaf, slot + i);
4995
4996         nritems = btrfs_header_nritems(leaf);
4997
4998         if (slot + nr != nritems) {
4999                 int data_end = leaf_data_end(leaf);
5000                 struct btrfs_map_token token;
5001
5002                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
5003                               data_end + dsize,
5004                               BTRFS_LEAF_DATA_OFFSET + data_end,
5005                               last_off - data_end);
5006
5007                 btrfs_init_map_token(&token, leaf);
5008                 for (i = slot + nr; i < nritems; i++) {
5009                         u32 ioff;
5010
5011                         item = btrfs_item_nr(i);
5012                         ioff = btrfs_token_item_offset(leaf, item, &token);
5013                         btrfs_set_token_item_offset(leaf, item,
5014                                                     ioff + dsize, &token);
5015                 }
5016
5017                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
5018                               btrfs_item_nr_offset(slot + nr),
5019                               sizeof(struct btrfs_item) *
5020                               (nritems - slot - nr));
5021         }
5022         btrfs_set_header_nritems(leaf, nritems - nr);
5023         nritems -= nr;
5024
5025         /* delete the leaf if we've emptied it */
5026         if (nritems == 0) {
5027                 if (leaf == root->node) {
5028                         btrfs_set_header_level(leaf, 0);
5029                 } else {
5030                         btrfs_set_path_blocking(path);
5031                         btrfs_clean_tree_block(leaf);
5032                         btrfs_del_leaf(trans, root, path, leaf);
5033                 }
5034         } else {
5035                 int used = leaf_space_used(leaf, 0, nritems);
5036                 if (slot == 0) {
5037                         struct btrfs_disk_key disk_key;
5038
5039                         btrfs_item_key(leaf, &disk_key, 0);
5040                         fixup_low_keys(path, &disk_key, 1);
5041                 }
5042
5043                 /* delete the leaf if it is mostly empty */
5044                 if (used < BTRFS_LEAF_DATA_SIZE(fs_info) / 3) {
5045                         /* push_leaf_left fixes the path.
5046                          * make sure the path still points to our leaf
5047                          * for possible call to del_ptr below
5048                          */
5049                         slot = path->slots[1];
5050                         extent_buffer_get(leaf);
5051
5052                         btrfs_set_path_blocking(path);
5053                         wret = push_leaf_left(trans, root, path, 1, 1,
5054                                               1, (u32)-1);
5055                         if (wret < 0 && wret != -ENOSPC)
5056                                 ret = wret;
5057
5058                         if (path->nodes[0] == leaf &&
5059                             btrfs_header_nritems(leaf)) {
5060                                 wret = push_leaf_right(trans, root, path, 1,
5061                                                        1, 1, 0);
5062                                 if (wret < 0 && wret != -ENOSPC)
5063                                         ret = wret;
5064                         }
5065
5066                         if (btrfs_header_nritems(leaf) == 0) {
5067                                 path->slots[1] = slot;
5068                                 btrfs_del_leaf(trans, root, path, leaf);
5069                                 free_extent_buffer(leaf);
5070                                 ret = 0;
5071                         } else {
5072                                 /* if we're still in the path, make sure
5073                                  * we're dirty.  Otherwise, one of the
5074                                  * push_leaf functions must have already
5075                                  * dirtied this buffer
5076                                  */
5077                                 if (path->nodes[0] == leaf)
5078                                         btrfs_mark_buffer_dirty(leaf);
5079                                 free_extent_buffer(leaf);
5080                         }
5081                 } else {
5082                         btrfs_mark_buffer_dirty(leaf);
5083                 }
5084         }
5085         return ret;
5086 }
5087
5088 /*
5089  * search the tree again to find a leaf with lesser keys
5090  * returns 0 if it found something or 1 if there are no lesser leaves.
5091  * returns < 0 on io errors.
5092  *
5093  * This may release the path, and so you may lose any locks held at the
5094  * time you call it.
5095  */
5096 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
5097 {
5098         struct btrfs_key key;
5099         struct btrfs_disk_key found_key;
5100         int ret;
5101
5102         btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
5103
5104         if (key.offset > 0) {
5105                 key.offset--;
5106         } else if (key.type > 0) {
5107                 key.type--;
5108                 key.offset = (u64)-1;
5109         } else if (key.objectid > 0) {
5110                 key.objectid--;
5111                 key.type = (u8)-1;
5112                 key.offset = (u64)-1;
5113         } else {
5114                 return 1;
5115         }
5116
5117         btrfs_release_path(path);
5118         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5119         if (ret < 0)
5120                 return ret;
5121         btrfs_item_key(path->nodes[0], &found_key, 0);
5122         ret = comp_keys(&found_key, &key);
5123         /*
5124          * We might have had an item with the previous key in the tree right
5125          * before we released our path. And after we released our path, that
5126          * item might have been pushed to the first slot (0) of the leaf we
5127          * were holding due to a tree balance. Alternatively, an item with the
5128          * previous key can exist as the only element of a leaf (big fat item).
5129          * Therefore account for these 2 cases, so that our callers (like
5130          * btrfs_previous_item) don't miss an existing item with a key matching
5131          * the previous key we computed above.
5132          */
5133         if (ret <= 0)
5134                 return 0;
5135         return 1;
5136 }
5137
5138 /*
5139  * A helper function to walk down the tree starting at min_key, and looking
5140  * for nodes or leaves that are have a minimum transaction id.
5141  * This is used by the btree defrag code, and tree logging
5142  *
5143  * This does not cow, but it does stuff the starting key it finds back
5144  * into min_key, so you can call btrfs_search_slot with cow=1 on the
5145  * key and get a writable path.
5146  *
5147  * This honors path->lowest_level to prevent descent past a given level
5148  * of the tree.
5149  *
5150  * min_trans indicates the oldest transaction that you are interested
5151  * in walking through.  Any nodes or leaves older than min_trans are
5152  * skipped over (without reading them).
5153  *
5154  * returns zero if something useful was found, < 0 on error and 1 if there
5155  * was nothing in the tree that matched the search criteria.
5156  */
5157 int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
5158                          struct btrfs_path *path,
5159                          u64 min_trans)
5160 {
5161         struct extent_buffer *cur;
5162         struct btrfs_key found_key;
5163         int slot;
5164         int sret;
5165         u32 nritems;
5166         int level;
5167         int ret = 1;
5168         int keep_locks = path->keep_locks;
5169
5170         path->keep_locks = 1;
5171 again:
5172         cur = btrfs_read_lock_root_node(root);
5173         level = btrfs_header_level(cur);
5174         WARN_ON(path->nodes[level]);
5175         path->nodes[level] = cur;
5176         path->locks[level] = BTRFS_READ_LOCK;
5177
5178         if (btrfs_header_generation(cur) < min_trans) {
5179                 ret = 1;
5180                 goto out;
5181         }
5182         while (1) {
5183                 nritems = btrfs_header_nritems(cur);
5184                 level = btrfs_header_level(cur);
5185                 sret = btrfs_bin_search(cur, min_key, level, &slot);
5186                 if (sret < 0) {
5187                         ret = sret;
5188                         goto out;
5189                 }
5190
5191                 /* at the lowest level, we're done, setup the path and exit */
5192                 if (level == path->lowest_level) {
5193                         if (slot >= nritems)
5194                                 goto find_next_key;
5195                         ret = 0;
5196                         path->slots[level] = slot;
5197                         btrfs_item_key_to_cpu(cur, &found_key, slot);
5198                         goto out;
5199                 }
5200                 if (sret && slot > 0)
5201                         slot--;
5202                 /*
5203                  * check this node pointer against the min_trans parameters.
5204                  * If it is too old, old, skip to the next one.
5205                  */
5206                 while (slot < nritems) {
5207                         u64 gen;
5208
5209                         gen = btrfs_node_ptr_generation(cur, slot);
5210                         if (gen < min_trans) {
5211                                 slot++;
5212                                 continue;
5213                         }
5214                         break;
5215                 }
5216 find_next_key:
5217                 /*
5218                  * we didn't find a candidate key in this node, walk forward
5219                  * and find another one
5220                  */
5221                 if (slot >= nritems) {
5222                         path->slots[level] = slot;
5223                         btrfs_set_path_blocking(path);
5224                         sret = btrfs_find_next_key(root, path, min_key, level,
5225                                                   min_trans);
5226                         if (sret == 0) {
5227                                 btrfs_release_path(path);
5228                                 goto again;
5229                         } else {
5230                                 goto out;
5231                         }
5232                 }
5233                 /* save our key for returning back */
5234                 btrfs_node_key_to_cpu(cur, &found_key, slot);
5235                 path->slots[level] = slot;
5236                 if (level == path->lowest_level) {
5237                         ret = 0;
5238                         goto out;
5239                 }
5240                 btrfs_set_path_blocking(path);
5241                 cur = btrfs_read_node_slot(cur, slot);
5242                 if (IS_ERR(cur)) {
5243                         ret = PTR_ERR(cur);
5244                         goto out;
5245                 }
5246
5247                 btrfs_tree_read_lock(cur);
5248
5249                 path->locks[level - 1] = BTRFS_READ_LOCK;
5250                 path->nodes[level - 1] = cur;
5251                 unlock_up(path, level, 1, 0, NULL);
5252         }
5253 out:
5254         path->keep_locks = keep_locks;
5255         if (ret == 0) {
5256                 btrfs_unlock_up_safe(path, path->lowest_level + 1);
5257                 btrfs_set_path_blocking(path);
5258                 memcpy(min_key, &found_key, sizeof(found_key));
5259         }
5260         return ret;
5261 }
5262
5263 /*
5264  * this is similar to btrfs_next_leaf, but does not try to preserve
5265  * and fixup the path.  It looks for and returns the next key in the
5266  * tree based on the current path and the min_trans parameters.
5267  *
5268  * 0 is returned if another key is found, < 0 if there are any errors
5269  * and 1 is returned if there are no higher keys in the tree
5270  *
5271  * path->keep_locks should be set to 1 on the search made before
5272  * calling this function.
5273  */
5274 int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
5275                         struct btrfs_key *key, int level, u64 min_trans)
5276 {
5277         int slot;
5278         struct extent_buffer *c;
5279
5280         WARN_ON(!path->keep_locks && !path->skip_locking);
5281         while (level < BTRFS_MAX_LEVEL) {
5282                 if (!path->nodes[level])
5283                         return 1;
5284
5285                 slot = path->slots[level] + 1;
5286                 c = path->nodes[level];
5287 next:
5288                 if (slot >= btrfs_header_nritems(c)) {
5289                         int ret;
5290                         int orig_lowest;
5291                         struct btrfs_key cur_key;
5292                         if (level + 1 >= BTRFS_MAX_LEVEL ||
5293                             !path->nodes[level + 1])
5294                                 return 1;
5295
5296                         if (path->locks[level + 1] || path->skip_locking) {
5297                                 level++;
5298                                 continue;
5299                         }
5300
5301                         slot = btrfs_header_nritems(c) - 1;
5302                         if (level == 0)
5303                                 btrfs_item_key_to_cpu(c, &cur_key, slot);
5304                         else
5305                                 btrfs_node_key_to_cpu(c, &cur_key, slot);
5306
5307                         orig_lowest = path->lowest_level;
5308                         btrfs_release_path(path);
5309                         path->lowest_level = level;
5310                         ret = btrfs_search_slot(NULL, root, &cur_key, path,
5311                                                 0, 0);
5312                         path->lowest_level = orig_lowest;
5313                         if (ret < 0)
5314                                 return ret;
5315
5316                         c = path->nodes[level];
5317                         slot = path->slots[level];
5318                         if (ret == 0)
5319                                 slot++;
5320                         goto next;
5321                 }
5322
5323                 if (level == 0)
5324                         btrfs_item_key_to_cpu(c, key, slot);
5325                 else {
5326                         u64 gen = btrfs_node_ptr_generation(c, slot);
5327
5328                         if (gen < min_trans) {
5329                                 slot++;
5330                                 goto next;
5331                         }
5332                         btrfs_node_key_to_cpu(c, key, slot);
5333                 }
5334                 return 0;
5335         }
5336         return 1;
5337 }
5338
5339 /*
5340  * search the tree again to find a leaf with greater keys
5341  * returns 0 if it found something or 1 if there are no greater leaves.
5342  * returns < 0 on io errors.
5343  */
5344 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
5345 {
5346         return btrfs_next_old_leaf(root, path, 0);
5347 }
5348
5349 int btrfs_next_old_leaf(struct btrfs_root *root, struct btrfs_path *path,
5350                         u64 time_seq)
5351 {
5352         int slot;
5353         int level;
5354         struct extent_buffer *c;
5355         struct extent_buffer *next;
5356         struct btrfs_key key;
5357         u32 nritems;
5358         int ret;
5359         int old_spinning = path->leave_spinning;
5360         int next_rw_lock = 0;
5361
5362         nritems = btrfs_header_nritems(path->nodes[0]);
5363         if (nritems == 0)
5364                 return 1;
5365
5366         btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
5367 again:
5368         level = 1;
5369         next = NULL;
5370         next_rw_lock = 0;
5371         btrfs_release_path(path);
5372
5373         path->keep_locks = 1;
5374         path->leave_spinning = 1;
5375
5376         if (time_seq)
5377                 ret = btrfs_search_old_slot(root, &key, path, time_seq);
5378         else
5379                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5380         path->keep_locks = 0;
5381
5382         if (ret < 0)
5383                 return ret;
5384
5385         nritems = btrfs_header_nritems(path->nodes[0]);
5386         /*
5387          * by releasing the path above we dropped all our locks.  A balance
5388          * could have added more items next to the key that used to be
5389          * at the very end of the block.  So, check again here and
5390          * advance the path if there are now more items available.
5391          */
5392         if (nritems > 0 && path->slots[0] < nritems - 1) {
5393                 if (ret == 0)
5394                         path->slots[0]++;
5395                 ret = 0;
5396                 goto done;
5397         }
5398         /*
5399          * So the above check misses one case:
5400          * - after releasing the path above, someone has removed the item that
5401          *   used to be at the very end of the block, and balance between leafs
5402          *   gets another one with bigger key.offset to replace it.
5403          *
5404          * This one should be returned as well, or we can get leaf corruption
5405          * later(esp. in __btrfs_drop_extents()).
5406          *
5407          * And a bit more explanation about this check,
5408          * with ret > 0, the key isn't found, the path points to the slot
5409          * where it should be inserted, so the path->slots[0] item must be the
5410          * bigger one.
5411          */
5412         if (nritems > 0 && ret > 0 && path->slots[0] == nritems - 1) {
5413                 ret = 0;
5414                 goto done;
5415         }
5416
5417         while (level < BTRFS_MAX_LEVEL) {
5418                 if (!path->nodes[level]) {
5419                         ret = 1;
5420                         goto done;
5421                 }
5422
5423                 slot = path->slots[level] + 1;
5424                 c = path->nodes[level];
5425                 if (slot >= btrfs_header_nritems(c)) {
5426                         level++;
5427                         if (level == BTRFS_MAX_LEVEL) {
5428                                 ret = 1;
5429                                 goto done;
5430                         }
5431                         continue;
5432                 }
5433
5434                 if (next) {
5435                         btrfs_tree_unlock_rw(next, next_rw_lock);
5436                         free_extent_buffer(next);
5437                 }
5438
5439                 next = c;
5440                 next_rw_lock = path->locks[level];
5441                 ret = read_block_for_search(root, path, &next, level,
5442                                             slot, &key);
5443                 if (ret == -EAGAIN)
5444                         goto again;
5445
5446                 if (ret < 0) {
5447                         btrfs_release_path(path);
5448                         goto done;
5449                 }
5450
5451                 if (!path->skip_locking) {
5452                         ret = btrfs_try_tree_read_lock(next);
5453                         if (!ret && time_seq) {
5454                                 /*
5455                                  * If we don't get the lock, we may be racing
5456                                  * with push_leaf_left, holding that lock while
5457                                  * itself waiting for the leaf we've currently
5458                                  * locked. To solve this situation, we give up
5459                                  * on our lock and cycle.
5460                                  */
5461                                 free_extent_buffer(next);
5462                                 btrfs_release_path(path);
5463                                 cond_resched();
5464                                 goto again;
5465                         }
5466                         if (!ret) {
5467                                 btrfs_set_path_blocking(path);
5468                                 btrfs_tree_read_lock(next);
5469                         }
5470                         next_rw_lock = BTRFS_READ_LOCK;
5471                 }
5472                 break;
5473         }
5474         path->slots[level] = slot;
5475         while (1) {
5476                 level--;
5477                 c = path->nodes[level];
5478                 if (path->locks[level])
5479                         btrfs_tree_unlock_rw(c, path->locks[level]);
5480
5481                 free_extent_buffer(c);
5482                 path->nodes[level] = next;
5483                 path->slots[level] = 0;
5484                 if (!path->skip_locking)
5485                         path->locks[level] = next_rw_lock;
5486                 if (!level)
5487                         break;
5488
5489                 ret = read_block_for_search(root, path, &next, level,
5490                                             0, &key);
5491                 if (ret == -EAGAIN)
5492                         goto again;
5493
5494                 if (ret < 0) {
5495                         btrfs_release_path(path);
5496                         goto done;
5497                 }
5498
5499                 if (!path->skip_locking) {
5500                         ret = btrfs_try_tree_read_lock(next);
5501                         if (!ret) {
5502                                 btrfs_set_path_blocking(path);
5503                                 btrfs_tree_read_lock(next);
5504                         }
5505                         next_rw_lock = BTRFS_READ_LOCK;
5506                 }
5507         }
5508         ret = 0;
5509 done:
5510         unlock_up(path, 0, 1, 0, NULL);
5511         path->leave_spinning = old_spinning;
5512         if (!old_spinning)
5513                 btrfs_set_path_blocking(path);
5514
5515         return ret;
5516 }
5517
5518 /*
5519  * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
5520  * searching until it gets past min_objectid or finds an item of 'type'
5521  *
5522  * returns 0 if something is found, 1 if nothing was found and < 0 on error
5523  */
5524 int btrfs_previous_item(struct btrfs_root *root,
5525                         struct btrfs_path *path, u64 min_objectid,
5526                         int type)
5527 {
5528         struct btrfs_key found_key;
5529         struct extent_buffer *leaf;
5530         u32 nritems;
5531         int ret;
5532
5533         while (1) {
5534                 if (path->slots[0] == 0) {
5535                         btrfs_set_path_blocking(path);
5536                         ret = btrfs_prev_leaf(root, path);
5537                         if (ret != 0)
5538                                 return ret;
5539                 } else {
5540                         path->slots[0]--;
5541                 }
5542                 leaf = path->nodes[0];
5543                 nritems = btrfs_header_nritems(leaf);
5544                 if (nritems == 0)
5545                         return 1;
5546                 if (path->slots[0] == nritems)
5547                         path->slots[0]--;
5548
5549                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5550                 if (found_key.objectid < min_objectid)
5551                         break;
5552                 if (found_key.type == type)
5553                         return 0;
5554                 if (found_key.objectid == min_objectid &&
5555                     found_key.type < type)
5556                         break;
5557         }
5558         return 1;
5559 }
5560
5561 /*
5562  * search in extent tree to find a previous Metadata/Data extent item with
5563  * min objecitd.
5564  *
5565  * returns 0 if something is found, 1 if nothing was found and < 0 on error
5566  */
5567 int btrfs_previous_extent_item(struct btrfs_root *root,
5568                         struct btrfs_path *path, u64 min_objectid)
5569 {
5570         struct btrfs_key found_key;
5571         struct extent_buffer *leaf;
5572         u32 nritems;
5573         int ret;
5574
5575         while (1) {
5576                 if (path->slots[0] == 0) {
5577                         btrfs_set_path_blocking(path);
5578                         ret = btrfs_prev_leaf(root, path);
5579                         if (ret != 0)
5580                                 return ret;
5581                 } else {
5582                         path->slots[0]--;
5583                 }
5584                 leaf = path->nodes[0];
5585                 nritems = btrfs_header_nritems(leaf);
5586                 if (nritems == 0)
5587                         return 1;
5588                 if (path->slots[0] == nritems)
5589                         path->slots[0]--;
5590
5591                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5592                 if (found_key.objectid < min_objectid)
5593                         break;
5594                 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
5595                     found_key.type == BTRFS_METADATA_ITEM_KEY)
5596                         return 0;
5597                 if (found_key.objectid == min_objectid &&
5598                     found_key.type < BTRFS_EXTENT_ITEM_KEY)
5599                         break;
5600         }
5601         return 1;
5602 }