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
  17 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  18                       *root, struct btrfs_path *path, int level);
  19 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  20                       const struct btrfs_key *ins_key, struct btrfs_path *path,
  21                       int data_size, int extend);
  22 static int push_node_left(struct btrfs_trans_handle *trans,
  23                           struct btrfs_fs_info *fs_info,
  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 btrfs_fs_info *fs_info,
  28                               struct extent_buffer *dst_buf,
  29                               struct extent_buffer *src_buf);
  30 static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
  31                     int level, int slot);
  32
  33 struct btrfs_path *btrfs_alloc_path(void)
  34 {
  35         return kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
  36 }
  37
  38 /*
  39  * set all locked nodes in the path to blocking locks.  This should
  40  * be done before scheduling
  41  */
  42 noinline void btrfs_set_path_blocking(struct btrfs_path *p)
  43 {
  44         int i;
  45         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  46                 if (!p->nodes[i] || !p->locks[i])
  47                         continue;
  48                 btrfs_set_lock_blocking_rw(p->nodes[i], p->locks[i]);
  49                 if (p->locks[i] == BTRFS_READ_LOCK)
  50                         p->locks[i] = BTRFS_READ_LOCK_BLOCKING;
  51                 else if (p->locks[i] == BTRFS_WRITE_LOCK)
  52                         p->locks[i] = BTRFS_WRITE_LOCK_BLOCKING;
  53         }
  54 }
  55
  56 /* this also releases the path */
  57 void btrfs_free_path(struct btrfs_path *p)
  58 {
  59         if (!p)
  60                 return;
  61         btrfs_release_path(p);
  62         kmem_cache_free(btrfs_path_cachep, p);
  63 }
  64
  65 /*
  66  * path release drops references on the extent buffers in the path
  67  * and it drops any locks held by this path
  68  *
  69  * It is safe to call this on paths that no locks or extent buffers held.
  70  */
  71 noinline void btrfs_release_path(struct btrfs_path *p)
  72 {
  73         int i;
  74
  75         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  76                 p->slots[i] = 0;
  77                 if (!p->nodes[i])
  78                         continue;
  79                 if (p->locks[i]) {
  80                         btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
  81                         p->locks[i] = 0;
  82                 }
  83                 free_extent_buffer(p->nodes[i]);
  84                 p->nodes[i] = NULL;
  85         }
  86 }
  87
  88 /*
  89  * safely gets a reference on the root node of a tree.  A lock
  90  * is not taken, so a concurrent writer may put a different node
  91  * at the root of the tree.  See btrfs_lock_root_node for the
  92  * looping required.
  93  *
  94  * The extent buffer returned by this has a reference taken, so
  95  * it won't disappear.  It may stop being the root of the tree
  96  * at any time because there are no locks held.
  97  */
  98 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
  99 {
 100         struct extent_buffer *eb;
 101
 102         while (1) {
 103                 rcu_read_lock();
 104                 eb = rcu_dereference(root->node);
 105
 106                 /*
 107                  * RCU really hurts here, we could free up the root node because
 108                  * it was COWed but we may not get the new root node yet so do
 109                  * the inc_not_zero dance and if it doesn't work then
 110                  * synchronize_rcu and try again.
 111                  */
 112                 if (atomic_inc_not_zero(&eb->refs)) {
 113                         rcu_read_unlock();
 114                         break;
 115                 }
 116                 rcu_read_unlock();
 117                 synchronize_rcu();
 118         }
 119         return eb;
 120 }
 121
 122 /* loop around taking references on and locking the root node of the
 123  * tree until you end up with a lock on the root.  A locked buffer
 124  * is returned, with a reference held.
 125  */
 126 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
 127 {
 128         struct extent_buffer *eb;
 129
 130         while (1) {
 131                 eb = btrfs_root_node(root);
 132                 btrfs_tree_lock(eb);
 133                 if (eb == root->node)
 134                         break;
 135                 btrfs_tree_unlock(eb);
 136                 free_extent_buffer(eb);
 137         }
 138         return eb;
 139 }
 140
 141 /* loop around taking references on and locking the root node of the
 142  * tree until you end up with a lock on the root.  A locked buffer
 143  * is returned, with a reference held.
 144  */
 145 struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
 146 {
 147         struct extent_buffer *eb;
 148
 149         while (1) {
 150                 eb = btrfs_root_node(root);
 151                 btrfs_tree_read_lock(eb);
 152                 if (eb == root->node)
 153                         break;
 154                 btrfs_tree_read_unlock(eb);
 155                 free_extent_buffer(eb);
 156         }
 157         return eb;
 158 }
 159
 160 /* cowonly root (everything not a reference counted cow subvolume), just get
 161  * put onto a simple dirty list.  transaction.c walks this to make sure they
 162  * get properly updated on disk.
 163  */
 164 static void add_root_to_dirty_list(struct btrfs_root *root)
 165 {
 166         struct btrfs_fs_info *fs_info = root->fs_info;
 167
 168         if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
 169             !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
 170                 return;
 171
 172         spin_lock(&fs_info->trans_lock);
 173         if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
 174                 /* Want the extent tree to be the last on the list */
 175                 if (root->root_key.objectid == BTRFS_EXTENT_TREE_OBJECTID)
 176                         list_move_tail(&root->dirty_list,
 177                                        &fs_info->dirty_cowonly_roots);
 178                 else
 179                         list_move(&root->dirty_list,
 180                                   &fs_info->dirty_cowonly_roots);
 181         }
 182         spin_unlock(&fs_info->trans_lock);
 183 }
 184
 185 /*
 186  * used by snapshot creation to make a copy of a root for a tree with
 187  * a given objectid.  The buffer with the new root node is returned in
 188  * cow_ret, and this func returns zero on success or a negative error code.
 189  */
 190 int btrfs_copy_root(struct btrfs_trans_handle *trans,
 191                       struct btrfs_root *root,
 192                       struct extent_buffer *buf,
 193                       struct extent_buffer **cow_ret, u64 new_root_objectid)
 194 {
 195         struct btrfs_fs_info *fs_info = root->fs_info;
 196         struct extent_buffer *cow;
 197         int ret = 0;
 198         int level;
 199         struct btrfs_disk_key disk_key;
 200
 201         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 202                 trans->transid != fs_info->running_transaction->transid);
 203         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 204                 trans->transid != root->last_trans);
 205
 206         level = btrfs_header_level(buf);
 207         if (level == 0)
 208                 btrfs_item_key(buf, &disk_key, 0);
 209         else
 210                 btrfs_node_key(buf, &disk_key, 0);
 211
 212         cow = btrfs_alloc_tree_block(trans, root, 0, new_root_objectid,
 213                         &disk_key, level, buf->start, 0);
 214         if (IS_ERR(cow))
 215                 return PTR_ERR(cow);
 216
 217         copy_extent_buffer_full(cow, buf);
 218         btrfs_set_header_bytenr(cow, cow->start);
 219         btrfs_set_header_generation(cow, trans->transid);
 220         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 221         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 222                                      BTRFS_HEADER_FLAG_RELOC);
 223         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 224                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 225         else
 226                 btrfs_set_header_owner(cow, new_root_objectid);
 227
 228         write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
 229
 230         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 231         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 232                 ret = btrfs_inc_ref(trans, root, cow, 1);
 233         else
 234                 ret = btrfs_inc_ref(trans, root, cow, 0);
 235
 236         if (ret)
 237                 return ret;
 238
 239         btrfs_mark_buffer_dirty(cow);
 240         *cow_ret = cow;
 241         return 0;
 242 }
 243
 244 enum mod_log_op {
 245         MOD_LOG_KEY_REPLACE,
 246         MOD_LOG_KEY_ADD,
 247         MOD_LOG_KEY_REMOVE,
 248         MOD_LOG_KEY_REMOVE_WHILE_FREEING,
 249         MOD_LOG_KEY_REMOVE_WHILE_MOVING,
 250         MOD_LOG_MOVE_KEYS,
 251         MOD_LOG_ROOT_REPLACE,
 252 };
 253
 254 struct tree_mod_root {
 255         u64 logical;
 256         u8 level;
 257 };
 258
 259 struct tree_mod_elem {
 260         struct rb_node node;
 261         u64 logical;
 262         u64 seq;
 263         enum mod_log_op op;
 264
 265         /* this is used for MOD_LOG_KEY_* and MOD_LOG_MOVE_KEYS operations */
 266         int slot;
 267
 268         /* this is used for MOD_LOG_KEY* and MOD_LOG_ROOT_REPLACE */
 269         u64 generation;
 270
 271         /* those are used for op == MOD_LOG_KEY_{REPLACE,REMOVE} */
 272         struct btrfs_disk_key key;
 273         u64 blockptr;
 274
 275         /* this is used for op == MOD_LOG_MOVE_KEYS */
 276         struct {
 277                 int dst_slot;
 278                 int nr_items;
 279         } move;
 280
 281         /* this is used for op == MOD_LOG_ROOT_REPLACE */
 282         struct tree_mod_root old_root;
 283 };
 284
 285 /*
 286  * Pull a new tree mod seq number for our operation.
 287  */
 288 static inline u64 btrfs_inc_tree_mod_seq(struct btrfs_fs_info *fs_info)
 289 {
 290         return atomic64_inc_return(&fs_info->tree_mod_seq);
 291 }
 292
 293 /*
 294  * This adds a new blocker to the tree mod log's blocker list if the @elem
 295  * passed does not already have a sequence number set. So when a caller expects
 296  * to record tree modifications, it should ensure to set elem->seq to zero
 297  * before calling btrfs_get_tree_mod_seq.
 298  * Returns a fresh, unused tree log modification sequence number, even if no new
 299  * blocker was added.
 300  */
 301 u64 btrfs_get_tree_mod_seq(struct btrfs_fs_info *fs_info,
 302                            struct seq_list *elem)
 303 {
 304         write_lock(&fs_info->tree_mod_log_lock);
 305         spin_lock(&fs_info->tree_mod_seq_lock);
 306         if (!elem->seq) {
 307                 elem->seq = btrfs_inc_tree_mod_seq(fs_info);
 308                 list_add_tail(&elem->list, &fs_info->tree_mod_seq_list);
 309         }
 310         spin_unlock(&fs_info->tree_mod_seq_lock);
 311         write_unlock(&fs_info->tree_mod_log_lock);
 312
 313         return elem->seq;
 314 }
 315
 316 void btrfs_put_tree_mod_seq(struct btrfs_fs_info *fs_info,
 317                             struct seq_list *elem)
 318 {
 319         struct rb_root *tm_root;
 320         struct rb_node *node;
 321         struct rb_node *next;
 322         struct seq_list *cur_elem;
 323         struct tree_mod_elem *tm;
 324         u64 min_seq = (u64)-1;
 325         u64 seq_putting = elem->seq;
 326
 327         if (!seq_putting)
 328                 return;
 329
 330         spin_lock(&fs_info->tree_mod_seq_lock);
 331         list_del(&elem->list);
 332         elem->seq = 0;
 333
 334         list_for_each_entry(cur_elem, &fs_info->tree_mod_seq_list, list) {
 335                 if (cur_elem->seq < min_seq) {
 336                         if (seq_putting > cur_elem->seq) {
 337                                 /*
 338                                  * blocker with lower sequence number exists, we
 339                                  * cannot remove anything from the log
 340                                  */
 341                                 spin_unlock(&fs_info->tree_mod_seq_lock);
 342                                 return;
 343                         }
 344                         min_seq = cur_elem->seq;
 345                 }
 346         }
 347         spin_unlock(&fs_info->tree_mod_seq_lock);
 348
 349         /*
 350          * anything that's lower than the lowest existing (read: blocked)
 351          * sequence number can be removed from the tree.
 352          */
 353         write_lock(&fs_info->tree_mod_log_lock);
 354         tm_root = &fs_info->tree_mod_log;
 355         for (node = rb_first(tm_root); node; node = next) {
 356                 next = rb_next(node);
 357                 tm = rb_entry(node, struct tree_mod_elem, node);
 358                 if (tm->seq > min_seq)
 359                         continue;
 360                 rb_erase(node, tm_root);
 361                 kfree(tm);
 362         }
 363         write_unlock(&fs_info->tree_mod_log_lock);
 364 }
 365
 366 /*
 367  * key order of the log:
 368  *       node/leaf start address -> sequence
 369  *
 370  * The 'start address' is the logical address of the *new* root node
 371  * for root replace operations, or the logical address of the affected
 372  * block for all other operations.
 373  *
 374  * Note: must be called with write lock for fs_info::tree_mod_log_lock.
 375  */
 376 static noinline int
 377 __tree_mod_log_insert(struct btrfs_fs_info *fs_info, struct tree_mod_elem *tm)
 378 {
 379         struct rb_root *tm_root;
 380         struct rb_node **new;
 381         struct rb_node *parent = NULL;
 382         struct tree_mod_elem *cur;
 383
 384         tm->seq = btrfs_inc_tree_mod_seq(fs_info);
 385
 386         tm_root = &fs_info->tree_mod_log;
 387         new = &tm_root->rb_node;
 388         while (*new) {
 389                 cur = rb_entry(*new, struct tree_mod_elem, node);
 390                 parent = *new;
 391                 if (cur->logical < tm->logical)
 392                         new = &((*new)->rb_left);
 393                 else if (cur->logical > tm->logical)
 394                         new = &((*new)->rb_right);
 395                 else if (cur->seq < tm->seq)
 396                         new = &((*new)->rb_left);
 397                 else if (cur->seq > tm->seq)
 398                         new = &((*new)->rb_right);
 399                 else
 400                         return -EEXIST;
 401         }
 402
 403         rb_link_node(&tm->node, parent, new);
 404         rb_insert_color(&tm->node, tm_root);
 405         return 0;
 406 }
 407
 408 /*
 409  * Determines if logging can be omitted. Returns 1 if it can. Otherwise, it
 410  * returns zero with the tree_mod_log_lock acquired. The caller must hold
 411  * this until all tree mod log insertions are recorded in the rb tree and then
 412  * write unlock fs_info::tree_mod_log_lock.
 413  */
 414 static inline int tree_mod_dont_log(struct btrfs_fs_info *fs_info,
 415                                     struct extent_buffer *eb) {
 416         smp_mb();
 417         if (list_empty(&(fs_info)->tree_mod_seq_list))
 418                 return 1;
 419         if (eb && btrfs_header_level(eb) == 0)
 420                 return 1;
 421
 422         write_lock(&fs_info->tree_mod_log_lock);
 423         if (list_empty(&(fs_info)->tree_mod_seq_list)) {
 424                 write_unlock(&fs_info->tree_mod_log_lock);
 425                 return 1;
 426         }
 427
 428         return 0;
 429 }
 430
 431 /* Similar to tree_mod_dont_log, but doesn't acquire any locks. */
 432 static inline int tree_mod_need_log(const struct btrfs_fs_info *fs_info,
 433                                     struct extent_buffer *eb)
 434 {
 435         smp_mb();
 436         if (list_empty(&(fs_info)->tree_mod_seq_list))
 437                 return 0;
 438         if (eb && btrfs_header_level(eb) == 0)
 439                 return 0;
 440
 441         return 1;
 442 }
 443
 444 static struct tree_mod_elem *
 445 alloc_tree_mod_elem(struct extent_buffer *eb, int slot,
 446                     enum mod_log_op op, gfp_t flags)
 447 {
 448         struct tree_mod_elem *tm;
 449
 450         tm = kzalloc(sizeof(*tm), flags);
 451         if (!tm)
 452                 return NULL;
 453
 454         tm->logical = eb->start;
 455         if (op != MOD_LOG_KEY_ADD) {
 456                 btrfs_node_key(eb, &tm->key, slot);
 457                 tm->blockptr = btrfs_node_blockptr(eb, slot);
 458         }
 459         tm->op = op;
 460         tm->slot = slot;
 461         tm->generation = btrfs_node_ptr_generation(eb, slot);
 462         RB_CLEAR_NODE(&tm->node);
 463
 464         return tm;
 465 }
 466
 467 static noinline int tree_mod_log_insert_key(struct extent_buffer *eb, int slot,
 468                 enum mod_log_op op, gfp_t flags)
 469 {
 470         struct tree_mod_elem *tm;
 471         int ret;
 472
 473         if (!tree_mod_need_log(eb->fs_info, eb))
 474                 return 0;
 475
 476         tm = alloc_tree_mod_elem(eb, slot, op, flags);
 477         if (!tm)
 478                 return -ENOMEM;
 479
 480         if (tree_mod_dont_log(eb->fs_info, eb)) {
 481                 kfree(tm);
 482                 return 0;
 483         }
 484
 485         ret = __tree_mod_log_insert(eb->fs_info, tm);
 486         write_unlock(&eb->fs_info->tree_mod_log_lock);
 487         if (ret)
 488                 kfree(tm);
 489
 490         return ret;
 491 }
 492
 493 static noinline int tree_mod_log_insert_move(struct extent_buffer *eb,
 494                 int dst_slot, int src_slot, int nr_items)
 495 {
 496         struct tree_mod_elem *tm = NULL;
 497         struct tree_mod_elem **tm_list = NULL;
 498         int ret = 0;
 499         int i;
 500         int locked = 0;
 501
 502         if (!tree_mod_need_log(eb->fs_info, eb))
 503                 return 0;
 504
 505         tm_list = kcalloc(nr_items, sizeof(struct tree_mod_elem *), GFP_NOFS);
 506         if (!tm_list)
 507                 return -ENOMEM;
 508
 509         tm = kzalloc(sizeof(*tm), GFP_NOFS);
 510         if (!tm) {
 511                 ret = -ENOMEM;
 512                 goto free_tms;
 513         }
 514
 515         tm->logical = eb->start;
 516         tm->slot = src_slot;
 517         tm->move.dst_slot = dst_slot;
 518         tm->move.nr_items = nr_items;
 519         tm->op = MOD_LOG_MOVE_KEYS;
 520
 521         for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
 522                 tm_list[i] = alloc_tree_mod_elem(eb, i + dst_slot,
 523                     MOD_LOG_KEY_REMOVE_WHILE_MOVING, GFP_NOFS);
 524                 if (!tm_list[i]) {
 525                         ret = -ENOMEM;
 526                         goto free_tms;
 527                 }
 528         }
 529
 530         if (tree_mod_dont_log(eb->fs_info, eb))
 531                 goto free_tms;
 532         locked = 1;
 533
 534         /*
 535          * When we override something during the move, we log these removals.
 536          * This can only happen when we move towards the beginning of the
 537          * buffer, i.e. dst_slot < src_slot.
 538          */
 539         for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
 540                 ret = __tree_mod_log_insert(eb->fs_info, tm_list[i]);
 541                 if (ret)
 542                         goto free_tms;
 543         }
 544
 545         ret = __tree_mod_log_insert(eb->fs_info, tm);
 546         if (ret)
 547                 goto free_tms;
 548         write_unlock(&eb->fs_info->tree_mod_log_lock);
 549         kfree(tm_list);
 550
 551         return 0;
 552 free_tms:
 553         for (i = 0; i < nr_items; i++) {
 554                 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
 555                         rb_erase(&tm_list[i]->node, &eb->fs_info->tree_mod_log);
 556                 kfree(tm_list[i]);
 557         }
 558         if (locked)
 559                 write_unlock(&eb->fs_info->tree_mod_log_lock);
 560         kfree(tm_list);
 561         kfree(tm);
 562
 563         return ret;
 564 }
 565
 566 static inline int
 567 __tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
 568                        struct tree_mod_elem **tm_list,
 569                        int nritems)
 570 {
 571         int i, j;
 572         int ret;
 573
 574         for (i = nritems - 1; i >= 0; i--) {
 575                 ret = __tree_mod_log_insert(fs_info, tm_list[i]);
 576                 if (ret) {
 577                         for (j = nritems - 1; j > i; j--)
 578                                 rb_erase(&tm_list[j]->node,
 579                                          &fs_info->tree_mod_log);
 580                         return ret;
 581                 }
 582         }
 583
 584         return 0;
 585 }
 586
 587 static noinline int tree_mod_log_insert_root(struct extent_buffer *old_root,
 588                          struct extent_buffer *new_root, int log_removal)
 589 {
 590         struct btrfs_fs_info *fs_info = old_root->fs_info;
 591         struct tree_mod_elem *tm = NULL;
 592         struct tree_mod_elem **tm_list = NULL;
 593         int nritems = 0;
 594         int ret = 0;
 595         int i;
 596
 597         if (!tree_mod_need_log(fs_info, NULL))
 598                 return 0;
 599
 600         if (log_removal && btrfs_header_level(old_root) > 0) {
 601                 nritems = btrfs_header_nritems(old_root);
 602                 tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *),
 603                                   GFP_NOFS);
 604                 if (!tm_list) {
 605                         ret = -ENOMEM;
 606                         goto free_tms;
 607                 }
 608                 for (i = 0; i < nritems; i++) {
 609                         tm_list[i] = alloc_tree_mod_elem(old_root, i,
 610                             MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
 611                         if (!tm_list[i]) {
 612                                 ret = -ENOMEM;
 613                                 goto free_tms;
 614                         }
 615                 }
 616         }
 617
 618         tm = kzalloc(sizeof(*tm), GFP_NOFS);
 619         if (!tm) {
 620                 ret = -ENOMEM;
 621                 goto free_tms;
 622         }
 623
 624         tm->logical = new_root->start;
 625         tm->old_root.logical = old_root->start;
 626         tm->old_root.level = btrfs_header_level(old_root);
 627         tm->generation = btrfs_header_generation(old_root);
 628         tm->op = MOD_LOG_ROOT_REPLACE;
 629
 630         if (tree_mod_dont_log(fs_info, NULL))
 631                 goto free_tms;
 632
 633         if (tm_list)
 634                 ret = __tree_mod_log_free_eb(fs_info, tm_list, nritems);
 635         if (!ret)
 636                 ret = __tree_mod_log_insert(fs_info, tm);
 637
 638         write_unlock(&fs_info->tree_mod_log_lock);
 639         if (ret)
 640                 goto free_tms;
 641         kfree(tm_list);
 642
 643         return ret;
 644
 645 free_tms:
 646         if (tm_list) {
 647                 for (i = 0; i < nritems; i++)
 648                         kfree(tm_list[i]);
 649                 kfree(tm_list);
 650         }
 651         kfree(tm);
 652
 653         return ret;
 654 }
 655
 656 static struct tree_mod_elem *
 657 __tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq,
 658                       int smallest)
 659 {
 660         struct rb_root *tm_root;
 661         struct rb_node *node;
 662         struct tree_mod_elem *cur = NULL;
 663         struct tree_mod_elem *found = NULL;
 664
 665         read_lock(&fs_info->tree_mod_log_lock);
 666         tm_root = &fs_info->tree_mod_log;
 667         node = tm_root->rb_node;
 668         while (node) {
 669                 cur = rb_entry(node, struct tree_mod_elem, node);
 670                 if (cur->logical < start) {
 671                         node = node->rb_left;
 672                 } else if (cur->logical > start) {
 673                         node = node->rb_right;
 674                 } else if (cur->seq < min_seq) {
 675                         node = node->rb_left;
 676                 } else if (!smallest) {
 677                         /* we want the node with the highest seq */
 678                         if (found)
 679                                 BUG_ON(found->seq > cur->seq);
 680                         found = cur;
 681                         node = node->rb_left;
 682                 } else if (cur->seq > min_seq) {
 683                         /* we want the node with the smallest seq */
 684                         if (found)
 685                                 BUG_ON(found->seq < cur->seq);
 686                         found = cur;
 687                         node = node->rb_right;
 688                 } else {
 689                         found = cur;
 690                         break;
 691                 }
 692         }
 693         read_unlock(&fs_info->tree_mod_log_lock);
 694
 695         return found;
 696 }
 697
 698 /*
 699  * this returns the element from the log with the smallest time sequence
 700  * value that's in the log (the oldest log item). any element with a time
 701  * sequence lower than min_seq will be ignored.
 702  */
 703 static struct tree_mod_elem *
 704 tree_mod_log_search_oldest(struct btrfs_fs_info *fs_info, u64 start,
 705                            u64 min_seq)
 706 {
 707         return __tree_mod_log_search(fs_info, start, min_seq, 1);
 708 }
 709
 710 /*
 711  * this returns the element from the log with the largest time sequence
 712  * value that's in the log (the most recent log item). any element with
 713  * a time sequence lower than min_seq will be ignored.
 714  */
 715 static struct tree_mod_elem *
 716 tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq)
 717 {
 718         return __tree_mod_log_search(fs_info, start, min_seq, 0);
 719 }
 720
 721 static noinline int
 722 tree_mod_log_eb_copy(struct btrfs_fs_info *fs_info, struct extent_buffer *dst,
 723                      struct extent_buffer *src, unsigned long dst_offset,
 724                      unsigned long src_offset, int nr_items)
 725 {
 726         int ret = 0;
 727         struct tree_mod_elem **tm_list = NULL;
 728         struct tree_mod_elem **tm_list_add, **tm_list_rem;
 729         int i;
 730         int locked = 0;
 731
 732         if (!tree_mod_need_log(fs_info, NULL))
 733                 return 0;
 734
 735         if (btrfs_header_level(dst) == 0 && btrfs_header_level(src) == 0)
 736                 return 0;
 737
 738         tm_list = kcalloc(nr_items * 2, sizeof(struct tree_mod_elem *),
 739                           GFP_NOFS);
 740         if (!tm_list)
 741                 return -ENOMEM;
 742
 743         tm_list_add = tm_list;
 744         tm_list_rem = tm_list + nr_items;
 745         for (i = 0; i < nr_items; i++) {
 746                 tm_list_rem[i] = alloc_tree_mod_elem(src, i + src_offset,
 747                     MOD_LOG_KEY_REMOVE, GFP_NOFS);
 748                 if (!tm_list_rem[i]) {
 749                         ret = -ENOMEM;
 750                         goto free_tms;
 751                 }
 752
 753                 tm_list_add[i] = alloc_tree_mod_elem(dst, i + dst_offset,
 754                     MOD_LOG_KEY_ADD, GFP_NOFS);
 755                 if (!tm_list_add[i]) {
 756                         ret = -ENOMEM;
 757                         goto free_tms;
 758                 }
 759         }
 760
 761         if (tree_mod_dont_log(fs_info, NULL))
 762                 goto free_tms;
 763         locked = 1;
 764
 765         for (i = 0; i < nr_items; i++) {
 766                 ret = __tree_mod_log_insert(fs_info, tm_list_rem[i]);
 767                 if (ret)
 768                         goto free_tms;
 769                 ret = __tree_mod_log_insert(fs_info, tm_list_add[i]);
 770                 if (ret)
 771                         goto free_tms;
 772         }
 773
 774         write_unlock(&fs_info->tree_mod_log_lock);
 775         kfree(tm_list);
 776
 777         return 0;
 778
 779 free_tms:
 780         for (i = 0; i < nr_items * 2; i++) {
 781                 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
 782                         rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
 783                 kfree(tm_list[i]);
 784         }
 785         if (locked)
 786                 write_unlock(&fs_info->tree_mod_log_lock);
 787         kfree(tm_list);
 788
 789         return ret;
 790 }
 791
 792 static noinline int tree_mod_log_free_eb(struct extent_buffer *eb)
 793 {
 794         struct tree_mod_elem **tm_list = NULL;
 795         int nritems = 0;
 796         int i;
 797         int ret = 0;
 798
 799         if (btrfs_header_level(eb) == 0)
 800                 return 0;
 801
 802         if (!tree_mod_need_log(eb->fs_info, NULL))
 803                 return 0;
 804
 805         nritems = btrfs_header_nritems(eb);
 806         tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *), GFP_NOFS);
 807         if (!tm_list)
 808                 return -ENOMEM;
 809
 810         for (i = 0; i < nritems; i++) {
 811                 tm_list[i] = alloc_tree_mod_elem(eb, i,
 812                     MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
 813                 if (!tm_list[i]) {
 814                         ret = -ENOMEM;
 815                         goto free_tms;
 816                 }
 817         }
 818
 819         if (tree_mod_dont_log(eb->fs_info, eb))
 820                 goto free_tms;
 821
 822         ret = __tree_mod_log_free_eb(eb->fs_info, tm_list, nritems);
 823         write_unlock(&eb->fs_info->tree_mod_log_lock);
 824         if (ret)
 825                 goto free_tms;
 826         kfree(tm_list);
 827
 828         return 0;
 829
 830 free_tms:
 831         for (i = 0; i < nritems; i++)
 832                 kfree(tm_list[i]);
 833         kfree(tm_list);
 834
 835         return ret;
 836 }
 837
 838 /*
 839  * check if the tree block can be shared by multiple trees
 840  */
 841 int btrfs_block_can_be_shared(struct btrfs_root *root,
 842                               struct extent_buffer *buf)
 843 {
 844         /*
 845          * Tree blocks not in reference counted trees and tree roots
 846          * are never shared. If a block was allocated after the last
 847          * snapshot and the block was not allocated by tree relocation,
 848          * we know the block is not shared.
 849          */
 850         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 851             buf != root->node && buf != root->commit_root &&
 852             (btrfs_header_generation(buf) <=
 853              btrfs_root_last_snapshot(&root->root_item) ||
 854              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 855                 return 1;
 856
 857         return 0;
 858 }
 859
 860 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
 861                                        struct btrfs_root *root,
 862                                        struct extent_buffer *buf,
 863                                        struct extent_buffer *cow,
 864                                        int *last_ref)
 865 {
 866         struct btrfs_fs_info *fs_info = root->fs_info;
 867         u64 refs;
 868         u64 owner;
 869         u64 flags;
 870         u64 new_flags = 0;
 871         int ret;
 872
 873         /*
 874          * Backrefs update rules:
 875          *
 876          * Always use full backrefs for extent pointers in tree block
 877          * allocated by tree relocation.
 878          *
 879          * If a shared tree block is no longer referenced by its owner
 880          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
 881          * use full backrefs for extent pointers in tree block.
 882          *
 883          * If a tree block is been relocating
 884          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
 885          * use full backrefs for extent pointers in tree block.
 886          * The reason for this is some operations (such as drop tree)
 887          * are only allowed for blocks use full backrefs.
 888          */
 889
 890         if (btrfs_block_can_be_shared(root, buf)) {
 891                 ret = btrfs_lookup_extent_info(trans, fs_info, buf->start,
 892                                                btrfs_header_level(buf), 1,
 893                                                &refs, &flags);
 894                 if (ret)
 895                         return ret;
 896                 if (refs == 0) {
 897                         ret = -EROFS;
 898                         btrfs_handle_fs_error(fs_info, ret, NULL);
 899                         return ret;
 900                 }
 901         } else {
 902                 refs = 1;
 903                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 904                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 905                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 906                 else
 907                         flags = 0;
 908         }
 909
 910         owner = btrfs_header_owner(buf);
 911         BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
 912                !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
 913
 914         if (refs > 1) {
 915                 if ((owner == root->root_key.objectid ||
 916                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
 917                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
 918                         ret = btrfs_inc_ref(trans, root, buf, 1);
 919                         if (ret)
 920                                 return ret;
 921
 922                         if (root->root_key.objectid ==
 923                             BTRFS_TREE_RELOC_OBJECTID) {
 924                                 ret = btrfs_dec_ref(trans, root, buf, 0);
 925                                 if (ret)
 926                                         return ret;
 927                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 928                                 if (ret)
 929                                         return ret;
 930                         }
 931                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
 932                 } else {
 933
 934                         if (root->root_key.objectid ==
 935                             BTRFS_TREE_RELOC_OBJECTID)
 936                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 937                         else
 938                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 939                         if (ret)
 940                                 return ret;
 941                 }
 942                 if (new_flags != 0) {
 943                         int level = btrfs_header_level(buf);
 944
 945                         ret = btrfs_set_disk_extent_flags(trans, fs_info,
 946                                                           buf->start,
 947                                                           buf->len,
 948                                                           new_flags, level, 0);
 949                         if (ret)
 950                                 return ret;
 951                 }
 952         } else {
 953                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
 954                         if (root->root_key.objectid ==
 955                             BTRFS_TREE_RELOC_OBJECTID)
 956                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 957                         else
 958                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 959                         if (ret)
 960                                 return ret;
 961                         ret = btrfs_dec_ref(trans, root, buf, 1);
 962                         if (ret)
 963                                 return ret;
 964                 }
 965                 clean_tree_block(fs_info, buf);
 966                 *last_ref = 1;
 967         }
 968         return 0;
 969 }
 970
 971 /*
 972  * does the dirty work in cow of a single block.  The parent block (if
 973  * supplied) is updated to point to the new cow copy.  The new buffer is marked
 974  * dirty and returned locked.  If you modify the block it needs to be marked
 975  * dirty again.
 976  *
 977  * search_start -- an allocation hint for the new block
 978  *
 979  * empty_size -- a hint that you plan on doing more cow.  This is the size in
 980  * bytes the allocator should try to find free next to the block it returns.
 981  * This is just a hint and may be ignored by the allocator.
 982  */
 983 static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 984                              struct btrfs_root *root,
 985                              struct extent_buffer *buf,
 986                              struct extent_buffer *parent, int parent_slot,
 987                              struct extent_buffer **cow_ret,
 988                              u64 search_start, u64 empty_size)
 989 {
 990         struct btrfs_fs_info *fs_info = root->fs_info;
 991         struct btrfs_disk_key disk_key;
 992         struct extent_buffer *cow;
 993         int level, ret;
 994         int last_ref = 0;
 995         int unlock_orig = 0;
 996         u64 parent_start = 0;
 997
 998         if (*cow_ret == buf)
 999                 unlock_orig = 1;
1000
1001         btrfs_assert_tree_locked(buf);
1002
1003         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1004                 trans->transid != fs_info->running_transaction->transid);
1005         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1006                 trans->transid != root->last_trans);
1007
1008         level = btrfs_header_level(buf);
1009
1010         if (level == 0)
1011                 btrfs_item_key(buf, &disk_key, 0);
1012         else
1013                 btrfs_node_key(buf, &disk_key, 0);
1014
1015         if ((root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) && parent)
1016                 parent_start = parent->start;
1017
1018         /*
1019          * If we are COWing a node/leaf from the extent, chunk or device trees,
1020          * make sure that we do not finish block group creation of pending block
1021          * groups. We do this to avoid a deadlock.
1022          * COWing can result in allocation of a new chunk, and flushing pending
1023          * block groups (btrfs_create_pending_block_groups()) can be triggered
1024          * when finishing allocation of a new chunk. Creation of a pending block
1025          * group modifies the extent, chunk and device trees, therefore we could
1026          * deadlock with ourselves since we are holding a lock on an extent
1027          * buffer that btrfs_create_pending_block_groups() may try to COW later.
1028          */
1029         if (root == fs_info->extent_root ||
1030             root == fs_info->chunk_root ||
1031             root == fs_info->dev_root)
1032                 trans->can_flush_pending_bgs = false;
1033
1034         cow = btrfs_alloc_tree_block(trans, root, parent_start,
1035                         root->root_key.objectid, &disk_key, level,
1036                         search_start, empty_size);
1037         trans->can_flush_pending_bgs = true;
1038         if (IS_ERR(cow))
1039                 return PTR_ERR(cow);
1040
1041         /* cow is set to blocking by btrfs_init_new_buffer */
1042
1043         copy_extent_buffer_full(cow, buf);
1044         btrfs_set_header_bytenr(cow, cow->start);
1045         btrfs_set_header_generation(cow, trans->transid);
1046         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
1047         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
1048                                      BTRFS_HEADER_FLAG_RELOC);
1049         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1050                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
1051         else
1052                 btrfs_set_header_owner(cow, root->root_key.objectid);
1053
1054         write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
1055
1056         ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
1057         if (ret) {
1058                 btrfs_abort_transaction(trans, ret);
1059                 return ret;
1060         }
1061
1062         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
1063                 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
1064                 if (ret) {
1065                         btrfs_abort_transaction(trans, ret);
1066                         return ret;
1067                 }
1068         }
1069
1070         if (buf == root->node) {
1071                 WARN_ON(parent && parent != buf);
1072                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1073                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1074                         parent_start = buf->start;
1075
1076                 extent_buffer_get(cow);
1077                 ret = tree_mod_log_insert_root(root->node, cow, 1);
1078                 BUG_ON(ret < 0);
1079                 rcu_assign_pointer(root->node, cow);
1080
1081                 btrfs_free_tree_block(trans, root, buf, parent_start,
1082                                       last_ref);
1083                 free_extent_buffer(buf);
1084                 add_root_to_dirty_list(root);
1085         } else {
1086                 WARN_ON(trans->transid != btrfs_header_generation(parent));
1087                 tree_mod_log_insert_key(parent, parent_slot,
1088                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
1089                 btrfs_set_node_blockptr(parent, parent_slot,
1090                                         cow->start);
1091                 btrfs_set_node_ptr_generation(parent, parent_slot,
1092                                               trans->transid);
1093                 btrfs_mark_buffer_dirty(parent);
1094                 if (last_ref) {
1095                         ret = tree_mod_log_free_eb(buf);
1096                         if (ret) {
1097                                 btrfs_abort_transaction(trans, ret);
1098                                 return ret;
1099                         }
1100                 }
1101                 btrfs_free_tree_block(trans, root, buf, parent_start,
1102                                       last_ref);
1103         }
1104         if (unlock_orig)
1105                 btrfs_tree_unlock(buf);
1106         free_extent_buffer_stale(buf);
1107         btrfs_mark_buffer_dirty(cow);
1108         *cow_ret = cow;
1109         return 0;
1110 }
1111
1112 /*
1113  * returns the logical address of the oldest predecessor of the given root.
1114  * entries older than time_seq are ignored.
1115  */
1116 static struct tree_mod_elem *__tree_mod_log_oldest_root(
1117                 struct extent_buffer *eb_root, u64 time_seq)
1118 {
1119         struct tree_mod_elem *tm;
1120         struct tree_mod_elem *found = NULL;
1121         u64 root_logical = eb_root->start;
1122         int looped = 0;
1123
1124         if (!time_seq)
1125                 return NULL;
1126
1127         /*
1128          * the very last operation that's logged for a root is the
1129          * replacement operation (if it is replaced at all). this has
1130          * the logical address of the *new* root, making it the very
1131          * first operation that's logged for this root.
1132          */
1133         while (1) {
1134                 tm = tree_mod_log_search_oldest(eb_root->fs_info, root_logical,
1135                                                 time_seq);
1136                 if (!looped && !tm)
1137                         return NULL;
1138                 /*
1139                  * if there are no tree operation for the oldest root, we simply
1140                  * return it. this should only happen if that (old) root is at
1141                  * level 0.
1142                  */
1143                 if (!tm)
1144                         break;
1145
1146                 /*
1147                  * if there's an operation that's not a root replacement, we
1148                  * found the oldest version of our root. normally, we'll find a
1149                  * MOD_LOG_KEY_REMOVE_WHILE_FREEING operation here.
1150                  */
1151                 if (tm->op != MOD_LOG_ROOT_REPLACE)
1152                         break;
1153
1154                 found = tm;
1155                 root_logical = tm->old_root.logical;
1156                 looped = 1;
1157         }
1158
1159         /* if there's no old root to return, return what we found instead */
1160         if (!found)
1161                 found = tm;
1162
1163         return found;
1164 }
1165
1166 /*
1167  * tm is a pointer to the first operation to rewind within eb. then, all
1168  * previous operations will be rewound (until we reach something older than
1169  * time_seq).
1170  */
1171 static void
1172 __tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct extent_buffer *eb,
1173                       u64 time_seq, struct tree_mod_elem *first_tm)
1174 {
1175         u32 n;
1176         struct rb_node *next;
1177         struct tree_mod_elem *tm = first_tm;
1178         unsigned long o_dst;
1179         unsigned long o_src;
1180         unsigned long p_size = sizeof(struct btrfs_key_ptr);
1181
1182         n = btrfs_header_nritems(eb);
1183         read_lock(&fs_info->tree_mod_log_lock);
1184         while (tm && tm->seq >= time_seq) {
1185                 /*
1186                  * all the operations are recorded with the operator used for
1187                  * the modification. as we're going backwards, we do the
1188                  * opposite of each operation here.
1189                  */
1190                 switch (tm->op) {
1191                 case MOD_LOG_KEY_REMOVE_WHILE_FREEING:
1192                         BUG_ON(tm->slot < n);
1193                         /* Fallthrough */
1194                 case MOD_LOG_KEY_REMOVE_WHILE_MOVING:
1195                 case MOD_LOG_KEY_REMOVE:
1196                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1197                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1198                         btrfs_set_node_ptr_generation(eb, tm->slot,
1199                                                       tm->generation);
1200                         n++;
1201                         break;
1202                 case MOD_LOG_KEY_REPLACE:
1203                         BUG_ON(tm->slot >= n);
1204                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1205                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1206                         btrfs_set_node_ptr_generation(eb, tm->slot,
1207                                                       tm->generation);
1208                         break;
1209                 case MOD_LOG_KEY_ADD:
1210                         /* if a move operation is needed it's in the log */
1211                         n--;
1212                         break;
1213                 case MOD_LOG_MOVE_KEYS:
1214                         o_dst = btrfs_node_key_ptr_offset(tm->slot);
1215                         o_src = btrfs_node_key_ptr_offset(tm->move.dst_slot);
1216                         memmove_extent_buffer(eb, o_dst, o_src,
1217                                               tm->move.nr_items * p_size);
1218                         break;
1219                 case MOD_LOG_ROOT_REPLACE:
1220                         /*
1221                          * this operation is special. for roots, this must be
1222                          * handled explicitly before rewinding.
1223                          * for non-roots, this operation may exist if the node
1224                          * was a root: root A -> child B; then A gets empty and
1225                          * B is promoted to the new root. in the mod log, we'll
1226                          * have a root-replace operation for B, a tree block
1227                          * that is no root. we simply ignore that operation.
1228                          */
1229                         break;
1230                 }
1231                 next = rb_next(&tm->node);
1232                 if (!next)
1233                         break;
1234                 tm = rb_entry(next, struct tree_mod_elem, node);
1235                 if (tm->logical != first_tm->logical)
1236                         break;
1237         }
1238         read_unlock(&fs_info->tree_mod_log_lock);
1239         btrfs_set_header_nritems(eb, n);
1240 }
1241
1242 /*
1243  * Called with eb read locked. If the buffer cannot be rewound, the same buffer
1244  * is returned. If rewind operations happen, a fresh buffer is returned. The
1245  * returned buffer is always read-locked. If the returned buffer is not the
1246  * input buffer, the lock on the input buffer is released and the input buffer
1247  * is freed (its refcount is decremented).
1248  */
1249 static struct extent_buffer *
1250 tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
1251                     struct extent_buffer *eb, u64 time_seq)
1252 {
1253         struct extent_buffer *eb_rewin;
1254         struct tree_mod_elem *tm;
1255
1256         if (!time_seq)
1257                 return eb;
1258
1259         if (btrfs_header_level(eb) == 0)
1260                 return eb;
1261
1262         tm = tree_mod_log_search(fs_info, eb->start, time_seq);
1263         if (!tm)
1264                 return eb;
1265
1266         btrfs_set_path_blocking(path);
1267         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1268
1269         if (tm->op == MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1270                 BUG_ON(tm->slot != 0);
1271                 eb_rewin = alloc_dummy_extent_buffer(fs_info, eb->start);
1272                 if (!eb_rewin) {
1273                         btrfs_tree_read_unlock_blocking(eb);
1274                         free_extent_buffer(eb);
1275                         return NULL;
1276                 }
1277                 btrfs_set_header_bytenr(eb_rewin, eb->start);
1278                 btrfs_set_header_backref_rev(eb_rewin,
1279                                              btrfs_header_backref_rev(eb));
1280                 btrfs_set_header_owner(eb_rewin, btrfs_header_owner(eb));
1281                 btrfs_set_header_level(eb_rewin, btrfs_header_level(eb));
1282         } else {
1283                 eb_rewin = btrfs_clone_extent_buffer(eb);
1284                 if (!eb_rewin) {
1285                         btrfs_tree_read_unlock_blocking(eb);
1286                         free_extent_buffer(eb);
1287                         return NULL;
1288                 }
1289         }
1290
1291         btrfs_tree_read_unlock_blocking(eb);
1292         free_extent_buffer(eb);
1293
1294         btrfs_tree_read_lock(eb_rewin);
1295         __tree_mod_log_rewind(fs_info, eb_rewin, time_seq, tm);
1296         WARN_ON(btrfs_header_nritems(eb_rewin) >
1297                 BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1298
1299         return eb_rewin;
1300 }
1301
1302 /*
1303  * get_old_root() rewinds the state of @root's root node to the given @time_seq
1304  * value. If there are no changes, the current root->root_node is returned. If
1305  * anything changed in between, there's a fresh buffer allocated on which the
1306  * rewind operations are done. In any case, the returned buffer is read locked.
1307  * Returns NULL on error (with no locks held).
1308  */
1309 static inline struct extent_buffer *
1310 get_old_root(struct btrfs_root *root, u64 time_seq)
1311 {
1312         struct btrfs_fs_info *fs_info = root->fs_info;
1313         struct tree_mod_elem *tm;
1314         struct extent_buffer *eb = NULL;
1315         struct extent_buffer *eb_root;
1316         struct extent_buffer *old;
1317         struct tree_mod_root *old_root = NULL;
1318         u64 old_generation = 0;
1319         u64 logical;
1320         int level;
1321
1322         eb_root = btrfs_read_lock_root_node(root);
1323         tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1324         if (!tm)
1325                 return eb_root;
1326
1327         if (tm->op == MOD_LOG_ROOT_REPLACE) {
1328                 old_root = &tm->old_root;
1329                 old_generation = tm->generation;
1330                 logical = old_root->logical;
1331                 level = old_root->level;
1332         } else {
1333                 logical = eb_root->start;
1334                 level = btrfs_header_level(eb_root);
1335         }
1336
1337         tm = tree_mod_log_search(fs_info, logical, time_seq);
1338         if (old_root && tm && tm->op != MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1339                 btrfs_tree_read_unlock(eb_root);
1340                 free_extent_buffer(eb_root);
1341                 old = read_tree_block(fs_info, logical, 0, level, NULL);
1342                 if (WARN_ON(IS_ERR(old) || !extent_buffer_uptodate(old))) {
1343                         if (!IS_ERR(old))
1344                                 free_extent_buffer(old);
1345                         btrfs_warn(fs_info,
1346                                    "failed to read tree block %llu from get_old_root",
1347                                    logical);
1348                 } else {
1349                         eb = btrfs_clone_extent_buffer(old);
1350                         free_extent_buffer(old);
1351                 }
1352         } else if (old_root) {
1353                 btrfs_tree_read_unlock(eb_root);
1354                 free_extent_buffer(eb_root);
1355                 eb = alloc_dummy_extent_buffer(fs_info, logical);
1356         } else {
1357                 btrfs_set_lock_blocking_rw(eb_root, BTRFS_READ_LOCK);
1358                 eb = btrfs_clone_extent_buffer(eb_root);
1359                 btrfs_tree_read_unlock_blocking(eb_root);
1360                 free_extent_buffer(eb_root);
1361         }
1362
1363         if (!eb)
1364                 return NULL;
1365         btrfs_tree_read_lock(eb);
1366         if (old_root) {
1367                 btrfs_set_header_bytenr(eb, eb->start);
1368                 btrfs_set_header_backref_rev(eb, BTRFS_MIXED_BACKREF_REV);
1369                 btrfs_set_header_owner(eb, btrfs_header_owner(eb_root));
1370                 btrfs_set_header_level(eb, old_root->level);
1371                 btrfs_set_header_generation(eb, old_generation);
1372         }
1373         if (tm)
1374                 __tree_mod_log_rewind(fs_info, eb, time_seq, tm);
1375         else
1376                 WARN_ON(btrfs_header_level(eb) != 0);
1377         WARN_ON(btrfs_header_nritems(eb) > BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1378
1379         return eb;
1380 }
1381
1382 int btrfs_old_root_level(struct btrfs_root *root, u64 time_seq)
1383 {
1384         struct tree_mod_elem *tm;
1385         int level;
1386         struct extent_buffer *eb_root = btrfs_root_node(root);
1387
1388         tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1389         if (tm && tm->op == MOD_LOG_ROOT_REPLACE) {
1390                 level = tm->old_root.level;
1391         } else {
1392                 level = btrfs_header_level(eb_root);
1393         }
1394         free_extent_buffer(eb_root);
1395
1396         return level;
1397 }
1398
1399 static inline int should_cow_block(struct btrfs_trans_handle *trans,
1400                                    struct btrfs_root *root,
1401                                    struct extent_buffer *buf)
1402 {
1403         if (btrfs_is_testing(root->fs_info))
1404                 return 0;
1405
1406         /* Ensure we can see the FORCE_COW bit */
1407         smp_mb__before_atomic();
1408
1409         /*
1410          * We do not need to cow a block if
1411          * 1) this block is not created or changed in this transaction;
1412          * 2) this block does not belong to TREE_RELOC tree;
1413          * 3) the root is not forced COW.
1414          *
1415          * What is forced COW:
1416          *    when we create snapshot during committing the transaction,
1417          *    after we've finished coping src root, we must COW the shared
1418          *    block to ensure the metadata consistency.
1419          */
1420         if (btrfs_header_generation(buf) == trans->transid &&
1421             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
1422             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
1423               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
1424             !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
1425                 return 0;
1426         return 1;
1427 }
1428
1429 /*
1430  * cows a single block, see __btrfs_cow_block for the real work.
1431  * This version of it has extra checks so that a block isn't COWed more than
1432  * once per transaction, as long as it hasn't been written yet
1433  */
1434 noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
1435                     struct btrfs_root *root, struct extent_buffer *buf,
1436                     struct extent_buffer *parent, int parent_slot,
1437                     struct extent_buffer **cow_ret)
1438 {
1439         struct btrfs_fs_info *fs_info = root->fs_info;
1440         u64 search_start;
1441         int ret;
1442
1443         if (trans->transaction != fs_info->running_transaction)
1444                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1445                        trans->transid,
1446                        fs_info->running_transaction->transid);
1447
1448         if (trans->transid != fs_info->generation)
1449                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1450                        trans->transid, fs_info->generation);
1451
1452         if (!should_cow_block(trans, root, buf)) {
1453                 trans->dirty = true;
1454                 *cow_ret = buf;
1455                 return 0;
1456         }
1457
1458         search_start = buf->start & ~((u64)SZ_1G - 1);
1459
1460         if (parent)
1461                 btrfs_set_lock_blocking(parent);
1462         btrfs_set_lock_blocking(buf);
1463
1464         ret = __btrfs_cow_block(trans, root, buf, parent,
1465                                  parent_slot, cow_ret, search_start, 0);
1466
1467         trace_btrfs_cow_block(root, buf, *cow_ret);
1468
1469         return ret;
1470 }
1471
1472 /*
1473  * helper function for defrag to decide if two blocks pointed to by a
1474  * node are actually close by
1475  */
1476 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
1477 {
1478         if (blocknr < other && other - (blocknr + blocksize) < 32768)
1479                 return 1;
1480         if (blocknr > other && blocknr - (other + blocksize) < 32768)
1481                 return 1;
1482         return 0;
1483 }
1484
1485 /*
1486  * compare two keys in a memcmp fashion
1487  */
1488 static int comp_keys(const struct btrfs_disk_key *disk,
1489                      const struct btrfs_key *k2)
1490 {
1491         struct btrfs_key k1;
1492
1493         btrfs_disk_key_to_cpu(&k1, disk);
1494
1495         return btrfs_comp_cpu_keys(&k1, k2);
1496 }
1497
1498 /*
1499  * same as comp_keys only with two btrfs_key's
1500  */
1501 int btrfs_comp_cpu_keys(const struct btrfs_key *k1, const struct btrfs_key *k2)
1502 {
1503         if (k1->objectid > k2->objectid)
1504                 return 1;
1505         if (k1->objectid < k2->objectid)
1506                 return -1;
1507         if (k1->type > k2->type)
1508                 return 1;
1509         if (k1->type < k2->type)
1510                 return -1;
1511         if (k1->offset > k2->offset)
1512                 return 1;
1513         if (k1->offset < k2->offset)
1514                 return -1;
1515         return 0;
1516 }
1517
1518 /*
1519  * this is used by the defrag code to go through all the
1520  * leaves pointed to by a node and reallocate them so that
1521  * disk order is close to key order
1522  */
1523 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
1524                        struct btrfs_root *root, struct extent_buffer *parent,
1525                        int start_slot, u64 *last_ret,
1526                        struct btrfs_key *progress)
1527 {
1528         struct btrfs_fs_info *fs_info = root->fs_info;
1529         struct extent_buffer *cur;
1530         u64 blocknr;
1531         u64 gen;
1532         u64 search_start = *last_ret;
1533         u64 last_block = 0;
1534         u64 other;
1535         u32 parent_nritems;
1536         int end_slot;
1537         int i;
1538         int err = 0;
1539         int parent_level;
1540         int uptodate;
1541         u32 blocksize;
1542         int progress_passed = 0;
1543         struct btrfs_disk_key disk_key;
1544
1545         parent_level = btrfs_header_level(parent);
1546
1547         WARN_ON(trans->transaction != fs_info->running_transaction);
1548         WARN_ON(trans->transid != fs_info->generation);
1549
1550         parent_nritems = btrfs_header_nritems(parent);
1551         blocksize = fs_info->nodesize;
1552         end_slot = parent_nritems - 1;
1553
1554         if (parent_nritems <= 1)
1555                 return 0;
1556
1557         btrfs_set_lock_blocking(parent);
1558
1559         for (i = start_slot; i <= end_slot; i++) {
1560                 struct btrfs_key first_key;
1561                 int close = 1;
1562
1563                 btrfs_node_key(parent, &disk_key, i);
1564                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
1565                         continue;
1566
1567                 progress_passed = 1;
1568                 blocknr = btrfs_node_blockptr(parent, i);
1569                 gen = btrfs_node_ptr_generation(parent, i);
1570                 btrfs_node_key_to_cpu(parent, &first_key, i);
1571                 if (last_block == 0)
1572                         last_block = blocknr;
1573
1574                 if (i > 0) {
1575                         other = btrfs_node_blockptr(parent, i - 1);
1576                         close = close_blocks(blocknr, other, blocksize);
1577                 }
1578                 if (!close && i < end_slot) {
1579                         other = btrfs_node_blockptr(parent, i + 1);
1580                         close = close_blocks(blocknr, other, blocksize);
1581                 }
1582                 if (close) {
1583                         last_block = blocknr;
1584                         continue;
1585                 }
1586
1587                 cur = find_extent_buffer(fs_info, blocknr);
1588                 if (cur)
1589                         uptodate = btrfs_buffer_uptodate(cur, gen, 0);
1590                 else
1591                         uptodate = 0;
1592                 if (!cur || !uptodate) {
1593                         if (!cur) {
1594                                 cur = read_tree_block(fs_info, blocknr, gen,
1595                                                       parent_level - 1,
1596                                                       &first_key);
1597                                 if (IS_ERR(cur)) {
1598                                         return PTR_ERR(cur);
1599                                 } else if (!extent_buffer_uptodate(cur)) {
1600                                         free_extent_buffer(cur);
1601                                         return -EIO;
1602                                 }
1603                         } else if (!uptodate) {
1604                                 err = btrfs_read_buffer(cur, gen,
1605                                                 parent_level - 1,&first_key);
1606                                 if (err) {
1607                                         free_extent_buffer(cur);
1608                                         return err;
1609                                 }
1610                         }
1611                 }
1612                 if (search_start == 0)
1613                         search_start = last_block;
1614
1615                 btrfs_tree_lock(cur);
1616                 btrfs_set_lock_blocking(cur);
1617                 err = __btrfs_cow_block(trans, root, cur, parent, i,
1618                                         &cur, search_start,
1619                                         min(16 * blocksize,
1620                                             (end_slot - i) * blocksize));
1621                 if (err) {
1622                         btrfs_tree_unlock(cur);
1623                         free_extent_buffer(cur);
1624                         break;
1625                 }
1626                 search_start = cur->start;
1627                 last_block = cur->start;
1628                 *last_ret = search_start;
1629                 btrfs_tree_unlock(cur);
1630                 free_extent_buffer(cur);
1631         }
1632         return err;
1633 }
1634
1635 /*
1636  * search for key in the extent_buffer.  The items start at offset p,
1637  * and they are item_size apart.  There are 'max' items in p.
1638  *
1639  * the slot in the array is returned via slot, and it points to
1640  * the place where you would insert key if it is not found in
1641  * the array.
1642  *
1643  * slot may point to max if the key is bigger than all of the keys
1644  */
1645 static noinline int generic_bin_search(struct extent_buffer *eb,
1646                                        unsigned long p, int item_size,
1647                                        const struct btrfs_key *key,
1648                                        int max, int *slot)
1649 {
1650         int low = 0;
1651         int high = max;
1652         int mid;
1653         int ret;
1654         struct btrfs_disk_key *tmp = NULL;
1655         struct btrfs_disk_key unaligned;
1656         unsigned long offset;
1657         char *kaddr = NULL;
1658         unsigned long map_start = 0;
1659         unsigned long map_len = 0;
1660         int err;
1661
1662         if (low > high) {
1663                 btrfs_err(eb->fs_info,
1664                  "%s: low (%d) > high (%d) eb %llu owner %llu level %d",
1665                           __func__, low, high, eb->start,
1666                           btrfs_header_owner(eb), btrfs_header_level(eb));
1667                 return -EINVAL;
1668         }
1669
1670         while (low < high) {
1671                 mid = (low + high) / 2;
1672                 offset = p + mid * item_size;
1673
1674                 if (!kaddr || offset < map_start ||
1675                     (offset + sizeof(struct btrfs_disk_key)) >
1676                     map_start + map_len) {
1677
1678                         err = map_private_extent_buffer(eb, offset,
1679                                                 sizeof(struct btrfs_disk_key),
1680                                                 &kaddr, &map_start, &map_len);
1681
1682                         if (!err) {
1683                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1684                                                         map_start);
1685                         } else if (err == 1) {
1686                                 read_extent_buffer(eb, &unaligned,
1687                                                    offset, sizeof(unaligned));
1688                                 tmp = &unaligned;
1689                         } else {
1690                                 return err;
1691                         }
1692
1693                 } else {
1694                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
1695                                                         map_start);
1696                 }
1697                 ret = comp_keys(tmp, key);
1698
1699                 if (ret < 0)
1700                         low = mid + 1;
1701                 else if (ret > 0)
1702                         high = mid;
1703                 else {
1704                         *slot = mid;
1705                         return 0;
1706                 }
1707         }
1708         *slot = low;
1709         return 1;
1710 }
1711
1712 /*
1713  * simple bin_search frontend that does the right thing for
1714  * leaves vs nodes
1715  */
1716 int btrfs_bin_search(struct extent_buffer *eb, const struct btrfs_key *key,
1717                      int level, int *slot)
1718 {
1719         if (level == 0)
1720                 return generic_bin_search(eb,
1721                                           offsetof(struct btrfs_leaf, items),
1722                                           sizeof(struct btrfs_item),
1723                                           key, btrfs_header_nritems(eb),
1724                                           slot);
1725         else
1726                 return generic_bin_search(eb,
1727                                           offsetof(struct btrfs_node, ptrs),
1728                                           sizeof(struct btrfs_key_ptr),
1729                                           key, btrfs_header_nritems(eb),
1730                                           slot);
1731 }
1732
1733 static void root_add_used(struct btrfs_root *root, u32 size)
1734 {
1735         spin_lock(&root->accounting_lock);
1736         btrfs_set_root_used(&root->root_item,
1737                             btrfs_root_used(&root->root_item) + size);
1738         spin_unlock(&root->accounting_lock);
1739 }
1740
1741 static void root_sub_used(struct btrfs_root *root, u32 size)
1742 {
1743         spin_lock(&root->accounting_lock);
1744         btrfs_set_root_used(&root->root_item,
1745                             btrfs_root_used(&root->root_item) - size);
1746         spin_unlock(&root->accounting_lock);
1747 }
1748
1749 /* given a node and slot number, this reads the blocks it points to.  The
1750  * extent buffer is returned with a reference taken (but unlocked).
1751  */
1752 static noinline struct extent_buffer *
1753 read_node_slot(struct btrfs_fs_info *fs_info, struct extent_buffer *parent,
1754                int slot)
1755 {
1756         int level = btrfs_header_level(parent);
1757         struct extent_buffer *eb;
1758         struct btrfs_key first_key;
1759
1760         if (slot < 0 || slot >= btrfs_header_nritems(parent))
1761                 return ERR_PTR(-ENOENT);
1762
1763         BUG_ON(level == 0);
1764
1765         btrfs_node_key_to_cpu(parent, &first_key, slot);
1766         eb = read_tree_block(fs_info, btrfs_node_blockptr(parent, slot),
1767                              btrfs_node_ptr_generation(parent, slot),
1768                              level - 1, &first_key);
1769         if (!IS_ERR(eb) && !extent_buffer_uptodate(eb)) {
1770                 free_extent_buffer(eb);
1771                 eb = ERR_PTR(-EIO);
1772         }
1773
1774         return eb;
1775 }
1776
1777 /*
1778  * node level balancing, used to make sure nodes are in proper order for
1779  * item deletion.  We balance from the top down, so we have to make sure
1780  * that a deletion won't leave an node completely empty later on.
1781  */
1782 static noinline int balance_level(struct btrfs_trans_handle *trans,
1783                          struct btrfs_root *root,
1784                          struct btrfs_path *path, int level)
1785 {
1786         struct btrfs_fs_info *fs_info = root->fs_info;
1787         struct extent_buffer *right = NULL;
1788         struct extent_buffer *mid;
1789         struct extent_buffer *left = NULL;
1790         struct extent_buffer *parent = NULL;
1791         int ret = 0;
1792         int wret;
1793         int pslot;
1794         int orig_slot = path->slots[level];
1795         u64 orig_ptr;
1796
1797         ASSERT(level > 0);
1798
1799         mid = path->nodes[level];
1800
1801         WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK &&
1802                 path->locks[level] != BTRFS_WRITE_LOCK_BLOCKING);
1803         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1804
1805         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1806
1807         if (level < BTRFS_MAX_LEVEL - 1) {
1808                 parent = path->nodes[level + 1];
1809                 pslot = path->slots[level + 1];
1810         }
1811
1812         /*
1813          * deal with the case where there is only one pointer in the root
1814          * by promoting the node below to a root
1815          */
1816         if (!parent) {
1817                 struct extent_buffer *child;
1818
1819                 if (btrfs_header_nritems(mid) != 1)
1820                         return 0;
1821
1822                 /* promote the child to a root */
1823                 child = read_node_slot(fs_info, mid, 0);
1824                 if (IS_ERR(child)) {
1825                         ret = PTR_ERR(child);
1826                         btrfs_handle_fs_error(fs_info, ret, NULL);
1827                         goto enospc;
1828                 }
1829
1830                 btrfs_tree_lock(child);
1831                 btrfs_set_lock_blocking(child);
1832                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
1833                 if (ret) {
1834                         btrfs_tree_unlock(child);
1835                         free_extent_buffer(child);
1836                         goto enospc;
1837                 }
1838
1839                 ret = tree_mod_log_insert_root(root->node, child, 1);
1840                 BUG_ON(ret < 0);
1841                 rcu_assign_pointer(root->node, child);
1842
1843                 add_root_to_dirty_list(root);
1844                 btrfs_tree_unlock(child);
1845
1846                 path->locks[level] = 0;
1847                 path->nodes[level] = NULL;
1848                 clean_tree_block(fs_info, mid);
1849                 btrfs_tree_unlock(mid);
1850                 /* once for the path */
1851                 free_extent_buffer(mid);
1852
1853                 root_sub_used(root, mid->len);
1854                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1855                 /* once for the root ptr */
1856                 free_extent_buffer_stale(mid);
1857                 return 0;
1858         }
1859         if (btrfs_header_nritems(mid) >
1860             BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 4)
1861                 return 0;
1862
1863         left = read_node_slot(fs_info, parent, pslot - 1);
1864         if (IS_ERR(left))
1865                 left = NULL;
1866
1867         if (left) {
1868                 btrfs_tree_lock(left);
1869                 btrfs_set_lock_blocking(left);
1870                 wret = btrfs_cow_block(trans, root, left,
1871                                        parent, pslot - 1, &left);
1872                 if (wret) {
1873                         ret = wret;
1874                         goto enospc;
1875                 }
1876         }
1877
1878         right = read_node_slot(fs_info, parent, pslot + 1);
1879         if (IS_ERR(right))
1880                 right = NULL;
1881
1882         if (right) {
1883                 btrfs_tree_lock(right);
1884                 btrfs_set_lock_blocking(right);
1885                 wret = btrfs_cow_block(trans, root, right,
1886                                        parent, pslot + 1, &right);
1887                 if (wret) {
1888                         ret = wret;
1889                         goto enospc;
1890                 }
1891         }
1892
1893         /* first, try to make some room in the middle buffer */
1894         if (left) {
1895                 orig_slot += btrfs_header_nritems(left);
1896                 wret = push_node_left(trans, fs_info, left, mid, 1);
1897                 if (wret < 0)
1898                         ret = wret;
1899         }
1900
1901         /*
1902          * then try to empty the right most buffer into the middle
1903          */
1904         if (right) {
1905                 wret = push_node_left(trans, fs_info, mid, right, 1);
1906                 if (wret < 0 && wret != -ENOSPC)
1907                         ret = wret;
1908                 if (btrfs_header_nritems(right) == 0) {
1909                         clean_tree_block(fs_info, right);
1910                         btrfs_tree_unlock(right);
1911                         del_ptr(root, path, level + 1, pslot + 1);
1912                         root_sub_used(root, right->len);
1913                         btrfs_free_tree_block(trans, root, right, 0, 1);
1914                         free_extent_buffer_stale(right);
1915                         right = NULL;
1916                 } else {
1917                         struct btrfs_disk_key right_key;
1918                         btrfs_node_key(right, &right_key, 0);
1919                         ret = tree_mod_log_insert_key(parent, pslot + 1,
1920                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
1921                         BUG_ON(ret < 0);
1922                         btrfs_set_node_key(parent, &right_key, pslot + 1);
1923                         btrfs_mark_buffer_dirty(parent);
1924                 }
1925         }
1926         if (btrfs_header_nritems(mid) == 1) {
1927                 /*
1928                  * we're not allowed to leave a node with one item in the
1929                  * tree during a delete.  A deletion from lower in the tree
1930                  * could try to delete the only pointer in this node.
1931                  * So, pull some keys from the left.
1932                  * There has to be a left pointer at this point because
1933                  * otherwise we would have pulled some pointers from the
1934                  * right
1935                  */
1936                 if (!left) {
1937                         ret = -EROFS;
1938                         btrfs_handle_fs_error(fs_info, ret, NULL);
1939                         goto enospc;
1940                 }
1941                 wret = balance_node_right(trans, fs_info, mid, left);
1942                 if (wret < 0) {
1943                         ret = wret;
1944                         goto enospc;
1945                 }
1946                 if (wret == 1) {
1947                         wret = push_node_left(trans, fs_info, left, mid, 1);
1948                         if (wret < 0)
1949                                 ret = wret;
1950                 }
1951                 BUG_ON(wret == 1);
1952         }
1953         if (btrfs_header_nritems(mid) == 0) {
1954                 clean_tree_block(fs_info, mid);
1955                 btrfs_tree_unlock(mid);
1956                 del_ptr(root, path, level + 1, pslot);
1957                 root_sub_used(root, mid->len);
1958                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1959                 free_extent_buffer_stale(mid);
1960                 mid = NULL;
1961         } else {
1962                 /* update the parent key to reflect our changes */
1963                 struct btrfs_disk_key mid_key;
1964                 btrfs_node_key(mid, &mid_key, 0);
1965                 ret = tree_mod_log_insert_key(parent, pslot,
1966                                 MOD_LOG_KEY_REPLACE, GFP_NOFS);
1967                 BUG_ON(ret < 0);
1968                 btrfs_set_node_key(parent, &mid_key, pslot);
1969                 btrfs_mark_buffer_dirty(parent);
1970         }
1971
1972         /* update the path */
1973         if (left) {
1974                 if (btrfs_header_nritems(left) > orig_slot) {
1975                         extent_buffer_get(left);
1976                         /* left was locked after cow */
1977                         path->nodes[level] = left;
1978                         path->slots[level + 1] -= 1;
1979                         path->slots[level] = orig_slot;
1980                         if (mid) {
1981                                 btrfs_tree_unlock(mid);
1982                                 free_extent_buffer(mid);
1983                         }
1984                 } else {
1985                         orig_slot -= btrfs_header_nritems(left);
1986                         path->slots[level] = orig_slot;
1987                 }
1988         }
1989         /* double check we haven't messed things up */
1990         if (orig_ptr !=
1991             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
1992                 BUG();
1993 enospc:
1994         if (right) {
1995                 btrfs_tree_unlock(right);
1996                 free_extent_buffer(right);
1997         }
1998         if (left) {
1999                 if (path->nodes[level] != left)
2000                         btrfs_tree_unlock(left);
2001                 free_extent_buffer(left);
2002         }
2003         return ret;
2004 }
2005
2006 /* Node balancing for insertion.  Here we only split or push nodes around
2007  * when they are completely full.  This is also done top down, so we
2008  * have to be pessimistic.
2009  */
2010 static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
2011                                           struct btrfs_root *root,
2012                                           struct btrfs_path *path, int level)
2013 {
2014         struct btrfs_fs_info *fs_info = root->fs_info;
2015         struct extent_buffer *right = NULL;
2016         struct extent_buffer *mid;
2017         struct extent_buffer *left = NULL;
2018         struct extent_buffer *parent = NULL;
2019         int ret = 0;
2020         int wret;
2021         int pslot;
2022         int orig_slot = path->slots[level];
2023
2024         if (level == 0)
2025                 return 1;
2026
2027         mid = path->nodes[level];
2028         WARN_ON(btrfs_header_generation(mid) != trans->transid);
2029
2030         if (level < BTRFS_MAX_LEVEL - 1) {
2031                 parent = path->nodes[level + 1];
2032                 pslot = path->slots[level + 1];
2033         }
2034
2035         if (!parent)
2036                 return 1;
2037
2038         left = read_node_slot(fs_info, parent, pslot - 1);
2039         if (IS_ERR(left))
2040                 left = NULL;
2041
2042         /* first, try to make some room in the middle buffer */
2043         if (left) {
2044                 u32 left_nr;
2045
2046                 btrfs_tree_lock(left);
2047                 btrfs_set_lock_blocking(left);
2048
2049                 left_nr = btrfs_header_nritems(left);
2050                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2051                         wret = 1;
2052                 } else {
2053                         ret = btrfs_cow_block(trans, root, left, parent,
2054                                               pslot - 1, &left);
2055                         if (ret)
2056                                 wret = 1;
2057                         else {
2058                                 wret = push_node_left(trans, fs_info,
2059                                                       left, mid, 0);
2060                         }
2061                 }
2062                 if (wret < 0)
2063                         ret = wret;
2064                 if (wret == 0) {
2065                         struct btrfs_disk_key disk_key;
2066                         orig_slot += left_nr;
2067                         btrfs_node_key(mid, &disk_key, 0);
2068                         ret = tree_mod_log_insert_key(parent, pslot,
2069                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
2070                         BUG_ON(ret < 0);
2071                         btrfs_set_node_key(parent, &disk_key, pslot);
2072                         btrfs_mark_buffer_dirty(parent);
2073                         if (btrfs_header_nritems(left) > orig_slot) {
2074                                 path->nodes[level] = left;
2075                                 path->slots[level + 1] -= 1;
2076                                 path->slots[level] = orig_slot;
2077                                 btrfs_tree_unlock(mid);
2078                                 free_extent_buffer(mid);
2079                         } else {
2080                                 orig_slot -=
2081                                         btrfs_header_nritems(left);
2082                                 path->slots[level] = orig_slot;
2083                                 btrfs_tree_unlock(left);
2084                                 free_extent_buffer(left);
2085                         }
2086                         return 0;
2087                 }
2088                 btrfs_tree_unlock(left);
2089                 free_extent_buffer(left);
2090         }
2091         right = read_node_slot(fs_info, parent, pslot + 1);
2092         if (IS_ERR(right))
2093                 right = NULL;
2094
2095         /*
2096          * then try to empty the right most buffer into the middle
2097          */
2098         if (right) {
2099                 u32 right_nr;
2100
2101                 btrfs_tree_lock(right);
2102                 btrfs_set_lock_blocking(right);
2103
2104                 right_nr = btrfs_header_nritems(right);
2105                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2106                         wret = 1;
2107                 } else {
2108                         ret = btrfs_cow_block(trans, root, right,
2109                                               parent, pslot + 1,
2110                                               &right);
2111                         if (ret)
2112                                 wret = 1;
2113                         else {
2114                                 wret = balance_node_right(trans, fs_info,
2115                                                           right, mid);
2116                         }
2117                 }
2118                 if (wret < 0)
2119                         ret = wret;
2120                 if (wret == 0) {
2121                         struct btrfs_disk_key disk_key;
2122
2123                         btrfs_node_key(right, &disk_key, 0);
2124                         ret = tree_mod_log_insert_key(parent, pslot + 1,
2125                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
2126                         BUG_ON(ret < 0);
2127                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
2128                         btrfs_mark_buffer_dirty(parent);
2129
2130                         if (btrfs_header_nritems(mid) <= orig_slot) {
2131                                 path->nodes[level] = right;
2132                                 path->slots[level + 1] += 1;
2133                                 path->slots[level] = orig_slot -
2134                                         btrfs_header_nritems(mid);
2135                                 btrfs_tree_unlock(mid);
2136                                 free_extent_buffer(mid);
2137                         } else {
2138                                 btrfs_tree_unlock(right);
2139                                 free_extent_buffer(right);
2140                         }
2141                         return 0;
2142                 }
2143                 btrfs_tree_unlock(right);
2144                 free_extent_buffer(right);
2145         }
2146         return 1;
2147 }
2148
2149 /*
2150  * readahead one full node of leaves, finding things that are close
2151  * to the block in 'slot', and triggering ra on them.
2152  */
2153 static void reada_for_search(struct btrfs_fs_info *fs_info,
2154                              struct btrfs_path *path,
2155                              int level, int slot, u64 objectid)
2156 {
2157         struct extent_buffer *node;
2158         struct btrfs_disk_key disk_key;
2159         u32 nritems;
2160         u64 search;
2161         u64 target;
2162         u64 nread = 0;
2163         struct extent_buffer *eb;
2164         u32 nr;
2165         u32 blocksize;
2166         u32 nscan = 0;
2167
2168         if (level != 1)
2169                 return;
2170
2171         if (!path->nodes[level])
2172                 return;
2173
2174         node = path->nodes[level];
2175
2176         search = btrfs_node_blockptr(node, slot);
2177         blocksize = fs_info->nodesize;
2178         eb = find_extent_buffer(fs_info, search);
2179         if (eb) {
2180                 free_extent_buffer(eb);
2181                 return;
2182         }
2183
2184         target = search;
2185
2186         nritems = btrfs_header_nritems(node);
2187         nr = slot;
2188
2189         while (1) {
2190                 if (path->reada == READA_BACK) {
2191                         if (nr == 0)
2192                                 break;
2193                         nr--;
2194                 } else if (path->reada == READA_FORWARD) {
2195                         nr++;
2196                         if (nr >= nritems)
2197                                 break;
2198                 }
2199                 if (path->reada == READA_BACK && objectid) {
2200                         btrfs_node_key(node, &disk_key, nr);
2201                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
2202                                 break;
2203                 }
2204                 search = btrfs_node_blockptr(node, nr);
2205                 if ((search <= target && target - search <= 65536) ||
2206                     (search > target && search - target <= 65536)) {
2207                         readahead_tree_block(fs_info, search);
2208                         nread += blocksize;
2209                 }
2210                 nscan++;
2211                 if ((nread > 65536 || nscan > 32))
2212                         break;
2213         }
2214 }
2215
2216 static noinline void reada_for_balance(struct btrfs_fs_info *fs_info,
2217                                        struct btrfs_path *path, int level)
2218 {
2219         int slot;
2220         int nritems;
2221         struct extent_buffer *parent;
2222         struct extent_buffer *eb;
2223         u64 gen;
2224         u64 block1 = 0;
2225         u64 block2 = 0;
2226
2227         parent = path->nodes[level + 1];
2228         if (!parent)
2229                 return;
2230
2231         nritems = btrfs_header_nritems(parent);
2232         slot = path->slots[level + 1];
2233
2234         if (slot > 0) {
2235                 block1 = btrfs_node_blockptr(parent, slot - 1);
2236                 gen = btrfs_node_ptr_generation(parent, slot - 1);
2237                 eb = find_extent_buffer(fs_info, block1);
2238                 /*
2239                  * if we get -eagain from btrfs_buffer_uptodate, we
2240                  * don't want to return eagain here.  That will loop
2241                  * forever
2242                  */
2243                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2244                         block1 = 0;
2245                 free_extent_buffer(eb);
2246         }
2247         if (slot + 1 < nritems) {
2248                 block2 = btrfs_node_blockptr(parent, slot + 1);
2249                 gen = btrfs_node_ptr_generation(parent, slot + 1);
2250                 eb = find_extent_buffer(fs_info, block2);
2251                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2252                         block2 = 0;
2253                 free_extent_buffer(eb);
2254         }
2255
2256         if (block1)
2257                 readahead_tree_block(fs_info, block1);
2258         if (block2)
2259                 readahead_tree_block(fs_info, block2);
2260 }
2261
2262
2263 /*
2264  * when we walk down the tree, it is usually safe to unlock the higher layers
2265  * in the tree.  The exceptions are when our path goes through slot 0, because
2266  * operations on the tree might require changing key pointers higher up in the
2267  * tree.
2268  *
2269  * callers might also have set path->keep_locks, which tells this code to keep
2270  * the lock if the path points to the last slot in the block.  This is part of
2271  * walking through the tree, and selecting the next slot in the higher block.
2272  *
2273  * lowest_unlock sets the lowest level in the tree we're allowed to unlock.  so
2274  * if lowest_unlock is 1, level 0 won't be unlocked
2275  */
2276 static noinline void unlock_up(struct btrfs_path *path, int level,
2277                                int lowest_unlock, int min_write_lock_level,
2278                                int *write_lock_level)
2279 {
2280         int i;
2281         int skip_level = level;
2282         int no_skips = 0;
2283         struct extent_buffer *t;
2284
2285         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2286                 if (!path->nodes[i])
2287                         break;
2288                 if (!path->locks[i])
2289                         break;
2290                 if (!no_skips && path->slots[i] == 0) {
2291                         skip_level = i + 1;
2292                         continue;
2293                 }
2294                 if (!no_skips && path->keep_locks) {
2295                         u32 nritems;
2296                         t = path->nodes[i];
2297                         nritems = btrfs_header_nritems(t);
2298                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
2299                                 skip_level = i + 1;
2300                                 continue;
2301                         }
2302                 }
2303                 if (skip_level < i && i >= lowest_unlock)
2304                         no_skips = 1;
2305
2306                 t = path->nodes[i];
2307                 if (i >= lowest_unlock && i > skip_level) {
2308                         btrfs_tree_unlock_rw(t, path->locks[i]);
2309                         path->locks[i] = 0;
2310                         if (write_lock_level &&
2311                             i > min_write_lock_level &&
2312                             i <= *write_lock_level) {
2313                                 *write_lock_level = i - 1;
2314                         }
2315                 }
2316         }
2317 }
2318
2319 /*
2320  * This releases any locks held in the path starting at level and
2321  * going all the way up to the root.
2322  *
2323  * btrfs_search_slot will keep the lock held on higher nodes in a few
2324  * corner cases, such as COW of the block at slot zero in the node.  This
2325  * ignores those rules, and it should only be called when there are no
2326  * more updates to be done higher up in the tree.
2327  */
2328 noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
2329 {
2330         int i;
2331
2332         if (path->keep_locks)
2333                 return;
2334
2335         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2336                 if (!path->nodes[i])
2337                         continue;
2338                 if (!path->locks[i])
2339                         continue;
2340                 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
2341                 path->locks[i] = 0;
2342         }
2343 }
2344
2345 /*
2346  * helper function for btrfs_search_slot.  The goal is to find a block
2347  * in cache without setting the path to blocking.  If we find the block
2348  * we return zero and the path is unchanged.
2349  *
2350  * If we can't find the block, we set the path blocking and do some
2351  * reada.  -EAGAIN is returned and the search must be repeated.
2352  */
2353 static int
2354 read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
2355                       struct extent_buffer **eb_ret, int level, int slot,
2356                       const struct btrfs_key *key)
2357 {
2358         struct btrfs_fs_info *fs_info = root->fs_info;
2359         u64 blocknr;
2360         u64 gen;
2361         struct extent_buffer *b = *eb_ret;
2362         struct extent_buffer *tmp;
2363         struct btrfs_key first_key;
2364         int ret;
2365         int parent_level;
2366
2367         blocknr = btrfs_node_blockptr(b, slot);
2368         gen = btrfs_node_ptr_generation(b, slot);
2369         parent_level = btrfs_header_level(b);
2370         btrfs_node_key_to_cpu(b, &first_key, slot);
2371
2372         tmp = find_extent_buffer(fs_info, blocknr);
2373         if (tmp) {
2374                 /* first we do an atomic uptodate check */
2375                 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
2376                         *eb_ret = tmp;
2377                         return 0;
2378                 }
2379
2380                 /* the pages were up to date, but we failed
2381                  * the generation number check.  Do a full
2382                  * read for the generation number that is correct.
2383                  * We must do this without dropping locks so
2384                  * we can trust our generation number
2385                  */
2386                 btrfs_set_path_blocking(p);
2387
2388                 /* now we're allowed to do a blocking uptodate check */
2389                 ret = btrfs_read_buffer(tmp, gen, parent_level - 1, &first_key);
2390                 if (!ret) {
2391                         *eb_ret = tmp;
2392                         return 0;
2393                 }
2394                 free_extent_buffer(tmp);
2395                 btrfs_release_path(p);
2396                 return -EIO;
2397         }
2398
2399         /*
2400          * reduce lock contention at high levels
2401          * of the btree by dropping locks before
2402          * we read.  Don't release the lock on the current
2403          * level because we need to walk this node to figure
2404          * out which blocks to read.
2405          */
2406         btrfs_unlock_up_safe(p, level + 1);
2407         btrfs_set_path_blocking(p);
2408
2409         if (p->reada != READA_NONE)
2410                 reada_for_search(fs_info, p, level, slot, key->objectid);
2411
2412         ret = -EAGAIN;
2413         tmp = read_tree_block(fs_info, blocknr, gen, parent_level - 1,
2414                               &first_key);
2415         if (!IS_ERR(tmp)) {
2416                 /*
2417                  * If the read above didn't mark this buffer up to date,
2418                  * it will never end up being up to date.  Set ret to EIO now
2419                  * and give up so that our caller doesn't loop forever
2420                  * on our EAGAINs.
2421                  */
2422                 if (!extent_buffer_uptodate(tmp))
2423                         ret = -EIO;
2424                 free_extent_buffer(tmp);
2425         } else {
2426                 ret = PTR_ERR(tmp);
2427         }
2428
2429         btrfs_release_path(p);
2430         return ret;
2431 }
2432
2433 /*
2434  * helper function for btrfs_search_slot.  This does all of the checks
2435  * for node-level blocks and does any balancing required based on
2436  * the ins_len.
2437  *
2438  * If no extra work was required, zero is returned.  If we had to
2439  * drop the path, -EAGAIN is returned and btrfs_search_slot must
2440  * start over
2441  */
2442 static int
2443 setup_nodes_for_search(struct btrfs_trans_handle *trans,
2444                        struct btrfs_root *root, struct btrfs_path *p,
2445                        struct extent_buffer *b, int level, int ins_len,
2446                        int *write_lock_level)
2447 {
2448         struct btrfs_fs_info *fs_info = root->fs_info;
2449         int ret;
2450
2451         if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
2452             BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3) {
2453                 int sret;
2454
2455                 if (*write_lock_level < level + 1) {
2456                         *write_lock_level = level + 1;
2457                         btrfs_release_path(p);
2458                         goto again;
2459                 }
2460
2461                 btrfs_set_path_blocking(p);
2462                 reada_for_balance(fs_info, p, level);
2463                 sret = split_node(trans, root, p, level);
2464
2465                 BUG_ON(sret > 0);
2466                 if (sret) {
2467                         ret = sret;
2468                         goto done;
2469                 }
2470                 b = p->nodes[level];
2471         } else if (ins_len < 0 && btrfs_header_nritems(b) <
2472                    BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 2) {
2473                 int sret;
2474
2475                 if (*write_lock_level < level + 1) {
2476                         *write_lock_level = level + 1;
2477                         btrfs_release_path(p);
2478                         goto again;
2479                 }
2480
2481                 btrfs_set_path_blocking(p);
2482                 reada_for_balance(fs_info, p, level);
2483                 sret = balance_level(trans, root, p, level);
2484
2485                 if (sret) {
2486                         ret = sret;
2487                         goto done;
2488                 }
2489                 b = p->nodes[level];
2490                 if (!b) {
2491                         btrfs_release_path(p);
2492                         goto again;
2493                 }
2494                 BUG_ON(btrfs_header_nritems(b) == 1);
2495         }
2496         return 0;
2497
2498 again:
2499         ret = -EAGAIN;
2500 done:
2501         return ret;
2502 }
2503
2504 static void key_search_validate(struct extent_buffer *b,
2505                                 const struct btrfs_key *key,
2506                                 int level)
2507 {
2508 #ifdef CONFIG_BTRFS_ASSERT
2509         struct btrfs_disk_key disk_key;
2510
2511         btrfs_cpu_key_to_disk(&disk_key, key);
2512
2513         if (level == 0)
2514                 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2515                     offsetof(struct btrfs_leaf, items[0].key),
2516                     sizeof(disk_key)));
2517         else
2518                 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2519                     offsetof(struct btrfs_node, ptrs[0].key),
2520                     sizeof(disk_key)));
2521 #endif
2522 }
2523
2524 static int key_search(struct extent_buffer *b, const struct btrfs_key *key,
2525                       int level, int *prev_cmp, int *slot)
2526 {
2527         if (*prev_cmp != 0) {
2528                 *prev_cmp = btrfs_bin_search(b, key, level, slot);
2529                 return *prev_cmp;
2530         }
2531
2532         key_search_validate(b, key, level);
2533         *slot = 0;
2534
2535         return 0;
2536 }
2537
2538 int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
2539                 u64 iobjectid, u64 ioff, u8 key_type,
2540                 struct btrfs_key *found_key)
2541 {
2542         int ret;
2543         struct btrfs_key key;
2544         struct extent_buffer *eb;
2545
2546         ASSERT(path);
2547         ASSERT(found_key);
2548
2549         key.type = key_type;
2550         key.objectid = iobjectid;
2551         key.offset = ioff;
2552
2553         ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
2554         if (ret < 0)
2555                 return ret;
2556
2557         eb = path->nodes[0];
2558         if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
2559                 ret = btrfs_next_leaf(fs_root, path);
2560                 if (ret)
2561                         return ret;
2562                 eb = path->nodes[0];
2563         }
2564
2565         btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
2566         if (found_key->type != key.type ||
2567                         found_key->objectid != key.objectid)
2568                 return 1;
2569
2570         return 0;
2571 }
2572
2573 static struct extent_buffer *btrfs_search_slot_get_root(struct btrfs_root *root,
2574                                                         struct btrfs_path *p,
2575                                                         int write_lock_level)
2576 {
2577         struct btrfs_fs_info *fs_info = root->fs_info;
2578         struct extent_buffer *b;
2579         int root_lock;
2580         int level = 0;
2581
2582         /* We try very hard to do read locks on the root */
2583         root_lock = BTRFS_READ_LOCK;
2584
2585         if (p->search_commit_root) {
2586                 /* The commit roots are read only so we always do read locks */
2587                 if (p->need_commit_sem)
2588                         down_read(&fs_info->commit_root_sem);
2589                 b = root->commit_root;
2590                 extent_buffer_get(b);
2591                 level = btrfs_header_level(b);
2592                 if (p->need_commit_sem)
2593                         up_read(&fs_info->commit_root_sem);
2594                 /*
2595                  * Ensure that all callers have set skip_locking when
2596                  * p->search_commit_root = 1.
2597                  */
2598                 ASSERT(p->skip_locking == 1);
2599
2600                 goto out;
2601         }
2602
2603         if (p->skip_locking) {
2604                 b = btrfs_root_node(root);
2605                 level = btrfs_header_level(b);
2606                 goto out;
2607         }
2608
2609         /*
2610          * If the level is set to maximum, we can skip trying to get the read
2611          * lock.
2612          */
2613         if (write_lock_level < BTRFS_MAX_LEVEL) {
2614                 /*
2615                  * We don't know the level of the root node until we actually
2616                  * have it read locked
2617                  */
2618                 b = btrfs_read_lock_root_node(root);
2619                 level = btrfs_header_level(b);
2620                 if (level > write_lock_level)
2621                         goto out;
2622
2623                 /* Whoops, must trade for write lock */
2624                 btrfs_tree_read_unlock(b);
2625                 free_extent_buffer(b);
2626         }
2627
2628         b = btrfs_lock_root_node(root);
2629         root_lock = BTRFS_WRITE_LOCK;
2630
2631         /* The level might have changed, check again */
2632         level = btrfs_header_level(b);
2633
2634 out:
2635         p->nodes[level] = b;
2636         if (!p->skip_locking)
2637                 p->locks[level] = root_lock;
2638         /*
2639          * Callers are responsible for dropping b's references.
2640          */
2641         return b;
2642 }
2643
2644
2645 /*
2646  * btrfs_search_slot - look for a key in a tree and perform necessary
2647  * modifications to preserve tree invariants.
2648  *
2649  * @trans:      Handle of transaction, used when modifying the tree
2650  * @p:          Holds all btree nodes along the search path
2651  * @root:       The root node of the tree
2652  * @key:        The key we are looking for
2653  * @ins_len:    Indicates purpose of search, for inserts it is 1, for
2654  *              deletions it's -1. 0 for plain searches
2655  * @cow:        boolean should CoW operations be performed. Must always be 1
2656  *              when modifying the tree.
2657  *
2658  * If @ins_len > 0, nodes and leaves will be split as we walk down the tree.
2659  * If @ins_len < 0, nodes will be merged as we walk down the tree (if possible)
2660  *
2661  * If @key is found, 0 is returned and you can find the item in the leaf level
2662  * of the path (level 0)
2663  *
2664  * If @key isn't found, 1 is returned and the leaf level of the path (level 0)
2665  * points to the slot where it should be inserted
2666  *
2667  * If an error is encountered while searching the tree a negative error number
2668  * is returned
2669  */
2670 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2671                       const struct btrfs_key *key, struct btrfs_path *p,
2672                       int ins_len, int cow)
2673 {
2674         struct btrfs_fs_info *fs_info = root->fs_info;
2675         struct extent_buffer *b;
2676         int slot;
2677         int ret;
2678         int err;
2679         int level;
2680         int lowest_unlock = 1;
2681         /* everything at write_lock_level or lower must be write locked */
2682         int write_lock_level = 0;
2683         u8 lowest_level = 0;
2684         int min_write_lock_level;
2685         int prev_cmp;
2686
2687         lowest_level = p->lowest_level;
2688         WARN_ON(lowest_level && ins_len > 0);
2689         WARN_ON(p->nodes[0] != NULL);
2690         BUG_ON(!cow && ins_len);
2691
2692         if (ins_len < 0) {
2693                 lowest_unlock = 2;
2694
2695                 /* when we are removing items, we might have to go up to level
2696                  * two as we update tree pointers  Make sure we keep write
2697                  * for those levels as well
2698                  */
2699                 write_lock_level = 2;
2700         } else if (ins_len > 0) {
2701                 /*
2702                  * for inserting items, make sure we have a write lock on
2703                  * level 1 so we can update keys
2704                  */
2705                 write_lock_level = 1;
2706         }
2707
2708         if (!cow)
2709                 write_lock_level = -1;
2710
2711         if (cow && (p->keep_locks || p->lowest_level))
2712                 write_lock_level = BTRFS_MAX_LEVEL;
2713
2714         min_write_lock_level = write_lock_level;
2715
2716 again:
2717         prev_cmp = -1;
2718         b = btrfs_search_slot_get_root(root, p, write_lock_level);
2719
2720         while (b) {
2721                 level = btrfs_header_level(b);
2722
2723                 /*
2724                  * setup the path here so we can release it under lock
2725                  * contention with the cow code
2726                  */
2727                 if (cow) {
2728                         bool last_level = (level == (BTRFS_MAX_LEVEL - 1));
2729
2730                         /*
2731                          * if we don't really need to cow this block
2732                          * then we don't want to set the path blocking,
2733                          * so we test it here
2734                          */
2735                         if (!should_cow_block(trans, root, b)) {
2736                                 trans->dirty = true;
2737                                 goto cow_done;
2738                         }
2739
2740                         /*
2741                          * must have write locks on this node and the
2742                          * parent
2743                          */
2744                         if (level > write_lock_level ||
2745                             (level + 1 > write_lock_level &&
2746                             level + 1 < BTRFS_MAX_LEVEL &&
2747                             p->nodes[level + 1])) {
2748                                 write_lock_level = level + 1;
2749                                 btrfs_release_path(p);
2750                                 goto again;
2751                         }
2752
2753                         btrfs_set_path_blocking(p);
2754                         if (last_level)
2755                                 err = btrfs_cow_block(trans, root, b, NULL, 0,
2756                                                       &b);
2757                         else
2758                                 err = btrfs_cow_block(trans, root, b,
2759                                                       p->nodes[level + 1],
2760                                                       p->slots[level + 1], &b);
2761                         if (err) {
2762                                 ret = err;
2763                                 goto done;
2764                         }
2765                 }
2766 cow_done:
2767                 p->nodes[level] = b;
2768                 /*
2769                  * Leave path with blocking locks to avoid massive
2770                  * lock context switch, this is made on purpose.
2771                  */
2772
2773                 /*
2774                  * we have a lock on b and as long as we aren't changing
2775                  * the tree, there is no way to for the items in b to change.
2776                  * It is safe to drop the lock on our parent before we
2777                  * go through the expensive btree search on b.
2778                  *
2779                  * If we're inserting or deleting (ins_len != 0), then we might
2780                  * be changing slot zero, which may require changing the parent.
2781                  * So, we can't drop the lock until after we know which slot
2782                  * we're operating on.
2783                  */
2784                 if (!ins_len && !p->keep_locks) {
2785                         int u = level + 1;
2786
2787                         if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
2788                                 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
2789                                 p->locks[u] = 0;
2790                         }
2791                 }
2792
2793                 ret = key_search(b, key, level, &prev_cmp, &slot);
2794                 if (ret < 0)
2795                         goto done;
2796
2797                 if (level != 0) {
2798                         int dec = 0;
2799                         if (ret && slot > 0) {
2800                                 dec = 1;
2801                                 slot -= 1;
2802                         }
2803                         p->slots[level] = slot;
2804                         err = setup_nodes_for_search(trans, root, p, b, level,
2805                                              ins_len, &write_lock_level);
2806                         if (err == -EAGAIN)
2807                                 goto again;
2808                         if (err) {
2809                                 ret = err;
2810                                 goto done;
2811                         }
2812                         b = p->nodes[level];
2813                         slot = p->slots[level];
2814
2815                         /*
2816                          * slot 0 is special, if we change the key
2817                          * we have to update the parent pointer
2818                          * which means we must have a write lock
2819                          * on the parent
2820                          */
2821                         if (slot == 0 && ins_len &&
2822                             write_lock_level < level + 1) {
2823                                 write_lock_level = level + 1;
2824                                 btrfs_release_path(p);
2825                                 goto again;
2826                         }
2827
2828                         unlock_up(p, level, lowest_unlock,
2829                                   min_write_lock_level, &write_lock_level);
2830
2831                         if (level == lowest_level) {
2832                                 if (dec)
2833                                         p->slots[level]++;
2834                                 goto done;
2835                         }
2836
2837                         err = read_block_for_search(root, p, &b, level,
2838                                                     slot, key);
2839                         if (err == -EAGAIN)
2840                                 goto again;
2841                         if (err) {
2842                                 ret = err;
2843                                 goto done;
2844                         }
2845
2846                         if (!p->skip_locking) {
2847                                 level = btrfs_header_level(b);
2848                                 if (level <= write_lock_level) {
2849                                         err = btrfs_try_tree_write_lock(b);
2850                                         if (!err) {
2851                                                 btrfs_set_path_blocking(p);
2852                                                 btrfs_tree_lock(b);
2853                                         }
2854                                         p->locks[level] = BTRFS_WRITE_LOCK;
2855                                 } else {
2856                                         err = btrfs_tree_read_lock_atomic(b);
2857                                         if (!err) {
2858                                                 btrfs_set_path_blocking(p);
2859                                                 btrfs_tree_read_lock(b);
2860                                         }
2861                                         p->locks[level] = BTRFS_READ_LOCK;
2862                                 }
2863                                 p->nodes[level] = b;
2864                         }
2865                 } else {
2866                         p->slots[level] = slot;
2867                         if (ins_len > 0 &&
2868                             btrfs_leaf_free_space(fs_info, b) < ins_len) {
2869                                 if (write_lock_level < 1) {
2870                                         write_lock_level = 1;
2871                                         btrfs_release_path(p);
2872                                         goto again;
2873                                 }
2874
2875                                 btrfs_set_path_blocking(p);
2876                                 err = split_leaf(trans, root, key,
2877                                                  p, ins_len, ret == 0);
2878
2879                                 BUG_ON(err > 0);
2880                                 if (err) {
2881                                         ret = err;
2882                                         goto done;
2883                                 }
2884                         }
2885                         if (!p->search_for_split)
2886                                 unlock_up(p, level, lowest_unlock,
2887                                           min_write_lock_level, NULL);
2888                         goto done;
2889                 }
2890         }
2891         ret = 1;
2892 done:
2893         /*
2894          * we don't really know what they plan on doing with the path
2895          * from here on, so for now just mark it as blocking
2896          */
2897         if (!p->leave_spinning)
2898                 btrfs_set_path_blocking(p);
2899         if (ret < 0 && !p->skip_release_on_error)
2900                 btrfs_release_path(p);
2901         return ret;
2902 }
2903
2904 /*
2905  * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2906  * current state of the tree together with the operations recorded in the tree
2907  * modification log to search for the key in a previous version of this tree, as
2908  * denoted by the time_seq parameter.
2909  *
2910  * Naturally, there is no support for insert, delete or cow operations.
2911  *
2912  * The resulting path and return value will be set up as if we called
2913  * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2914  */
2915 int btrfs_search_old_slot(struct btrfs_root *root, const struct btrfs_key *key,
2916                           struct btrfs_path *p, u64 time_seq)
2917 {
2918         struct btrfs_fs_info *fs_info = root->fs_info;
2919         struct extent_buffer *b;
2920         int slot;
2921         int ret;
2922         int err;
2923         int level;
2924         int lowest_unlock = 1;
2925         u8 lowest_level = 0;
2926         int prev_cmp = -1;
2927
2928         lowest_level = p->lowest_level;
2929         WARN_ON(p->nodes[0] != NULL);
2930
2931         if (p->search_commit_root) {
2932                 BUG_ON(time_seq);
2933                 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2934         }
2935
2936 again:
2937         b = get_old_root(root, time_seq);
2938         if (!b) {
2939                 ret = -EIO;
2940                 goto done;
2941         }
2942         level = btrfs_header_level(b);
2943         p->locks[level] = BTRFS_READ_LOCK;
2944
2945         while (b) {
2946                 level = btrfs_header_level(b);
2947                 p->nodes[level] = b;
2948
2949                 /*
2950                  * we have a lock on b and as long as we aren't changing
2951                  * the tree, there is no way to for the items in b to change.
2952                  * It is safe to drop the lock on our parent before we
2953                  * go through the expensive btree search on b.
2954                  */
2955                 btrfs_unlock_up_safe(p, level + 1);
2956
2957                 /*
2958                  * Since we can unwind ebs we want to do a real search every
2959                  * time.
2960                  */
2961                 prev_cmp = -1;
2962                 ret = key_search(b, key, level, &prev_cmp, &slot);
2963
2964                 if (level != 0) {
2965                         int dec = 0;
2966                         if (ret && slot > 0) {
2967                                 dec = 1;
2968                                 slot -= 1;
2969                         }
2970                         p->slots[level] = slot;
2971                         unlock_up(p, level, lowest_unlock, 0, NULL);
2972
2973                         if (level == lowest_level) {
2974                                 if (dec)
2975                                         p->slots[level]++;
2976                                 goto done;
2977                         }
2978
2979                         err = read_block_for_search(root, p, &b, level,
2980                                                     slot, key);
2981                         if (err == -EAGAIN)
2982                                 goto again;
2983                         if (err) {
2984                                 ret = err;
2985                                 goto done;
2986                         }
2987
2988                         level = btrfs_header_level(b);
2989                         err = btrfs_tree_read_lock_atomic(b);
2990                         if (!err) {
2991                                 btrfs_set_path_blocking(p);
2992                                 btrfs_tree_read_lock(b);
2993                         }
2994                         b = tree_mod_log_rewind(fs_info, p, b, time_seq);
2995                         if (!b) {
2996                                 ret = -ENOMEM;
2997                                 goto done;
2998                         }
2999                         p->locks[level] = BTRFS_READ_LOCK;
3000                         p->nodes[level] = b;
3001                 } else {
3002                         p->slots[level] = slot;
3003                         unlock_up(p, level, lowest_unlock, 0, NULL);
3004                         goto done;
3005                 }
3006         }
3007         ret = 1;
3008 done:
3009         if (!p->leave_spinning)
3010                 btrfs_set_path_blocking(p);
3011         if (ret < 0)
3012                 btrfs_release_path(p);
3013
3014         return ret;
3015 }
3016
3017 /*
3018  * helper to use instead of search slot if no exact match is needed but
3019  * instead the next or previous item should be returned.
3020  * When find_higher is true, the next higher item is returned, the next lower
3021  * otherwise.
3022  * When return_any and find_higher are both true, and no higher item is found,
3023  * return the next lower instead.
3024  * When return_any is true and find_higher is false, and no lower item is found,
3025  * return the next higher instead.
3026  * It returns 0 if any item is found, 1 if none is found (tree empty), and
3027  * < 0 on error
3028  */
3029 int btrfs_search_slot_for_read(struct btrfs_root *root,
3030                                const struct btrfs_key *key,
3031                                struct btrfs_path *p, int find_higher,
3032                                int return_any)
3033 {
3034         int ret;
3035         struct extent_buffer *leaf;
3036
3037 again:
3038         ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
3039         if (ret <= 0)
3040                 return ret;
3041         /*
3042          * a return value of 1 means the path is at the position where the
3043          * item should be inserted. Normally this is the next bigger item,
3044          * but in case the previous item is the last in a leaf, path points
3045          * to the first free slot in the previous leaf, i.e. at an invalid
3046          * item.
3047          */
3048         leaf = p->nodes[0];
3049
3050         if (find_higher) {
3051                 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
3052                         ret = btrfs_next_leaf(root, p);
3053                         if (ret <= 0)
3054                                 return ret;
3055                         if (!return_any)
3056                                 return 1;
3057                         /*
3058                          * no higher item found, return the next
3059                          * lower instead
3060                          */
3061                         return_any = 0;
3062                         find_higher = 0;
3063                         btrfs_release_path(p);
3064                         goto again;
3065                 }
3066         } else {
3067                 if (p->slots[0] == 0) {
3068                         ret = btrfs_prev_leaf(root, p);
3069                         if (ret < 0)
3070                                 return ret;
3071                         if (!ret) {
3072                                 leaf = p->nodes[0];
3073                                 if (p->slots[0] == btrfs_header_nritems(leaf))
3074                                         p->slots[0]--;
3075                                 return 0;
3076                         }
3077                         if (!return_any)
3078                                 return 1;
3079                         /*
3080                          * no lower item found, return the next
3081                          * higher instead
3082                          */
3083                         return_any = 0;
3084                         find_higher = 1;
3085                         btrfs_release_path(p);
3086                         goto again;
3087                 } else {
3088                         --p->slots[0];
3089                 }
3090         }
3091         return 0;
3092 }
3093
3094 /*
3095  * adjust the pointers going up the tree, starting at level
3096  * making sure the right key of each node is points to 'key'.
3097  * This is used after shifting pointers to the left, so it stops
3098  * fixing up pointers when a given leaf/node is not in slot 0 of the
3099  * higher levels
3100  *
3101  */
3102 static void fixup_low_keys(struct btrfs_path *path,
3103                            struct btrfs_disk_key *key, int level)
3104 {
3105         int i;
3106         struct extent_buffer *t;
3107         int ret;
3108
3109         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
3110                 int tslot = path->slots[i];
3111
3112                 if (!path->nodes[i])
3113                         break;
3114                 t = path->nodes[i];
3115                 ret = tree_mod_log_insert_key(t, tslot, MOD_LOG_KEY_REPLACE,
3116                                 GFP_ATOMIC);
3117                 BUG_ON(ret < 0);
3118                 btrfs_set_node_key(t, key, tslot);
3119                 btrfs_mark_buffer_dirty(path->nodes[i]);
3120                 if (tslot != 0)
3121                         break;
3122         }
3123 }
3124
3125 /*
3126  * update item key.
3127  *
3128  * This function isn't completely safe. It's the caller's responsibility
3129  * that the new key won't break the order
3130  */
3131 void btrfs_set_item_key_safe(struct btrfs_fs_info *fs_info,
3132                              struct btrfs_path *path,
3133                              const struct btrfs_key *new_key)
3134 {
3135         struct btrfs_disk_key disk_key;
3136         struct extent_buffer *eb;
3137         int slot;
3138
3139         eb = path->nodes[0];
3140         slot = path->slots[0];
3141         if (slot > 0) {
3142                 btrfs_item_key(eb, &disk_key, slot - 1);
3143                 BUG_ON(comp_keys(&disk_key, new_key) >= 0);
3144         }
3145         if (slot < btrfs_header_nritems(eb) - 1) {
3146                 btrfs_item_key(eb, &disk_key, slot + 1);
3147                 BUG_ON(comp_keys(&disk_key, new_key) <= 0);
3148         }
3149
3150         btrfs_cpu_key_to_disk(&disk_key, new_key);
3151         btrfs_set_item_key(eb, &disk_key, slot);
3152         btrfs_mark_buffer_dirty(eb);
3153         if (slot == 0)
3154                 fixup_low_keys(path, &disk_key, 1);
3155 }
3156
3157 /*
3158  * try to push data from one node into the next node left in the
3159  * tree.
3160  *
3161  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
3162  * error, and > 0 if there was no room in the left hand block.
3163  */
3164 static int push_node_left(struct btrfs_trans_handle *trans,
3165                           struct btrfs_fs_info *fs_info,
3166                           struct extent_buffer *dst,
3167                           struct extent_buffer *src, int empty)
3168 {
3169         int push_items = 0;
3170         int src_nritems;
3171         int dst_nritems;
3172         int ret = 0;
3173
3174         src_nritems = btrfs_header_nritems(src);
3175         dst_nritems = btrfs_header_nritems(dst);
3176         push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3177         WARN_ON(btrfs_header_generation(src) != trans->transid);
3178         WARN_ON(btrfs_header_generation(dst) != trans->transid);
3179
3180         if (!empty && src_nritems <= 8)
3181                 return 1;
3182
3183         if (push_items <= 0)
3184                 return 1;
3185
3186         if (empty) {
3187                 push_items = min(src_nritems, push_items);
3188                 if (push_items < src_nritems) {
3189                         /* leave at least 8 pointers in the node if
3190                          * we aren't going to empty it
3191                          */
3192                         if (src_nritems - push_items < 8) {
3193                                 if (push_items <= 8)
3194                                         return 1;
3195                                 push_items -= 8;
3196                         }
3197                 }
3198         } else
3199                 push_items = min(src_nritems - 8, push_items);
3200
3201         ret = tree_mod_log_eb_copy(fs_info, dst, src, dst_nritems, 0,
3202                                    push_items);
3203         if (ret) {
3204                 btrfs_abort_transaction(trans, ret);
3205                 return ret;
3206         }
3207         copy_extent_buffer(dst, src,
3208                            btrfs_node_key_ptr_offset(dst_nritems),
3209                            btrfs_node_key_ptr_offset(0),
3210                            push_items * sizeof(struct btrfs_key_ptr));
3211
3212         if (push_items < src_nritems) {
3213                 /*
3214                  * Don't call tree_mod_log_insert_move here, key removal was
3215                  * already fully logged by tree_mod_log_eb_copy above.
3216                  */
3217                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
3218                                       btrfs_node_key_ptr_offset(push_items),
3219                                       (src_nritems - push_items) *
3220                                       sizeof(struct btrfs_key_ptr));
3221         }
3222         btrfs_set_header_nritems(src, src_nritems - push_items);
3223         btrfs_set_header_nritems(dst, dst_nritems + push_items);
3224         btrfs_mark_buffer_dirty(src);
3225         btrfs_mark_buffer_dirty(dst);
3226
3227         return ret;
3228 }
3229
3230 /*
3231  * try to push data from one node into the next node right in the
3232  * tree.
3233  *
3234  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
3235  * error, and > 0 if there was no room in the right hand block.
3236  *
3237  * this will  only push up to 1/2 the contents of the left node over
3238  */
3239 static int balance_node_right(struct btrfs_trans_handle *trans,
3240                               struct btrfs_fs_info *fs_info,
3241                               struct extent_buffer *dst,
3242                               struct extent_buffer *src)
3243 {
3244         int push_items = 0;
3245         int max_push;
3246         int src_nritems;
3247         int dst_nritems;
3248         int ret = 0;
3249
3250         WARN_ON(btrfs_header_generation(src) != trans->transid);
3251         WARN_ON(btrfs_header_generation(dst) != trans->transid);
3252
3253         src_nritems = btrfs_header_nritems(src);
3254         dst_nritems = btrfs_header_nritems(dst);
3255         push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3256         if (push_items <= 0)
3257                 return 1;
3258
3259         if (src_nritems < 4)
3260                 return 1;
3261
3262         max_push = src_nritems / 2 + 1;
3263         /* don't try to empty the node */
3264         if (max_push >= src_nritems)
3265                 return 1;
3266
3267         if (max_push < push_items)
3268                 push_items = max_push;
3269
3270         ret = tree_mod_log_insert_move(dst, push_items, 0, dst_nritems);
3271         BUG_ON(ret < 0);
3272         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
3273                                       btrfs_node_key_ptr_offset(0),
3274                                       (dst_nritems) *
3275                                       sizeof(struct btrfs_key_ptr));
3276
3277         ret = tree_mod_log_eb_copy(fs_info, dst, src, 0,
3278                                    src_nritems - push_items, push_items);
3279         if (ret) {
3280                 btrfs_abort_transaction(trans, ret);
3281                 return ret;
3282         }
3283         copy_extent_buffer(dst, src,
3284                            btrfs_node_key_ptr_offset(0),
3285                            btrfs_node_key_ptr_offset(src_nritems - push_items),
3286                            push_items * sizeof(struct btrfs_key_ptr));
3287
3288         btrfs_set_header_nritems(src, src_nritems - push_items);
3289         btrfs_set_header_nritems(dst, dst_nritems + push_items);
3290
3291         btrfs_mark_buffer_dirty(src);
3292         btrfs_mark_buffer_dirty(dst);
3293
3294         return ret;
3295 }
3296
3297 /*
3298  * helper function to insert a new root level in the tree.
3299  * A new node is allocated, and a single item is inserted to
3300  * point to the existing root
3301  *
3302  * returns zero on success or < 0 on failure.
3303  */
3304 static noinline int insert_new_root(struct btrfs_trans_handle *trans,
3305                            struct btrfs_root *root,
3306                            struct btrfs_path *path, int level)
3307 {
3308         struct btrfs_fs_info *fs_info = root->fs_info;
3309         u64 lower_gen;
3310         struct extent_buffer *lower;
3311         struct extent_buffer *c;
3312         struct extent_buffer *old;
3313         struct btrfs_disk_key lower_key;
3314         int ret;
3315
3316         BUG_ON(path->nodes[level]);
3317         BUG_ON(path->nodes[level-1] != root->node);
3318
3319         lower = path->nodes[level-1];
3320         if (level == 1)
3321                 btrfs_item_key(lower, &lower_key, 0);
3322         else
3323                 btrfs_node_key(lower, &lower_key, 0);
3324
3325         c = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
3326                                    &lower_key, level, root->node->start, 0);
3327         if (IS_ERR(c))
3328                 return PTR_ERR(c);
3329
3330         root_add_used(root, fs_info->nodesize);
3331
3332         btrfs_set_header_nritems(c, 1);
3333         btrfs_set_node_key(c, &lower_key, 0);
3334         btrfs_set_node_blockptr(c, 0, lower->start);
3335         lower_gen = btrfs_header_generation(lower);
3336         WARN_ON(lower_gen != trans->transid);
3337
3338         btrfs_set_node_ptr_generation(c, 0, lower_gen);
3339
3340         btrfs_mark_buffer_dirty(c);
3341
3342         old = root->node;
3343         ret = tree_mod_log_insert_root(root->node, c, 0);
3344         BUG_ON(ret < 0);
3345         rcu_assign_pointer(root->node, c);
3346
3347         /* the super has an extra ref to root->node */
3348         free_extent_buffer(old);
3349
3350         add_root_to_dirty_list(root);
3351         extent_buffer_get(c);
3352         path->nodes[level] = c;
3353         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
3354         path->slots[level] = 0;
3355         return 0;
3356 }
3357
3358 /*
3359  * worker function to insert a single pointer in a node.
3360  * the node should have enough room for the pointer already
3361  *
3362  * slot and level indicate where you want the key to go, and
3363  * blocknr is the block the key points to.
3364  */
3365 static void insert_ptr(struct btrfs_trans_handle *trans,
3366                        struct btrfs_fs_info *fs_info, struct btrfs_path *path,
3367                        struct btrfs_disk_key *key, u64 bytenr,
3368                        int slot, int level)
3369 {
3370         struct extent_buffer *lower;
3371         int nritems;
3372         int ret;
3373
3374         BUG_ON(!path->nodes[level]);
3375         btrfs_assert_tree_locked(path->nodes[level]);
3376         lower = path->nodes[level];
3377         nritems = btrfs_header_nritems(lower);
3378         BUG_ON(slot > nritems);
3379         BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(fs_info));
3380         if (slot != nritems) {
3381                 if (level) {
3382                         ret = tree_mod_log_insert_move(lower, slot + 1, slot,
3383                                         nritems - slot);
3384                         BUG_ON(ret < 0);
3385                 }
3386                 memmove_extent_buffer(lower,
3387                               btrfs_node_key_ptr_offset(slot + 1),
3388                               btrfs_node_key_ptr_offset(slot),
3389                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
3390         }
3391         if (level) {
3392                 ret = tree_mod_log_insert_key(lower, slot, MOD_LOG_KEY_ADD,
3393                                 GFP_NOFS);
3394                 BUG_ON(ret < 0);
3395         }
3396         btrfs_set_node_key(lower, key, slot);
3397         btrfs_set_node_blockptr(lower, slot, bytenr);
3398         WARN_ON(trans->transid == 0);
3399         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
3400         btrfs_set_header_nritems(lower, nritems + 1);
3401         btrfs_mark_buffer_dirty(lower);
3402 }
3403
3404 /*
3405  * split the node at the specified level in path in two.
3406  * The path is corrected to point to the appropriate node after the split
3407  *
3408  * Before splitting this tries to make some room in the node by pushing
3409  * left and right, if either one works, it returns right away.
3410  *
3411  * returns 0 on success and < 0 on failure
3412  */
3413 static noinline int split_node(struct btrfs_trans_handle *trans,
3414                                struct btrfs_root *root,
3415                                struct btrfs_path *path, int level)
3416 {
3417         struct btrfs_fs_info *fs_info = root->fs_info;
3418         struct extent_buffer *c;
3419         struct extent_buffer *split;
3420         struct btrfs_disk_key disk_key;
3421         int mid;
3422         int ret;
3423         u32 c_nritems;
3424
3425         c = path->nodes[level];
3426         WARN_ON(btrfs_header_generation(c) != trans->transid);
3427         if (c == root->node) {
3428                 /*
3429                  * trying to split the root, lets make a new one
3430                  *
3431                  * tree mod log: We don't log_removal old root in
3432                  * insert_new_root, because that root buffer will be kept as a
3433                  * normal node. We are going to log removal of half of the
3434                  * elements below with tree_mod_log_eb_copy. We're holding a
3435                  * tree lock on the buffer, which is why we cannot race with
3436                  * other tree_mod_log users.
3437                  */
3438                 ret = insert_new_root(trans, root, path, level + 1);
3439                 if (ret)
3440                         return ret;
3441         } else {
3442                 ret = push_nodes_for_insert(trans, root, path, level);
3443                 c = path->nodes[level];
3444                 if (!ret && btrfs_header_nritems(c) <
3445                     BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3)
3446                         return 0;
3447                 if (ret < 0)
3448                         return ret;
3449         }
3450
3451         c_nritems = btrfs_header_nritems(c);
3452         mid = (c_nritems + 1) / 2;
3453         btrfs_node_key(c, &disk_key, mid);
3454
3455         split = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
3456                         &disk_key, level, c->start, 0);
3457         if (IS_ERR(split))
3458                 return PTR_ERR(split);
3459
3460         root_add_used(root, fs_info->nodesize);
3461         ASSERT(btrfs_header_level(c) == level);
3462
3463         ret = tree_mod_log_eb_copy(fs_info, split, c, 0, mid, c_nritems - mid);
3464         if (ret) {
3465                 btrfs_abort_transaction(trans, ret);
3466                 return ret;
3467         }
3468         copy_extent_buffer(split, c,
3469                            btrfs_node_key_ptr_offset(0),
3470                            btrfs_node_key_ptr_offset(mid),
3471                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
3472         btrfs_set_header_nritems(split, c_nritems - mid);
3473         btrfs_set_header_nritems(c, mid);
3474         ret = 0;
3475
3476         btrfs_mark_buffer_dirty(c);
3477         btrfs_mark_buffer_dirty(split);
3478
3479         insert_ptr(trans, fs_info, path, &disk_key, split->start,
3480                    path->slots[level + 1] + 1, level + 1);
3481
3482         if (path->slots[level] >= mid) {
3483                 path->slots[level] -= mid;
3484                 btrfs_tree_unlock(c);
3485                 free_extent_buffer(c);
3486                 path->nodes[level] = split;
3487                 path->slots[level + 1] += 1;
3488         } else {
3489                 btrfs_tree_unlock(split);
3490                 free_extent_buffer(split);
3491         }
3492         return ret;
3493 }
3494
3495 /*
3496  * how many bytes are required to store the items in a leaf.  start
3497  * and nr indicate which items in the leaf to check.  This totals up the
3498  * space used both by the item structs and the item data
3499  */
3500 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
3501 {
3502         struct btrfs_item *start_item;
3503         struct btrfs_item *end_item;
3504         struct btrfs_map_token token;
3505         int data_len;
3506         int nritems = btrfs_header_nritems(l);
3507         int end = min(nritems, start + nr) - 1;
3508
3509         if (!nr)
3510                 return 0;
3511         btrfs_init_map_token(&token);
3512         start_item = btrfs_item_nr(start);
3513         end_item = btrfs_item_nr(end);
3514         data_len = btrfs_token_item_offset(l, start_item, &token) +
3515                 btrfs_token_item_size(l, start_item, &token);
3516         data_len = data_len - btrfs_token_item_offset(l, end_item, &token);
3517         data_len += sizeof(struct btrfs_item) * nr;
3518         WARN_ON(data_len < 0);
3519         return data_len;
3520 }
3521
3522 /*
3523  * The space between the end of the leaf items and
3524  * the start of the leaf data.  IOW, how much room
3525  * the leaf has left for both items and data
3526  */
3527 noinline int btrfs_leaf_free_space(struct btrfs_fs_info *fs_info,
3528                                    struct extent_buffer *leaf)
3529 {
3530         int nritems = btrfs_header_nritems(leaf);
3531         int ret;
3532
3533         ret = BTRFS_LEAF_DATA_SIZE(fs_info) - leaf_space_used(leaf, 0, nritems);
3534         if (ret < 0) {
3535                 btrfs_crit(fs_info,
3536                            "leaf free space ret %d, leaf data size %lu, used %d nritems %d",
3537                            ret,
3538                            (unsigned long) BTRFS_LEAF_DATA_SIZE(fs_info),
3539                            leaf_space_used(leaf, 0, nritems), nritems);
3540         }
3541         return ret;
3542 }
3543
3544 /*
3545  * min slot controls the lowest index we're willing to push to the
3546  * right.  We'll push up to and including min_slot, but no lower
3547  */
3548 static noinline int __push_leaf_right(struct btrfs_fs_info *fs_info,
3549                                       struct btrfs_path *path,
3550                                       int data_size, int empty,
3551                                       struct extent_buffer *right,
3552                                       int free_space, u32 left_nritems,
3553                                       u32 min_slot)
3554 {
3555         struct extent_buffer *left = path->nodes[0];
3556         struct extent_buffer *upper = path->nodes[1];
3557         struct btrfs_map_token token;
3558         struct btrfs_disk_key disk_key;
3559         int slot;
3560         u32 i;
3561         int push_space = 0;
3562         int push_items = 0;
3563         struct btrfs_item *item;
3564         u32 nr;
3565         u32 right_nritems;
3566         u32 data_end;
3567         u32 this_item_size;
3568
3569         btrfs_init_map_token(&token);
3570
3571         if (empty)
3572                 nr = 0;
3573         else
3574                 nr = max_t(u32, 1, min_slot);
3575
3576         if (path->slots[0] >= left_nritems)
3577                 push_space += data_size;
3578
3579         slot = path->slots[1];
3580         i = left_nritems - 1;
3581         while (i >= nr) {
3582                 item = btrfs_item_nr(i);
3583
3584                 if (!empty && push_items > 0) {
3585                         if (path->slots[0] > i)
3586                                 break;
3587                         if (path->slots[0] == i) {
3588                                 int space = btrfs_leaf_free_space(fs_info, left);
3589                                 if (space + push_space * 2 > free_space)
3590                                         break;
3591                         }
3592                 }
3593
3594                 if (path->slots[0] == i)
3595                         push_space += data_size;
3596
3597                 this_item_size = btrfs_item_size(left, item);
3598                 if (this_item_size + sizeof(*item) + push_space > free_space)
3599                         break;
3600
3601                 push_items++;
3602                 push_space += this_item_size + sizeof(*item);
3603                 if (i == 0)
3604                         break;
3605                 i--;
3606         }
3607
3608         if (push_items == 0)
3609                 goto out_unlock;
3610
3611         WARN_ON(!empty && push_items == left_nritems);
3612
3613         /* push left to right */
3614         right_nritems = btrfs_header_nritems(right);
3615
3616         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
3617         push_space -= leaf_data_end(fs_info, left);
3618
3619         /* make room in the right data area */
3620         data_end = leaf_data_end(fs_info, right);
3621         memmove_extent_buffer(right,
3622                               BTRFS_LEAF_DATA_OFFSET + data_end - push_space,
3623                               BTRFS_LEAF_DATA_OFFSET + data_end,
3624                               BTRFS_LEAF_DATA_SIZE(fs_info) - data_end);
3625
3626         /* copy from the left data area */
3627         copy_extent_buffer(right, left, BTRFS_LEAF_DATA_OFFSET +
3628                      BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3629                      BTRFS_LEAF_DATA_OFFSET + leaf_data_end(fs_info, left),
3630                      push_space);
3631
3632         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
3633                               btrfs_item_nr_offset(0),
3634                               right_nritems * sizeof(struct btrfs_item));
3635
3636         /* copy the items from left to right */
3637         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
3638                    btrfs_item_nr_offset(left_nritems - push_items),
3639                    push_items * sizeof(struct btrfs_item));
3640
3641         /* update the item pointers */
3642         right_nritems += push_items;
3643         btrfs_set_header_nritems(right, right_nritems);
3644         push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3645         for (i = 0; i < right_nritems; i++) {
3646                 item = btrfs_item_nr(i);
3647                 push_space -= btrfs_token_item_size(right, item, &token);
3648                 btrfs_set_token_item_offset(right, item, push_space, &token);
3649         }
3650
3651         left_nritems -= push_items;
3652         btrfs_set_header_nritems(left, left_nritems);
3653
3654         if (left_nritems)
3655                 btrfs_mark_buffer_dirty(left);
3656         else
3657                 clean_tree_block(fs_info, left);
3658
3659         btrfs_mark_buffer_dirty(right);
3660
3661         btrfs_item_key(right, &disk_key, 0);
3662         btrfs_set_node_key(upper, &disk_key, slot + 1);
3663         btrfs_mark_buffer_dirty(upper);
3664
3665         /* then fixup the leaf pointer in the path */
3666         if (path->slots[0] >= left_nritems) {
3667                 path->slots[0] -= left_nritems;
3668                 if (btrfs_header_nritems(path->nodes[0]) == 0)
3669                         clean_tree_block(fs_info, path->nodes[0]);
3670                 btrfs_tree_unlock(path->nodes[0]);
3671                 free_extent_buffer(path->nodes[0]);
3672                 path->nodes[0] = right;
3673                 path->slots[1] += 1;
3674         } else {
3675                 btrfs_tree_unlock(right);
3676                 free_extent_buffer(right);
3677         }
3678         return 0;
3679
3680 out_unlock:
3681         btrfs_tree_unlock(right);
3682         free_extent_buffer(right);
3683         return 1;
3684 }
3685
3686 /*
3687  * push some data in the path leaf to the right, trying to free up at
3688  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
3689  *
3690  * returns 1 if the push failed because the other node didn't have enough
3691  * room, 0 if everything worked out and < 0 if there were major errors.
3692  *
3693  * this will push starting from min_slot to the end of the leaf.  It won't
3694  * push any slot lower than min_slot
3695  */
3696 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
3697                            *root, struct btrfs_path *path,
3698                            int min_data_size, int data_size,
3699                            int empty, u32 min_slot)
3700 {
3701         struct btrfs_fs_info *fs_info = root->fs_info;
3702         struct extent_buffer *left = path->nodes[0];
3703         struct extent_buffer *right;
3704         struct extent_buffer *upper;
3705         int slot;
3706         int free_space;
3707         u32 left_nritems;
3708         int ret;
3709
3710         if (!path->nodes[1])
3711                 return 1;
3712
3713         slot = path->slots[1];
3714         upper = path->nodes[1];
3715         if (slot >= btrfs_header_nritems(upper) - 1)
3716                 return 1;
3717
3718         btrfs_assert_tree_locked(path->nodes[1]);
3719
3720         right = read_node_slot(fs_info, upper, slot + 1);
3721         /*
3722          * slot + 1 is not valid or we fail to read the right node,
3723          * no big deal, just return.
3724          */
3725         if (IS_ERR(right))
3726                 return 1;
3727
3728         btrfs_tree_lock(right);
3729         btrfs_set_lock_blocking(right);
3730
3731         free_space = btrfs_leaf_free_space(fs_info, right);
3732         if (free_space < data_size)
3733                 goto out_unlock;
3734
3735         /* cow and double check */
3736         ret = btrfs_cow_block(trans, root, right, upper,
3737                               slot + 1, &right);
3738         if (ret)
3739                 goto out_unlock;
3740
3741         free_space = btrfs_leaf_free_space(fs_info, right);
3742         if (free_space < data_size)
3743                 goto out_unlock;
3744
3745         left_nritems = btrfs_header_nritems(left);
3746         if (left_nritems == 0)
3747                 goto out_unlock;
3748
3749         if (path->slots[0] == left_nritems && !empty) {
3750                 /* Key greater than all keys in the leaf, right neighbor has
3751                  * enough room for it and we're not emptying our leaf to delete
3752                  * it, therefore use right neighbor to insert the new item and
3753                  * no need to touch/dirty our left leaft. */
3754                 btrfs_tree_unlock(left);
3755                 free_extent_buffer(left);
3756                 path->nodes[0] = right;
3757                 path->slots[0] = 0;
3758                 path->slots[1]++;
3759                 return 0;
3760         }
3761
3762         return __push_leaf_right(fs_info, path, min_data_size, empty,
3763                                 right, free_space, left_nritems, min_slot);
3764 out_unlock:
3765         btrfs_tree_unlock(right);
3766         free_extent_buffer(right);
3767         return 1;
3768 }
3769
3770 /*
3771  * push some data in the path leaf to the left, trying to free up at
3772  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
3773  *
3774  * max_slot can put a limit on how far into the leaf we'll push items.  The
3775  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us do all the
3776  * items
3777  */
3778 static noinline int __push_leaf_left(struct btrfs_fs_info *fs_info,
3779                                      struct btrfs_path *path, int data_size,
3780                                      int empty, struct extent_buffer *left,
3781                                      int free_space, u32 right_nritems,
3782                                      u32 max_slot)
3783 {
3784         struct btrfs_disk_key disk_key;
3785         struct extent_buffer *right = path->nodes[0];
3786         int i;
3787         int push_space = 0;
3788         int push_items = 0;
3789         struct btrfs_item *item;
3790         u32 old_left_nritems;
3791         u32 nr;
3792         int ret = 0;
3793         u32 this_item_size;
3794         u32 old_left_item_size;
3795         struct btrfs_map_token token;
3796
3797         btrfs_init_map_token(&token);
3798
3799         if (empty)
3800                 nr = min(right_nritems, max_slot);
3801         else
3802                 nr = min(right_nritems - 1, max_slot);
3803
3804         for (i = 0; i < nr; i++) {
3805                 item = btrfs_item_nr(i);
3806
3807                 if (!empty && push_items > 0) {
3808                         if (path->slots[0] < i)
3809                                 break;
3810                         if (path->slots[0] == i) {
3811                                 int space = btrfs_leaf_free_space(fs_info, right);
3812                                 if (space + push_space * 2 > free_space)
3813                                         break;
3814                         }
3815                 }
3816
3817                 if (path->slots[0] == i)
3818                         push_space += data_size;
3819
3820                 this_item_size = btrfs_item_size(right, item);
3821                 if (this_item_size + sizeof(*item) + push_space > free_space)
3822                         break;
3823
3824                 push_items++;
3825                 push_space += this_item_size + sizeof(*item);
3826         }
3827
3828         if (push_items == 0) {
3829                 ret = 1;
3830                 goto out;
3831         }
3832         WARN_ON(!empty && push_items == btrfs_header_nritems(right));
3833
3834         /* push data from right to left */
3835         copy_extent_buffer(left, right,
3836                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
3837                            btrfs_item_nr_offset(0),
3838                            push_items * sizeof(struct btrfs_item));
3839
3840         push_space = BTRFS_LEAF_DATA_SIZE(fs_info) -
3841                      btrfs_item_offset_nr(right, push_items - 1);
3842
3843         copy_extent_buffer(left, right, BTRFS_LEAF_DATA_OFFSET +
3844                      leaf_data_end(fs_info, left) - push_space,
3845                      BTRFS_LEAF_DATA_OFFSET +
3846                      btrfs_item_offset_nr(right, push_items - 1),
3847                      push_space);
3848         old_left_nritems = btrfs_header_nritems(left);
3849         BUG_ON(old_left_nritems <= 0);
3850
3851         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
3852         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
3853                 u32 ioff;
3854
3855                 item = btrfs_item_nr(i);
3856
3857                 ioff = btrfs_token_item_offset(left, item, &token);
3858                 btrfs_set_token_item_offset(left, item,
3859                       ioff - (BTRFS_LEAF_DATA_SIZE(fs_info) - old_left_item_size),
3860                       &token);
3861         }
3862         btrfs_set_header_nritems(left, old_left_nritems + push_items);
3863
3864         /* fixup right node */
3865         if (push_items > right_nritems)
3866                 WARN(1, KERN_CRIT "push items %d nr %u\n", push_items,
3867                        right_nritems);
3868
3869         if (push_items < right_nritems) {
3870                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
3871                                                   leaf_data_end(fs_info, right);
3872                 memmove_extent_buffer(right, BTRFS_LEAF_DATA_OFFSET +
3873                                       BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3874                                       BTRFS_LEAF_DATA_OFFSET +
3875                                       leaf_data_end(fs_info, right), push_space);
3876
3877                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
3878                               btrfs_item_nr_offset(push_items),
3879                              (btrfs_header_nritems(right) - push_items) *
3880                              sizeof(struct btrfs_item));
3881         }
3882         right_nritems -= push_items;
3883         btrfs_set_header_nritems(right, right_nritems);
3884         push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3885         for (i = 0; i < right_nritems; i++) {
3886                 item = btrfs_item_nr(i);
3887
3888                 push_space = push_space - btrfs_token_item_size(right,
3889                                                                 item, &token);
3890                 btrfs_set_token_item_offset(right, item, push_space, &token);
3891         }
3892
3893         btrfs_mark_buffer_dirty(left);
3894         if (right_nritems)
3895                 btrfs_mark_buffer_dirty(right);
3896         else
3897                 clean_tree_block(fs_info, right);
3898
3899         btrfs_item_key(right, &disk_key, 0);
3900         fixup_low_keys(path, &disk_key, 1);
3901
3902         /* then fixup the leaf pointer in the path */
3903         if (path->slots[0] < push_items) {
3904                 path->slots[0] += old_left_nritems;
3905                 btrfs_tree_unlock(path->nodes[0]);
3906                 free_extent_buffer(path->nodes[0]);
3907                 path->nodes[0] = left;
3908                 path->slots[1] -= 1;
3909         } else {
3910                 btrfs_tree_unlock(left);
3911                 free_extent_buffer(left);
3912                 path->slots[0] -= push_items;
3913         }
3914         BUG_ON(path->slots[0] < 0);
3915         return ret;
3916 out:
3917         btrfs_tree_unlock(left);
3918         free_extent_buffer(left);
3919         return ret;
3920 }
3921
3922 /*
3923  * push some data in the path leaf to the left, trying to free up at
3924  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
3925  *
3926  * max_slot can put a limit on how far into the leaf we'll push items.  The
3927  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us push all the
3928  * items
3929  */
3930 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
3931                           *root, struct btrfs_path *path, int min_data_size,
3932                           int data_size, int empty, u32 max_slot)
3933 {
3934         struct btrfs_fs_info *fs_info = root->fs_info;
3935         struct extent_buffer *right = path->nodes[0];
3936         struct extent_buffer *left;
3937         int slot;
3938         int free_space;
3939         u32 right_nritems;
3940         int ret = 0;
3941
3942         slot = path->slots[1];
3943         if (slot == 0)
3944                 return 1;
3945         if (!path->nodes[1])
3946                 return 1;
3947
3948         right_nritems = btrfs_header_nritems(right);
3949         if (right_nritems == 0)
3950                 return 1;
3951
3952         btrfs_assert_tree_locked(path->nodes[1]);
3953
3954         left = read_node_slot(fs_info, path->nodes[1], slot - 1);
3955         /*
3956          * slot - 1 is not valid or we fail to read the left node,
3957          * no big deal, just return.
3958          */
3959         if (IS_ERR(left))
3960                 return 1;
3961
3962         btrfs_tree_lock(left);
3963         btrfs_set_lock_blocking(left);
3964
3965         free_space = btrfs_leaf_free_space(fs_info, left);
3966         if (free_space < data_size) {
3967                 ret = 1;
3968                 goto out;
3969         }
3970
3971         /* cow and double check */
3972         ret = btrfs_cow_block(trans, root, left,
3973                               path->nodes[1], slot - 1, &left);
3974         if (ret) {
3975                 /* we hit -ENOSPC, but it isn't fatal here */
3976                 if (ret == -ENOSPC)
3977                         ret = 1;
3978                 goto out;
3979         }
3980
3981         free_space = btrfs_leaf_free_space(fs_info, left);
3982         if (free_space < data_size) {
3983                 ret = 1;
3984                 goto out;
3985         }
3986
3987         return __push_leaf_left(fs_info, path, min_data_size,
3988                                empty, left, free_space, right_nritems,
3989                                max_slot);
3990 out:
3991         btrfs_tree_unlock(left);
3992         free_extent_buffer(left);
3993         return ret;
3994 }
3995
3996 /*
3997  * split the path's leaf in two, making sure there is at least data_size
3998  * available for the resulting leaf level of the path.
3999  */
4000 static noinline void copy_for_split(struct btrfs_trans_handle *trans,
4001                                     struct btrfs_fs_info *fs_info,
4002                                     struct btrfs_path *path,
4003                                     struct extent_buffer *l,
4004                                     struct extent_buffer *right,
4005                                     int slot, int mid, int nritems)
4006 {
4007         int data_copy_size;
4008         int rt_data_off;
4009         int i;
4010         struct btrfs_disk_key disk_key;
4011         struct btrfs_map_token token;
4012
4013         btrfs_init_map_token(&token);
4014
4015         nritems = nritems - mid;
4016         btrfs_set_header_nritems(right, nritems);
4017         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(fs_info, l);
4018
4019         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
4020                            btrfs_item_nr_offset(mid),
4021                            nritems * sizeof(struct btrfs_item));
4022
4023         copy_extent_buffer(right, l,
4024                      BTRFS_LEAF_DATA_OFFSET + BTRFS_LEAF_DATA_SIZE(fs_info) -
4025                      data_copy_size, BTRFS_LEAF_DATA_OFFSET +
4026                      leaf_data_end(fs_info, l), data_copy_size);
4027
4028         rt_data_off = BTRFS_LEAF_DATA_SIZE(fs_info) - btrfs_item_end_nr(l, mid);
4029
4030         for (i = 0; i < nritems; i++) {
4031                 struct btrfs_item *item = btrfs_item_nr(i);
4032                 u32 ioff;
4033
4034                 ioff = btrfs_token_item_offset(right, item, &token);
4035                 btrfs_set_token_item_offset(right, item,
4036                                             ioff + rt_data_off, &token);
4037         }
4038
4039         btrfs_set_header_nritems(l, mid);
4040         btrfs_item_key(right, &disk_key, 0);
4041         insert_ptr(trans, fs_info, path, &disk_key, right->start,
4042                    path->slots[1] + 1, 1);
4043
4044         btrfs_mark_buffer_dirty(right);
4045         btrfs_mark_buffer_dirty(l);
4046         BUG_ON(path->slots[0] != slot);
4047
4048         if (mid <= slot) {
4049                 btrfs_tree_unlock(path->nodes[0]);
4050                 free_extent_buffer(path->nodes[0]);
4051                 path->nodes[0] = right;
4052                 path->slots[0] -= mid;
4053                 path->slots[1] += 1;
4054         } else {
4055                 btrfs_tree_unlock(right);
4056                 free_extent_buffer(right);
4057         }
4058
4059         BUG_ON(path->slots[0] < 0);
4060 }
4061
4062 /*
4063  * double splits happen when we need to insert a big item in the middle
4064  * of a leaf.  A double split can leave us with 3 mostly empty leaves:
4065  * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
4066  *          A                 B                 C
4067  *
4068  * We avoid this by trying to push the items on either side of our target
4069  * into the adjacent leaves.  If all goes well we can avoid the double split
4070  * completely.
4071  */
4072 static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
4073                                           struct btrfs_root *root,
4074                                           struct btrfs_path *path,
4075                                           int data_size)
4076 {
4077         struct btrfs_fs_info *fs_info = root->fs_info;
4078         int ret;
4079         int progress = 0;
4080         int slot;
4081         u32 nritems;
4082         int space_needed = data_size;
4083
4084         slot = path->slots[0];
4085         if (slot < btrfs_header_nritems(path->nodes[0]))
4086                 space_needed -= btrfs_leaf_free_space(fs_info, path->nodes[0]);
4087
4088         /*
4089          * try to push all the items after our slot into the
4090          * right leaf
4091          */
4092         ret = push_leaf_right(trans, root, path, 1, space_needed, 0, slot);
4093         if (ret < 0)
4094                 return ret;
4095
4096         if (ret == 0)
4097                 progress++;
4098
4099         nritems = btrfs_header_nritems(path->nodes[0]);
4100         /*
4101          * our goal is to get our slot at the start or end of a leaf.  If
4102          * we've done so we're done
4103          */
4104         if (path->slots[0] == 0 || path->slots[0] == nritems)
4105                 return 0;
4106
4107         if (btrfs_leaf_free_space(fs_info, path->nodes[0]) >= data_size)
4108                 return 0;
4109
4110         /* try to push all the items before our slot into the next leaf */
4111         slot = path->slots[0];
4112         space_needed = data_size;
4113         if (slot > 0)
4114                 space_needed -= btrfs_leaf_free_space(fs_info, path->nodes[0]);
4115         ret = push_leaf_left(trans, root, path, 1, space_needed, 0, slot);
4116         if (ret < 0)
4117                 return ret;
4118
4119         if (ret == 0)
4120                 progress++;
4121
4122         if (progress)
4123                 return 0;
4124         return 1;
4125 }
4126
4127 /*
4128  * split the path's leaf in two, making sure there is at least data_size
4129  * available for the resulting leaf level of the path.
4130  *
4131  * returns 0 if all went well and < 0 on failure.
4132  */
4133 static noinline int split_leaf(struct btrfs_trans_handle *trans,
4134                                struct btrfs_root *root,
4135                                const struct btrfs_key *ins_key,
4136                                struct btrfs_path *path, int data_size,
4137                                int extend)
4138 {
4139         struct btrfs_disk_key disk_key;
4140         struct extent_buffer *l;
4141         u32 nritems;
4142         int mid;
4143         int slot;
4144         struct extent_buffer *right;
4145         struct btrfs_fs_info *fs_info = root->fs_info;
4146         int ret = 0;
4147         int wret;
4148         int split;
4149         int num_doubles = 0;
4150         int tried_avoid_double = 0;
4151
4152         l = path->nodes[0];
4153         slot = path->slots[0];
4154         if (extend && data_size + btrfs_item_size_nr(l, slot) +
4155             sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(fs_info))
4156                 return -EOVERFLOW;
4157
4158         /* first try to make some room by pushing left and right */
4159         if (data_size && path->nodes[1]) {
4160                 int space_needed = data_size;
4161
4162                 if (slot < btrfs_header_nritems(l))
4163                         space_needed -= btrfs_leaf_free_space(fs_info, l);
4164
4165                 wret = push_leaf_right(trans, root, path, space_needed,
4166                                        space_needed, 0, 0);
4167                 if (wret < 0)
4168                         return wret;
4169                 if (wret) {
4170                         space_needed = data_size;
4171                         if (slot > 0)
4172                                 space_needed -= btrfs_leaf_free_space(fs_info,
4173                                                                       l);
4174                         wret = push_leaf_left(trans, root, path, space_needed,
4175                                               space_needed, 0, (u32)-1);
4176                         if (wret < 0)
4177                                 return wret;
4178                 }
4179                 l = path->nodes[0];
4180
4181                 /* did the pushes work? */
4182                 if (btrfs_leaf_free_space(fs_info, l) >= data_size)
4183                         return 0;
4184         }
4185
4186         if (!path->nodes[1]) {
4187                 ret = insert_new_root(trans, root, path, 1);
4188                 if (ret)
4189                         return ret;
4190         }
4191 again:
4192         split = 1;
4193         l = path->nodes[0];
4194         slot = path->slots[0];
4195         nritems = btrfs_header_nritems(l);
4196         mid = (nritems + 1) / 2;
4197
4198         if (mid <= slot) {
4199                 if (nritems == 1 ||
4200                     leaf_space_used(l, mid, nritems - mid) + data_size >
4201                         BTRFS_LEAF_DATA_SIZE(fs_info)) {
4202                         if (slot >= nritems) {
4203                                 split = 0;
4204                         } else {
4205                                 mid = slot;
4206                                 if (mid != nritems &&
4207                                     leaf_space_used(l, mid, nritems - mid) +
4208                                     data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
4209                                         if (data_size && !tried_avoid_double)
4210                                                 goto push_for_double;
4211                                         split = 2;
4212                                 }
4213                         }
4214                 }
4215         } else {
4216                 if (leaf_space_used(l, 0, mid) + data_size >
4217                         BTRFS_LEAF_DATA_SIZE(fs_info)) {
4218                         if (!extend && data_size && slot == 0) {
4219                                 split = 0;
4220                         } else if ((extend || !data_size) && slot == 0) {
4221                                 mid = 1;
4222                         } else {
4223                                 mid = slot;
4224                                 if (mid != nritems &&
4225                                     leaf_space_used(l, mid, nritems - mid) +
4226                                     data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
4227                                         if (data_size && !tried_avoid_double)
4228                                                 goto push_for_double;
4229                                         split = 2;
4230                                 }
4231                         }
4232                 }
4233         }
4234
4235         if (split == 0)
4236                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
4237         else
4238                 btrfs_item_key(l, &disk_key, mid);
4239
4240         right = btrfs_alloc_tree_block(trans, root, 0, root->root_key.objectid,
4241                         &disk_key, 0, l->start, 0);
4242         if (IS_ERR(right))
4243                 return PTR_ERR(right);
4244
4245         root_add_used(root, fs_info->nodesize);
4246
4247         if (split == 0) {
4248                 if (mid <= slot) {
4249                         btrfs_set_header_nritems(right, 0);
4250                         insert_ptr(trans, fs_info, path, &disk_key,
4251                                    right->start, path->slots[1] + 1, 1);
4252                         btrfs_tree_unlock(path->nodes[0]);
4253                         free_extent_buffer(path->nodes[0]);
4254                         path->nodes[0] = right;
4255                         path->slots[0] = 0;
4256                         path->slots[1] += 1;
4257                 } else {
4258                         btrfs_set_header_nritems(right, 0);
4259                         insert_ptr(trans, fs_info, path, &disk_key,
4260                                    right->start, path->slots[1], 1);
4261                         btrfs_tree_unlock(path->nodes[0]);
4262                         free_extent_buffer(path->nodes[0]);
4263                         path->nodes[0] = right;
4264                         path->slots[0] = 0;
4265                         if (path->slots[1] == 0)
4266                                 fixup_low_keys(path, &disk_key, 1);
4267                 }
4268                 /*
4269                  * We create a new leaf 'right' for the required ins_len and
4270                  * we'll do btrfs_mark_buffer_dirty() on this leaf after copying
4271                  * the content of ins_len to 'right'.
4272                  */
4273                 return ret;
4274         }
4275
4276         copy_for_split(trans, fs_info, path, l, right, slot, mid, nritems);
4277
4278         if (split == 2) {
4279                 BUG_ON(num_doubles != 0);
4280                 num_doubles++;
4281                 goto again;
4282         }
4283
4284         return 0;
4285
4286 push_for_double:
4287         push_for_double_split(trans, root, path, data_size);
4288         tried_avoid_double = 1;
4289         if (btrfs_leaf_free_space(fs_info, path->nodes[0]) >= data_size)
4290                 return 0;
4291         goto again;
4292 }
4293
4294 static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
4295                                          struct btrfs_root *root,
4296                                          struct btrfs_path *path, int ins_len)
4297 {
4298         struct btrfs_fs_info *fs_info = root->fs_info;
4299         struct btrfs_key key;
4300         struct extent_buffer *leaf;
4301         struct btrfs_file_extent_item *fi;
4302         u64 extent_len = 0;
4303         u32 item_size;
4304         int ret;
4305
4306         leaf = path->nodes[0];
4307         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4308
4309         BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
4310                key.type != BTRFS_EXTENT_CSUM_KEY);
4311
4312         if (btrfs_leaf_free_space(fs_info, leaf) >= ins_len)
4313                 return 0;
4314
4315         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4316         if (key.type == BTRFS_EXTENT_DATA_KEY) {
4317                 fi = btrfs_item_ptr(leaf, path->slots[0],
4318                                     struct btrfs_file_extent_item);
4319                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
4320         }
4321         btrfs_release_path(path);
4322
4323         path->keep_locks = 1;
4324         path->search_for_split = 1;
4325         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
4326         path->search_for_split = 0;
4327         if (ret > 0)
4328                 ret = -EAGAIN;
4329         if (ret < 0)
4330                 goto err;
4331
4332         ret = -EAGAIN;
4333         leaf = path->nodes[0];
4334         /* if our item isn't there, return now */
4335         if (item_size != btrfs_item_size_nr(leaf, path->slots[0]))
4336                 goto err;
4337
4338         /* the leaf has  changed, it now has room.  return now */
4339         if (btrfs_leaf_free_space(fs_info, path->nodes[0]) >= ins_len)
4340                 goto err;
4341
4342         if (key.type == BTRFS_EXTENT_DATA_KEY) {
4343                 fi = btrfs_item_ptr(leaf, path->slots[0],
4344                                     struct btrfs_file_extent_item);
4345                 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
4346                         goto err;
4347         }
4348
4349         btrfs_set_path_blocking(path);
4350         ret = split_leaf(trans, root, &key, path, ins_len, 1);
4351         if (ret)
4352                 goto err;
4353
4354         path->keep_locks = 0;
4355         btrfs_unlock_up_safe(path, 1);
4356         return 0;
4357 err:
4358         path->keep_locks = 0;
4359         return ret;
4360 }
4361
4362 static noinline int split_item(struct btrfs_fs_info *fs_info,
4363                                struct btrfs_path *path,
4364                                const struct btrfs_key *new_key,
4365                                unsigned long split_offset)
4366 {
4367         struct extent_buffer *leaf;
4368         struct btrfs_item *item;
4369         struct btrfs_item *new_item;
4370         int slot;
4371         char *buf;
4372         u32 nritems;
4373         u32 item_size;
4374         u32 orig_offset;
4375         struct btrfs_disk_key disk_key;
4376
4377         leaf = path->nodes[0];
4378         BUG_ON(btrfs_leaf_free_space(fs_info, leaf) < sizeof(struct btrfs_item));
4379
4380         btrfs_set_path_blocking(path);
4381
4382         item = btrfs_item_nr(path->slots[0]);
4383         orig_offset = btrfs_item_offset(leaf, item);
4384         item_size = btrfs_item_size(leaf, item);
4385
4386         buf = kmalloc(item_size, GFP_NOFS);
4387         if (!buf)
4388                 return -ENOMEM;
4389
4390         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
4391                             path->slots[0]), item_size);
4392
4393         slot = path->slots[0] + 1;
4394         nritems = btrfs_header_nritems(leaf);
4395         if (slot != nritems) {
4396                 /* shift the items */
4397                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
4398                                 btrfs_item_nr_offset(slot),
4399                                 (nritems - slot) * sizeof(struct btrfs_item));
4400         }
4401
4402         btrfs_cpu_key_to_disk(&disk_key, new_key);
4403         btrfs_set_item_key(leaf, &disk_key, slot);
4404
4405         new_item = btrfs_item_nr(slot);
4406
4407         btrfs_set_item_offset(leaf, new_item, orig_offset);
4408         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
4409
4410         btrfs_set_item_offset(leaf, item,
4411                               orig_offset + item_size - split_offset);
4412         btrfs_set_item_size(leaf, item, split_offset);
4413
4414         btrfs_set_header_nritems(leaf, nritems + 1);
4415
4416         /* write the data for the start of the original item */
4417         write_extent_buffer(leaf, buf,
4418                             btrfs_item_ptr_offset(leaf, path->slots[0]),
4419                             split_offset);
4420
4421         /* write the data for the new item */
4422         write_extent_buffer(leaf, buf + split_offset,
4423                             btrfs_item_ptr_offset(leaf, slot),
4424                             item_size - split_offset);
4425         btrfs_mark_buffer_dirty(leaf);
4426
4427         BUG_ON(btrfs_leaf_free_space(fs_info, leaf) < 0);
4428         kfree(buf);
4429         return 0;
4430 }
4431
4432 /*
4433  * This function splits a single item into two items,
4434  * giving 'new_key' to the new item and splitting the
4435  * old one at split_offset (from the start of the item).
4436  *
4437  * The path may be released by this operation.  After
4438  * the split, the path is pointing to the old item.  The
4439  * new item is going to be in the same node as the old one.
4440  *
4441  * Note, the item being split must be smaller enough to live alone on
4442  * a tree block with room for one extra struct btrfs_item
4443  *
4444  * This allows us to split the item in place, keeping a lock on the
4445  * leaf the entire time.
4446  */
4447 int btrfs_split_item(struct btrfs_trans_handle *trans,
4448                      struct btrfs_root *root,
4449                      struct btrfs_path *path,
4450                      const struct btrfs_key *new_key,
4451                      unsigned long split_offset)
4452 {
4453         int ret;
4454         ret = setup_leaf_for_split(trans, root, path,
4455                                    sizeof(struct btrfs_item));
4456         if (ret)
4457                 return ret;
4458
4459         ret = split_item(root->fs_info, path, new_key, split_offset);
4460         return ret;
4461 }
4462
4463 /*
4464  * This function duplicate a item, giving 'new_key' to the new item.
4465  * It guarantees both items live in the same tree leaf and the new item
4466  * is contiguous with the original item.
4467  *
4468  * This allows us to split file extent in place, keeping a lock on the
4469  * leaf the entire time.
4470  */
4471 int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
4472                          struct btrfs_root *root,
4473                          struct btrfs_path *path,
4474                          const struct btrfs_key *new_key)
4475 {
4476         struct extent_buffer *leaf;
4477         int ret;
4478         u32 item_size;
4479
4480         leaf = path->nodes[0];
4481         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4482         ret = setup_leaf_for_split(trans, root, path,
4483                                    item_size + sizeof(struct btrfs_item));
4484         if (ret)
4485                 return ret;
4486
4487         path->slots[0]++;
4488         setup_items_for_insert(root, path, new_key, &item_size,
4489                                item_size, item_size +
4490                                sizeof(struct btrfs_item), 1);
4491         leaf = path->nodes[0];
4492         memcpy_extent_buffer(leaf,
4493                              btrfs_item_ptr_offset(leaf, path->slots[0]),
4494                              btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
4495                              item_size);
4496         return 0;
4497 }
4498
4499 /*
4500  * make the item pointed to by the path smaller.  new_size indicates
4501  * how small to make it, and from_end tells us if we just chop bytes
4502  * off the end of the item or if we shift the item to chop bytes off
4503  * the front.
4504  */
4505 void btrfs_truncate_item(struct btrfs_fs_info *fs_info,
4506                          struct btrfs_path *path, u32 new_size, int from_end)
4507 {
4508         int slot;
4509         struct extent_buffer *leaf;
4510         struct btrfs_item *item;
4511         u32 nritems;
4512         unsigned int data_end;
4513         unsigned int old_data_start;
4514         unsigned int old_size;
4515         unsigned int size_diff;
4516         int i;
4517         struct btrfs_map_token token;
4518
4519         btrfs_init_map_token(&token);
4520
4521         leaf = path->nodes[0];
4522         slot = path->slots[0];
4523
4524         old_size = btrfs_item_size_nr(leaf, slot);
4525         if (old_size == new_size)
4526                 return;
4527
4528         nritems = btrfs_header_nritems(leaf);
4529         data_end = leaf_data_end(fs_info, leaf);
4530
4531         old_data_start = btrfs_item_offset_nr(leaf, slot);
4532
4533         size_diff = old_size - new_size;
4534
4535         BUG_ON(slot < 0);
4536         BUG_ON(slot >= nritems);
4537
4538         /*
4539          * item0..itemN ... dataN.offset..dataN.size .. data0.size
4540          */
4541         /* first correct the data pointers */
4542         for (i = slot; i < nritems; i++) {
4543                 u32 ioff;
4544                 item = btrfs_item_nr(i);
4545
4546                 ioff = btrfs_token_item_offset(leaf, item, &token);
4547                 btrfs_set_token_item_offset(leaf, item,
4548                                             ioff + size_diff, &token);
4549         }
4550
4551         /* shift the data */
4552         if (from_end) {
4553                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4554                               data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
4555                               data_end, old_data_start + new_size - data_end);
4556         } else {
4557                 struct btrfs_disk_key disk_key;
4558                 u64 offset;
4559
4560                 btrfs_item_key(leaf, &disk_key, slot);
4561
4562                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
4563                         unsigned long ptr;
4564                         struct btrfs_file_extent_item *fi;
4565
4566                         fi = btrfs_item_ptr(leaf, slot,
4567                                             struct btrfs_file_extent_item);
4568                         fi = (struct btrfs_file_extent_item *)(
4569                              (unsigned long)fi - size_diff);
4570
4571                         if (btrfs_file_extent_type(leaf, fi) ==
4572                             BTRFS_FILE_EXTENT_INLINE) {
4573                                 ptr = btrfs_item_ptr_offset(leaf, slot);
4574                                 memmove_extent_buffer(leaf, ptr,
4575                                       (unsigned long)fi,
4576                                       BTRFS_FILE_EXTENT_INLINE_DATA_START);
4577                         }
4578                 }
4579
4580                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4581                               data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
4582                               data_end, old_data_start - data_end);
4583
4584                 offset = btrfs_disk_key_offset(&disk_key);
4585                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
4586                 btrfs_set_item_key(leaf, &disk_key, slot);
4587                 if (slot == 0)
4588                         fixup_low_keys(path, &disk_key, 1);
4589         }
4590
4591         item = btrfs_item_nr(slot);
4592         btrfs_set_item_size(leaf, item, new_size);
4593         btrfs_mark_buffer_dirty(leaf);
4594
4595         if (btrfs_leaf_free_space(fs_info, leaf) < 0) {
4596                 btrfs_print_leaf(leaf);
4597                 BUG();
4598         }
4599 }
4600
4601 /*
4602  * make the item pointed to by the path bigger, data_size is the added size.
4603  */
4604 void btrfs_extend_item(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
4605                        u32 data_size)
4606 {
4607         int slot;
4608         struct extent_buffer *leaf;
4609         struct btrfs_item *item;
4610         u32 nritems;
4611         unsigned int data_end;
4612         unsigned int old_data;
4613         unsigned int old_size;
4614         int i;
4615         struct btrfs_map_token token;
4616
4617         btrfs_init_map_token(&token);
4618
4619         leaf = path->nodes[0];
4620
4621         nritems = btrfs_header_nritems(leaf);
4622         data_end = leaf_data_end(fs_info, leaf);
4623
4624         if (btrfs_leaf_free_space(fs_info, leaf) < data_size) {
4625                 btrfs_print_leaf(leaf);
4626                 BUG();
4627         }
4628         slot = path->slots[0];
4629         old_data = btrfs_item_end_nr(leaf, slot);
4630
4631         BUG_ON(slot < 0);
4632         if (slot >= nritems) {
4633                 btrfs_print_leaf(leaf);
4634                 btrfs_crit(fs_info, "slot %d too large, nritems %d",
4635                            slot, nritems);
4636                 BUG_ON(1);
4637         }
4638
4639         /*
4640          * item0..itemN ... dataN.offset..dataN.size .. data0.size
4641          */
4642         /* first correct the data pointers */
4643         for (i = slot; i < nritems; i++) {
4644                 u32 ioff;
4645                 item = btrfs_item_nr(i);
4646
4647                 ioff = btrfs_token_item_offset(leaf, item, &token);
4648                 btrfs_set_token_item_offset(leaf, item,
4649                                             ioff - data_size, &token);
4650         }
4651
4652         /* shift the data */
4653         memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4654                       data_end - data_size, BTRFS_LEAF_DATA_OFFSET +
4655                       data_end, old_data - data_end);
4656
4657         data_end = old_data;
4658         old_size = btrfs_item_size_nr(leaf, slot);
4659         item = btrfs_item_nr(slot);
4660         btrfs_set_item_size(leaf, item, old_size + data_size);
4661         btrfs_mark_buffer_dirty(leaf);
4662
4663         if (btrfs_leaf_free_space(fs_info, leaf) < 0) {
4664                 btrfs_print_leaf(leaf);
4665                 BUG();
4666         }
4667 }
4668
4669 /*
4670  * this is a helper for btrfs_insert_empty_items, the main goal here is
4671  * to save stack depth by doing the bulk of the work in a function
4672  * that doesn't call btrfs_search_slot
4673  */
4674 void setup_items_for_insert(struct btrfs_root *root, struct btrfs_path *path,
4675                             const struct btrfs_key *cpu_key, u32 *data_size,
4676                             u32 total_data, u32 total_size, int nr)
4677 {
4678         struct btrfs_fs_info *fs_info = root->fs_info;
4679         struct btrfs_item *item;
4680         int i;
4681         u32 nritems;
4682         unsigned int data_end;
4683         struct btrfs_disk_key disk_key;
4684         struct extent_buffer *leaf;
4685         int slot;
4686         struct btrfs_map_token token;
4687
4688         if (path->slots[0] == 0) {
4689                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
4690                 fixup_low_keys(path, &disk_key, 1);
4691         }
4692         btrfs_unlock_up_safe(path, 1);
4693
4694         btrfs_init_map_token(&token);
4695
4696         leaf = path->nodes[0];
4697         slot = path->slots[0];
4698
4699         nritems = btrfs_header_nritems(leaf);
4700         data_end = leaf_data_end(fs_info, leaf);
4701
4702         if (btrfs_leaf_free_space(fs_info, leaf) < total_size) {
4703                 btrfs_print_leaf(leaf);
4704                 btrfs_crit(fs_info, "not enough freespace need %u have %d",
4705                            total_size, btrfs_leaf_free_space(fs_info, leaf));
4706                 BUG();
4707         }
4708
4709         if (slot != nritems) {
4710                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
4711
4712                 if (old_data < data_end) {
4713                         btrfs_print_leaf(leaf);
4714                         btrfs_crit(fs_info, "slot %d old_data %d data_end %d",
4715                                    slot, old_data, data_end);
4716                         BUG_ON(1);
4717                 }
4718                 /*
4719                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
4720                  */
4721                 /* first correct the data pointers */
4722                 for (i = slot; i < nritems; i++) {
4723                         u32 ioff;
4724
4725                         item = btrfs_item_nr(i);
4726                         ioff = btrfs_token_item_offset(leaf, item, &token);
4727                         btrfs_set_token_item_offset(leaf, item,
4728                                                     ioff - total_data, &token);
4729                 }
4730                 /* shift the items */
4731                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
4732                               btrfs_item_nr_offset(slot),
4733                               (nritems - slot) * sizeof(struct btrfs_item));
4734
4735                 /* shift the data */
4736                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4737                               data_end - total_data, BTRFS_LEAF_DATA_OFFSET +
4738                               data_end, old_data - data_end);
4739                 data_end = old_data;
4740         }
4741
4742         /* setup the item for the new data */
4743         for (i = 0; i < nr; i++) {
4744                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
4745                 btrfs_set_item_key(leaf, &disk_key, slot + i);
4746                 item = btrfs_item_nr(slot + i);
4747                 btrfs_set_token_item_offset(leaf, item,
4748                                             data_end - data_size[i], &token);
4749                 data_end -= data_size[i];
4750                 btrfs_set_token_item_size(leaf, item, data_size[i], &token);
4751         }
4752
4753         btrfs_set_header_nritems(leaf, nritems + nr);
4754         btrfs_mark_buffer_dirty(leaf);
4755
4756         if (btrfs_leaf_free_space(fs_info, leaf) < 0) {
4757                 btrfs_print_leaf(leaf);
4758                 BUG();
4759         }
4760 }
4761
4762 /*
4763  * Given a key and some data, insert items into the tree.
4764  * This does all the path init required, making room in the tree if needed.
4765  */
4766 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
4767                             struct btrfs_root *root,
4768                             struct btrfs_path *path,
4769                             const struct btrfs_key *cpu_key, u32 *data_size,
4770                             int nr)
4771 {
4772         int ret = 0;
4773         int slot;
4774         int i;
4775         u32 total_size = 0;
4776         u32 total_data = 0;
4777
4778         for (i = 0; i < nr; i++)
4779                 total_data += data_size[i];
4780
4781         total_size = total_data + (nr * sizeof(struct btrfs_item));
4782         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
4783         if (ret == 0)
4784                 return -EEXIST;
4785         if (ret < 0)
4786                 return ret;
4787
4788         slot = path->slots[0];
4789         BUG_ON(slot < 0);
4790
4791         setup_items_for_insert(root, path, cpu_key, data_size,
4792                                total_data, total_size, nr);
4793         return 0;
4794 }
4795
4796 /*
4797  * Given a key and some data, insert an item into the tree.
4798  * This does all the path init required, making room in the tree if needed.
4799  */
4800 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4801                       const struct btrfs_key *cpu_key, void *data,
4802                       u32 data_size)
4803 {
4804         int ret = 0;
4805         struct btrfs_path *path;
4806         struct extent_buffer *leaf;
4807         unsigned long ptr;
4808
4809         path = btrfs_alloc_path();
4810         if (!path)
4811                 return -ENOMEM;
4812         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
4813         if (!ret) {
4814                 leaf = path->nodes[0];
4815                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
4816                 write_extent_buffer(leaf, data, ptr, data_size);
4817                 btrfs_mark_buffer_dirty(leaf);
4818         }
4819         btrfs_free_path(path);
4820         return ret;
4821 }
4822
4823 /*
4824  * delete the pointer from a given node.
4825  *
4826  * the tree should have been previously balanced so the deletion does not
4827  * empty a node.
4828  */
4829 static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
4830                     int level, int slot)
4831 {
4832         struct extent_buffer *parent = path->nodes[level];
4833         u32 nritems;
4834         int ret;
4835
4836         nritems = btrfs_header_nritems(parent);
4837         if (slot != nritems - 1) {
4838                 if (level) {
4839                         ret = tree_mod_log_insert_move(parent, slot, slot + 1,
4840                                         nritems - slot - 1);
4841                         BUG_ON(ret < 0);
4842                 }
4843                 memmove_extent_buffer(parent,
4844                               btrfs_node_key_ptr_offset(slot),
4845                               btrfs_node_key_ptr_offset(slot + 1),
4846                               sizeof(struct btrfs_key_ptr) *
4847                               (nritems - slot - 1));
4848         } else if (level) {
4849                 ret = tree_mod_log_insert_key(parent, slot, MOD_LOG_KEY_REMOVE,
4850                                 GFP_NOFS);
4851                 BUG_ON(ret < 0);
4852         }
4853
4854         nritems--;
4855         btrfs_set_header_nritems(parent, nritems);
4856         if (nritems == 0 && parent == root->node) {
4857                 BUG_ON(btrfs_header_level(root->node) != 1);
4858                 /* just turn the root into a leaf and break */
4859                 btrfs_set_header_level(root->node, 0);
4860         } else if (slot == 0) {
4861                 struct btrfs_disk_key disk_key;
4862
4863                 btrfs_node_key(parent, &disk_key, 0);
4864                 fixup_low_keys(path, &disk_key, level + 1);
4865         }
4866         btrfs_mark_buffer_dirty(parent);
4867 }
4868
4869 /*
4870  * a helper function to delete the leaf pointed to by path->slots[1] and
4871  * path->nodes[1].
4872  *
4873  * This deletes the pointer in path->nodes[1] and frees the leaf
4874  * block extent.  zero is returned if it all worked out, < 0 otherwise.
4875  *
4876  * The path must have already been setup for deleting the leaf, including
4877  * all the proper balancing.  path->nodes[1] must be locked.
4878  */
4879 static noinline void btrfs_del_leaf(struct btrfs_trans_handle *trans,
4880                                     struct btrfs_root *root,
4881                                     struct btrfs_path *path,
4882                                     struct extent_buffer *leaf)
4883 {
4884         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
4885         del_ptr(root, path, 1, path->slots[1]);
4886
4887         /*
4888          * btrfs_free_extent is expensive, we want to make sure we
4889          * aren't holding any locks when we call it
4890          */
4891         btrfs_unlock_up_safe(path, 0);
4892
4893         root_sub_used(root, leaf->len);
4894
4895         extent_buffer_get(leaf);
4896         btrfs_free_tree_block(trans, root, leaf, 0, 1);
4897         free_extent_buffer_stale(leaf);
4898 }
4899 /*
4900  * delete the item at the leaf level in path.  If that empties
4901  * the leaf, remove it from the tree
4902  */
4903 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4904                     struct btrfs_path *path, int slot, int nr)
4905 {
4906         struct btrfs_fs_info *fs_info = root->fs_info;
4907         struct extent_buffer *leaf;
4908         struct btrfs_item *item;
4909         u32 last_off;
4910         u32 dsize = 0;
4911         int ret = 0;
4912         int wret;
4913         int i;
4914         u32 nritems;
4915         struct btrfs_map_token token;
4916
4917         btrfs_init_map_token(&token);
4918
4919         leaf = path->nodes[0];
4920         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
4921
4922         for (i = 0; i < nr; i++)
4923                 dsize += btrfs_item_size_nr(leaf, slot + i);
4924
4925         nritems = btrfs_header_nritems(leaf);
4926
4927         if (slot + nr != nritems) {
4928                 int data_end = leaf_data_end(fs_info, leaf);
4929
4930                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4931                               data_end + dsize,
4932                               BTRFS_LEAF_DATA_OFFSET + data_end,
4933                               last_off - data_end);
4934
4935                 for (i = slot + nr; i < nritems; i++) {
4936                         u32 ioff;
4937
4938                         item = btrfs_item_nr(i);
4939                         ioff = btrfs_token_item_offset(leaf, item, &token);
4940                         btrfs_set_token_item_offset(leaf, item,
4941                                                     ioff + dsize, &token);
4942                 }
4943
4944                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
4945                               btrfs_item_nr_offset(slot + nr),
4946                               sizeof(struct btrfs_item) *
4947                               (nritems - slot - nr));
4948         }
4949         btrfs_set_header_nritems(leaf, nritems - nr);
4950         nritems -= nr;
4951
4952         /* delete the leaf if we've emptied it */
4953         if (nritems == 0) {
4954                 if (leaf == root->node) {
4955                         btrfs_set_header_level(leaf, 0);
4956                 } else {
4957                         btrfs_set_path_blocking(path);
4958                         clean_tree_block(fs_info, leaf);
4959                         btrfs_del_leaf(trans, root, path, leaf);
4960                 }
4961         } else {
4962                 int used = leaf_space_used(leaf, 0, nritems);
4963                 if (slot == 0) {
4964                         struct btrfs_disk_key disk_key;
4965
4966                         btrfs_item_key(leaf, &disk_key, 0);
4967                         fixup_low_keys(path, &disk_key, 1);
4968                 }
4969
4970                 /* delete the leaf if it is mostly empty */
4971                 if (used < BTRFS_LEAF_DATA_SIZE(fs_info) / 3) {
4972                         /* push_leaf_left fixes the path.
4973                          * make sure the path still points to our leaf
4974                          * for possible call to del_ptr below
4975                          */
4976                         slot = path->slots[1];
4977                         extent_buffer_get(leaf);
4978
4979                         btrfs_set_path_blocking(path);
4980                         wret = push_leaf_left(trans, root, path, 1, 1,
4981                                               1, (u32)-1);
4982                         if (wret < 0 && wret != -ENOSPC)
4983                                 ret = wret;
4984
4985                         if (path->nodes[0] == leaf &&
4986                             btrfs_header_nritems(leaf)) {
4987                                 wret = push_leaf_right(trans, root, path, 1,
4988                                                        1, 1, 0);
4989                                 if (wret < 0 && wret != -ENOSPC)
4990                                         ret = wret;
4991                         }
4992
4993                         if (btrfs_header_nritems(leaf) == 0) {
4994                                 path->slots[1] = slot;
4995                                 btrfs_del_leaf(trans, root, path, leaf);
4996                                 free_extent_buffer(leaf);
4997                                 ret = 0;
4998                         } else {
4999                                 /* if we're still in the path, make sure
5000                                  * we're dirty.  Otherwise, one of the
5001                                  * push_leaf functions must have already
5002                                  * dirtied this buffer
5003                                  */
5004                                 if (path->nodes[0] == leaf)
5005                                         btrfs_mark_buffer_dirty(leaf);
5006                                 free_extent_buffer(leaf);
5007                         }
5008                 } else {
5009                         btrfs_mark_buffer_dirty(leaf);
5010                 }
5011         }
5012         return ret;
5013 }
5014
5015 /*
5016  * search the tree again to find a leaf with lesser keys
5017  * returns 0 if it found something or 1 if there are no lesser leaves.
5018  * returns < 0 on io errors.
5019  *
5020  * This may release the path, and so you may lose any locks held at the
5021  * time you call it.
5022  */
5023 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
5024 {
5025         struct btrfs_key key;
5026         struct btrfs_disk_key found_key;
5027         int ret;
5028
5029         btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
5030
5031         if (key.offset > 0) {
5032                 key.offset--;
5033         } else if (key.type > 0) {
5034                 key.type--;
5035                 key.offset = (u64)-1;
5036         } else if (key.objectid > 0) {
5037                 key.objectid--;
5038                 key.type = (u8)-1;
5039                 key.offset = (u64)-1;
5040         } else {
5041                 return 1;
5042         }
5043
5044         btrfs_release_path(path);
5045         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5046         if (ret < 0)
5047                 return ret;
5048         btrfs_item_key(path->nodes[0], &found_key, 0);
5049         ret = comp_keys(&found_key, &key);
5050         /*
5051          * We might have had an item with the previous key in the tree right
5052          * before we released our path. And after we released our path, that
5053          * item might have been pushed to the first slot (0) of the leaf we
5054          * were holding due to a tree balance. Alternatively, an item with the
5055          * previous key can exist as the only element of a leaf (big fat item).
5056          * Therefore account for these 2 cases, so that our callers (like
5057          * btrfs_previous_item) don't miss an existing item with a key matching
5058          * the previous key we computed above.
5059          */
5060         if (ret <= 0)
5061                 return 0;
5062         return 1;
5063 }
5064
5065 /*
5066  * A helper function to walk down the tree starting at min_key, and looking
5067  * for nodes or leaves that are have a minimum transaction id.
5068  * This is used by the btree defrag code, and tree logging
5069  *
5070  * This does not cow, but it does stuff the starting key it finds back
5071  * into min_key, so you can call btrfs_search_slot with cow=1 on the
5072  * key and get a writable path.
5073  *
5074  * This honors path->lowest_level to prevent descent past a given level
5075  * of the tree.
5076  *
5077  * min_trans indicates the oldest transaction that you are interested
5078  * in walking through.  Any nodes or leaves older than min_trans are
5079  * skipped over (without reading them).
5080  *
5081  * returns zero if something useful was found, < 0 on error and 1 if there
5082  * was nothing in the tree that matched the search criteria.
5083  */
5084 int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
5085                          struct btrfs_path *path,
5086                          u64 min_trans)
5087 {
5088         struct btrfs_fs_info *fs_info = root->fs_info;
5089         struct extent_buffer *cur;
5090         struct btrfs_key found_key;
5091         int slot;
5092         int sret;
5093         u32 nritems;
5094         int level;
5095         int ret = 1;
5096         int keep_locks = path->keep_locks;
5097
5098         path->keep_locks = 1;
5099 again:
5100         cur = btrfs_read_lock_root_node(root);
5101         level = btrfs_header_level(cur);
5102         WARN_ON(path->nodes[level]);
5103         path->nodes[level] = cur;
5104         path->locks[level] = BTRFS_READ_LOCK;
5105
5106         if (btrfs_header_generation(cur) < min_trans) {
5107                 ret = 1;
5108                 goto out;
5109         }
5110         while (1) {
5111                 nritems = btrfs_header_nritems(cur);
5112                 level = btrfs_header_level(cur);
5113                 sret = btrfs_bin_search(cur, min_key, level, &slot);
5114
5115                 /* at the lowest level, we're done, setup the path and exit */
5116                 if (level == path->lowest_level) {
5117                         if (slot >= nritems)
5118                                 goto find_next_key;
5119                         ret = 0;
5120                         path->slots[level] = slot;
5121                         btrfs_item_key_to_cpu(cur, &found_key, slot);
5122                         goto out;
5123                 }
5124                 if (sret && slot > 0)
5125                         slot--;
5126                 /*
5127                  * check this node pointer against the min_trans parameters.
5128                  * If it is too old, old, skip to the next one.
5129                  */
5130                 while (slot < nritems) {
5131                         u64 gen;
5132
5133                         gen = btrfs_node_ptr_generation(cur, slot);
5134                         if (gen < min_trans) {
5135                                 slot++;
5136                                 continue;
5137                         }
5138                         break;
5139                 }
5140 find_next_key:
5141                 /*
5142                  * we didn't find a candidate key in this node, walk forward
5143                  * and find another one
5144                  */
5145                 if (slot >= nritems) {
5146                         path->slots[level] = slot;
5147                         btrfs_set_path_blocking(path);
5148                         sret = btrfs_find_next_key(root, path, min_key, level,
5149                                                   min_trans);
5150                         if (sret == 0) {
5151                                 btrfs_release_path(path);
5152                                 goto again;
5153                         } else {
5154                                 goto out;
5155                         }
5156                 }
5157                 /* save our key for returning back */
5158                 btrfs_node_key_to_cpu(cur, &found_key, slot);
5159                 path->slots[level] = slot;
5160                 if (level == path->lowest_level) {
5161                         ret = 0;
5162                         goto out;
5163                 }
5164                 btrfs_set_path_blocking(path);
5165                 cur = read_node_slot(fs_info, cur, slot);
5166                 if (IS_ERR(cur)) {
5167                         ret = PTR_ERR(cur);
5168                         goto out;
5169                 }
5170
5171                 btrfs_tree_read_lock(cur);
5172
5173                 path->locks[level - 1] = BTRFS_READ_LOCK;
5174                 path->nodes[level - 1] = cur;
5175                 unlock_up(path, level, 1, 0, NULL);
5176         }
5177 out:
5178         path->keep_locks = keep_locks;
5179         if (ret == 0) {
5180                 btrfs_unlock_up_safe(path, path->lowest_level + 1);
5181                 btrfs_set_path_blocking(path);
5182                 memcpy(min_key, &found_key, sizeof(found_key));
5183         }
5184         return ret;
5185 }
5186
5187 static int tree_move_down(struct btrfs_fs_info *fs_info,
5188                            struct btrfs_path *path,
5189                            int *level)
5190 {
5191         struct extent_buffer *eb;
5192
5193         BUG_ON(*level == 0);
5194         eb = read_node_slot(fs_info, path->nodes[*level], path->slots[*level]);
5195         if (IS_ERR(eb))
5196                 return PTR_ERR(eb);
5197
5198         path->nodes[*level - 1] = eb;
5199         path->slots[*level - 1] = 0;
5200         (*level)--;
5201         return 0;
5202 }
5203
5204 static int tree_move_next_or_upnext(struct btrfs_path *path,
5205                                     int *level, int root_level)
5206 {
5207         int ret = 0;
5208         int nritems;
5209         nritems = btrfs_header_nritems(path->nodes[*level]);
5210
5211         path->slots[*level]++;
5212
5213         while (path->slots[*level] >= nritems) {
5214                 if (*level == root_level)
5215                         return -1;
5216
5217                 /* move upnext */
5218                 path->slots[*level] = 0;
5219                 free_extent_buffer(path->nodes[*level]);
5220                 path->nodes[*level] = NULL;
5221                 (*level)++;
5222                 path->slots[*level]++;
5223
5224                 nritems = btrfs_header_nritems(path->nodes[*level]);
5225                 ret = 1;
5226         }
5227         return ret;
5228 }
5229
5230 /*
5231  * Returns 1 if it had to move up and next. 0 is returned if it moved only next
5232  * or down.
5233  */
5234 static int tree_advance(struct btrfs_fs_info *fs_info,
5235                         struct btrfs_path *path,
5236                         int *level, int root_level,
5237                         int allow_down,
5238                         struct btrfs_key *key)
5239 {
5240         int ret;
5241
5242         if (*level == 0 || !allow_down) {
5243                 ret = tree_move_next_or_upnext(path, level, root_level);
5244         } else {
5245                 ret = tree_move_down(fs_info, path, level);
5246         }
5247         if (ret >= 0) {
5248                 if (*level == 0)
5249                         btrfs_item_key_to_cpu(path->nodes[*level], key,
5250                                         path->slots[*level]);
5251                 else
5252                         btrfs_node_key_to_cpu(path->nodes[*level], key,
5253                                         path->slots[*level]);
5254         }
5255         return ret;
5256 }
5257
5258 static int tree_compare_item(struct btrfs_path *left_path,
5259                              struct btrfs_path *right_path,
5260                              char *tmp_buf)
5261 {
5262         int cmp;
5263         int len1, len2;
5264         unsigned long off1, off2;
5265
5266         len1 = btrfs_item_size_nr(left_path->nodes[0], left_path->slots[0]);
5267         len2 = btrfs_item_size_nr(right_path->nodes[0], right_path->slots[0]);
5268         if (len1 != len2)
5269                 return 1;
5270
5271         off1 = btrfs_item_ptr_offset(left_path->nodes[0], left_path->slots[0]);
5272         off2 = btrfs_item_ptr_offset(right_path->nodes[0],
5273                                 right_path->slots[0]);
5274
5275         read_extent_buffer(left_path->nodes[0], tmp_buf, off1, len1);
5276
5277         cmp = memcmp_extent_buffer(right_path->nodes[0], tmp_buf, off2, len1);
5278         if (cmp)
5279                 return 1;
5280         return 0;
5281 }
5282
5283 #define ADVANCE 1
5284 #define ADVANCE_ONLY_NEXT -1
5285
5286 /*
5287  * This function compares two trees and calls the provided callback for
5288  * every changed/new/deleted item it finds.
5289  * If shared tree blocks are encountered, whole subtrees are skipped, making
5290  * the compare pretty fast on snapshotted subvolumes.
5291  *
5292  * This currently works on commit roots only. As commit roots are read only,
5293  * we don't do any locking. The commit roots are protected with transactions.
5294  * Transactions are ended and rejoined when a commit is tried in between.
5295  *
5296  * This function checks for modifications done to the trees while comparing.
5297  * If it detects a change, it aborts immediately.
5298  */
5299 int btrfs_compare_trees(struct btrfs_root *left_root,
5300                         struct btrfs_root *right_root,
5301                         btrfs_changed_cb_t changed_cb, void *ctx)
5302 {
5303         struct btrfs_fs_info *fs_info = left_root->fs_info;
5304         int ret;
5305         int cmp;
5306         struct btrfs_path *left_path = NULL;
5307         struct btrfs_path *right_path = NULL;
5308         struct btrfs_key left_key;
5309         struct btrfs_key right_key;
5310         char *tmp_buf = NULL;
5311         int left_root_level;
5312         int right_root_level;
5313         int left_level;
5314         int right_level;
5315         int left_end_reached;
5316         int right_end_reached;
5317         int advance_left;
5318         int advance_right;
5319         u64 left_blockptr;
5320         u64 right_blockptr;
5321         u64 left_gen;
5322         u64 right_gen;
5323
5324         left_path = btrfs_alloc_path();
5325         if (!left_path) {
5326                 ret = -ENOMEM;
5327                 goto out;
5328         }
5329         right_path = btrfs_alloc_path();
5330         if (!right_path) {
5331                 ret = -ENOMEM;
5332                 goto out;
5333         }
5334
5335         tmp_buf = kvmalloc(fs_info->nodesize, GFP_KERNEL);
5336         if (!tmp_buf) {
5337                 ret = -ENOMEM;
5338                 goto out;
5339         }
5340
5341         left_path->search_commit_root = 1;
5342         left_path->skip_locking = 1;
5343         right_path->search_commit_root = 1;
5344         right_path->skip_locking = 1;
5345
5346         /*
5347          * Strategy: Go to the first items of both trees. Then do
5348          *
5349          * If both trees are at level 0
5350          *   Compare keys of current items
5351          *     If left < right treat left item as new, advance left tree
5352          *       and repeat
5353          *     If left > right treat right item as deleted, advance right tree
5354          *       and repeat
5355          *     If left == right do deep compare of items, treat as changed if
5356          *       needed, advance both trees and repeat
5357          * If both trees are at the same level but not at level 0
5358          *   Compare keys of current nodes/leafs
5359          *     If left < right advance left tree and repeat
5360          *     If left > right advance right tree and repeat
5361          *     If left == right compare blockptrs of the next nodes/leafs
5362          *       If they match advance both trees but stay at the same level
5363          *         and repeat
5364          *       If they don't match advance both trees while allowing to go
5365          *         deeper and repeat
5366          * If tree levels are different
5367          *   Advance the tree that needs it and repeat
5368          *
5369          * Advancing a tree means:
5370          *   If we are at level 0, try to go to the next slot. If that's not
5371          *   possible, go one level up and repeat. Stop when we found a level
5372          *   where we could go to the next slot. We may at this point be on a
5373          *   node or a leaf.
5374          *
5375          *   If we are not at level 0 and not on shared tree blocks, go one
5376          *   level deeper.
5377          *
5378          *   If we are not at level 0 and on shared tree blocks, go one slot to
5379          *   the right if possible or go up and right.
5380          */
5381
5382         down_read(&fs_info->commit_root_sem);
5383         left_level = btrfs_header_level(left_root->commit_root);
5384         left_root_level = left_level;
5385         left_path->nodes[left_level] =
5386                         btrfs_clone_extent_buffer(left_root->commit_root);
5387         if (!left_path->nodes[left_level]) {
5388                 up_read(&fs_info->commit_root_sem);
5389                 ret = -ENOMEM;
5390                 goto out;
5391         }
5392
5393         right_level = btrfs_header_level(right_root->commit_root);
5394         right_root_level = right_level;
5395         right_path->nodes[right_level] =
5396                         btrfs_clone_extent_buffer(right_root->commit_root);
5397         if (!right_path->nodes[right_level]) {
5398                 up_read(&fs_info->commit_root_sem);
5399                 ret = -ENOMEM;
5400                 goto out;
5401         }
5402         up_read(&fs_info->commit_root_sem);
5403
5404         if (left_level == 0)
5405                 btrfs_item_key_to_cpu(left_path->nodes[left_level],
5406                                 &left_key, left_path->slots[left_level]);
5407         else
5408                 btrfs_node_key_to_cpu(left_path->nodes[left_level],
5409                                 &left_key, left_path->slots[left_level]);
5410         if (right_level == 0)
5411                 btrfs_item_key_to_cpu(right_path->nodes[right_level],
5412                                 &right_key, right_path->slots[right_level]);
5413         else
5414                 btrfs_node_key_to_cpu(right_path->nodes[right_level],
5415                                 &right_key, right_path->slots[right_level]);
5416
5417         left_end_reached = right_end_reached = 0;
5418         advance_left = advance_right = 0;
5419
5420         while (1) {
5421                 if (advance_left && !left_end_reached) {
5422                         ret = tree_advance(fs_info, left_path, &left_level,
5423                                         left_root_level,
5424                                         advance_left != ADVANCE_ONLY_NEXT,
5425                                         &left_key);
5426                         if (ret == -1)
5427                                 left_end_reached = ADVANCE;
5428                         else if (ret < 0)
5429                                 goto out;
5430                         advance_left = 0;
5431                 }
5432                 if (advance_right && !right_end_reached) {
5433                         ret = tree_advance(fs_info, right_path, &right_level,
5434                                         right_root_level,
5435                                         advance_right != ADVANCE_ONLY_NEXT,
5436                                         &right_key);
5437                         if (ret == -1)
5438                                 right_end_reached = ADVANCE;
5439                         else if (ret < 0)
5440                                 goto out;
5441                         advance_right = 0;
5442                 }
5443
5444                 if (left_end_reached && right_end_reached) {
5445                         ret = 0;
5446                         goto out;
5447                 } else if (left_end_reached) {
5448                         if (right_level == 0) {
5449                                 ret = changed_cb(left_path, right_path,
5450                                                 &right_key,
5451                                                 BTRFS_COMPARE_TREE_DELETED,
5452                                                 ctx);
5453                                 if (ret < 0)
5454                                         goto out;
5455                         }
5456                         advance_right = ADVANCE;
5457                         continue;
5458                 } else if (right_end_reached) {
5459                         if (left_level == 0) {
5460                                 ret = changed_cb(left_path, right_path,
5461                                                 &left_key,
5462                                                 BTRFS_COMPARE_TREE_NEW,
5463                                                 ctx);
5464                                 if (ret < 0)
5465                                         goto out;
5466                         }
5467                         advance_left = ADVANCE;
5468                         continue;
5469                 }
5470
5471                 if (left_level == 0 && right_level == 0) {
5472                         cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
5473                         if (cmp < 0) {
5474                                 ret = changed_cb(left_path, right_path,
5475                                                 &left_key,
5476                                                 BTRFS_COMPARE_TREE_NEW,
5477                                                 ctx);
5478                                 if (ret < 0)
5479                                         goto out;
5480                                 advance_left = ADVANCE;
5481                         } else if (cmp > 0) {
5482                                 ret = changed_cb(left_path, right_path,
5483                                                 &right_key,
5484                                                 BTRFS_COMPARE_TREE_DELETED,
5485                                                 ctx);
5486                                 if (ret < 0)
5487                                         goto out;
5488                                 advance_right = ADVANCE;
5489                         } else {
5490                                 enum btrfs_compare_tree_result result;
5491
5492                                 WARN_ON(!extent_buffer_uptodate(left_path->nodes[0]));
5493                                 ret = tree_compare_item(left_path, right_path,
5494                                                         tmp_buf);
5495                                 if (ret)
5496                                         result = BTRFS_COMPARE_TREE_CHANGED;
5497                                 else
5498                                         result = BTRFS_COMPARE_TREE_SAME;
5499                                 ret = changed_cb(left_path, right_path,
5500                                                  &left_key, result, ctx);
5501                                 if (ret < 0)
5502                                         goto out;
5503                                 advance_left = ADVANCE;
5504                                 advance_right = ADVANCE;
5505                         }
5506                 } else if (left_level == right_level) {
5507                         cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
5508                         if (cmp < 0) {
5509                                 advance_left = ADVANCE;
5510                         } else if (cmp > 0) {
5511                                 advance_right = ADVANCE;
5512                         } else {
5513                                 left_blockptr = btrfs_node_blockptr(
5514                                                 left_path->nodes[left_level],
5515                                                 left_path->slots[left_level]);
5516                                 right_blockptr = btrfs_node_blockptr(
5517                                                 right_path->nodes[right_level],
5518                                                 right_path->slots[right_level]);
5519                                 left_gen = btrfs_node_ptr_generation(
5520                                                 left_path->nodes[left_level],
5521                                                 left_path->slots[left_level]);
5522                                 right_gen = btrfs_node_ptr_generation(
5523                                                 right_path->nodes[right_level],
5524                                                 right_path->slots[right_level]);
5525                                 if (left_blockptr == right_blockptr &&
5526                                     left_gen == right_gen) {
5527                                         /*
5528                                          * As we're on a shared block, don't
5529                                          * allow to go deeper.
5530                                          */
5531                                         advance_left = ADVANCE_ONLY_NEXT;
5532                                         advance_right = ADVANCE_ONLY_NEXT;
5533                                 } else {
5534                                         advance_left = ADVANCE;
5535                                         advance_right = ADVANCE;
5536                                 }
5537                         }
5538                 } else if (left_level < right_level) {
5539                         advance_right = ADVANCE;
5540                 } else {
5541                         advance_left = ADVANCE;
5542                 }
5543         }
5544
5545 out:
5546         btrfs_free_path(left_path);
5547         btrfs_free_path(right_path);
5548         kvfree(tmp_buf);
5549         return ret;
5550 }
5551
5552 /*
5553  * this is similar to btrfs_next_leaf, but does not try to preserve
5554  * and fixup the path.  It looks for and returns the next key in the
5555  * tree based on the current path and the min_trans parameters.
5556  *
5557  * 0 is returned if another key is found, < 0 if there are any errors
5558  * and 1 is returned if there are no higher keys in the tree
5559  *
5560  * path->keep_locks should be set to 1 on the search made before
5561  * calling this function.
5562  */
5563 int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
5564                         struct btrfs_key *key, int level, u64 min_trans)
5565 {
5566         int slot;
5567         struct extent_buffer *c;
5568
5569         WARN_ON(!path->keep_locks);
5570         while (level < BTRFS_MAX_LEVEL) {
5571                 if (!path->nodes[level])
5572                         return 1;
5573
5574                 slot = path->slots[level] + 1;
5575                 c = path->nodes[level];
5576 next:
5577                 if (slot >= btrfs_header_nritems(c)) {
5578                         int ret;
5579                         int orig_lowest;
5580                         struct btrfs_key cur_key;
5581                         if (level + 1 >= BTRFS_MAX_LEVEL ||
5582                             !path->nodes[level + 1])
5583                                 return 1;
5584
5585                         if (path->locks[level + 1]) {
5586                                 level++;
5587                                 continue;
5588                         }
5589
5590                         slot = btrfs_header_nritems(c) - 1;
5591                         if (level == 0)
5592                                 btrfs_item_key_to_cpu(c, &cur_key, slot);
5593                         else
5594                                 btrfs_node_key_to_cpu(c, &cur_key, slot);
5595
5596                         orig_lowest = path->lowest_level;
5597                         btrfs_release_path(path);
5598                         path->lowest_level = level;
5599                         ret = btrfs_search_slot(NULL, root, &cur_key, path,
5600                                                 0, 0);
5601                         path->lowest_level = orig_lowest;
5602                         if (ret < 0)
5603                                 return ret;
5604
5605                         c = path->nodes[level];
5606                         slot = path->slots[level];
5607                         if (ret == 0)
5608                                 slot++;
5609                         goto next;
5610                 }
5611
5612                 if (level == 0)
5613                         btrfs_item_key_to_cpu(c, key, slot);
5614                 else {
5615                         u64 gen = btrfs_node_ptr_generation(c, slot);
5616
5617                         if (gen < min_trans) {
5618                                 slot++;
5619                                 goto next;
5620                         }
5621                         btrfs_node_key_to_cpu(c, key, slot);
5622                 }
5623                 return 0;
5624         }
5625         return 1;
5626 }
5627
5628 /*
5629  * search the tree again to find a leaf with greater keys
5630  * returns 0 if it found something or 1 if there are no greater leaves.
5631  * returns < 0 on io errors.
5632  */
5633 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
5634 {
5635         return btrfs_next_old_leaf(root, path, 0);
5636 }
5637
5638 int btrfs_next_old_leaf(struct btrfs_root *root, struct btrfs_path *path,
5639                         u64 time_seq)
5640 {
5641         int slot;
5642         int level;
5643         struct extent_buffer *c;
5644         struct extent_buffer *next;
5645         struct btrfs_key key;
5646         u32 nritems;
5647         int ret;
5648         int old_spinning = path->leave_spinning;
5649         int next_rw_lock = 0;
5650
5651         nritems = btrfs_header_nritems(path->nodes[0]);
5652         if (nritems == 0)
5653                 return 1;
5654
5655         btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
5656 again:
5657         level = 1;
5658         next = NULL;
5659         next_rw_lock = 0;
5660         btrfs_release_path(path);
5661
5662         path->keep_locks = 1;
5663         path->leave_spinning = 1;
5664
5665         if (time_seq)
5666                 ret = btrfs_search_old_slot(root, &key, path, time_seq);
5667         else
5668                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5669         path->keep_locks = 0;
5670
5671         if (ret < 0)
5672                 return ret;
5673
5674         nritems = btrfs_header_nritems(path->nodes[0]);
5675         /*
5676          * by releasing the path above we dropped all our locks.  A balance
5677          * could have added more items next to the key that used to be
5678          * at the very end of the block.  So, check again here and
5679          * advance the path if there are now more items available.
5680          */
5681         if (nritems > 0 && path->slots[0] < nritems - 1) {
5682                 if (ret == 0)
5683                         path->slots[0]++;
5684                 ret = 0;
5685                 goto done;
5686         }
5687         /*
5688          * So the above check misses one case:
5689          * - after releasing the path above, someone has removed the item that
5690          *   used to be at the very end of the block, and balance between leafs
5691          *   gets another one with bigger key.offset to replace it.
5692          *
5693          * This one should be returned as well, or we can get leaf corruption
5694          * later(esp. in __btrfs_drop_extents()).
5695          *
5696          * And a bit more explanation about this check,
5697          * with ret > 0, the key isn't found, the path points to the slot
5698          * where it should be inserted, so the path->slots[0] item must be the
5699          * bigger one.
5700          */
5701         if (nritems > 0 && ret > 0 && path->slots[0] == nritems - 1) {
5702                 ret = 0;
5703                 goto done;
5704         }
5705
5706         while (level < BTRFS_MAX_LEVEL) {
5707                 if (!path->nodes[level]) {
5708                         ret = 1;
5709                         goto done;
5710                 }
5711
5712                 slot = path->slots[level] + 1;
5713                 c = path->nodes[level];
5714                 if (slot >= btrfs_header_nritems(c)) {
5715                         level++;
5716                         if (level == BTRFS_MAX_LEVEL) {
5717                                 ret = 1;
5718                                 goto done;
5719                         }
5720                         continue;
5721                 }
5722
5723                 if (next) {
5724                         btrfs_tree_unlock_rw(next, next_rw_lock);
5725                         free_extent_buffer(next);
5726                 }
5727
5728                 next = c;
5729                 next_rw_lock = path->locks[level];
5730                 ret = read_block_for_search(root, path, &next, level,
5731                                             slot, &key);
5732                 if (ret == -EAGAIN)
5733                         goto again;
5734
5735                 if (ret < 0) {
5736                         btrfs_release_path(path);
5737                         goto done;
5738                 }
5739
5740                 if (!path->skip_locking) {
5741                         ret = btrfs_try_tree_read_lock(next);
5742                         if (!ret && time_seq) {
5743                                 /*
5744                                  * If we don't get the lock, we may be racing
5745                                  * with push_leaf_left, holding that lock while
5746                                  * itself waiting for the leaf we've currently
5747                                  * locked. To solve this situation, we give up
5748                                  * on our lock and cycle.
5749                                  */
5750                                 free_extent_buffer(next);
5751                                 btrfs_release_path(path);
5752                                 cond_resched();
5753                                 goto again;
5754                         }
5755                         if (!ret) {
5756                                 btrfs_set_path_blocking(path);
5757                                 btrfs_tree_read_lock(next);
5758                         }
5759                         next_rw_lock = BTRFS_READ_LOCK;
5760                 }
5761                 break;
5762         }
5763         path->slots[level] = slot;
5764         while (1) {
5765                 level--;
5766                 c = path->nodes[level];
5767                 if (path->locks[level])
5768                         btrfs_tree_unlock_rw(c, path->locks[level]);
5769
5770                 free_extent_buffer(c);
5771                 path->nodes[level] = next;
5772                 path->slots[level] = 0;
5773                 if (!path->skip_locking)
5774                         path->locks[level] = next_rw_lock;
5775                 if (!level)
5776                         break;
5777
5778                 ret = read_block_for_search(root, path, &next, level,
5779                                             0, &key);
5780                 if (ret == -EAGAIN)
5781                         goto again;
5782
5783                 if (ret < 0) {
5784                         btrfs_release_path(path);
5785                         goto done;
5786                 }
5787
5788                 if (!path->skip_locking) {
5789                         ret = btrfs_try_tree_read_lock(next);
5790                         if (!ret) {
5791                                 btrfs_set_path_blocking(path);
5792                                 btrfs_tree_read_lock(next);
5793                         }
5794                         next_rw_lock = BTRFS_READ_LOCK;
5795                 }
5796         }
5797         ret = 0;
5798 done:
5799         unlock_up(path, 0, 1, 0, NULL);
5800         path->leave_spinning = old_spinning;
5801         if (!old_spinning)
5802                 btrfs_set_path_blocking(path);
5803
5804         return ret;
5805 }
5806
5807 /*
5808  * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
5809  * searching until it gets past min_objectid or finds an item of 'type'
5810  *
5811  * returns 0 if something is found, 1 if nothing was found and < 0 on error
5812  */
5813 int btrfs_previous_item(struct btrfs_root *root,
5814                         struct btrfs_path *path, u64 min_objectid,
5815                         int type)
5816 {
5817         struct btrfs_key found_key;
5818         struct extent_buffer *leaf;
5819         u32 nritems;
5820         int ret;
5821
5822         while (1) {
5823                 if (path->slots[0] == 0) {
5824                         btrfs_set_path_blocking(path);
5825                         ret = btrfs_prev_leaf(root, path);
5826                         if (ret != 0)
5827                                 return ret;
5828                 } else {
5829                         path->slots[0]--;
5830                 }
5831                 leaf = path->nodes[0];
5832                 nritems = btrfs_header_nritems(leaf);
5833                 if (nritems == 0)
5834                         return 1;
5835                 if (path->slots[0] == nritems)
5836                         path->slots[0]--;
5837
5838                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5839                 if (found_key.objectid < min_objectid)
5840                         break;
5841                 if (found_key.type == type)
5842                         return 0;
5843                 if (found_key.objectid == min_objectid &&
5844                     found_key.type < type)
5845                         break;
5846         }
5847         return 1;
5848 }
5849
5850 /*
5851  * search in extent tree to find a previous Metadata/Data extent item with
5852  * min objecitd.
5853  *
5854  * returns 0 if something is found, 1 if nothing was found and < 0 on error
5855  */
5856 int btrfs_previous_extent_item(struct btrfs_root *root,
5857                         struct btrfs_path *path, u64 min_objectid)
5858 {
5859         struct btrfs_key found_key;
5860         struct extent_buffer *leaf;
5861         u32 nritems;
5862         int ret;
5863
5864         while (1) {
5865                 if (path->slots[0] == 0) {
5866                         btrfs_set_path_blocking(path);
5867                         ret = btrfs_prev_leaf(root, path);
5868                         if (ret != 0)
5869                                 return ret;
5870                 } else {
5871                         path->slots[0]--;
5872                 }
5873                 leaf = path->nodes[0];
5874                 nritems = btrfs_header_nritems(leaf);
5875                 if (nritems == 0)
5876                         return 1;
5877                 if (path->slots[0] == nritems)
5878                         path->slots[0]--;
5879
5880                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5881                 if (found_key.objectid < min_objectid)
5882                         break;
5883                 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
5884                     found_key.type == BTRFS_METADATA_ITEM_KEY)
5885                         return 0;
5886                 if (found_key.objectid == min_objectid &&
5887                     found_key.type < BTRFS_EXTENT_ITEM_KEY)
5888                         break;
5889         }
5890         return 1;
5891 }