fs/btrfs/ctree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007,2008 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/slab.h>
   8 #include <linux/rbtree.h>
   9 #include <linux/mm.h>
  10 #include "ctree.h"
  11 #include "disk-io.h"
  12 #include "transaction.h"
  13 #include "print-tree.h"
  14 #include "locking.h"
  15 #include "volumes.h"
  16 #include "qgroup.h"
  17
  18 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  19                       *root, struct btrfs_path *path, int level);
  20 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  21                       const struct btrfs_key *ins_key, struct btrfs_path *path,
  22                       int data_size, int extend);
  23 static int push_node_left(struct btrfs_trans_handle *trans,
  24                           struct extent_buffer *dst,
  25                           struct extent_buffer *src, int empty);
  26 static int balance_node_right(struct btrfs_trans_handle *trans,
  27                               struct extent_buffer *dst_buf,
  28                               struct extent_buffer *src_buf);
  29 static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
  30                     int level, int slot);
  31
  32 static const struct btrfs_csums {
  33         u16             size;
  34         const char      name[10];
  35         const char      driver[12];
  36 } btrfs_csums[] = {
  37         [BTRFS_CSUM_TYPE_CRC32] = { .size = 4, .name = "crc32c" },
  38         [BTRFS_CSUM_TYPE_XXHASH] = { .size = 8, .name = "xxhash64" },
  39         [BTRFS_CSUM_TYPE_SHA256] = { .size = 32, .name = "sha256" },
  40         [BTRFS_CSUM_TYPE_BLAKE2] = { .size = 32, .name = "blake2b",
  41                                      .driver = "blake2b-256" },
  42 };
  43
  44 int btrfs_super_csum_size(const struct btrfs_super_block *s)
  45 {
  46         u16 t = btrfs_super_csum_type(s);
  47         /*
  48          * csum type is validated at mount time
  49          */
  50         return btrfs_csums[t].size;
  51 }
  52
  53 const char *btrfs_super_csum_name(u16 csum_type)
  54 {
  55         /* csum type is validated at mount time */
  56         return btrfs_csums[csum_type].name;
  57 }
  58
  59 /*
  60  * Return driver name if defined, otherwise the name that's also a valid driver
  61  * name
  62  */
  63 const char *btrfs_super_csum_driver(u16 csum_type)
  64 {
  65         /* csum type is validated at mount time */
  66         return btrfs_csums[csum_type].driver[0] ?
  67                 btrfs_csums[csum_type].driver :
  68                 btrfs_csums[csum_type].name;
  69 }
  70
  71 size_t __const btrfs_get_num_csums(void)
  72 {
  73         return ARRAY_SIZE(btrfs_csums);
  74 }
  75
  76 struct btrfs_path *btrfs_alloc_path(void)
  77 {
  78         return kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
  79 }
  80
  81 /* this also releases the path */
  82 void btrfs_free_path(struct btrfs_path *p)
  83 {
  84         if (!p)
  85                 return;
  86         btrfs_release_path(p);
  87         kmem_cache_free(btrfs_path_cachep, p);
  88 }
  89
  90 /*
  91  * path release drops references on the extent buffers in the path
  92  * and it drops any locks held by this path
  93  *
  94  * It is safe to call this on paths that no locks or extent buffers held.
  95  */
  96 noinline void btrfs_release_path(struct btrfs_path *p)
  97 {
  98         int i;
  99
 100         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
 101                 p->slots[i] = 0;
 102                 if (!p->nodes[i])
 103                         continue;
 104                 if (p->locks[i]) {
 105                         btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
 106                         p->locks[i] = 0;
 107                 }
 108                 free_extent_buffer(p->nodes[i]);
 109                 p->nodes[i] = NULL;
 110         }
 111 }
 112
 113 /*
 114  * safely gets a reference on the root node of a tree.  A lock
 115  * is not taken, so a concurrent writer may put a different node
 116  * at the root of the tree.  See btrfs_lock_root_node for the
 117  * looping required.
 118  *
 119  * The extent buffer returned by this has a reference taken, so
 120  * it won't disappear.  It may stop being the root of the tree
 121  * at any time because there are no locks held.
 122  */
 123 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
 124 {
 125         struct extent_buffer *eb;
 126
 127         while (1) {
 128                 rcu_read_lock();
 129                 eb = rcu_dereference(root->node);
 130
 131                 /*
 132                  * RCU really hurts here, we could free up the root node because
 133                  * it was COWed but we may not get the new root node yet so do
 134                  * the inc_not_zero dance and if it doesn't work then
 135                  * synchronize_rcu and try again.
 136                  */
 137                 if (atomic_inc_not_zero(&eb->refs)) {
 138                         rcu_read_unlock();
 139                         break;
 140                 }
 141                 rcu_read_unlock();
 142                 synchronize_rcu();
 143         }
 144         return eb;
 145 }
 146
 147 /*
 148  * Cowonly root (not-shareable trees, everything not subvolume or reloc roots),
 149  * just get put onto a simple dirty list.  Transaction walks this list to make
 150  * sure they get properly updated on disk.
 151  */
 152 static void add_root_to_dirty_list(struct btrfs_root *root)
 153 {
 154         struct btrfs_fs_info *fs_info = root->fs_info;
 155
 156         if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
 157             !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
 158                 return;
 159
 160         spin_lock(&fs_info->trans_lock);
 161         if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
 162                 /* Want the extent tree to be the last on the list */
 163                 if (root->root_key.objectid == BTRFS_EXTENT_TREE_OBJECTID)
 164                         list_move_tail(&root->dirty_list,
 165                                        &fs_info->dirty_cowonly_roots);
 166                 else
 167                         list_move(&root->dirty_list,
 168                                   &fs_info->dirty_cowonly_roots);
 169         }
 170         spin_unlock(&fs_info->trans_lock);
 171 }
 172
 173 /*
 174  * used by snapshot creation to make a copy of a root for a tree with
 175  * a given objectid.  The buffer with the new root node is returned in
 176  * cow_ret, and this func returns zero on success or a negative error code.
 177  */
 178 int btrfs_copy_root(struct btrfs_trans_handle *trans,
 179                       struct btrfs_root *root,
 180                       struct extent_buffer *buf,
 181                       struct extent_buffer **cow_ret, u64 new_root_objectid)
 182 {
 183         struct btrfs_fs_info *fs_info = root->fs_info;
 184         struct extent_buffer *cow;
 185         int ret = 0;
 186         int level;
 187         struct btrfs_disk_key disk_key;
 188
 189         WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 190                 trans->transid != fs_info->running_transaction->transid);
 191         WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 192                 trans->transid != root->last_trans);
 193
 194         level = btrfs_header_level(buf);
 195         if (level == 0)
 196                 btrfs_item_key(buf, &disk_key, 0);
 197         else
 198                 btrfs_node_key(buf, &disk_key, 0);
 199
 200         cow = btrfs_alloc_tree_block(trans, root, 0, new_root_objectid,
 201                         &disk_key, level, buf->start, 0);
 202         if (IS_ERR(cow))
 203                 return PTR_ERR(cow);
 204
 205         copy_extent_buffer_full(cow, buf);
 206         btrfs_set_header_bytenr(cow, cow->start);
 207         btrfs_set_header_generation(cow, trans->transid);
 208         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 209         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 210                                      BTRFS_HEADER_FLAG_RELOC);
 211         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 212                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 213         else
 214                 btrfs_set_header_owner(cow, new_root_objectid);
 215
 216         write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
 217
 218         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 219         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 220                 ret = btrfs_inc_ref(trans, root, cow, 1);
 221         else
 222                 ret = btrfs_inc_ref(trans, root, cow, 0);
 223
 224         if (ret)
 225                 return ret;
 226
 227         btrfs_mark_buffer_dirty(cow);
 228         *cow_ret = cow;
 229         return 0;
 230 }
 231
 232 enum mod_log_op {
 233         MOD_LOG_KEY_REPLACE,
 234         MOD_LOG_KEY_ADD,
 235         MOD_LOG_KEY_REMOVE,
 236         MOD_LOG_KEY_REMOVE_WHILE_FREEING,
 237         MOD_LOG_KEY_REMOVE_WHILE_MOVING,
 238         MOD_LOG_MOVE_KEYS,
 239         MOD_LOG_ROOT_REPLACE,
 240 };
 241
 242 struct tree_mod_root {
 243         u64 logical;
 244         u8 level;
 245 };
 246
 247 struct tree_mod_elem {
 248         struct rb_node node;
 249         u64 logical;
 250         u64 seq;
 251         enum mod_log_op op;
 252
 253         /* this is used for MOD_LOG_KEY_* and MOD_LOG_MOVE_KEYS operations */
 254         int slot;
 255
 256         /* this is used for MOD_LOG_KEY* and MOD_LOG_ROOT_REPLACE */
 257         u64 generation;
 258
 259         /* those are used for op == MOD_LOG_KEY_{REPLACE,REMOVE} */
 260         struct btrfs_disk_key key;
 261         u64 blockptr;
 262
 263         /* this is used for op == MOD_LOG_MOVE_KEYS */
 264         struct {
 265                 int dst_slot;
 266                 int nr_items;
 267         } move;
 268
 269         /* this is used for op == MOD_LOG_ROOT_REPLACE */
 270         struct tree_mod_root old_root;
 271 };
 272
 273 /*
 274  * Pull a new tree mod seq number for our operation.
 275  */
 276 static inline u64 btrfs_inc_tree_mod_seq(struct btrfs_fs_info *fs_info)
 277 {
 278         return atomic64_inc_return(&fs_info->tree_mod_seq);
 279 }
 280
 281 /*
 282  * This adds a new blocker to the tree mod log's blocker list if the @elem
 283  * passed does not already have a sequence number set. So when a caller expects
 284  * to record tree modifications, it should ensure to set elem->seq to zero
 285  * before calling btrfs_get_tree_mod_seq.
 286  * Returns a fresh, unused tree log modification sequence number, even if no new
 287  * blocker was added.
 288  */
 289 u64 btrfs_get_tree_mod_seq(struct btrfs_fs_info *fs_info,
 290                            struct seq_list *elem)
 291 {
 292         write_lock(&fs_info->tree_mod_log_lock);
 293         if (!elem->seq) {
 294                 elem->seq = btrfs_inc_tree_mod_seq(fs_info);
 295                 list_add_tail(&elem->list, &fs_info->tree_mod_seq_list);
 296         }
 297         write_unlock(&fs_info->tree_mod_log_lock);
 298
 299         return elem->seq;
 300 }
 301
 302 void btrfs_put_tree_mod_seq(struct btrfs_fs_info *fs_info,
 303                             struct seq_list *elem)
 304 {
 305         struct rb_root *tm_root;
 306         struct rb_node *node;
 307         struct rb_node *next;
 308         struct tree_mod_elem *tm;
 309         u64 min_seq = (u64)-1;
 310         u64 seq_putting = elem->seq;
 311
 312         if (!seq_putting)
 313                 return;
 314
 315         write_lock(&fs_info->tree_mod_log_lock);
 316         list_del(&elem->list);
 317         elem->seq = 0;
 318
 319         if (!list_empty(&fs_info->tree_mod_seq_list)) {
 320                 struct seq_list *first;
 321
 322                 first = list_first_entry(&fs_info->tree_mod_seq_list,
 323                                          struct seq_list, list);
 324                 if (seq_putting > first->seq) {
 325                         /*
 326                          * Blocker with lower sequence number exists, we
 327                          * cannot remove anything from the log.
 328                          */
 329                         write_unlock(&fs_info->tree_mod_log_lock);
 330                         return;
 331                 }
 332                 min_seq = first->seq;
 333         }
 334
 335         /*
 336          * anything that's lower than the lowest existing (read: blocked)
 337          * sequence number can be removed from the tree.
 338          */
 339         tm_root = &fs_info->tree_mod_log;
 340         for (node = rb_first(tm_root); node; node = next) {
 341                 next = rb_next(node);
 342                 tm = rb_entry(node, struct tree_mod_elem, node);
 343                 if (tm->seq >= min_seq)
 344                         continue;
 345                 rb_erase(node, tm_root);
 346                 kfree(tm);
 347         }
 348         write_unlock(&fs_info->tree_mod_log_lock);
 349 }
 350
 351 /*
 352  * key order of the log:
 353  *       node/leaf start address -> sequence
 354  *
 355  * The 'start address' is the logical address of the *new* root node
 356  * for root replace operations, or the logical address of the affected
 357  * block for all other operations.
 358  */
 359 static noinline int
 360 __tree_mod_log_insert(struct btrfs_fs_info *fs_info, struct tree_mod_elem *tm)
 361 {
 362         struct rb_root *tm_root;
 363         struct rb_node **new;
 364         struct rb_node *parent = NULL;
 365         struct tree_mod_elem *cur;
 366
 367         lockdep_assert_held_write(&fs_info->tree_mod_log_lock);
 368
 369         tm->seq = btrfs_inc_tree_mod_seq(fs_info);
 370
 371         tm_root = &fs_info->tree_mod_log;
 372         new = &tm_root->rb_node;
 373         while (*new) {
 374                 cur = rb_entry(*new, struct tree_mod_elem, node);
 375                 parent = *new;
 376                 if (cur->logical < tm->logical)
 377                         new = &((*new)->rb_left);
 378                 else if (cur->logical > tm->logical)
 379                         new = &((*new)->rb_right);
 380                 else if (cur->seq < tm->seq)
 381                         new = &((*new)->rb_left);
 382                 else if (cur->seq > tm->seq)
 383                         new = &((*new)->rb_right);
 384                 else
 385                         return -EEXIST;
 386         }
 387
 388         rb_link_node(&tm->node, parent, new);
 389         rb_insert_color(&tm->node, tm_root);
 390         return 0;
 391 }
 392
 393 /*
 394  * Determines if logging can be omitted. Returns 1 if it can. Otherwise, it
 395  * returns zero with the tree_mod_log_lock acquired. The caller must hold
 396  * this until all tree mod log insertions are recorded in the rb tree and then
 397  * write unlock fs_info::tree_mod_log_lock.
 398  */
 399 static inline int tree_mod_dont_log(struct btrfs_fs_info *fs_info,
 400                                     struct extent_buffer *eb) {
 401         smp_mb();
 402         if (list_empty(&(fs_info)->tree_mod_seq_list))
 403                 return 1;
 404         if (eb && btrfs_header_level(eb) == 0)
 405                 return 1;
 406
 407         write_lock(&fs_info->tree_mod_log_lock);
 408         if (list_empty(&(fs_info)->tree_mod_seq_list)) {
 409                 write_unlock(&fs_info->tree_mod_log_lock);
 410                 return 1;
 411         }
 412
 413         return 0;
 414 }
 415
 416 /* Similar to tree_mod_dont_log, but doesn't acquire any locks. */
 417 static inline int tree_mod_need_log(const struct btrfs_fs_info *fs_info,
 418                                     struct extent_buffer *eb)
 419 {
 420         smp_mb();
 421         if (list_empty(&(fs_info)->tree_mod_seq_list))
 422                 return 0;
 423         if (eb && btrfs_header_level(eb) == 0)
 424                 return 0;
 425
 426         return 1;
 427 }
 428
 429 static struct tree_mod_elem *
 430 alloc_tree_mod_elem(struct extent_buffer *eb, int slot,
 431                     enum mod_log_op op, gfp_t flags)
 432 {
 433         struct tree_mod_elem *tm;
 434
 435         tm = kzalloc(sizeof(*tm), flags);
 436         if (!tm)
 437                 return NULL;
 438
 439         tm->logical = eb->start;
 440         if (op != MOD_LOG_KEY_ADD) {
 441                 btrfs_node_key(eb, &tm->key, slot);
 442                 tm->blockptr = btrfs_node_blockptr(eb, slot);
 443         }
 444         tm->op = op;
 445         tm->slot = slot;
 446         tm->generation = btrfs_node_ptr_generation(eb, slot);
 447         RB_CLEAR_NODE(&tm->node);
 448
 449         return tm;
 450 }
 451
 452 static noinline int tree_mod_log_insert_key(struct extent_buffer *eb, int slot,
 453                 enum mod_log_op op, gfp_t flags)
 454 {
 455         struct tree_mod_elem *tm;
 456         int ret;
 457
 458         if (!tree_mod_need_log(eb->fs_info, eb))
 459                 return 0;
 460
 461         tm = alloc_tree_mod_elem(eb, slot, op, flags);
 462         if (!tm)
 463                 return -ENOMEM;
 464
 465         if (tree_mod_dont_log(eb->fs_info, eb)) {
 466                 kfree(tm);
 467                 return 0;
 468         }
 469
 470         ret = __tree_mod_log_insert(eb->fs_info, tm);
 471         write_unlock(&eb->fs_info->tree_mod_log_lock);
 472         if (ret)
 473                 kfree(tm);
 474
 475         return ret;
 476 }
 477
 478 static noinline int tree_mod_log_insert_move(struct extent_buffer *eb,
 479                 int dst_slot, int src_slot, int nr_items)
 480 {
 481         struct tree_mod_elem *tm = NULL;
 482         struct tree_mod_elem **tm_list = NULL;
 483         int ret = 0;
 484         int i;
 485         int locked = 0;
 486
 487         if (!tree_mod_need_log(eb->fs_info, eb))
 488                 return 0;
 489
 490         tm_list = kcalloc(nr_items, sizeof(struct tree_mod_elem *), GFP_NOFS);
 491         if (!tm_list)
 492                 return -ENOMEM;
 493
 494         tm = kzalloc(sizeof(*tm), GFP_NOFS);
 495         if (!tm) {
 496                 ret = -ENOMEM;
 497                 goto free_tms;
 498         }
 499
 500         tm->logical = eb->start;
 501         tm->slot = src_slot;
 502         tm->move.dst_slot = dst_slot;
 503         tm->move.nr_items = nr_items;
 504         tm->op = MOD_LOG_MOVE_KEYS;
 505
 506         for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
 507                 tm_list[i] = alloc_tree_mod_elem(eb, i + dst_slot,
 508                     MOD_LOG_KEY_REMOVE_WHILE_MOVING, GFP_NOFS);
 509                 if (!tm_list[i]) {
 510                         ret = -ENOMEM;
 511                         goto free_tms;
 512                 }
 513         }
 514
 515         if (tree_mod_dont_log(eb->fs_info, eb))
 516                 goto free_tms;
 517         locked = 1;
 518
 519         /*
 520          * When we override something during the move, we log these removals.
 521          * This can only happen when we move towards the beginning of the
 522          * buffer, i.e. dst_slot < src_slot.
 523          */
 524         for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
 525                 ret = __tree_mod_log_insert(eb->fs_info, tm_list[i]);
 526                 if (ret)
 527                         goto free_tms;
 528         }
 529
 530         ret = __tree_mod_log_insert(eb->fs_info, tm);
 531         if (ret)
 532                 goto free_tms;
 533         write_unlock(&eb->fs_info->tree_mod_log_lock);
 534         kfree(tm_list);
 535
 536         return 0;
 537 free_tms:
 538         for (i = 0; i < nr_items; i++) {
 539                 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
 540                         rb_erase(&tm_list[i]->node, &eb->fs_info->tree_mod_log);
 541                 kfree(tm_list[i]);
 542         }
 543         if (locked)
 544                 write_unlock(&eb->fs_info->tree_mod_log_lock);
 545         kfree(tm_list);
 546         kfree(tm);
 547
 548         return ret;
 549 }
 550
 551 static inline int
 552 __tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
 553                        struct tree_mod_elem **tm_list,
 554                        int nritems)
 555 {
 556         int i, j;
 557         int ret;
 558
 559         for (i = nritems - 1; i >= 0; i--) {
 560                 ret = __tree_mod_log_insert(fs_info, tm_list[i]);
 561                 if (ret) {
 562                         for (j = nritems - 1; j > i; j--)
 563                                 rb_erase(&tm_list[j]->node,
 564                                          &fs_info->tree_mod_log);
 565                         return ret;
 566                 }
 567         }
 568
 569         return 0;
 570 }
 571
 572 static noinline int tree_mod_log_insert_root(struct extent_buffer *old_root,
 573                          struct extent_buffer *new_root, int log_removal)
 574 {
 575         struct btrfs_fs_info *fs_info = old_root->fs_info;
 576         struct tree_mod_elem *tm = NULL;
 577         struct tree_mod_elem **tm_list = NULL;
 578         int nritems = 0;
 579         int ret = 0;
 580         int i;
 581
 582         if (!tree_mod_need_log(fs_info, NULL))
 583                 return 0;
 584
 585         if (log_removal && btrfs_header_level(old_root) > 0) {
 586                 nritems = btrfs_header_nritems(old_root);
 587                 tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *),
 588                                   GFP_NOFS);
 589                 if (!tm_list) {
 590                         ret = -ENOMEM;
 591                         goto free_tms;
 592                 }
 593                 for (i = 0; i < nritems; i++) {
 594                         tm_list[i] = alloc_tree_mod_elem(old_root, i,
 595                             MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
 596                         if (!tm_list[i]) {
 597                                 ret = -ENOMEM;
 598                                 goto free_tms;
 599                         }
 600                 }
 601         }
 602
 603         tm = kzalloc(sizeof(*tm), GFP_NOFS);
 604         if (!tm) {
 605                 ret = -ENOMEM;
 606                 goto free_tms;
 607         }
 608
 609         tm->logical = new_root->start;
 610         tm->old_root.logical = old_root->start;
 611         tm->old_root.level = btrfs_header_level(old_root);
 612         tm->generation = btrfs_header_generation(old_root);
 613         tm->op = MOD_LOG_ROOT_REPLACE;
 614
 615         if (tree_mod_dont_log(fs_info, NULL))
 616                 goto free_tms;
 617
 618         if (tm_list)
 619                 ret = __tree_mod_log_free_eb(fs_info, tm_list, nritems);
 620         if (!ret)
 621                 ret = __tree_mod_log_insert(fs_info, tm);
 622
 623         write_unlock(&fs_info->tree_mod_log_lock);
 624         if (ret)
 625                 goto free_tms;
 626         kfree(tm_list);
 627
 628         return ret;
 629
 630 free_tms:
 631         if (tm_list) {
 632                 for (i = 0; i < nritems; i++)
 633                         kfree(tm_list[i]);
 634                 kfree(tm_list);
 635         }
 636         kfree(tm);
 637
 638         return ret;
 639 }
 640
 641 static struct tree_mod_elem *
 642 __tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq,
 643                       int smallest)
 644 {
 645         struct rb_root *tm_root;
 646         struct rb_node *node;
 647         struct tree_mod_elem *cur = NULL;
 648         struct tree_mod_elem *found = NULL;
 649
 650         read_lock(&fs_info->tree_mod_log_lock);
 651         tm_root = &fs_info->tree_mod_log;
 652         node = tm_root->rb_node;
 653         while (node) {
 654                 cur = rb_entry(node, struct tree_mod_elem, node);
 655                 if (cur->logical < start) {
 656                         node = node->rb_left;
 657                 } else if (cur->logical > start) {
 658                         node = node->rb_right;
 659                 } else if (cur->seq < min_seq) {
 660                         node = node->rb_left;
 661                 } else if (!smallest) {
 662                         /* we want the node with the highest seq */
 663                         if (found)
 664                                 BUG_ON(found->seq > cur->seq);
 665                         found = cur;
 666                         node = node->rb_left;
 667                 } else if (cur->seq > min_seq) {
 668                         /* we want the node with the smallest seq */
 669                         if (found)
 670                                 BUG_ON(found->seq < cur->seq);
 671                         found = cur;
 672                         node = node->rb_right;
 673                 } else {
 674                         found = cur;
 675                         break;
 676                 }
 677         }
 678         read_unlock(&fs_info->tree_mod_log_lock);
 679
 680         return found;
 681 }
 682
 683 /*
 684  * this returns the element from the log with the smallest time sequence
 685  * value that's in the log (the oldest log item). any element with a time
 686  * sequence lower than min_seq will be ignored.
 687  */
 688 static struct tree_mod_elem *
 689 tree_mod_log_search_oldest(struct btrfs_fs_info *fs_info, u64 start,
 690                            u64 min_seq)
 691 {
 692         return __tree_mod_log_search(fs_info, start, min_seq, 1);
 693 }
 694
 695 /*
 696  * this returns the element from the log with the largest time sequence
 697  * value that's in the log (the most recent log item). any element with
 698  * a time sequence lower than min_seq will be ignored.
 699  */
 700 static struct tree_mod_elem *
 701 tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq)
 702 {
 703         return __tree_mod_log_search(fs_info, start, min_seq, 0);
 704 }
 705
 706 static noinline int tree_mod_log_eb_copy(struct extent_buffer *dst,
 707                      struct extent_buffer *src, unsigned long dst_offset,
 708                      unsigned long src_offset, int nr_items)
 709 {
 710         struct btrfs_fs_info *fs_info = dst->fs_info;
 711         int ret = 0;
 712         struct tree_mod_elem **tm_list = NULL;
 713         struct tree_mod_elem **tm_list_add, **tm_list_rem;
 714         int i;
 715         int locked = 0;
 716
 717         if (!tree_mod_need_log(fs_info, NULL))
 718                 return 0;
 719
 720         if (btrfs_header_level(dst) == 0 && btrfs_header_level(src) == 0)
 721                 return 0;
 722
 723         tm_list = kcalloc(nr_items * 2, sizeof(struct tree_mod_elem *),
 724                           GFP_NOFS);
 725         if (!tm_list)
 726                 return -ENOMEM;
 727
 728         tm_list_add = tm_list;
 729         tm_list_rem = tm_list + nr_items;
 730         for (i = 0; i < nr_items; i++) {
 731                 tm_list_rem[i] = alloc_tree_mod_elem(src, i + src_offset,
 732                     MOD_LOG_KEY_REMOVE, GFP_NOFS);
 733                 if (!tm_list_rem[i]) {
 734                         ret = -ENOMEM;
 735                         goto free_tms;
 736                 }
 737
 738                 tm_list_add[i] = alloc_tree_mod_elem(dst, i + dst_offset,
 739                     MOD_LOG_KEY_ADD, GFP_NOFS);
 740                 if (!tm_list_add[i]) {
 741                         ret = -ENOMEM;
 742                         goto free_tms;
 743                 }
 744         }
 745
 746         if (tree_mod_dont_log(fs_info, NULL))
 747                 goto free_tms;
 748         locked = 1;
 749
 750         for (i = 0; i < nr_items; i++) {
 751                 ret = __tree_mod_log_insert(fs_info, tm_list_rem[i]);
 752                 if (ret)
 753                         goto free_tms;
 754                 ret = __tree_mod_log_insert(fs_info, tm_list_add[i]);
 755                 if (ret)
 756                         goto free_tms;
 757         }
 758
 759         write_unlock(&fs_info->tree_mod_log_lock);
 760         kfree(tm_list);
 761
 762         return 0;
 763
 764 free_tms:
 765         for (i = 0; i < nr_items * 2; i++) {
 766                 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
 767                         rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
 768                 kfree(tm_list[i]);
 769         }
 770         if (locked)
 771                 write_unlock(&fs_info->tree_mod_log_lock);
 772         kfree(tm_list);
 773
 774         return ret;
 775 }
 776
 777 static noinline int tree_mod_log_free_eb(struct extent_buffer *eb)
 778 {
 779         struct tree_mod_elem **tm_list = NULL;
 780         int nritems = 0;
 781         int i;
 782         int ret = 0;
 783
 784         if (btrfs_header_level(eb) == 0)
 785                 return 0;
 786
 787         if (!tree_mod_need_log(eb->fs_info, NULL))
 788                 return 0;
 789
 790         nritems = btrfs_header_nritems(eb);
 791         tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *), GFP_NOFS);
 792         if (!tm_list)
 793                 return -ENOMEM;
 794
 795         for (i = 0; i < nritems; i++) {
 796                 tm_list[i] = alloc_tree_mod_elem(eb, i,
 797                     MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
 798                 if (!tm_list[i]) {
 799                         ret = -ENOMEM;
 800                         goto free_tms;
 801                 }
 802         }
 803
 804         if (tree_mod_dont_log(eb->fs_info, eb))
 805                 goto free_tms;
 806
 807         ret = __tree_mod_log_free_eb(eb->fs_info, tm_list, nritems);
 808         write_unlock(&eb->fs_info->tree_mod_log_lock);
 809         if (ret)
 810                 goto free_tms;
 811         kfree(tm_list);
 812
 813         return 0;
 814
 815 free_tms:
 816         for (i = 0; i < nritems; i++)
 817                 kfree(tm_list[i]);
 818         kfree(tm_list);
 819
 820         return ret;
 821 }
 822
 823 /*
 824  * check if the tree block can be shared by multiple trees
 825  */
 826 int btrfs_block_can_be_shared(struct btrfs_root *root,
 827                               struct extent_buffer *buf)
 828 {
 829         /*
 830          * Tree blocks not in shareable trees and tree roots are never shared.
 831          * If a block was allocated after the last snapshot and the block was
 832          * not allocated by tree relocation, we know the block is not shared.
 833          */
 834         if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 835             buf != root->node && buf != root->commit_root &&
 836             (btrfs_header_generation(buf) <=
 837              btrfs_root_last_snapshot(&root->root_item) ||
 838              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 839                 return 1;
 840
 841         return 0;
 842 }
 843
 844 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
 845                                        struct btrfs_root *root,
 846                                        struct extent_buffer *buf,
 847                                        struct extent_buffer *cow,
 848                                        int *last_ref)
 849 {
 850         struct btrfs_fs_info *fs_info = root->fs_info;
 851         u64 refs;
 852         u64 owner;
 853         u64 flags;
 854         u64 new_flags = 0;
 855         int ret;
 856
 857         /*
 858          * Backrefs update rules:
 859          *
 860          * Always use full backrefs for extent pointers in tree block
 861          * allocated by tree relocation.
 862          *
 863          * If a shared tree block is no longer referenced by its owner
 864          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
 865          * use full backrefs for extent pointers in tree block.
 866          *
 867          * If a tree block is been relocating
 868          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
 869          * use full backrefs for extent pointers in tree block.
 870          * The reason for this is some operations (such as drop tree)
 871          * are only allowed for blocks use full backrefs.
 872          */
 873
 874         if (btrfs_block_can_be_shared(root, buf)) {
 875                 ret = btrfs_lookup_extent_info(trans, fs_info, buf->start,
 876                                                btrfs_header_level(buf), 1,
 877                                                &refs, &flags);
 878                 if (ret)
 879                         return ret;
 880                 if (refs == 0) {
 881                         ret = -EROFS;
 882                         btrfs_handle_fs_error(fs_info, ret, NULL);
 883                         return ret;
 884                 }
 885         } else {
 886                 refs = 1;
 887                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 888                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 889                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 890                 else
 891                         flags = 0;
 892         }
 893
 894         owner = btrfs_header_owner(buf);
 895         BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
 896                !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
 897
 898         if (refs > 1) {
 899                 if ((owner == root->root_key.objectid ||
 900                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
 901                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
 902                         ret = btrfs_inc_ref(trans, root, buf, 1);
 903                         if (ret)
 904                                 return ret;
 905
 906                         if (root->root_key.objectid ==
 907                             BTRFS_TREE_RELOC_OBJECTID) {
 908                                 ret = btrfs_dec_ref(trans, root, buf, 0);
 909                                 if (ret)
 910                                         return ret;
 911                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 912                                 if (ret)
 913                                         return ret;
 914                         }
 915                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
 916                 } else {
 917
 918                         if (root->root_key.objectid ==
 919                             BTRFS_TREE_RELOC_OBJECTID)
 920                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 921                         else
 922                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 923                         if (ret)
 924                                 return ret;
 925                 }
 926                 if (new_flags != 0) {
 927                         int level = btrfs_header_level(buf);
 928
 929                         ret = btrfs_set_disk_extent_flags(trans, buf,
 930                                                           new_flags, level, 0);
 931                         if (ret)
 932                                 return ret;
 933                 }
 934         } else {
 935                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
 936                         if (root->root_key.objectid ==
 937                             BTRFS_TREE_RELOC_OBJECTID)
 938                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 939                         else
 940                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 941                         if (ret)
 942                                 return ret;
 943                         ret = btrfs_dec_ref(trans, root, buf, 1);
 944                         if (ret)
 945                                 return ret;
 946                 }
 947                 btrfs_clean_tree_block(buf);
 948                 *last_ref = 1;
 949         }
 950         return 0;
 951 }
 952
 953 static struct extent_buffer *alloc_tree_block_no_bg_flush(
 954                                           struct btrfs_trans_handle *trans,
 955                                           struct btrfs_root *root,
 956                                           u64 parent_start,
 957                                           const struct btrfs_disk_key *disk_key,
 958                                           int level,
 959                                           u64 hint,
 960                                           u64 empty_size)
 961 {
 962         struct btrfs_fs_info *fs_info = root->fs_info;
 963         struct extent_buffer *ret;
 964
 965         /*
 966          * If we are COWing a node/leaf from the extent, chunk, device or free
 967          * space trees, make sure that we do not finish block group creation of
 968          * pending block groups. We do this to avoid a deadlock.
 969          * COWing can result in allocation of a new chunk, and flushing pending
 970          * block groups (btrfs_create_pending_block_groups()) can be triggered
 971          * when finishing allocation of a new chunk. Creation of a pending block
 972          * group modifies the extent, chunk, device and free space trees,
 973          * therefore we could deadlock with ourselves since we are holding a
 974          * lock on an extent buffer that btrfs_create_pending_block_groups() may
 975          * try to COW later.
 976          * For similar reasons, we also need to delay flushing pending block
 977          * groups when splitting a leaf or node, from one of those trees, since
 978          * we are holding a write lock on it and its parent or when inserting a
 979          * new root node for one of those trees.
 980          */
 981         if (root == fs_info->extent_root ||
 982             root == fs_info->chunk_root ||
 983             root == fs_info->dev_root ||
 984             root == fs_info->free_space_root)
 985                 trans->can_flush_pending_bgs = false;
 986
 987         ret = btrfs_alloc_tree_block(trans, root, parent_start,
 988                                      root->root_key.objectid, disk_key, level,
 989                                      hint, empty_size);
 990         trans->can_flush_pending_bgs = true;
 991
 992         return ret;
 993 }
 994
 995 /*
 996  * does the dirty work in cow of a single block.  The parent block (if
 997  * supplied) is updated to point to the new cow copy.  The new buffer is marked
 998  * dirty and returned locked.  If you modify the block it needs to be marked
 999  * dirty again.
1000  *
1001  * search_start -- an allocation hint for the new block
1002  *
1003  * empty_size -- a hint that you plan on doing more cow.  This is the size in
1004  * bytes the allocator should try to find free next to the block it returns.
1005  * This is just a hint and may be ignored by the allocator.
1006  */
1007 static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
1008                              struct btrfs_root *root,
1009                              struct extent_buffer *buf,
1010                              struct extent_buffer *parent, int parent_slot,
1011                              struct extent_buffer **cow_ret,
1012                              u64 search_start, u64 empty_size)
1013 {
1014         struct btrfs_fs_info *fs_info = root->fs_info;
1015         struct btrfs_disk_key disk_key;
1016         struct extent_buffer *cow;
1017         int level, ret;
1018         int last_ref = 0;
1019         int unlock_orig = 0;
1020         u64 parent_start = 0;
1021
1022         if (*cow_ret == buf)
1023                 unlock_orig = 1;
1024
1025         btrfs_assert_tree_locked(buf);
1026
1027         WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
1028                 trans->transid != fs_info->running_transaction->transid);
1029         WARN_ON(test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
1030                 trans->transid != root->last_trans);
1031
1032         level = btrfs_header_level(buf);
1033
1034         if (level == 0)
1035                 btrfs_item_key(buf, &disk_key, 0);
1036         else
1037                 btrfs_node_key(buf, &disk_key, 0);
1038
1039         if ((root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) && parent)
1040                 parent_start = parent->start;
1041
1042         cow = alloc_tree_block_no_bg_flush(trans, root, parent_start, &disk_key,
1043                                            level, search_start, empty_size);
1044         if (IS_ERR(cow))
1045                 return PTR_ERR(cow);
1046
1047         /* cow is set to blocking by btrfs_init_new_buffer */
1048
1049         copy_extent_buffer_full(cow, buf);
1050         btrfs_set_header_bytenr(cow, cow->start);
1051         btrfs_set_header_generation(cow, trans->transid);
1052         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
1053         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
1054                                      BTRFS_HEADER_FLAG_RELOC);
1055         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1056                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
1057         else
1058                 btrfs_set_header_owner(cow, root->root_key.objectid);
1059
1060         write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
1061
1062         ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
1063         if (ret) {
1064                 btrfs_abort_transaction(trans, ret);
1065                 return ret;
1066         }
1067
1068         if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
1069                 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
1070                 if (ret) {
1071                         btrfs_abort_transaction(trans, ret);
1072                         return ret;
1073                 }
1074         }
1075
1076         if (buf == root->node) {
1077                 WARN_ON(parent && parent != buf);
1078                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1079                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1080                         parent_start = buf->start;
1081
1082                 atomic_inc(&cow->refs);
1083                 ret = tree_mod_log_insert_root(root->node, cow, 1);
1084                 BUG_ON(ret < 0);
1085                 rcu_assign_pointer(root->node, cow);
1086
1087                 btrfs_free_tree_block(trans, root, buf, parent_start,
1088                                       last_ref);
1089                 free_extent_buffer(buf);
1090                 add_root_to_dirty_list(root);
1091         } else {
1092                 WARN_ON(trans->transid != btrfs_header_generation(parent));
1093                 tree_mod_log_insert_key(parent, parent_slot,
1094                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
1095                 btrfs_set_node_blockptr(parent, parent_slot,
1096                                         cow->start);
1097                 btrfs_set_node_ptr_generation(parent, parent_slot,
1098                                               trans->transid);
1099                 btrfs_mark_buffer_dirty(parent);
1100                 if (last_ref) {
1101                         ret = tree_mod_log_free_eb(buf);
1102                         if (ret) {
1103                                 btrfs_abort_transaction(trans, ret);
1104                                 return ret;
1105                         }
1106                 }
1107                 btrfs_free_tree_block(trans, root, buf, parent_start,
1108                                       last_ref);
1109         }
1110         if (unlock_orig)
1111                 btrfs_tree_unlock(buf);
1112         free_extent_buffer_stale(buf);
1113         btrfs_mark_buffer_dirty(cow);
1114         *cow_ret = cow;
1115         return 0;
1116 }
1117
1118 /*
1119  * returns the logical address of the oldest predecessor of the given root.
1120  * entries older than time_seq are ignored.
1121  */
1122 static struct tree_mod_elem *__tree_mod_log_oldest_root(
1123                 struct extent_buffer *eb_root, u64 time_seq)
1124 {
1125         struct tree_mod_elem *tm;
1126         struct tree_mod_elem *found = NULL;
1127         u64 root_logical = eb_root->start;
1128         int looped = 0;
1129
1130         if (!time_seq)
1131                 return NULL;
1132
1133         /*
1134          * the very last operation that's logged for a root is the
1135          * replacement operation (if it is replaced at all). this has
1136          * the logical address of the *new* root, making it the very
1137          * first operation that's logged for this root.
1138          */
1139         while (1) {
1140                 tm = tree_mod_log_search_oldest(eb_root->fs_info, root_logical,
1141                                                 time_seq);
1142                 if (!looped && !tm)
1143                         return NULL;
1144                 /*
1145                  * if there are no tree operation for the oldest root, we simply
1146                  * return it. this should only happen if that (old) root is at
1147                  * level 0.
1148                  */
1149                 if (!tm)
1150                         break;
1151
1152                 /*
1153                  * if there's an operation that's not a root replacement, we
1154                  * found the oldest version of our root. normally, we'll find a
1155                  * MOD_LOG_KEY_REMOVE_WHILE_FREEING operation here.
1156                  */
1157                 if (tm->op != MOD_LOG_ROOT_REPLACE)
1158                         break;
1159
1160                 found = tm;
1161                 root_logical = tm->old_root.logical;
1162                 looped = 1;
1163         }
1164
1165         /* if there's no old root to return, return what we found instead */
1166         if (!found)
1167                 found = tm;
1168
1169         return found;
1170 }
1171
1172 /*
1173  * tm is a pointer to the first operation to rewind within eb. then, all
1174  * previous operations will be rewound (until we reach something older than
1175  * time_seq).
1176  */
1177 static void
1178 __tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct extent_buffer *eb,
1179                       u64 time_seq, struct tree_mod_elem *first_tm)
1180 {
1181         u32 n;
1182         struct rb_node *next;
1183         struct tree_mod_elem *tm = first_tm;
1184         unsigned long o_dst;
1185         unsigned long o_src;
1186         unsigned long p_size = sizeof(struct btrfs_key_ptr);
1187
1188         n = btrfs_header_nritems(eb);
1189         read_lock(&fs_info->tree_mod_log_lock);
1190         while (tm && tm->seq >= time_seq) {
1191                 /*
1192                  * all the operations are recorded with the operator used for
1193                  * the modification. as we're going backwards, we do the
1194                  * opposite of each operation here.
1195                  */
1196                 switch (tm->op) {
1197                 case MOD_LOG_KEY_REMOVE_WHILE_FREEING:
1198                         BUG_ON(tm->slot < n);
1199                         fallthrough;
1200                 case MOD_LOG_KEY_REMOVE_WHILE_MOVING:
1201                 case MOD_LOG_KEY_REMOVE:
1202                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1203                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1204                         btrfs_set_node_ptr_generation(eb, tm->slot,
1205                                                       tm->generation);
1206                         n++;
1207                         break;
1208                 case MOD_LOG_KEY_REPLACE:
1209                         BUG_ON(tm->slot >= n);
1210                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1211                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1212                         btrfs_set_node_ptr_generation(eb, tm->slot,
1213                                                       tm->generation);
1214                         break;
1215                 case MOD_LOG_KEY_ADD:
1216                         /* if a move operation is needed it's in the log */
1217                         n--;
1218                         break;
1219                 case MOD_LOG_MOVE_KEYS:
1220                         o_dst = btrfs_node_key_ptr_offset(tm->slot);
1221                         o_src = btrfs_node_key_ptr_offset(tm->move.dst_slot);
1222                         memmove_extent_buffer(eb, o_dst, o_src,
1223                                               tm->move.nr_items * p_size);
1224                         break;
1225                 case MOD_LOG_ROOT_REPLACE:
1226                         /*
1227                          * this operation is special. for roots, this must be
1228                          * handled explicitly before rewinding.
1229                          * for non-roots, this operation may exist if the node
1230                          * was a root: root A -> child B; then A gets empty and
1231                          * B is promoted to the new root. in the mod log, we'll
1232                          * have a root-replace operation for B, a tree block
1233                          * that is no root. we simply ignore that operation.
1234                          */
1235                         break;
1236                 }
1237                 next = rb_next(&tm->node);
1238                 if (!next)
1239                         break;
1240                 tm = rb_entry(next, struct tree_mod_elem, node);
1241                 if (tm->logical != first_tm->logical)
1242                         break;
1243         }
1244         read_unlock(&fs_info->tree_mod_log_lock);
1245         btrfs_set_header_nritems(eb, n);
1246 }
1247
1248 /*
1249  * Called with eb read locked. If the buffer cannot be rewound, the same buffer
1250  * is returned. If rewind operations happen, a fresh buffer is returned. The
1251  * returned buffer is always read-locked. If the returned buffer is not the
1252  * input buffer, the lock on the input buffer is released and the input buffer
1253  * is freed (its refcount is decremented).
1254  */
1255 static struct extent_buffer *
1256 tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
1257                     struct extent_buffer *eb, u64 time_seq)
1258 {
1259         struct extent_buffer *eb_rewin;
1260         struct tree_mod_elem *tm;
1261
1262         if (!time_seq)
1263                 return eb;
1264
1265         if (btrfs_header_level(eb) == 0)
1266                 return eb;
1267
1268         tm = tree_mod_log_search(fs_info, eb->start, time_seq);
1269         if (!tm)
1270                 return eb;
1271
1272         btrfs_set_path_blocking(path);
1273         btrfs_set_lock_blocking_read(eb);
1274
1275         if (tm->op == MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1276                 BUG_ON(tm->slot != 0);
1277                 eb_rewin = alloc_dummy_extent_buffer(fs_info, eb->start);
1278                 if (!eb_rewin) {
1279                         btrfs_tree_read_unlock_blocking(eb);
1280                         free_extent_buffer(eb);
1281                         return NULL;
1282                 }
1283                 btrfs_set_header_bytenr(eb_rewin, eb->start);
1284                 btrfs_set_header_backref_rev(eb_rewin,
1285                                              btrfs_header_backref_rev(eb));
1286                 btrfs_set_header_owner(eb_rewin, btrfs_header_owner(eb));
1287                 btrfs_set_header_level(eb_rewin, btrfs_header_level(eb));
1288         } else {
1289                 eb_rewin = btrfs_clone_extent_buffer(eb);
1290                 if (!eb_rewin) {
1291                         btrfs_tree_read_unlock_blocking(eb);
1292                         free_extent_buffer(eb);
1293                         return NULL;
1294                 }
1295         }
1296
1297         btrfs_tree_read_unlock_blocking(eb);
1298         free_extent_buffer(eb);
1299
1300         btrfs_tree_read_lock(eb_rewin);
1301         __tree_mod_log_rewind(fs_info, eb_rewin, time_seq, tm);
1302         WARN_ON(btrfs_header_nritems(eb_rewin) >
1303                 BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1304
1305         return eb_rewin;
1306 }
1307
1308 /*
1309  * get_old_root() rewinds the state of @root's root node to the given @time_seq
1310  * value. If there are no changes, the current root->root_node is returned. If
1311  * anything changed in between, there's a fresh buffer allocated on which the
1312  * rewind operations are done. In any case, the returned buffer is read locked.
1313  * Returns NULL on error (with no locks held).
1314  */
1315 static inline struct extent_buffer *
1316 get_old_root(struct btrfs_root *root, u64 time_seq)
1317 {
1318         struct btrfs_fs_info *fs_info = root->fs_info;
1319         struct tree_mod_elem *tm;
1320         struct extent_buffer *eb = NULL;
1321         struct extent_buffer *eb_root;
1322         u64 eb_root_owner = 0;
1323         struct extent_buffer *old;
1324         struct tree_mod_root *old_root = NULL;
1325         u64 old_generation = 0;
1326         u64 logical;
1327         int level;
1328
1329         eb_root = btrfs_read_lock_root_node(root);
1330         tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1331         if (!tm)
1332                 return eb_root;
1333
1334         if (tm->op == MOD_LOG_ROOT_REPLACE) {
1335                 old_root = &tm->old_root;
1336                 old_generation = tm->generation;
1337                 logical = old_root->logical;
1338                 level = old_root->level;
1339         } else {
1340                 logical = eb_root->start;
1341                 level = btrfs_header_level(eb_root);
1342         }
1343
1344         tm = tree_mod_log_search(fs_info, logical, time_seq);
1345         if (old_root && tm && tm->op != MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1346                 btrfs_tree_read_unlock(eb_root);
1347                 free_extent_buffer(eb_root);
1348                 old = read_tree_block(fs_info, logical, 0, level, NULL);
1349                 if (WARN_ON(IS_ERR(old) || !extent_buffer_uptodate(old))) {
1350                         if (!IS_ERR(old))
1351                                 free_extent_buffer(old);
1352                         btrfs_warn(fs_info,
1353                                    "failed to read tree block %llu from get_old_root",
1354                                    logical);
1355                 } else {
1356                         eb = btrfs_clone_extent_buffer(old);
1357                         free_extent_buffer(old);
1358                 }
1359         } else if (old_root) {
1360                 eb_root_owner = btrfs_header_owner(eb_root);
1361                 btrfs_tree_read_unlock(eb_root);
1362                 free_extent_buffer(eb_root);
1363                 eb = alloc_dummy_extent_buffer(fs_info, logical);
1364         } else {
1365                 btrfs_set_lock_blocking_read(eb_root);
1366                 eb = btrfs_clone_extent_buffer(eb_root);
1367                 btrfs_tree_read_unlock_blocking(eb_root);
1368                 free_extent_buffer(eb_root);
1369         }
1370
1371         if (!eb)
1372                 return NULL;
1373         btrfs_tree_read_lock(eb);
1374         if (old_root) {
1375                 btrfs_set_header_bytenr(eb, eb->start);
1376                 btrfs_set_header_backref_rev(eb, BTRFS_MIXED_BACKREF_REV);
1377                 btrfs_set_header_owner(eb, eb_root_owner);
1378                 btrfs_set_header_level(eb, old_root->level);
1379                 btrfs_set_header_generation(eb, old_generation);
1380         }
1381         if (tm)
1382                 __tree_mod_log_rewind(fs_info, eb, time_seq, tm);
1383         else
1384                 WARN_ON(btrfs_header_level(eb) != 0);
1385         WARN_ON(btrfs_header_nritems(eb) > BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1386
1387         return eb;
1388 }
1389
1390 int btrfs_old_root_level(struct btrfs_root *root, u64 time_seq)
1391 {
1392         struct tree_mod_elem *tm;
1393         int level;
1394         struct extent_buffer *eb_root = btrfs_root_node(root);
1395
1396         tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1397         if (tm && tm->op == MOD_LOG_ROOT_REPLACE) {
1398                 level = tm->old_root.level;
1399         } else {
1400                 level = btrfs_header_level(eb_root);
1401         }
1402         free_extent_buffer(eb_root);
1403
1404         return level;
1405 }
1406
1407 static inline int should_cow_block(struct btrfs_trans_handle *trans,
1408                                    struct btrfs_root *root,
1409                                    struct extent_buffer *buf)
1410 {
1411         if (btrfs_is_testing(root->fs_info))
1412                 return 0;
1413
1414         /* Ensure we can see the FORCE_COW bit */
1415         smp_mb__before_atomic();
1416
1417         /*
1418          * We do not need to cow a block if
1419          * 1) this block is not created or changed in this transaction;
1420          * 2) this block does not belong to TREE_RELOC tree;
1421          * 3) the root is not forced COW.
1422          *
1423          * What is forced COW:
1424          *    when we create snapshot during committing the transaction,
1425          *    after we've finished copying src root, we must COW the shared
1426          *    block to ensure the metadata consistency.
1427          */
1428         if (btrfs_header_generation(buf) == trans->transid &&
1429             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
1430             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
1431               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
1432             !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
1433                 return 0;
1434         return 1;
1435 }
1436
1437 /*
1438  * cows a single block, see __btrfs_cow_block for the real work.
1439  * This version of it has extra checks so that a block isn't COWed more than
1440  * once per transaction, as long as it hasn't been written yet
1441  */
1442 noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
1443                     struct btrfs_root *root, struct extent_buffer *buf,
1444                     struct extent_buffer *parent, int parent_slot,
1445                     struct extent_buffer **cow_ret)
1446 {
1447         struct btrfs_fs_info *fs_info = root->fs_info;
1448         u64 search_start;
1449         int ret;
1450
1451         if (test_bit(BTRFS_ROOT_DELETING, &root->state))
1452                 btrfs_err(fs_info,
1453                         "COW'ing blocks on a fs root that's being dropped");
1454
1455         if (trans->transaction != fs_info->running_transaction)
1456                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1457                        trans->transid,
1458                        fs_info->running_transaction->transid);
1459
1460         if (trans->transid != fs_info->generation)
1461                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1462                        trans->transid, fs_info->generation);
1463
1464         if (!should_cow_block(trans, root, buf)) {
1465                 trans->dirty = true;
1466                 *cow_ret = buf;
1467                 return 0;
1468         }
1469
1470         search_start = buf->start & ~((u64)SZ_1G - 1);
1471
1472         if (parent)
1473                 btrfs_set_lock_blocking_write(parent);
1474         btrfs_set_lock_blocking_write(buf);
1475
1476         /*
1477          * Before CoWing this block for later modification, check if it's
1478          * the subtree root and do the delayed subtree trace if needed.
1479          *
1480          * Also We don't care about the error, as it's handled internally.
1481          */
1482         btrfs_qgroup_trace_subtree_after_cow(trans, root, buf);
1483         ret = __btrfs_cow_block(trans, root, buf, parent,
1484                                  parent_slot, cow_ret, search_start, 0);
1485
1486         trace_btrfs_cow_block(root, buf, *cow_ret);
1487
1488         return ret;
1489 }
1490
1491 /*
1492  * helper function for defrag to decide if two blocks pointed to by a
1493  * node are actually close by
1494  */
1495 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
1496 {
1497         if (blocknr < other && other - (blocknr + blocksize) < 32768)
1498                 return 1;
1499         if (blocknr > other && blocknr - (other + blocksize) < 32768)
1500                 return 1;
1501         return 0;
1502 }
1503
1504 #ifdef __LITTLE_ENDIAN
1505
1506 /*
1507  * Compare two keys, on little-endian the disk order is same as CPU order and
1508  * we can avoid the conversion.
1509  */
1510 static int comp_keys(const struct btrfs_disk_key *disk_key,
1511                      const struct btrfs_key *k2)
1512 {
1513         const struct btrfs_key *k1 = (const struct btrfs_key *)disk_key;
1514
1515         return btrfs_comp_cpu_keys(k1, k2);
1516 }
1517
1518 #else
1519
1520 /*
1521  * compare two keys in a memcmp fashion
1522  */
1523 static int comp_keys(const struct btrfs_disk_key *disk,
1524                      const struct btrfs_key *k2)
1525 {
1526         struct btrfs_key k1;
1527
1528         btrfs_disk_key_to_cpu(&k1, disk);
1529
1530         return btrfs_comp_cpu_keys(&k1, k2);
1531 }
1532 #endif
1533
1534 /*
1535  * same as comp_keys only with two btrfs_key's
1536  */
1537 int __pure btrfs_comp_cpu_keys(const struct btrfs_key *k1, const struct btrfs_key *k2)
1538 {
1539         if (k1->objectid > k2->objectid)
1540                 return 1;
1541         if (k1->objectid < k2->objectid)
1542                 return -1;
1543         if (k1->type > k2->type)
1544                 return 1;
1545         if (k1->type < k2->type)
1546                 return -1;
1547         if (k1->offset > k2->offset)
1548                 return 1;
1549         if (k1->offset < k2->offset)
1550                 return -1;
1551         return 0;
1552 }
1553
1554 /*
1555  * this is used by the defrag code to go through all the
1556  * leaves pointed to by a node and reallocate them so that
1557  * disk order is close to key order
1558  */
1559 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
1560                        struct btrfs_root *root, struct extent_buffer *parent,
1561                        int start_slot, u64 *last_ret,
1562                        struct btrfs_key *progress)
1563 {
1564         struct btrfs_fs_info *fs_info = root->fs_info;
1565         struct extent_buffer *cur;
1566         u64 blocknr;
1567         u64 gen;
1568         u64 search_start = *last_ret;
1569         u64 last_block = 0;
1570         u64 other;
1571         u32 parent_nritems;
1572         int end_slot;
1573         int i;
1574         int err = 0;
1575         int parent_level;
1576         int uptodate;
1577         u32 blocksize;
1578         int progress_passed = 0;
1579         struct btrfs_disk_key disk_key;
1580
1581         parent_level = btrfs_header_level(parent);
1582
1583         WARN_ON(trans->transaction != fs_info->running_transaction);
1584         WARN_ON(trans->transid != fs_info->generation);
1585
1586         parent_nritems = btrfs_header_nritems(parent);
1587         blocksize = fs_info->nodesize;
1588         end_slot = parent_nritems - 1;
1589
1590         if (parent_nritems <= 1)
1591                 return 0;
1592
1593         btrfs_set_lock_blocking_write(parent);
1594
1595         for (i = start_slot; i <= end_slot; i++) {
1596                 struct btrfs_key first_key;
1597                 int close = 1;
1598
1599                 btrfs_node_key(parent, &disk_key, i);
1600                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
1601                         continue;
1602
1603                 progress_passed = 1;
1604                 blocknr = btrfs_node_blockptr(parent, i);
1605                 gen = btrfs_node_ptr_generation(parent, i);
1606                 btrfs_node_key_to_cpu(parent, &first_key, i);
1607                 if (last_block == 0)
1608                         last_block = blocknr;
1609
1610                 if (i > 0) {
1611                         other = btrfs_node_blockptr(parent, i - 1);
1612                         close = close_blocks(blocknr, other, blocksize);
1613                 }
1614                 if (!close && i < end_slot) {
1615                         other = btrfs_node_blockptr(parent, i + 1);
1616                         close = close_blocks(blocknr, other, blocksize);
1617                 }
1618                 if (close) {
1619                         last_block = blocknr;
1620                         continue;
1621                 }
1622
1623                 cur = find_extent_buffer(fs_info, blocknr);
1624                 if (cur)
1625                         uptodate = btrfs_buffer_uptodate(cur, gen, 0);
1626                 else
1627                         uptodate = 0;
1628                 if (!cur || !uptodate) {
1629                         if (!cur) {
1630                                 cur = read_tree_block(fs_info, blocknr, gen,
1631                                                       parent_level - 1,
1632                                                       &first_key);
1633                                 if (IS_ERR(cur)) {
1634                                         return PTR_ERR(cur);
1635                                 } else if (!extent_buffer_uptodate(cur)) {
1636                                         free_extent_buffer(cur);
1637                                         return -EIO;
1638                                 }
1639                         } else if (!uptodate) {
1640                                 err = btrfs_read_buffer(cur, gen,
1641                                                 parent_level - 1,&first_key);
1642                                 if (err) {
1643                                         free_extent_buffer(cur);
1644                                         return err;
1645                                 }
1646                         }
1647                 }
1648                 if (search_start == 0)
1649                         search_start = last_block;
1650
1651                 btrfs_tree_lock(cur);
1652                 btrfs_set_lock_blocking_write(cur);
1653                 err = __btrfs_cow_block(trans, root, cur, parent, i,
1654                                         &cur, search_start,
1655                                         min(16 * blocksize,
1656                                             (end_slot - i) * blocksize));
1657                 if (err) {
1658                         btrfs_tree_unlock(cur);
1659                         free_extent_buffer(cur);
1660                         break;
1661                 }
1662                 search_start = cur->start;
1663                 last_block = cur->start;
1664                 *last_ret = search_start;
1665                 btrfs_tree_unlock(cur);
1666                 free_extent_buffer(cur);
1667         }
1668         return err;
1669 }
1670
1671 /*
1672  * search for key in the extent_buffer.  The items start at offset p,
1673  * and they are item_size apart.  There are 'max' items in p.
1674  *
1675  * the slot in the array is returned via slot, and it points to
1676  * the place where you would insert key if it is not found in
1677  * the array.
1678  *
1679  * slot may point to max if the key is bigger than all of the keys
1680  */
1681 static noinline int generic_bin_search(struct extent_buffer *eb,
1682                                        unsigned long p, int item_size,
1683                                        const struct btrfs_key *key,
1684                                        int max, int *slot)
1685 {
1686         int low = 0;
1687         int high = max;
1688         int ret;
1689         const int key_size = sizeof(struct btrfs_disk_key);
1690
1691         if (low > high) {
1692                 btrfs_err(eb->fs_info,
1693                  "%s: low (%d) > high (%d) eb %llu owner %llu level %d",
1694                           __func__, low, high, eb->start,
1695                           btrfs_header_owner(eb), btrfs_header_level(eb));
1696                 return -EINVAL;
1697         }
1698
1699         while (low < high) {
1700                 unsigned long oip;
1701                 unsigned long offset;
1702                 struct btrfs_disk_key *tmp;
1703                 struct btrfs_disk_key unaligned;
1704                 int mid;
1705
1706                 mid = (low + high) / 2;
1707                 offset = p + mid * item_size;
1708                 oip = offset_in_page(offset);
1709
1710                 if (oip + key_size <= PAGE_SIZE) {
1711                         const unsigned long idx = offset >> PAGE_SHIFT;
1712                         char *kaddr = page_address(eb->pages[idx]);
1713
1714                         tmp = (struct btrfs_disk_key *)(kaddr + oip);
1715                 } else {
1716                         read_extent_buffer(eb, &unaligned, offset, key_size);
1717                         tmp = &unaligned;
1718                 }
1719
1720                 ret = comp_keys(tmp, key);
1721
1722                 if (ret < 0)
1723                         low = mid + 1;
1724                 else if (ret > 0)
1725                         high = mid;
1726                 else {
1727                         *slot = mid;
1728                         return 0;
1729                 }
1730         }
1731         *slot = low;
1732         return 1;
1733 }
1734
1735 /*
1736  * simple bin_search frontend that does the right thing for
1737  * leaves vs nodes
1738  */
1739 int btrfs_bin_search(struct extent_buffer *eb, const struct btrfs_key *key,
1740                      int *slot)
1741 {
1742         if (btrfs_header_level(eb) == 0)
1743                 return generic_bin_search(eb,
1744                                           offsetof(struct btrfs_leaf, items),
1745                                           sizeof(struct btrfs_item),
1746                                           key, btrfs_header_nritems(eb),
1747                                           slot);
1748         else
1749                 return generic_bin_search(eb,
1750                                           offsetof(struct btrfs_node, ptrs),
1751                                           sizeof(struct btrfs_key_ptr),
1752                                           key, btrfs_header_nritems(eb),
1753                                           slot);
1754 }
1755
1756 static void root_add_used(struct btrfs_root *root, u32 size)
1757 {
1758         spin_lock(&root->accounting_lock);
1759         btrfs_set_root_used(&root->root_item,
1760                             btrfs_root_used(&root->root_item) + size);
1761         spin_unlock(&root->accounting_lock);
1762 }
1763
1764 static void root_sub_used(struct btrfs_root *root, u32 size)
1765 {
1766         spin_lock(&root->accounting_lock);
1767         btrfs_set_root_used(&root->root_item,
1768                             btrfs_root_used(&root->root_item) - size);
1769         spin_unlock(&root->accounting_lock);
1770 }
1771
1772 /* given a node and slot number, this reads the blocks it points to.  The
1773  * extent buffer is returned with a reference taken (but unlocked).
1774  */
1775 struct extent_buffer *btrfs_read_node_slot(struct extent_buffer *parent,
1776                                            int slot)
1777 {
1778         int level = btrfs_header_level(parent);
1779         struct extent_buffer *eb;
1780         struct btrfs_key first_key;
1781
1782         if (slot < 0 || slot >= btrfs_header_nritems(parent))
1783                 return ERR_PTR(-ENOENT);
1784
1785         BUG_ON(level == 0);
1786
1787         btrfs_node_key_to_cpu(parent, &first_key, slot);
1788         eb = read_tree_block(parent->fs_info, btrfs_node_blockptr(parent, slot),
1789                              btrfs_node_ptr_generation(parent, slot),
1790                              level - 1, &first_key);
1791         if (!IS_ERR(eb) && !extent_buffer_uptodate(eb)) {
1792                 free_extent_buffer(eb);
1793                 eb = ERR_PTR(-EIO);
1794         }
1795
1796         return eb;
1797 }
1798
1799 /*
1800  * node level balancing, used to make sure nodes are in proper order for
1801  * item deletion.  We balance from the top down, so we have to make sure
1802  * that a deletion won't leave an node completely empty later on.
1803  */
1804 static noinline int balance_level(struct btrfs_trans_handle *trans,
1805                          struct btrfs_root *root,
1806                          struct btrfs_path *path, int level)
1807 {
1808         struct btrfs_fs_info *fs_info = root->fs_info;
1809         struct extent_buffer *right = NULL;
1810         struct extent_buffer *mid;
1811         struct extent_buffer *left = NULL;
1812         struct extent_buffer *parent = NULL;
1813         int ret = 0;
1814         int wret;
1815         int pslot;
1816         int orig_slot = path->slots[level];
1817         u64 orig_ptr;
1818
1819         ASSERT(level > 0);
1820
1821         mid = path->nodes[level];
1822
1823         WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK &&
1824                 path->locks[level] != BTRFS_WRITE_LOCK_BLOCKING);
1825         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1826
1827         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1828
1829         if (level < BTRFS_MAX_LEVEL - 1) {
1830                 parent = path->nodes[level + 1];
1831                 pslot = path->slots[level + 1];
1832         }
1833
1834         /*
1835          * deal with the case where there is only one pointer in the root
1836          * by promoting the node below to a root
1837          */
1838         if (!parent) {
1839                 struct extent_buffer *child;
1840
1841                 if (btrfs_header_nritems(mid) != 1)
1842                         return 0;
1843
1844                 /* promote the child to a root */
1845                 child = btrfs_read_node_slot(mid, 0);
1846                 if (IS_ERR(child)) {
1847                         ret = PTR_ERR(child);
1848                         btrfs_handle_fs_error(fs_info, ret, NULL);
1849                         goto enospc;
1850                 }
1851
1852                 btrfs_tree_lock(child);
1853                 btrfs_set_lock_blocking_write(child);
1854                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
1855                 if (ret) {
1856                         btrfs_tree_unlock(child);
1857                         free_extent_buffer(child);
1858                         goto enospc;
1859                 }
1860
1861                 ret = tree_mod_log_insert_root(root->node, child, 1);
1862                 BUG_ON(ret < 0);
1863                 rcu_assign_pointer(root->node, child);
1864
1865                 add_root_to_dirty_list(root);
1866                 btrfs_tree_unlock(child);
1867
1868                 path->locks[level] = 0;
1869                 path->nodes[level] = NULL;
1870                 btrfs_clean_tree_block(mid);
1871                 btrfs_tree_unlock(mid);
1872                 /* once for the path */
1873                 free_extent_buffer(mid);
1874
1875                 root_sub_used(root, mid->len);
1876                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1877                 /* once for the root ptr */
1878                 free_extent_buffer_stale(mid);
1879                 return 0;
1880         }
1881         if (btrfs_header_nritems(mid) >
1882             BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 4)
1883                 return 0;
1884
1885         left = btrfs_read_node_slot(parent, pslot - 1);
1886         if (IS_ERR(left))
1887                 left = NULL;
1888
1889         if (left) {
1890                 btrfs_tree_lock(left);
1891                 btrfs_set_lock_blocking_write(left);
1892                 wret = btrfs_cow_block(trans, root, left,
1893                                        parent, pslot - 1, &left);
1894                 if (wret) {
1895                         ret = wret;
1896                         goto enospc;
1897                 }
1898         }
1899
1900         right = btrfs_read_node_slot(parent, pslot + 1);
1901         if (IS_ERR(right))
1902                 right = NULL;
1903
1904         if (right) {
1905                 btrfs_tree_lock(right);
1906                 btrfs_set_lock_blocking_write(right);
1907                 wret = btrfs_cow_block(trans, root, right,
1908                                        parent, pslot + 1, &right);
1909                 if (wret) {
1910                         ret = wret;
1911                         goto enospc;
1912                 }
1913         }
1914
1915         /* first, try to make some room in the middle buffer */
1916         if (left) {
1917                 orig_slot += btrfs_header_nritems(left);
1918                 wret = push_node_left(trans, left, mid, 1);
1919                 if (wret < 0)
1920                         ret = wret;
1921         }
1922
1923         /*
1924          * then try to empty the right most buffer into the middle
1925          */
1926         if (right) {
1927                 wret = push_node_left(trans, mid, right, 1);
1928                 if (wret < 0 && wret != -ENOSPC)
1929                         ret = wret;
1930                 if (btrfs_header_nritems(right) == 0) {
1931                         btrfs_clean_tree_block(right);
1932                         btrfs_tree_unlock(right);
1933                         del_ptr(root, path, level + 1, pslot + 1);
1934                         root_sub_used(root, right->len);
1935                         btrfs_free_tree_block(trans, root, right, 0, 1);
1936                         free_extent_buffer_stale(right);
1937                         right = NULL;
1938                 } else {
1939                         struct btrfs_disk_key right_key;
1940                         btrfs_node_key(right, &right_key, 0);
1941                         ret = tree_mod_log_insert_key(parent, pslot + 1,
1942                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
1943                         BUG_ON(ret < 0);
1944                         btrfs_set_node_key(parent, &right_key, pslot + 1);
1945                         btrfs_mark_buffer_dirty(parent);
1946                 }
1947         }
1948         if (btrfs_header_nritems(mid) == 1) {
1949                 /*
1950                  * we're not allowed to leave a node with one item in the
1951                  * tree during a delete.  A deletion from lower in the tree
1952                  * could try to delete the only pointer in this node.
1953                  * So, pull some keys from the left.
1954                  * There has to be a left pointer at this point because
1955                  * otherwise we would have pulled some pointers from the
1956                  * right
1957                  */
1958                 if (!left) {
1959                         ret = -EROFS;
1960                         btrfs_handle_fs_error(fs_info, ret, NULL);
1961                         goto enospc;
1962                 }
1963                 wret = balance_node_right(trans, mid, left);
1964                 if (wret < 0) {
1965                         ret = wret;
1966                         goto enospc;
1967                 }
1968                 if (wret == 1) {
1969                         wret = push_node_left(trans, left, mid, 1);
1970                         if (wret < 0)
1971                                 ret = wret;
1972                 }
1973                 BUG_ON(wret == 1);
1974         }
1975         if (btrfs_header_nritems(mid) == 0) {
1976                 btrfs_clean_tree_block(mid);
1977                 btrfs_tree_unlock(mid);
1978                 del_ptr(root, path, level + 1, pslot);
1979                 root_sub_used(root, mid->len);
1980                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1981                 free_extent_buffer_stale(mid);
1982                 mid = NULL;
1983         } else {
1984                 /* update the parent key to reflect our changes */
1985                 struct btrfs_disk_key mid_key;
1986                 btrfs_node_key(mid, &mid_key, 0);
1987                 ret = tree_mod_log_insert_key(parent, pslot,
1988                                 MOD_LOG_KEY_REPLACE, GFP_NOFS);
1989                 BUG_ON(ret < 0);
1990                 btrfs_set_node_key(parent, &mid_key, pslot);
1991                 btrfs_mark_buffer_dirty(parent);
1992         }
1993
1994         /* update the path */
1995         if (left) {
1996                 if (btrfs_header_nritems(left) > orig_slot) {
1997                         atomic_inc(&left->refs);
1998                         /* left was locked after cow */
1999                         path->nodes[level] = left;
2000                         path->slots[level + 1] -= 1;
2001                         path->slots[level] = orig_slot;
2002                         if (mid) {
2003                                 btrfs_tree_unlock(mid);
2004                                 free_extent_buffer(mid);
2005                         }
2006                 } else {
2007                         orig_slot -= btrfs_header_nritems(left);
2008                         path->slots[level] = orig_slot;
2009                 }
2010         }
2011         /* double check we haven't messed things up */
2012         if (orig_ptr !=
2013             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
2014                 BUG();
2015 enospc:
2016         if (right) {
2017                 btrfs_tree_unlock(right);
2018                 free_extent_buffer(right);
2019         }
2020         if (left) {
2021                 if (path->nodes[level] != left)
2022                         btrfs_tree_unlock(left);
2023                 free_extent_buffer(left);
2024         }
2025         return ret;
2026 }
2027
2028 /* Node balancing for insertion.  Here we only split or push nodes around
2029  * when they are completely full.  This is also done top down, so we
2030  * have to be pessimistic.
2031  */
2032 static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
2033                                           struct btrfs_root *root,
2034                                           struct btrfs_path *path, int level)
2035 {
2036         struct btrfs_fs_info *fs_info = root->fs_info;
2037         struct extent_buffer *right = NULL;
2038         struct extent_buffer *mid;
2039         struct extent_buffer *left = NULL;
2040         struct extent_buffer *parent = NULL;
2041         int ret = 0;
2042         int wret;
2043         int pslot;
2044         int orig_slot = path->slots[level];
2045
2046         if (level == 0)
2047                 return 1;
2048
2049         mid = path->nodes[level];
2050         WARN_ON(btrfs_header_generation(mid) != trans->transid);
2051
2052         if (level < BTRFS_MAX_LEVEL - 1) {
2053                 parent = path->nodes[level + 1];
2054                 pslot = path->slots[level + 1];
2055         }
2056
2057         if (!parent)
2058                 return 1;
2059
2060         left = btrfs_read_node_slot(parent, pslot - 1);
2061         if (IS_ERR(left))
2062                 left = NULL;
2063
2064         /* first, try to make some room in the middle buffer */
2065         if (left) {
2066                 u32 left_nr;
2067
2068                 btrfs_tree_lock(left);
2069                 btrfs_set_lock_blocking_write(left);
2070
2071                 left_nr = btrfs_header_nritems(left);
2072                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2073                         wret = 1;
2074                 } else {
2075                         ret = btrfs_cow_block(trans, root, left, parent,
2076                                               pslot - 1, &left);
2077                         if (ret)
2078                                 wret = 1;
2079                         else {
2080                                 wret = push_node_left(trans, left, mid, 0);
2081                         }
2082                 }
2083                 if (wret < 0)
2084                         ret = wret;
2085                 if (wret == 0) {
2086                         struct btrfs_disk_key disk_key;
2087                         orig_slot += left_nr;
2088                         btrfs_node_key(mid, &disk_key, 0);
2089                         ret = tree_mod_log_insert_key(parent, pslot,
2090                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
2091                         BUG_ON(ret < 0);
2092                         btrfs_set_node_key(parent, &disk_key, pslot);
2093                         btrfs_mark_buffer_dirty(parent);
2094                         if (btrfs_header_nritems(left) > orig_slot) {
2095                                 path->nodes[level] = left;
2096                                 path->slots[level + 1] -= 1;
2097                                 path->slots[level] = orig_slot;
2098                                 btrfs_tree_unlock(mid);
2099                                 free_extent_buffer(mid);
2100                         } else {
2101                                 orig_slot -=
2102                                         btrfs_header_nritems(left);
2103                                 path->slots[level] = orig_slot;
2104                                 btrfs_tree_unlock(left);
2105                                 free_extent_buffer(left);
2106                         }
2107                         return 0;
2108                 }
2109                 btrfs_tree_unlock(left);
2110                 free_extent_buffer(left);
2111         }
2112         right = btrfs_read_node_slot(parent, pslot + 1);
2113         if (IS_ERR(right))
2114                 right = NULL;
2115
2116         /*
2117          * then try to empty the right most buffer into the middle
2118          */
2119         if (right) {
2120                 u32 right_nr;
2121
2122                 btrfs_tree_lock(right);
2123                 btrfs_set_lock_blocking_write(right);
2124
2125                 right_nr = btrfs_header_nritems(right);
2126                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2127                         wret = 1;
2128                 } else {
2129                         ret = btrfs_cow_block(trans, root, right,
2130                                               parent, pslot + 1,
2131                                               &right);
2132                         if (ret)
2133                                 wret = 1;
2134                         else {
2135                                 wret = balance_node_right(trans, right, mid);
2136                         }
2137                 }
2138                 if (wret < 0)
2139                         ret = wret;
2140                 if (wret == 0) {
2141                         struct btrfs_disk_key disk_key;
2142
2143                         btrfs_node_key(right, &disk_key, 0);
2144                         ret = tree_mod_log_insert_key(parent, pslot + 1,
2145                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
2146                         BUG_ON(ret < 0);
2147                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
2148                         btrfs_mark_buffer_dirty(parent);
2149
2150                         if (btrfs_header_nritems(mid) <= orig_slot) {
2151                                 path->nodes[level] = right;
2152                                 path->slots[level + 1] += 1;
2153                                 path->slots[level] = orig_slot -
2154                                         btrfs_header_nritems(mid);
2155                                 btrfs_tree_unlock(mid);
2156                                 free_extent_buffer(mid);
2157                         } else {
2158                                 btrfs_tree_unlock(right);
2159                                 free_extent_buffer(right);
2160                         }
2161                         return 0;
2162                 }
2163                 btrfs_tree_unlock(right);
2164                 free_extent_buffer(right);
2165         }
2166         return 1;
2167 }
2168
2169 /*
2170  * readahead one full node of leaves, finding things that are close
2171  * to the block in 'slot', and triggering ra on them.
2172  */
2173 static void reada_for_search(struct btrfs_fs_info *fs_info,
2174                              struct btrfs_path *path,
2175                              int level, int slot, u64 objectid)
2176 {
2177         struct extent_buffer *node;
2178         struct btrfs_disk_key disk_key;
2179         u32 nritems;
2180         u64 search;
2181         u64 target;
2182         u64 nread = 0;
2183         struct extent_buffer *eb;
2184         u32 nr;
2185         u32 blocksize;
2186         u32 nscan = 0;
2187
2188         if (level != 1)
2189                 return;
2190
2191         if (!path->nodes[level])
2192                 return;
2193
2194         node = path->nodes[level];
2195
2196         search = btrfs_node_blockptr(node, slot);
2197         blocksize = fs_info->nodesize;
2198         eb = find_extent_buffer(fs_info, search);
2199         if (eb) {
2200                 free_extent_buffer(eb);
2201                 return;
2202         }
2203
2204         target = search;
2205
2206         nritems = btrfs_header_nritems(node);
2207         nr = slot;
2208
2209         while (1) {
2210                 if (path->reada == READA_BACK) {
2211                         if (nr == 0)
2212                                 break;
2213                         nr--;
2214                 } else if (path->reada == READA_FORWARD) {
2215                         nr++;
2216                         if (nr >= nritems)
2217                                 break;
2218                 }
2219                 if (path->reada == READA_BACK && objectid) {
2220                         btrfs_node_key(node, &disk_key, nr);
2221                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
2222                                 break;
2223                 }
2224                 search = btrfs_node_blockptr(node, nr);
2225                 if ((search <= target && target - search <= 65536) ||
2226                     (search > target && search - target <= 65536)) {
2227                         readahead_tree_block(fs_info, search);
2228                         nread += blocksize;
2229                 }
2230                 nscan++;
2231                 if ((nread > 65536 || nscan > 32))
2232                         break;
2233         }
2234 }
2235
2236 static noinline void reada_for_balance(struct btrfs_fs_info *fs_info,
2237                                        struct btrfs_path *path, int level)
2238 {
2239         int slot;
2240         int nritems;
2241         struct extent_buffer *parent;
2242         struct extent_buffer *eb;
2243         u64 gen;
2244         u64 block1 = 0;
2245         u64 block2 = 0;
2246
2247         parent = path->nodes[level + 1];
2248         if (!parent)
2249                 return;
2250
2251         nritems = btrfs_header_nritems(parent);
2252         slot = path->slots[level + 1];
2253
2254         if (slot > 0) {
2255                 block1 = btrfs_node_blockptr(parent, slot - 1);
2256                 gen = btrfs_node_ptr_generation(parent, slot - 1);
2257                 eb = find_extent_buffer(fs_info, block1);
2258                 /*
2259                  * if we get -eagain from btrfs_buffer_uptodate, we
2260                  * don't want to return eagain here.  That will loop
2261                  * forever
2262                  */
2263                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2264                         block1 = 0;
2265                 free_extent_buffer(eb);
2266         }
2267         if (slot + 1 < nritems) {
2268                 block2 = btrfs_node_blockptr(parent, slot + 1);
2269                 gen = btrfs_node_ptr_generation(parent, slot + 1);
2270                 eb = find_extent_buffer(fs_info, block2);
2271                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2272                         block2 = 0;
2273                 free_extent_buffer(eb);
2274         }
2275
2276         if (block1)
2277                 readahead_tree_block(fs_info, block1);
2278         if (block2)
2279                 readahead_tree_block(fs_info, block2);
2280 }
2281
2282
2283 /*
2284  * when we walk down the tree, it is usually safe to unlock the higher layers
2285  * in the tree.  The exceptions are when our path goes through slot 0, because
2286  * operations on the tree might require changing key pointers higher up in the
2287  * tree.
2288  *
2289  * callers might also have set path->keep_locks, which tells this code to keep
2290  * the lock if the path points to the last slot in the block.  This is part of
2291  * walking through the tree, and selecting the next slot in the higher block.
2292  *
2293  * lowest_unlock sets the lowest level in the tree we're allowed to unlock.  so
2294  * if lowest_unlock is 1, level 0 won't be unlocked
2295  */
2296 static noinline void unlock_up(struct btrfs_path *path, int level,
2297                                int lowest_unlock, int min_write_lock_level,
2298                                int *write_lock_level)
2299 {
2300         int i;
2301         int skip_level = level;
2302         int no_skips = 0;
2303         struct extent_buffer *t;
2304
2305         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2306                 if (!path->nodes[i])
2307                         break;
2308                 if (!path->locks[i])
2309                         break;
2310                 if (!no_skips && path->slots[i] == 0) {
2311                         skip_level = i + 1;
2312                         continue;
2313                 }
2314                 if (!no_skips && path->keep_locks) {
2315                         u32 nritems;
2316                         t = path->nodes[i];
2317                         nritems = btrfs_header_nritems(t);
2318                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
2319                                 skip_level = i + 1;
2320                                 continue;
2321                         }
2322                 }
2323                 if (skip_level < i && i >= lowest_unlock)
2324                         no_skips = 1;
2325
2326                 t = path->nodes[i];
2327                 if (i >= lowest_unlock && i > skip_level) {
2328                         btrfs_tree_unlock_rw(t, path->locks[i]);
2329                         path->locks[i] = 0;
2330                         if (write_lock_level &&
2331                             i > min_write_lock_level &&
2332                             i <= *write_lock_level) {
2333                                 *write_lock_level = i - 1;
2334                         }
2335                 }
2336         }
2337 }
2338
2339 /*
2340  * helper function for btrfs_search_slot.  The goal is to find a block
2341  * in cache without setting the path to blocking.  If we find the block
2342  * we return zero and the path is unchanged.
2343  *
2344  * If we can't find the block, we set the path blocking and do some
2345  * reada.  -EAGAIN is returned and the search must be repeated.
2346  */
2347 static int
2348 read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
2349                       struct extent_buffer **eb_ret, int level, int slot,
2350                       const struct btrfs_key *key)
2351 {
2352         struct btrfs_fs_info *fs_info = root->fs_info;
2353         u64 blocknr;
2354         u64 gen;
2355         struct extent_buffer *tmp;
2356         struct btrfs_key first_key;
2357         int ret;
2358         int parent_level;
2359
2360         blocknr = btrfs_node_blockptr(*eb_ret, slot);
2361         gen = btrfs_node_ptr_generation(*eb_ret, slot);
2362         parent_level = btrfs_header_level(*eb_ret);
2363         btrfs_node_key_to_cpu(*eb_ret, &first_key, slot);
2364
2365         tmp = find_extent_buffer(fs_info, blocknr);
2366         if (tmp) {
2367                 /* first we do an atomic uptodate check */
2368                 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
2369                         /*
2370                          * Do extra check for first_key, eb can be stale due to
2371                          * being cached, read from scrub, or have multiple
2372                          * parents (shared tree blocks).
2373                          */
2374                         if (btrfs_verify_level_key(tmp,
2375                                         parent_level - 1, &first_key, gen)) {
2376                                 free_extent_buffer(tmp);
2377                                 return -EUCLEAN;
2378                         }
2379                         *eb_ret = tmp;
2380                         return 0;
2381                 }
2382
2383                 /* the pages were up to date, but we failed
2384                  * the generation number check.  Do a full
2385                  * read for the generation number that is correct.
2386                  * We must do this without dropping locks so
2387                  * we can trust our generation number
2388                  */
2389                 btrfs_set_path_blocking(p);
2390
2391                 /* now we're allowed to do a blocking uptodate check */
2392                 ret = btrfs_read_buffer(tmp, gen, parent_level - 1, &first_key);
2393                 if (!ret) {
2394                         *eb_ret = tmp;
2395                         return 0;
2396                 }
2397                 free_extent_buffer(tmp);
2398                 btrfs_release_path(p);
2399                 return -EIO;
2400         }
2401
2402         /*
2403          * reduce lock contention at high levels
2404          * of the btree by dropping locks before
2405          * we read.  Don't release the lock on the current
2406          * level because we need to walk this node to figure
2407          * out which blocks to read.
2408          */
2409         btrfs_unlock_up_safe(p, level + 1);
2410         btrfs_set_path_blocking(p);
2411
2412         if (p->reada != READA_NONE)
2413                 reada_for_search(fs_info, p, level, slot, key->objectid);
2414
2415         ret = -EAGAIN;
2416         tmp = read_tree_block(fs_info, blocknr, gen, parent_level - 1,
2417                               &first_key);
2418         if (!IS_ERR(tmp)) {
2419                 /*
2420                  * If the read above didn't mark this buffer up to date,
2421                  * it will never end up being up to date.  Set ret to EIO now
2422                  * and give up so that our caller doesn't loop forever
2423                  * on our EAGAINs.
2424                  */
2425                 if (!extent_buffer_uptodate(tmp))
2426                         ret = -EIO;
2427                 free_extent_buffer(tmp);
2428         } else {
2429                 ret = PTR_ERR(tmp);
2430         }
2431
2432         btrfs_release_path(p);
2433         return ret;
2434 }
2435
2436 /*
2437  * helper function for btrfs_search_slot.  This does all of the checks
2438  * for node-level blocks and does any balancing required based on
2439  * the ins_len.
2440  *
2441  * If no extra work was required, zero is returned.  If we had to
2442  * drop the path, -EAGAIN is returned and btrfs_search_slot must
2443  * start over
2444  */
2445 static int
2446 setup_nodes_for_search(struct btrfs_trans_handle *trans,
2447                        struct btrfs_root *root, struct btrfs_path *p,
2448                        struct extent_buffer *b, int level, int ins_len,
2449                        int *write_lock_level)
2450 {
2451         struct btrfs_fs_info *fs_info = root->fs_info;
2452         int ret;
2453
2454         if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
2455             BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3) {
2456                 int sret;
2457
2458                 if (*write_lock_level < level + 1) {
2459                         *write_lock_level = level + 1;
2460                         btrfs_release_path(p);
2461                         goto again;
2462                 }
2463
2464                 btrfs_set_path_blocking(p);
2465                 reada_for_balance(fs_info, p, level);
2466                 sret = split_node(trans, root, p, level);
2467
2468                 BUG_ON(sret > 0);
2469                 if (sret) {
2470                         ret = sret;
2471                         goto done;
2472                 }
2473                 b = p->nodes[level];
2474         } else if (ins_len < 0 && btrfs_header_nritems(b) <
2475                    BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 2) {
2476                 int sret;
2477
2478                 if (*write_lock_level < level + 1) {
2479                         *write_lock_level = level + 1;
2480                         btrfs_release_path(p);
2481                         goto again;
2482                 }
2483
2484                 btrfs_set_path_blocking(p);
2485                 reada_for_balance(fs_info, p, level);
2486                 sret = balance_level(trans, root, p, level);
2487
2488                 if (sret) {
2489                         ret = sret;
2490                         goto done;
2491                 }
2492                 b = p->nodes[level];
2493                 if (!b) {
2494                         btrfs_release_path(p);
2495                         goto again;
2496                 }
2497                 BUG_ON(btrfs_header_nritems(b) == 1);
2498         }
2499         return 0;
2500
2501 again:
2502         ret = -EAGAIN;
2503 done:
2504         return ret;
2505 }
2506
2507 int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
2508                 u64 iobjectid, u64 ioff, u8 key_type,
2509                 struct btrfs_key *found_key)
2510 {
2511         int ret;
2512         struct btrfs_key key;
2513         struct extent_buffer *eb;
2514
2515         ASSERT(path);
2516         ASSERT(found_key);
2517
2518         key.type = key_type;
2519         key.objectid = iobjectid;
2520         key.offset = ioff;
2521
2522         ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
2523         if (ret < 0)
2524                 return ret;
2525
2526         eb = path->nodes[0];
2527         if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
2528                 ret = btrfs_next_leaf(fs_root, path);
2529                 if (ret)
2530                         return ret;
2531                 eb = path->nodes[0];
2532         }
2533
2534         btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
2535         if (found_key->type != key.type ||
2536                         found_key->objectid != key.objectid)
2537                 return 1;
2538
2539         return 0;
2540 }
2541
2542 static struct extent_buffer *btrfs_search_slot_get_root(struct btrfs_root *root,
2543                                                         struct btrfs_path *p,
2544                                                         int write_lock_level)
2545 {
2546         struct btrfs_fs_info *fs_info = root->fs_info;
2547         struct extent_buffer *b;
2548         int root_lock;
2549         int level = 0;
2550
2551         /* We try very hard to do read locks on the root */
2552         root_lock = BTRFS_READ_LOCK;
2553
2554         if (p->search_commit_root) {
2555                 /*
2556                  * The commit roots are read only so we always do read locks,
2557                  * and we always must hold the commit_root_sem when doing
2558                  * searches on them, the only exception is send where we don't
2559                  * want to block transaction commits for a long time, so
2560                  * we need to clone the commit root in order to avoid races
2561                  * with transaction commits that create a snapshot of one of
2562                  * the roots used by a send operation.
2563                  */
2564                 if (p->need_commit_sem) {
2565                         down_read(&fs_info->commit_root_sem);
2566                         b = btrfs_clone_extent_buffer(root->commit_root);
2567                         up_read(&fs_info->commit_root_sem);
2568                         if (!b)
2569                                 return ERR_PTR(-ENOMEM);
2570
2571                 } else {
2572                         b = root->commit_root;
2573                         atomic_inc(&b->refs);
2574                 }
2575                 level = btrfs_header_level(b);
2576                 /*
2577                  * Ensure that all callers have set skip_locking when
2578                  * p->search_commit_root = 1.
2579                  */
2580                 ASSERT(p->skip_locking == 1);
2581
2582                 goto out;
2583         }
2584
2585         if (p->skip_locking) {
2586                 b = btrfs_root_node(root);
2587                 level = btrfs_header_level(b);
2588                 goto out;
2589         }
2590
2591         /*
2592          * If the level is set to maximum, we can skip trying to get the read
2593          * lock.
2594          */
2595         if (write_lock_level < BTRFS_MAX_LEVEL) {
2596                 /*
2597                  * We don't know the level of the root node until we actually
2598                  * have it read locked
2599                  */
2600                 b = btrfs_read_lock_root_node(root);
2601                 level = btrfs_header_level(b);
2602                 if (level > write_lock_level)
2603                         goto out;
2604
2605                 /* Whoops, must trade for write lock */
2606                 btrfs_tree_read_unlock(b);
2607                 free_extent_buffer(b);
2608         }
2609
2610         b = btrfs_lock_root_node(root);
2611         root_lock = BTRFS_WRITE_LOCK;
2612
2613         /* The level might have changed, check again */
2614         level = btrfs_header_level(b);
2615
2616 out:
2617         p->nodes[level] = b;
2618         if (!p->skip_locking)
2619                 p->locks[level] = root_lock;
2620         /*
2621          * Callers are responsible for dropping b's references.
2622          */
2623         return b;
2624 }
2625
2626
2627 /*
2628  * btrfs_search_slot - look for a key in a tree and perform necessary
2629  * modifications to preserve tree invariants.
2630  *
2631  * @trans:      Handle of transaction, used when modifying the tree
2632  * @p:          Holds all btree nodes along the search path
2633  * @root:       The root node of the tree
2634  * @key:        The key we are looking for
2635  * @ins_len:    Indicates purpose of search, for inserts it is 1, for
2636  *              deletions it's -1. 0 for plain searches
2637  * @cow:        boolean should CoW operations be performed. Must always be 1
2638  *              when modifying the tree.
2639  *
2640  * If @ins_len > 0, nodes and leaves will be split as we walk down the tree.
2641  * If @ins_len < 0, nodes will be merged as we walk down the tree (if possible)
2642  *
2643  * If @key is found, 0 is returned and you can find the item in the leaf level
2644  * of the path (level 0)
2645  *
2646  * If @key isn't found, 1 is returned and the leaf level of the path (level 0)
2647  * points to the slot where it should be inserted
2648  *
2649  * If an error is encountered while searching the tree a negative error number
2650  * is returned
2651  */
2652 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2653                       const struct btrfs_key *key, struct btrfs_path *p,
2654                       int ins_len, int cow)
2655 {
2656         struct extent_buffer *b;
2657         int slot;
2658         int ret;
2659         int err;
2660         int level;
2661         int lowest_unlock = 1;
2662         /* everything at write_lock_level or lower must be write locked */
2663         int write_lock_level = 0;
2664         u8 lowest_level = 0;
2665         int min_write_lock_level;
2666         int prev_cmp;
2667
2668         lowest_level = p->lowest_level;
2669         WARN_ON(lowest_level && ins_len > 0);
2670         WARN_ON(p->nodes[0] != NULL);
2671         BUG_ON(!cow && ins_len);
2672
2673         if (ins_len < 0) {
2674                 lowest_unlock = 2;
2675
2676                 /* when we are removing items, we might have to go up to level
2677                  * two as we update tree pointers  Make sure we keep write
2678                  * for those levels as well
2679                  */
2680                 write_lock_level = 2;
2681         } else if (ins_len > 0) {
2682                 /*
2683                  * for inserting items, make sure we have a write lock on
2684                  * level 1 so we can update keys
2685                  */
2686                 write_lock_level = 1;
2687         }
2688
2689         if (!cow)
2690                 write_lock_level = -1;
2691
2692         if (cow && (p->keep_locks || p->lowest_level))
2693                 write_lock_level = BTRFS_MAX_LEVEL;
2694
2695         min_write_lock_level = write_lock_level;
2696
2697 again:
2698         prev_cmp = -1;
2699         b = btrfs_search_slot_get_root(root, p, write_lock_level);
2700         if (IS_ERR(b)) {
2701                 ret = PTR_ERR(b);
2702                 goto done;
2703         }
2704
2705         while (b) {
2706                 int dec = 0;
2707
2708                 level = btrfs_header_level(b);
2709
2710                 if (cow) {
2711                         bool last_level = (level == (BTRFS_MAX_LEVEL - 1));
2712
2713                         /*
2714                          * if we don't really need to cow this block
2715                          * then we don't want to set the path blocking,
2716                          * so we test it here
2717                          */
2718                         if (!should_cow_block(trans, root, b)) {
2719                                 trans->dirty = true;
2720                                 goto cow_done;
2721                         }
2722
2723                         /*
2724                          * must have write locks on this node and the
2725                          * parent
2726                          */
2727                         if (level > write_lock_level ||
2728                             (level + 1 > write_lock_level &&
2729                             level + 1 < BTRFS_MAX_LEVEL &&
2730                             p->nodes[level + 1])) {
2731                                 write_lock_level = level + 1;
2732                                 btrfs_release_path(p);
2733                                 goto again;
2734                         }
2735
2736                         btrfs_set_path_blocking(p);
2737                         if (last_level)
2738                                 err = btrfs_cow_block(trans, root, b, NULL, 0,
2739                                                       &b);
2740                         else
2741                                 err = btrfs_cow_block(trans, root, b,
2742                                                       p->nodes[level + 1],
2743                                                       p->slots[level + 1], &b);
2744                         if (err) {
2745                                 ret = err;
2746                                 goto done;
2747                         }
2748                 }
2749 cow_done:
2750                 p->nodes[level] = b;
2751                 /*
2752                  * Leave path with blocking locks to avoid massive
2753                  * lock context switch, this is made on purpose.
2754                  */
2755
2756                 /*
2757                  * we have a lock on b and as long as we aren't changing
2758                  * the tree, there is no way to for the items in b to change.
2759                  * It is safe to drop the lock on our parent before we
2760                  * go through the expensive btree search on b.
2761                  *
2762                  * If we're inserting or deleting (ins_len != 0), then we might
2763                  * be changing slot zero, which may require changing the parent.
2764                  * So, we can't drop the lock until after we know which slot
2765                  * we're operating on.
2766                  */
2767                 if (!ins_len && !p->keep_locks) {
2768                         int u = level + 1;
2769
2770                         if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
2771                                 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
2772                                 p->locks[u] = 0;
2773                         }
2774                 }
2775
2776                 /*
2777                  * If btrfs_bin_search returns an exact match (prev_cmp == 0)
2778                  * we can safely assume the target key will always be in slot 0
2779                  * on lower levels due to the invariants BTRFS' btree provides,
2780                  * namely that a btrfs_key_ptr entry always points to the
2781                  * lowest key in the child node, thus we can skip searching
2782                  * lower levels
2783                  */
2784                 if (prev_cmp == 0) {
2785                         slot = 0;
2786                         ret = 0;
2787                 } else {
2788                         ret = btrfs_bin_search(b, key, &slot);
2789                         prev_cmp = ret;
2790                         if (ret < 0)
2791                                 goto done;
2792                 }
2793
2794                 if (level == 0) {
2795                         p->slots[level] = slot;
2796                         if (ins_len > 0 &&
2797                             btrfs_leaf_free_space(b) < ins_len) {
2798                                 if (write_lock_level < 1) {
2799                                         write_lock_level = 1;
2800                                         btrfs_release_path(p);
2801                                         goto again;
2802                                 }
2803
2804                                 btrfs_set_path_blocking(p);
2805                                 err = split_leaf(trans, root, key,
2806                                                  p, ins_len, ret == 0);
2807
2808                                 BUG_ON(err > 0);
2809                                 if (err) {
2810                                         ret = err;
2811                                         goto done;
2812                                 }
2813                         }
2814                         if (!p->search_for_split)
2815                                 unlock_up(p, level, lowest_unlock,
2816                                           min_write_lock_level, NULL);
2817                         goto done;
2818                 }
2819                 if (ret && slot > 0) {
2820                         dec = 1;
2821                         slot--;
2822                 }
2823                 p->slots[level] = slot;
2824                 err = setup_nodes_for_search(trans, root, p, b, level, ins_len,
2825                                              &write_lock_level);
2826                 if (err == -EAGAIN)
2827                         goto again;
2828                 if (err) {
2829                         ret = err;
2830                         goto done;
2831                 }
2832                 b = p->nodes[level];
2833                 slot = p->slots[level];
2834
2835                 /*
2836                  * Slot 0 is special, if we change the key we have to update
2837                  * the parent pointer which means we must have a write lock on
2838                  * the parent
2839                  */
2840                 if (slot == 0 && ins_len && write_lock_level < level + 1) {
2841                         write_lock_level = level + 1;
2842                         btrfs_release_path(p);
2843                         goto again;
2844                 }
2845
2846                 unlock_up(p, level, lowest_unlock, min_write_lock_level,
2847                           &write_lock_level);
2848
2849                 if (level == lowest_level) {
2850                         if (dec)
2851                                 p->slots[level]++;
2852                         goto done;
2853                 }
2854
2855                 err = read_block_for_search(root, p, &b, level, slot, key);
2856                 if (err == -EAGAIN)
2857                         goto again;
2858                 if (err) {
2859                         ret = err;
2860                         goto done;
2861                 }
2862
2863                 if (!p->skip_locking) {
2864                         level = btrfs_header_level(b);
2865                         if (level <= write_lock_level) {
2866                                 if (!btrfs_try_tree_write_lock(b)) {
2867                                         btrfs_set_path_blocking(p);
2868                                         btrfs_tree_lock(b);
2869                                 }
2870                                 p->locks[level] = BTRFS_WRITE_LOCK;
2871                         } else {
2872                                 if (!btrfs_tree_read_lock_atomic(b)) {
2873                                         btrfs_set_path_blocking(p);
2874                                         btrfs_tree_read_lock(b);
2875                                 }
2876                                 p->locks[level] = BTRFS_READ_LOCK;
2877                         }
2878                         p->nodes[level] = b;
2879                 }
2880         }
2881         ret = 1;
2882 done:
2883         /*
2884          * we don't really know what they plan on doing with the path
2885          * from here on, so for now just mark it as blocking
2886          */
2887         if (!p->leave_spinning)
2888                 btrfs_set_path_blocking(p);
2889         if (ret < 0 && !p->skip_release_on_error)
2890                 btrfs_release_path(p);
2891         return ret;
2892 }
2893
2894 /*
2895  * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2896  * current state of the tree together with the operations recorded in the tree
2897  * modification log to search for the key in a previous version of this tree, as
2898  * denoted by the time_seq parameter.
2899  *
2900  * Naturally, there is no support for insert, delete or cow operations.
2901  *
2902  * The resulting path and return value will be set up as if we called
2903  * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2904  */
2905 int btrfs_search_old_slot(struct btrfs_root *root, const struct btrfs_key *key,
2906                           struct btrfs_path *p, u64 time_seq)
2907 {
2908         struct btrfs_fs_info *fs_info = root->fs_info;
2909         struct extent_buffer *b;
2910         int slot;
2911         int ret;
2912         int err;
2913         int level;
2914         int lowest_unlock = 1;
2915         u8 lowest_level = 0;
2916
2917         lowest_level = p->lowest_level;
2918         WARN_ON(p->nodes[0] != NULL);
2919
2920         if (p->search_commit_root) {
2921                 BUG_ON(time_seq);
2922                 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2923         }
2924
2925 again:
2926         b = get_old_root(root, time_seq);
2927         if (!b) {
2928                 ret = -EIO;
2929                 goto done;
2930         }
2931         level = btrfs_header_level(b);
2932         p->locks[level] = BTRFS_READ_LOCK;
2933
2934         while (b) {
2935                 int dec = 0;
2936
2937                 level = btrfs_header_level(b);
2938                 p->nodes[level] = b;
2939
2940                 /*
2941                  * we have a lock on b and as long as we aren't changing
2942                  * the tree, there is no way to for the items in b to change.
2943                  * It is safe to drop the lock on our parent before we
2944                  * go through the expensive btree search on b.
2945                  */
2946                 btrfs_unlock_up_safe(p, level + 1);
2947
2948                 ret = btrfs_bin_search(b, key, &slot);
2949                 if (ret < 0)
2950                         goto done;
2951
2952                 if (level == 0) {
2953                         p->slots[level] = slot;
2954                         unlock_up(p, level, lowest_unlock, 0, NULL);
2955                         goto done;
2956                 }
2957
2958                 if (ret && slot > 0) {
2959                         dec = 1;
2960                         slot--;
2961                 }
2962                 p->slots[level] = slot;
2963                 unlock_up(p, level, lowest_unlock, 0, NULL);
2964
2965                 if (level == lowest_level) {
2966                         if (dec)
2967                                 p->slots[level]++;
2968                         goto done;
2969                 }
2970
2971                 err = read_block_for_search(root, p, &b, level, slot, key);
2972                 if (err == -EAGAIN)
2973                         goto again;
2974                 if (err) {
2975                         ret = err;
2976                         goto done;
2977                 }
2978
2979                 level = btrfs_header_level(b);
2980                 if (!btrfs_tree_read_lock_atomic(b)) {
2981                         btrfs_set_path_blocking(p);
2982                         btrfs_tree_read_lock(b);
2983                 }
2984                 b = tree_mod_log_rewind(fs_info, p, b, time_seq);
2985                 if (!b) {
2986                         ret = -ENOMEM;
2987                         goto done;
2988                 }
2989                 p->locks[level] = BTRFS_READ_LOCK;
2990                 p->nodes[level] = b;
2991         }
2992         ret = 1;
2993 done:
2994         if (!p->leave_spinning)
2995                 btrfs_set_path_blocking(p);
2996         if (ret < 0)
2997                 btrfs_release_path(p);
2998
2999         return ret;
3000 }
3001
3002 /*
3003  * helper to use instead of search slot if no exact match is needed but
3004  * instead the next or previous item should be returned.
3005  * When find_higher is true, the next higher item is returned, the next lower
3006  * otherwise.
3007  * When return_any and find_higher are both true, and no higher item is found,
3008  * return the next lower instead.
3009  * When return_any is true and find_higher is false, and no lower item is found,
3010  * return the next higher instead.
3011  * It returns 0 if any item is found, 1 if none is found (tree empty), and
3012  * < 0 on error
3013  */
3014 int btrfs_search_slot_for_read(struct btrfs_root *root,
3015                                const struct btrfs_key *key,
3016                                struct btrfs_path *p, int find_higher,
3017                                int return_any)
3018 {
3019         int ret;
3020         struct extent_buffer *leaf;
3021
3022 again:
3023         ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
3024         if (ret <= 0)
3025                 return ret;
3026         /*
3027          * a return value of 1 means the path is at the position where the
3028          * item should be inserted. Normally this is the next bigger item,
3029          * but in case the previous item is the last in a leaf, path points
3030          * to the first free slot in the previous leaf, i.e. at an invalid
3031          * item.
3032          */
3033         leaf = p->nodes[0];
3034
3035         if (find_higher) {
3036                 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
3037                         ret = btrfs_next_leaf(root, p);
3038                         if (ret <= 0)
3039                                 return ret;
3040                         if (!return_any)
3041                                 return 1;
3042                         /*
3043                          * no higher item found, return the next
3044                          * lower instead
3045                          */
3046                         return_any = 0;
3047                         find_higher = 0;
3048                         btrfs_release_path(p);
3049                         goto again;
3050                 }
3051         } else {
3052                 if (p->slots[0] == 0) {
3053                         ret = btrfs_prev_leaf(root, p);
3054                         if (ret < 0)
3055                                 return ret;
3056                         if (!ret) {
3057                                 leaf = p->nodes[0];
3058                                 if (p->slots[0] == btrfs_header_nritems(leaf))
3059                                         p->slots[0]--;
3060                                 return 0;
3061                         }
3062                         if (!return_any)
3063                                 return 1;
3064                         /*
3065                          * no lower item found, return the next
3066                          * higher instead
3067                          */
3068                         return_any = 0;
3069                         find_higher = 1;
3070                         btrfs_release_path(p);
3071                         goto again;
3072                 } else {
3073                         --p->slots[0];
3074                 }
3075         }
3076         return 0;
3077 }
3078
3079 /*
3080  * adjust the pointers going up the tree, starting at level
3081  * making sure the right key of each node is points to 'key'.
3082  * This is used after shifting pointers to the left, so it stops
3083  * fixing up pointers when a given leaf/node is not in slot 0 of the
3084  * higher levels
3085  *
3086  */
3087 static void fixup_low_keys(struct btrfs_path *path,
3088                            struct btrfs_disk_key *key, int level)
3089 {
3090         int i;
3091         struct extent_buffer *t;
3092         int ret;
3093
3094         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
3095                 int tslot = path->slots[i];
3096
3097                 if (!path->nodes[i])
3098                         break;
3099                 t = path->nodes[i];
3100                 ret = tree_mod_log_insert_key(t, tslot, MOD_LOG_KEY_REPLACE,
3101                                 GFP_ATOMIC);
3102                 BUG_ON(ret < 0);
3103                 btrfs_set_node_key(t, key, tslot);
3104                 btrfs_mark_buffer_dirty(path->nodes[i]);
3105                 if (tslot != 0)
3106                         break;
3107         }
3108 }
3109
3110 /*
3111  * update item key.
3112  *
3113  * This function isn't completely safe. It's the caller's responsibility
3114  * that the new key won't break the order
3115  */
3116 void btrfs_set_item_key_safe(struct btrfs_fs_info *fs_info,
3117                              struct btrfs_path *path,
3118                              const struct btrfs_key *new_key)
3119 {
3120         struct btrfs_disk_key disk_key;
3121         struct extent_buffer *eb;
3122         int slot;
3123
3124         eb = path->nodes[0];
3125         slot = path->slots[0];
3126         if (slot > 0) {
3127                 btrfs_item_key(eb, &disk_key, slot - 1);
3128                 if (unlikely(comp_keys(&disk_key, new_key) >= 0)) {
3129                         btrfs_crit(fs_info,
3130                 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
3131                                    slot, btrfs_disk_key_objectid(&disk_key),
3132                                    btrfs_disk_key_type(&disk_key),
3133                                    btrfs_disk_key_offset(&disk_key),
3134                                    new_key->objectid, new_key->type,
3135                                    new_key->offset);
3136                         btrfs_print_leaf(eb);
3137                         BUG();
3138                 }
3139         }
3140         if (slot < btrfs_header_nritems(eb) - 1) {
3141                 btrfs_item_key(eb, &disk_key, slot + 1);
3142                 if (unlikely(comp_keys(&disk_key, new_key) <= 0)) {
3143                         btrfs_crit(fs_info,
3144                 "slot %u key (%llu %u %llu) new key (%llu %u %llu)",
3145                                    slot, btrfs_disk_key_objectid(&disk_key),
3146                                    btrfs_disk_key_type(&disk_key),
3147                                    btrfs_disk_key_offset(&disk_key),
3148                                    new_key->objectid, new_key->type,
3149                                    new_key->offset);
3150                         btrfs_print_leaf(eb);
3151                         BUG();
3152                 }
3153         }
3154
3155         btrfs_cpu_key_to_disk(&disk_key, new_key);
3156         btrfs_set_item_key(eb, &disk_key, slot);
3157         btrfs_mark_buffer_dirty(eb);
3158         if (slot == 0)
3159                 fixup_low_keys(path, &disk_key, 1);
3160 }
3161
3162 /*
3163  * try to push data from one node into the next node left in the
3164  * tree.
3165  *
3166  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
3167  * error, and > 0 if there was no room in the left hand block.
3168  */
3169 static int push_node_left(struct btrfs_trans_handle *trans,
3170                           struct extent_buffer *dst,
3171                           struct extent_buffer *src, int empty)
3172 {
3173         struct btrfs_fs_info *fs_info = trans->fs_info;
3174         int push_items = 0;
3175         int src_nritems;
3176         int dst_nritems;
3177         int ret = 0;
3178
3179         src_nritems = btrfs_header_nritems(src);
3180         dst_nritems = btrfs_header_nritems(dst);
3181         push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3182         WARN_ON(btrfs_header_generation(src) != trans->transid);
3183         WARN_ON(btrfs_header_generation(dst) != trans->transid);
3184
3185         if (!empty && src_nritems <= 8)
3186                 return 1;
3187
3188         if (push_items <= 0)
3189                 return 1;
3190
3191         if (empty) {
3192                 push_items = min(src_nritems, push_items);
3193                 if (push_items < src_nritems) {
3194                         /* leave at least 8 pointers in the node if
3195                          * we aren't going to empty it
3196                          */
3197                         if (src_nritems - push_items < 8) {
3198                                 if (push_items <= 8)
3199                                         return 1;
3200                                 push_items -= 8;
3201                         }
3202                 }
3203         } else
3204                 push_items = min(src_nritems - 8, push_items);
3205
3206         ret = tree_mod_log_eb_copy(dst, src, dst_nritems, 0, push_items);
3207         if (ret) {
3208                 btrfs_abort_transaction(trans, ret);
3209                 return ret;
3210         }
3211         copy_extent_buffer(dst, src,
3212                            btrfs_node_key_ptr_offset(dst_nritems),
3213                            btrfs_node_key_ptr_offset(0),
3214                            push_items * sizeof(struct btrfs_key_ptr));
3215
3216         if (push_items < src_nritems) {
3217                 /*
3218                  * Don't call tree_mod_log_insert_move here, key removal was
3219                  * already fully logged by tree_mod_log_eb_copy above.
3220                  */
3221                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
3222                                       btrfs_node_key_ptr_offset(push_items),
3223                                       (src_nritems - push_items) *
3224                                       sizeof(struct btrfs_key_ptr));
3225         }
3226         btrfs_set_header_nritems(src, src_nritems - push_items);
3227         btrfs_set_header_nritems(dst, dst_nritems + push_items);
3228         btrfs_mark_buffer_dirty(src);
3229         btrfs_mark_buffer_dirty(dst);
3230
3231         return ret;
3232 }
3233
3234 /*
3235  * try to push data from one node into the next node right in the
3236  * tree.
3237  *
3238  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
3239  * error, and > 0 if there was no room in the right hand block.
3240  *
3241  * this will  only push up to 1/2 the contents of the left node over
3242  */
3243 static int balance_node_right(struct btrfs_trans_handle *trans,
3244                               struct extent_buffer *dst,
3245                               struct extent_buffer *src)
3246 {
3247         struct btrfs_fs_info *fs_info = trans->fs_info;
3248         int push_items = 0;
3249         int max_push;
3250         int src_nritems;
3251         int dst_nritems;
3252         int ret = 0;
3253
3254         WARN_ON(btrfs_header_generation(src) != trans->transid);
3255         WARN_ON(btrfs_header_generation(dst) != trans->transid);
3256
3257         src_nritems = btrfs_header_nritems(src);
3258         dst_nritems = btrfs_header_nritems(dst);
3259         push_items = BTRFS_NODEPTRS_PER_BLOCK(fs_info) - dst_nritems;
3260         if (push_items <= 0)
3261                 return 1;
3262
3263         if (src_nritems < 4)
3264                 return 1;
3265
3266         max_push = src_nritems / 2 + 1;
3267         /* don't try to empty the node */
3268         if (max_push >= src_nritems)
3269                 return 1;
3270
3271         if (max_push < push_items)
3272                 push_items = max_push;
3273
3274         ret = tree_mod_log_insert_move(dst, push_items, 0, dst_nritems);
3275         BUG_ON(ret < 0);
3276         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
3277                                       btrfs_node_key_ptr_offset(0),
3278                                       (dst_nritems) *
3279                                       sizeof(struct btrfs_key_ptr));
3280
3281         ret = tree_mod_log_eb_copy(dst, src, 0, src_nritems - push_items,
3282                                    push_items);
3283         if (ret) {
3284                 btrfs_abort_transaction(trans, ret);
3285                 return ret;
3286         }
3287         copy_extent_buffer(dst, src,
3288                            btrfs_node_key_ptr_offset(0),
3289                            btrfs_node_key_ptr_offset(src_nritems - push_items),
3290                            push_items * sizeof(struct btrfs_key_ptr));
3291
3292         btrfs_set_header_nritems(src, src_nritems - push_items);
3293         btrfs_set_header_nritems(dst, dst_nritems + push_items);
3294
3295         btrfs_mark_buffer_dirty(src);
3296         btrfs_mark_buffer_dirty(dst);
3297
3298         return ret;
3299 }
3300
3301 /*
3302  * helper function to insert a new root level in the tree.
3303  * A new node is allocated, and a single item is inserted to
3304  * point to the existing root
3305  *
3306  * returns zero on success or < 0 on failure.
3307  */
3308 static noinline int insert_new_root(struct btrfs_trans_handle *trans,
3309                            struct btrfs_root *root,
3310                            struct btrfs_path *path, int level)
3311 {
3312         struct btrfs_fs_info *fs_info = root->fs_info;
3313         u64 lower_gen;
3314         struct extent_buffer *lower;
3315         struct extent_buffer *c;
3316         struct extent_buffer *old;
3317         struct btrfs_disk_key lower_key;
3318         int ret;
3319
3320         BUG_ON(path->nodes[level]);
3321         BUG_ON(path->nodes[level-1] != root->node);
3322
3323         lower = path->nodes[level-1];
3324         if (level == 1)
3325                 btrfs_item_key(lower, &lower_key, 0);
3326         else
3327                 btrfs_node_key(lower, &lower_key, 0);
3328
3329         c = alloc_tree_block_no_bg_flush(trans, root, 0, &lower_key, level,
3330                                          root->node->start, 0);
3331         if (IS_ERR(c))
3332                 return PTR_ERR(c);
3333
3334         root_add_used(root, fs_info->nodesize);
3335
3336         btrfs_set_header_nritems(c, 1);
3337         btrfs_set_node_key(c, &lower_key, 0);
3338         btrfs_set_node_blockptr(c, 0, lower->start);
3339         lower_gen = btrfs_header_generation(lower);
3340         WARN_ON(lower_gen != trans->transid);
3341
3342         btrfs_set_node_ptr_generation(c, 0, lower_gen);
3343
3344         btrfs_mark_buffer_dirty(c);
3345
3346         old = root->node;
3347         ret = tree_mod_log_insert_root(root->node, c, 0);
3348         BUG_ON(ret < 0);
3349         rcu_assign_pointer(root->node, c);
3350
3351         /* the super has an extra ref to root->node */
3352         free_extent_buffer(old);
3353
3354         add_root_to_dirty_list(root);
3355         atomic_inc(&c->refs);
3356         path->nodes[level] = c;
3357         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
3358         path->slots[level] = 0;
3359         return 0;
3360 }
3361
3362 /*
3363  * worker function to insert a single pointer in a node.
3364  * the node should have enough room for the pointer already
3365  *
3366  * slot and level indicate where you want the key to go, and
3367  * blocknr is the block the key points to.
3368  */
3369 static void insert_ptr(struct btrfs_trans_handle *trans,
3370                        struct btrfs_path *path,
3371                        struct btrfs_disk_key *key, u64 bytenr,
3372                        int slot, int level)
3373 {
3374         struct extent_buffer *lower;
3375         int nritems;
3376         int ret;
3377
3378         BUG_ON(!path->nodes[level]);
3379         btrfs_assert_tree_locked(path->nodes[level]);
3380         lower = path->nodes[level];
3381         nritems = btrfs_header_nritems(lower);
3382         BUG_ON(slot > nritems);
3383         BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(trans->fs_info));
3384         if (slot != nritems) {
3385                 if (level) {
3386                         ret = tree_mod_log_insert_move(lower, slot + 1, slot,
3387                                         nritems - slot);
3388                         BUG_ON(ret < 0);
3389                 }
3390                 memmove_extent_buffer(lower,
3391                               btrfs_node_key_ptr_offset(slot + 1),
3392                               btrfs_node_key_ptr_offset(slot),
3393                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
3394         }
3395         if (level) {
3396                 ret = tree_mod_log_insert_key(lower, slot, MOD_LOG_KEY_ADD,
3397                                 GFP_NOFS);
3398                 BUG_ON(ret < 0);
3399         }
3400         btrfs_set_node_key(lower, key, slot);
3401         btrfs_set_node_blockptr(lower, slot, bytenr);
3402         WARN_ON(trans->transid == 0);
3403         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
3404         btrfs_set_header_nritems(lower, nritems + 1);
3405         btrfs_mark_buffer_dirty(lower);
3406 }
3407
3408 /*
3409  * split the node at the specified level in path in two.
3410  * The path is corrected to point to the appropriate node after the split
3411  *
3412  * Before splitting this tries to make some room in the node by pushing
3413  * left and right, if either one works, it returns right away.
3414  *
3415  * returns 0 on success and < 0 on failure
3416  */
3417 static noinline int split_node(struct btrfs_trans_handle *trans,
3418                                struct btrfs_root *root,
3419                                struct btrfs_path *path, int level)
3420 {
3421         struct btrfs_fs_info *fs_info = root->fs_info;
3422         struct extent_buffer *c;
3423         struct extent_buffer *split;
3424         struct btrfs_disk_key disk_key;
3425         int mid;
3426         int ret;
3427         u32 c_nritems;
3428
3429         c = path->nodes[level];
3430         WARN_ON(btrfs_header_generation(c) != trans->transid);
3431         if (c == root->node) {
3432                 /*
3433                  * trying to split the root, lets make a new one
3434                  *
3435                  * tree mod log: We don't log_removal old root in
3436                  * insert_new_root, because that root buffer will be kept as a
3437                  * normal node. We are going to log removal of half of the
3438                  * elements below with tree_mod_log_eb_copy. We're holding a
3439                  * tree lock on the buffer, which is why we cannot race with
3440                  * other tree_mod_log users.
3441                  */
3442                 ret = insert_new_root(trans, root, path, level + 1);
3443                 if (ret)
3444                         return ret;
3445         } else {
3446                 ret = push_nodes_for_insert(trans, root, path, level);
3447                 c = path->nodes[level];
3448                 if (!ret && btrfs_header_nritems(c) <
3449                     BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3)
3450                         return 0;
3451                 if (ret < 0)
3452                         return ret;
3453         }
3454
3455         c_nritems = btrfs_header_nritems(c);
3456         mid = (c_nritems + 1) / 2;
3457         btrfs_node_key(c, &disk_key, mid);
3458
3459         split = alloc_tree_block_no_bg_flush(trans, root, 0, &disk_key, level,
3460                                              c->start, 0);
3461         if (IS_ERR(split))
3462                 return PTR_ERR(split);
3463
3464         root_add_used(root, fs_info->nodesize);
3465         ASSERT(btrfs_header_level(c) == level);
3466
3467         ret = tree_mod_log_eb_copy(split, c, 0, mid, c_nritems - mid);
3468         if (ret) {
3469                 btrfs_abort_transaction(trans, ret);
3470                 return ret;
3471         }
3472         copy_extent_buffer(split, c,
3473                            btrfs_node_key_ptr_offset(0),
3474                            btrfs_node_key_ptr_offset(mid),
3475                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
3476         btrfs_set_header_nritems(split, c_nritems - mid);
3477         btrfs_set_header_nritems(c, mid);
3478         ret = 0;
3479
3480         btrfs_mark_buffer_dirty(c);
3481         btrfs_mark_buffer_dirty(split);
3482
3483         insert_ptr(trans, path, &disk_key, split->start,
3484                    path->slots[level + 1] + 1, level + 1);
3485
3486         if (path->slots[level] >= mid) {
3487                 path->slots[level] -= mid;
3488                 btrfs_tree_unlock(c);
3489                 free_extent_buffer(c);
3490                 path->nodes[level] = split;
3491                 path->slots[level + 1] += 1;
3492         } else {
3493                 btrfs_tree_unlock(split);
3494                 free_extent_buffer(split);
3495         }
3496         return ret;
3497 }
3498
3499 /*
3500  * how many bytes are required to store the items in a leaf.  start
3501  * and nr indicate which items in the leaf to check.  This totals up the
3502  * space used both by the item structs and the item data
3503  */
3504 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
3505 {
3506         struct btrfs_item *start_item;
3507         struct btrfs_item *end_item;
3508         int data_len;
3509         int nritems = btrfs_header_nritems(l);
3510         int end = min(nritems, start + nr) - 1;
3511
3512         if (!nr)
3513                 return 0;
3514         start_item = btrfs_item_nr(start);
3515         end_item = btrfs_item_nr(end);
3516         data_len = btrfs_item_offset(l, start_item) +
3517                    btrfs_item_size(l, start_item);
3518         data_len = data_len - btrfs_item_offset(l, end_item);
3519         data_len += sizeof(struct btrfs_item) * nr;
3520         WARN_ON(data_len < 0);
3521         return data_len;
3522 }
3523
3524 /*
3525  * The space between the end of the leaf items and
3526  * the start of the leaf data.  IOW, how much room
3527  * the leaf has left for both items and data
3528  */
3529 noinline int btrfs_leaf_free_space(struct extent_buffer *leaf)
3530 {
3531         struct btrfs_fs_info *fs_info = leaf->fs_info;
3532         int nritems = btrfs_header_nritems(leaf);
3533         int ret;
3534
3535         ret = BTRFS_LEAF_DATA_SIZE(fs_info) - leaf_space_used(leaf, 0, nritems);
3536         if (ret < 0) {
3537                 btrfs_crit(fs_info,
3538                            "leaf free space ret %d, leaf data size %lu, used %d nritems %d",
3539                            ret,
3540                            (unsigned long) BTRFS_LEAF_DATA_SIZE(fs_info),
3541                            leaf_space_used(leaf, 0, nritems), nritems);
3542         }
3543         return ret;
3544 }
3545
3546 /*
3547  * min slot controls the lowest index we're willing to push to the
3548  * right.  We'll push up to and including min_slot, but no lower
3549  */
3550 static noinline int __push_leaf_right(struct btrfs_path *path,
3551                                       int data_size, int empty,
3552                                       struct extent_buffer *right,
3553                                       int free_space, u32 left_nritems,
3554                                       u32 min_slot)
3555 {
3556         struct btrfs_fs_info *fs_info = right->fs_info;
3557         struct extent_buffer *left = path->nodes[0];
3558         struct extent_buffer *upper = path->nodes[1];
3559         struct btrfs_map_token token;
3560         struct btrfs_disk_key disk_key;
3561         int slot;
3562         u32 i;
3563         int push_space = 0;
3564         int push_items = 0;
3565         struct btrfs_item *item;
3566         u32 nr;
3567         u32 right_nritems;
3568         u32 data_end;
3569         u32 this_item_size;
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(left);
3589
3590                                 if (space + push_space * 2 > free_space)
3591                                         break;
3592                         }
3593                 }
3594
3595                 if (path->slots[0] == i)
3596                         push_space += data_size;
3597
3598                 this_item_size = btrfs_item_size(left, item);
3599                 if (this_item_size + sizeof(*item) + push_space > free_space)
3600                         break;
3601
3602                 push_items++;
3603                 push_space += this_item_size + sizeof(*item);
3604                 if (i == 0)
3605                         break;
3606                 i--;
3607         }
3608
3609         if (push_items == 0)
3610                 goto out_unlock;
3611
3612         WARN_ON(!empty && push_items == left_nritems);
3613
3614         /* push left to right */
3615         right_nritems = btrfs_header_nritems(right);
3616
3617         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
3618         push_space -= leaf_data_end(left);
3619
3620         /* make room in the right data area */
3621         data_end = leaf_data_end(right);
3622         memmove_extent_buffer(right,
3623                               BTRFS_LEAF_DATA_OFFSET + data_end - push_space,
3624                               BTRFS_LEAF_DATA_OFFSET + data_end,
3625                               BTRFS_LEAF_DATA_SIZE(fs_info) - data_end);
3626
3627         /* copy from the left data area */
3628         copy_extent_buffer(right, left, BTRFS_LEAF_DATA_OFFSET +
3629                      BTRFS_LEAF_DATA_SIZE(fs_info) - push_space,
3630                      BTRFS_LEAF_DATA_OFFSET + leaf_data_end(left),
3631                      push_space);
3632
3633         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
3634                               btrfs_item_nr_offset(0),
3635                               right_nritems * sizeof(struct btrfs_item));
3636
3637         /* copy the items from left to right */
3638         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
3639                    btrfs_item_nr_offset(left_nritems - push_items),
3640                    push_items * sizeof(struct btrfs_item));
3641
3642         /* update the item pointers */
3643         btrfs_init_map_token(&token, right);
3644         right_nritems += push_items;
3645         btrfs_set_header_nritems(right, right_nritems);
3646         push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3647         for (i = 0; i < right_nritems; i++) {
3648                 item = btrfs_item_nr(i);
3649                 push_space -= btrfs_token_item_size(&token, item);
3650                 btrfs_set_token_item_offset(&token, item, push_space);
3651         }
3652
3653         left_nritems -= push_items;
3654         btrfs_set_header_nritems(left, left_nritems);
3655
3656         if (left_nritems)
3657                 btrfs_mark_buffer_dirty(left);
3658         else
3659                 btrfs_clean_tree_block(left);
3660
3661         btrfs_mark_buffer_dirty(right);
3662
3663         btrfs_item_key(right, &disk_key, 0);
3664         btrfs_set_node_key(upper, &disk_key, slot + 1);
3665         btrfs_mark_buffer_dirty(upper);
3666
3667         /* then fixup the leaf pointer in the path */
3668         if (path->slots[0] >= left_nritems) {
3669                 path->slots[0] -= left_nritems;
3670                 if (btrfs_header_nritems(path->nodes[0]) == 0)
3671                         btrfs_clean_tree_block(path->nodes[0]);
3672                 btrfs_tree_unlock(path->nodes[0]);
3673                 free_extent_buffer(path->nodes[0]);
3674                 path->nodes[0] = right;
3675                 path->slots[1] += 1;
3676         } else {
3677                 btrfs_tree_unlock(right);
3678                 free_extent_buffer(right);
3679         }
3680         return 0;
3681
3682 out_unlock:
3683         btrfs_tree_unlock(right);
3684         free_extent_buffer(right);
3685         return 1;
3686 }
3687
3688 /*
3689  * push some data in the path leaf to the right, trying to free up at
3690  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
3691  *
3692  * returns 1 if the push failed because the other node didn't have enough
3693  * room, 0 if everything worked out and < 0 if there were major errors.
3694  *
3695  * this will push starting from min_slot to the end of the leaf.  It won't
3696  * push any slot lower than min_slot
3697  */
3698 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
3699                            *root, struct btrfs_path *path,
3700                            int min_data_size, int data_size,
3701                            int empty, u32 min_slot)
3702 {
3703         struct extent_buffer *left = path->nodes[0];
3704         struct extent_buffer *right;
3705         struct extent_buffer *upper;
3706         int slot;
3707         int free_space;
3708         u32 left_nritems;
3709         int ret;
3710
3711         if (!path->nodes[1])
3712                 return 1;
3713
3714         slot = path->slots[1];
3715         upper = path->nodes[1];
3716         if (slot >= btrfs_header_nritems(upper) - 1)
3717                 return 1;
3718
3719         btrfs_assert_tree_locked(path->nodes[1]);
3720
3721         right = btrfs_read_node_slot(upper, slot + 1);
3722         /*
3723          * slot + 1 is not valid or we fail to read the right node,
3724          * no big deal, just return.
3725          */
3726         if (IS_ERR(right))
3727                 return 1;
3728
3729         btrfs_tree_lock(right);
3730         btrfs_set_lock_blocking_write(right);
3731
3732         free_space = btrfs_leaf_free_space(right);
3733         if (free_space < data_size)
3734                 goto out_unlock;
3735
3736         /* cow and double check */
3737         ret = btrfs_cow_block(trans, root, right, upper,
3738                               slot + 1, &right);
3739         if (ret)
3740                 goto out_unlock;
3741
3742         free_space = btrfs_leaf_free_space(right);
3743         if (free_space < data_size)
3744                 goto out_unlock;
3745
3746         left_nritems = btrfs_header_nritems(left);
3747         if (left_nritems == 0)
3748                 goto out_unlock;
3749
3750         if (path->slots[0] == left_nritems && !empty) {
3751                 /* Key greater than all keys in the leaf, right neighbor has
3752                  * enough room for it and we're not emptying our leaf to delete
3753                  * it, therefore use right neighbor to insert the new item and
3754                  * no need to touch/dirty our left leaf. */
3755                 btrfs_tree_unlock(left);
3756                 free_extent_buffer(left);
3757                 path->nodes[0] = right;
3758                 path->slots[0] = 0;
3759                 path->slots[1]++;
3760                 return 0;
3761         }
3762
3763         return __push_leaf_right(path, min_data_size, empty,
3764                                 right, free_space, left_nritems, min_slot);
3765 out_unlock:
3766         btrfs_tree_unlock(right);
3767         free_extent_buffer(right);
3768         return 1;
3769 }
3770
3771 /*
3772  * push some data in the path leaf to the left, trying to free up at
3773  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
3774  *
3775  * max_slot can put a limit on how far into the leaf we'll push items.  The
3776  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us do all the
3777  * items
3778  */
3779 static noinline int __push_leaf_left(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_fs_info *fs_info = left->fs_info;
3785         struct btrfs_disk_key disk_key;
3786         struct extent_buffer *right = path->nodes[0];
3787         int i;
3788         int push_space = 0;
3789         int push_items = 0;
3790         struct btrfs_item *item;
3791         u32 old_left_nritems;
3792         u32 nr;
3793         int ret = 0;
3794         u32 this_item_size;
3795         u32 old_left_item_size;
3796         struct btrfs_map_token token;
3797
3798         if (empty)
3799                 nr = min(right_nritems, max_slot);
3800         else
3801                 nr = min(right_nritems - 1, max_slot);
3802
3803         for (i = 0; i < nr; i++) {
3804                 item = btrfs_item_nr(i);
3805
3806                 if (!empty && push_items > 0) {
3807                         if (path->slots[0] < i)
3808                                 break;
3809                         if (path->slots[0] == i) {
3810                                 int space = btrfs_leaf_free_space(right);
3811
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(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         btrfs_init_map_token(&token, left);
3852         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
3853         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
3854                 u32 ioff;
3855
3856                 item = btrfs_item_nr(i);
3857
3858                 ioff = btrfs_token_item_offset(&token, item);
3859                 btrfs_set_token_item_offset(&token, item,
3860                       ioff - (BTRFS_LEAF_DATA_SIZE(fs_info) - old_left_item_size));
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(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(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
3883         btrfs_init_map_token(&token, right);
3884         right_nritems -= push_items;
3885         btrfs_set_header_nritems(right, right_nritems);
3886         push_space = BTRFS_LEAF_DATA_SIZE(fs_info);
3887         for (i = 0; i < right_nritems; i++) {
3888                 item = btrfs_item_nr(i);
3889
3890                 push_space = push_space - btrfs_token_item_size(&token, item);
3891                 btrfs_set_token_item_offset(&token, item, push_space);
3892         }
3893
3894         btrfs_mark_buffer_dirty(left);
3895         if (right_nritems)
3896                 btrfs_mark_buffer_dirty(right);
3897         else
3898                 btrfs_clean_tree_block(right);
3899
3900         btrfs_item_key(right, &disk_key, 0);
3901         fixup_low_keys(path, &disk_key, 1);
3902
3903         /* then fixup the leaf pointer in the path */
3904         if (path->slots[0] < push_items) {
3905                 path->slots[0] += old_left_nritems;
3906                 btrfs_tree_unlock(path->nodes[0]);
3907                 free_extent_buffer(path->nodes[0]);
3908                 path->nodes[0] = left;
3909                 path->slots[1] -= 1;
3910         } else {
3911                 btrfs_tree_unlock(left);
3912                 free_extent_buffer(left);
3913                 path->slots[0] -= push_items;
3914         }
3915         BUG_ON(path->slots[0] < 0);
3916         return ret;
3917 out:
3918         btrfs_tree_unlock(left);
3919         free_extent_buffer(left);
3920         return ret;
3921 }
3922
3923 /*
3924  * push some data in the path leaf to the left, trying to free up at
3925  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
3926  *
3927  * max_slot can put a limit on how far into the leaf we'll push items.  The
3928  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us push all the
3929  * items
3930  */
3931 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
3932                           *root, struct btrfs_path *path, int min_data_size,
3933                           int data_size, int empty, u32 max_slot)
3934 {
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 = btrfs_read_node_slot(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_write(left);
3964
3965         free_space = btrfs_leaf_free_space(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(left);
3982         if (free_space < data_size) {
3983                 ret = 1;
3984                 goto out;
3985         }
3986
3987         return __push_leaf_left(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_path *path,
4002                                     struct extent_buffer *l,
4003                                     struct extent_buffer *right,
4004                                     int slot, int mid, int nritems)
4005 {
4006         struct btrfs_fs_info *fs_info = trans->fs_info;
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         nritems = nritems - mid;
4014         btrfs_set_header_nritems(right, nritems);
4015         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(l);
4016
4017         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
4018                            btrfs_item_nr_offset(mid),
4019                            nritems * sizeof(struct btrfs_item));
4020
4021         copy_extent_buffer(right, l,
4022                      BTRFS_LEAF_DATA_OFFSET + BTRFS_LEAF_DATA_SIZE(fs_info) -
4023                      data_copy_size, BTRFS_LEAF_DATA_OFFSET +
4024                      leaf_data_end(l), data_copy_size);
4025
4026         rt_data_off = BTRFS_LEAF_DATA_SIZE(fs_info) - btrfs_item_end_nr(l, mid);
4027
4028         btrfs_init_map_token(&token, right);
4029         for (i = 0; i < nritems; i++) {
4030                 struct btrfs_item *item = btrfs_item_nr(i);
4031                 u32 ioff;
4032
4033                 ioff = btrfs_token_item_offset(&token, item);
4034                 btrfs_set_token_item_offset(&token, item, ioff + rt_data_off);
4035         }
4036
4037         btrfs_set_header_nritems(l, mid);
4038         btrfs_item_key(right, &disk_key, 0);
4039         insert_ptr(trans, path, &disk_key, right->start, path->slots[1] + 1, 1);
4040
4041         btrfs_mark_buffer_dirty(right);
4042         btrfs_mark_buffer_dirty(l);
4043         BUG_ON(path->slots[0] != slot);
4044
4045         if (mid <= slot) {
4046                 btrfs_tree_unlock(path->nodes[0]);
4047                 free_extent_buffer(path->nodes[0]);
4048                 path->nodes[0] = right;
4049                 path->slots[0] -= mid;
4050                 path->slots[1] += 1;
4051         } else {
4052                 btrfs_tree_unlock(right);
4053                 free_extent_buffer(right);
4054         }
4055
4056         BUG_ON(path->slots[0] < 0);
4057 }
4058
4059 /*
4060  * double splits happen when we need to insert a big item in the middle
4061  * of a leaf.  A double split can leave us with 3 mostly empty leaves:
4062  * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
4063  *          A                 B                 C
4064  *
4065  * We avoid this by trying to push the items on either side of our target
4066  * into the adjacent leaves.  If all goes well we can avoid the double split
4067  * completely.
4068  */
4069 static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
4070                                           struct btrfs_root *root,
4071                                           struct btrfs_path *path,
4072                                           int data_size)
4073 {
4074         int ret;
4075         int progress = 0;
4076         int slot;
4077         u32 nritems;
4078         int space_needed = data_size;
4079
4080         slot = path->slots[0];
4081         if (slot < btrfs_header_nritems(path->nodes[0]))
4082                 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
4083
4084         /*
4085          * try to push all the items after our slot into the
4086          * right leaf
4087          */
4088         ret = push_leaf_right(trans, root, path, 1, space_needed, 0, slot);
4089         if (ret < 0)
4090                 return ret;
4091
4092         if (ret == 0)
4093                 progress++;
4094
4095         nritems = btrfs_header_nritems(path->nodes[0]);
4096         /*
4097          * our goal is to get our slot at the start or end of a leaf.  If
4098          * we've done so we're done
4099          */
4100         if (path->slots[0] == 0 || path->slots[0] == nritems)
4101                 return 0;
4102
4103         if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
4104                 return 0;
4105
4106         /* try to push all the items before our slot into the next leaf */
4107         slot = path->slots[0];
4108         space_needed = data_size;
4109         if (slot > 0)
4110                 space_needed -= btrfs_leaf_free_space(path->nodes[0]);
4111         ret = push_leaf_left(trans, root, path, 1, space_needed, 0, slot);
4112         if (ret < 0)
4113                 return ret;
4114
4115         if (ret == 0)
4116                 progress++;
4117
4118         if (progress)
4119                 return 0;
4120         return 1;
4121 }
4122
4123 /*
4124  * split the path's leaf in two, making sure there is at least data_size
4125  * available for the resulting leaf level of the path.
4126  *
4127  * returns 0 if all went well and < 0 on failure.
4128  */
4129 static noinline int split_leaf(struct btrfs_trans_handle *trans,
4130                                struct btrfs_root *root,
4131                                const struct btrfs_key *ins_key,
4132                                struct btrfs_path *path, int data_size,
4133                                int extend)
4134 {
4135         struct btrfs_disk_key disk_key;
4136         struct extent_buffer *l;
4137         u32 nritems;
4138         int mid;
4139         int slot;
4140         struct extent_buffer *right;
4141         struct btrfs_fs_info *fs_info = root->fs_info;
4142         int ret = 0;
4143         int wret;
4144         int split;
4145         int num_doubles = 0;
4146         int tried_avoid_double = 0;
4147
4148         l = path->nodes[0];
4149         slot = path->slots[0];
4150         if (extend && data_size + btrfs_item_size_nr(l, slot) +
4151             sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(fs_info))
4152                 return -EOVERFLOW;
4153
4154         /* first try to make some room by pushing left and right */
4155         if (data_size && path->nodes[1]) {
4156                 int space_needed = data_size;
4157
4158                 if (slot < btrfs_header_nritems(l))
4159                         space_needed -= btrfs_leaf_free_space(l);
4160
4161                 wret = push_leaf_right(trans, root, path, space_needed,
4162                                        space_needed, 0, 0);
4163                 if (wret < 0)
4164                         return wret;
4165                 if (wret) {
4166                         space_needed = data_size;
4167                         if (slot > 0)
4168                                 space_needed -= btrfs_leaf_free_space(l);
4169                         wret = push_leaf_left(trans, root, path, space_needed,
4170                                               space_needed, 0, (u32)-1);
4171                         if (wret < 0)
4172                                 return wret;
4173                 }
4174                 l = path->nodes[0];
4175
4176                 /* did the pushes work? */
4177                 if (btrfs_leaf_free_space(l) >= data_size)
4178                         return 0;
4179         }
4180
4181         if (!path->nodes[1]) {
4182                 ret = insert_new_root(trans, root, path, 1);
4183                 if (ret)
4184                         return ret;
4185         }
4186 again:
4187         split = 1;
4188         l = path->nodes[0];
4189         slot = path->slots[0];
4190         nritems = btrfs_header_nritems(l);
4191         mid = (nritems + 1) / 2;
4192
4193         if (mid <= slot) {
4194                 if (nritems == 1 ||
4195                     leaf_space_used(l, mid, nritems - mid) + data_size >
4196                         BTRFS_LEAF_DATA_SIZE(fs_info)) {
4197                         if (slot >= nritems) {
4198                                 split = 0;
4199                         } else {
4200                                 mid = slot;
4201                                 if (mid != nritems &&
4202                                     leaf_space_used(l, mid, nritems - mid) +
4203                                     data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
4204                                         if (data_size && !tried_avoid_double)
4205                                                 goto push_for_double;
4206                                         split = 2;
4207                                 }
4208                         }
4209                 }
4210         } else {
4211                 if (leaf_space_used(l, 0, mid) + data_size >
4212                         BTRFS_LEAF_DATA_SIZE(fs_info)) {
4213                         if (!extend && data_size && slot == 0) {
4214                                 split = 0;
4215                         } else if ((extend || !data_size) && slot == 0) {
4216                                 mid = 1;
4217                         } else {
4218                                 mid = slot;
4219                                 if (mid != nritems &&
4220                                     leaf_space_used(l, mid, nritems - mid) +
4221                                     data_size > BTRFS_LEAF_DATA_SIZE(fs_info)) {
4222                                         if (data_size && !tried_avoid_double)
4223                                                 goto push_for_double;
4224                                         split = 2;
4225                                 }
4226                         }
4227                 }
4228         }
4229
4230         if (split == 0)
4231                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
4232         else
4233                 btrfs_item_key(l, &disk_key, mid);
4234
4235         right = alloc_tree_block_no_bg_flush(trans, root, 0, &disk_key, 0,
4236                                              l->start, 0);
4237         if (IS_ERR(right))
4238                 return PTR_ERR(right);
4239
4240         root_add_used(root, fs_info->nodesize);
4241
4242         if (split == 0) {
4243                 if (mid <= slot) {
4244                         btrfs_set_header_nritems(right, 0);
4245                         insert_ptr(trans, path, &disk_key,
4246                                    right->start, path->slots[1] + 1, 1);
4247                         btrfs_tree_unlock(path->nodes[0]);
4248                         free_extent_buffer(path->nodes[0]);
4249                         path->nodes[0] = right;
4250                         path->slots[0] = 0;
4251                         path->slots[1] += 1;
4252                 } else {
4253                         btrfs_set_header_nritems(right, 0);
4254                         insert_ptr(trans, path, &disk_key,
4255                                    right->start, path->slots[1], 1);
4256                         btrfs_tree_unlock(path->nodes[0]);
4257                         free_extent_buffer(path->nodes[0]);
4258                         path->nodes[0] = right;
4259                         path->slots[0] = 0;
4260                         if (path->slots[1] == 0)
4261                                 fixup_low_keys(path, &disk_key, 1);
4262                 }
4263                 /*
4264                  * We create a new leaf 'right' for the required ins_len and
4265                  * we'll do btrfs_mark_buffer_dirty() on this leaf after copying
4266                  * the content of ins_len to 'right'.
4267                  */
4268                 return ret;
4269         }
4270
4271         copy_for_split(trans, path, l, right, slot, mid, nritems);
4272
4273         if (split == 2) {
4274                 BUG_ON(num_doubles != 0);
4275                 num_doubles++;
4276                 goto again;
4277         }
4278
4279         return 0;
4280
4281 push_for_double:
4282         push_for_double_split(trans, root, path, data_size);
4283         tried_avoid_double = 1;
4284         if (btrfs_leaf_free_space(path->nodes[0]) >= data_size)
4285                 return 0;
4286         goto again;
4287 }
4288
4289 static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
4290                                          struct btrfs_root *root,
4291                                          struct btrfs_path *path, int ins_len)
4292 {
4293         struct btrfs_key key;
4294         struct extent_buffer *leaf;
4295         struct btrfs_file_extent_item *fi;
4296         u64 extent_len = 0;
4297         u32 item_size;
4298         int ret;
4299
4300         leaf = path->nodes[0];
4301         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4302
4303         BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
4304                key.type != BTRFS_EXTENT_CSUM_KEY);
4305
4306         if (btrfs_leaf_free_space(leaf) >= ins_len)
4307                 return 0;
4308
4309         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4310         if (key.type == BTRFS_EXTENT_DATA_KEY) {
4311                 fi = btrfs_item_ptr(leaf, path->slots[0],
4312                                     struct btrfs_file_extent_item);
4313                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
4314         }
4315         btrfs_release_path(path);
4316
4317         path->keep_locks = 1;
4318         path->search_for_split = 1;
4319         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
4320         path->search_for_split = 0;
4321         if (ret > 0)
4322                 ret = -EAGAIN;
4323         if (ret < 0)
4324                 goto err;
4325
4326         ret = -EAGAIN;
4327         leaf = path->nodes[0];
4328         /* if our item isn't there, return now */
4329         if (item_size != btrfs_item_size_nr(leaf, path->slots[0]))
4330                 goto err;
4331
4332         /* the leaf has  changed, it now has room.  return now */
4333         if (btrfs_leaf_free_space(path->nodes[0]) >= ins_len)
4334                 goto err;
4335
4336         if (key.type == BTRFS_EXTENT_DATA_KEY) {
4337                 fi = btrfs_item_ptr(leaf, path->slots[0],
4338                                     struct btrfs_file_extent_item);
4339                 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
4340                         goto err;
4341         }
4342
4343         btrfs_set_path_blocking(path);
4344         ret = split_leaf(trans, root, &key, path, ins_len, 1);
4345         if (ret)
4346                 goto err;
4347
4348         path->keep_locks = 0;
4349         btrfs_unlock_up_safe(path, 1);
4350         return 0;
4351 err:
4352         path->keep_locks = 0;
4353         return ret;
4354 }
4355
4356 static noinline int split_item(struct btrfs_path *path,
4357                                const struct btrfs_key *new_key,
4358                                unsigned long split_offset)
4359 {
4360         struct extent_buffer *leaf;
4361         struct btrfs_item *item;
4362         struct btrfs_item *new_item;
4363         int slot;
4364         char *buf;
4365         u32 nritems;
4366         u32 item_size;
4367         u32 orig_offset;
4368         struct btrfs_disk_key disk_key;
4369
4370         leaf = path->nodes[0];
4371         BUG_ON(btrfs_leaf_free_space(leaf) < sizeof(struct btrfs_item));
4372
4373         btrfs_set_path_blocking(path);
4374
4375         item = btrfs_item_nr(path->slots[0]);
4376         orig_offset = btrfs_item_offset(leaf, item);
4377         item_size = btrfs_item_size(leaf, item);
4378
4379         buf = kmalloc(item_size, GFP_NOFS);
4380         if (!buf)
4381                 return -ENOMEM;
4382
4383         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
4384                             path->slots[0]), item_size);
4385
4386         slot = path->slots[0] + 1;
4387         nritems = btrfs_header_nritems(leaf);
4388         if (slot != nritems) {
4389                 /* shift the items */
4390                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
4391                                 btrfs_item_nr_offset(slot),
4392                                 (nritems - slot) * sizeof(struct btrfs_item));
4393         }
4394
4395         btrfs_cpu_key_to_disk(&disk_key, new_key);
4396         btrfs_set_item_key(leaf, &disk_key, slot);
4397
4398         new_item = btrfs_item_nr(slot);
4399
4400         btrfs_set_item_offset(leaf, new_item, orig_offset);
4401         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
4402
4403         btrfs_set_item_offset(leaf, item,
4404                               orig_offset + item_size - split_offset);
4405         btrfs_set_item_size(leaf, item, split_offset);
4406
4407         btrfs_set_header_nritems(leaf, nritems + 1);
4408
4409         /* write the data for the start of the original item */
4410         write_extent_buffer(leaf, buf,
4411                             btrfs_item_ptr_offset(leaf, path->slots[0]),
4412                             split_offset);
4413
4414         /* write the data for the new item */
4415         write_extent_buffer(leaf, buf + split_offset,
4416                             btrfs_item_ptr_offset(leaf, slot),
4417                             item_size - split_offset);
4418         btrfs_mark_buffer_dirty(leaf);
4419
4420         BUG_ON(btrfs_leaf_free_space(leaf) < 0);
4421         kfree(buf);
4422         return 0;
4423 }
4424
4425 /*
4426  * This function splits a single item into two items,
4427  * giving 'new_key' to the new item and splitting the
4428  * old one at split_offset (from the start of the item).
4429  *
4430  * The path may be released by this operation.  After
4431  * the split, the path is pointing to the old item.  The
4432  * new item is going to be in the same node as the old one.
4433  *
4434  * Note, the item being split must be smaller enough to live alone on
4435  * a tree block with room for one extra struct btrfs_item
4436  *
4437  * This allows us to split the item in place, keeping a lock on the
4438  * leaf the entire time.
4439  */
4440 int btrfs_split_item(struct btrfs_trans_handle *trans,
4441                      struct btrfs_root *root,
4442                      struct btrfs_path *path,
4443                      const struct btrfs_key *new_key,
4444                      unsigned long split_offset)
4445 {
4446         int ret;
4447         ret = setup_leaf_for_split(trans, root, path,
4448                                    sizeof(struct btrfs_item));
4449         if (ret)
4450                 return ret;
4451
4452         ret = split_item(path, new_key, split_offset);
4453         return ret;
4454 }
4455
4456 /*
4457  * This function duplicate a item, giving 'new_key' to the new item.
4458  * It guarantees both items live in the same tree leaf and the new item
4459  * is contiguous with the original item.
4460  *
4461  * This allows us to split file extent in place, keeping a lock on the
4462  * leaf the entire time.
4463  */
4464 int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
4465                          struct btrfs_root *root,
4466                          struct btrfs_path *path,
4467                          const struct btrfs_key *new_key)
4468 {
4469         struct extent_buffer *leaf;
4470         int ret;
4471         u32 item_size;
4472
4473         leaf = path->nodes[0];
4474         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
4475         ret = setup_leaf_for_split(trans, root, path,
4476                                    item_size + sizeof(struct btrfs_item));
4477         if (ret)
4478                 return ret;
4479
4480         path->slots[0]++;
4481         setup_items_for_insert(root, path, new_key, &item_size,
4482                                item_size, item_size +
4483                                sizeof(struct btrfs_item), 1);
4484         leaf = path->nodes[0];
4485         memcpy_extent_buffer(leaf,
4486                              btrfs_item_ptr_offset(leaf, path->slots[0]),
4487                              btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
4488                              item_size);
4489         return 0;
4490 }
4491
4492 /*
4493  * make the item pointed to by the path smaller.  new_size indicates
4494  * how small to make it, and from_end tells us if we just chop bytes
4495  * off the end of the item or if we shift the item to chop bytes off
4496  * the front.
4497  */
4498 void btrfs_truncate_item(struct btrfs_path *path, u32 new_size, int from_end)
4499 {
4500         int slot;
4501         struct extent_buffer *leaf;
4502         struct btrfs_item *item;
4503         u32 nritems;
4504         unsigned int data_end;
4505         unsigned int old_data_start;
4506         unsigned int old_size;
4507         unsigned int size_diff;
4508         int i;
4509         struct btrfs_map_token token;
4510
4511         leaf = path->nodes[0];
4512         slot = path->slots[0];
4513
4514         old_size = btrfs_item_size_nr(leaf, slot);
4515         if (old_size == new_size)
4516                 return;
4517
4518         nritems = btrfs_header_nritems(leaf);
4519         data_end = leaf_data_end(leaf);
4520
4521         old_data_start = btrfs_item_offset_nr(leaf, slot);
4522
4523         size_diff = old_size - new_size;
4524
4525         BUG_ON(slot < 0);
4526         BUG_ON(slot >= nritems);
4527
4528         /*
4529          * item0..itemN ... dataN.offset..dataN.size .. data0.size
4530          */
4531         /* first correct the data pointers */
4532         btrfs_init_map_token(&token, leaf);
4533         for (i = slot; i < nritems; i++) {
4534                 u32 ioff;
4535                 item = btrfs_item_nr(i);
4536
4537                 ioff = btrfs_token_item_offset(&token, item);
4538                 btrfs_set_token_item_offset(&token, item, ioff + size_diff);
4539         }
4540
4541         /* shift the data */
4542         if (from_end) {
4543                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4544                               data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
4545                               data_end, old_data_start + new_size - data_end);
4546         } else {
4547                 struct btrfs_disk_key disk_key;
4548                 u64 offset;
4549
4550                 btrfs_item_key(leaf, &disk_key, slot);
4551
4552                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
4553                         unsigned long ptr;
4554                         struct btrfs_file_extent_item *fi;
4555
4556                         fi = btrfs_item_ptr(leaf, slot,
4557                                             struct btrfs_file_extent_item);
4558                         fi = (struct btrfs_file_extent_item *)(
4559                              (unsigned long)fi - size_diff);
4560
4561                         if (btrfs_file_extent_type(leaf, fi) ==
4562                             BTRFS_FILE_EXTENT_INLINE) {
4563                                 ptr = btrfs_item_ptr_offset(leaf, slot);
4564                                 memmove_extent_buffer(leaf, ptr,
4565                                       (unsigned long)fi,
4566                                       BTRFS_FILE_EXTENT_INLINE_DATA_START);
4567                         }
4568                 }
4569
4570                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4571                               data_end + size_diff, BTRFS_LEAF_DATA_OFFSET +
4572                               data_end, old_data_start - data_end);
4573
4574                 offset = btrfs_disk_key_offset(&disk_key);
4575                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
4576                 btrfs_set_item_key(leaf, &disk_key, slot);
4577                 if (slot == 0)
4578                         fixup_low_keys(path, &disk_key, 1);
4579         }
4580
4581         item = btrfs_item_nr(slot);
4582         btrfs_set_item_size(leaf, item, new_size);
4583         btrfs_mark_buffer_dirty(leaf);
4584
4585         if (btrfs_leaf_free_space(leaf) < 0) {
4586                 btrfs_print_leaf(leaf);
4587                 BUG();
4588         }
4589 }
4590
4591 /*
4592  * make the item pointed to by the path bigger, data_size is the added size.
4593  */
4594 void btrfs_extend_item(struct btrfs_path *path, u32 data_size)
4595 {
4596         int slot;
4597         struct extent_buffer *leaf;
4598         struct btrfs_item *item;
4599         u32 nritems;
4600         unsigned int data_end;
4601         unsigned int old_data;
4602         unsigned int old_size;
4603         int i;
4604         struct btrfs_map_token token;
4605
4606         leaf = path->nodes[0];
4607
4608         nritems = btrfs_header_nritems(leaf);
4609         data_end = leaf_data_end(leaf);
4610
4611         if (btrfs_leaf_free_space(leaf) < data_size) {
4612                 btrfs_print_leaf(leaf);
4613                 BUG();
4614         }
4615         slot = path->slots[0];
4616         old_data = btrfs_item_end_nr(leaf, slot);
4617
4618         BUG_ON(slot < 0);
4619         if (slot >= nritems) {
4620                 btrfs_print_leaf(leaf);
4621                 btrfs_crit(leaf->fs_info, "slot %d too large, nritems %d",
4622                            slot, nritems);
4623                 BUG();
4624         }
4625
4626         /*
4627          * item0..itemN ... dataN.offset..dataN.size .. data0.size
4628          */
4629         /* first correct the data pointers */
4630         btrfs_init_map_token(&token, leaf);
4631         for (i = slot; i < nritems; i++) {
4632                 u32 ioff;
4633                 item = btrfs_item_nr(i);
4634
4635                 ioff = btrfs_token_item_offset(&token, item);
4636                 btrfs_set_token_item_offset(&token, item, ioff - data_size);
4637         }
4638
4639         /* shift the data */
4640         memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4641                       data_end - data_size, BTRFS_LEAF_DATA_OFFSET +
4642                       data_end, old_data - data_end);
4643
4644         data_end = old_data;
4645         old_size = btrfs_item_size_nr(leaf, slot);
4646         item = btrfs_item_nr(slot);
4647         btrfs_set_item_size(leaf, item, old_size + data_size);
4648         btrfs_mark_buffer_dirty(leaf);
4649
4650         if (btrfs_leaf_free_space(leaf) < 0) {
4651                 btrfs_print_leaf(leaf);
4652                 BUG();
4653         }
4654 }
4655
4656 /*
4657  * this is a helper for btrfs_insert_empty_items, the main goal here is
4658  * to save stack depth by doing the bulk of the work in a function
4659  * that doesn't call btrfs_search_slot
4660  */
4661 void setup_items_for_insert(struct btrfs_root *root, struct btrfs_path *path,
4662                             const struct btrfs_key *cpu_key, u32 *data_size,
4663                             u32 total_data, u32 total_size, int nr)
4664 {
4665         struct btrfs_fs_info *fs_info = root->fs_info;
4666         struct btrfs_item *item;
4667         int i;
4668         u32 nritems;
4669         unsigned int data_end;
4670         struct btrfs_disk_key disk_key;
4671         struct extent_buffer *leaf;
4672         int slot;
4673         struct btrfs_map_token token;
4674
4675         if (path->slots[0] == 0) {
4676                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
4677                 fixup_low_keys(path, &disk_key, 1);
4678         }
4679         btrfs_unlock_up_safe(path, 1);
4680
4681         leaf = path->nodes[0];
4682         slot = path->slots[0];
4683
4684         nritems = btrfs_header_nritems(leaf);
4685         data_end = leaf_data_end(leaf);
4686
4687         if (btrfs_leaf_free_space(leaf) < total_size) {
4688                 btrfs_print_leaf(leaf);
4689                 btrfs_crit(fs_info, "not enough freespace need %u have %d",
4690                            total_size, btrfs_leaf_free_space(leaf));
4691                 BUG();
4692         }
4693
4694         btrfs_init_map_token(&token, leaf);
4695         if (slot != nritems) {
4696                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
4697
4698                 if (old_data < data_end) {
4699                         btrfs_print_leaf(leaf);
4700                         btrfs_crit(fs_info, "slot %d old_data %d data_end %d",
4701                                    slot, old_data, data_end);
4702                         BUG();
4703                 }
4704                 /*
4705                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
4706                  */
4707                 /* first correct the data pointers */
4708                 for (i = slot; i < nritems; i++) {
4709                         u32 ioff;
4710
4711                         item = btrfs_item_nr(i);
4712                         ioff = btrfs_token_item_offset(&token, item);
4713                         btrfs_set_token_item_offset(&token, item,
4714                                                     ioff - total_data);
4715                 }
4716                 /* shift the items */
4717                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
4718                               btrfs_item_nr_offset(slot),
4719                               (nritems - slot) * sizeof(struct btrfs_item));
4720
4721                 /* shift the data */
4722                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4723                               data_end - total_data, BTRFS_LEAF_DATA_OFFSET +
4724                               data_end, old_data - data_end);
4725                 data_end = old_data;
4726         }
4727
4728         /* setup the item for the new data */
4729         for (i = 0; i < nr; i++) {
4730                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
4731                 btrfs_set_item_key(leaf, &disk_key, slot + i);
4732                 item = btrfs_item_nr(slot + i);
4733                 btrfs_set_token_item_offset(&token, item, data_end - data_size[i]);
4734                 data_end -= data_size[i];
4735                 btrfs_set_token_item_size(&token, item, data_size[i]);
4736         }
4737
4738         btrfs_set_header_nritems(leaf, nritems + nr);
4739         btrfs_mark_buffer_dirty(leaf);
4740
4741         if (btrfs_leaf_free_space(leaf) < 0) {
4742                 btrfs_print_leaf(leaf);
4743                 BUG();
4744         }
4745 }
4746
4747 /*
4748  * Given a key and some data, insert items into the tree.
4749  * This does all the path init required, making room in the tree if needed.
4750  */
4751 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
4752                             struct btrfs_root *root,
4753                             struct btrfs_path *path,
4754                             const struct btrfs_key *cpu_key, u32 *data_size,
4755                             int nr)
4756 {
4757         int ret = 0;
4758         int slot;
4759         int i;
4760         u32 total_size = 0;
4761         u32 total_data = 0;
4762
4763         for (i = 0; i < nr; i++)
4764                 total_data += data_size[i];
4765
4766         total_size = total_data + (nr * sizeof(struct btrfs_item));
4767         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
4768         if (ret == 0)
4769                 return -EEXIST;
4770         if (ret < 0)
4771                 return ret;
4772
4773         slot = path->slots[0];
4774         BUG_ON(slot < 0);
4775
4776         setup_items_for_insert(root, path, cpu_key, data_size,
4777                                total_data, total_size, nr);
4778         return 0;
4779 }
4780
4781 /*
4782  * Given a key and some data, insert an item into the tree.
4783  * This does all the path init required, making room in the tree if needed.
4784  */
4785 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4786                       const struct btrfs_key *cpu_key, void *data,
4787                       u32 data_size)
4788 {
4789         int ret = 0;
4790         struct btrfs_path *path;
4791         struct extent_buffer *leaf;
4792         unsigned long ptr;
4793
4794         path = btrfs_alloc_path();
4795         if (!path)
4796                 return -ENOMEM;
4797         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
4798         if (!ret) {
4799                 leaf = path->nodes[0];
4800                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
4801                 write_extent_buffer(leaf, data, ptr, data_size);
4802                 btrfs_mark_buffer_dirty(leaf);
4803         }
4804         btrfs_free_path(path);
4805         return ret;
4806 }
4807
4808 /*
4809  * delete the pointer from a given node.
4810  *
4811  * the tree should have been previously balanced so the deletion does not
4812  * empty a node.
4813  */
4814 static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
4815                     int level, int slot)
4816 {
4817         struct extent_buffer *parent = path->nodes[level];
4818         u32 nritems;
4819         int ret;
4820
4821         nritems = btrfs_header_nritems(parent);
4822         if (slot != nritems - 1) {
4823                 if (level) {
4824                         ret = tree_mod_log_insert_move(parent, slot, slot + 1,
4825                                         nritems - slot - 1);
4826                         BUG_ON(ret < 0);
4827                 }
4828                 memmove_extent_buffer(parent,
4829                               btrfs_node_key_ptr_offset(slot),
4830                               btrfs_node_key_ptr_offset(slot + 1),
4831                               sizeof(struct btrfs_key_ptr) *
4832                               (nritems - slot - 1));
4833         } else if (level) {
4834                 ret = tree_mod_log_insert_key(parent, slot, MOD_LOG_KEY_REMOVE,
4835                                 GFP_NOFS);
4836                 BUG_ON(ret < 0);
4837         }
4838
4839         nritems--;
4840         btrfs_set_header_nritems(parent, nritems);
4841         if (nritems == 0 && parent == root->node) {
4842                 BUG_ON(btrfs_header_level(root->node) != 1);
4843                 /* just turn the root into a leaf and break */
4844                 btrfs_set_header_level(root->node, 0);
4845         } else if (slot == 0) {
4846                 struct btrfs_disk_key disk_key;
4847
4848                 btrfs_node_key(parent, &disk_key, 0);
4849                 fixup_low_keys(path, &disk_key, level + 1);
4850         }
4851         btrfs_mark_buffer_dirty(parent);
4852 }
4853
4854 /*
4855  * a helper function to delete the leaf pointed to by path->slots[1] and
4856  * path->nodes[1].
4857  *
4858  * This deletes the pointer in path->nodes[1] and frees the leaf
4859  * block extent.  zero is returned if it all worked out, < 0 otherwise.
4860  *
4861  * The path must have already been setup for deleting the leaf, including
4862  * all the proper balancing.  path->nodes[1] must be locked.
4863  */
4864 static noinline void btrfs_del_leaf(struct btrfs_trans_handle *trans,
4865                                     struct btrfs_root *root,
4866                                     struct btrfs_path *path,
4867                                     struct extent_buffer *leaf)
4868 {
4869         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
4870         del_ptr(root, path, 1, path->slots[1]);
4871
4872         /*
4873          * btrfs_free_extent is expensive, we want to make sure we
4874          * aren't holding any locks when we call it
4875          */
4876         btrfs_unlock_up_safe(path, 0);
4877
4878         root_sub_used(root, leaf->len);
4879
4880         atomic_inc(&leaf->refs);
4881         btrfs_free_tree_block(trans, root, leaf, 0, 1);
4882         free_extent_buffer_stale(leaf);
4883 }
4884 /*
4885  * delete the item at the leaf level in path.  If that empties
4886  * the leaf, remove it from the tree
4887  */
4888 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4889                     struct btrfs_path *path, int slot, int nr)
4890 {
4891         struct btrfs_fs_info *fs_info = root->fs_info;
4892         struct extent_buffer *leaf;
4893         struct btrfs_item *item;
4894         u32 last_off;
4895         u32 dsize = 0;
4896         int ret = 0;
4897         int wret;
4898         int i;
4899         u32 nritems;
4900
4901         leaf = path->nodes[0];
4902         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
4903
4904         for (i = 0; i < nr; i++)
4905                 dsize += btrfs_item_size_nr(leaf, slot + i);
4906
4907         nritems = btrfs_header_nritems(leaf);
4908
4909         if (slot + nr != nritems) {
4910                 int data_end = leaf_data_end(leaf);
4911                 struct btrfs_map_token token;
4912
4913                 memmove_extent_buffer(leaf, BTRFS_LEAF_DATA_OFFSET +
4914                               data_end + dsize,
4915                               BTRFS_LEAF_DATA_OFFSET + data_end,
4916                               last_off - data_end);
4917
4918                 btrfs_init_map_token(&token, leaf);
4919                 for (i = slot + nr; i < nritems; i++) {
4920                         u32 ioff;
4921
4922                         item = btrfs_item_nr(i);
4923                         ioff = btrfs_token_item_offset(&token, item);
4924                         btrfs_set_token_item_offset(&token, item, ioff + dsize);
4925                 }
4926
4927                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
4928                               btrfs_item_nr_offset(slot + nr),
4929                               sizeof(struct btrfs_item) *
4930                               (nritems - slot - nr));
4931         }
4932         btrfs_set_header_nritems(leaf, nritems - nr);
4933         nritems -= nr;
4934
4935         /* delete the leaf if we've emptied it */
4936         if (nritems == 0) {
4937                 if (leaf == root->node) {
4938                         btrfs_set_header_level(leaf, 0);
4939                 } else {
4940                         btrfs_set_path_blocking(path);
4941                         btrfs_clean_tree_block(leaf);
4942                         btrfs_del_leaf(trans, root, path, leaf);
4943                 }
4944         } else {
4945                 int used = leaf_space_used(leaf, 0, nritems);
4946                 if (slot == 0) {
4947                         struct btrfs_disk_key disk_key;
4948
4949                         btrfs_item_key(leaf, &disk_key, 0);
4950                         fixup_low_keys(path, &disk_key, 1);
4951                 }
4952
4953                 /* delete the leaf if it is mostly empty */
4954                 if (used < BTRFS_LEAF_DATA_SIZE(fs_info) / 3) {
4955                         /* push_leaf_left fixes the path.
4956                          * make sure the path still points to our leaf
4957                          * for possible call to del_ptr below
4958                          */
4959                         slot = path->slots[1];
4960                         atomic_inc(&leaf->refs);
4961
4962                         btrfs_set_path_blocking(path);
4963                         wret = push_leaf_left(trans, root, path, 1, 1,
4964                                               1, (u32)-1);
4965                         if (wret < 0 && wret != -ENOSPC)
4966                                 ret = wret;
4967
4968                         if (path->nodes[0] == leaf &&
4969                             btrfs_header_nritems(leaf)) {
4970                                 wret = push_leaf_right(trans, root, path, 1,
4971                                                        1, 1, 0);
4972                                 if (wret < 0 && wret != -ENOSPC)
4973                                         ret = wret;
4974                         }
4975
4976                         if (btrfs_header_nritems(leaf) == 0) {
4977                                 path->slots[1] = slot;
4978                                 btrfs_del_leaf(trans, root, path, leaf);
4979                                 free_extent_buffer(leaf);
4980                                 ret = 0;
4981                         } else {
4982                                 /* if we're still in the path, make sure
4983                                  * we're dirty.  Otherwise, one of the
4984                                  * push_leaf functions must have already
4985                                  * dirtied this buffer
4986                                  */
4987                                 if (path->nodes[0] == leaf)
4988                                         btrfs_mark_buffer_dirty(leaf);
4989                                 free_extent_buffer(leaf);
4990                         }
4991                 } else {
4992                         btrfs_mark_buffer_dirty(leaf);
4993                 }
4994         }
4995         return ret;
4996 }
4997
4998 /*
4999  * search the tree again to find a leaf with lesser keys
5000  * returns 0 if it found something or 1 if there are no lesser leaves.
5001  * returns < 0 on io errors.
5002  *
5003  * This may release the path, and so you may lose any locks held at the
5004  * time you call it.
5005  */
5006 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
5007 {
5008         struct btrfs_key key;
5009         struct btrfs_disk_key found_key;
5010         int ret;
5011
5012         btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
5013
5014         if (key.offset > 0) {
5015                 key.offset--;
5016         } else if (key.type > 0) {
5017                 key.type--;
5018                 key.offset = (u64)-1;
5019         } else if (key.objectid > 0) {
5020                 key.objectid--;
5021                 key.type = (u8)-1;
5022                 key.offset = (u64)-1;
5023         } else {
5024                 return 1;
5025         }
5026
5027         btrfs_release_path(path);
5028         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5029         if (ret < 0)
5030                 return ret;
5031         btrfs_item_key(path->nodes[0], &found_key, 0);
5032         ret = comp_keys(&found_key, &key);
5033         /*
5034          * We might have had an item with the previous key in the tree right
5035          * before we released our path. And after we released our path, that
5036          * item might have been pushed to the first slot (0) of the leaf we
5037          * were holding due to a tree balance. Alternatively, an item with the
5038          * previous key can exist as the only element of a leaf (big fat item).
5039          * Therefore account for these 2 cases, so that our callers (like
5040          * btrfs_previous_item) don't miss an existing item with a key matching
5041          * the previous key we computed above.
5042          */
5043         if (ret <= 0)
5044                 return 0;
5045         return 1;
5046 }
5047
5048 /*
5049  * A helper function to walk down the tree starting at min_key, and looking
5050  * for nodes or leaves that are have a minimum transaction id.
5051  * This is used by the btree defrag code, and tree logging
5052  *
5053  * This does not cow, but it does stuff the starting key it finds back
5054  * into min_key, so you can call btrfs_search_slot with cow=1 on the
5055  * key and get a writable path.
5056  *
5057  * This honors path->lowest_level to prevent descent past a given level
5058  * of the tree.
5059  *
5060  * min_trans indicates the oldest transaction that you are interested
5061  * in walking through.  Any nodes or leaves older than min_trans are
5062  * skipped over (without reading them).
5063  *
5064  * returns zero if something useful was found, < 0 on error and 1 if there
5065  * was nothing in the tree that matched the search criteria.
5066  */
5067 int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
5068                          struct btrfs_path *path,
5069                          u64 min_trans)
5070 {
5071         struct extent_buffer *cur;
5072         struct btrfs_key found_key;
5073         int slot;
5074         int sret;
5075         u32 nritems;
5076         int level;
5077         int ret = 1;
5078         int keep_locks = path->keep_locks;
5079
5080         path->keep_locks = 1;
5081 again:
5082         cur = btrfs_read_lock_root_node(root);
5083         level = btrfs_header_level(cur);
5084         WARN_ON(path->nodes[level]);
5085         path->nodes[level] = cur;
5086         path->locks[level] = BTRFS_READ_LOCK;
5087
5088         if (btrfs_header_generation(cur) < min_trans) {
5089                 ret = 1;
5090                 goto out;
5091         }
5092         while (1) {
5093                 nritems = btrfs_header_nritems(cur);
5094                 level = btrfs_header_level(cur);
5095                 sret = btrfs_bin_search(cur, min_key, &slot);
5096                 if (sret < 0) {
5097                         ret = sret;
5098                         goto out;
5099                 }
5100
5101                 /* at the lowest level, we're done, setup the path and exit */
5102                 if (level == path->lowest_level) {
5103                         if (slot >= nritems)
5104                                 goto find_next_key;
5105                         ret = 0;
5106                         path->slots[level] = slot;
5107                         btrfs_item_key_to_cpu(cur, &found_key, slot);
5108                         goto out;
5109                 }
5110                 if (sret && slot > 0)
5111                         slot--;
5112                 /*
5113                  * check this node pointer against the min_trans parameters.
5114                  * If it is too old, old, skip to the next one.
5115                  */
5116                 while (slot < nritems) {
5117                         u64 gen;
5118
5119                         gen = btrfs_node_ptr_generation(cur, slot);
5120                         if (gen < min_trans) {
5121                                 slot++;
5122                                 continue;
5123                         }
5124                         break;
5125                 }
5126 find_next_key:
5127                 /*
5128                  * we didn't find a candidate key in this node, walk forward
5129                  * and find another one
5130                  */
5131                 if (slot >= nritems) {
5132                         path->slots[level] = slot;
5133                         btrfs_set_path_blocking(path);
5134                         sret = btrfs_find_next_key(root, path, min_key, level,
5135                                                   min_trans);
5136                         if (sret == 0) {
5137                                 btrfs_release_path(path);
5138                                 goto again;
5139                         } else {
5140                                 goto out;
5141                         }
5142                 }
5143                 /* save our key for returning back */
5144                 btrfs_node_key_to_cpu(cur, &found_key, slot);
5145                 path->slots[level] = slot;
5146                 if (level == path->lowest_level) {
5147                         ret = 0;
5148                         goto out;
5149                 }
5150                 btrfs_set_path_blocking(path);
5151                 cur = btrfs_read_node_slot(cur, slot);
5152                 if (IS_ERR(cur)) {
5153                         ret = PTR_ERR(cur);
5154                         goto out;
5155                 }
5156
5157                 btrfs_tree_read_lock(cur);
5158
5159                 path->locks[level - 1] = BTRFS_READ_LOCK;
5160                 path->nodes[level - 1] = cur;
5161                 unlock_up(path, level, 1, 0, NULL);
5162         }
5163 out:
5164         path->keep_locks = keep_locks;
5165         if (ret == 0) {
5166                 btrfs_unlock_up_safe(path, path->lowest_level + 1);
5167                 btrfs_set_path_blocking(path);
5168                 memcpy(min_key, &found_key, sizeof(found_key));
5169         }
5170         return ret;
5171 }
5172
5173 /*
5174  * this is similar to btrfs_next_leaf, but does not try to preserve
5175  * and fixup the path.  It looks for and returns the next key in the
5176  * tree based on the current path and the min_trans parameters.
5177  *
5178  * 0 is returned if another key is found, < 0 if there are any errors
5179  * and 1 is returned if there are no higher keys in the tree
5180  *
5181  * path->keep_locks should be set to 1 on the search made before
5182  * calling this function.
5183  */
5184 int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
5185                         struct btrfs_key *key, int level, u64 min_trans)
5186 {
5187         int slot;
5188         struct extent_buffer *c;
5189
5190         WARN_ON(!path->keep_locks && !path->skip_locking);
5191         while (level < BTRFS_MAX_LEVEL) {
5192                 if (!path->nodes[level])
5193                         return 1;
5194
5195                 slot = path->slots[level] + 1;
5196                 c = path->nodes[level];
5197 next:
5198                 if (slot >= btrfs_header_nritems(c)) {
5199                         int ret;
5200                         int orig_lowest;
5201                         struct btrfs_key cur_key;
5202                         if (level + 1 >= BTRFS_MAX_LEVEL ||
5203                             !path->nodes[level + 1])
5204                                 return 1;
5205
5206                         if (path->locks[level + 1] || path->skip_locking) {
5207                                 level++;
5208                                 continue;
5209                         }
5210
5211                         slot = btrfs_header_nritems(c) - 1;
5212                         if (level == 0)
5213                                 btrfs_item_key_to_cpu(c, &cur_key, slot);
5214                         else
5215                                 btrfs_node_key_to_cpu(c, &cur_key, slot);
5216
5217                         orig_lowest = path->lowest_level;
5218                         btrfs_release_path(path);
5219                         path->lowest_level = level;
5220                         ret = btrfs_search_slot(NULL, root, &cur_key, path,
5221                                                 0, 0);
5222                         path->lowest_level = orig_lowest;
5223                         if (ret < 0)
5224                                 return ret;
5225
5226                         c = path->nodes[level];
5227                         slot = path->slots[level];
5228                         if (ret == 0)
5229                                 slot++;
5230                         goto next;
5231                 }
5232
5233                 if (level == 0)
5234                         btrfs_item_key_to_cpu(c, key, slot);
5235                 else {
5236                         u64 gen = btrfs_node_ptr_generation(c, slot);
5237
5238                         if (gen < min_trans) {
5239                                 slot++;
5240                                 goto next;
5241                         }
5242                         btrfs_node_key_to_cpu(c, key, slot);
5243                 }
5244                 return 0;
5245         }
5246         return 1;
5247 }
5248
5249 /*
5250  * search the tree again to find a leaf with greater keys
5251  * returns 0 if it found something or 1 if there are no greater leaves.
5252  * returns < 0 on io errors.
5253  */
5254 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
5255 {
5256         return btrfs_next_old_leaf(root, path, 0);
5257 }
5258
5259 int btrfs_next_old_leaf(struct btrfs_root *root, struct btrfs_path *path,
5260                         u64 time_seq)
5261 {
5262         int slot;
5263         int level;
5264         struct extent_buffer *c;
5265         struct extent_buffer *next;
5266         struct btrfs_key key;
5267         u32 nritems;
5268         int ret;
5269         int old_spinning = path->leave_spinning;
5270         int next_rw_lock = 0;
5271
5272         nritems = btrfs_header_nritems(path->nodes[0]);
5273         if (nritems == 0)
5274                 return 1;
5275
5276         btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
5277 again:
5278         level = 1;
5279         next = NULL;
5280         next_rw_lock = 0;
5281         btrfs_release_path(path);
5282
5283         path->keep_locks = 1;
5284         path->leave_spinning = 1;
5285
5286         if (time_seq)
5287                 ret = btrfs_search_old_slot(root, &key, path, time_seq);
5288         else
5289                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5290         path->keep_locks = 0;
5291
5292         if (ret < 0)
5293                 return ret;
5294
5295         nritems = btrfs_header_nritems(path->nodes[0]);
5296         /*
5297          * by releasing the path above we dropped all our locks.  A balance
5298          * could have added more items next to the key that used to be
5299          * at the very end of the block.  So, check again here and
5300          * advance the path if there are now more items available.
5301          */
5302         if (nritems > 0 && path->slots[0] < nritems - 1) {
5303                 if (ret == 0)
5304                         path->slots[0]++;
5305                 ret = 0;
5306                 goto done;
5307         }
5308         /*
5309          * So the above check misses one case:
5310          * - after releasing the path above, someone has removed the item that
5311          *   used to be at the very end of the block, and balance between leafs
5312          *   gets another one with bigger key.offset to replace it.
5313          *
5314          * This one should be returned as well, or we can get leaf corruption
5315          * later(esp. in __btrfs_drop_extents()).
5316          *
5317          * And a bit more explanation about this check,
5318          * with ret > 0, the key isn't found, the path points to the slot
5319          * where it should be inserted, so the path->slots[0] item must be the
5320          * bigger one.
5321          */
5322         if (nritems > 0 && ret > 0 && path->slots[0] == nritems - 1) {
5323                 ret = 0;
5324                 goto done;
5325         }
5326
5327         while (level < BTRFS_MAX_LEVEL) {
5328                 if (!path->nodes[level]) {
5329                         ret = 1;
5330                         goto done;
5331                 }
5332
5333                 slot = path->slots[level] + 1;
5334                 c = path->nodes[level];
5335                 if (slot >= btrfs_header_nritems(c)) {
5336                         level++;
5337                         if (level == BTRFS_MAX_LEVEL) {
5338                                 ret = 1;
5339                                 goto done;
5340                         }
5341                         continue;
5342                 }
5343
5344                 if (next) {
5345                         btrfs_tree_unlock_rw(next, next_rw_lock);
5346                         free_extent_buffer(next);
5347                 }
5348
5349                 next = c;
5350                 next_rw_lock = path->locks[level];
5351                 ret = read_block_for_search(root, path, &next, level,
5352                                             slot, &key);
5353                 if (ret == -EAGAIN)
5354                         goto again;
5355
5356                 if (ret < 0) {
5357                         btrfs_release_path(path);
5358                         goto done;
5359                 }
5360
5361                 if (!path->skip_locking) {
5362                         ret = btrfs_try_tree_read_lock(next);
5363                         if (!ret && time_seq) {
5364                                 /*
5365                                  * If we don't get the lock, we may be racing
5366                                  * with push_leaf_left, holding that lock while
5367                                  * itself waiting for the leaf we've currently
5368                                  * locked. To solve this situation, we give up
5369                                  * on our lock and cycle.
5370                                  */
5371                                 free_extent_buffer(next);
5372                                 btrfs_release_path(path);
5373                                 cond_resched();
5374                                 goto again;
5375                         }
5376                         if (!ret) {
5377                                 btrfs_set_path_blocking(path);
5378                                 btrfs_tree_read_lock(next);
5379                         }
5380                         next_rw_lock = BTRFS_READ_LOCK;
5381                 }
5382                 break;
5383         }
5384         path->slots[level] = slot;
5385         while (1) {
5386                 level--;
5387                 c = path->nodes[level];
5388                 if (path->locks[level])
5389                         btrfs_tree_unlock_rw(c, path->locks[level]);
5390
5391                 free_extent_buffer(c);
5392                 path->nodes[level] = next;
5393                 path->slots[level] = 0;
5394                 if (!path->skip_locking)
5395                         path->locks[level] = next_rw_lock;
5396                 if (!level)
5397                         break;
5398
5399                 ret = read_block_for_search(root, path, &next, level,
5400                                             0, &key);
5401                 if (ret == -EAGAIN)
5402                         goto again;
5403
5404                 if (ret < 0) {
5405                         btrfs_release_path(path);
5406                         goto done;
5407                 }
5408
5409                 if (!path->skip_locking) {
5410                         ret = btrfs_try_tree_read_lock(next);
5411                         if (!ret) {
5412                                 btrfs_set_path_blocking(path);
5413                                 btrfs_tree_read_lock(next);
5414                         }
5415                         next_rw_lock = BTRFS_READ_LOCK;
5416                 }
5417         }
5418         ret = 0;
5419 done:
5420         unlock_up(path, 0, 1, 0, NULL);
5421         path->leave_spinning = old_spinning;
5422         if (!old_spinning)
5423                 btrfs_set_path_blocking(path);
5424
5425         return ret;
5426 }
5427
5428 /*
5429  * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
5430  * searching until it gets past min_objectid or finds an item of 'type'
5431  *
5432  * returns 0 if something is found, 1 if nothing was found and < 0 on error
5433  */
5434 int btrfs_previous_item(struct btrfs_root *root,
5435                         struct btrfs_path *path, u64 min_objectid,
5436                         int type)
5437 {
5438         struct btrfs_key found_key;
5439         struct extent_buffer *leaf;
5440         u32 nritems;
5441         int ret;
5442
5443         while (1) {
5444                 if (path->slots[0] == 0) {
5445                         btrfs_set_path_blocking(path);
5446                         ret = btrfs_prev_leaf(root, path);
5447                         if (ret != 0)
5448                                 return ret;
5449                 } else {
5450                         path->slots[0]--;
5451                 }
5452                 leaf = path->nodes[0];
5453                 nritems = btrfs_header_nritems(leaf);
5454                 if (nritems == 0)
5455                         return 1;
5456                 if (path->slots[0] == nritems)
5457                         path->slots[0]--;
5458
5459                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5460                 if (found_key.objectid < min_objectid)
5461                         break;
5462                 if (found_key.type == type)
5463                         return 0;
5464                 if (found_key.objectid == min_objectid &&
5465                     found_key.type < type)
5466                         break;
5467         }
5468         return 1;
5469 }
5470
5471 /*
5472  * search in extent tree to find a previous Metadata/Data extent item with
5473  * min objecitd.
5474  *
5475  * returns 0 if something is found, 1 if nothing was found and < 0 on error
5476  */
5477 int btrfs_previous_extent_item(struct btrfs_root *root,
5478                         struct btrfs_path *path, u64 min_objectid)
5479 {
5480         struct btrfs_key found_key;
5481         struct extent_buffer *leaf;
5482         u32 nritems;
5483         int ret;
5484
5485         while (1) {
5486                 if (path->slots[0] == 0) {
5487                         btrfs_set_path_blocking(path);
5488                         ret = btrfs_prev_leaf(root, path);
5489                         if (ret != 0)
5490                                 return ret;
5491                 } else {
5492                         path->slots[0]--;
5493                 }
5494                 leaf = path->nodes[0];
5495                 nritems = btrfs_header_nritems(leaf);
5496                 if (nritems == 0)
5497                         return 1;
5498                 if (path->slots[0] == nritems)
5499                         path->slots[0]--;
5500
5501                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5502                 if (found_key.objectid < min_objectid)
5503                         break;
5504                 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
5505                     found_key.type == BTRFS_METADATA_ITEM_KEY)
5506                         return 0;
5507                 if (found_key.objectid == min_objectid &&
5508                     found_key.type < BTRFS_EXTENT_ITEM_KEY)
5509                         break;
5510         }
5511         return 1;
5512 }