ctree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include "ctree.h"
  19 #include "disk-io.h"
  20 #include "transaction.h"
  21 #include "print-tree.h"
  22 #include "repair.h"
  23
  24 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  25                       *root, struct btrfs_path *path, int level);
  26 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  27                       *root, struct btrfs_key *ins_key,
  28                       struct btrfs_path *path, int data_size, int extend);
  29 static int push_node_left(struct btrfs_trans_handle *trans,
  30                           struct btrfs_root *root, struct extent_buffer *dst,
  31                           struct extent_buffer *src, int empty);
  32 static int balance_node_right(struct btrfs_trans_handle *trans,
  33                               struct btrfs_root *root,
  34                               struct extent_buffer *dst_buf,
  35                               struct extent_buffer *src_buf);
  36
  37 inline void btrfs_init_path(struct btrfs_path *p)
  38 {
  39         memset(p, 0, sizeof(*p));
  40 }
  41
  42 struct btrfs_path *btrfs_alloc_path(void)
  43 {
  44         struct btrfs_path *path;
  45         path = kzalloc(sizeof(struct btrfs_path), GFP_NOFS);
  46         return path;
  47 }
  48
  49 void btrfs_free_path(struct btrfs_path *p)
  50 {
  51         if (!p)
  52                 return;
  53         btrfs_release_path(p);
  54         kfree(p);
  55 }
  56
  57 void btrfs_release_path(struct btrfs_path *p)
  58 {
  59         int i;
  60         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  61                 if (!p->nodes[i])
  62                         continue;
  63                 free_extent_buffer(p->nodes[i]);
  64         }
  65         memset(p, 0, sizeof(*p));
  66 }
  67
  68 void add_root_to_dirty_list(struct btrfs_root *root)
  69 {
  70         if (root->track_dirty && list_empty(&root->dirty_list)) {
  71                 list_add(&root->dirty_list,
  72                          &root->fs_info->dirty_cowonly_roots);
  73         }
  74 }
  75
  76 int btrfs_copy_root(struct btrfs_trans_handle *trans,
  77                       struct btrfs_root *root,
  78                       struct extent_buffer *buf,
  79                       struct extent_buffer **cow_ret, u64 new_root_objectid)
  80 {
  81         struct extent_buffer *cow;
  82         int ret = 0;
  83         int level;
  84         struct btrfs_root *new_root;
  85         struct btrfs_disk_key disk_key;
  86
  87         new_root = kmalloc(sizeof(*new_root), GFP_NOFS);
  88         if (!new_root)
  89                 return -ENOMEM;
  90
  91         memcpy(new_root, root, sizeof(*new_root));
  92         new_root->root_key.objectid = new_root_objectid;
  93
  94         WARN_ON(root->ref_cows && trans->transid !=
  95                 root->fs_info->running_transaction->transid);
  96         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
  97
  98         level = btrfs_header_level(buf);
  99         if (level == 0)
 100                 btrfs_item_key(buf, &disk_key, 0);
 101         else
 102                 btrfs_node_key(buf, &disk_key, 0);
 103         cow = btrfs_alloc_free_block(trans, new_root, buf->len,
 104                                      new_root_objectid, &disk_key,
 105                                      level, buf->start, 0);
 106         if (IS_ERR(cow)) {
 107                 kfree(new_root);
 108                 return PTR_ERR(cow);
 109         }
 110
 111         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 112         btrfs_set_header_bytenr(cow, cow->start);
 113         btrfs_set_header_generation(cow, trans->transid);
 114         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 115         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 116                                      BTRFS_HEADER_FLAG_RELOC);
 117         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 118                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 119         else
 120                 btrfs_set_header_owner(cow, new_root_objectid);
 121
 122         write_extent_buffer(cow, root->fs_info->fsid,
 123                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
 124
 125         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 126         ret = btrfs_inc_ref(trans, new_root, cow, 0);
 127         kfree(new_root);
 128
 129         if (ret)
 130                 return ret;
 131
 132         btrfs_mark_buffer_dirty(cow);
 133         *cow_ret = cow;
 134         return 0;
 135 }
 136
 137 /*
 138  * check if the tree block can be shared by multiple trees
 139  */
 140 static int btrfs_block_can_be_shared(struct btrfs_root *root,
 141                                      struct extent_buffer *buf)
 142 {
 143         /*
 144          * Tree blocks not in refernece counted trees and tree roots
 145          * are never shared. If a block was allocated after the last
 146          * snapshot and the block was not allocated by tree relocation,
 147          * we know the block is not shared.
 148          */
 149         if (root->ref_cows &&
 150             buf != root->node && buf != root->commit_root &&
 151             (btrfs_header_generation(buf) <=
 152              btrfs_root_last_snapshot(&root->root_item) ||
 153              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 154                 return 1;
 155 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 156         if (root->ref_cows &&
 157             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 158                 return 1;
 159 #endif
 160         return 0;
 161 }
 162
 163 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
 164                                        struct btrfs_root *root,
 165                                        struct extent_buffer *buf,
 166                                        struct extent_buffer *cow)
 167 {
 168         u64 refs;
 169         u64 owner;
 170         u64 flags;
 171         u64 new_flags = 0;
 172         int ret;
 173
 174         /*
 175          * Backrefs update rules:
 176          *
 177          * Always use full backrefs for extent pointers in tree block
 178          * allocated by tree relocation.
 179          *
 180          * If a shared tree block is no longer referenced by its owner
 181          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
 182          * use full backrefs for extent pointers in tree block.
 183          *
 184          * If a tree block is been relocating
 185          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
 186          * use full backrefs for extent pointers in tree block.
 187          * The reason for this is some operations (such as drop tree)
 188          * are only allowed for blocks use full backrefs.
 189          */
 190
 191         if (btrfs_block_can_be_shared(root, buf)) {
 192                 ret = btrfs_lookup_extent_info(trans, root, buf->start,
 193                                                btrfs_header_level(buf), 1,
 194                                                &refs, &flags);
 195                 BUG_ON(ret);
 196                 BUG_ON(refs == 0);
 197         } else {
 198                 refs = 1;
 199                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 200                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 201                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 202                 else
 203                         flags = 0;
 204         }
 205
 206         owner = btrfs_header_owner(buf);
 207         BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) &&
 208                owner == BTRFS_TREE_RELOC_OBJECTID);
 209
 210         if (refs > 1) {
 211                 if ((owner == root->root_key.objectid ||
 212                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
 213                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
 214                         ret = btrfs_inc_ref(trans, root, buf, 1);
 215                         BUG_ON(ret);
 216
 217                         if (root->root_key.objectid ==
 218                             BTRFS_TREE_RELOC_OBJECTID) {
 219                                 ret = btrfs_dec_ref(trans, root, buf, 0);
 220                                 BUG_ON(ret);
 221                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 222                                 BUG_ON(ret);
 223                         }
 224                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
 225                 } else {
 226
 227                         if (root->root_key.objectid ==
 228                             BTRFS_TREE_RELOC_OBJECTID)
 229                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 230                         else
 231                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 232                         BUG_ON(ret);
 233                 }
 234                 if (new_flags != 0) {
 235                         ret = btrfs_set_block_flags(trans, root, buf->start,
 236                                                     btrfs_header_level(buf),
 237                                                     new_flags);
 238                         BUG_ON(ret);
 239                 }
 240         } else {
 241                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
 242                         if (root->root_key.objectid ==
 243                             BTRFS_TREE_RELOC_OBJECTID)
 244                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 245                         else
 246                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 247                         BUG_ON(ret);
 248                         ret = btrfs_dec_ref(trans, root, buf, 1);
 249                         BUG_ON(ret);
 250                 }
 251                 clean_tree_block(trans, root, buf);
 252         }
 253         return 0;
 254 }
 255
 256 int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 257                              struct btrfs_root *root,
 258                              struct extent_buffer *buf,
 259                              struct extent_buffer *parent, int parent_slot,
 260                              struct extent_buffer **cow_ret,
 261                              u64 search_start, u64 empty_size)
 262 {
 263         struct extent_buffer *cow;
 264         struct btrfs_disk_key disk_key;
 265         int level;
 266
 267         WARN_ON(root->ref_cows && trans->transid !=
 268                 root->fs_info->running_transaction->transid);
 269         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 270
 271         level = btrfs_header_level(buf);
 272
 273         if (level == 0)
 274                 btrfs_item_key(buf, &disk_key, 0);
 275         else
 276                 btrfs_node_key(buf, &disk_key, 0);
 277
 278         cow = btrfs_alloc_free_block(trans, root, buf->len,
 279                                      root->root_key.objectid, &disk_key,
 280                                      level, search_start, empty_size);
 281         if (IS_ERR(cow))
 282                 return PTR_ERR(cow);
 283
 284         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 285         btrfs_set_header_bytenr(cow, cow->start);
 286         btrfs_set_header_generation(cow, trans->transid);
 287         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 288         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 289                                      BTRFS_HEADER_FLAG_RELOC);
 290         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 291                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 292         else
 293                 btrfs_set_header_owner(cow, root->root_key.objectid);
 294
 295         write_extent_buffer(cow, root->fs_info->fsid,
 296                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
 297
 298         WARN_ON(!(buf->flags & EXTENT_BAD_TRANSID) &&
 299                 btrfs_header_generation(buf) > trans->transid);
 300
 301         update_ref_for_cow(trans, root, buf, cow);
 302
 303         if (buf == root->node) {
 304                 root->node = cow;
 305                 extent_buffer_get(cow);
 306
 307                 btrfs_free_extent(trans, root, buf->start, buf->len,
 308                                   0, root->root_key.objectid, level, 0);
 309                 free_extent_buffer(buf);
 310                 add_root_to_dirty_list(root);
 311         } else {
 312                 btrfs_set_node_blockptr(parent, parent_slot,
 313                                         cow->start);
 314                 WARN_ON(trans->transid == 0);
 315                 btrfs_set_node_ptr_generation(parent, parent_slot,
 316                                               trans->transid);
 317                 btrfs_mark_buffer_dirty(parent);
 318                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 319
 320                 btrfs_free_extent(trans, root, buf->start, buf->len,
 321                                   0, root->root_key.objectid, level, 1);
 322         }
 323         if (!list_empty(&buf->recow)) {
 324                 list_del_init(&buf->recow);
 325                 free_extent_buffer(buf);
 326         }
 327         free_extent_buffer(buf);
 328         btrfs_mark_buffer_dirty(cow);
 329         *cow_ret = cow;
 330         return 0;
 331 }
 332
 333 static inline int should_cow_block(struct btrfs_trans_handle *trans,
 334                                    struct btrfs_root *root,
 335                                    struct extent_buffer *buf)
 336 {
 337         if (btrfs_header_generation(buf) == trans->transid &&
 338             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
 339             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
 340               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 341                 return 0;
 342         return 1;
 343 }
 344
 345 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 346                     struct btrfs_root *root, struct extent_buffer *buf,
 347                     struct extent_buffer *parent, int parent_slot,
 348                     struct extent_buffer **cow_ret)
 349 {
 350         u64 search_start;
 351         int ret;
 352         /*
 353         if (trans->transaction != root->fs_info->running_transaction) {
 354                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 355                        root->fs_info->running_transaction->transid);
 356                 WARN_ON(1);
 357         }
 358         */
 359         if (trans->transid != root->fs_info->generation) {
 360                 printk(KERN_CRIT "trans %llu running %llu\n",
 361                         (unsigned long long)trans->transid,
 362                         (unsigned long long)root->fs_info->generation);
 363                 WARN_ON(1);
 364         }
 365         if (!should_cow_block(trans, root, buf)) {
 366                 *cow_ret = buf;
 367                 return 0;
 368         }
 369
 370         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
 371         ret = __btrfs_cow_block(trans, root, buf, parent,
 372                                  parent_slot, cow_ret, search_start, 0);
 373         return ret;
 374 }
 375
 376 int btrfs_comp_cpu_keys(struct btrfs_key *k1, struct btrfs_key *k2)
 377 {
 378         if (k1->objectid > k2->objectid)
 379                 return 1;
 380         if (k1->objectid < k2->objectid)
 381                 return -1;
 382         if (k1->type > k2->type)
 383                 return 1;
 384         if (k1->type < k2->type)
 385                 return -1;
 386         if (k1->offset > k2->offset)
 387                 return 1;
 388         if (k1->offset < k2->offset)
 389                 return -1;
 390         return 0;
 391 }
 392
 393 /*
 394  * compare two keys in a memcmp fashion
 395  */
 396 static int btrfs_comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 397 {
 398         struct btrfs_key k1;
 399
 400         btrfs_disk_key_to_cpu(&k1, disk);
 401         return btrfs_comp_cpu_keys(&k1, k2);
 402 }
 403
 404 /*
 405  * The leaf data grows from end-to-front in the node.
 406  * this returns the address of the start of the last item,
 407  * which is the stop of the leaf data stack
 408  */
 409 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 410                                          struct extent_buffer *leaf)
 411 {
 412         u32 nr = btrfs_header_nritems(leaf);
 413         if (nr == 0)
 414                 return BTRFS_LEAF_DATA_SIZE(root);
 415         return btrfs_item_offset_nr(leaf, nr - 1);
 416 }
 417
 418 enum btrfs_tree_block_status
 419 btrfs_check_node(struct btrfs_root *root, struct btrfs_disk_key *parent_key,
 420                  struct extent_buffer *buf)
 421 {
 422         int i;
 423         struct btrfs_key cpukey;
 424         struct btrfs_disk_key key;
 425         u32 nritems = btrfs_header_nritems(buf);
 426         enum btrfs_tree_block_status ret = BTRFS_TREE_BLOCK_INVALID_NRITEMS;
 427
 428         if (nritems == 0 || nritems > BTRFS_NODEPTRS_PER_BLOCK(root))
 429                 goto fail;
 430
 431         ret = BTRFS_TREE_BLOCK_INVALID_PARENT_KEY;
 432         if (parent_key && parent_key->type) {
 433                 btrfs_node_key(buf, &key, 0);
 434                 if (memcmp(parent_key, &key, sizeof(key)))
 435                         goto fail;
 436         }
 437         ret = BTRFS_TREE_BLOCK_BAD_KEY_ORDER;
 438         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 439                 btrfs_node_key(buf, &key, i);
 440                 btrfs_node_key_to_cpu(buf, &cpukey, i + 1);
 441                 if (btrfs_comp_keys(&key, &cpukey) >= 0)
 442                         goto fail;
 443         }
 444         return BTRFS_TREE_BLOCK_CLEAN;
 445 fail:
 446         if (btrfs_header_owner(buf) == BTRFS_EXTENT_TREE_OBJECTID) {
 447                 if (parent_key)
 448                         btrfs_disk_key_to_cpu(&cpukey, parent_key);
 449                 else
 450                         btrfs_node_key_to_cpu(buf, &cpukey, 0);
 451                 btrfs_add_corrupt_extent_record(root->fs_info, &cpukey,
 452                                                 buf->start, buf->len,
 453                                                 btrfs_header_level(buf));
 454         }
 455         return ret;
 456 }
 457
 458 enum btrfs_tree_block_status
 459 btrfs_check_leaf(struct btrfs_root *root, struct btrfs_disk_key *parent_key,
 460                  struct extent_buffer *buf)
 461 {
 462         int i;
 463         struct btrfs_key cpukey;
 464         struct btrfs_disk_key key;
 465         u32 nritems = btrfs_header_nritems(buf);
 466         enum btrfs_tree_block_status ret = BTRFS_TREE_BLOCK_INVALID_NRITEMS;
 467
 468         if (nritems * sizeof(struct btrfs_item) > buf->len)  {
 469                 fprintf(stderr, "invalid number of items %llu\n",
 470                         (unsigned long long)buf->start);
 471                 goto fail;
 472         }
 473
 474         if (btrfs_header_level(buf) != 0) {
 475                 ret = BTRFS_TREE_BLOCK_INVALID_LEVEL;
 476                 fprintf(stderr, "leaf is not a leaf %llu\n",
 477                        (unsigned long long)btrfs_header_bytenr(buf));
 478                 goto fail;
 479         }
 480         if (btrfs_leaf_free_space(root, buf) < 0) {
 481                 ret = BTRFS_TREE_BLOCK_INVALID_FREE_SPACE;
 482                 fprintf(stderr, "leaf free space incorrect %llu %d\n",
 483                         (unsigned long long)btrfs_header_bytenr(buf),
 484                         btrfs_leaf_free_space(root, buf));
 485                 goto fail;
 486         }
 487
 488         if (nritems == 0)
 489                 return BTRFS_TREE_BLOCK_CLEAN;
 490
 491         btrfs_item_key(buf, &key, 0);
 492         if (parent_key && parent_key->type &&
 493             memcmp(parent_key, &key, sizeof(key))) {
 494                 ret = BTRFS_TREE_BLOCK_INVALID_PARENT_KEY;
 495                 fprintf(stderr, "leaf parent key incorrect %llu\n",
 496                        (unsigned long long)btrfs_header_bytenr(buf));
 497                 goto fail;
 498         }
 499         for (i = 0; nritems > 1 && i < nritems - 1; i++) {
 500                 btrfs_item_key(buf, &key, i);
 501                 btrfs_item_key_to_cpu(buf, &cpukey, i + 1);
 502                 if (btrfs_comp_keys(&key, &cpukey) >= 0) {
 503                         ret = BTRFS_TREE_BLOCK_BAD_KEY_ORDER;
 504                         fprintf(stderr, "bad key ordering %d %d\n", i, i+1);
 505                         goto fail;
 506                 }
 507                 if (btrfs_item_offset_nr(buf, i) !=
 508                         btrfs_item_end_nr(buf, i + 1)) {
 509                         ret = BTRFS_TREE_BLOCK_INVALID_OFFSETS;
 510                         fprintf(stderr, "incorrect offsets %u %u\n",
 511                                 btrfs_item_offset_nr(buf, i),
 512                                 btrfs_item_end_nr(buf, i + 1));
 513                         goto fail;
 514                 }
 515                 if (i == 0 && btrfs_item_end_nr(buf, i) !=
 516                     BTRFS_LEAF_DATA_SIZE(root)) {
 517                         ret = BTRFS_TREE_BLOCK_INVALID_OFFSETS;
 518                         fprintf(stderr, "bad item end %u wanted %u\n",
 519                                 btrfs_item_end_nr(buf, i),
 520                                 (unsigned)BTRFS_LEAF_DATA_SIZE(root));
 521                         goto fail;
 522                 }
 523         }
 524
 525         for (i = 0; i < nritems; i++) {
 526                 if (btrfs_item_end_nr(buf, i) > BTRFS_LEAF_DATA_SIZE(root)) {
 527                         btrfs_item_key(buf, &key, 0);
 528                         btrfs_print_key(&key);
 529                         fflush(stdout);
 530                         ret = BTRFS_TREE_BLOCK_INVALID_OFFSETS;
 531                         fprintf(stderr, "slot end outside of leaf %llu > %llu\n",
 532                                 (unsigned long long)btrfs_item_end_nr(buf, i),
 533                                 (unsigned long long)BTRFS_LEAF_DATA_SIZE(root));
 534                         goto fail;
 535                 }
 536         }
 537
 538         return BTRFS_TREE_BLOCK_CLEAN;
 539 fail:
 540         if (btrfs_header_owner(buf) == BTRFS_EXTENT_TREE_OBJECTID) {
 541                 if (parent_key)
 542                         btrfs_disk_key_to_cpu(&cpukey, parent_key);
 543                 else
 544                         btrfs_item_key_to_cpu(buf, &cpukey, 0);
 545
 546                 btrfs_add_corrupt_extent_record(root->fs_info, &cpukey,
 547                                                 buf->start, buf->len, 0);
 548         }
 549         return ret;
 550 }
 551
 552 static int noinline check_block(struct btrfs_root *root,
 553                                 struct btrfs_path *path, int level)
 554 {
 555         struct btrfs_disk_key key;
 556         struct btrfs_disk_key *key_ptr = NULL;
 557         struct extent_buffer *parent;
 558         enum btrfs_tree_block_status ret;
 559
 560         if (path->skip_check_block)
 561                 return 0;
 562         if (path->nodes[level + 1]) {
 563                 parent = path->nodes[level + 1];
 564                 btrfs_node_key(parent, &key, path->slots[level + 1]);
 565                 key_ptr = &key;
 566         }
 567         if (level == 0)
 568                 ret =  btrfs_check_leaf(root, key_ptr, path->nodes[0]);
 569         else
 570                 ret = btrfs_check_node(root, key_ptr, path->nodes[level]);
 571         if (ret == BTRFS_TREE_BLOCK_CLEAN)
 572                 return 0;
 573         return -EIO;
 574 }
 575
 576 /*
 577  * search for key in the extent_buffer.  The items start at offset p,
 578  * and they are item_size apart.  There are 'max' items in p.
 579  *
 580  * the slot in the array is returned via slot, and it points to
 581  * the place where you would insert key if it is not found in
 582  * the array.
 583  *
 584  * slot may point to max if the key is bigger than all of the keys
 585  */
 586 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 587                               int item_size, struct btrfs_key *key,
 588                               int max, int *slot)
 589 {
 590         int low = 0;
 591         int high = max;
 592         int mid;
 593         int ret;
 594         unsigned long offset;
 595         struct btrfs_disk_key *tmp;
 596
 597         while(low < high) {
 598                 mid = (low + high) / 2;
 599                 offset = p + mid * item_size;
 600
 601                 tmp = (struct btrfs_disk_key *)(eb->data + offset);
 602                 ret = btrfs_comp_keys(tmp, key);
 603
 604                 if (ret < 0)
 605                         low = mid + 1;
 606                 else if (ret > 0)
 607                         high = mid;
 608                 else {
 609                         *slot = mid;
 610                         return 0;
 611                 }
 612         }
 613         *slot = low;
 614         return 1;
 615 }
 616
 617 /*
 618  * simple bin_search frontend that does the right thing for
 619  * leaves vs nodes
 620  */
 621 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 622                       int level, int *slot)
 623 {
 624         if (level == 0)
 625                 return generic_bin_search(eb,
 626                                           offsetof(struct btrfs_leaf, items),
 627                                           sizeof(struct btrfs_item),
 628                                           key, btrfs_header_nritems(eb),
 629                                           slot);
 630         else
 631                 return generic_bin_search(eb,
 632                                           offsetof(struct btrfs_node, ptrs),
 633                                           sizeof(struct btrfs_key_ptr),
 634                                           key, btrfs_header_nritems(eb),
 635                                           slot);
 636 }
 637
 638 struct extent_buffer *read_node_slot(struct btrfs_root *root,
 639                                    struct extent_buffer *parent, int slot)
 640 {
 641         int level = btrfs_header_level(parent);
 642         if (slot < 0)
 643                 return NULL;
 644         if (slot >= btrfs_header_nritems(parent))
 645                 return NULL;
 646
 647         if (level == 0)
 648                 return NULL;
 649
 650         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 651                        btrfs_level_size(root, level - 1),
 652                        btrfs_node_ptr_generation(parent, slot));
 653 }
 654
 655 static int balance_level(struct btrfs_trans_handle *trans,
 656                          struct btrfs_root *root,
 657                          struct btrfs_path *path, int level)
 658 {
 659         struct extent_buffer *right = NULL;
 660         struct extent_buffer *mid;
 661         struct extent_buffer *left = NULL;
 662         struct extent_buffer *parent = NULL;
 663         int ret = 0;
 664         int wret;
 665         int pslot;
 666         int orig_slot = path->slots[level];
 667         u64 orig_ptr;
 668
 669         if (level == 0)
 670                 return 0;
 671
 672         mid = path->nodes[level];
 673         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 674
 675         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 676
 677         if (level < BTRFS_MAX_LEVEL - 1) {
 678                 parent = path->nodes[level + 1];
 679                 pslot = path->slots[level + 1];
 680         }
 681
 682         /*
 683          * deal with the case where there is only one pointer in the root
 684          * by promoting the node below to a root
 685          */
 686         if (!parent) {
 687                 struct extent_buffer *child;
 688
 689                 if (btrfs_header_nritems(mid) != 1)
 690                         return 0;
 691
 692                 /* promote the child to a root */
 693                 child = read_node_slot(root, mid, 0);
 694                 BUG_ON(!extent_buffer_uptodate(child));
 695                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
 696                 BUG_ON(ret);
 697
 698                 root->node = child;
 699                 add_root_to_dirty_list(root);
 700                 path->nodes[level] = NULL;
 701                 clean_tree_block(trans, root, mid);
 702                 wait_on_tree_block_writeback(root, mid);
 703                 /* once for the path */
 704                 free_extent_buffer(mid);
 705
 706                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 707                                         0, root->root_key.objectid,
 708                                         level, 1);
 709                 /* once for the root ptr */
 710                 free_extent_buffer(mid);
 711                 return ret;
 712         }
 713         if (btrfs_header_nritems(mid) >
 714             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 715                 return 0;
 716
 717         left = read_node_slot(root, parent, pslot - 1);
 718         if (extent_buffer_uptodate(left)) {
 719                 wret = btrfs_cow_block(trans, root, left,
 720                                        parent, pslot - 1, &left);
 721                 if (wret) {
 722                         ret = wret;
 723                         goto enospc;
 724                 }
 725         }
 726         right = read_node_slot(root, parent, pslot + 1);
 727         if (extent_buffer_uptodate(right)) {
 728                 wret = btrfs_cow_block(trans, root, right,
 729                                        parent, pslot + 1, &right);
 730                 if (wret) {
 731                         ret = wret;
 732                         goto enospc;
 733                 }
 734         }
 735
 736         /* first, try to make some room in the middle buffer */
 737         if (left) {
 738                 orig_slot += btrfs_header_nritems(left);
 739                 wret = push_node_left(trans, root, left, mid, 1);
 740                 if (wret < 0)
 741                         ret = wret;
 742         }
 743
 744         /*
 745          * then try to empty the right most buffer into the middle
 746          */
 747         if (right) {
 748                 wret = push_node_left(trans, root, mid, right, 1);
 749                 if (wret < 0 && wret != -ENOSPC)
 750                         ret = wret;
 751                 if (btrfs_header_nritems(right) == 0) {
 752                         u64 bytenr = right->start;
 753                         u32 blocksize = right->len;
 754
 755                         clean_tree_block(trans, root, right);
 756                         wait_on_tree_block_writeback(root, right);
 757                         free_extent_buffer(right);
 758                         right = NULL;
 759                         wret = btrfs_del_ptr(trans, root, path,
 760                                              level + 1, pslot + 1);
 761                         if (wret)
 762                                 ret = wret;
 763                         wret = btrfs_free_extent(trans, root, bytenr,
 764                                                  blocksize, 0,
 765                                                  root->root_key.objectid,
 766                                                  level, 0);
 767                         if (wret)
 768                                 ret = wret;
 769                 } else {
 770                         struct btrfs_disk_key right_key;
 771                         btrfs_node_key(right, &right_key, 0);
 772                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 773                         btrfs_mark_buffer_dirty(parent);
 774                 }
 775         }
 776         if (btrfs_header_nritems(mid) == 1) {
 777                 /*
 778                  * we're not allowed to leave a node with one item in the
 779                  * tree during a delete.  A deletion from lower in the tree
 780                  * could try to delete the only pointer in this node.
 781                  * So, pull some keys from the left.
 782                  * There has to be a left pointer at this point because
 783                  * otherwise we would have pulled some pointers from the
 784                  * right
 785                  */
 786                 BUG_ON(!left);
 787                 wret = balance_node_right(trans, root, mid, left);
 788                 if (wret < 0) {
 789                         ret = wret;
 790                         goto enospc;
 791                 }
 792                 if (wret == 1) {
 793                         wret = push_node_left(trans, root, left, mid, 1);
 794                         if (wret < 0)
 795                                 ret = wret;
 796                 }
 797                 BUG_ON(wret == 1);
 798         }
 799         if (btrfs_header_nritems(mid) == 0) {
 800                 /* we've managed to empty the middle node, drop it */
 801                 u64 bytenr = mid->start;
 802                 u32 blocksize = mid->len;
 803                 clean_tree_block(trans, root, mid);
 804                 wait_on_tree_block_writeback(root, mid);
 805                 free_extent_buffer(mid);
 806                 mid = NULL;
 807                 wret = btrfs_del_ptr(trans, root, path, level + 1, pslot);
 808                 if (wret)
 809                         ret = wret;
 810                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 811                                          0, root->root_key.objectid,
 812                                          level, 0);
 813                 if (wret)
 814                         ret = wret;
 815         } else {
 816                 /* update the parent key to reflect our changes */
 817                 struct btrfs_disk_key mid_key;
 818                 btrfs_node_key(mid, &mid_key, 0);
 819                 btrfs_set_node_key(parent, &mid_key, pslot);
 820                 btrfs_mark_buffer_dirty(parent);
 821         }
 822
 823         /* update the path */
 824         if (left) {
 825                 if (btrfs_header_nritems(left) > orig_slot) {
 826                         extent_buffer_get(left);
 827                         path->nodes[level] = left;
 828                         path->slots[level + 1] -= 1;
 829                         path->slots[level] = orig_slot;
 830                         if (mid)
 831                                 free_extent_buffer(mid);
 832                 } else {
 833                         orig_slot -= btrfs_header_nritems(left);
 834                         path->slots[level] = orig_slot;
 835                 }
 836         }
 837         /* double check we haven't messed things up */
 838         check_block(root, path, level);
 839         if (orig_ptr !=
 840             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 841                 BUG();
 842 enospc:
 843         if (right)
 844                 free_extent_buffer(right);
 845         if (left)
 846                 free_extent_buffer(left);
 847         return ret;
 848 }
 849
 850 /* returns zero if the push worked, non-zero otherwise */
 851 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
 852                                           struct btrfs_root *root,
 853                                           struct btrfs_path *path, int level)
 854 {
 855         struct extent_buffer *right = NULL;
 856         struct extent_buffer *mid;
 857         struct extent_buffer *left = NULL;
 858         struct extent_buffer *parent = NULL;
 859         int ret = 0;
 860         int wret;
 861         int pslot;
 862         int orig_slot = path->slots[level];
 863
 864         if (level == 0)
 865                 return 1;
 866
 867         mid = path->nodes[level];
 868         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 869
 870         if (level < BTRFS_MAX_LEVEL - 1) {
 871                 parent = path->nodes[level + 1];
 872                 pslot = path->slots[level + 1];
 873         }
 874
 875         if (!parent)
 876                 return 1;
 877
 878         left = read_node_slot(root, parent, pslot - 1);
 879
 880         /* first, try to make some room in the middle buffer */
 881         if (extent_buffer_uptodate(left)) {
 882                 u32 left_nr;
 883                 left_nr = btrfs_header_nritems(left);
 884                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 885                         wret = 1;
 886                 } else {
 887                         ret = btrfs_cow_block(trans, root, left, parent,
 888                                               pslot - 1, &left);
 889                         if (ret)
 890                                 wret = 1;
 891                         else {
 892                                 wret = push_node_left(trans, root,
 893                                                       left, mid, 0);
 894                         }
 895                 }
 896                 if (wret < 0)
 897                         ret = wret;
 898                 if (wret == 0) {
 899                         struct btrfs_disk_key disk_key;
 900                         orig_slot += left_nr;
 901                         btrfs_node_key(mid, &disk_key, 0);
 902                         btrfs_set_node_key(parent, &disk_key, pslot);
 903                         btrfs_mark_buffer_dirty(parent);
 904                         if (btrfs_header_nritems(left) > orig_slot) {
 905                                 path->nodes[level] = left;
 906                                 path->slots[level + 1] -= 1;
 907                                 path->slots[level] = orig_slot;
 908                                 free_extent_buffer(mid);
 909                         } else {
 910                                 orig_slot -=
 911                                         btrfs_header_nritems(left);
 912                                 path->slots[level] = orig_slot;
 913                                 free_extent_buffer(left);
 914                         }
 915                         return 0;
 916                 }
 917                 free_extent_buffer(left);
 918         }
 919         right= read_node_slot(root, parent, pslot + 1);
 920
 921         /*
 922          * then try to empty the right most buffer into the middle
 923          */
 924         if (extent_buffer_uptodate(right)) {
 925                 u32 right_nr;
 926                 right_nr = btrfs_header_nritems(right);
 927                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 928                         wret = 1;
 929                 } else {
 930                         ret = btrfs_cow_block(trans, root, right,
 931                                               parent, pslot + 1,
 932                                               &right);
 933                         if (ret)
 934                                 wret = 1;
 935                         else {
 936                                 wret = balance_node_right(trans, root,
 937                                                           right, mid);
 938                         }
 939                 }
 940                 if (wret < 0)
 941                         ret = wret;
 942                 if (wret == 0) {
 943                         struct btrfs_disk_key disk_key;
 944
 945                         btrfs_node_key(right, &disk_key, 0);
 946                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
 947                         btrfs_mark_buffer_dirty(parent);
 948
 949                         if (btrfs_header_nritems(mid) <= orig_slot) {
 950                                 path->nodes[level] = right;
 951                                 path->slots[level + 1] += 1;
 952                                 path->slots[level] = orig_slot -
 953                                         btrfs_header_nritems(mid);
 954                                 free_extent_buffer(mid);
 955                         } else {
 956                                 free_extent_buffer(right);
 957                         }
 958                         return 0;
 959                 }
 960                 free_extent_buffer(right);
 961         }
 962         return 1;
 963 }
 964
 965 /*
 966  * readahead one full node of leaves
 967  */
 968 void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
 969                              int level, int slot, u64 objectid)
 970 {
 971         struct extent_buffer *node;
 972         struct btrfs_disk_key disk_key;
 973         u32 nritems;
 974         u64 search;
 975         u64 lowest_read;
 976         u64 highest_read;
 977         u64 nread = 0;
 978         int direction = path->reada;
 979         struct extent_buffer *eb;
 980         u32 nr;
 981         u32 blocksize;
 982         u32 nscan = 0;
 983
 984         if (level != 1)
 985                 return;
 986
 987         if (!path->nodes[level])
 988                 return;
 989
 990         node = path->nodes[level];
 991         search = btrfs_node_blockptr(node, slot);
 992         blocksize = btrfs_level_size(root, level - 1);
 993         eb = btrfs_find_tree_block(root, search, blocksize);
 994         if (eb) {
 995                 free_extent_buffer(eb);
 996                 return;
 997         }
 998
 999         highest_read = search;
1000         lowest_read = search;
1001
1002         nritems = btrfs_header_nritems(node);
1003         nr = slot;
1004         while(1) {
1005                 if (direction < 0) {
1006                         if (nr == 0)
1007                                 break;
1008                         nr--;
1009                 } else if (direction > 0) {
1010                         nr++;
1011                         if (nr >= nritems)
1012                                 break;
1013                 }
1014                 if (path->reada < 0 && objectid) {
1015                         btrfs_node_key(node, &disk_key, nr);
1016                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1017                                 break;
1018                 }
1019                 search = btrfs_node_blockptr(node, nr);
1020                 if ((search >= lowest_read && search <= highest_read) ||
1021                     (search < lowest_read && lowest_read - search <= 32768) ||
1022                     (search > highest_read && search - highest_read <= 32768)) {
1023                         readahead_tree_block(root, search, blocksize,
1024                                      btrfs_node_ptr_generation(node, nr));
1025                         nread += blocksize;
1026                 }
1027                 nscan++;
1028                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
1029                         break;
1030                 if(nread > (1024 * 1024) || nscan > 128)
1031                         break;
1032
1033                 if (search < lowest_read)
1034                         lowest_read = search;
1035                 if (search > highest_read)
1036                         highest_read = search;
1037         }
1038 }
1039
1040 int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *found_path,
1041                 u64 iobjectid, u64 ioff, u8 key_type,
1042                 struct btrfs_key *found_key)
1043 {
1044         int ret;
1045         struct btrfs_key key;
1046         struct extent_buffer *eb;
1047         struct btrfs_path *path;
1048
1049         key.type = key_type;
1050         key.objectid = iobjectid;
1051         key.offset = ioff;
1052
1053         if (found_path == NULL) {
1054                 path = btrfs_alloc_path();
1055                 if (!path)
1056                         return -ENOMEM;
1057         } else
1058                 path = found_path;
1059
1060         ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
1061         if ((ret < 0) || (found_key == NULL))
1062                 goto out;
1063
1064         eb = path->nodes[0];
1065         if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
1066                 ret = btrfs_next_leaf(fs_root, path);
1067                 if (ret)
1068                         goto out;
1069                 eb = path->nodes[0];
1070         }
1071
1072         btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
1073         if (found_key->type != key.type ||
1074                         found_key->objectid != key.objectid) {
1075                 ret = 1;
1076                 goto out;
1077         }
1078
1079 out:
1080         if (path != found_path)
1081                 btrfs_free_path(path);
1082         return ret;
1083 }
1084
1085 /*
1086  * look for key in the tree.  path is filled in with nodes along the way
1087  * if key is found, we return zero and you can find the item in the leaf
1088  * level of the path (level 0)
1089  *
1090  * If the key isn't found, the path points to the slot where it should
1091  * be inserted, and 1 is returned.  If there are other errors during the
1092  * search a negative error number is returned.
1093  *
1094  * if ins_len > 0, nodes and leaves will be split as we walk down the
1095  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1096  * possible)
1097  */
1098 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1099                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1100                       ins_len, int cow)
1101 {
1102         struct extent_buffer *b;
1103         int slot;
1104         int ret;
1105         int level;
1106         int should_reada = p->reada;
1107         u8 lowest_level = 0;
1108
1109         lowest_level = p->lowest_level;
1110         WARN_ON(lowest_level && ins_len > 0);
1111         WARN_ON(p->nodes[0] != NULL);
1112         /*
1113         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
1114         */
1115 again:
1116         b = root->node;
1117         extent_buffer_get(b);
1118         while (b) {
1119                 level = btrfs_header_level(b);
1120                 if (cow) {
1121                         int wret;
1122                         wret = btrfs_cow_block(trans, root, b,
1123                                                p->nodes[level + 1],
1124                                                p->slots[level + 1],
1125                                                &b);
1126                         if (wret) {
1127                                 free_extent_buffer(b);
1128                                 return wret;
1129                         }
1130                 }
1131                 BUG_ON(!cow && ins_len);
1132                 if (level != btrfs_header_level(b))
1133                         WARN_ON(1);
1134                 level = btrfs_header_level(b);
1135                 p->nodes[level] = b;
1136                 ret = check_block(root, p, level);
1137                 if (ret)
1138                         return -1;
1139                 ret = bin_search(b, key, level, &slot);
1140                 if (level != 0) {
1141                         if (ret && slot > 0)
1142                                 slot -= 1;
1143                         p->slots[level] = slot;
1144                         if ((p->search_for_split || ins_len > 0) &&
1145                             btrfs_header_nritems(b) >=
1146                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1147                                 int sret = split_node(trans, root, p, level);
1148                                 BUG_ON(sret > 0);
1149                                 if (sret)
1150                                         return sret;
1151                                 b = p->nodes[level];
1152                                 slot = p->slots[level];
1153                         } else if (ins_len < 0) {
1154                                 int sret = balance_level(trans, root, p,
1155                                                          level);
1156                                 if (sret)
1157                                         return sret;
1158                                 b = p->nodes[level];
1159                                 if (!b) {
1160                                         btrfs_release_path(p);
1161                                         goto again;
1162                                 }
1163                                 slot = p->slots[level];
1164                                 BUG_ON(btrfs_header_nritems(b) == 1);
1165                         }
1166                         /* this is only true while dropping a snapshot */
1167                         if (level == lowest_level)
1168                                 break;
1169
1170                         if (should_reada)
1171                                 reada_for_search(root, p, level, slot,
1172                                                  key->objectid);
1173
1174                         b = read_node_slot(root, b, slot);
1175                         if (!extent_buffer_uptodate(b))
1176                                 return -EIO;
1177                 } else {
1178                         p->slots[level] = slot;
1179                         if (ins_len > 0 &&
1180                             ins_len > btrfs_leaf_free_space(root, b)) {
1181                                 int sret = split_leaf(trans, root, key,
1182                                                       p, ins_len, ret == 0);
1183                                 BUG_ON(sret > 0);
1184                                 if (sret)
1185                                         return sret;
1186                         }
1187                         return ret;
1188                 }
1189         }
1190         return 1;
1191 }
1192
1193 /*
1194  * adjust the pointers going up the tree, starting at level
1195  * making sure the right key of each node is points to 'key'.
1196  * This is used after shifting pointers to the left, so it stops
1197  * fixing up pointers when a given leaf/node is not in slot 0 of the
1198  * higher levels
1199  */
1200 void btrfs_fixup_low_keys(struct btrfs_root *root, struct btrfs_path *path,
1201                           struct btrfs_disk_key *key, int level)
1202 {
1203         int i;
1204         struct extent_buffer *t;
1205
1206         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1207                 int tslot = path->slots[i];
1208                 if (!path->nodes[i])
1209                         break;
1210                 t = path->nodes[i];
1211                 btrfs_set_node_key(t, key, tslot);
1212                 btrfs_mark_buffer_dirty(path->nodes[i]);
1213                 if (tslot != 0)
1214                         break;
1215         }
1216 }
1217
1218 /*
1219  * update item key.
1220  *
1221  * This function isn't completely safe. It's the caller's responsibility
1222  * that the new key won't break the order
1223  */
1224 int btrfs_set_item_key_safe(struct btrfs_root *root, struct btrfs_path *path,
1225                             struct btrfs_key *new_key)
1226 {
1227         struct btrfs_disk_key disk_key;
1228         struct extent_buffer *eb;
1229         int slot;
1230
1231         eb = path->nodes[0];
1232         slot = path->slots[0];
1233         if (slot > 0) {
1234                 btrfs_item_key(eb, &disk_key, slot - 1);
1235                 if (btrfs_comp_keys(&disk_key, new_key) >= 0)
1236                         return -1;
1237         }
1238         if (slot < btrfs_header_nritems(eb) - 1) {
1239                 btrfs_item_key(eb, &disk_key, slot + 1);
1240                 if (btrfs_comp_keys(&disk_key, new_key) <= 0)
1241                         return -1;
1242         }
1243
1244         btrfs_cpu_key_to_disk(&disk_key, new_key);
1245         btrfs_set_item_key(eb, &disk_key, slot);
1246         btrfs_mark_buffer_dirty(eb);
1247         if (slot == 0)
1248                 btrfs_fixup_low_keys(root, path, &disk_key, 1);
1249         return 0;
1250 }
1251
1252 /*
1253  * update an item key without the safety checks.  This is meant to be called by
1254  * fsck only.
1255  */
1256 void btrfs_set_item_key_unsafe(struct btrfs_root *root,
1257                                struct btrfs_path *path,
1258                                struct btrfs_key *new_key)
1259 {
1260         struct btrfs_disk_key disk_key;
1261         struct extent_buffer *eb;
1262         int slot;
1263
1264         eb = path->nodes[0];
1265         slot = path->slots[0];
1266
1267         btrfs_cpu_key_to_disk(&disk_key, new_key);
1268         btrfs_set_item_key(eb, &disk_key, slot);
1269         btrfs_mark_buffer_dirty(eb);
1270         if (slot == 0)
1271                 btrfs_fixup_low_keys(root, path, &disk_key, 1);
1272 }
1273
1274 /*
1275  * try to push data from one node into the next node left in the
1276  * tree.
1277  *
1278  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1279  * error, and > 0 if there was no room in the left hand block.
1280  */
1281 static int push_node_left(struct btrfs_trans_handle *trans,
1282                           struct btrfs_root *root, struct extent_buffer *dst,
1283                           struct extent_buffer *src, int empty)
1284 {
1285         int push_items = 0;
1286         int src_nritems;
1287         int dst_nritems;
1288         int ret = 0;
1289
1290         src_nritems = btrfs_header_nritems(src);
1291         dst_nritems = btrfs_header_nritems(dst);
1292         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1293         WARN_ON(btrfs_header_generation(src) != trans->transid);
1294         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1295
1296         if (!empty && src_nritems <= 8)
1297                 return 1;
1298
1299         if (push_items <= 0) {
1300                 return 1;
1301         }
1302
1303         if (empty) {
1304                 push_items = min(src_nritems, push_items);
1305                 if (push_items < src_nritems) {
1306                         /* leave at least 8 pointers in the node if
1307                          * we aren't going to empty it
1308                          */
1309                         if (src_nritems - push_items < 8) {
1310                                 if (push_items <= 8)
1311                                         return 1;
1312                                 push_items -= 8;
1313                         }
1314                 }
1315         } else
1316                 push_items = min(src_nritems - 8, push_items);
1317
1318         copy_extent_buffer(dst, src,
1319                            btrfs_node_key_ptr_offset(dst_nritems),
1320                            btrfs_node_key_ptr_offset(0),
1321                            push_items * sizeof(struct btrfs_key_ptr));
1322
1323         if (push_items < src_nritems) {
1324                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1325                                       btrfs_node_key_ptr_offset(push_items),
1326                                       (src_nritems - push_items) *
1327                                       sizeof(struct btrfs_key_ptr));
1328         }
1329         btrfs_set_header_nritems(src, src_nritems - push_items);
1330         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1331         btrfs_mark_buffer_dirty(src);
1332         btrfs_mark_buffer_dirty(dst);
1333
1334         return ret;
1335 }
1336
1337 /*
1338  * try to push data from one node into the next node right in the
1339  * tree.
1340  *
1341  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1342  * error, and > 0 if there was no room in the right hand block.
1343  *
1344  * this will  only push up to 1/2 the contents of the left node over
1345  */
1346 static int balance_node_right(struct btrfs_trans_handle *trans,
1347                               struct btrfs_root *root,
1348                               struct extent_buffer *dst,
1349                               struct extent_buffer *src)
1350 {
1351         int push_items = 0;
1352         int max_push;
1353         int src_nritems;
1354         int dst_nritems;
1355         int ret = 0;
1356
1357         WARN_ON(btrfs_header_generation(src) != trans->transid);
1358         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1359
1360         src_nritems = btrfs_header_nritems(src);
1361         dst_nritems = btrfs_header_nritems(dst);
1362         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1363         if (push_items <= 0) {
1364                 return 1;
1365         }
1366
1367         if (src_nritems < 4) {
1368                 return 1;
1369         }
1370
1371         max_push = src_nritems / 2 + 1;
1372         /* don't try to empty the node */
1373         if (max_push >= src_nritems) {
1374                 return 1;
1375         }
1376
1377         if (max_push < push_items)
1378                 push_items = max_push;
1379
1380         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1381                                       btrfs_node_key_ptr_offset(0),
1382                                       (dst_nritems) *
1383                                       sizeof(struct btrfs_key_ptr));
1384
1385         copy_extent_buffer(dst, src,
1386                            btrfs_node_key_ptr_offset(0),
1387                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1388                            push_items * sizeof(struct btrfs_key_ptr));
1389
1390         btrfs_set_header_nritems(src, src_nritems - push_items);
1391         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1392
1393         btrfs_mark_buffer_dirty(src);
1394         btrfs_mark_buffer_dirty(dst);
1395
1396         return ret;
1397 }
1398
1399 /*
1400  * helper function to insert a new root level in the tree.
1401  * A new node is allocated, and a single item is inserted to
1402  * point to the existing root
1403  *
1404  * returns zero on success or < 0 on failure.
1405  */
1406 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1407                            struct btrfs_root *root,
1408                            struct btrfs_path *path, int level)
1409 {
1410         u64 lower_gen;
1411         struct extent_buffer *lower;
1412         struct extent_buffer *c;
1413         struct extent_buffer *old;
1414         struct btrfs_disk_key lower_key;
1415
1416         BUG_ON(path->nodes[level]);
1417         BUG_ON(path->nodes[level-1] != root->node);
1418
1419         lower = path->nodes[level-1];
1420         if (level == 1)
1421                 btrfs_item_key(lower, &lower_key, 0);
1422         else
1423                 btrfs_node_key(lower, &lower_key, 0);
1424
1425         c = btrfs_alloc_free_block(trans, root, root->nodesize,
1426                                    root->root_key.objectid, &lower_key,
1427                                    level, root->node->start, 0);
1428
1429         if (IS_ERR(c))
1430                 return PTR_ERR(c);
1431
1432         memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
1433         btrfs_set_header_nritems(c, 1);
1434         btrfs_set_header_level(c, level);
1435         btrfs_set_header_bytenr(c, c->start);
1436         btrfs_set_header_generation(c, trans->transid);
1437         btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
1438         btrfs_set_header_owner(c, root->root_key.objectid);
1439
1440         write_extent_buffer(c, root->fs_info->fsid,
1441                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
1442
1443         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1444                             btrfs_header_chunk_tree_uuid(c),
1445                             BTRFS_UUID_SIZE);
1446
1447         btrfs_set_node_key(c, &lower_key, 0);
1448         btrfs_set_node_blockptr(c, 0, lower->start);
1449         lower_gen = btrfs_header_generation(lower);
1450         WARN_ON(lower_gen != trans->transid);
1451
1452         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1453
1454         btrfs_mark_buffer_dirty(c);
1455
1456         old = root->node;
1457         root->node = c;
1458
1459         /* the super has an extra ref to root->node */
1460         free_extent_buffer(old);
1461
1462         add_root_to_dirty_list(root);
1463         extent_buffer_get(c);
1464         path->nodes[level] = c;
1465         path->slots[level] = 0;
1466         return 0;
1467 }
1468
1469 /*
1470  * worker function to insert a single pointer in a node.
1471  * the node should have enough room for the pointer already
1472  *
1473  * slot and level indicate where you want the key to go, and
1474  * blocknr is the block the key points to.
1475  *
1476  * returns zero on success and < 0 on any error
1477  */
1478 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1479                       *root, struct btrfs_path *path, struct btrfs_disk_key
1480                       *key, u64 bytenr, int slot, int level)
1481 {
1482         struct extent_buffer *lower;
1483         int nritems;
1484
1485         BUG_ON(!path->nodes[level]);
1486         lower = path->nodes[level];
1487         nritems = btrfs_header_nritems(lower);
1488         if (slot > nritems)
1489                 BUG();
1490         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1491                 BUG();
1492         if (slot != nritems) {
1493                 memmove_extent_buffer(lower,
1494                               btrfs_node_key_ptr_offset(slot + 1),
1495                               btrfs_node_key_ptr_offset(slot),
1496                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1497         }
1498         btrfs_set_node_key(lower, key, slot);
1499         btrfs_set_node_blockptr(lower, slot, bytenr);
1500         WARN_ON(trans->transid == 0);
1501         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1502         btrfs_set_header_nritems(lower, nritems + 1);
1503         btrfs_mark_buffer_dirty(lower);
1504         return 0;
1505 }
1506
1507 /*
1508  * split the node at the specified level in path in two.
1509  * The path is corrected to point to the appropriate node after the split
1510  *
1511  * Before splitting this tries to make some room in the node by pushing
1512  * left and right, if either one works, it returns right away.
1513  *
1514  * returns 0 on success and < 0 on failure
1515  */
1516 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1517                       *root, struct btrfs_path *path, int level)
1518 {
1519         struct extent_buffer *c;
1520         struct extent_buffer *split;
1521         struct btrfs_disk_key disk_key;
1522         int mid;
1523         int ret;
1524         int wret;
1525         u32 c_nritems;
1526
1527         c = path->nodes[level];
1528         WARN_ON(btrfs_header_generation(c) != trans->transid);
1529         if (c == root->node) {
1530                 /* trying to split the root, lets make a new one */
1531                 ret = insert_new_root(trans, root, path, level + 1);
1532                 if (ret)
1533                         return ret;
1534         } else {
1535                 ret = push_nodes_for_insert(trans, root, path, level);
1536                 c = path->nodes[level];
1537                 if (!ret && btrfs_header_nritems(c) <
1538                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
1539                         return 0;
1540                 if (ret < 0)
1541                         return ret;
1542         }
1543
1544         c_nritems = btrfs_header_nritems(c);
1545         mid = (c_nritems + 1) / 2;
1546         btrfs_node_key(c, &disk_key, mid);
1547
1548         split = btrfs_alloc_free_block(trans, root, root->nodesize,
1549                                         root->root_key.objectid,
1550                                         &disk_key, level, c->start, 0);
1551         if (IS_ERR(split))
1552                 return PTR_ERR(split);
1553
1554         memset_extent_buffer(split, 0, 0, sizeof(struct btrfs_header));
1555         btrfs_set_header_level(split, btrfs_header_level(c));
1556         btrfs_set_header_bytenr(split, split->start);
1557         btrfs_set_header_generation(split, trans->transid);
1558         btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
1559         btrfs_set_header_owner(split, root->root_key.objectid);
1560         write_extent_buffer(split, root->fs_info->fsid,
1561                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
1562         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1563                             btrfs_header_chunk_tree_uuid(split),
1564                             BTRFS_UUID_SIZE);
1565
1566
1567         copy_extent_buffer(split, c,
1568                            btrfs_node_key_ptr_offset(0),
1569                            btrfs_node_key_ptr_offset(mid),
1570                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1571         btrfs_set_header_nritems(split, c_nritems - mid);
1572         btrfs_set_header_nritems(c, mid);
1573         ret = 0;
1574
1575         btrfs_mark_buffer_dirty(c);
1576         btrfs_mark_buffer_dirty(split);
1577
1578         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1579                           path->slots[level + 1] + 1,
1580                           level + 1);
1581         if (wret)
1582                 ret = wret;
1583
1584         if (path->slots[level] >= mid) {
1585                 path->slots[level] -= mid;
1586                 free_extent_buffer(c);
1587                 path->nodes[level] = split;
1588                 path->slots[level + 1] += 1;
1589         } else {
1590                 free_extent_buffer(split);
1591         }
1592         return ret;
1593 }
1594
1595 /*
1596  * how many bytes are required to store the items in a leaf.  start
1597  * and nr indicate which items in the leaf to check.  This totals up the
1598  * space used both by the item structs and the item data
1599  */
1600 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1601 {
1602         int data_len;
1603         int nritems = btrfs_header_nritems(l);
1604         int end = min(nritems, start + nr) - 1;
1605
1606         if (!nr)
1607                 return 0;
1608         data_len = btrfs_item_end_nr(l, start);
1609         data_len = data_len - btrfs_item_offset_nr(l, end);
1610         data_len += sizeof(struct btrfs_item) * nr;
1611         WARN_ON(data_len < 0);
1612         return data_len;
1613 }
1614
1615 /*
1616  * The space between the end of the leaf items and
1617  * the start of the leaf data.  IOW, how much room
1618  * the leaf has left for both items and data
1619  */
1620 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1621 {
1622         int nritems = btrfs_header_nritems(leaf);
1623         int ret;
1624         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1625         if (ret < 0) {
1626                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1627                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1628                        leaf_space_used(leaf, 0, nritems), nritems);
1629         }
1630         return ret;
1631 }
1632
1633 /*
1634  * push some data in the path leaf to the right, trying to free up at
1635  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1636  *
1637  * returns 1 if the push failed because the other node didn't have enough
1638  * room, 0 if everything worked out and < 0 if there were major errors.
1639  */
1640 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1641                            *root, struct btrfs_path *path, int data_size,
1642                            int empty)
1643 {
1644         struct extent_buffer *left = path->nodes[0];
1645         struct extent_buffer *right;
1646         struct extent_buffer *upper;
1647         struct btrfs_disk_key disk_key;
1648         int slot;
1649         u32 i;
1650         int free_space;
1651         int push_space = 0;
1652         int push_items = 0;
1653         struct btrfs_item *item;
1654         u32 left_nritems;
1655         u32 nr;
1656         u32 right_nritems;
1657         u32 data_end;
1658         u32 this_item_size;
1659         int ret;
1660
1661         slot = path->slots[1];
1662         if (!path->nodes[1]) {
1663                 return 1;
1664         }
1665         upper = path->nodes[1];
1666         if (slot >= btrfs_header_nritems(upper) - 1)
1667                 return 1;
1668
1669         right = read_node_slot(root, upper, slot + 1);
1670         if (!extent_buffer_uptodate(right)) {
1671                 if (IS_ERR(right))
1672                         return PTR_ERR(right);
1673                 return -EIO;
1674         }
1675         free_space = btrfs_leaf_free_space(root, right);
1676         if (free_space < data_size) {
1677                 free_extent_buffer(right);
1678                 return 1;
1679         }
1680
1681         /* cow and double check */
1682         ret = btrfs_cow_block(trans, root, right, upper,
1683                               slot + 1, &right);
1684         if (ret) {
1685                 free_extent_buffer(right);
1686                 return 1;
1687         }
1688         free_space = btrfs_leaf_free_space(root, right);
1689         if (free_space < data_size) {
1690                 free_extent_buffer(right);
1691                 return 1;
1692         }
1693
1694         left_nritems = btrfs_header_nritems(left);
1695         if (left_nritems == 0) {
1696                 free_extent_buffer(right);
1697                 return 1;
1698         }
1699
1700         if (empty)
1701                 nr = 0;
1702         else
1703                 nr = 1;
1704
1705         i = left_nritems - 1;
1706         while (i >= nr) {
1707                 item = btrfs_item_nr(i);
1708
1709                 if (path->slots[0] == i)
1710                         push_space += data_size + sizeof(*item);
1711
1712                 this_item_size = btrfs_item_size(left, item);
1713                 if (this_item_size + sizeof(*item) + push_space > free_space)
1714                         break;
1715                 push_items++;
1716                 push_space += this_item_size + sizeof(*item);
1717                 if (i == 0)
1718                         break;
1719                 i--;
1720         }
1721
1722         if (push_items == 0) {
1723                 free_extent_buffer(right);
1724                 return 1;
1725         }
1726
1727         if (!empty && push_items == left_nritems)
1728                 WARN_ON(1);
1729
1730         /* push left to right */
1731         right_nritems = btrfs_header_nritems(right);
1732
1733         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1734         push_space -= leaf_data_end(root, left);
1735
1736         /* make room in the right data area */
1737         data_end = leaf_data_end(root, right);
1738         memmove_extent_buffer(right,
1739                               btrfs_leaf_data(right) + data_end - push_space,
1740                               btrfs_leaf_data(right) + data_end,
1741                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1742
1743         /* copy from the left data area */
1744         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1745                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1746                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1747                      push_space);
1748
1749         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1750                               btrfs_item_nr_offset(0),
1751                               right_nritems * sizeof(struct btrfs_item));
1752
1753         /* copy the items from left to right */
1754         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1755                    btrfs_item_nr_offset(left_nritems - push_items),
1756                    push_items * sizeof(struct btrfs_item));
1757
1758         /* update the item pointers */
1759         right_nritems += push_items;
1760         btrfs_set_header_nritems(right, right_nritems);
1761         push_space = BTRFS_LEAF_DATA_SIZE(root);
1762         for (i = 0; i < right_nritems; i++) {
1763                 item = btrfs_item_nr(i);
1764                 push_space -= btrfs_item_size(right, item);
1765                 btrfs_set_item_offset(right, item, push_space);
1766         }
1767
1768         left_nritems -= push_items;
1769         btrfs_set_header_nritems(left, left_nritems);
1770
1771         if (left_nritems)
1772                 btrfs_mark_buffer_dirty(left);
1773         btrfs_mark_buffer_dirty(right);
1774
1775         btrfs_item_key(right, &disk_key, 0);
1776         btrfs_set_node_key(upper, &disk_key, slot + 1);
1777         btrfs_mark_buffer_dirty(upper);
1778
1779         /* then fixup the leaf pointer in the path */
1780         if (path->slots[0] >= left_nritems) {
1781                 path->slots[0] -= left_nritems;
1782                 free_extent_buffer(path->nodes[0]);
1783                 path->nodes[0] = right;
1784                 path->slots[1] += 1;
1785         } else {
1786                 free_extent_buffer(right);
1787         }
1788         return 0;
1789 }
1790 /*
1791  * push some data in the path leaf to the left, trying to free up at
1792  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1793  */
1794 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1795                           *root, struct btrfs_path *path, int data_size,
1796                           int empty)
1797 {
1798         struct btrfs_disk_key disk_key;
1799         struct extent_buffer *right = path->nodes[0];
1800         struct extent_buffer *left;
1801         int slot;
1802         int i;
1803         int free_space;
1804         int push_space = 0;
1805         int push_items = 0;
1806         struct btrfs_item *item;
1807         u32 old_left_nritems;
1808         u32 right_nritems;
1809         u32 nr;
1810         int ret = 0;
1811         u32 this_item_size;
1812         u32 old_left_item_size;
1813
1814         slot = path->slots[1];
1815         if (slot == 0)
1816                 return 1;
1817         if (!path->nodes[1])
1818                 return 1;
1819
1820         right_nritems = btrfs_header_nritems(right);
1821         if (right_nritems == 0) {
1822                 return 1;
1823         }
1824
1825         left = read_node_slot(root, path->nodes[1], slot - 1);
1826         free_space = btrfs_leaf_free_space(root, left);
1827         if (free_space < data_size) {
1828                 free_extent_buffer(left);
1829                 return 1;
1830         }
1831
1832         /* cow and double check */
1833         ret = btrfs_cow_block(trans, root, left,
1834                               path->nodes[1], slot - 1, &left);
1835         if (ret) {
1836                 /* we hit -ENOSPC, but it isn't fatal here */
1837                 free_extent_buffer(left);
1838                 return 1;
1839         }
1840
1841         free_space = btrfs_leaf_free_space(root, left);
1842         if (free_space < data_size) {
1843                 free_extent_buffer(left);
1844                 return 1;
1845         }
1846
1847         if (empty)
1848                 nr = right_nritems;
1849         else
1850                 nr = right_nritems - 1;
1851
1852         for (i = 0; i < nr; i++) {
1853                 item = btrfs_item_nr(i);
1854
1855                 if (path->slots[0] == i)
1856                         push_space += data_size + sizeof(*item);
1857
1858                 this_item_size = btrfs_item_size(right, item);
1859                 if (this_item_size + sizeof(*item) + push_space > free_space)
1860                         break;
1861
1862                 push_items++;
1863                 push_space += this_item_size + sizeof(*item);
1864         }
1865
1866         if (push_items == 0) {
1867                 free_extent_buffer(left);
1868                 return 1;
1869         }
1870         if (!empty && push_items == btrfs_header_nritems(right))
1871                 WARN_ON(1);
1872
1873         /* push data from right to left */
1874         copy_extent_buffer(left, right,
1875                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1876                            btrfs_item_nr_offset(0),
1877                            push_items * sizeof(struct btrfs_item));
1878
1879         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1880                      btrfs_item_offset_nr(right, push_items -1);
1881
1882         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1883                      leaf_data_end(root, left) - push_space,
1884                      btrfs_leaf_data(right) +
1885                      btrfs_item_offset_nr(right, push_items - 1),
1886                      push_space);
1887         old_left_nritems = btrfs_header_nritems(left);
1888         BUG_ON(old_left_nritems == 0);
1889
1890         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1891         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1892                 u32 ioff;
1893
1894                 item = btrfs_item_nr(i);
1895                 ioff = btrfs_item_offset(left, item);
1896                 btrfs_set_item_offset(left, item,
1897                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1898         }
1899         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1900
1901         /* fixup right node */
1902         if (push_items > right_nritems) {
1903                 printk("push items %d nr %u\n", push_items, right_nritems);
1904                 WARN_ON(1);
1905         }
1906
1907         if (push_items < right_nritems) {
1908                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1909                                                   leaf_data_end(root, right);
1910                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
1911                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1912                                       btrfs_leaf_data(right) +
1913                                       leaf_data_end(root, right), push_space);
1914
1915                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1916                               btrfs_item_nr_offset(push_items),
1917                              (btrfs_header_nritems(right) - push_items) *
1918                              sizeof(struct btrfs_item));
1919         }
1920         right_nritems -= push_items;
1921         btrfs_set_header_nritems(right, right_nritems);
1922         push_space = BTRFS_LEAF_DATA_SIZE(root);
1923         for (i = 0; i < right_nritems; i++) {
1924                 item = btrfs_item_nr(i);
1925                 push_space = push_space - btrfs_item_size(right, item);
1926                 btrfs_set_item_offset(right, item, push_space);
1927         }
1928
1929         btrfs_mark_buffer_dirty(left);
1930         if (right_nritems)
1931                 btrfs_mark_buffer_dirty(right);
1932
1933         btrfs_item_key(right, &disk_key, 0);
1934         btrfs_fixup_low_keys(root, path, &disk_key, 1);
1935
1936         /* then fixup the leaf pointer in the path */
1937         if (path->slots[0] < push_items) {
1938                 path->slots[0] += old_left_nritems;
1939                 free_extent_buffer(path->nodes[0]);
1940                 path->nodes[0] = left;
1941                 path->slots[1] -= 1;
1942         } else {
1943                 free_extent_buffer(left);
1944                 path->slots[0] -= push_items;
1945         }
1946         BUG_ON(path->slots[0] < 0);
1947         return ret;
1948 }
1949
1950 /*
1951  * split the path's leaf in two, making sure there is at least data_size
1952  * available for the resulting leaf level of the path.
1953  *
1954  * returns 0 if all went well and < 0 on failure.
1955  */
1956 static noinline int copy_for_split(struct btrfs_trans_handle *trans,
1957                                struct btrfs_root *root,
1958                                struct btrfs_path *path,
1959                                struct extent_buffer *l,
1960                                struct extent_buffer *right,
1961                                int slot, int mid, int nritems)
1962 {
1963         int data_copy_size;
1964         int rt_data_off;
1965         int i;
1966         int ret = 0;
1967         int wret;
1968         struct btrfs_disk_key disk_key;
1969
1970         nritems = nritems - mid;
1971         btrfs_set_header_nritems(right, nritems);
1972         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
1973
1974         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
1975                            btrfs_item_nr_offset(mid),
1976                            nritems * sizeof(struct btrfs_item));
1977
1978         copy_extent_buffer(right, l,
1979                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
1980                      data_copy_size, btrfs_leaf_data(l) +
1981                      leaf_data_end(root, l), data_copy_size);
1982
1983         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
1984                       btrfs_item_end_nr(l, mid);
1985
1986         for (i = 0; i < nritems; i++) {
1987                 struct btrfs_item *item = btrfs_item_nr(i);
1988                 u32 ioff = btrfs_item_offset(right, item);
1989                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
1990         }
1991
1992         btrfs_set_header_nritems(l, mid);
1993         ret = 0;
1994         btrfs_item_key(right, &disk_key, 0);
1995         wret = insert_ptr(trans, root, path, &disk_key, right->start,
1996                           path->slots[1] + 1, 1);
1997         if (wret)
1998                 ret = wret;
1999
2000         btrfs_mark_buffer_dirty(right);
2001         btrfs_mark_buffer_dirty(l);
2002         BUG_ON(path->slots[0] != slot);
2003
2004         if (mid <= slot) {
2005                 free_extent_buffer(path->nodes[0]);
2006                 path->nodes[0] = right;
2007                 path->slots[0] -= mid;
2008                 path->slots[1] += 1;
2009         } else {
2010                 free_extent_buffer(right);
2011         }
2012
2013         BUG_ON(path->slots[0] < 0);
2014
2015         return ret;
2016 }
2017
2018 /*
2019  * split the path's leaf in two, making sure there is at least data_size
2020  * available for the resulting leaf level of the path.
2021  *
2022  * returns 0 if all went well and < 0 on failure.
2023  */
2024 static noinline int split_leaf(struct btrfs_trans_handle *trans,
2025                                struct btrfs_root *root,
2026                                struct btrfs_key *ins_key,
2027                                struct btrfs_path *path, int data_size,
2028                                int extend)
2029 {
2030         struct btrfs_disk_key disk_key;
2031         struct extent_buffer *l;
2032         u32 nritems;
2033         int mid;
2034         int slot;
2035         struct extent_buffer *right;
2036         int ret = 0;
2037         int wret;
2038         int split;
2039         int num_doubles = 0;
2040
2041         l = path->nodes[0];
2042         slot = path->slots[0];
2043         if (extend && data_size + btrfs_item_size_nr(l, slot) +
2044             sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(root))
2045                 return -EOVERFLOW;
2046
2047         /* first try to make some room by pushing left and right */
2048         if (data_size && ins_key->type != BTRFS_DIR_ITEM_KEY) {
2049                 wret = push_leaf_right(trans, root, path, data_size, 0);
2050                 if (wret < 0)
2051                         return wret;
2052                 if (wret) {
2053                         wret = push_leaf_left(trans, root, path, data_size, 0);
2054                         if (wret < 0)
2055                                 return wret;
2056                 }
2057                 l = path->nodes[0];
2058
2059                 /* did the pushes work? */
2060                 if (btrfs_leaf_free_space(root, l) >= data_size)
2061                         return 0;
2062         }
2063
2064         if (!path->nodes[1]) {
2065                 ret = insert_new_root(trans, root, path, 1);
2066                 if (ret)
2067                         return ret;
2068         }
2069 again:
2070         split = 1;
2071         l = path->nodes[0];
2072         slot = path->slots[0];
2073         nritems = btrfs_header_nritems(l);
2074         mid = (nritems + 1) / 2;
2075
2076         if (mid <= slot) {
2077                 if (nritems == 1 ||
2078                     leaf_space_used(l, mid, nritems - mid) + data_size >
2079                         BTRFS_LEAF_DATA_SIZE(root)) {
2080                         if (slot >= nritems) {
2081                                 split = 0;
2082                         } else {
2083                                 mid = slot;
2084                                 if (mid != nritems &&
2085                                     leaf_space_used(l, mid, nritems - mid) +
2086                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2087                                         split = 2;
2088                                 }
2089                         }
2090                 }
2091         } else {
2092                 if (leaf_space_used(l, 0, mid) + data_size >
2093                         BTRFS_LEAF_DATA_SIZE(root)) {
2094                         if (!extend && data_size && slot == 0) {
2095                                 split = 0;
2096                         } else if ((extend || !data_size) && slot == 0) {
2097                                 mid = 1;
2098                         } else {
2099                                 mid = slot;
2100                                 if (mid != nritems &&
2101                                     leaf_space_used(l, mid, nritems - mid) +
2102                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2103                                         split = 2 ;
2104                                 }
2105                         }
2106                 }
2107         }
2108
2109         if (split == 0)
2110                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2111         else
2112                 btrfs_item_key(l, &disk_key, mid);
2113
2114         right = btrfs_alloc_free_block(trans, root, root->leafsize,
2115                                         root->root_key.objectid,
2116                                         &disk_key, 0, l->start, 0);
2117         if (IS_ERR(right)) {
2118                 BUG_ON(1);
2119                 return PTR_ERR(right);
2120         }
2121
2122         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2123         btrfs_set_header_bytenr(right, right->start);
2124         btrfs_set_header_generation(right, trans->transid);
2125         btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
2126         btrfs_set_header_owner(right, root->root_key.objectid);
2127         btrfs_set_header_level(right, 0);
2128         write_extent_buffer(right, root->fs_info->fsid,
2129                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
2130
2131         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2132                             btrfs_header_chunk_tree_uuid(right),
2133                             BTRFS_UUID_SIZE);
2134
2135         if (split == 0) {
2136                 if (mid <= slot) {
2137                         btrfs_set_header_nritems(right, 0);
2138                         wret = insert_ptr(trans, root, path,
2139                                           &disk_key, right->start,
2140                                           path->slots[1] + 1, 1);
2141                         if (wret)
2142                                 ret = wret;
2143
2144                         free_extent_buffer(path->nodes[0]);
2145                         path->nodes[0] = right;
2146                         path->slots[0] = 0;
2147                         path->slots[1] += 1;
2148                 } else {
2149                         btrfs_set_header_nritems(right, 0);
2150                         wret = insert_ptr(trans, root, path,
2151                                           &disk_key,
2152                                           right->start,
2153                                           path->slots[1], 1);
2154                         if (wret)
2155                                 ret = wret;
2156                         free_extent_buffer(path->nodes[0]);
2157                         path->nodes[0] = right;
2158                         path->slots[0] = 0;
2159                         if (path->slots[1] == 0) {
2160                                 btrfs_fixup_low_keys(root, path,
2161                                                      &disk_key, 1);
2162                         }
2163                 }
2164                 btrfs_mark_buffer_dirty(right);
2165                 return ret;
2166         }
2167
2168         ret = copy_for_split(trans, root, path, l, right, slot, mid, nritems);
2169         BUG_ON(ret);
2170
2171         if (split == 2) {
2172                 BUG_ON(num_doubles != 0);
2173                 num_doubles++;
2174                 goto again;
2175         }
2176
2177         return ret;
2178 }
2179
2180 /*
2181  * This function splits a single item into two items,
2182  * giving 'new_key' to the new item and splitting the
2183  * old one at split_offset (from the start of the item).
2184  *
2185  * The path may be released by this operation.  After
2186  * the split, the path is pointing to the old item.  The
2187  * new item is going to be in the same node as the old one.
2188  *
2189  * Note, the item being split must be smaller enough to live alone on
2190  * a tree block with room for one extra struct btrfs_item
2191  *
2192  * This allows us to split the item in place, keeping a lock on the
2193  * leaf the entire time.
2194  */
2195 int btrfs_split_item(struct btrfs_trans_handle *trans,
2196                      struct btrfs_root *root,
2197                      struct btrfs_path *path,
2198                      struct btrfs_key *new_key,
2199                      unsigned long split_offset)
2200 {
2201         u32 item_size;
2202         struct extent_buffer *leaf;
2203         struct btrfs_key orig_key;
2204         struct btrfs_item *item;
2205         struct btrfs_item *new_item;
2206         int ret = 0;
2207         int slot;
2208         u32 nritems;
2209         u32 orig_offset;
2210         struct btrfs_disk_key disk_key;
2211         char *buf;
2212
2213         leaf = path->nodes[0];
2214         btrfs_item_key_to_cpu(leaf, &orig_key, path->slots[0]);
2215         if (btrfs_leaf_free_space(root, leaf) >= sizeof(struct btrfs_item))
2216                 goto split;
2217
2218         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2219         btrfs_release_path(path);
2220
2221         path->search_for_split = 1;
2222
2223         ret = btrfs_search_slot(trans, root, &orig_key, path, 0, 1);
2224         path->search_for_split = 0;
2225
2226         /* if our item isn't there or got smaller, return now */
2227         if (ret != 0 || item_size != btrfs_item_size_nr(path->nodes[0],
2228                                                         path->slots[0])) {
2229                 return -EAGAIN;
2230         }
2231
2232         ret = split_leaf(trans, root, &orig_key, path, 0, 0);
2233         BUG_ON(ret);
2234
2235         BUG_ON(btrfs_leaf_free_space(root, leaf) < sizeof(struct btrfs_item));
2236         leaf = path->nodes[0];
2237
2238 split:
2239         item = btrfs_item_nr(path->slots[0]);
2240         orig_offset = btrfs_item_offset(leaf, item);
2241         item_size = btrfs_item_size(leaf, item);
2242
2243
2244         buf = kmalloc(item_size, GFP_NOFS);
2245         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
2246                             path->slots[0]), item_size);
2247         slot = path->slots[0] + 1;
2248         leaf = path->nodes[0];
2249
2250         nritems = btrfs_header_nritems(leaf);
2251
2252         if (slot != nritems) {
2253                 /* shift the items */
2254                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
2255                               btrfs_item_nr_offset(slot),
2256                               (nritems - slot) * sizeof(struct btrfs_item));
2257
2258         }
2259
2260         btrfs_cpu_key_to_disk(&disk_key, new_key);
2261         btrfs_set_item_key(leaf, &disk_key, slot);
2262
2263         new_item = btrfs_item_nr(slot);
2264
2265         btrfs_set_item_offset(leaf, new_item, orig_offset);
2266         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
2267
2268         btrfs_set_item_offset(leaf, item,
2269                               orig_offset + item_size - split_offset);
2270         btrfs_set_item_size(leaf, item, split_offset);
2271
2272         btrfs_set_header_nritems(leaf, nritems + 1);
2273
2274         /* write the data for the start of the original item */
2275         write_extent_buffer(leaf, buf,
2276                             btrfs_item_ptr_offset(leaf, path->slots[0]),
2277                             split_offset);
2278
2279         /* write the data for the new item */
2280         write_extent_buffer(leaf, buf + split_offset,
2281                             btrfs_item_ptr_offset(leaf, slot),
2282                             item_size - split_offset);
2283         btrfs_mark_buffer_dirty(leaf);
2284
2285         ret = 0;
2286         if (btrfs_leaf_free_space(root, leaf) < 0) {
2287                 btrfs_print_leaf(root, leaf);
2288                 BUG();
2289         }
2290         kfree(buf);
2291         return ret;
2292 }
2293
2294 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2295                         struct btrfs_root *root,
2296                         struct btrfs_path *path,
2297                         u32 new_size, int from_end)
2298 {
2299         int ret = 0;
2300         int slot;
2301         struct extent_buffer *leaf;
2302         struct btrfs_item *item;
2303         u32 nritems;
2304         unsigned int data_end;
2305         unsigned int old_data_start;
2306         unsigned int old_size;
2307         unsigned int size_diff;
2308         int i;
2309
2310         leaf = path->nodes[0];
2311         slot = path->slots[0];
2312
2313         old_size = btrfs_item_size_nr(leaf, slot);
2314         if (old_size == new_size)
2315                 return 0;
2316
2317         nritems = btrfs_header_nritems(leaf);
2318         data_end = leaf_data_end(root, leaf);
2319
2320         old_data_start = btrfs_item_offset_nr(leaf, slot);
2321
2322         size_diff = old_size - new_size;
2323
2324         BUG_ON(slot < 0);
2325         BUG_ON(slot >= nritems);
2326
2327         /*
2328          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2329          */
2330         /* first correct the data pointers */
2331         for (i = slot; i < nritems; i++) {
2332                 u32 ioff;
2333                 item = btrfs_item_nr(i);
2334                 ioff = btrfs_item_offset(leaf, item);
2335                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2336         }
2337
2338         /* shift the data */
2339         if (from_end) {
2340                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2341                               data_end + size_diff, btrfs_leaf_data(leaf) +
2342                               data_end, old_data_start + new_size - data_end);
2343         } else {
2344                 struct btrfs_disk_key disk_key;
2345                 u64 offset;
2346
2347                 btrfs_item_key(leaf, &disk_key, slot);
2348
2349                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2350                         unsigned long ptr;
2351                         struct btrfs_file_extent_item *fi;
2352
2353                         fi = btrfs_item_ptr(leaf, slot,
2354                                             struct btrfs_file_extent_item);
2355                         fi = (struct btrfs_file_extent_item *)(
2356                              (unsigned long)fi - size_diff);
2357
2358                         if (btrfs_file_extent_type(leaf, fi) ==
2359                             BTRFS_FILE_EXTENT_INLINE) {
2360                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2361                                 memmove_extent_buffer(leaf, ptr,
2362                                         (unsigned long)fi,
2363                                         offsetof(struct btrfs_file_extent_item,
2364                                                  disk_bytenr));
2365                         }
2366                 }
2367
2368                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2369                               data_end + size_diff, btrfs_leaf_data(leaf) +
2370                               data_end, old_data_start - data_end);
2371
2372                 offset = btrfs_disk_key_offset(&disk_key);
2373                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2374                 btrfs_set_item_key(leaf, &disk_key, slot);
2375                 if (slot == 0)
2376                         btrfs_fixup_low_keys(root, path, &disk_key, 1);
2377         }
2378
2379         item = btrfs_item_nr(slot);
2380         btrfs_set_item_size(leaf, item, new_size);
2381         btrfs_mark_buffer_dirty(leaf);
2382
2383         ret = 0;
2384         if (btrfs_leaf_free_space(root, leaf) < 0) {
2385                 btrfs_print_leaf(root, leaf);
2386                 BUG();
2387         }
2388         return ret;
2389 }
2390
2391 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2392                       struct btrfs_root *root, struct btrfs_path *path,
2393                       u32 data_size)
2394 {
2395         int ret = 0;
2396         int slot;
2397         struct extent_buffer *leaf;
2398         struct btrfs_item *item;
2399         u32 nritems;
2400         unsigned int data_end;
2401         unsigned int old_data;
2402         unsigned int old_size;
2403         int i;
2404
2405         leaf = path->nodes[0];
2406
2407         nritems = btrfs_header_nritems(leaf);
2408         data_end = leaf_data_end(root, leaf);
2409
2410         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2411                 btrfs_print_leaf(root, leaf);
2412                 BUG();
2413         }
2414         slot = path->slots[0];
2415         old_data = btrfs_item_end_nr(leaf, slot);
2416
2417         BUG_ON(slot < 0);
2418         if (slot >= nritems) {
2419                 btrfs_print_leaf(root, leaf);
2420                 printk("slot %d too large, nritems %d\n", slot, nritems);
2421                 BUG_ON(1);
2422         }
2423
2424         /*
2425          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2426          */
2427         /* first correct the data pointers */
2428         for (i = slot; i < nritems; i++) {
2429                 u32 ioff;
2430                 item = btrfs_item_nr(i);
2431                 ioff = btrfs_item_offset(leaf, item);
2432                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2433         }
2434
2435         /* shift the data */
2436         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2437                       data_end - data_size, btrfs_leaf_data(leaf) +
2438                       data_end, old_data - data_end);
2439
2440         data_end = old_data;
2441         old_size = btrfs_item_size_nr(leaf, slot);
2442         item = btrfs_item_nr(slot);
2443         btrfs_set_item_size(leaf, item, old_size + data_size);
2444         btrfs_mark_buffer_dirty(leaf);
2445
2446         ret = 0;
2447         if (btrfs_leaf_free_space(root, leaf) < 0) {
2448                 btrfs_print_leaf(root, leaf);
2449                 BUG();
2450         }
2451         return ret;
2452 }
2453
2454 /*
2455  * Given a key and some data, insert an item into the tree.
2456  * This does all the path init required, making room in the tree if needed.
2457  */
2458 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2459                             struct btrfs_root *root,
2460                             struct btrfs_path *path,
2461                             struct btrfs_key *cpu_key, u32 *data_size,
2462                             int nr)
2463 {
2464         struct extent_buffer *leaf;
2465         struct btrfs_item *item;
2466         int ret = 0;
2467         int slot;
2468         int i;
2469         u32 nritems;
2470         u32 total_size = 0;
2471         u32 total_data = 0;
2472         unsigned int data_end;
2473         struct btrfs_disk_key disk_key;
2474
2475         for (i = 0; i < nr; i++) {
2476                 total_data += data_size[i];
2477         }
2478
2479         /* create a root if there isn't one */
2480         if (!root->node)
2481                 BUG();
2482
2483         total_size = total_data + nr * sizeof(struct btrfs_item);
2484         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2485         if (ret == 0) {
2486                 return -EEXIST;
2487         }
2488         if (ret < 0)
2489                 goto out;
2490
2491         leaf = path->nodes[0];
2492
2493         nritems = btrfs_header_nritems(leaf);
2494         data_end = leaf_data_end(root, leaf);
2495
2496         if (btrfs_leaf_free_space(root, leaf) < total_size) {
2497                 btrfs_print_leaf(root, leaf);
2498                 printk("not enough freespace need %u have %d\n",
2499                        total_size, btrfs_leaf_free_space(root, leaf));
2500                 BUG();
2501         }
2502
2503         slot = path->slots[0];
2504         BUG_ON(slot < 0);
2505
2506         if (slot != nritems) {
2507                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2508
2509                 if (old_data < data_end) {
2510                         btrfs_print_leaf(root, leaf);
2511                         printk("slot %d old_data %d data_end %d\n",
2512                                slot, old_data, data_end);
2513                         BUG_ON(1);
2514                 }
2515                 /*
2516                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2517                  */
2518                 /* first correct the data pointers */
2519                 for (i = slot; i < nritems; i++) {
2520                         u32 ioff;
2521
2522                         item = btrfs_item_nr(i);
2523                         ioff = btrfs_item_offset(leaf, item);
2524                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2525                 }
2526
2527                 /* shift the items */
2528                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2529                               btrfs_item_nr_offset(slot),
2530                               (nritems - slot) * sizeof(struct btrfs_item));
2531
2532                 /* shift the data */
2533                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2534                               data_end - total_data, btrfs_leaf_data(leaf) +
2535                               data_end, old_data - data_end);
2536                 data_end = old_data;
2537         }
2538
2539         /* setup the item for the new data */
2540         for (i = 0; i < nr; i++) {
2541                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2542                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2543                 item = btrfs_item_nr(slot + i);
2544                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2545                 data_end -= data_size[i];
2546                 btrfs_set_item_size(leaf, item, data_size[i]);
2547         }
2548         btrfs_set_header_nritems(leaf, nritems + nr);
2549         btrfs_mark_buffer_dirty(leaf);
2550
2551         ret = 0;
2552         if (slot == 0) {
2553                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2554                 btrfs_fixup_low_keys(root, path, &disk_key, 1);
2555         }
2556
2557         if (btrfs_leaf_free_space(root, leaf) < 0) {
2558                 btrfs_print_leaf(root, leaf);
2559                 BUG();
2560         }
2561
2562 out:
2563         return ret;
2564 }
2565
2566 /*
2567  * Given a key and some data, insert an item into the tree.
2568  * This does all the path init required, making room in the tree if needed.
2569  */
2570 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2571                       *root, struct btrfs_key *cpu_key, void *data, u32
2572                       data_size)
2573 {
2574         int ret = 0;
2575         struct btrfs_path *path;
2576         struct extent_buffer *leaf;
2577         unsigned long ptr;
2578
2579         path = btrfs_alloc_path();
2580         BUG_ON(!path);
2581         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2582         if (!ret) {
2583                 leaf = path->nodes[0];
2584                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2585                 write_extent_buffer(leaf, data, ptr, data_size);
2586                 btrfs_mark_buffer_dirty(leaf);
2587         }
2588         btrfs_free_path(path);
2589         return ret;
2590 }
2591
2592 /*
2593  * delete the pointer from a given node.
2594  *
2595  * If the delete empties a node, the node is removed from the tree,
2596  * continuing all the way the root if required.  The root is converted into
2597  * a leaf if all the nodes are emptied.
2598  */
2599 int btrfs_del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2600                    struct btrfs_path *path, int level, int slot)
2601 {
2602         struct extent_buffer *parent = path->nodes[level];
2603         u32 nritems;
2604         int ret = 0;
2605
2606         nritems = btrfs_header_nritems(parent);
2607         if (slot != nritems -1) {
2608                 memmove_extent_buffer(parent,
2609                               btrfs_node_key_ptr_offset(slot),
2610                               btrfs_node_key_ptr_offset(slot + 1),
2611                               sizeof(struct btrfs_key_ptr) *
2612                               (nritems - slot - 1));
2613         }
2614         nritems--;
2615         btrfs_set_header_nritems(parent, nritems);
2616         if (nritems == 0 && parent == root->node) {
2617                 BUG_ON(btrfs_header_level(root->node) != 1);
2618                 /* just turn the root into a leaf and break */
2619                 btrfs_set_header_level(root->node, 0);
2620         } else if (slot == 0) {
2621                 struct btrfs_disk_key disk_key;
2622
2623                 btrfs_node_key(parent, &disk_key, 0);
2624                 btrfs_fixup_low_keys(root, path, &disk_key, level + 1);
2625         }
2626         btrfs_mark_buffer_dirty(parent);
2627         return ret;
2628 }
2629
2630 /*
2631  * a helper function to delete the leaf pointed to by path->slots[1] and
2632  * path->nodes[1].
2633  *
2634  * This deletes the pointer in path->nodes[1] and frees the leaf
2635  * block extent.  zero is returned if it all worked out, < 0 otherwise.
2636  *
2637  * The path must have already been setup for deleting the leaf, including
2638  * all the proper balancing.  path->nodes[1] must be locked.
2639  */
2640 static noinline int btrfs_del_leaf(struct btrfs_trans_handle *trans,
2641                                    struct btrfs_root *root,
2642                                    struct btrfs_path *path,
2643                                    struct extent_buffer *leaf)
2644 {
2645         int ret;
2646
2647         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
2648         ret = btrfs_del_ptr(trans, root, path, 1, path->slots[1]);
2649         if (ret)
2650                 return ret;
2651
2652         ret = btrfs_free_extent(trans, root, leaf->start, leaf->len,
2653                                 0, root->root_key.objectid, 0, 0);
2654         return ret;
2655 }
2656
2657 /*
2658  * delete the item at the leaf level in path.  If that empties
2659  * the leaf, remove it from the tree
2660  */
2661 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2662                     struct btrfs_path *path, int slot, int nr)
2663 {
2664         struct extent_buffer *leaf;
2665         struct btrfs_item *item;
2666         int last_off;
2667         int dsize = 0;
2668         int ret = 0;
2669         int wret;
2670         int i;
2671         u32 nritems;
2672
2673         leaf = path->nodes[0];
2674         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2675
2676         for (i = 0; i < nr; i++)
2677                 dsize += btrfs_item_size_nr(leaf, slot + i);
2678
2679         nritems = btrfs_header_nritems(leaf);
2680
2681         if (slot + nr != nritems) {
2682                 int data_end = leaf_data_end(root, leaf);
2683
2684                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2685                               data_end + dsize,
2686                               btrfs_leaf_data(leaf) + data_end,
2687                               last_off - data_end);
2688
2689                 for (i = slot + nr; i < nritems; i++) {
2690                         u32 ioff;
2691
2692                         item = btrfs_item_nr(i);
2693                         ioff = btrfs_item_offset(leaf, item);
2694                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2695                 }
2696
2697                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2698                               btrfs_item_nr_offset(slot + nr),
2699                               sizeof(struct btrfs_item) *
2700                               (nritems - slot - nr));
2701         }
2702         btrfs_set_header_nritems(leaf, nritems - nr);
2703         nritems -= nr;
2704
2705         /* delete the leaf if we've emptied it */
2706         if (nritems == 0) {
2707                 if (leaf == root->node) {
2708                         btrfs_set_header_level(leaf, 0);
2709                 } else {
2710                         clean_tree_block(trans, root, leaf);
2711                         wait_on_tree_block_writeback(root, leaf);
2712
2713                         wret = btrfs_del_leaf(trans, root, path, leaf);
2714                         BUG_ON(ret);
2715                         if (wret)
2716                                 ret = wret;
2717                 }
2718         } else {
2719                 int used = leaf_space_used(leaf, 0, nritems);
2720                 if (slot == 0) {
2721                         struct btrfs_disk_key disk_key;
2722
2723                         btrfs_item_key(leaf, &disk_key, 0);
2724                         btrfs_fixup_low_keys(root, path, &disk_key, 1);
2725                 }
2726
2727                 /* delete the leaf if it is mostly empty */
2728                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2729                         /* push_leaf_left fixes the path.
2730                          * make sure the path still points to our leaf
2731                          * for possible call to del_ptr below
2732                          */
2733                         slot = path->slots[1];
2734                         extent_buffer_get(leaf);
2735
2736                         wret = push_leaf_left(trans, root, path, 1, 1);
2737                         if (wret < 0 && wret != -ENOSPC)
2738                                 ret = wret;
2739
2740                         if (path->nodes[0] == leaf &&
2741                             btrfs_header_nritems(leaf)) {
2742                                 wret = push_leaf_right(trans, root, path, 1, 1);
2743                                 if (wret < 0 && wret != -ENOSPC)
2744                                         ret = wret;
2745                         }
2746
2747                         if (btrfs_header_nritems(leaf) == 0) {
2748                                 clean_tree_block(trans, root, leaf);
2749                                 wait_on_tree_block_writeback(root, leaf);
2750
2751                                 path->slots[1] = slot;
2752                                 ret = btrfs_del_leaf(trans, root, path, leaf);
2753                                 BUG_ON(ret);
2754                                 free_extent_buffer(leaf);
2755
2756                         } else {
2757                                 btrfs_mark_buffer_dirty(leaf);
2758                                 free_extent_buffer(leaf);
2759                         }
2760                 } else {
2761                         btrfs_mark_buffer_dirty(leaf);
2762                 }
2763         }
2764         return ret;
2765 }
2766
2767 /*
2768  * walk up the tree as far as required to find the previous leaf.
2769  * returns 0 if it found something or 1 if there are no lesser leaves.
2770  * returns < 0 on io errors.
2771  */
2772 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2773 {
2774         int slot;
2775         int level = 1;
2776         struct extent_buffer *c;
2777         struct extent_buffer *next = NULL;
2778
2779         while(level < BTRFS_MAX_LEVEL) {
2780                 if (!path->nodes[level])
2781                         return 1;
2782
2783                 slot = path->slots[level];
2784                 c = path->nodes[level];
2785                 if (slot == 0) {
2786                         level++;
2787                         if (level == BTRFS_MAX_LEVEL)
2788                                 return 1;
2789                         continue;
2790                 }
2791                 slot--;
2792
2793                 next = read_node_slot(root, c, slot);
2794                 if (!extent_buffer_uptodate(next)) {
2795                         if (IS_ERR(next))
2796                                 return PTR_ERR(next);
2797                         return -EIO;
2798                 }
2799                 break;
2800         }
2801         path->slots[level] = slot;
2802         while(1) {
2803                 level--;
2804                 c = path->nodes[level];
2805                 free_extent_buffer(c);
2806                 slot = btrfs_header_nritems(next);
2807                 if (slot != 0)
2808                         slot--;
2809                 path->nodes[level] = next;
2810                 path->slots[level] = slot;
2811                 if (!level)
2812                         break;
2813                 next = read_node_slot(root, next, slot);
2814                 if (!extent_buffer_uptodate(next)) {
2815                         if (IS_ERR(next))
2816                                 return PTR_ERR(next);
2817                         return -EIO;
2818                 }
2819         }
2820         return 0;
2821 }
2822
2823 /*
2824  * walk up the tree as far as required to find the next leaf.
2825  * returns 0 if it found something or 1 if there are no greater leaves.
2826  * returns < 0 on io errors.
2827  */
2828 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2829 {
2830         int slot;
2831         int level = 1;
2832         struct extent_buffer *c;
2833         struct extent_buffer *next = NULL;
2834
2835         while(level < BTRFS_MAX_LEVEL) {
2836                 if (!path->nodes[level])
2837                         return 1;
2838
2839                 slot = path->slots[level] + 1;
2840                 c = path->nodes[level];
2841                 if (slot >= btrfs_header_nritems(c)) {
2842                         level++;
2843                         if (level == BTRFS_MAX_LEVEL)
2844                                 return 1;
2845                         continue;
2846                 }
2847
2848                 if (path->reada)
2849                         reada_for_search(root, path, level, slot, 0);
2850
2851                 next = read_node_slot(root, c, slot);
2852                 if (!extent_buffer_uptodate(next))
2853                         return -EIO;
2854                 break;
2855         }
2856         path->slots[level] = slot;
2857         while(1) {
2858                 level--;
2859                 c = path->nodes[level];
2860                 free_extent_buffer(c);
2861                 path->nodes[level] = next;
2862                 path->slots[level] = 0;
2863                 if (!level)
2864                         break;
2865                 if (path->reada)
2866                         reada_for_search(root, path, level, 0, 0);
2867                 next = read_node_slot(root, next, 0);
2868                 if (!extent_buffer_uptodate(next))
2869                         return -EIO;
2870         }
2871         return 0;
2872 }
2873
2874 int btrfs_previous_item(struct btrfs_root *root,
2875                         struct btrfs_path *path, u64 min_objectid,
2876                         int type)
2877 {
2878         struct btrfs_key found_key;
2879         struct extent_buffer *leaf;
2880         int ret;
2881
2882         while(1) {
2883                 if (path->slots[0] == 0) {
2884                         ret = btrfs_prev_leaf(root, path);
2885                         if (ret != 0)
2886                                 return ret;
2887                 } else {
2888                         path->slots[0]--;
2889                 }
2890                 leaf = path->nodes[0];
2891                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2892                 if (found_key.type == type)
2893                         return 0;
2894         }
2895         return 1;
2896 }
2897
2898 /*
2899  * search in extent tree to find a previous Metadata/Data extent item with
2900  * min objecitd.
2901  *
2902  * returns 0 if something is found, 1 if nothing was found and < 0 on error
2903  */
2904 int btrfs_previous_extent_item(struct btrfs_root *root,
2905                         struct btrfs_path *path, u64 min_objectid)
2906 {
2907         struct btrfs_key found_key;
2908         struct extent_buffer *leaf;
2909         u32 nritems;
2910         int ret;
2911
2912         while (1) {
2913                 if (path->slots[0] == 0) {
2914                         ret = btrfs_prev_leaf(root, path);
2915                         if (ret != 0)
2916                                 return ret;
2917                 } else {
2918                         path->slots[0]--;
2919                 }
2920                 leaf = path->nodes[0];
2921                 nritems = btrfs_header_nritems(leaf);
2922                 if (nritems == 0)
2923                         return 1;
2924                 if (path->slots[0] == nritems)
2925                         path->slots[0]--;
2926
2927                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2928                 if (found_key.objectid < min_objectid)
2929                         break;
2930                 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
2931                     found_key.type == BTRFS_METADATA_ITEM_KEY)
2932                         return 0;
2933                 if (found_key.objectid == min_objectid &&
2934                     found_key.type < BTRFS_EXTENT_ITEM_KEY)
2935                         break;
2936         }
2937         return 1;
2938 }