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 #include "internal.h"
  24 #include "sizes.h"
  25
  26 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  27                       *root, struct btrfs_path *path, int level);
  28 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  29                       *root, struct btrfs_key *ins_key,
  30                       struct btrfs_path *path, int data_size, int extend);
  31 static int push_node_left(struct btrfs_trans_handle *trans,
  32                           struct btrfs_root *root, struct extent_buffer *dst,
  33                           struct extent_buffer *src, int empty);
  34 static int balance_node_right(struct btrfs_trans_handle *trans,
  35                               struct btrfs_root *root,
  36                               struct extent_buffer *dst_buf,
  37                               struct extent_buffer *src_buf);
  38
  39 inline void btrfs_init_path(struct btrfs_path *p)
  40 {
  41         memset(p, 0, sizeof(*p));
  42 }
  43
  44 struct btrfs_path *btrfs_alloc_path(void)
  45 {
  46         struct btrfs_path *path;
  47         path = kzalloc(sizeof(struct btrfs_path), GFP_NOFS);
  48         return path;
  49 }
  50
  51 void btrfs_free_path(struct btrfs_path *p)
  52 {
  53         if (!p)
  54                 return;
  55         btrfs_release_path(p);
  56         kfree(p);
  57 }
  58
  59 void btrfs_release_path(struct btrfs_path *p)
  60 {
  61         int i;
  62         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  63                 if (!p->nodes[i])
  64                         continue;
  65                 free_extent_buffer(p->nodes[i]);
  66         }
  67         memset(p, 0, sizeof(*p));
  68 }
  69
  70 void add_root_to_dirty_list(struct btrfs_root *root)
  71 {
  72         if (root->track_dirty && list_empty(&root->dirty_list)) {
  73                 list_add(&root->dirty_list,
  74                          &root->fs_info->dirty_cowonly_roots);
  75         }
  76 }
  77
  78 int btrfs_copy_root(struct btrfs_trans_handle *trans,
  79                       struct btrfs_root *root,
  80                       struct extent_buffer *buf,
  81                       struct extent_buffer **cow_ret, u64 new_root_objectid)
  82 {
  83         struct extent_buffer *cow;
  84         int ret = 0;
  85         int level;
  86         struct btrfs_root *new_root;
  87         struct btrfs_disk_key disk_key;
  88
  89         new_root = kmalloc(sizeof(*new_root), GFP_NOFS);
  90         if (!new_root)
  91                 return -ENOMEM;
  92
  93         memcpy(new_root, root, sizeof(*new_root));
  94         new_root->root_key.objectid = new_root_objectid;
  95
  96         WARN_ON(root->ref_cows && trans->transid !=
  97                 root->fs_info->running_transaction->transid);
  98         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
  99
 100         level = btrfs_header_level(buf);
 101         if (level == 0)
 102                 btrfs_item_key(buf, &disk_key, 0);
 103         else
 104                 btrfs_node_key(buf, &disk_key, 0);
 105         cow = btrfs_alloc_free_block(trans, new_root, buf->len,
 106                                      new_root_objectid, &disk_key,
 107                                      level, buf->start, 0);
 108         if (IS_ERR(cow)) {
 109                 kfree(new_root);
 110                 return PTR_ERR(cow);
 111         }
 112
 113         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 114         btrfs_set_header_bytenr(cow, cow->start);
 115         btrfs_set_header_generation(cow, trans->transid);
 116         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 117         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 118                                      BTRFS_HEADER_FLAG_RELOC);
 119         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
 120                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 121         else
 122                 btrfs_set_header_owner(cow, new_root_objectid);
 123
 124         write_extent_buffer(cow, root->fs_info->fsid,
 125                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
 126
 127         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 128         ret = btrfs_inc_ref(trans, new_root, cow, 0);
 129         kfree(new_root);
 130
 131         if (ret)
 132                 return ret;
 133
 134         btrfs_mark_buffer_dirty(cow);
 135         *cow_ret = cow;
 136         return 0;
 137 }
 138
 139 /*
 140  * check if the tree block can be shared by multiple trees
 141  */
 142 static int btrfs_block_can_be_shared(struct btrfs_root *root,
 143                                      struct extent_buffer *buf)
 144 {
 145         /*
 146          * Tree blocks not in reference counted trees and tree roots
 147          * are never shared. If a block was allocated after the last
 148          * snapshot and the block was not allocated by tree relocation,
 149          * we know the block is not shared.
 150          */
 151         if (root->ref_cows &&
 152             buf != root->node && buf != root->commit_root &&
 153             (btrfs_header_generation(buf) <=
 154              btrfs_root_last_snapshot(&root->root_item) ||
 155              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 156                 return 1;
 157 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 158         if (root->ref_cows &&
 159             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 160                 return 1;
 161 #endif
 162         return 0;
 163 }
 164
 165 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
 166                                        struct btrfs_root *root,
 167                                        struct extent_buffer *buf,
 168                                        struct extent_buffer *cow)
 169 {
 170         u64 refs;
 171         u64 owner;
 172         u64 flags;
 173         u64 new_flags = 0;
 174         int ret;
 175
 176         /*
 177          * Backrefs update rules:
 178          *
 179          * Always use full backrefs for extent pointers in tree block
 180          * allocated by tree relocation.
 181          *
 182          * If a shared tree block is no longer referenced by its owner
 183          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
 184          * use full backrefs for extent pointers in tree block.
 185          *
 186          * If a tree block is been relocating
 187          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
 188          * use full backrefs for extent pointers in tree block.
 189          * The reason for this is some operations (such as drop tree)
 190          * are only allowed for blocks use full backrefs.
 191          */
 192
 193         if (btrfs_block_can_be_shared(root, buf)) {
 194                 ret = btrfs_lookup_extent_info(trans, root, buf->start,
 195                                                btrfs_header_level(buf), 1,
 196                                                &refs, &flags);
 197                 BUG_ON(ret);
 198                 BUG_ON(refs == 0);
 199         } else {
 200                 refs = 1;
 201                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 202                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 203                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 204                 else
 205                         flags = 0;
 206         }
 207
 208         owner = btrfs_header_owner(buf);
 209         BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) &&
 210                owner == BTRFS_TREE_RELOC_OBJECTID);
 211
 212         if (refs > 1) {
 213                 if ((owner == root->root_key.objectid ||
 214                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
 215                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
 216                         ret = btrfs_inc_ref(trans, root, buf, 1);
 217                         BUG_ON(ret);
 218
 219                         if (root->root_key.objectid ==
 220                             BTRFS_TREE_RELOC_OBJECTID) {
 221                                 ret = btrfs_dec_ref(trans, root, buf, 0);
 222                                 BUG_ON(ret);
 223                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 224                                 BUG_ON(ret);
 225                         }
 226                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
 227                 } else {
 228
 229                         if (root->root_key.objectid ==
 230                             BTRFS_TREE_RELOC_OBJECTID)
 231                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 232                         else
 233                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 234                         BUG_ON(ret);
 235                 }
 236                 if (new_flags != 0) {
 237                         ret = btrfs_set_block_flags(trans, root, buf->start,
 238                                                     btrfs_header_level(buf),
 239                                                     new_flags);
 240                         BUG_ON(ret);
 241                 }
 242         } else {
 243                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
 244                         if (root->root_key.objectid ==
 245                             BTRFS_TREE_RELOC_OBJECTID)
 246                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 247                         else
 248                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 249                         BUG_ON(ret);
 250                         ret = btrfs_dec_ref(trans, root, buf, 1);
 251                         BUG_ON(ret);
 252                 }
 253                 clean_tree_block(trans, root, buf);
 254         }
 255         return 0;
 256 }
 257
 258 int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 259                              struct btrfs_root *root,
 260                              struct extent_buffer *buf,
 261                              struct extent_buffer *parent, int parent_slot,
 262                              struct extent_buffer **cow_ret,
 263                              u64 search_start, u64 empty_size)
 264 {
 265         struct extent_buffer *cow;
 266         struct btrfs_disk_key disk_key;
 267         int level;
 268
 269         WARN_ON(root->ref_cows && trans->transid !=
 270                 root->fs_info->running_transaction->transid);
 271         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 272
 273         level = btrfs_header_level(buf);
 274
 275         if (level == 0)
 276                 btrfs_item_key(buf, &disk_key, 0);
 277         else
 278                 btrfs_node_key(buf, &disk_key, 0);
 279
 280         cow = btrfs_alloc_free_block(trans, root, buf->len,
 281                                      root->root_key.objectid, &disk_key,
 282                                      level, search_start, empty_size);
 283         if (IS_ERR(cow))
 284                 return PTR_ERR(cow);
 285
 286         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 287         btrfs_set_header_bytenr(cow, cow->start);
 288         btrfs_set_header_generation(cow, trans->transid);
 289         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 290         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 291                                      BTRFS_HEADER_FLAG_RELOC);
 292         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 293                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 294         else
 295                 btrfs_set_header_owner(cow, root->root_key.objectid);
 296
 297         write_extent_buffer(cow, root->fs_info->fsid,
 298                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
 299
 300         WARN_ON(!(buf->flags & EXTENT_BAD_TRANSID) &&
 301                 btrfs_header_generation(buf) > trans->transid);
 302
 303         update_ref_for_cow(trans, root, buf, cow);
 304
 305         if (buf == root->node) {
 306                 root->node = cow;
 307                 extent_buffer_get(cow);
 308
 309                 btrfs_free_extent(trans, root, buf->start, buf->len,
 310                                   0, root->root_key.objectid, level, 0);
 311                 free_extent_buffer(buf);
 312                 add_root_to_dirty_list(root);
 313         } else {
 314                 btrfs_set_node_blockptr(parent, parent_slot,
 315                                         cow->start);
 316                 WARN_ON(trans->transid == 0);
 317                 btrfs_set_node_ptr_generation(parent, parent_slot,
 318                                               trans->transid);
 319                 btrfs_mark_buffer_dirty(parent);
 320                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 321
 322                 btrfs_free_extent(trans, root, buf->start, buf->len,
 323                                   0, root->root_key.objectid, level, 1);
 324         }
 325         if (!list_empty(&buf->recow)) {
 326                 list_del_init(&buf->recow);
 327                 free_extent_buffer(buf);
 328         }
 329         free_extent_buffer(buf);
 330         btrfs_mark_buffer_dirty(cow);
 331         *cow_ret = cow;
 332         return 0;
 333 }
 334
 335 static inline int should_cow_block(struct btrfs_trans_handle *trans,
 336                                    struct btrfs_root *root,
 337                                    struct extent_buffer *buf)
 338 {
 339         if (btrfs_header_generation(buf) == trans->transid &&
 340             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
 341             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
 342               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 343                 return 0;
 344         return 1;
 345 }
 346
 347 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 348                     struct btrfs_root *root, struct extent_buffer *buf,
 349                     struct extent_buffer *parent, int parent_slot,
 350                     struct extent_buffer **cow_ret)
 351 {
 352         u64 search_start;
 353         int ret;
 354         /*
 355         if (trans->transaction != root->fs_info->running_transaction) {
 356                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 357                        root->fs_info->running_transaction->transid);
 358                 WARN_ON(1);
 359         }
 360         */
 361         if (trans->transid != root->fs_info->generation) {
 362                 printk(KERN_CRIT "trans %llu running %llu\n",
 363                         (unsigned long long)trans->transid,
 364                         (unsigned long long)root->fs_info->generation);
 365                 WARN_ON(1);
 366         }
 367         if (!should_cow_block(trans, root, buf)) {
 368                 *cow_ret = buf;
 369                 return 0;
 370         }
 371
 372         search_start = buf->start & ~((u64)SZ_1G - 1);
 373         ret = __btrfs_cow_block(trans, root, buf, parent,
 374                                  parent_slot, cow_ret, search_start, 0);
 375         return ret;
 376 }
 377
 378 int btrfs_comp_cpu_keys(struct btrfs_key *k1, struct btrfs_key *k2)
 379 {
 380         if (k1->objectid > k2->objectid)
 381                 return 1;
 382         if (k1->objectid < k2->objectid)
 383                 return -1;
 384         if (k1->type > k2->type)
 385                 return 1;
 386         if (k1->type < k2->type)
 387                 return -1;
 388         if (k1->offset > k2->offset)
 389                 return 1;
 390         if (k1->offset < k2->offset)
 391                 return -1;
 392         return 0;
 393 }
 394
 395 /*
 396  * compare two keys in a memcmp fashion
 397  */
 398 static int btrfs_comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 399 {
 400         struct btrfs_key k1;
 401
 402         btrfs_disk_key_to_cpu(&k1, disk);
 403         return btrfs_comp_cpu_keys(&k1, k2);
 404 }
 405
 406 /*
 407  * The leaf data grows from end-to-front in the node.
 408  * this returns the address of the start of the last item,
 409  * which is the stop of the leaf data stack
 410  */
 411 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 412                                          struct extent_buffer *leaf)
 413 {
 414         u32 nr = btrfs_header_nritems(leaf);
 415         if (nr == 0)
 416                 return BTRFS_LEAF_DATA_SIZE(root);
 417         return btrfs_item_offset_nr(leaf, nr - 1);
 418 }
 419
 420 enum btrfs_tree_block_status
 421 btrfs_check_node(struct btrfs_root *root, struct btrfs_disk_key *parent_key,
 422                  struct extent_buffer *buf)
 423 {
 424         int i;
 425         struct btrfs_key cpukey;
 426         struct btrfs_disk_key key;
 427         u32 nritems = btrfs_header_nritems(buf);
 428         enum btrfs_tree_block_status ret = BTRFS_TREE_BLOCK_INVALID_NRITEMS;
 429
 430         if (nritems == 0 || nritems > BTRFS_NODEPTRS_PER_BLOCK(root))
 431                 goto fail;
 432
 433         ret = BTRFS_TREE_BLOCK_INVALID_PARENT_KEY;
 434         if (parent_key && parent_key->type) {
 435                 btrfs_node_key(buf, &key, 0);
 436                 if (memcmp(parent_key, &key, sizeof(key)))
 437                         goto fail;
 438         }
 439         ret = BTRFS_TREE_BLOCK_BAD_KEY_ORDER;
 440         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 441                 btrfs_node_key(buf, &key, i);
 442                 btrfs_node_key_to_cpu(buf, &cpukey, i + 1);
 443                 if (btrfs_comp_keys(&key, &cpukey) >= 0)
 444                         goto fail;
 445         }
 446         return BTRFS_TREE_BLOCK_CLEAN;
 447 fail:
 448         if (btrfs_header_owner(buf) == BTRFS_EXTENT_TREE_OBJECTID) {
 449                 if (parent_key)
 450                         btrfs_disk_key_to_cpu(&cpukey, parent_key);
 451                 else
 452                         btrfs_node_key_to_cpu(buf, &cpukey, 0);
 453                 btrfs_add_corrupt_extent_record(root->fs_info, &cpukey,
 454                                                 buf->start, buf->len,
 455                                                 btrfs_header_level(buf));
 456         }
 457         return ret;
 458 }
 459
 460 enum btrfs_tree_block_status
 461 btrfs_check_leaf(struct btrfs_root *root, struct btrfs_disk_key *parent_key,
 462                  struct extent_buffer *buf)
 463 {
 464         int i;
 465         struct btrfs_key cpukey;
 466         struct btrfs_disk_key key;
 467         u32 nritems = btrfs_header_nritems(buf);
 468         enum btrfs_tree_block_status ret = BTRFS_TREE_BLOCK_INVALID_NRITEMS;
 469
 470         if (nritems * sizeof(struct btrfs_item) > buf->len)  {
 471                 fprintf(stderr, "invalid number of items %llu\n",
 472                         (unsigned long long)buf->start);
 473                 goto fail;
 474         }
 475
 476         if (btrfs_header_level(buf) != 0) {
 477                 ret = BTRFS_TREE_BLOCK_INVALID_LEVEL;
 478                 fprintf(stderr, "leaf is not a leaf %llu\n",
 479                        (unsigned long long)btrfs_header_bytenr(buf));
 480                 goto fail;
 481         }
 482         if (btrfs_leaf_free_space(root, buf) < 0) {
 483                 ret = BTRFS_TREE_BLOCK_INVALID_FREE_SPACE;
 484                 fprintf(stderr, "leaf free space incorrect %llu %d\n",
 485                         (unsigned long long)btrfs_header_bytenr(buf),
 486                         btrfs_leaf_free_space(root, buf));
 487                 goto fail;
 488         }
 489
 490         if (nritems == 0)
 491                 return BTRFS_TREE_BLOCK_CLEAN;
 492
 493         btrfs_item_key(buf, &key, 0);
 494         if (parent_key && parent_key->type &&
 495             memcmp(parent_key, &key, sizeof(key))) {
 496                 ret = BTRFS_TREE_BLOCK_INVALID_PARENT_KEY;
 497                 fprintf(stderr, "leaf parent key incorrect %llu\n",
 498                        (unsigned long long)btrfs_header_bytenr(buf));
 499                 goto fail;
 500         }
 501         for (i = 0; nritems > 1 && i < nritems - 1; i++) {
 502                 btrfs_item_key(buf, &key, i);
 503                 btrfs_item_key_to_cpu(buf, &cpukey, i + 1);
 504                 if (btrfs_comp_keys(&key, &cpukey) >= 0) {
 505                         ret = BTRFS_TREE_BLOCK_BAD_KEY_ORDER;
 506                         fprintf(stderr, "bad key ordering %d %d\n", i, i+1);
 507                         goto fail;
 508                 }
 509                 if (btrfs_item_offset_nr(buf, i) !=
 510                         btrfs_item_end_nr(buf, i + 1)) {
 511                         ret = BTRFS_TREE_BLOCK_INVALID_OFFSETS;
 512                         fprintf(stderr, "incorrect offsets %u %u\n",
 513                                 btrfs_item_offset_nr(buf, i),
 514                                 btrfs_item_end_nr(buf, i + 1));
 515                         goto fail;
 516                 }
 517                 if (i == 0 && btrfs_item_end_nr(buf, i) !=
 518                     BTRFS_LEAF_DATA_SIZE(root)) {
 519                         ret = BTRFS_TREE_BLOCK_INVALID_OFFSETS;
 520                         fprintf(stderr, "bad item end %u wanted %u\n",
 521                                 btrfs_item_end_nr(buf, i),
 522                                 (unsigned)BTRFS_LEAF_DATA_SIZE(root));
 523                         goto fail;
 524                 }
 525         }
 526
 527         for (i = 0; i < nritems; i++) {
 528                 if (btrfs_item_end_nr(buf, i) > BTRFS_LEAF_DATA_SIZE(root)) {
 529                         btrfs_item_key(buf, &key, 0);
 530                         btrfs_print_key(&key);
 531                         fflush(stdout);
 532                         ret = BTRFS_TREE_BLOCK_INVALID_OFFSETS;
 533                         fprintf(stderr, "slot end outside of leaf %llu > %llu\n",
 534                                 (unsigned long long)btrfs_item_end_nr(buf, i),
 535                                 (unsigned long long)BTRFS_LEAF_DATA_SIZE(root));
 536                         goto fail;
 537                 }
 538         }
 539
 540         return BTRFS_TREE_BLOCK_CLEAN;
 541 fail:
 542         if (btrfs_header_owner(buf) == BTRFS_EXTENT_TREE_OBJECTID) {
 543                 if (parent_key)
 544                         btrfs_disk_key_to_cpu(&cpukey, parent_key);
 545                 else
 546                         btrfs_item_key_to_cpu(buf, &cpukey, 0);
 547
 548                 btrfs_add_corrupt_extent_record(root->fs_info, &cpukey,
 549                                                 buf->start, buf->len, 0);
 550         }
 551         return ret;
 552 }
 553
 554 static int noinline check_block(struct btrfs_root *root,
 555                                 struct btrfs_path *path, int level)
 556 {
 557         struct btrfs_disk_key key;
 558         struct btrfs_disk_key *key_ptr = NULL;
 559         struct extent_buffer *parent;
 560         enum btrfs_tree_block_status ret;
 561
 562         if (path->skip_check_block)
 563                 return 0;
 564         if (path->nodes[level + 1]) {
 565                 parent = path->nodes[level + 1];
 566                 btrfs_node_key(parent, &key, path->slots[level + 1]);
 567                 key_ptr = &key;
 568         }
 569         if (level == 0)
 570                 ret =  btrfs_check_leaf(root, key_ptr, path->nodes[0]);
 571         else
 572                 ret = btrfs_check_node(root, key_ptr, path->nodes[level]);
 573         if (ret == BTRFS_TREE_BLOCK_CLEAN)
 574                 return 0;
 575         return -EIO;
 576 }
 577
 578 /*
 579  * search for key in the extent_buffer.  The items start at offset p,
 580  * and they are item_size apart.  There are 'max' items in p.
 581  *
 582  * the slot in the array is returned via slot, and it points to
 583  * the place where you would insert key if it is not found in
 584  * the array.
 585  *
 586  * slot may point to max if the key is bigger than all of the keys
 587  */
 588 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 589                               int item_size, struct btrfs_key *key,
 590                               int max, int *slot)
 591 {
 592         int low = 0;
 593         int high = max;
 594         int mid;
 595         int ret;
 596         unsigned long offset;
 597         struct btrfs_disk_key *tmp;
 598
 599         while(low < high) {
 600                 mid = (low + high) / 2;
 601                 offset = p + mid * item_size;
 602
 603                 tmp = (struct btrfs_disk_key *)(eb->data + offset);
 604                 ret = btrfs_comp_keys(tmp, key);
 605
 606                 if (ret < 0)
 607                         low = mid + 1;
 608                 else if (ret > 0)
 609                         high = mid;
 610                 else {
 611                         *slot = mid;
 612                         return 0;
 613                 }
 614         }
 615         *slot = low;
 616         return 1;
 617 }
 618
 619 /*
 620  * simple bin_search frontend that does the right thing for
 621  * leaves vs nodes
 622  */
 623 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 624                       int level, int *slot)
 625 {
 626         if (level == 0)
 627                 return generic_bin_search(eb,
 628                                           offsetof(struct btrfs_leaf, items),
 629                                           sizeof(struct btrfs_item),
 630                                           key, btrfs_header_nritems(eb),
 631                                           slot);
 632         else
 633                 return generic_bin_search(eb,
 634                                           offsetof(struct btrfs_node, ptrs),
 635                                           sizeof(struct btrfs_key_ptr),
 636                                           key, btrfs_header_nritems(eb),
 637                                           slot);
 638 }
 639
 640 struct extent_buffer *read_node_slot(struct btrfs_fs_info *fs_info,
 641                                    struct extent_buffer *parent, int slot)
 642 {
 643         int level = btrfs_header_level(parent);
 644         if (slot < 0)
 645                 return NULL;
 646         if (slot >= btrfs_header_nritems(parent))
 647                 return NULL;
 648
 649         if (level == 0)
 650                 return NULL;
 651
 652         return read_tree_block(fs_info, btrfs_node_blockptr(parent, slot),
 653                        fs_info->nodesize,
 654                        btrfs_node_ptr_generation(parent, slot));
 655 }
 656
 657 static int balance_level(struct btrfs_trans_handle *trans,
 658                          struct btrfs_root *root,
 659                          struct btrfs_path *path, int level)
 660 {
 661         struct extent_buffer *right = NULL;
 662         struct extent_buffer *mid;
 663         struct extent_buffer *left = NULL;
 664         struct extent_buffer *parent = NULL;
 665         struct btrfs_fs_info *fs_info = root->fs_info;
 666         int ret = 0;
 667         int wret;
 668         int pslot;
 669         int orig_slot = path->slots[level];
 670         u64 orig_ptr;
 671
 672         if (level == 0)
 673                 return 0;
 674
 675         mid = path->nodes[level];
 676         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 677
 678         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 679
 680         if (level < BTRFS_MAX_LEVEL - 1) {
 681                 parent = path->nodes[level + 1];
 682                 pslot = path->slots[level + 1];
 683         }
 684
 685         /*
 686          * deal with the case where there is only one pointer in the root
 687          * by promoting the node below to a root
 688          */
 689         if (!parent) {
 690                 struct extent_buffer *child;
 691
 692                 if (btrfs_header_nritems(mid) != 1)
 693                         return 0;
 694
 695                 /* promote the child to a root */
 696                 child = read_node_slot(fs_info, mid, 0);
 697                 BUG_ON(!extent_buffer_uptodate(child));
 698                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
 699                 BUG_ON(ret);
 700
 701                 root->node = child;
 702                 add_root_to_dirty_list(root);
 703                 path->nodes[level] = NULL;
 704                 clean_tree_block(trans, root, mid);
 705                 /* once for the path */
 706                 free_extent_buffer(mid);
 707
 708                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 709                                         0, root->root_key.objectid,
 710                                         level, 1);
 711                 /* once for the root ptr */
 712                 free_extent_buffer(mid);
 713                 return ret;
 714         }
 715         if (btrfs_header_nritems(mid) >
 716             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 717                 return 0;
 718
 719         left = read_node_slot(fs_info, parent, pslot - 1);
 720         if (extent_buffer_uptodate(left)) {
 721                 wret = btrfs_cow_block(trans, root, left,
 722                                        parent, pslot - 1, &left);
 723                 if (wret) {
 724                         ret = wret;
 725                         goto enospc;
 726                 }
 727         }
 728         right = read_node_slot(fs_info, parent, pslot + 1);
 729         if (extent_buffer_uptodate(right)) {
 730                 wret = btrfs_cow_block(trans, root, right,
 731                                        parent, pslot + 1, &right);
 732                 if (wret) {
 733                         ret = wret;
 734                         goto enospc;
 735                 }
 736         }
 737
 738         /* first, try to make some room in the middle buffer */
 739         if (left) {
 740                 orig_slot += btrfs_header_nritems(left);
 741                 wret = push_node_left(trans, root, left, mid, 1);
 742                 if (wret < 0)
 743                         ret = wret;
 744         }
 745
 746         /*
 747          * then try to empty the right most buffer into the middle
 748          */
 749         if (right) {
 750                 wret = push_node_left(trans, root, mid, right, 1);
 751                 if (wret < 0 && wret != -ENOSPC)
 752                         ret = wret;
 753                 if (btrfs_header_nritems(right) == 0) {
 754                         u64 bytenr = right->start;
 755                         u32 blocksize = right->len;
 756
 757                         clean_tree_block(trans, root, right);
 758                         free_extent_buffer(right);
 759                         right = NULL;
 760                         wret = btrfs_del_ptr(root, path, 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                 free_extent_buffer(mid);
 805                 mid = NULL;
 806                 wret = btrfs_del_ptr(root, path, level + 1, pslot);
 807                 if (wret)
 808                         ret = wret;
 809                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 810                                          0, root->root_key.objectid,
 811                                          level, 0);
 812                 if (wret)
 813                         ret = wret;
 814         } else {
 815                 /* update the parent key to reflect our changes */
 816                 struct btrfs_disk_key mid_key;
 817                 btrfs_node_key(mid, &mid_key, 0);
 818                 btrfs_set_node_key(parent, &mid_key, pslot);
 819                 btrfs_mark_buffer_dirty(parent);
 820         }
 821
 822         /* update the path */
 823         if (left) {
 824                 if (btrfs_header_nritems(left) > orig_slot) {
 825                         extent_buffer_get(left);
 826                         path->nodes[level] = left;
 827                         path->slots[level + 1] -= 1;
 828                         path->slots[level] = orig_slot;
 829                         if (mid)
 830                                 free_extent_buffer(mid);
 831                 } else {
 832                         orig_slot -= btrfs_header_nritems(left);
 833                         path->slots[level] = orig_slot;
 834                 }
 835         }
 836         /* double check we haven't messed things up */
 837         check_block(root, path, level);
 838         if (orig_ptr !=
 839             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 840                 BUG();
 841 enospc:
 842         if (right)
 843                 free_extent_buffer(right);
 844         if (left)
 845                 free_extent_buffer(left);
 846         return ret;
 847 }
 848
 849 /* returns zero if the push worked, non-zero otherwise */
 850 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
 851                                           struct btrfs_root *root,
 852                                           struct btrfs_path *path, int level)
 853 {
 854         struct extent_buffer *right = NULL;
 855         struct extent_buffer *mid;
 856         struct extent_buffer *left = NULL;
 857         struct extent_buffer *parent = NULL;
 858         struct btrfs_fs_info *fs_info = root->fs_info;
 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(fs_info, 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(fs_info, 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 = root->fs_info->nodesize;
 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 > SZ_256K || nscan > 32))
1029                         break;
1030                 if(nread > SZ_1M || 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         struct btrfs_fs_info *fs_info = root->fs_info;
1108         u8 lowest_level = 0;
1109
1110         lowest_level = p->lowest_level;
1111         WARN_ON(lowest_level && ins_len > 0);
1112         WARN_ON(p->nodes[0] != NULL);
1113         /*
1114         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
1115         */
1116 again:
1117         b = root->node;
1118         extent_buffer_get(b);
1119         while (b) {
1120                 level = btrfs_header_level(b);
1121                 if (cow) {
1122                         int wret;
1123                         wret = btrfs_cow_block(trans, root, b,
1124                                                p->nodes[level + 1],
1125                                                p->slots[level + 1],
1126                                                &b);
1127                         if (wret) {
1128                                 free_extent_buffer(b);
1129                                 return wret;
1130                         }
1131                 }
1132                 BUG_ON(!cow && ins_len);
1133                 if (level != btrfs_header_level(b))
1134                         WARN_ON(1);
1135                 level = btrfs_header_level(b);
1136                 p->nodes[level] = b;
1137                 ret = check_block(root, p, level);
1138                 if (ret)
1139                         return -1;
1140                 ret = bin_search(b, key, level, &slot);
1141                 if (level != 0) {
1142                         if (ret && slot > 0)
1143                                 slot -= 1;
1144                         p->slots[level] = slot;
1145                         if ((p->search_for_split || ins_len > 0) &&
1146                             btrfs_header_nritems(b) >=
1147                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1148                                 int sret = split_node(trans, root, p, level);
1149                                 BUG_ON(sret > 0);
1150                                 if (sret)
1151                                         return sret;
1152                                 b = p->nodes[level];
1153                                 slot = p->slots[level];
1154                         } else if (ins_len < 0) {
1155                                 int sret = balance_level(trans, root, p,
1156                                                          level);
1157                                 if (sret)
1158                                         return sret;
1159                                 b = p->nodes[level];
1160                                 if (!b) {
1161                                         btrfs_release_path(p);
1162                                         goto again;
1163                                 }
1164                                 slot = p->slots[level];
1165                                 BUG_ON(btrfs_header_nritems(b) == 1);
1166                         }
1167                         /* this is only true while dropping a snapshot */
1168                         if (level == lowest_level)
1169                                 break;
1170
1171                         if (should_reada)
1172                                 reada_for_search(root, p, level, slot,
1173                                                  key->objectid);
1174
1175                         b = read_node_slot(fs_info, b, slot);
1176                         if (!extent_buffer_uptodate(b))
1177                                 return -EIO;
1178                 } else {
1179                         p->slots[level] = slot;
1180                         if (ins_len > 0 &&
1181                             ins_len > btrfs_leaf_free_space(root, b)) {
1182                                 int sret = split_leaf(trans, root, key,
1183                                                       p, ins_len, ret == 0);
1184                                 BUG_ON(sret > 0);
1185                                 if (sret)
1186                                         return sret;
1187                         }
1188                         return ret;
1189                 }
1190         }
1191         return 1;
1192 }
1193
1194 /*
1195  * adjust the pointers going up the tree, starting at level
1196  * making sure the right key of each node is points to 'key'.
1197  * This is used after shifting pointers to the left, so it stops
1198  * fixing up pointers when a given leaf/node is not in slot 0 of the
1199  * higher levels
1200  */
1201 void btrfs_fixup_low_keys(struct btrfs_root *root, struct btrfs_path *path,
1202                           struct btrfs_disk_key *key, int level)
1203 {
1204         int i;
1205         struct extent_buffer *t;
1206
1207         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1208                 int tslot = path->slots[i];
1209                 if (!path->nodes[i])
1210                         break;
1211                 t = path->nodes[i];
1212                 btrfs_set_node_key(t, key, tslot);
1213                 btrfs_mark_buffer_dirty(path->nodes[i]);
1214                 if (tslot != 0)
1215                         break;
1216         }
1217 }
1218
1219 /*
1220  * update item key.
1221  *
1222  * This function isn't completely safe. It's the caller's responsibility
1223  * that the new key won't break the order
1224  */
1225 int btrfs_set_item_key_safe(struct btrfs_root *root, struct btrfs_path *path,
1226                             struct btrfs_key *new_key)
1227 {
1228         struct btrfs_disk_key disk_key;
1229         struct extent_buffer *eb;
1230         int slot;
1231
1232         eb = path->nodes[0];
1233         slot = path->slots[0];
1234         if (slot > 0) {
1235                 btrfs_item_key(eb, &disk_key, slot - 1);
1236                 if (btrfs_comp_keys(&disk_key, new_key) >= 0)
1237                         return -1;
1238         }
1239         if (slot < btrfs_header_nritems(eb) - 1) {
1240                 btrfs_item_key(eb, &disk_key, slot + 1);
1241                 if (btrfs_comp_keys(&disk_key, new_key) <= 0)
1242                         return -1;
1243         }
1244
1245         btrfs_cpu_key_to_disk(&disk_key, new_key);
1246         btrfs_set_item_key(eb, &disk_key, slot);
1247         btrfs_mark_buffer_dirty(eb);
1248         if (slot == 0)
1249                 btrfs_fixup_low_keys(root, path, &disk_key, 1);
1250         return 0;
1251 }
1252
1253 /*
1254  * update an item key without the safety checks.  This is meant to be called by
1255  * fsck only.
1256  */
1257 void btrfs_set_item_key_unsafe(struct btrfs_root *root,
1258                                struct btrfs_path *path,
1259                                struct btrfs_key *new_key)
1260 {
1261         struct btrfs_disk_key disk_key;
1262         struct extent_buffer *eb;
1263         int slot;
1264
1265         eb = path->nodes[0];
1266         slot = path->slots[0];
1267
1268         btrfs_cpu_key_to_disk(&disk_key, new_key);
1269         btrfs_set_item_key(eb, &disk_key, slot);
1270         btrfs_mark_buffer_dirty(eb);
1271         if (slot == 0)
1272                 btrfs_fixup_low_keys(root, path, &disk_key, 1);
1273 }
1274
1275 /*
1276  * try to push data from one node into the next node left in the
1277  * tree.
1278  *
1279  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1280  * error, and > 0 if there was no room in the left hand block.
1281  */
1282 static int push_node_left(struct btrfs_trans_handle *trans,
1283                           struct btrfs_root *root, struct extent_buffer *dst,
1284                           struct extent_buffer *src, int empty)
1285 {
1286         int push_items = 0;
1287         int src_nritems;
1288         int dst_nritems;
1289         int ret = 0;
1290
1291         src_nritems = btrfs_header_nritems(src);
1292         dst_nritems = btrfs_header_nritems(dst);
1293         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1294         WARN_ON(btrfs_header_generation(src) != trans->transid);
1295         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1296
1297         if (!empty && src_nritems <= 8)
1298                 return 1;
1299
1300         if (push_items <= 0) {
1301                 return 1;
1302         }
1303
1304         if (empty) {
1305                 push_items = min(src_nritems, push_items);
1306                 if (push_items < src_nritems) {
1307                         /* leave at least 8 pointers in the node if
1308                          * we aren't going to empty it
1309                          */
1310                         if (src_nritems - push_items < 8) {
1311                                 if (push_items <= 8)
1312                                         return 1;
1313                                 push_items -= 8;
1314                         }
1315                 }
1316         } else
1317                 push_items = min(src_nritems - 8, push_items);
1318
1319         copy_extent_buffer(dst, src,
1320                            btrfs_node_key_ptr_offset(dst_nritems),
1321                            btrfs_node_key_ptr_offset(0),
1322                            push_items * sizeof(struct btrfs_key_ptr));
1323
1324         if (push_items < src_nritems) {
1325                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1326                                       btrfs_node_key_ptr_offset(push_items),
1327                                       (src_nritems - push_items) *
1328                                       sizeof(struct btrfs_key_ptr));
1329         }
1330         btrfs_set_header_nritems(src, src_nritems - push_items);
1331         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1332         btrfs_mark_buffer_dirty(src);
1333         btrfs_mark_buffer_dirty(dst);
1334
1335         return ret;
1336 }
1337
1338 /*
1339  * try to push data from one node into the next node right in the
1340  * tree.
1341  *
1342  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1343  * error, and > 0 if there was no room in the right hand block.
1344  *
1345  * this will  only push up to 1/2 the contents of the left node over
1346  */
1347 static int balance_node_right(struct btrfs_trans_handle *trans,
1348                               struct btrfs_root *root,
1349                               struct extent_buffer *dst,
1350                               struct extent_buffer *src)
1351 {
1352         int push_items = 0;
1353         int max_push;
1354         int src_nritems;
1355         int dst_nritems;
1356         int ret = 0;
1357
1358         WARN_ON(btrfs_header_generation(src) != trans->transid);
1359         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1360
1361         src_nritems = btrfs_header_nritems(src);
1362         dst_nritems = btrfs_header_nritems(dst);
1363         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1364         if (push_items <= 0) {
1365                 return 1;
1366         }
1367
1368         if (src_nritems < 4) {
1369                 return 1;
1370         }
1371
1372         max_push = src_nritems / 2 + 1;
1373         /* don't try to empty the node */
1374         if (max_push >= src_nritems) {
1375                 return 1;
1376         }
1377
1378         if (max_push < push_items)
1379                 push_items = max_push;
1380
1381         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1382                                       btrfs_node_key_ptr_offset(0),
1383                                       (dst_nritems) *
1384                                       sizeof(struct btrfs_key_ptr));
1385
1386         copy_extent_buffer(dst, src,
1387                            btrfs_node_key_ptr_offset(0),
1388                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1389                            push_items * sizeof(struct btrfs_key_ptr));
1390
1391         btrfs_set_header_nritems(src, src_nritems - push_items);
1392         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1393
1394         btrfs_mark_buffer_dirty(src);
1395         btrfs_mark_buffer_dirty(dst);
1396
1397         return ret;
1398 }
1399
1400 /*
1401  * helper function to insert a new root level in the tree.
1402  * A new node is allocated, and a single item is inserted to
1403  * point to the existing root
1404  *
1405  * returns zero on success or < 0 on failure.
1406  */
1407 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1408                            struct btrfs_root *root,
1409                            struct btrfs_path *path, int level)
1410 {
1411         u64 lower_gen;
1412         struct extent_buffer *lower;
1413         struct extent_buffer *c;
1414         struct extent_buffer *old;
1415         struct btrfs_disk_key lower_key;
1416
1417         BUG_ON(path->nodes[level]);
1418         BUG_ON(path->nodes[level-1] != root->node);
1419
1420         lower = path->nodes[level-1];
1421         if (level == 1)
1422                 btrfs_item_key(lower, &lower_key, 0);
1423         else
1424                 btrfs_node_key(lower, &lower_key, 0);
1425
1426         c = btrfs_alloc_free_block(trans, root, root->fs_info->nodesize,
1427                                    root->root_key.objectid, &lower_key,
1428                                    level, root->node->start, 0);
1429
1430         if (IS_ERR(c))
1431                 return PTR_ERR(c);
1432
1433         memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
1434         btrfs_set_header_nritems(c, 1);
1435         btrfs_set_header_level(c, level);
1436         btrfs_set_header_bytenr(c, c->start);
1437         btrfs_set_header_generation(c, trans->transid);
1438         btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
1439         btrfs_set_header_owner(c, root->root_key.objectid);
1440
1441         write_extent_buffer(c, root->fs_info->fsid,
1442                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
1443
1444         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1445                             btrfs_header_chunk_tree_uuid(c),
1446                             BTRFS_UUID_SIZE);
1447
1448         btrfs_set_node_key(c, &lower_key, 0);
1449         btrfs_set_node_blockptr(c, 0, lower->start);
1450         lower_gen = btrfs_header_generation(lower);
1451         WARN_ON(lower_gen != trans->transid);
1452
1453         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1454
1455         btrfs_mark_buffer_dirty(c);
1456
1457         old = root->node;
1458         root->node = c;
1459
1460         /* the super has an extra ref to root->node */
1461         free_extent_buffer(old);
1462
1463         add_root_to_dirty_list(root);
1464         extent_buffer_get(c);
1465         path->nodes[level] = c;
1466         path->slots[level] = 0;
1467         return 0;
1468 }
1469
1470 /*
1471  * worker function to insert a single pointer in a node.
1472  * the node should have enough room for the pointer already
1473  *
1474  * slot and level indicate where you want the key to go, and
1475  * blocknr is the block the key points to.
1476  *
1477  * returns zero on success and < 0 on any error
1478  */
1479 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1480                       *root, struct btrfs_path *path, struct btrfs_disk_key
1481                       *key, u64 bytenr, int slot, int level)
1482 {
1483         struct extent_buffer *lower;
1484         int nritems;
1485
1486         BUG_ON(!path->nodes[level]);
1487         lower = path->nodes[level];
1488         nritems = btrfs_header_nritems(lower);
1489         if (slot > nritems)
1490                 BUG();
1491         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1492                 BUG();
1493         if (slot < nritems) {
1494                 /* shift the items */
1495                 memmove_extent_buffer(lower,
1496                               btrfs_node_key_ptr_offset(slot + 1),
1497                               btrfs_node_key_ptr_offset(slot),
1498                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1499         }
1500         btrfs_set_node_key(lower, key, slot);
1501         btrfs_set_node_blockptr(lower, slot, bytenr);
1502         WARN_ON(trans->transid == 0);
1503         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1504         btrfs_set_header_nritems(lower, nritems + 1);
1505         btrfs_mark_buffer_dirty(lower);
1506         return 0;
1507 }
1508
1509 /*
1510  * split the node at the specified level in path in two.
1511  * The path is corrected to point to the appropriate node after the split
1512  *
1513  * Before splitting this tries to make some room in the node by pushing
1514  * left and right, if either one works, it returns right away.
1515  *
1516  * returns 0 on success and < 0 on failure
1517  */
1518 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1519                       *root, struct btrfs_path *path, int level)
1520 {
1521         struct extent_buffer *c;
1522         struct extent_buffer *split;
1523         struct btrfs_disk_key disk_key;
1524         int mid;
1525         int ret;
1526         int wret;
1527         u32 c_nritems;
1528
1529         c = path->nodes[level];
1530         WARN_ON(btrfs_header_generation(c) != trans->transid);
1531         if (c == root->node) {
1532                 /* trying to split the root, lets make a new one */
1533                 ret = insert_new_root(trans, root, path, level + 1);
1534                 if (ret)
1535                         return ret;
1536         } else {
1537                 ret = push_nodes_for_insert(trans, root, path, level);
1538                 c = path->nodes[level];
1539                 if (!ret && btrfs_header_nritems(c) <
1540                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
1541                         return 0;
1542                 if (ret < 0)
1543                         return ret;
1544         }
1545
1546         c_nritems = btrfs_header_nritems(c);
1547         mid = (c_nritems + 1) / 2;
1548         btrfs_node_key(c, &disk_key, mid);
1549
1550         split = btrfs_alloc_free_block(trans, root, root->fs_info->nodesize,
1551                                         root->root_key.objectid,
1552                                         &disk_key, level, c->start, 0);
1553         if (IS_ERR(split))
1554                 return PTR_ERR(split);
1555
1556         memset_extent_buffer(split, 0, 0, sizeof(struct btrfs_header));
1557         btrfs_set_header_level(split, btrfs_header_level(c));
1558         btrfs_set_header_bytenr(split, split->start);
1559         btrfs_set_header_generation(split, trans->transid);
1560         btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
1561         btrfs_set_header_owner(split, root->root_key.objectid);
1562         write_extent_buffer(split, root->fs_info->fsid,
1563                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
1564         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1565                             btrfs_header_chunk_tree_uuid(split),
1566                             BTRFS_UUID_SIZE);
1567
1568
1569         copy_extent_buffer(split, c,
1570                            btrfs_node_key_ptr_offset(0),
1571                            btrfs_node_key_ptr_offset(mid),
1572                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1573         btrfs_set_header_nritems(split, c_nritems - mid);
1574         btrfs_set_header_nritems(c, mid);
1575         ret = 0;
1576
1577         btrfs_mark_buffer_dirty(c);
1578         btrfs_mark_buffer_dirty(split);
1579
1580         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1581                           path->slots[level + 1] + 1,
1582                           level + 1);
1583         if (wret)
1584                 ret = wret;
1585
1586         if (path->slots[level] >= mid) {
1587                 path->slots[level] -= mid;
1588                 free_extent_buffer(c);
1589                 path->nodes[level] = split;
1590                 path->slots[level + 1] += 1;
1591         } else {
1592                 free_extent_buffer(split);
1593         }
1594         return ret;
1595 }
1596
1597 /*
1598  * how many bytes are required to store the items in a leaf.  start
1599  * and nr indicate which items in the leaf to check.  This totals up the
1600  * space used both by the item structs and the item data
1601  */
1602 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1603 {
1604         int data_len;
1605         int nritems = btrfs_header_nritems(l);
1606         int end = min(nritems, start + nr) - 1;
1607
1608         if (!nr)
1609                 return 0;
1610         data_len = btrfs_item_end_nr(l, start);
1611         data_len = data_len - btrfs_item_offset_nr(l, end);
1612         data_len += sizeof(struct btrfs_item) * nr;
1613         WARN_ON(data_len < 0);
1614         return data_len;
1615 }
1616
1617 /*
1618  * The space between the end of the leaf items and
1619  * the start of the leaf data.  IOW, how much room
1620  * the leaf has left for both items and data
1621  */
1622 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1623 {
1624         u32 nodesize = (root ? BTRFS_LEAF_DATA_SIZE(root) : leaf->len);
1625         int nritems = btrfs_header_nritems(leaf);
1626         int ret;
1627         ret = nodesize - leaf_space_used(leaf, 0, nritems);
1628         if (ret < 0) {
1629                 printk("leaf free space ret %d, leaf data size %u, used %d nritems %d\n",
1630                        ret, nodesize, leaf_space_used(leaf, 0, nritems),
1631                        nritems);
1632         }
1633         return ret;
1634 }
1635
1636 /*
1637  * push some data in the path leaf to the right, trying to free up at
1638  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1639  *
1640  * returns 1 if the push failed because the other node didn't have enough
1641  * room, 0 if everything worked out and < 0 if there were major errors.
1642  */
1643 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1644                            *root, struct btrfs_path *path, int data_size,
1645                            int empty)
1646 {
1647         struct extent_buffer *left = path->nodes[0];
1648         struct extent_buffer *right;
1649         struct extent_buffer *upper;
1650         struct btrfs_disk_key disk_key;
1651         struct btrfs_fs_info *fs_info = root->fs_info;
1652         int slot;
1653         u32 i;
1654         int free_space;
1655         int push_space = 0;
1656         int push_items = 0;
1657         struct btrfs_item *item;
1658         u32 left_nritems;
1659         u32 nr;
1660         u32 right_nritems;
1661         u32 data_end;
1662         u32 this_item_size;
1663         int ret;
1664
1665         slot = path->slots[1];
1666         if (!path->nodes[1]) {
1667                 return 1;
1668         }
1669         upper = path->nodes[1];
1670         if (slot >= btrfs_header_nritems(upper) - 1)
1671                 return 1;
1672
1673         right = read_node_slot(fs_info, upper, slot + 1);
1674         if (!extent_buffer_uptodate(right)) {
1675                 if (IS_ERR(right))
1676                         return PTR_ERR(right);
1677                 return -EIO;
1678         }
1679         free_space = btrfs_leaf_free_space(root, right);
1680         if (free_space < data_size) {
1681                 free_extent_buffer(right);
1682                 return 1;
1683         }
1684
1685         /* cow and double check */
1686         ret = btrfs_cow_block(trans, root, right, upper,
1687                               slot + 1, &right);
1688         if (ret) {
1689                 free_extent_buffer(right);
1690                 return 1;
1691         }
1692         free_space = btrfs_leaf_free_space(root, right);
1693         if (free_space < data_size) {
1694                 free_extent_buffer(right);
1695                 return 1;
1696         }
1697
1698         left_nritems = btrfs_header_nritems(left);
1699         if (left_nritems == 0) {
1700                 free_extent_buffer(right);
1701                 return 1;
1702         }
1703
1704         if (empty)
1705                 nr = 0;
1706         else
1707                 nr = 1;
1708
1709         i = left_nritems - 1;
1710         while (i >= nr) {
1711                 item = btrfs_item_nr(i);
1712
1713                 if (path->slots[0] == i)
1714                         push_space += data_size + sizeof(*item);
1715
1716                 this_item_size = btrfs_item_size(left, item);
1717                 if (this_item_size + sizeof(*item) + push_space > free_space)
1718                         break;
1719                 push_items++;
1720                 push_space += this_item_size + sizeof(*item);
1721                 if (i == 0)
1722                         break;
1723                 i--;
1724         }
1725
1726         if (push_items == 0) {
1727                 free_extent_buffer(right);
1728                 return 1;
1729         }
1730
1731         if (!empty && push_items == left_nritems)
1732                 WARN_ON(1);
1733
1734         /* push left to right */
1735         right_nritems = btrfs_header_nritems(right);
1736
1737         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1738         push_space -= leaf_data_end(root, left);
1739
1740         /* make room in the right data area */
1741         data_end = leaf_data_end(root, right);
1742         memmove_extent_buffer(right,
1743                               btrfs_leaf_data(right) + data_end - push_space,
1744                               btrfs_leaf_data(right) + data_end,
1745                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1746
1747         /* copy from the left data area */
1748         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1749                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1750                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1751                      push_space);
1752
1753         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1754                               btrfs_item_nr_offset(0),
1755                               right_nritems * sizeof(struct btrfs_item));
1756
1757         /* copy the items from left to right */
1758         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1759                    btrfs_item_nr_offset(left_nritems - push_items),
1760                    push_items * sizeof(struct btrfs_item));
1761
1762         /* update the item pointers */
1763         right_nritems += push_items;
1764         btrfs_set_header_nritems(right, right_nritems);
1765         push_space = BTRFS_LEAF_DATA_SIZE(root);
1766         for (i = 0; i < right_nritems; i++) {
1767                 item = btrfs_item_nr(i);
1768                 push_space -= btrfs_item_size(right, item);
1769                 btrfs_set_item_offset(right, item, push_space);
1770         }
1771
1772         left_nritems -= push_items;
1773         btrfs_set_header_nritems(left, left_nritems);
1774
1775         if (left_nritems)
1776                 btrfs_mark_buffer_dirty(left);
1777         btrfs_mark_buffer_dirty(right);
1778
1779         btrfs_item_key(right, &disk_key, 0);
1780         btrfs_set_node_key(upper, &disk_key, slot + 1);
1781         btrfs_mark_buffer_dirty(upper);
1782
1783         /* then fixup the leaf pointer in the path */
1784         if (path->slots[0] >= left_nritems) {
1785                 path->slots[0] -= left_nritems;
1786                 free_extent_buffer(path->nodes[0]);
1787                 path->nodes[0] = right;
1788                 path->slots[1] += 1;
1789         } else {
1790                 free_extent_buffer(right);
1791         }
1792         return 0;
1793 }
1794 /*
1795  * push some data in the path leaf to the left, trying to free up at
1796  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1797  */
1798 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1799                           *root, struct btrfs_path *path, int data_size,
1800                           int empty)
1801 {
1802         struct btrfs_disk_key disk_key;
1803         struct extent_buffer *right = path->nodes[0];
1804         struct extent_buffer *left;
1805         struct btrfs_fs_info *fs_info = root->fs_info;
1806         int slot;
1807         int i;
1808         int free_space;
1809         int push_space = 0;
1810         int push_items = 0;
1811         struct btrfs_item *item;
1812         u32 old_left_nritems;
1813         u32 right_nritems;
1814         u32 nr;
1815         int ret = 0;
1816         u32 this_item_size;
1817         u32 old_left_item_size;
1818
1819         slot = path->slots[1];
1820         if (slot == 0)
1821                 return 1;
1822         if (!path->nodes[1])
1823                 return 1;
1824
1825         right_nritems = btrfs_header_nritems(right);
1826         if (right_nritems == 0) {
1827                 return 1;
1828         }
1829
1830         left = read_node_slot(fs_info, path->nodes[1], slot - 1);
1831         free_space = btrfs_leaf_free_space(root, left);
1832         if (free_space < data_size) {
1833                 free_extent_buffer(left);
1834                 return 1;
1835         }
1836
1837         /* cow and double check */
1838         ret = btrfs_cow_block(trans, root, left,
1839                               path->nodes[1], slot - 1, &left);
1840         if (ret) {
1841                 /* we hit -ENOSPC, but it isn't fatal here */
1842                 free_extent_buffer(left);
1843                 return 1;
1844         }
1845
1846         free_space = btrfs_leaf_free_space(root, left);
1847         if (free_space < data_size) {
1848                 free_extent_buffer(left);
1849                 return 1;
1850         }
1851
1852         if (empty)
1853                 nr = right_nritems;
1854         else
1855                 nr = right_nritems - 1;
1856
1857         for (i = 0; i < nr; i++) {
1858                 item = btrfs_item_nr(i);
1859
1860                 if (path->slots[0] == i)
1861                         push_space += data_size + sizeof(*item);
1862
1863                 this_item_size = btrfs_item_size(right, item);
1864                 if (this_item_size + sizeof(*item) + push_space > free_space)
1865                         break;
1866
1867                 push_items++;
1868                 push_space += this_item_size + sizeof(*item);
1869         }
1870
1871         if (push_items == 0) {
1872                 free_extent_buffer(left);
1873                 return 1;
1874         }
1875         if (!empty && push_items == btrfs_header_nritems(right))
1876                 WARN_ON(1);
1877
1878         /* push data from right to left */
1879         copy_extent_buffer(left, right,
1880                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1881                            btrfs_item_nr_offset(0),
1882                            push_items * sizeof(struct btrfs_item));
1883
1884         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1885                      btrfs_item_offset_nr(right, push_items -1);
1886
1887         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1888                      leaf_data_end(root, left) - push_space,
1889                      btrfs_leaf_data(right) +
1890                      btrfs_item_offset_nr(right, push_items - 1),
1891                      push_space);
1892         old_left_nritems = btrfs_header_nritems(left);
1893         BUG_ON(old_left_nritems == 0);
1894
1895         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1896         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1897                 u32 ioff;
1898
1899                 item = btrfs_item_nr(i);
1900                 ioff = btrfs_item_offset(left, item);
1901                 btrfs_set_item_offset(left, item,
1902                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1903         }
1904         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1905
1906         /* fixup right node */
1907         if (push_items > right_nritems) {
1908                 printk("push items %d nr %u\n", push_items, right_nritems);
1909                 WARN_ON(1);
1910         }
1911
1912         if (push_items < right_nritems) {
1913                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1914                                                   leaf_data_end(root, right);
1915                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
1916                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1917                                       btrfs_leaf_data(right) +
1918                                       leaf_data_end(root, right), push_space);
1919
1920                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1921                               btrfs_item_nr_offset(push_items),
1922                              (btrfs_header_nritems(right) - push_items) *
1923                              sizeof(struct btrfs_item));
1924         }
1925         right_nritems -= push_items;
1926         btrfs_set_header_nritems(right, right_nritems);
1927         push_space = BTRFS_LEAF_DATA_SIZE(root);
1928         for (i = 0; i < right_nritems; i++) {
1929                 item = btrfs_item_nr(i);
1930                 push_space = push_space - btrfs_item_size(right, item);
1931                 btrfs_set_item_offset(right, item, push_space);
1932         }
1933
1934         btrfs_mark_buffer_dirty(left);
1935         if (right_nritems)
1936                 btrfs_mark_buffer_dirty(right);
1937
1938         btrfs_item_key(right, &disk_key, 0);
1939         btrfs_fixup_low_keys(root, path, &disk_key, 1);
1940
1941         /* then fixup the leaf pointer in the path */
1942         if (path->slots[0] < push_items) {
1943                 path->slots[0] += old_left_nritems;
1944                 free_extent_buffer(path->nodes[0]);
1945                 path->nodes[0] = left;
1946                 path->slots[1] -= 1;
1947         } else {
1948                 free_extent_buffer(left);
1949                 path->slots[0] -= push_items;
1950         }
1951         BUG_ON(path->slots[0] < 0);
1952         return ret;
1953 }
1954
1955 /*
1956  * split the path's leaf in two, making sure there is at least data_size
1957  * available for the resulting leaf level of the path.
1958  *
1959  * returns 0 if all went well and < 0 on failure.
1960  */
1961 static noinline int copy_for_split(struct btrfs_trans_handle *trans,
1962                                struct btrfs_root *root,
1963                                struct btrfs_path *path,
1964                                struct extent_buffer *l,
1965                                struct extent_buffer *right,
1966                                int slot, int mid, int nritems)
1967 {
1968         int data_copy_size;
1969         int rt_data_off;
1970         int i;
1971         int ret = 0;
1972         int wret;
1973         struct btrfs_disk_key disk_key;
1974
1975         nritems = nritems - mid;
1976         btrfs_set_header_nritems(right, nritems);
1977         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
1978
1979         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
1980                            btrfs_item_nr_offset(mid),
1981                            nritems * sizeof(struct btrfs_item));
1982
1983         copy_extent_buffer(right, l,
1984                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
1985                      data_copy_size, btrfs_leaf_data(l) +
1986                      leaf_data_end(root, l), data_copy_size);
1987
1988         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
1989                       btrfs_item_end_nr(l, mid);
1990
1991         for (i = 0; i < nritems; i++) {
1992                 struct btrfs_item *item = btrfs_item_nr(i);
1993                 u32 ioff = btrfs_item_offset(right, item);
1994                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
1995         }
1996
1997         btrfs_set_header_nritems(l, mid);
1998         ret = 0;
1999         btrfs_item_key(right, &disk_key, 0);
2000         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2001                           path->slots[1] + 1, 1);
2002         if (wret)
2003                 ret = wret;
2004
2005         btrfs_mark_buffer_dirty(right);
2006         btrfs_mark_buffer_dirty(l);
2007         BUG_ON(path->slots[0] != slot);
2008
2009         if (mid <= slot) {
2010                 free_extent_buffer(path->nodes[0]);
2011                 path->nodes[0] = right;
2012                 path->slots[0] -= mid;
2013                 path->slots[1] += 1;
2014         } else {
2015                 free_extent_buffer(right);
2016         }
2017
2018         BUG_ON(path->slots[0] < 0);
2019
2020         return ret;
2021 }
2022
2023 /*
2024  * split the path's leaf in two, making sure there is at least data_size
2025  * available for the resulting leaf level of the path.
2026  *
2027  * returns 0 if all went well and < 0 on failure.
2028  */
2029 static noinline int split_leaf(struct btrfs_trans_handle *trans,
2030                                struct btrfs_root *root,
2031                                struct btrfs_key *ins_key,
2032                                struct btrfs_path *path, int data_size,
2033                                int extend)
2034 {
2035         struct btrfs_disk_key disk_key;
2036         struct extent_buffer *l;
2037         u32 nritems;
2038         int mid;
2039         int slot;
2040         struct extent_buffer *right;
2041         int ret = 0;
2042         int wret;
2043         int split;
2044         int num_doubles = 0;
2045
2046         l = path->nodes[0];
2047         slot = path->slots[0];
2048         if (extend && data_size + btrfs_item_size_nr(l, slot) +
2049             sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(root))
2050                 return -EOVERFLOW;
2051
2052         /* first try to make some room by pushing left and right */
2053         if (data_size && ins_key->type != BTRFS_DIR_ITEM_KEY) {
2054                 wret = push_leaf_right(trans, root, path, data_size, 0);
2055                 if (wret < 0)
2056                         return wret;
2057                 if (wret) {
2058                         wret = push_leaf_left(trans, root, path, data_size, 0);
2059                         if (wret < 0)
2060                                 return wret;
2061                 }
2062                 l = path->nodes[0];
2063
2064                 /* did the pushes work? */
2065                 if (btrfs_leaf_free_space(root, l) >= data_size)
2066                         return 0;
2067         }
2068
2069         if (!path->nodes[1]) {
2070                 ret = insert_new_root(trans, root, path, 1);
2071                 if (ret)
2072                         return ret;
2073         }
2074 again:
2075         split = 1;
2076         l = path->nodes[0];
2077         slot = path->slots[0];
2078         nritems = btrfs_header_nritems(l);
2079         mid = (nritems + 1) / 2;
2080
2081         if (mid <= slot) {
2082                 if (nritems == 1 ||
2083                     leaf_space_used(l, mid, nritems - mid) + data_size >
2084                         BTRFS_LEAF_DATA_SIZE(root)) {
2085                         if (slot >= nritems) {
2086                                 split = 0;
2087                         } else {
2088                                 mid = slot;
2089                                 if (mid != nritems &&
2090                                     leaf_space_used(l, mid, nritems - mid) +
2091                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2092                                         split = 2;
2093                                 }
2094                         }
2095                 }
2096         } else {
2097                 if (leaf_space_used(l, 0, mid) + data_size >
2098                         BTRFS_LEAF_DATA_SIZE(root)) {
2099                         if (!extend && data_size && slot == 0) {
2100                                 split = 0;
2101                         } else if ((extend || !data_size) && slot == 0) {
2102                                 mid = 1;
2103                         } else {
2104                                 mid = slot;
2105                                 if (mid != nritems &&
2106                                     leaf_space_used(l, mid, nritems - mid) +
2107                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2108                                         split = 2 ;
2109                                 }
2110                         }
2111                 }
2112         }
2113
2114         if (split == 0)
2115                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2116         else
2117                 btrfs_item_key(l, &disk_key, mid);
2118
2119         right = btrfs_alloc_free_block(trans, root, root->fs_info->nodesize,
2120                                         root->root_key.objectid,
2121                                         &disk_key, 0, l->start, 0);
2122         if (IS_ERR(right)) {
2123                 BUG_ON(1);
2124                 return PTR_ERR(right);
2125         }
2126
2127         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2128         btrfs_set_header_bytenr(right, right->start);
2129         btrfs_set_header_generation(right, trans->transid);
2130         btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
2131         btrfs_set_header_owner(right, root->root_key.objectid);
2132         btrfs_set_header_level(right, 0);
2133         write_extent_buffer(right, root->fs_info->fsid,
2134                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
2135
2136         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2137                             btrfs_header_chunk_tree_uuid(right),
2138                             BTRFS_UUID_SIZE);
2139
2140         if (split == 0) {
2141                 if (mid <= slot) {
2142                         btrfs_set_header_nritems(right, 0);
2143                         wret = insert_ptr(trans, root, path,
2144                                           &disk_key, right->start,
2145                                           path->slots[1] + 1, 1);
2146                         if (wret)
2147                                 ret = wret;
2148
2149                         free_extent_buffer(path->nodes[0]);
2150                         path->nodes[0] = right;
2151                         path->slots[0] = 0;
2152                         path->slots[1] += 1;
2153                 } else {
2154                         btrfs_set_header_nritems(right, 0);
2155                         wret = insert_ptr(trans, root, path,
2156                                           &disk_key,
2157                                           right->start,
2158                                           path->slots[1], 1);
2159                         if (wret)
2160                                 ret = wret;
2161                         free_extent_buffer(path->nodes[0]);
2162                         path->nodes[0] = right;
2163                         path->slots[0] = 0;
2164                         if (path->slots[1] == 0) {
2165                                 btrfs_fixup_low_keys(root, path,
2166                                                      &disk_key, 1);
2167                         }
2168                 }
2169                 btrfs_mark_buffer_dirty(right);
2170                 return ret;
2171         }
2172
2173         ret = copy_for_split(trans, root, path, l, right, slot, mid, nritems);
2174         BUG_ON(ret);
2175
2176         if (split == 2) {
2177                 BUG_ON(num_doubles != 0);
2178                 num_doubles++;
2179                 goto again;
2180         }
2181
2182         return ret;
2183 }
2184
2185 /*
2186  * This function splits a single item into two items,
2187  * giving 'new_key' to the new item and splitting the
2188  * old one at split_offset (from the start of the item).
2189  *
2190  * The path may be released by this operation.  After
2191  * the split, the path is pointing to the old item.  The
2192  * new item is going to be in the same node as the old one.
2193  *
2194  * Note, the item being split must be smaller enough to live alone on
2195  * a tree block with room for one extra struct btrfs_item
2196  *
2197  * This allows us to split the item in place, keeping a lock on the
2198  * leaf the entire time.
2199  */
2200 int btrfs_split_item(struct btrfs_trans_handle *trans,
2201                      struct btrfs_root *root,
2202                      struct btrfs_path *path,
2203                      struct btrfs_key *new_key,
2204                      unsigned long split_offset)
2205 {
2206         u32 item_size;
2207         struct extent_buffer *leaf;
2208         struct btrfs_key orig_key;
2209         struct btrfs_item *item;
2210         struct btrfs_item *new_item;
2211         int ret = 0;
2212         int slot;
2213         u32 nritems;
2214         u32 orig_offset;
2215         struct btrfs_disk_key disk_key;
2216         char *buf;
2217
2218         leaf = path->nodes[0];
2219         btrfs_item_key_to_cpu(leaf, &orig_key, path->slots[0]);
2220         if (btrfs_leaf_free_space(root, leaf) >= sizeof(struct btrfs_item))
2221                 goto split;
2222
2223         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2224         btrfs_release_path(path);
2225
2226         path->search_for_split = 1;
2227
2228         ret = btrfs_search_slot(trans, root, &orig_key, path, 0, 1);
2229         path->search_for_split = 0;
2230
2231         /* if our item isn't there or got smaller, return now */
2232         if (ret != 0 || item_size != btrfs_item_size_nr(path->nodes[0],
2233                                                         path->slots[0])) {
2234                 return -EAGAIN;
2235         }
2236
2237         ret = split_leaf(trans, root, &orig_key, path, 0, 0);
2238         BUG_ON(ret);
2239
2240         BUG_ON(btrfs_leaf_free_space(root, leaf) < sizeof(struct btrfs_item));
2241         leaf = path->nodes[0];
2242
2243 split:
2244         item = btrfs_item_nr(path->slots[0]);
2245         orig_offset = btrfs_item_offset(leaf, item);
2246         item_size = btrfs_item_size(leaf, item);
2247
2248
2249         buf = kmalloc(item_size, GFP_NOFS);
2250         BUG_ON(!buf);
2251         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
2252                             path->slots[0]), item_size);
2253         slot = path->slots[0] + 1;
2254         leaf = path->nodes[0];
2255
2256         nritems = btrfs_header_nritems(leaf);
2257
2258         if (slot < nritems) {
2259                 /* shift the items */
2260                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
2261                               btrfs_item_nr_offset(slot),
2262                               (nritems - slot) * sizeof(struct btrfs_item));
2263
2264         }
2265
2266         btrfs_cpu_key_to_disk(&disk_key, new_key);
2267         btrfs_set_item_key(leaf, &disk_key, slot);
2268
2269         new_item = btrfs_item_nr(slot);
2270
2271         btrfs_set_item_offset(leaf, new_item, orig_offset);
2272         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
2273
2274         btrfs_set_item_offset(leaf, item,
2275                               orig_offset + item_size - split_offset);
2276         btrfs_set_item_size(leaf, item, split_offset);
2277
2278         btrfs_set_header_nritems(leaf, nritems + 1);
2279
2280         /* write the data for the start of the original item */
2281         write_extent_buffer(leaf, buf,
2282                             btrfs_item_ptr_offset(leaf, path->slots[0]),
2283                             split_offset);
2284
2285         /* write the data for the new item */
2286         write_extent_buffer(leaf, buf + split_offset,
2287                             btrfs_item_ptr_offset(leaf, slot),
2288                             item_size - split_offset);
2289         btrfs_mark_buffer_dirty(leaf);
2290
2291         ret = 0;
2292         if (btrfs_leaf_free_space(root, leaf) < 0) {
2293                 btrfs_print_leaf(root, leaf);
2294                 BUG();
2295         }
2296         kfree(buf);
2297         return ret;
2298 }
2299
2300 int btrfs_truncate_item(struct btrfs_root *root, struct btrfs_path *path,
2301                         u32 new_size, int from_end)
2302 {
2303         int ret = 0;
2304         int slot;
2305         struct extent_buffer *leaf;
2306         struct btrfs_item *item;
2307         u32 nritems;
2308         unsigned int data_end;
2309         unsigned int old_data_start;
2310         unsigned int old_size;
2311         unsigned int size_diff;
2312         int i;
2313
2314         leaf = path->nodes[0];
2315         slot = path->slots[0];
2316
2317         old_size = btrfs_item_size_nr(leaf, slot);
2318         if (old_size == new_size)
2319                 return 0;
2320
2321         nritems = btrfs_header_nritems(leaf);
2322         data_end = leaf_data_end(root, leaf);
2323
2324         old_data_start = btrfs_item_offset_nr(leaf, slot);
2325
2326         size_diff = old_size - new_size;
2327
2328         BUG_ON(slot < 0);
2329         BUG_ON(slot >= nritems);
2330
2331         /*
2332          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2333          */
2334         /* first correct the data pointers */
2335         for (i = slot; i < nritems; i++) {
2336                 u32 ioff;
2337                 item = btrfs_item_nr(i);
2338                 ioff = btrfs_item_offset(leaf, item);
2339                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2340         }
2341
2342         /* shift the data */
2343         if (from_end) {
2344                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2345                               data_end + size_diff, btrfs_leaf_data(leaf) +
2346                               data_end, old_data_start + new_size - data_end);
2347         } else {
2348                 struct btrfs_disk_key disk_key;
2349                 u64 offset;
2350
2351                 btrfs_item_key(leaf, &disk_key, slot);
2352
2353                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2354                         unsigned long ptr;
2355                         struct btrfs_file_extent_item *fi;
2356
2357                         fi = btrfs_item_ptr(leaf, slot,
2358                                             struct btrfs_file_extent_item);
2359                         fi = (struct btrfs_file_extent_item *)(
2360                              (unsigned long)fi - size_diff);
2361
2362                         if (btrfs_file_extent_type(leaf, fi) ==
2363                             BTRFS_FILE_EXTENT_INLINE) {
2364                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2365                                 memmove_extent_buffer(leaf, ptr,
2366                                         (unsigned long)fi,
2367                                         offsetof(struct btrfs_file_extent_item,
2368                                                  disk_bytenr));
2369                         }
2370                 }
2371
2372                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2373                               data_end + size_diff, btrfs_leaf_data(leaf) +
2374                               data_end, old_data_start - data_end);
2375
2376                 offset = btrfs_disk_key_offset(&disk_key);
2377                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2378                 btrfs_set_item_key(leaf, &disk_key, slot);
2379                 if (slot == 0)
2380                         btrfs_fixup_low_keys(root, path, &disk_key, 1);
2381         }
2382
2383         item = btrfs_item_nr(slot);
2384         btrfs_set_item_size(leaf, item, new_size);
2385         btrfs_mark_buffer_dirty(leaf);
2386
2387         ret = 0;
2388         if (btrfs_leaf_free_space(root, leaf) < 0) {
2389                 btrfs_print_leaf(root, leaf);
2390                 BUG();
2391         }
2392         return ret;
2393 }
2394
2395 int btrfs_extend_item(struct btrfs_root *root, struct btrfs_path *path,
2396                       u32 data_size)
2397 {
2398         int ret = 0;
2399         int slot;
2400         struct extent_buffer *leaf;
2401         struct btrfs_item *item;
2402         u32 nritems;
2403         unsigned int data_end;
2404         unsigned int old_data;
2405         unsigned int old_size;
2406         int i;
2407
2408         leaf = path->nodes[0];
2409
2410         nritems = btrfs_header_nritems(leaf);
2411         data_end = leaf_data_end(root, leaf);
2412
2413         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2414                 btrfs_print_leaf(root, leaf);
2415                 BUG();
2416         }
2417         slot = path->slots[0];
2418         old_data = btrfs_item_end_nr(leaf, slot);
2419
2420         BUG_ON(slot < 0);
2421         if (slot >= nritems) {
2422                 btrfs_print_leaf(root, leaf);
2423                 printk("slot %d too large, nritems %d\n", slot, nritems);
2424                 BUG_ON(1);
2425         }
2426
2427         /*
2428          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2429          */
2430         /* first correct the data pointers */
2431         for (i = slot; i < nritems; i++) {
2432                 u32 ioff;
2433                 item = btrfs_item_nr(i);
2434                 ioff = btrfs_item_offset(leaf, item);
2435                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2436         }
2437
2438         /* shift the data */
2439         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2440                       data_end - data_size, btrfs_leaf_data(leaf) +
2441                       data_end, old_data - data_end);
2442
2443         data_end = old_data;
2444         old_size = btrfs_item_size_nr(leaf, slot);
2445         item = btrfs_item_nr(slot);
2446         btrfs_set_item_size(leaf, item, old_size + data_size);
2447         btrfs_mark_buffer_dirty(leaf);
2448
2449         ret = 0;
2450         if (btrfs_leaf_free_space(root, leaf) < 0) {
2451                 btrfs_print_leaf(root, leaf);
2452                 BUG();
2453         }
2454         return ret;
2455 }
2456
2457 /*
2458  * Given a key and some data, insert an item into the tree.
2459  * This does all the path init required, making room in the tree if needed.
2460  */
2461 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2462                             struct btrfs_root *root,
2463                             struct btrfs_path *path,
2464                             struct btrfs_key *cpu_key, u32 *data_size,
2465                             int nr)
2466 {
2467         struct extent_buffer *leaf;
2468         struct btrfs_item *item;
2469         int ret = 0;
2470         int slot;
2471         int i;
2472         u32 nritems;
2473         u32 total_size = 0;
2474         u32 total_data = 0;
2475         unsigned int data_end;
2476         struct btrfs_disk_key disk_key;
2477
2478         for (i = 0; i < nr; i++) {
2479                 total_data += data_size[i];
2480         }
2481
2482         /* create a root if there isn't one */
2483         if (!root->node)
2484                 BUG();
2485
2486         total_size = total_data + nr * sizeof(struct btrfs_item);
2487         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2488         if (ret == 0) {
2489                 return -EEXIST;
2490         }
2491         if (ret < 0)
2492                 goto out;
2493
2494         leaf = path->nodes[0];
2495
2496         nritems = btrfs_header_nritems(leaf);
2497         data_end = leaf_data_end(root, leaf);
2498
2499         if (btrfs_leaf_free_space(root, leaf) < total_size) {
2500                 btrfs_print_leaf(root, leaf);
2501                 printk("not enough freespace need %u have %d\n",
2502                        total_size, btrfs_leaf_free_space(root, leaf));
2503                 BUG();
2504         }
2505
2506         slot = path->slots[0];
2507         BUG_ON(slot < 0);
2508
2509         if (slot < nritems) {
2510                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2511
2512                 if (old_data < data_end) {
2513                         btrfs_print_leaf(root, leaf);
2514                         printk("slot %d old_data %d data_end %d\n",
2515                                slot, old_data, data_end);
2516                         BUG_ON(1);
2517                 }
2518                 /*
2519                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2520                  */
2521                 /* first correct the data pointers */
2522                 for (i = slot; i < nritems; i++) {
2523                         u32 ioff;
2524
2525                         item = btrfs_item_nr(i);
2526                         ioff = btrfs_item_offset(leaf, item);
2527                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2528                 }
2529
2530                 /* shift the items */
2531                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2532                               btrfs_item_nr_offset(slot),
2533                               (nritems - slot) * sizeof(struct btrfs_item));
2534
2535                 /* shift the data */
2536                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2537                               data_end - total_data, btrfs_leaf_data(leaf) +
2538                               data_end, old_data - data_end);
2539                 data_end = old_data;
2540         }
2541
2542         /* setup the item for the new data */
2543         for (i = 0; i < nr; i++) {
2544                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2545                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2546                 item = btrfs_item_nr(slot + i);
2547                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2548                 data_end -= data_size[i];
2549                 btrfs_set_item_size(leaf, item, data_size[i]);
2550         }
2551         btrfs_set_header_nritems(leaf, nritems + nr);
2552         btrfs_mark_buffer_dirty(leaf);
2553
2554         ret = 0;
2555         if (slot == 0) {
2556                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2557                 btrfs_fixup_low_keys(root, path, &disk_key, 1);
2558         }
2559
2560         if (btrfs_leaf_free_space(root, leaf) < 0) {
2561                 btrfs_print_leaf(root, leaf);
2562                 BUG();
2563         }
2564
2565 out:
2566         return ret;
2567 }
2568
2569 /*
2570  * Given a key and some data, insert an item into the tree.
2571  * This does all the path init required, making room in the tree if needed.
2572  */
2573 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2574                       *root, struct btrfs_key *cpu_key, void *data, u32
2575                       data_size)
2576 {
2577         int ret = 0;
2578         struct btrfs_path *path;
2579         struct extent_buffer *leaf;
2580         unsigned long ptr;
2581
2582         path = btrfs_alloc_path();
2583         if (!path)
2584                 return -ENOMEM;
2585
2586         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2587         if (!ret) {
2588                 leaf = path->nodes[0];
2589                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2590                 write_extent_buffer(leaf, data, ptr, data_size);
2591                 btrfs_mark_buffer_dirty(leaf);
2592         }
2593         btrfs_free_path(path);
2594         return ret;
2595 }
2596
2597 /*
2598  * delete the pointer from a given node.
2599  *
2600  * If the delete empties a node, the node is removed from the tree,
2601  * continuing all the way the root if required.  The root is converted into
2602  * a leaf if all the nodes are emptied.
2603  */
2604 int btrfs_del_ptr(struct btrfs_root *root, struct btrfs_path *path,
2605                 int level, int slot)
2606 {
2607         struct extent_buffer *parent = path->nodes[level];
2608         u32 nritems;
2609         int ret = 0;
2610
2611         nritems = btrfs_header_nritems(parent);
2612         if (slot < nritems - 1) {
2613                 /* shift the items */
2614                 memmove_extent_buffer(parent,
2615                               btrfs_node_key_ptr_offset(slot),
2616                               btrfs_node_key_ptr_offset(slot + 1),
2617                               sizeof(struct btrfs_key_ptr) *
2618                               (nritems - slot - 1));
2619         }
2620         nritems--;
2621         btrfs_set_header_nritems(parent, nritems);
2622         if (nritems == 0 && parent == root->node) {
2623                 BUG_ON(btrfs_header_level(root->node) != 1);
2624                 /* just turn the root into a leaf and break */
2625                 btrfs_set_header_level(root->node, 0);
2626         } else if (slot == 0) {
2627                 struct btrfs_disk_key disk_key;
2628
2629                 btrfs_node_key(parent, &disk_key, 0);
2630                 btrfs_fixup_low_keys(root, path, &disk_key, level + 1);
2631         }
2632         btrfs_mark_buffer_dirty(parent);
2633         return ret;
2634 }
2635
2636 /*
2637  * a helper function to delete the leaf pointed to by path->slots[1] and
2638  * path->nodes[1].
2639  *
2640  * This deletes the pointer in path->nodes[1] and frees the leaf
2641  * block extent.  zero is returned if it all worked out, < 0 otherwise.
2642  *
2643  * The path must have already been setup for deleting the leaf, including
2644  * all the proper balancing.  path->nodes[1] must be locked.
2645  */
2646 static noinline int btrfs_del_leaf(struct btrfs_trans_handle *trans,
2647                                    struct btrfs_root *root,
2648                                    struct btrfs_path *path,
2649                                    struct extent_buffer *leaf)
2650 {
2651         int ret;
2652
2653         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
2654         ret = btrfs_del_ptr(root, path, 1, path->slots[1]);
2655         if (ret)
2656                 return ret;
2657
2658         ret = btrfs_free_extent(trans, root, leaf->start, leaf->len,
2659                                 0, root->root_key.objectid, 0, 0);
2660         return ret;
2661 }
2662
2663 /*
2664  * delete the item at the leaf level in path.  If that empties
2665  * the leaf, remove it from the tree
2666  */
2667 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2668                     struct btrfs_path *path, int slot, int nr)
2669 {
2670         struct extent_buffer *leaf;
2671         struct btrfs_item *item;
2672         int last_off;
2673         int dsize = 0;
2674         int ret = 0;
2675         int wret;
2676         int i;
2677         u32 nritems;
2678
2679         leaf = path->nodes[0];
2680         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2681
2682         for (i = 0; i < nr; i++)
2683                 dsize += btrfs_item_size_nr(leaf, slot + i);
2684
2685         nritems = btrfs_header_nritems(leaf);
2686
2687         if (slot + nr != nritems) {
2688                 int data_end = leaf_data_end(root, leaf);
2689
2690                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2691                               data_end + dsize,
2692                               btrfs_leaf_data(leaf) + data_end,
2693                               last_off - data_end);
2694
2695                 for (i = slot + nr; i < nritems; i++) {
2696                         u32 ioff;
2697
2698                         item = btrfs_item_nr(i);
2699                         ioff = btrfs_item_offset(leaf, item);
2700                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2701                 }
2702
2703                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2704                               btrfs_item_nr_offset(slot + nr),
2705                               sizeof(struct btrfs_item) *
2706                               (nritems - slot - nr));
2707         }
2708         btrfs_set_header_nritems(leaf, nritems - nr);
2709         nritems -= nr;
2710
2711         /* delete the leaf if we've emptied it */
2712         if (nritems == 0) {
2713                 if (leaf == root->node) {
2714                         btrfs_set_header_level(leaf, 0);
2715                 } else {
2716                         clean_tree_block(trans, root, leaf);
2717                         wret = btrfs_del_leaf(trans, root, path, leaf);
2718                         BUG_ON(ret);
2719                         if (wret)
2720                                 ret = wret;
2721                 }
2722         } else {
2723                 int used = leaf_space_used(leaf, 0, nritems);
2724                 if (slot == 0) {
2725                         struct btrfs_disk_key disk_key;
2726
2727                         btrfs_item_key(leaf, &disk_key, 0);
2728                         btrfs_fixup_low_keys(root, path, &disk_key, 1);
2729                 }
2730
2731                 /* delete the leaf if it is mostly empty */
2732                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2733                         /* push_leaf_left fixes the path.
2734                          * make sure the path still points to our leaf
2735                          * for possible call to del_ptr below
2736                          */
2737                         slot = path->slots[1];
2738                         extent_buffer_get(leaf);
2739
2740                         wret = push_leaf_left(trans, root, path, 1, 1);
2741                         if (wret < 0 && wret != -ENOSPC)
2742                                 ret = wret;
2743
2744                         if (path->nodes[0] == leaf &&
2745                             btrfs_header_nritems(leaf)) {
2746                                 wret = push_leaf_right(trans, root, path, 1, 1);
2747                                 if (wret < 0 && wret != -ENOSPC)
2748                                         ret = wret;
2749                         }
2750
2751                         if (btrfs_header_nritems(leaf) == 0) {
2752                                 clean_tree_block(trans, root, leaf);
2753                                 path->slots[1] = slot;
2754                                 ret = btrfs_del_leaf(trans, root, path, leaf);
2755                                 BUG_ON(ret);
2756                                 free_extent_buffer(leaf);
2757
2758                         } else {
2759                                 btrfs_mark_buffer_dirty(leaf);
2760                                 free_extent_buffer(leaf);
2761                         }
2762                 } else {
2763                         btrfs_mark_buffer_dirty(leaf);
2764                 }
2765         }
2766         return ret;
2767 }
2768
2769 /*
2770  * walk up the tree as far as required to find the previous leaf.
2771  * returns 0 if it found something or 1 if there are no lesser leaves.
2772  * returns < 0 on io errors.
2773  */
2774 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2775 {
2776         int slot;
2777         int level = 1;
2778         struct extent_buffer *c;
2779         struct extent_buffer *next = NULL;
2780         struct btrfs_fs_info *fs_info = root->fs_info;
2781
2782         while(level < BTRFS_MAX_LEVEL) {
2783                 if (!path->nodes[level])
2784                         return 1;
2785
2786                 slot = path->slots[level];
2787                 c = path->nodes[level];
2788                 if (slot == 0) {
2789                         level++;
2790                         if (level == BTRFS_MAX_LEVEL)
2791                                 return 1;
2792                         continue;
2793                 }
2794                 slot--;
2795
2796                 next = read_node_slot(fs_info, c, slot);
2797                 if (!extent_buffer_uptodate(next)) {
2798                         if (IS_ERR(next))
2799                                 return PTR_ERR(next);
2800                         return -EIO;
2801                 }
2802                 break;
2803         }
2804         path->slots[level] = slot;
2805         while(1) {
2806                 level--;
2807                 c = path->nodes[level];
2808                 free_extent_buffer(c);
2809                 slot = btrfs_header_nritems(next);
2810                 if (slot != 0)
2811                         slot--;
2812                 path->nodes[level] = next;
2813                 path->slots[level] = slot;
2814                 if (!level)
2815                         break;
2816                 next = read_node_slot(fs_info, next, slot);
2817                 if (!extent_buffer_uptodate(next)) {
2818                         if (IS_ERR(next))
2819                                 return PTR_ERR(next);
2820                         return -EIO;
2821                 }
2822         }
2823         return 0;
2824 }
2825
2826 /*
2827  * walk up the tree as far as required to find the next leaf.
2828  * returns 0 if it found something or 1 if there are no greater leaves.
2829  * returns < 0 on io errors.
2830  */
2831 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2832 {
2833         int slot;
2834         int level = 1;
2835         struct extent_buffer *c;
2836         struct extent_buffer *next = NULL;
2837         struct btrfs_fs_info *fs_info = root->fs_info;
2838
2839         while(level < BTRFS_MAX_LEVEL) {
2840                 if (!path->nodes[level])
2841                         return 1;
2842
2843                 slot = path->slots[level] + 1;
2844                 c = path->nodes[level];
2845                 if (slot >= btrfs_header_nritems(c)) {
2846                         level++;
2847                         if (level == BTRFS_MAX_LEVEL)
2848                                 return 1;
2849                         continue;
2850                 }
2851
2852                 if (path->reada)
2853                         reada_for_search(root, path, level, slot, 0);
2854
2855                 next = read_node_slot(fs_info, c, slot);
2856                 if (!extent_buffer_uptodate(next))
2857                         return -EIO;
2858                 break;
2859         }
2860         path->slots[level] = slot;
2861         while(1) {
2862                 level--;
2863                 c = path->nodes[level];
2864                 free_extent_buffer(c);
2865                 path->nodes[level] = next;
2866                 path->slots[level] = 0;
2867                 if (!level)
2868                         break;
2869                 if (path->reada)
2870                         reada_for_search(root, path, level, 0, 0);
2871                 next = read_node_slot(fs_info, next, 0);
2872                 if (!extent_buffer_uptodate(next))
2873                         return -EIO;
2874         }
2875         return 0;
2876 }
2877
2878 int btrfs_previous_item(struct btrfs_root *root,
2879                         struct btrfs_path *path, u64 min_objectid,
2880                         int type)
2881 {
2882         struct btrfs_key found_key;
2883         struct extent_buffer *leaf;
2884         u32 nritems;
2885         int ret;
2886
2887         while(1) {
2888                 if (path->slots[0] == 0) {
2889                         ret = btrfs_prev_leaf(root, path);
2890                         if (ret != 0)
2891                                 return ret;
2892                 } else {
2893                         path->slots[0]--;
2894                 }
2895                 leaf = path->nodes[0];
2896                 nritems = btrfs_header_nritems(leaf);
2897                 if (nritems == 0)
2898                         return 1;
2899                 if (path->slots[0] == nritems)
2900                         path->slots[0]--;
2901
2902                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2903                 if (found_key.objectid < min_objectid)
2904                         break;
2905                 if (found_key.type == type)
2906                         return 0;
2907                 if (found_key.objectid == min_objectid &&
2908                     found_key.type < type)
2909                         break;
2910         }
2911         return 1;
2912 }
2913
2914 /*
2915  * search in extent tree to find a previous Metadata/Data extent item with
2916  * min objecitd.
2917  *
2918  * returns 0 if something is found, 1 if nothing was found and < 0 on error
2919  */
2920 int btrfs_previous_extent_item(struct btrfs_root *root,
2921                         struct btrfs_path *path, u64 min_objectid)
2922 {
2923         struct btrfs_key found_key;
2924         struct extent_buffer *leaf;
2925         u32 nritems;
2926         int ret;
2927
2928         while (1) {
2929                 if (path->slots[0] == 0) {
2930                         ret = btrfs_prev_leaf(root, path);
2931                         if (ret != 0)
2932                                 return ret;
2933                 } else {
2934                         path->slots[0]--;
2935                 }
2936                 leaf = path->nodes[0];
2937                 nritems = btrfs_header_nritems(leaf);
2938                 if (nritems == 0)
2939                         return 1;
2940                 if (path->slots[0] == nritems)
2941                         path->slots[0]--;
2942
2943                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2944                 if (found_key.objectid < min_objectid)
2945                         break;
2946                 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
2947                     found_key.type == BTRFS_METADATA_ITEM_KEY)
2948                         return 0;
2949                 if (found_key.objectid == min_objectid &&
2950                     found_key.type < BTRFS_EXTENT_ITEM_KEY)
2951                         break;
2952         }
2953         return 1;
2954 }
2955
2956 /*
2957  * Search in extent tree to found next meta/data extent
2958  * Caller needs to check for no-hole or skinny metadata features.
2959  */
2960 int btrfs_next_extent_item(struct btrfs_root *root,
2961                         struct btrfs_path *path, u64 max_objectid)
2962 {
2963         struct btrfs_key found_key;
2964         int ret;
2965
2966         while (1) {
2967                 ret = btrfs_next_item(root, path);
2968                 if (ret)
2969                         return ret;
2970                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2971                                       path->slots[0]);
2972                 if (found_key.objectid > max_objectid)
2973                         return 1;
2974                 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
2975                     found_key.type == BTRFS_METADATA_ITEM_KEY)
2976                 return 0;
2977         }
2978 }