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 btrfs_fs_info *fs_info = root->fs_info;
 972         struct extent_buffer *node;
 973         struct btrfs_disk_key disk_key;
 974         u32 nritems;
 975         u64 search;
 976         u64 lowest_read;
 977         u64 highest_read;
 978         u64 nread = 0;
 979         int direction = path->reada;
 980         struct extent_buffer *eb;
 981         u32 nr;
 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         eb = btrfs_find_tree_block(fs_info, search, fs_info->nodesize);
 993         if (eb) {
 994                 free_extent_buffer(eb);
 995                 return;
 996         }
 997
 998         highest_read = search;
 999         lowest_read = search;
1000
1001         nritems = btrfs_header_nritems(node);
1002         nr = slot;
1003         while(1) {
1004                 if (direction < 0) {
1005                         if (nr == 0)
1006                                 break;
1007                         nr--;
1008                 } else if (direction > 0) {
1009                         nr++;
1010                         if (nr >= nritems)
1011                                 break;
1012                 }
1013                 if (path->reada < 0 && objectid) {
1014                         btrfs_node_key(node, &disk_key, nr);
1015                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1016                                 break;
1017                 }
1018                 search = btrfs_node_blockptr(node, nr);
1019                 if ((search >= lowest_read && search <= highest_read) ||
1020                     (search < lowest_read && lowest_read - search <= 32768) ||
1021                     (search > highest_read && search - highest_read <= 32768)) {
1022                         readahead_tree_block(fs_info, search, fs_info->nodesize,
1023                                      btrfs_node_ptr_generation(node, nr));
1024                         nread += fs_info->nodesize;
1025                 }
1026                 nscan++;
1027                 if (path->reada < 2 && (nread > SZ_256K || nscan > 32))
1028                         break;
1029                 if(nread > SZ_1M || nscan > 128)
1030                         break;
1031
1032                 if (search < lowest_read)
1033                         lowest_read = search;
1034                 if (search > highest_read)
1035                         highest_read = search;
1036         }
1037 }
1038
1039 int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *found_path,
1040                 u64 iobjectid, u64 ioff, u8 key_type,
1041                 struct btrfs_key *found_key)
1042 {
1043         int ret;
1044         struct btrfs_key key;
1045         struct extent_buffer *eb;
1046         struct btrfs_path *path;
1047
1048         key.type = key_type;
1049         key.objectid = iobjectid;
1050         key.offset = ioff;
1051
1052         if (found_path == NULL) {
1053                 path = btrfs_alloc_path();
1054                 if (!path)
1055                         return -ENOMEM;
1056         } else
1057                 path = found_path;
1058
1059         ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
1060         if ((ret < 0) || (found_key == NULL))
1061                 goto out;
1062
1063         eb = path->nodes[0];
1064         if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
1065                 ret = btrfs_next_leaf(fs_root, path);
1066                 if (ret)
1067                         goto out;
1068                 eb = path->nodes[0];
1069         }
1070
1071         btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
1072         if (found_key->type != key.type ||
1073                         found_key->objectid != key.objectid) {
1074                 ret = 1;
1075                 goto out;
1076         }
1077
1078 out:
1079         if (path != found_path)
1080                 btrfs_free_path(path);
1081         return ret;
1082 }
1083
1084 /*
1085  * look for key in the tree.  path is filled in with nodes along the way
1086  * if key is found, we return zero and you can find the item in the leaf
1087  * level of the path (level 0)
1088  *
1089  * If the key isn't found, the path points to the slot where it should
1090  * be inserted, and 1 is returned.  If there are other errors during the
1091  * search a negative error number is returned.
1092  *
1093  * if ins_len > 0, nodes and leaves will be split as we walk down the
1094  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1095  * possible)
1096  */
1097 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1098                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1099                       ins_len, int cow)
1100 {
1101         struct extent_buffer *b;
1102         int slot;
1103         int ret;
1104         int level;
1105         int should_reada = p->reada;
1106         struct btrfs_fs_info *fs_info = root->fs_info;
1107         u8 lowest_level = 0;
1108
1109         lowest_level = p->lowest_level;
1110         WARN_ON(lowest_level && ins_len > 0);
1111         WARN_ON(p->nodes[0] != NULL);
1112         /*
1113         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
1114         */
1115 again:
1116         b = root->node;
1117         extent_buffer_get(b);
1118         while (b) {
1119                 level = btrfs_header_level(b);
1120                 if (cow) {
1121                         int wret;
1122                         wret = btrfs_cow_block(trans, root, b,
1123                                                p->nodes[level + 1],
1124                                                p->slots[level + 1],
1125                                                &b);
1126                         if (wret) {
1127                                 free_extent_buffer(b);
1128                                 return wret;
1129                         }
1130                 }
1131                 BUG_ON(!cow && ins_len);
1132                 if (level != btrfs_header_level(b))
1133                         WARN_ON(1);
1134                 level = btrfs_header_level(b);
1135                 p->nodes[level] = b;
1136                 ret = check_block(root, p, level);
1137                 if (ret)
1138                         return -1;
1139                 ret = bin_search(b, key, level, &slot);
1140                 if (level != 0) {
1141                         if (ret && slot > 0)
1142                                 slot -= 1;
1143                         p->slots[level] = slot;
1144                         if ((p->search_for_split || ins_len > 0) &&
1145                             btrfs_header_nritems(b) >=
1146                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1147                                 int sret = split_node(trans, root, p, level);
1148                                 BUG_ON(sret > 0);
1149                                 if (sret)
1150                                         return sret;
1151                                 b = p->nodes[level];
1152                                 slot = p->slots[level];
1153                         } else if (ins_len < 0) {
1154                                 int sret = balance_level(trans, root, p,
1155                                                          level);
1156                                 if (sret)
1157                                         return sret;
1158                                 b = p->nodes[level];
1159                                 if (!b) {
1160                                         btrfs_release_path(p);
1161                                         goto again;
1162                                 }
1163                                 slot = p->slots[level];
1164                                 BUG_ON(btrfs_header_nritems(b) == 1);
1165                         }
1166                         /* this is only true while dropping a snapshot */
1167                         if (level == lowest_level)
1168                                 break;
1169
1170                         if (should_reada)
1171                                 reada_for_search(root, p, level, slot,
1172                                                  key->objectid);
1173
1174                         b = read_node_slot(fs_info, b, slot);
1175                         if (!extent_buffer_uptodate(b))
1176                                 return -EIO;
1177                 } else {
1178                         p->slots[level] = slot;
1179                         if (ins_len > 0 &&
1180                             ins_len > btrfs_leaf_free_space(root, b)) {
1181                                 int sret = split_leaf(trans, root, key,
1182                                                       p, ins_len, ret == 0);
1183                                 BUG_ON(sret > 0);
1184                                 if (sret)
1185                                         return sret;
1186                         }
1187                         return ret;
1188                 }
1189         }
1190         return 1;
1191 }
1192
1193 /*
1194  * adjust the pointers going up the tree, starting at level
1195  * making sure the right key of each node is points to 'key'.
1196  * This is used after shifting pointers to the left, so it stops
1197  * fixing up pointers when a given leaf/node is not in slot 0 of the
1198  * higher levels
1199  */
1200 void btrfs_fixup_low_keys(struct btrfs_root *root, struct btrfs_path *path,
1201                           struct btrfs_disk_key *key, int level)
1202 {
1203         int i;
1204         struct extent_buffer *t;
1205
1206         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1207                 int tslot = path->slots[i];
1208                 if (!path->nodes[i])
1209                         break;
1210                 t = path->nodes[i];
1211                 btrfs_set_node_key(t, key, tslot);
1212                 btrfs_mark_buffer_dirty(path->nodes[i]);
1213                 if (tslot != 0)
1214                         break;
1215         }
1216 }
1217
1218 /*
1219  * update item key.
1220  *
1221  * This function isn't completely safe. It's the caller's responsibility
1222  * that the new key won't break the order
1223  */
1224 int btrfs_set_item_key_safe(struct btrfs_root *root, struct btrfs_path *path,
1225                             struct btrfs_key *new_key)
1226 {
1227         struct btrfs_disk_key disk_key;
1228         struct extent_buffer *eb;
1229         int slot;
1230
1231         eb = path->nodes[0];
1232         slot = path->slots[0];
1233         if (slot > 0) {
1234                 btrfs_item_key(eb, &disk_key, slot - 1);
1235                 if (btrfs_comp_keys(&disk_key, new_key) >= 0)
1236                         return -1;
1237         }
1238         if (slot < btrfs_header_nritems(eb) - 1) {
1239                 btrfs_item_key(eb, &disk_key, slot + 1);
1240                 if (btrfs_comp_keys(&disk_key, new_key) <= 0)
1241                         return -1;
1242         }
1243
1244         btrfs_cpu_key_to_disk(&disk_key, new_key);
1245         btrfs_set_item_key(eb, &disk_key, slot);
1246         btrfs_mark_buffer_dirty(eb);
1247         if (slot == 0)
1248                 btrfs_fixup_low_keys(root, path, &disk_key, 1);
1249         return 0;
1250 }
1251
1252 /*
1253  * update an item key without the safety checks.  This is meant to be called by
1254  * fsck only.
1255  */
1256 void btrfs_set_item_key_unsafe(struct btrfs_root *root,
1257                                struct btrfs_path *path,
1258                                struct btrfs_key *new_key)
1259 {
1260         struct btrfs_disk_key disk_key;
1261         struct extent_buffer *eb;
1262         int slot;
1263
1264         eb = path->nodes[0];
1265         slot = path->slots[0];
1266
1267         btrfs_cpu_key_to_disk(&disk_key, new_key);
1268         btrfs_set_item_key(eb, &disk_key, slot);
1269         btrfs_mark_buffer_dirty(eb);
1270         if (slot == 0)
1271                 btrfs_fixup_low_keys(root, path, &disk_key, 1);
1272 }
1273
1274 /*
1275  * try to push data from one node into the next node left in the
1276  * tree.
1277  *
1278  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1279  * error, and > 0 if there was no room in the left hand block.
1280  */
1281 static int push_node_left(struct btrfs_trans_handle *trans,
1282                           struct btrfs_root *root, struct extent_buffer *dst,
1283                           struct extent_buffer *src, int empty)
1284 {
1285         int push_items = 0;
1286         int src_nritems;
1287         int dst_nritems;
1288         int ret = 0;
1289
1290         src_nritems = btrfs_header_nritems(src);
1291         dst_nritems = btrfs_header_nritems(dst);
1292         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1293         WARN_ON(btrfs_header_generation(src) != trans->transid);
1294         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1295
1296         if (!empty && src_nritems <= 8)
1297                 return 1;
1298
1299         if (push_items <= 0) {
1300                 return 1;
1301         }
1302
1303         if (empty) {
1304                 push_items = min(src_nritems, push_items);
1305                 if (push_items < src_nritems) {
1306                         /* leave at least 8 pointers in the node if
1307                          * we aren't going to empty it
1308                          */
1309                         if (src_nritems - push_items < 8) {
1310                                 if (push_items <= 8)
1311                                         return 1;
1312                                 push_items -= 8;
1313                         }
1314                 }
1315         } else
1316                 push_items = min(src_nritems - 8, push_items);
1317
1318         copy_extent_buffer(dst, src,
1319                            btrfs_node_key_ptr_offset(dst_nritems),
1320                            btrfs_node_key_ptr_offset(0),
1321                            push_items * sizeof(struct btrfs_key_ptr));
1322
1323         if (push_items < src_nritems) {
1324                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1325                                       btrfs_node_key_ptr_offset(push_items),
1326                                       (src_nritems - push_items) *
1327                                       sizeof(struct btrfs_key_ptr));
1328         }
1329         btrfs_set_header_nritems(src, src_nritems - push_items);
1330         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1331         btrfs_mark_buffer_dirty(src);
1332         btrfs_mark_buffer_dirty(dst);
1333
1334         return ret;
1335 }
1336
1337 /*
1338  * try to push data from one node into the next node right in the
1339  * tree.
1340  *
1341  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1342  * error, and > 0 if there was no room in the right hand block.
1343  *
1344  * this will  only push up to 1/2 the contents of the left node over
1345  */
1346 static int balance_node_right(struct btrfs_trans_handle *trans,
1347                               struct btrfs_root *root,
1348                               struct extent_buffer *dst,
1349                               struct extent_buffer *src)
1350 {
1351         int push_items = 0;
1352         int max_push;
1353         int src_nritems;
1354         int dst_nritems;
1355         int ret = 0;
1356
1357         WARN_ON(btrfs_header_generation(src) != trans->transid);
1358         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1359
1360         src_nritems = btrfs_header_nritems(src);
1361         dst_nritems = btrfs_header_nritems(dst);
1362         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1363         if (push_items <= 0) {
1364                 return 1;
1365         }
1366
1367         if (src_nritems < 4) {
1368                 return 1;
1369         }
1370
1371         max_push = src_nritems / 2 + 1;
1372         /* don't try to empty the node */
1373         if (max_push >= src_nritems) {
1374                 return 1;
1375         }
1376
1377         if (max_push < push_items)
1378                 push_items = max_push;
1379
1380         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1381                                       btrfs_node_key_ptr_offset(0),
1382                                       (dst_nritems) *
1383                                       sizeof(struct btrfs_key_ptr));
1384
1385         copy_extent_buffer(dst, src,
1386                            btrfs_node_key_ptr_offset(0),
1387                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1388                            push_items * sizeof(struct btrfs_key_ptr));
1389
1390         btrfs_set_header_nritems(src, src_nritems - push_items);
1391         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1392
1393         btrfs_mark_buffer_dirty(src);
1394         btrfs_mark_buffer_dirty(dst);
1395
1396         return ret;
1397 }
1398
1399 /*
1400  * helper function to insert a new root level in the tree.
1401  * A new node is allocated, and a single item is inserted to
1402  * point to the existing root
1403  *
1404  * returns zero on success or < 0 on failure.
1405  */
1406 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1407                            struct btrfs_root *root,
1408                            struct btrfs_path *path, int level)
1409 {
1410         u64 lower_gen;
1411         struct extent_buffer *lower;
1412         struct extent_buffer *c;
1413         struct extent_buffer *old;
1414         struct btrfs_disk_key lower_key;
1415
1416         BUG_ON(path->nodes[level]);
1417         BUG_ON(path->nodes[level-1] != root->node);
1418
1419         lower = path->nodes[level-1];
1420         if (level == 1)
1421                 btrfs_item_key(lower, &lower_key, 0);
1422         else
1423                 btrfs_node_key(lower, &lower_key, 0);
1424
1425         c = btrfs_alloc_free_block(trans, root, root->fs_info->nodesize,
1426                                    root->root_key.objectid, &lower_key,
1427                                    level, root->node->start, 0);
1428
1429         if (IS_ERR(c))
1430                 return PTR_ERR(c);
1431
1432         memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
1433         btrfs_set_header_nritems(c, 1);
1434         btrfs_set_header_level(c, level);
1435         btrfs_set_header_bytenr(c, c->start);
1436         btrfs_set_header_generation(c, trans->transid);
1437         btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
1438         btrfs_set_header_owner(c, root->root_key.objectid);
1439
1440         write_extent_buffer(c, root->fs_info->fsid,
1441                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
1442
1443         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1444                             btrfs_header_chunk_tree_uuid(c),
1445                             BTRFS_UUID_SIZE);
1446
1447         btrfs_set_node_key(c, &lower_key, 0);
1448         btrfs_set_node_blockptr(c, 0, lower->start);
1449         lower_gen = btrfs_header_generation(lower);
1450         WARN_ON(lower_gen != trans->transid);
1451
1452         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1453
1454         btrfs_mark_buffer_dirty(c);
1455
1456         old = root->node;
1457         root->node = c;
1458
1459         /* the super has an extra ref to root->node */
1460         free_extent_buffer(old);
1461
1462         add_root_to_dirty_list(root);
1463         extent_buffer_get(c);
1464         path->nodes[level] = c;
1465         path->slots[level] = 0;
1466         return 0;
1467 }
1468
1469 /*
1470  * worker function to insert a single pointer in a node.
1471  * the node should have enough room for the pointer already
1472  *
1473  * slot and level indicate where you want the key to go, and
1474  * blocknr is the block the key points to.
1475  *
1476  * returns zero on success and < 0 on any error
1477  */
1478 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1479                       *root, struct btrfs_path *path, struct btrfs_disk_key
1480                       *key, u64 bytenr, int slot, int level)
1481 {
1482         struct extent_buffer *lower;
1483         int nritems;
1484
1485         BUG_ON(!path->nodes[level]);
1486         lower = path->nodes[level];
1487         nritems = btrfs_header_nritems(lower);
1488         if (slot > nritems)
1489                 BUG();
1490         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1491                 BUG();
1492         if (slot < nritems) {
1493                 /* shift the items */
1494                 memmove_extent_buffer(lower,
1495                               btrfs_node_key_ptr_offset(slot + 1),
1496                               btrfs_node_key_ptr_offset(slot),
1497                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1498         }
1499         btrfs_set_node_key(lower, key, slot);
1500         btrfs_set_node_blockptr(lower, slot, bytenr);
1501         WARN_ON(trans->transid == 0);
1502         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1503         btrfs_set_header_nritems(lower, nritems + 1);
1504         btrfs_mark_buffer_dirty(lower);
1505         return 0;
1506 }
1507
1508 /*
1509  * split the node at the specified level in path in two.
1510  * The path is corrected to point to the appropriate node after the split
1511  *
1512  * Before splitting this tries to make some room in the node by pushing
1513  * left and right, if either one works, it returns right away.
1514  *
1515  * returns 0 on success and < 0 on failure
1516  */
1517 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1518                       *root, struct btrfs_path *path, int level)
1519 {
1520         struct extent_buffer *c;
1521         struct extent_buffer *split;
1522         struct btrfs_disk_key disk_key;
1523         int mid;
1524         int ret;
1525         int wret;
1526         u32 c_nritems;
1527
1528         c = path->nodes[level];
1529         WARN_ON(btrfs_header_generation(c) != trans->transid);
1530         if (c == root->node) {
1531                 /* trying to split the root, lets make a new one */
1532                 ret = insert_new_root(trans, root, path, level + 1);
1533                 if (ret)
1534                         return ret;
1535         } else {
1536                 ret = push_nodes_for_insert(trans, root, path, level);
1537                 c = path->nodes[level];
1538                 if (!ret && btrfs_header_nritems(c) <
1539                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
1540                         return 0;
1541                 if (ret < 0)
1542                         return ret;
1543         }
1544
1545         c_nritems = btrfs_header_nritems(c);
1546         mid = (c_nritems + 1) / 2;
1547         btrfs_node_key(c, &disk_key, mid);
1548
1549         split = btrfs_alloc_free_block(trans, root, root->fs_info->nodesize,
1550                                         root->root_key.objectid,
1551                                         &disk_key, level, c->start, 0);
1552         if (IS_ERR(split))
1553                 return PTR_ERR(split);
1554
1555         memset_extent_buffer(split, 0, 0, sizeof(struct btrfs_header));
1556         btrfs_set_header_level(split, btrfs_header_level(c));
1557         btrfs_set_header_bytenr(split, split->start);
1558         btrfs_set_header_generation(split, trans->transid);
1559         btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
1560         btrfs_set_header_owner(split, root->root_key.objectid);
1561         write_extent_buffer(split, root->fs_info->fsid,
1562                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
1563         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1564                             btrfs_header_chunk_tree_uuid(split),
1565                             BTRFS_UUID_SIZE);
1566
1567
1568         copy_extent_buffer(split, c,
1569                            btrfs_node_key_ptr_offset(0),
1570                            btrfs_node_key_ptr_offset(mid),
1571                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1572         btrfs_set_header_nritems(split, c_nritems - mid);
1573         btrfs_set_header_nritems(c, mid);
1574         ret = 0;
1575
1576         btrfs_mark_buffer_dirty(c);
1577         btrfs_mark_buffer_dirty(split);
1578
1579         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1580                           path->slots[level + 1] + 1,
1581                           level + 1);
1582         if (wret)
1583                 ret = wret;
1584
1585         if (path->slots[level] >= mid) {
1586                 path->slots[level] -= mid;
1587                 free_extent_buffer(c);
1588                 path->nodes[level] = split;
1589                 path->slots[level + 1] += 1;
1590         } else {
1591                 free_extent_buffer(split);
1592         }
1593         return ret;
1594 }
1595
1596 /*
1597  * how many bytes are required to store the items in a leaf.  start
1598  * and nr indicate which items in the leaf to check.  This totals up the
1599  * space used both by the item structs and the item data
1600  */
1601 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1602 {
1603         int data_len;
1604         int nritems = btrfs_header_nritems(l);
1605         int end = min(nritems, start + nr) - 1;
1606
1607         if (!nr)
1608                 return 0;
1609         data_len = btrfs_item_end_nr(l, start);
1610         data_len = data_len - btrfs_item_offset_nr(l, end);
1611         data_len += sizeof(struct btrfs_item) * nr;
1612         WARN_ON(data_len < 0);
1613         return data_len;
1614 }
1615
1616 /*
1617  * The space between the end of the leaf items and
1618  * the start of the leaf data.  IOW, how much room
1619  * the leaf has left for both items and data
1620  */
1621 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1622 {
1623         u32 nodesize = (root ? BTRFS_LEAF_DATA_SIZE(root) : leaf->len);
1624         int nritems = btrfs_header_nritems(leaf);
1625         int ret;
1626         ret = nodesize - leaf_space_used(leaf, 0, nritems);
1627         if (ret < 0) {
1628                 printk("leaf free space ret %d, leaf data size %u, used %d nritems %d\n",
1629                        ret, nodesize, leaf_space_used(leaf, 0, nritems),
1630                        nritems);
1631         }
1632         return ret;
1633 }
1634
1635 /*
1636  * push some data in the path leaf to the right, trying to free up at
1637  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1638  *
1639  * returns 1 if the push failed because the other node didn't have enough
1640  * room, 0 if everything worked out and < 0 if there were major errors.
1641  */
1642 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1643                            *root, struct btrfs_path *path, int data_size,
1644                            int empty)
1645 {
1646         struct extent_buffer *left = path->nodes[0];
1647         struct extent_buffer *right;
1648         struct extent_buffer *upper;
1649         struct btrfs_disk_key disk_key;
1650         struct btrfs_fs_info *fs_info = root->fs_info;
1651         int slot;
1652         u32 i;
1653         int free_space;
1654         int push_space = 0;
1655         int push_items = 0;
1656         struct btrfs_item *item;
1657         u32 left_nritems;
1658         u32 nr;
1659         u32 right_nritems;
1660         u32 data_end;
1661         u32 this_item_size;
1662         int ret;
1663
1664         slot = path->slots[1];
1665         if (!path->nodes[1]) {
1666                 return 1;
1667         }
1668         upper = path->nodes[1];
1669         if (slot >= btrfs_header_nritems(upper) - 1)
1670                 return 1;
1671
1672         right = read_node_slot(fs_info, upper, slot + 1);
1673         if (!extent_buffer_uptodate(right)) {
1674                 if (IS_ERR(right))
1675                         return PTR_ERR(right);
1676                 return -EIO;
1677         }
1678         free_space = btrfs_leaf_free_space(root, right);
1679         if (free_space < data_size) {
1680                 free_extent_buffer(right);
1681                 return 1;
1682         }
1683
1684         /* cow and double check */
1685         ret = btrfs_cow_block(trans, root, right, upper,
1686                               slot + 1, &right);
1687         if (ret) {
1688                 free_extent_buffer(right);
1689                 return 1;
1690         }
1691         free_space = btrfs_leaf_free_space(root, right);
1692         if (free_space < data_size) {
1693                 free_extent_buffer(right);
1694                 return 1;
1695         }
1696
1697         left_nritems = btrfs_header_nritems(left);
1698         if (left_nritems == 0) {
1699                 free_extent_buffer(right);
1700                 return 1;
1701         }
1702
1703         if (empty)
1704                 nr = 0;
1705         else
1706                 nr = 1;
1707
1708         i = left_nritems - 1;
1709         while (i >= nr) {
1710                 item = btrfs_item_nr(i);
1711
1712                 if (path->slots[0] == i)
1713                         push_space += data_size + sizeof(*item);
1714
1715                 this_item_size = btrfs_item_size(left, item);
1716                 if (this_item_size + sizeof(*item) + push_space > free_space)
1717                         break;
1718                 push_items++;
1719                 push_space += this_item_size + sizeof(*item);
1720                 if (i == 0)
1721                         break;
1722                 i--;
1723         }
1724
1725         if (push_items == 0) {
1726                 free_extent_buffer(right);
1727                 return 1;
1728         }
1729
1730         if (!empty && push_items == left_nritems)
1731                 WARN_ON(1);
1732
1733         /* push left to right */
1734         right_nritems = btrfs_header_nritems(right);
1735
1736         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1737         push_space -= leaf_data_end(root, left);
1738
1739         /* make room in the right data area */
1740         data_end = leaf_data_end(root, right);
1741         memmove_extent_buffer(right,
1742                               btrfs_leaf_data(right) + data_end - push_space,
1743                               btrfs_leaf_data(right) + data_end,
1744                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1745
1746         /* copy from the left data area */
1747         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1748                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1749                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1750                      push_space);
1751
1752         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1753                               btrfs_item_nr_offset(0),
1754                               right_nritems * sizeof(struct btrfs_item));
1755
1756         /* copy the items from left to right */
1757         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1758                    btrfs_item_nr_offset(left_nritems - push_items),
1759                    push_items * sizeof(struct btrfs_item));
1760
1761         /* update the item pointers */
1762         right_nritems += push_items;
1763         btrfs_set_header_nritems(right, right_nritems);
1764         push_space = BTRFS_LEAF_DATA_SIZE(root);
1765         for (i = 0; i < right_nritems; i++) {
1766                 item = btrfs_item_nr(i);
1767                 push_space -= btrfs_item_size(right, item);
1768                 btrfs_set_item_offset(right, item, push_space);
1769         }
1770
1771         left_nritems -= push_items;
1772         btrfs_set_header_nritems(left, left_nritems);
1773
1774         if (left_nritems)
1775                 btrfs_mark_buffer_dirty(left);
1776         btrfs_mark_buffer_dirty(right);
1777
1778         btrfs_item_key(right, &disk_key, 0);
1779         btrfs_set_node_key(upper, &disk_key, slot + 1);
1780         btrfs_mark_buffer_dirty(upper);
1781
1782         /* then fixup the leaf pointer in the path */
1783         if (path->slots[0] >= left_nritems) {
1784                 path->slots[0] -= left_nritems;
1785                 free_extent_buffer(path->nodes[0]);
1786                 path->nodes[0] = right;
1787                 path->slots[1] += 1;
1788         } else {
1789                 free_extent_buffer(right);
1790         }
1791         return 0;
1792 }
1793 /*
1794  * push some data in the path leaf to the left, trying to free up at
1795  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1796  */
1797 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1798                           *root, struct btrfs_path *path, int data_size,
1799                           int empty)
1800 {
1801         struct btrfs_disk_key disk_key;
1802         struct extent_buffer *right = path->nodes[0];
1803         struct extent_buffer *left;
1804         struct btrfs_fs_info *fs_info = root->fs_info;
1805         int slot;
1806         int i;
1807         int free_space;
1808         int push_space = 0;
1809         int push_items = 0;
1810         struct btrfs_item *item;
1811         u32 old_left_nritems;
1812         u32 right_nritems;
1813         u32 nr;
1814         int ret = 0;
1815         u32 this_item_size;
1816         u32 old_left_item_size;
1817
1818         slot = path->slots[1];
1819         if (slot == 0)
1820                 return 1;
1821         if (!path->nodes[1])
1822                 return 1;
1823
1824         right_nritems = btrfs_header_nritems(right);
1825         if (right_nritems == 0) {
1826                 return 1;
1827         }
1828
1829         left = read_node_slot(fs_info, path->nodes[1], slot - 1);
1830         free_space = btrfs_leaf_free_space(root, left);
1831         if (free_space < data_size) {
1832                 free_extent_buffer(left);
1833                 return 1;
1834         }
1835
1836         /* cow and double check */
1837         ret = btrfs_cow_block(trans, root, left,
1838                               path->nodes[1], slot - 1, &left);
1839         if (ret) {
1840                 /* we hit -ENOSPC, but it isn't fatal here */
1841                 free_extent_buffer(left);
1842                 return 1;
1843         }
1844
1845         free_space = btrfs_leaf_free_space(root, left);
1846         if (free_space < data_size) {
1847                 free_extent_buffer(left);
1848                 return 1;
1849         }
1850
1851         if (empty)
1852                 nr = right_nritems;
1853         else
1854                 nr = right_nritems - 1;
1855
1856         for (i = 0; i < nr; i++) {
1857                 item = btrfs_item_nr(i);
1858
1859                 if (path->slots[0] == i)
1860                         push_space += data_size + sizeof(*item);
1861
1862                 this_item_size = btrfs_item_size(right, item);
1863                 if (this_item_size + sizeof(*item) + push_space > free_space)
1864                         break;
1865
1866                 push_items++;
1867                 push_space += this_item_size + sizeof(*item);
1868         }
1869
1870         if (push_items == 0) {
1871                 free_extent_buffer(left);
1872                 return 1;
1873         }
1874         if (!empty && push_items == btrfs_header_nritems(right))
1875                 WARN_ON(1);
1876
1877         /* push data from right to left */
1878         copy_extent_buffer(left, right,
1879                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1880                            btrfs_item_nr_offset(0),
1881                            push_items * sizeof(struct btrfs_item));
1882
1883         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1884                      btrfs_item_offset_nr(right, push_items -1);
1885
1886         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1887                      leaf_data_end(root, left) - push_space,
1888                      btrfs_leaf_data(right) +
1889                      btrfs_item_offset_nr(right, push_items - 1),
1890                      push_space);
1891         old_left_nritems = btrfs_header_nritems(left);
1892         BUG_ON(old_left_nritems == 0);
1893
1894         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1895         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1896                 u32 ioff;
1897
1898                 item = btrfs_item_nr(i);
1899                 ioff = btrfs_item_offset(left, item);
1900                 btrfs_set_item_offset(left, item,
1901                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1902         }
1903         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1904
1905         /* fixup right node */
1906         if (push_items > right_nritems) {
1907                 printk("push items %d nr %u\n", push_items, right_nritems);
1908                 WARN_ON(1);
1909         }
1910
1911         if (push_items < right_nritems) {
1912                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1913                                                   leaf_data_end(root, right);
1914                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
1915                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1916                                       btrfs_leaf_data(right) +
1917                                       leaf_data_end(root, right), push_space);
1918
1919                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1920                               btrfs_item_nr_offset(push_items),
1921                              (btrfs_header_nritems(right) - push_items) *
1922                              sizeof(struct btrfs_item));
1923         }
1924         right_nritems -= push_items;
1925         btrfs_set_header_nritems(right, right_nritems);
1926         push_space = BTRFS_LEAF_DATA_SIZE(root);
1927         for (i = 0; i < right_nritems; i++) {
1928                 item = btrfs_item_nr(i);
1929                 push_space = push_space - btrfs_item_size(right, item);
1930                 btrfs_set_item_offset(right, item, push_space);
1931         }
1932
1933         btrfs_mark_buffer_dirty(left);
1934         if (right_nritems)
1935                 btrfs_mark_buffer_dirty(right);
1936
1937         btrfs_item_key(right, &disk_key, 0);
1938         btrfs_fixup_low_keys(root, path, &disk_key, 1);
1939
1940         /* then fixup the leaf pointer in the path */
1941         if (path->slots[0] < push_items) {
1942                 path->slots[0] += old_left_nritems;
1943                 free_extent_buffer(path->nodes[0]);
1944                 path->nodes[0] = left;
1945                 path->slots[1] -= 1;
1946         } else {
1947                 free_extent_buffer(left);
1948                 path->slots[0] -= push_items;
1949         }
1950         BUG_ON(path->slots[0] < 0);
1951         return ret;
1952 }
1953
1954 /*
1955  * split the path's leaf in two, making sure there is at least data_size
1956  * available for the resulting leaf level of the path.
1957  *
1958  * returns 0 if all went well and < 0 on failure.
1959  */
1960 static noinline int copy_for_split(struct btrfs_trans_handle *trans,
1961                                struct btrfs_root *root,
1962                                struct btrfs_path *path,
1963                                struct extent_buffer *l,
1964                                struct extent_buffer *right,
1965                                int slot, int mid, int nritems)
1966 {
1967         int data_copy_size;
1968         int rt_data_off;
1969         int i;
1970         int ret = 0;
1971         int wret;
1972         struct btrfs_disk_key disk_key;
1973
1974         nritems = nritems - mid;
1975         btrfs_set_header_nritems(right, nritems);
1976         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
1977
1978         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
1979                            btrfs_item_nr_offset(mid),
1980                            nritems * sizeof(struct btrfs_item));
1981
1982         copy_extent_buffer(right, l,
1983                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
1984                      data_copy_size, btrfs_leaf_data(l) +
1985                      leaf_data_end(root, l), data_copy_size);
1986
1987         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
1988                       btrfs_item_end_nr(l, mid);
1989
1990         for (i = 0; i < nritems; i++) {
1991                 struct btrfs_item *item = btrfs_item_nr(i);
1992                 u32 ioff = btrfs_item_offset(right, item);
1993                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
1994         }
1995
1996         btrfs_set_header_nritems(l, mid);
1997         ret = 0;
1998         btrfs_item_key(right, &disk_key, 0);
1999         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2000                           path->slots[1] + 1, 1);
2001         if (wret)
2002                 ret = wret;
2003
2004         btrfs_mark_buffer_dirty(right);
2005         btrfs_mark_buffer_dirty(l);
2006         BUG_ON(path->slots[0] != slot);
2007
2008         if (mid <= slot) {
2009                 free_extent_buffer(path->nodes[0]);
2010                 path->nodes[0] = right;
2011                 path->slots[0] -= mid;
2012                 path->slots[1] += 1;
2013         } else {
2014                 free_extent_buffer(right);
2015         }
2016
2017         BUG_ON(path->slots[0] < 0);
2018
2019         return ret;
2020 }
2021
2022 /*
2023  * split the path's leaf in two, making sure there is at least data_size
2024  * available for the resulting leaf level of the path.
2025  *
2026  * returns 0 if all went well and < 0 on failure.
2027  */
2028 static noinline int split_leaf(struct btrfs_trans_handle *trans,
2029                                struct btrfs_root *root,
2030                                struct btrfs_key *ins_key,
2031                                struct btrfs_path *path, int data_size,
2032                                int extend)
2033 {
2034         struct btrfs_disk_key disk_key;
2035         struct extent_buffer *l;
2036         u32 nritems;
2037         int mid;
2038         int slot;
2039         struct extent_buffer *right;
2040         int ret = 0;
2041         int wret;
2042         int split;
2043         int num_doubles = 0;
2044
2045         l = path->nodes[0];
2046         slot = path->slots[0];
2047         if (extend && data_size + btrfs_item_size_nr(l, slot) +
2048             sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(root))
2049                 return -EOVERFLOW;
2050
2051         /* first try to make some room by pushing left and right */
2052         if (data_size && ins_key->type != BTRFS_DIR_ITEM_KEY) {
2053                 wret = push_leaf_right(trans, root, path, data_size, 0);
2054                 if (wret < 0)
2055                         return wret;
2056                 if (wret) {
2057                         wret = push_leaf_left(trans, root, path, data_size, 0);
2058                         if (wret < 0)
2059                                 return wret;
2060                 }
2061                 l = path->nodes[0];
2062
2063                 /* did the pushes work? */
2064                 if (btrfs_leaf_free_space(root, l) >= data_size)
2065                         return 0;
2066         }
2067
2068         if (!path->nodes[1]) {
2069                 ret = insert_new_root(trans, root, path, 1);
2070                 if (ret)
2071                         return ret;
2072         }
2073 again:
2074         split = 1;
2075         l = path->nodes[0];
2076         slot = path->slots[0];
2077         nritems = btrfs_header_nritems(l);
2078         mid = (nritems + 1) / 2;
2079
2080         if (mid <= slot) {
2081                 if (nritems == 1 ||
2082                     leaf_space_used(l, mid, nritems - mid) + data_size >
2083                         BTRFS_LEAF_DATA_SIZE(root)) {
2084                         if (slot >= nritems) {
2085                                 split = 0;
2086                         } else {
2087                                 mid = slot;
2088                                 if (mid != nritems &&
2089                                     leaf_space_used(l, mid, nritems - mid) +
2090                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2091                                         split = 2;
2092                                 }
2093                         }
2094                 }
2095         } else {
2096                 if (leaf_space_used(l, 0, mid) + data_size >
2097                         BTRFS_LEAF_DATA_SIZE(root)) {
2098                         if (!extend && data_size && slot == 0) {
2099                                 split = 0;
2100                         } else if ((extend || !data_size) && slot == 0) {
2101                                 mid = 1;
2102                         } else {
2103                                 mid = slot;
2104                                 if (mid != nritems &&
2105                                     leaf_space_used(l, mid, nritems - mid) +
2106                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2107                                         split = 2 ;
2108                                 }
2109                         }
2110                 }
2111         }
2112
2113         if (split == 0)
2114                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2115         else
2116                 btrfs_item_key(l, &disk_key, mid);
2117
2118         right = btrfs_alloc_free_block(trans, root, root->fs_info->nodesize,
2119                                         root->root_key.objectid,
2120                                         &disk_key, 0, l->start, 0);
2121         if (IS_ERR(right)) {
2122                 BUG_ON(1);
2123                 return PTR_ERR(right);
2124         }
2125
2126         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2127         btrfs_set_header_bytenr(right, right->start);
2128         btrfs_set_header_generation(right, trans->transid);
2129         btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
2130         btrfs_set_header_owner(right, root->root_key.objectid);
2131         btrfs_set_header_level(right, 0);
2132         write_extent_buffer(right, root->fs_info->fsid,
2133                             btrfs_header_fsid(), BTRFS_FSID_SIZE);
2134
2135         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2136                             btrfs_header_chunk_tree_uuid(right),
2137                             BTRFS_UUID_SIZE);
2138
2139         if (split == 0) {
2140                 if (mid <= slot) {
2141                         btrfs_set_header_nritems(right, 0);
2142                         wret = insert_ptr(trans, root, path,
2143                                           &disk_key, right->start,
2144                                           path->slots[1] + 1, 1);
2145                         if (wret)
2146                                 ret = wret;
2147
2148                         free_extent_buffer(path->nodes[0]);
2149                         path->nodes[0] = right;
2150                         path->slots[0] = 0;
2151                         path->slots[1] += 1;
2152                 } else {
2153                         btrfs_set_header_nritems(right, 0);
2154                         wret = insert_ptr(trans, root, path,
2155                                           &disk_key,
2156                                           right->start,
2157                                           path->slots[1], 1);
2158                         if (wret)
2159                                 ret = wret;
2160                         free_extent_buffer(path->nodes[0]);
2161                         path->nodes[0] = right;
2162                         path->slots[0] = 0;
2163                         if (path->slots[1] == 0) {
2164                                 btrfs_fixup_low_keys(root, path,
2165                                                      &disk_key, 1);
2166                         }
2167                 }
2168                 btrfs_mark_buffer_dirty(right);
2169                 return ret;
2170         }
2171
2172         ret = copy_for_split(trans, root, path, l, right, slot, mid, nritems);
2173         BUG_ON(ret);
2174
2175         if (split == 2) {
2176                 BUG_ON(num_doubles != 0);
2177                 num_doubles++;
2178                 goto again;
2179         }
2180
2181         return ret;
2182 }
2183
2184 /*
2185  * This function splits a single item into two items,
2186  * giving 'new_key' to the new item and splitting the
2187  * old one at split_offset (from the start of the item).
2188  *
2189  * The path may be released by this operation.  After
2190  * the split, the path is pointing to the old item.  The
2191  * new item is going to be in the same node as the old one.
2192  *
2193  * Note, the item being split must be smaller enough to live alone on
2194  * a tree block with room for one extra struct btrfs_item
2195  *
2196  * This allows us to split the item in place, keeping a lock on the
2197  * leaf the entire time.
2198  */
2199 int btrfs_split_item(struct btrfs_trans_handle *trans,
2200                      struct btrfs_root *root,
2201                      struct btrfs_path *path,
2202                      struct btrfs_key *new_key,
2203                      unsigned long split_offset)
2204 {
2205         u32 item_size;
2206         struct extent_buffer *leaf;
2207         struct btrfs_key orig_key;
2208         struct btrfs_item *item;
2209         struct btrfs_item *new_item;
2210         int ret = 0;
2211         int slot;
2212         u32 nritems;
2213         u32 orig_offset;
2214         struct btrfs_disk_key disk_key;
2215         char *buf;
2216
2217         leaf = path->nodes[0];
2218         btrfs_item_key_to_cpu(leaf, &orig_key, path->slots[0]);
2219         if (btrfs_leaf_free_space(root, leaf) >= sizeof(struct btrfs_item))
2220                 goto split;
2221
2222         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2223         btrfs_release_path(path);
2224
2225         path->search_for_split = 1;
2226
2227         ret = btrfs_search_slot(trans, root, &orig_key, path, 0, 1);
2228         path->search_for_split = 0;
2229
2230         /* if our item isn't there or got smaller, return now */
2231         if (ret != 0 || item_size != btrfs_item_size_nr(path->nodes[0],
2232                                                         path->slots[0])) {
2233                 return -EAGAIN;
2234         }
2235
2236         ret = split_leaf(trans, root, &orig_key, path, 0, 0);
2237         BUG_ON(ret);
2238
2239         BUG_ON(btrfs_leaf_free_space(root, leaf) < sizeof(struct btrfs_item));
2240         leaf = path->nodes[0];
2241
2242 split:
2243         item = btrfs_item_nr(path->slots[0]);
2244         orig_offset = btrfs_item_offset(leaf, item);
2245         item_size = btrfs_item_size(leaf, item);
2246
2247
2248         buf = kmalloc(item_size, GFP_NOFS);
2249         BUG_ON(!buf);
2250         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
2251                             path->slots[0]), item_size);
2252         slot = path->slots[0] + 1;
2253         leaf = path->nodes[0];
2254
2255         nritems = btrfs_header_nritems(leaf);
2256
2257         if (slot < nritems) {
2258                 /* shift the items */
2259                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
2260                               btrfs_item_nr_offset(slot),
2261                               (nritems - slot) * sizeof(struct btrfs_item));
2262
2263         }
2264
2265         btrfs_cpu_key_to_disk(&disk_key, new_key);
2266         btrfs_set_item_key(leaf, &disk_key, slot);
2267
2268         new_item = btrfs_item_nr(slot);
2269
2270         btrfs_set_item_offset(leaf, new_item, orig_offset);
2271         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
2272
2273         btrfs_set_item_offset(leaf, item,
2274                               orig_offset + item_size - split_offset);
2275         btrfs_set_item_size(leaf, item, split_offset);
2276
2277         btrfs_set_header_nritems(leaf, nritems + 1);
2278
2279         /* write the data for the start of the original item */
2280         write_extent_buffer(leaf, buf,
2281                             btrfs_item_ptr_offset(leaf, path->slots[0]),
2282                             split_offset);
2283
2284         /* write the data for the new item */
2285         write_extent_buffer(leaf, buf + split_offset,
2286                             btrfs_item_ptr_offset(leaf, slot),
2287                             item_size - split_offset);
2288         btrfs_mark_buffer_dirty(leaf);
2289
2290         ret = 0;
2291         if (btrfs_leaf_free_space(root, leaf) < 0) {
2292                 btrfs_print_leaf(root, leaf);
2293                 BUG();
2294         }
2295         kfree(buf);
2296         return ret;
2297 }
2298
2299 int btrfs_truncate_item(struct btrfs_root *root, struct btrfs_path *path,
2300                         u32 new_size, int from_end)
2301 {
2302         int ret = 0;
2303         int slot;
2304         struct extent_buffer *leaf;
2305         struct btrfs_item *item;
2306         u32 nritems;
2307         unsigned int data_end;
2308         unsigned int old_data_start;
2309         unsigned int old_size;
2310         unsigned int size_diff;
2311         int i;
2312
2313         leaf = path->nodes[0];
2314         slot = path->slots[0];
2315
2316         old_size = btrfs_item_size_nr(leaf, slot);
2317         if (old_size == new_size)
2318                 return 0;
2319
2320         nritems = btrfs_header_nritems(leaf);
2321         data_end = leaf_data_end(root, leaf);
2322
2323         old_data_start = btrfs_item_offset_nr(leaf, slot);
2324
2325         size_diff = old_size - new_size;
2326
2327         BUG_ON(slot < 0);
2328         BUG_ON(slot >= nritems);
2329
2330         /*
2331          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2332          */
2333         /* first correct the data pointers */
2334         for (i = slot; i < nritems; i++) {
2335                 u32 ioff;
2336                 item = btrfs_item_nr(i);
2337                 ioff = btrfs_item_offset(leaf, item);
2338                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2339         }
2340
2341         /* shift the data */
2342         if (from_end) {
2343                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2344                               data_end + size_diff, btrfs_leaf_data(leaf) +
2345                               data_end, old_data_start + new_size - data_end);
2346         } else {
2347                 struct btrfs_disk_key disk_key;
2348                 u64 offset;
2349
2350                 btrfs_item_key(leaf, &disk_key, slot);
2351
2352                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2353                         unsigned long ptr;
2354                         struct btrfs_file_extent_item *fi;
2355
2356                         fi = btrfs_item_ptr(leaf, slot,
2357                                             struct btrfs_file_extent_item);
2358                         fi = (struct btrfs_file_extent_item *)(
2359                              (unsigned long)fi - size_diff);
2360
2361                         if (btrfs_file_extent_type(leaf, fi) ==
2362                             BTRFS_FILE_EXTENT_INLINE) {
2363                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2364                                 memmove_extent_buffer(leaf, ptr,
2365                                         (unsigned long)fi,
2366                                         offsetof(struct btrfs_file_extent_item,
2367                                                  disk_bytenr));
2368                         }
2369                 }
2370
2371                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2372                               data_end + size_diff, btrfs_leaf_data(leaf) +
2373                               data_end, old_data_start - data_end);
2374
2375                 offset = btrfs_disk_key_offset(&disk_key);
2376                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2377                 btrfs_set_item_key(leaf, &disk_key, slot);
2378                 if (slot == 0)
2379                         btrfs_fixup_low_keys(root, path, &disk_key, 1);
2380         }
2381
2382         item = btrfs_item_nr(slot);
2383         btrfs_set_item_size(leaf, item, new_size);
2384         btrfs_mark_buffer_dirty(leaf);
2385
2386         ret = 0;
2387         if (btrfs_leaf_free_space(root, leaf) < 0) {
2388                 btrfs_print_leaf(root, leaf);
2389                 BUG();
2390         }
2391         return ret;
2392 }
2393
2394 int btrfs_extend_item(struct btrfs_root *root, struct btrfs_path *path,
2395                       u32 data_size)
2396 {
2397         int ret = 0;
2398         int slot;
2399         struct extent_buffer *leaf;
2400         struct btrfs_item *item;
2401         u32 nritems;
2402         unsigned int data_end;
2403         unsigned int old_data;
2404         unsigned int old_size;
2405         int i;
2406
2407         leaf = path->nodes[0];
2408
2409         nritems = btrfs_header_nritems(leaf);
2410         data_end = leaf_data_end(root, leaf);
2411
2412         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2413                 btrfs_print_leaf(root, leaf);
2414                 BUG();
2415         }
2416         slot = path->slots[0];
2417         old_data = btrfs_item_end_nr(leaf, slot);
2418
2419         BUG_ON(slot < 0);
2420         if (slot >= nritems) {
2421                 btrfs_print_leaf(root, leaf);
2422                 printk("slot %d too large, nritems %d\n", slot, nritems);
2423                 BUG_ON(1);
2424         }
2425
2426         /*
2427          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2428          */
2429         /* first correct the data pointers */
2430         for (i = slot; i < nritems; i++) {
2431                 u32 ioff;
2432                 item = btrfs_item_nr(i);
2433                 ioff = btrfs_item_offset(leaf, item);
2434                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2435         }
2436
2437         /* shift the data */
2438         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2439                       data_end - data_size, btrfs_leaf_data(leaf) +
2440                       data_end, old_data - data_end);
2441
2442         data_end = old_data;
2443         old_size = btrfs_item_size_nr(leaf, slot);
2444         item = btrfs_item_nr(slot);
2445         btrfs_set_item_size(leaf, item, old_size + data_size);
2446         btrfs_mark_buffer_dirty(leaf);
2447
2448         ret = 0;
2449         if (btrfs_leaf_free_space(root, leaf) < 0) {
2450                 btrfs_print_leaf(root, leaf);
2451                 BUG();
2452         }
2453         return ret;
2454 }
2455
2456 /*
2457  * Given a key and some data, insert an item into the tree.
2458  * This does all the path init required, making room in the tree if needed.
2459  */
2460 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2461                             struct btrfs_root *root,
2462                             struct btrfs_path *path,
2463                             struct btrfs_key *cpu_key, u32 *data_size,
2464                             int nr)
2465 {
2466         struct extent_buffer *leaf;
2467         struct btrfs_item *item;
2468         int ret = 0;
2469         int slot;
2470         int i;
2471         u32 nritems;
2472         u32 total_size = 0;
2473         u32 total_data = 0;
2474         unsigned int data_end;
2475         struct btrfs_disk_key disk_key;
2476
2477         for (i = 0; i < nr; i++) {
2478                 total_data += data_size[i];
2479         }
2480
2481         /* create a root if there isn't one */
2482         if (!root->node)
2483                 BUG();
2484
2485         total_size = total_data + nr * sizeof(struct btrfs_item);
2486         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2487         if (ret == 0) {
2488                 return -EEXIST;
2489         }
2490         if (ret < 0)
2491                 goto out;
2492
2493         leaf = path->nodes[0];
2494
2495         nritems = btrfs_header_nritems(leaf);
2496         data_end = leaf_data_end(root, leaf);
2497
2498         if (btrfs_leaf_free_space(root, leaf) < total_size) {
2499                 btrfs_print_leaf(root, leaf);
2500                 printk("not enough freespace need %u have %d\n",
2501                        total_size, btrfs_leaf_free_space(root, leaf));
2502                 BUG();
2503         }
2504
2505         slot = path->slots[0];
2506         BUG_ON(slot < 0);
2507
2508         if (slot < nritems) {
2509                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2510
2511                 if (old_data < data_end) {
2512                         btrfs_print_leaf(root, leaf);
2513                         printk("slot %d old_data %d data_end %d\n",
2514                                slot, old_data, data_end);
2515                         BUG_ON(1);
2516                 }
2517                 /*
2518                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2519                  */
2520                 /* first correct the data pointers */
2521                 for (i = slot; i < nritems; i++) {
2522                         u32 ioff;
2523
2524                         item = btrfs_item_nr(i);
2525                         ioff = btrfs_item_offset(leaf, item);
2526                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2527                 }
2528
2529                 /* shift the items */
2530                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2531                               btrfs_item_nr_offset(slot),
2532                               (nritems - slot) * sizeof(struct btrfs_item));
2533
2534                 /* shift the data */
2535                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2536                               data_end - total_data, btrfs_leaf_data(leaf) +
2537                               data_end, old_data - data_end);
2538                 data_end = old_data;
2539         }
2540
2541         /* setup the item for the new data */
2542         for (i = 0; i < nr; i++) {
2543                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2544                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2545                 item = btrfs_item_nr(slot + i);
2546                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2547                 data_end -= data_size[i];
2548                 btrfs_set_item_size(leaf, item, data_size[i]);
2549         }
2550         btrfs_set_header_nritems(leaf, nritems + nr);
2551         btrfs_mark_buffer_dirty(leaf);
2552
2553         ret = 0;
2554         if (slot == 0) {
2555                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2556                 btrfs_fixup_low_keys(root, path, &disk_key, 1);
2557         }
2558
2559         if (btrfs_leaf_free_space(root, leaf) < 0) {
2560                 btrfs_print_leaf(root, leaf);
2561                 BUG();
2562         }
2563
2564 out:
2565         return ret;
2566 }
2567
2568 /*
2569  * Given a key and some data, insert an item into the tree.
2570  * This does all the path init required, making room in the tree if needed.
2571  */
2572 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2573                       *root, struct btrfs_key *cpu_key, void *data, u32
2574                       data_size)
2575 {
2576         int ret = 0;
2577         struct btrfs_path *path;
2578         struct extent_buffer *leaf;
2579         unsigned long ptr;
2580
2581         path = btrfs_alloc_path();
2582         if (!path)
2583                 return -ENOMEM;
2584
2585         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2586         if (!ret) {
2587                 leaf = path->nodes[0];
2588                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2589                 write_extent_buffer(leaf, data, ptr, data_size);
2590                 btrfs_mark_buffer_dirty(leaf);
2591         }
2592         btrfs_free_path(path);
2593         return ret;
2594 }
2595
2596 /*
2597  * delete the pointer from a given node.
2598  *
2599  * If the delete empties a node, the node is removed from the tree,
2600  * continuing all the way the root if required.  The root is converted into
2601  * a leaf if all the nodes are emptied.
2602  */
2603 int btrfs_del_ptr(struct btrfs_root *root, struct btrfs_path *path,
2604                 int level, int slot)
2605 {
2606         struct extent_buffer *parent = path->nodes[level];
2607         u32 nritems;
2608         int ret = 0;
2609
2610         nritems = btrfs_header_nritems(parent);
2611         if (slot < nritems - 1) {
2612                 /* shift the items */
2613                 memmove_extent_buffer(parent,
2614                               btrfs_node_key_ptr_offset(slot),
2615                               btrfs_node_key_ptr_offset(slot + 1),
2616                               sizeof(struct btrfs_key_ptr) *
2617                               (nritems - slot - 1));
2618         }
2619         nritems--;
2620         btrfs_set_header_nritems(parent, nritems);
2621         if (nritems == 0 && parent == root->node) {
2622                 BUG_ON(btrfs_header_level(root->node) != 1);
2623                 /* just turn the root into a leaf and break */
2624                 btrfs_set_header_level(root->node, 0);
2625         } else if (slot == 0) {
2626                 struct btrfs_disk_key disk_key;
2627
2628                 btrfs_node_key(parent, &disk_key, 0);
2629                 btrfs_fixup_low_keys(root, path, &disk_key, level + 1);
2630         }
2631         btrfs_mark_buffer_dirty(parent);
2632         return ret;
2633 }
2634
2635 /*
2636  * a helper function to delete the leaf pointed to by path->slots[1] and
2637  * path->nodes[1].
2638  *
2639  * This deletes the pointer in path->nodes[1] and frees the leaf
2640  * block extent.  zero is returned if it all worked out, < 0 otherwise.
2641  *
2642  * The path must have already been setup for deleting the leaf, including
2643  * all the proper balancing.  path->nodes[1] must be locked.
2644  */
2645 static noinline int btrfs_del_leaf(struct btrfs_trans_handle *trans,
2646                                    struct btrfs_root *root,
2647                                    struct btrfs_path *path,
2648                                    struct extent_buffer *leaf)
2649 {
2650         int ret;
2651
2652         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
2653         ret = btrfs_del_ptr(root, path, 1, path->slots[1]);
2654         if (ret)
2655                 return ret;
2656
2657         ret = btrfs_free_extent(trans, root, leaf->start, leaf->len,
2658                                 0, root->root_key.objectid, 0, 0);
2659         return ret;
2660 }
2661
2662 /*
2663  * delete the item at the leaf level in path.  If that empties
2664  * the leaf, remove it from the tree
2665  */
2666 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2667                     struct btrfs_path *path, int slot, int nr)
2668 {
2669         struct extent_buffer *leaf;
2670         struct btrfs_item *item;
2671         int last_off;
2672         int dsize = 0;
2673         int ret = 0;
2674         int wret;
2675         int i;
2676         u32 nritems;
2677
2678         leaf = path->nodes[0];
2679         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2680
2681         for (i = 0; i < nr; i++)
2682                 dsize += btrfs_item_size_nr(leaf, slot + i);
2683
2684         nritems = btrfs_header_nritems(leaf);
2685
2686         if (slot + nr != nritems) {
2687                 int data_end = leaf_data_end(root, leaf);
2688
2689                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2690                               data_end + dsize,
2691                               btrfs_leaf_data(leaf) + data_end,
2692                               last_off - data_end);
2693
2694                 for (i = slot + nr; i < nritems; i++) {
2695                         u32 ioff;
2696
2697                         item = btrfs_item_nr(i);
2698                         ioff = btrfs_item_offset(leaf, item);
2699                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2700                 }
2701
2702                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2703                               btrfs_item_nr_offset(slot + nr),
2704                               sizeof(struct btrfs_item) *
2705                               (nritems - slot - nr));
2706         }
2707         btrfs_set_header_nritems(leaf, nritems - nr);
2708         nritems -= nr;
2709
2710         /* delete the leaf if we've emptied it */
2711         if (nritems == 0) {
2712                 if (leaf == root->node) {
2713                         btrfs_set_header_level(leaf, 0);
2714                 } else {
2715                         clean_tree_block(trans, root, leaf);
2716                         wret = btrfs_del_leaf(trans, root, path, leaf);
2717                         BUG_ON(ret);
2718                         if (wret)
2719                                 ret = wret;
2720                 }
2721         } else {
2722                 int used = leaf_space_used(leaf, 0, nritems);
2723                 if (slot == 0) {
2724                         struct btrfs_disk_key disk_key;
2725
2726                         btrfs_item_key(leaf, &disk_key, 0);
2727                         btrfs_fixup_low_keys(root, path, &disk_key, 1);
2728                 }
2729
2730                 /* delete the leaf if it is mostly empty */
2731                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2732                         /* push_leaf_left fixes the path.
2733                          * make sure the path still points to our leaf
2734                          * for possible call to del_ptr below
2735                          */
2736                         slot = path->slots[1];
2737                         extent_buffer_get(leaf);
2738
2739                         wret = push_leaf_left(trans, root, path, 1, 1);
2740                         if (wret < 0 && wret != -ENOSPC)
2741                                 ret = wret;
2742
2743                         if (path->nodes[0] == leaf &&
2744                             btrfs_header_nritems(leaf)) {
2745                                 wret = push_leaf_right(trans, root, path, 1, 1);
2746                                 if (wret < 0 && wret != -ENOSPC)
2747                                         ret = wret;
2748                         }
2749
2750                         if (btrfs_header_nritems(leaf) == 0) {
2751                                 clean_tree_block(trans, root, leaf);
2752                                 path->slots[1] = slot;
2753                                 ret = btrfs_del_leaf(trans, root, path, leaf);
2754                                 BUG_ON(ret);
2755                                 free_extent_buffer(leaf);
2756
2757                         } else {
2758                                 btrfs_mark_buffer_dirty(leaf);
2759                                 free_extent_buffer(leaf);
2760                         }
2761                 } else {
2762                         btrfs_mark_buffer_dirty(leaf);
2763                 }
2764         }
2765         return ret;
2766 }
2767
2768 /*
2769  * walk up the tree as far as required to find the previous leaf.
2770  * returns 0 if it found something or 1 if there are no lesser leaves.
2771  * returns < 0 on io errors.
2772  */
2773 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2774 {
2775         int slot;
2776         int level = 1;
2777         struct extent_buffer *c;
2778         struct extent_buffer *next = NULL;
2779         struct btrfs_fs_info *fs_info = root->fs_info;
2780
2781         while(level < BTRFS_MAX_LEVEL) {
2782                 if (!path->nodes[level])
2783                         return 1;
2784
2785                 slot = path->slots[level];
2786                 c = path->nodes[level];
2787                 if (slot == 0) {
2788                         level++;
2789                         if (level == BTRFS_MAX_LEVEL)
2790                                 return 1;
2791                         continue;
2792                 }
2793                 slot--;
2794
2795                 next = read_node_slot(fs_info, c, slot);
2796                 if (!extent_buffer_uptodate(next)) {
2797                         if (IS_ERR(next))
2798                                 return PTR_ERR(next);
2799                         return -EIO;
2800                 }
2801                 break;
2802         }
2803         path->slots[level] = slot;
2804         while(1) {
2805                 level--;
2806                 c = path->nodes[level];
2807                 free_extent_buffer(c);
2808                 slot = btrfs_header_nritems(next);
2809                 if (slot != 0)
2810                         slot--;
2811                 path->nodes[level] = next;
2812                 path->slots[level] = slot;
2813                 if (!level)
2814                         break;
2815                 next = read_node_slot(fs_info, next, slot);
2816                 if (!extent_buffer_uptodate(next)) {
2817                         if (IS_ERR(next))
2818                                 return PTR_ERR(next);
2819                         return -EIO;
2820                 }
2821         }
2822         return 0;
2823 }
2824
2825 /*
2826  * walk up the tree as far as required to find the next leaf.
2827  * returns 0 if it found something or 1 if there are no greater leaves.
2828  * returns < 0 on io errors.
2829  */
2830 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2831 {
2832         int slot;
2833         int level = 1;
2834         struct extent_buffer *c;
2835         struct extent_buffer *next = NULL;
2836         struct btrfs_fs_info *fs_info = root->fs_info;
2837
2838         while(level < BTRFS_MAX_LEVEL) {
2839                 if (!path->nodes[level])
2840                         return 1;
2841
2842                 slot = path->slots[level] + 1;
2843                 c = path->nodes[level];
2844                 if (slot >= btrfs_header_nritems(c)) {
2845                         level++;
2846                         if (level == BTRFS_MAX_LEVEL)
2847                                 return 1;
2848                         continue;
2849                 }
2850
2851                 if (path->reada)
2852                         reada_for_search(root, path, level, slot, 0);
2853
2854                 next = read_node_slot(fs_info, c, slot);
2855                 if (!extent_buffer_uptodate(next))
2856                         return -EIO;
2857                 break;
2858         }
2859         path->slots[level] = slot;
2860         while(1) {
2861                 level--;
2862                 c = path->nodes[level];
2863                 free_extent_buffer(c);
2864                 path->nodes[level] = next;
2865                 path->slots[level] = 0;
2866                 if (!level)
2867                         break;
2868                 if (path->reada)
2869                         reada_for_search(root, path, level, 0, 0);
2870                 next = read_node_slot(fs_info, next, 0);
2871                 if (!extent_buffer_uptodate(next))
2872                         return -EIO;
2873         }
2874         return 0;
2875 }
2876
2877 int btrfs_previous_item(struct btrfs_root *root,
2878                         struct btrfs_path *path, u64 min_objectid,
2879                         int type)
2880 {
2881         struct btrfs_key found_key;
2882         struct extent_buffer *leaf;
2883         u32 nritems;
2884         int ret;
2885
2886         while(1) {
2887                 if (path->slots[0] == 0) {
2888                         ret = btrfs_prev_leaf(root, path);
2889                         if (ret != 0)
2890                                 return ret;
2891                 } else {
2892                         path->slots[0]--;
2893                 }
2894                 leaf = path->nodes[0];
2895                 nritems = btrfs_header_nritems(leaf);
2896                 if (nritems == 0)
2897                         return 1;
2898                 if (path->slots[0] == nritems)
2899                         path->slots[0]--;
2900
2901                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2902                 if (found_key.objectid < min_objectid)
2903                         break;
2904                 if (found_key.type == type)
2905                         return 0;
2906                 if (found_key.objectid == min_objectid &&
2907                     found_key.type < type)
2908                         break;
2909         }
2910         return 1;
2911 }
2912
2913 /*
2914  * search in extent tree to find a previous Metadata/Data extent item with
2915  * min objecitd.
2916  *
2917  * returns 0 if something is found, 1 if nothing was found and < 0 on error
2918  */
2919 int btrfs_previous_extent_item(struct btrfs_root *root,
2920                         struct btrfs_path *path, u64 min_objectid)
2921 {
2922         struct btrfs_key found_key;
2923         struct extent_buffer *leaf;
2924         u32 nritems;
2925         int ret;
2926
2927         while (1) {
2928                 if (path->slots[0] == 0) {
2929                         ret = btrfs_prev_leaf(root, path);
2930                         if (ret != 0)
2931                                 return ret;
2932                 } else {
2933                         path->slots[0]--;
2934                 }
2935                 leaf = path->nodes[0];
2936                 nritems = btrfs_header_nritems(leaf);
2937                 if (nritems == 0)
2938                         return 1;
2939                 if (path->slots[0] == nritems)
2940                         path->slots[0]--;
2941
2942                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2943                 if (found_key.objectid < min_objectid)
2944                         break;
2945                 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
2946                     found_key.type == BTRFS_METADATA_ITEM_KEY)
2947                         return 0;
2948                 if (found_key.objectid == min_objectid &&
2949                     found_key.type < BTRFS_EXTENT_ITEM_KEY)
2950                         break;
2951         }
2952         return 1;
2953 }
2954
2955 /*
2956  * Search in extent tree to found next meta/data extent
2957  * Caller needs to check for no-hole or skinny metadata features.
2958  */
2959 int btrfs_next_extent_item(struct btrfs_root *root,
2960                         struct btrfs_path *path, u64 max_objectid)
2961 {
2962         struct btrfs_key found_key;
2963         int ret;
2964
2965         while (1) {
2966                 ret = btrfs_next_item(root, path);
2967                 if (ret)
2968                         return ret;
2969                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2970                                       path->slots[0]);
2971                 if (found_key.objectid > max_objectid)
2972                         return 1;
2973                 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
2974                     found_key.type == BTRFS_METADATA_ITEM_KEY)
2975                 return 0;
2976         }
2977 }