ctree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include "ctree.h"
  19 #include "disk-io.h"
  20 #include "transaction.h"
  21 #include "print-tree.h"
  22 #include "repair.h"
  23
  24 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  25                       *root, struct btrfs_path *path, int level);
  26 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  27                       *root, struct btrfs_key *ins_key,
  28                       struct btrfs_path *path, int data_size, int extend);
  29 static int push_node_left(struct btrfs_trans_handle *trans,
  30                           struct btrfs_root *root, struct extent_buffer *dst,
  31                           struct extent_buffer *src, int empty);
  32 static int balance_node_right(struct btrfs_trans_handle *trans,
  33                               struct btrfs_root *root,
  34                               struct extent_buffer *dst_buf,
  35                               struct extent_buffer *src_buf);
  36
  37 inline void btrfs_init_path(struct btrfs_path *p)
  38 {
  39         memset(p, 0, sizeof(*p));
  40 }
  41
  42 struct btrfs_path *btrfs_alloc_path(void)
  43 {
  44         struct btrfs_path *path;
  45         path = kmalloc(sizeof(struct btrfs_path), GFP_NOFS);
  46         if (path) {
  47                 btrfs_init_path(path);
  48                 path->reada = 0;
  49         }
  50         return path;
  51 }
  52
  53 void btrfs_free_path(struct btrfs_path *p)
  54 {
  55         btrfs_release_path(NULL, p);
  56         kfree(p);
  57 }
  58
  59 void btrfs_release_path(struct btrfs_root *root, 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 static 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                             (unsigned long)btrfs_header_fsid(cow),
 126                             BTRFS_FSID_SIZE);
 127
 128         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 129         ret = btrfs_inc_ref(trans, new_root, cow, 0);
 130         kfree(new_root);
 131
 132         if (ret)
 133                 return ret;
 134
 135         btrfs_mark_buffer_dirty(cow);
 136         *cow_ret = cow;
 137         return 0;
 138 }
 139
 140 int btrfs_fsck_reinit_root(struct btrfs_trans_handle *trans,
 141                            struct btrfs_root *root, int overwrite)
 142 {
 143         struct extent_buffer *c;
 144         struct extent_buffer *old = root->node;
 145         int level;
 146         struct btrfs_disk_key disk_key = {0,0,0};
 147
 148         level = 0;
 149
 150         if (overwrite) {
 151                 c = old;
 152                 extent_buffer_get(c);
 153                 goto init;
 154         }
 155         c = btrfs_alloc_free_block(trans, root,
 156                                    btrfs_level_size(root, 0),
 157                                    root->root_key.objectid,
 158                                    &disk_key, level, 0, 0);
 159         if (IS_ERR(c)) {
 160                 c = old;
 161                 extent_buffer_get(c);
 162         }
 163 init:
 164         memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
 165         btrfs_set_header_level(c, level);
 166         btrfs_set_header_bytenr(c, c->start);
 167         btrfs_set_header_generation(c, trans->transid);
 168         btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
 169         btrfs_set_header_owner(c, root->root_key.objectid);
 170
 171         write_extent_buffer(c, root->fs_info->fsid,
 172                             (unsigned long)btrfs_header_fsid(c),
 173                             BTRFS_FSID_SIZE);
 174
 175         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
 176                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
 177                             BTRFS_UUID_SIZE);
 178
 179         btrfs_mark_buffer_dirty(c);
 180
 181         free_extent_buffer(old);
 182         root->node = c;
 183         add_root_to_dirty_list(root);
 184         return 0;
 185 }
 186
 187 /*
 188  * check if the tree block can be shared by multiple trees
 189  */
 190 int btrfs_block_can_be_shared(struct btrfs_root *root,
 191                               struct extent_buffer *buf)
 192 {
 193         /*
 194          * Tree blocks not in refernece counted trees and tree roots
 195          * are never shared. If a block was allocated after the last
 196          * snapshot and the block was not allocated by tree relocation,
 197          * we know the block is not shared.
 198          */
 199         if (root->ref_cows &&
 200             buf != root->node && buf != root->commit_root &&
 201             (btrfs_header_generation(buf) <=
 202              btrfs_root_last_snapshot(&root->root_item) ||
 203              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 204                 return 1;
 205 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 206         if (root->ref_cows &&
 207             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 208                 return 1;
 209 #endif
 210         return 0;
 211 }
 212
 213 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
 214                                        struct btrfs_root *root,
 215                                        struct extent_buffer *buf,
 216                                        struct extent_buffer *cow)
 217 {
 218         u64 refs;
 219         u64 owner;
 220         u64 flags;
 221         u64 new_flags = 0;
 222         int ret;
 223
 224         /*
 225          * Backrefs update rules:
 226          *
 227          * Always use full backrefs for extent pointers in tree block
 228          * allocated by tree relocation.
 229          *
 230          * If a shared tree block is no longer referenced by its owner
 231          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
 232          * use full backrefs for extent pointers in tree block.
 233          *
 234          * If a tree block is been relocating
 235          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
 236          * use full backrefs for extent pointers in tree block.
 237          * The reason for this is some operations (such as drop tree)
 238          * are only allowed for blocks use full backrefs.
 239          */
 240
 241         if (btrfs_block_can_be_shared(root, buf)) {
 242                 ret = btrfs_lookup_extent_info(trans, root, buf->start,
 243                                                btrfs_header_level(buf), 1,
 244                                                &refs, &flags);
 245                 BUG_ON(ret);
 246                 BUG_ON(refs == 0);
 247         } else {
 248                 refs = 1;
 249                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 250                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
 251                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 252                 else
 253                         flags = 0;
 254         }
 255
 256         owner = btrfs_header_owner(buf);
 257         BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) &&
 258                owner == BTRFS_TREE_RELOC_OBJECTID);
 259
 260         if (refs > 1) {
 261                 if ((owner == root->root_key.objectid ||
 262                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
 263                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
 264                         ret = btrfs_inc_ref(trans, root, buf, 1);
 265                         BUG_ON(ret);
 266
 267                         if (root->root_key.objectid ==
 268                             BTRFS_TREE_RELOC_OBJECTID) {
 269                                 ret = btrfs_dec_ref(trans, root, buf, 0);
 270                                 BUG_ON(ret);
 271                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 272                                 BUG_ON(ret);
 273                         }
 274                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
 275                 } else {
 276
 277                         if (root->root_key.objectid ==
 278                             BTRFS_TREE_RELOC_OBJECTID)
 279                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 280                         else
 281                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 282                         BUG_ON(ret);
 283                 }
 284                 if (new_flags != 0) {
 285                         ret = btrfs_set_block_flags(trans, root, buf->start,
 286                                                     btrfs_header_level(buf),
 287                                                     new_flags);
 288                         BUG_ON(ret);
 289                 }
 290         } else {
 291                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
 292                         if (root->root_key.objectid ==
 293                             BTRFS_TREE_RELOC_OBJECTID)
 294                                 ret = btrfs_inc_ref(trans, root, cow, 1);
 295                         else
 296                                 ret = btrfs_inc_ref(trans, root, cow, 0);
 297                         BUG_ON(ret);
 298                         ret = btrfs_dec_ref(trans, root, buf, 1);
 299                         BUG_ON(ret);
 300                 }
 301                 clean_tree_block(trans, root, buf);
 302         }
 303         return 0;
 304 }
 305
 306 int __btrfs_cow_block(struct btrfs_trans_handle *trans,
 307                              struct btrfs_root *root,
 308                              struct extent_buffer *buf,
 309                              struct extent_buffer *parent, int parent_slot,
 310                              struct extent_buffer **cow_ret,
 311                              u64 search_start, u64 empty_size)
 312 {
 313         struct extent_buffer *cow;
 314         struct btrfs_disk_key disk_key;
 315         int level;
 316
 317         WARN_ON(root->ref_cows && trans->transid !=
 318                 root->fs_info->running_transaction->transid);
 319         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
 320
 321         level = btrfs_header_level(buf);
 322
 323         if (level == 0)
 324                 btrfs_item_key(buf, &disk_key, 0);
 325         else
 326                 btrfs_node_key(buf, &disk_key, 0);
 327
 328         cow = btrfs_alloc_free_block(trans, root, buf->len,
 329                                      root->root_key.objectid, &disk_key,
 330                                      level, search_start, empty_size);
 331         if (IS_ERR(cow))
 332                 return PTR_ERR(cow);
 333
 334         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 335         btrfs_set_header_bytenr(cow, cow->start);
 336         btrfs_set_header_generation(cow, trans->transid);
 337         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
 338         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
 339                                      BTRFS_HEADER_FLAG_RELOC);
 340         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
 341                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
 342         else
 343                 btrfs_set_header_owner(cow, root->root_key.objectid);
 344
 345         write_extent_buffer(cow, root->fs_info->fsid,
 346                             (unsigned long)btrfs_header_fsid(cow),
 347                             BTRFS_FSID_SIZE);
 348
 349         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 350
 351         update_ref_for_cow(trans, root, buf, cow);
 352
 353         if (buf == root->node) {
 354                 root->node = cow;
 355                 extent_buffer_get(cow);
 356
 357                 btrfs_free_extent(trans, root, buf->start, buf->len,
 358                                   0, root->root_key.objectid, level, 0);
 359                 free_extent_buffer(buf);
 360                 add_root_to_dirty_list(root);
 361         } else {
 362                 btrfs_set_node_blockptr(parent, parent_slot,
 363                                         cow->start);
 364                 WARN_ON(trans->transid == 0);
 365                 btrfs_set_node_ptr_generation(parent, parent_slot,
 366                                               trans->transid);
 367                 btrfs_mark_buffer_dirty(parent);
 368                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 369
 370                 btrfs_free_extent(trans, root, buf->start, buf->len,
 371                                   0, root->root_key.objectid, level, 1);
 372         }
 373         free_extent_buffer(buf);
 374         btrfs_mark_buffer_dirty(cow);
 375         *cow_ret = cow;
 376         return 0;
 377 }
 378
 379 static inline int should_cow_block(struct btrfs_trans_handle *trans,
 380                                    struct btrfs_root *root,
 381                                    struct extent_buffer *buf)
 382 {
 383         if (btrfs_header_generation(buf) == trans->transid &&
 384             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
 385             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
 386               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
 387                 return 0;
 388         return 1;
 389 }
 390
 391 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 392                     struct btrfs_root *root, struct extent_buffer *buf,
 393                     struct extent_buffer *parent, int parent_slot,
 394                     struct extent_buffer **cow_ret)
 395 {
 396         u64 search_start;
 397         int ret;
 398         /*
 399         if (trans->transaction != root->fs_info->running_transaction) {
 400                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 401                        root->fs_info->running_transaction->transid);
 402                 WARN_ON(1);
 403         }
 404         */
 405         if (trans->transid != root->fs_info->generation) {
 406                 printk(KERN_CRIT "trans %llu running %llu\n",
 407                         (unsigned long long)trans->transid,
 408                         (unsigned long long)root->fs_info->generation);
 409                 WARN_ON(1);
 410         }
 411         if (!should_cow_block(trans, root, buf)) {
 412                 *cow_ret = buf;
 413                 return 0;
 414         }
 415
 416         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
 417         ret = __btrfs_cow_block(trans, root, buf, parent,
 418                                  parent_slot, cow_ret, search_start, 0);
 419         return ret;
 420 }
 421
 422 /*
 423 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 424 {
 425         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 426                 return 1;
 427         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 428                 return 1;
 429         return 0;
 430 }
 431 */
 432
 433 /*
 434  * compare two keys in a memcmp fashion
 435  */
 436 int btrfs_comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 437 {
 438         struct btrfs_key k1;
 439
 440         btrfs_disk_key_to_cpu(&k1, disk);
 441
 442         if (k1.objectid > k2->objectid)
 443                 return 1;
 444         if (k1.objectid < k2->objectid)
 445                 return -1;
 446         if (k1.type > k2->type)
 447                 return 1;
 448         if (k1.type < k2->type)
 449                 return -1;
 450         if (k1.offset > k2->offset)
 451                 return 1;
 452         if (k1.offset < k2->offset)
 453                 return -1;
 454         return 0;
 455 }
 456
 457
 458 #if 0
 459 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 460                        struct btrfs_root *root, struct extent_buffer *parent,
 461                        int start_slot, int cache_only, u64 *last_ret,
 462                        struct btrfs_key *progress)
 463 {
 464         struct extent_buffer *cur;
 465         struct extent_buffer *tmp;
 466         u64 blocknr;
 467         u64 gen;
 468         u64 search_start = *last_ret;
 469         u64 last_block = 0;
 470         u64 other;
 471         u32 parent_nritems;
 472         int end_slot;
 473         int i;
 474         int err = 0;
 475         int parent_level;
 476         int uptodate;
 477         u32 blocksize;
 478         int progress_passed = 0;
 479         struct btrfs_disk_key disk_key;
 480
 481         parent_level = btrfs_header_level(parent);
 482         if (cache_only && parent_level != 1)
 483                 return 0;
 484
 485         if (trans->transaction != root->fs_info->running_transaction) {
 486                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 487                        root->fs_info->running_transaction->transid);
 488                 WARN_ON(1);
 489         }
 490         if (trans->transid != root->fs_info->generation) {
 491                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 492                        root->fs_info->generation);
 493                 WARN_ON(1);
 494         }
 495
 496         parent_nritems = btrfs_header_nritems(parent);
 497         blocksize = btrfs_level_size(root, parent_level - 1);
 498         end_slot = parent_nritems;
 499
 500         if (parent_nritems == 1)
 501                 return 0;
 502
 503         for (i = start_slot; i < end_slot; i++) {
 504                 int close = 1;
 505
 506                 if (!parent->map_token) {
 507                         map_extent_buffer(parent,
 508                                         btrfs_node_key_ptr_offset(i),
 509                                         sizeof(struct btrfs_key_ptr),
 510                                         &parent->map_token, &parent->kaddr,
 511                                         &parent->map_start, &parent->map_len,
 512                                         KM_USER1);
 513                 }
 514                 btrfs_node_key(parent, &disk_key, i);
 515                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
 516                         continue;
 517
 518                 progress_passed = 1;
 519                 blocknr = btrfs_node_blockptr(parent, i);
 520                 gen = btrfs_node_ptr_generation(parent, i);
 521                 if (last_block == 0)
 522                         last_block = blocknr;
 523
 524                 if (i > 0) {
 525                         other = btrfs_node_blockptr(parent, i - 1);
 526                         close = close_blocks(blocknr, other, blocksize);
 527                 }
 528                 if (close && i < end_slot - 2) {
 529                         other = btrfs_node_blockptr(parent, i + 1);
 530                         close = close_blocks(blocknr, other, blocksize);
 531                 }
 532                 if (close) {
 533                         last_block = blocknr;
 534                         continue;
 535                 }
 536                 if (parent->map_token) {
 537                         unmap_extent_buffer(parent, parent->map_token,
 538                                             KM_USER1);
 539                         parent->map_token = NULL;
 540                 }
 541
 542                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
 543                 if (cur)
 544                         uptodate = btrfs_buffer_uptodate(cur, gen);
 545                 else
 546                         uptodate = 0;
 547                 if (!cur || !uptodate) {
 548                         if (cache_only) {
 549                                 free_extent_buffer(cur);
 550                                 continue;
 551                         }
 552                         if (!cur) {
 553                                 cur = read_tree_block(root, blocknr,
 554                                                          blocksize, gen);
 555                         } else if (!uptodate) {
 556                                 btrfs_read_buffer(cur, gen);
 557                         }
 558                 }
 559                 if (search_start == 0)
 560                         search_start = last_block;
 561
 562                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 563                                         &tmp, search_start,
 564                                         min(16 * blocksize,
 565                                             (end_slot - i) * blocksize));
 566                 if (err) {
 567                         free_extent_buffer(cur);
 568                         break;
 569                 }
 570                 search_start = tmp->start;
 571                 last_block = tmp->start;
 572                 *last_ret = search_start;
 573                 if (parent_level == 1)
 574                         btrfs_clear_buffer_defrag(tmp);
 575                 free_extent_buffer(tmp);
 576         }
 577         if (parent->map_token) {
 578                 unmap_extent_buffer(parent, parent->map_token,
 579                                     KM_USER1);
 580                 parent->map_token = NULL;
 581         }
 582         return err;
 583 }
 584 #endif
 585
 586 /*
 587  * The leaf data grows from end-to-front in the node.
 588  * this returns the address of the start of the last item,
 589  * which is the stop of the leaf data stack
 590  */
 591 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 592                                          struct extent_buffer *leaf)
 593 {
 594         u32 nr = btrfs_header_nritems(leaf);
 595         if (nr == 0)
 596                 return BTRFS_LEAF_DATA_SIZE(root);
 597         return btrfs_item_offset_nr(leaf, nr - 1);
 598 }
 599
 600 int btrfs_check_node(struct btrfs_root *root,
 601                       struct btrfs_disk_key *parent_key,
 602                       struct extent_buffer *buf)
 603 {
 604         int i;
 605         struct btrfs_key cpukey;
 606         struct btrfs_disk_key key;
 607         u32 nritems = btrfs_header_nritems(buf);
 608
 609         if (nritems == 0 || nritems > BTRFS_NODEPTRS_PER_BLOCK(root))
 610                 goto fail;
 611
 612         if (parent_key && parent_key->type) {
 613                 btrfs_node_key(buf, &key, 0);
 614                 if (memcmp(parent_key, &key, sizeof(key)))
 615                         goto fail;
 616         }
 617         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 618                 btrfs_node_key(buf, &key, i);
 619                 btrfs_node_key_to_cpu(buf, &cpukey, i + 1);
 620                 if (btrfs_comp_keys(&key, &cpukey) >= 0)
 621                         goto fail;
 622         }
 623         return 0;
 624 fail:
 625         if (btrfs_header_owner(buf) == BTRFS_EXTENT_TREE_OBJECTID) {
 626                 if (parent_key)
 627                         btrfs_disk_key_to_cpu(&cpukey, parent_key);
 628                 else
 629                         btrfs_node_key_to_cpu(buf, &cpukey, 0);
 630                 btrfs_add_corrupt_extent_record(root->fs_info, &cpukey,
 631                                                 buf->start, buf->len,
 632                                                 btrfs_header_level(buf));
 633         }
 634         return -EIO;
 635 }
 636
 637 int btrfs_check_leaf(struct btrfs_root *root,
 638                       struct btrfs_disk_key *parent_key,
 639                       struct extent_buffer *buf)
 640 {
 641         int i;
 642         struct btrfs_key cpukey;
 643         struct btrfs_disk_key key;
 644         u32 nritems = btrfs_header_nritems(buf);
 645
 646         if (nritems * sizeof(struct btrfs_item) > buf->len)  {
 647                 fprintf(stderr, "invalid number of items %llu\n",
 648                         (unsigned long long)buf->start);
 649                 goto fail;
 650         }
 651
 652         if (btrfs_header_level(buf) != 0) {
 653                 fprintf(stderr, "leaf is not a leaf %llu\n",
 654                        (unsigned long long)btrfs_header_bytenr(buf));
 655                 goto fail;
 656         }
 657         if (btrfs_leaf_free_space(root, buf) < 0) {
 658                 fprintf(stderr, "leaf free space incorrect %llu %d\n",
 659                         (unsigned long long)btrfs_header_bytenr(buf),
 660                         btrfs_leaf_free_space(root, buf));
 661                 goto fail;
 662         }
 663
 664         if (nritems == 0)
 665                 return 0;
 666
 667         btrfs_item_key(buf, &key, 0);
 668         if (parent_key && parent_key->type &&
 669             memcmp(parent_key, &key, sizeof(key))) {
 670                 fprintf(stderr, "leaf parent key incorrect %llu\n",
 671                        (unsigned long long)btrfs_header_bytenr(buf));
 672                 goto fail;
 673         }
 674         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 675                 btrfs_item_key(buf, &key, i);
 676                 btrfs_item_key_to_cpu(buf, &cpukey, i + 1);
 677                 if (btrfs_comp_keys(&key, &cpukey) >= 0) {
 678                         fprintf(stderr, "bad key ordering %d %d\n", i, i+1);
 679                         goto fail;
 680                 }
 681                 if (btrfs_item_offset_nr(buf, i) !=
 682                         btrfs_item_end_nr(buf, i + 1)) {
 683                         fprintf(stderr, "incorrect offsets %u %u\n",
 684                                 btrfs_item_offset_nr(buf, i),
 685                                 btrfs_item_end_nr(buf, i + 1));
 686                         goto fail;
 687                 }
 688                 if (i == 0 && btrfs_item_end_nr(buf, i) !=
 689                     BTRFS_LEAF_DATA_SIZE(root)) {
 690                         fprintf(stderr, "bad item end %u wanted %u\n",
 691                                 btrfs_item_end_nr(buf, i),
 692                                 (unsigned)BTRFS_LEAF_DATA_SIZE(root));
 693                         goto fail;
 694                 }
 695         }
 696         return 0;
 697 fail:
 698         if (btrfs_header_owner(buf) == BTRFS_EXTENT_TREE_OBJECTID) {
 699                 if (parent_key)
 700                         btrfs_disk_key_to_cpu(&cpukey, parent_key);
 701                 else
 702                         btrfs_item_key_to_cpu(buf, &cpukey, 0);
 703
 704                 btrfs_add_corrupt_extent_record(root->fs_info, &cpukey,
 705                                                 buf->start, buf->len, 0);
 706         }
 707         return -EIO;
 708 }
 709
 710 static int noinline check_block(struct btrfs_root *root,
 711                                 struct btrfs_path *path, int level)
 712 {
 713         struct btrfs_disk_key key;
 714         struct btrfs_disk_key *key_ptr = NULL;
 715         struct extent_buffer *parent;
 716
 717         if (path->nodes[level + 1]) {
 718                 parent = path->nodes[level + 1];
 719                 btrfs_node_key(parent, &key, path->slots[level + 1]);
 720                 key_ptr = &key;
 721         }
 722         if (level == 0)
 723                 return btrfs_check_leaf(root, key_ptr, path->nodes[0]);
 724         return btrfs_check_node(root, key_ptr, path->nodes[level]);
 725 }
 726
 727 /*
 728  * search for key in the extent_buffer.  The items start at offset p,
 729  * and they are item_size apart.  There are 'max' items in p.
 730  *
 731  * the slot in the array is returned via slot, and it points to
 732  * the place where you would insert key if it is not found in
 733  * the array.
 734  *
 735  * slot may point to max if the key is bigger than all of the keys
 736  */
 737 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 738                               int item_size, struct btrfs_key *key,
 739                               int max, int *slot)
 740 {
 741         int low = 0;
 742         int high = max;
 743         int mid;
 744         int ret;
 745         unsigned long offset;
 746         struct btrfs_disk_key *tmp;
 747
 748         while(low < high) {
 749                 mid = (low + high) / 2;
 750                 offset = p + mid * item_size;
 751
 752                 tmp = (struct btrfs_disk_key *)(eb->data + offset);
 753                 ret = btrfs_comp_keys(tmp, key);
 754
 755                 if (ret < 0)
 756                         low = mid + 1;
 757                 else if (ret > 0)
 758                         high = mid;
 759                 else {
 760                         *slot = mid;
 761                         return 0;
 762                 }
 763         }
 764         *slot = low;
 765         return 1;
 766 }
 767
 768 /*
 769  * simple bin_search frontend that does the right thing for
 770  * leaves vs nodes
 771  */
 772 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 773                       int level, int *slot)
 774 {
 775         if (level == 0) {
 776                 return generic_bin_search(eb,
 777                                           offsetof(struct btrfs_leaf, items),
 778                                           sizeof(struct btrfs_item),
 779                                           key, btrfs_header_nritems(eb),
 780                                           slot);
 781         } else {
 782                 return generic_bin_search(eb,
 783                                           offsetof(struct btrfs_node, ptrs),
 784                                           sizeof(struct btrfs_key_ptr),
 785                                           key, btrfs_header_nritems(eb),
 786                                           slot);
 787         }
 788         return -1;
 789 }
 790
 791 struct extent_buffer *read_node_slot(struct btrfs_root *root,
 792                                    struct extent_buffer *parent, int slot)
 793 {
 794         int level = btrfs_header_level(parent);
 795         if (slot < 0)
 796                 return NULL;
 797         if (slot >= btrfs_header_nritems(parent))
 798                 return NULL;
 799
 800         if (level == 0)
 801                 return NULL;
 802
 803         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 804                        btrfs_level_size(root, level - 1),
 805                        btrfs_node_ptr_generation(parent, slot));
 806 }
 807
 808 static int balance_level(struct btrfs_trans_handle *trans,
 809                          struct btrfs_root *root,
 810                          struct btrfs_path *path, int level)
 811 {
 812         struct extent_buffer *right = NULL;
 813         struct extent_buffer *mid;
 814         struct extent_buffer *left = NULL;
 815         struct extent_buffer *parent = NULL;
 816         int ret = 0;
 817         int wret;
 818         int pslot;
 819         int orig_slot = path->slots[level];
 820         u64 orig_ptr;
 821
 822         if (level == 0)
 823                 return 0;
 824
 825         mid = path->nodes[level];
 826         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 827
 828         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 829
 830         if (level < BTRFS_MAX_LEVEL - 1)
 831                 parent = path->nodes[level + 1];
 832         pslot = path->slots[level + 1];
 833
 834         /*
 835          * deal with the case where there is only one pointer in the root
 836          * by promoting the node below to a root
 837          */
 838         if (!parent) {
 839                 struct extent_buffer *child;
 840
 841                 if (btrfs_header_nritems(mid) != 1)
 842                         return 0;
 843
 844                 /* promote the child to a root */
 845                 child = read_node_slot(root, mid, 0);
 846                 BUG_ON(!child);
 847                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
 848                 BUG_ON(ret);
 849
 850                 root->node = child;
 851                 add_root_to_dirty_list(root);
 852                 path->nodes[level] = NULL;
 853                 clean_tree_block(trans, root, mid);
 854                 wait_on_tree_block_writeback(root, mid);
 855                 /* once for the path */
 856                 free_extent_buffer(mid);
 857
 858                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 859                                         0, root->root_key.objectid,
 860                                         level, 1);
 861                 /* once for the root ptr */
 862                 free_extent_buffer(mid);
 863                 return ret;
 864         }
 865         if (btrfs_header_nritems(mid) >
 866             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 867                 return 0;
 868
 869         left = read_node_slot(root, parent, pslot - 1);
 870         if (left) {
 871                 wret = btrfs_cow_block(trans, root, left,
 872                                        parent, pslot - 1, &left);
 873                 if (wret) {
 874                         ret = wret;
 875                         goto enospc;
 876                 }
 877         }
 878         right = read_node_slot(root, parent, pslot + 1);
 879         if (right) {
 880                 wret = btrfs_cow_block(trans, root, right,
 881                                        parent, pslot + 1, &right);
 882                 if (wret) {
 883                         ret = wret;
 884                         goto enospc;
 885                 }
 886         }
 887
 888         /* first, try to make some room in the middle buffer */
 889         if (left) {
 890                 orig_slot += btrfs_header_nritems(left);
 891                 wret = push_node_left(trans, root, left, mid, 1);
 892                 if (wret < 0)
 893                         ret = wret;
 894         }
 895
 896         /*
 897          * then try to empty the right most buffer into the middle
 898          */
 899         if (right) {
 900                 wret = push_node_left(trans, root, mid, right, 1);
 901                 if (wret < 0 && wret != -ENOSPC)
 902                         ret = wret;
 903                 if (btrfs_header_nritems(right) == 0) {
 904                         u64 bytenr = right->start;
 905                         u32 blocksize = right->len;
 906
 907                         clean_tree_block(trans, root, right);
 908                         wait_on_tree_block_writeback(root, right);
 909                         free_extent_buffer(right);
 910                         right = NULL;
 911                         wret = btrfs_del_ptr(trans, root, path,
 912                                              level + 1, pslot + 1);
 913                         if (wret)
 914                                 ret = wret;
 915                         wret = btrfs_free_extent(trans, root, bytenr,
 916                                                  blocksize, 0,
 917                                                  root->root_key.objectid,
 918                                                  level, 0);
 919                         if (wret)
 920                                 ret = wret;
 921                 } else {
 922                         struct btrfs_disk_key right_key;
 923                         btrfs_node_key(right, &right_key, 0);
 924                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 925                         btrfs_mark_buffer_dirty(parent);
 926                 }
 927         }
 928         if (btrfs_header_nritems(mid) == 1) {
 929                 /*
 930                  * we're not allowed to leave a node with one item in the
 931                  * tree during a delete.  A deletion from lower in the tree
 932                  * could try to delete the only pointer in this node.
 933                  * So, pull some keys from the left.
 934                  * There has to be a left pointer at this point because
 935                  * otherwise we would have pulled some pointers from the
 936                  * right
 937                  */
 938                 BUG_ON(!left);
 939                 wret = balance_node_right(trans, root, mid, left);
 940                 if (wret < 0) {
 941                         ret = wret;
 942                         goto enospc;
 943                 }
 944                 if (wret == 1) {
 945                         wret = push_node_left(trans, root, left, mid, 1);
 946                         if (wret < 0)
 947                                 ret = wret;
 948                 }
 949                 BUG_ON(wret == 1);
 950         }
 951         if (btrfs_header_nritems(mid) == 0) {
 952                 /* we've managed to empty the middle node, drop it */
 953                 u64 bytenr = mid->start;
 954                 u32 blocksize = mid->len;
 955                 clean_tree_block(trans, root, mid);
 956                 wait_on_tree_block_writeback(root, mid);
 957                 free_extent_buffer(mid);
 958                 mid = NULL;
 959                 wret = btrfs_del_ptr(trans, root, path, level + 1, pslot);
 960                 if (wret)
 961                         ret = wret;
 962                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 963                                          0, root->root_key.objectid,
 964                                          level, 0);
 965                 if (wret)
 966                         ret = wret;
 967         } else {
 968                 /* update the parent key to reflect our changes */
 969                 struct btrfs_disk_key mid_key;
 970                 btrfs_node_key(mid, &mid_key, 0);
 971                 btrfs_set_node_key(parent, &mid_key, pslot);
 972                 btrfs_mark_buffer_dirty(parent);
 973         }
 974
 975         /* update the path */
 976         if (left) {
 977                 if (btrfs_header_nritems(left) > orig_slot) {
 978                         extent_buffer_get(left);
 979                         path->nodes[level] = left;
 980                         path->slots[level + 1] -= 1;
 981                         path->slots[level] = orig_slot;
 982                         if (mid)
 983                                 free_extent_buffer(mid);
 984                 } else {
 985                         orig_slot -= btrfs_header_nritems(left);
 986                         path->slots[level] = orig_slot;
 987                 }
 988         }
 989         /* double check we haven't messed things up */
 990         check_block(root, path, level);
 991         if (orig_ptr !=
 992             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 993                 BUG();
 994 enospc:
 995         if (right)
 996                 free_extent_buffer(right);
 997         if (left)
 998                 free_extent_buffer(left);
 999         return ret;
1000 }
1001
1002 /* returns zero if the push worked, non-zero otherwise */
1003 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
1004                                           struct btrfs_root *root,
1005                                           struct btrfs_path *path, int level)
1006 {
1007         struct extent_buffer *right = NULL;
1008         struct extent_buffer *mid;
1009         struct extent_buffer *left = NULL;
1010         struct extent_buffer *parent = NULL;
1011         int ret = 0;
1012         int wret;
1013         int pslot;
1014         int orig_slot = path->slots[level];
1015
1016         if (level == 0)
1017                 return 1;
1018
1019         mid = path->nodes[level];
1020         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1021
1022         if (level < BTRFS_MAX_LEVEL - 1)
1023                 parent = path->nodes[level + 1];
1024         pslot = path->slots[level + 1];
1025
1026         if (!parent)
1027                 return 1;
1028
1029         left = read_node_slot(root, parent, pslot - 1);
1030
1031         /* first, try to make some room in the middle buffer */
1032         if (left) {
1033                 u32 left_nr;
1034                 left_nr = btrfs_header_nritems(left);
1035                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1036                         wret = 1;
1037                 } else {
1038                         ret = btrfs_cow_block(trans, root, left, parent,
1039                                               pslot - 1, &left);
1040                         if (ret)
1041                                 wret = 1;
1042                         else {
1043                                 wret = push_node_left(trans, root,
1044                                                       left, mid, 0);
1045                         }
1046                 }
1047                 if (wret < 0)
1048                         ret = wret;
1049                 if (wret == 0) {
1050                         struct btrfs_disk_key disk_key;
1051                         orig_slot += left_nr;
1052                         btrfs_node_key(mid, &disk_key, 0);
1053                         btrfs_set_node_key(parent, &disk_key, pslot);
1054                         btrfs_mark_buffer_dirty(parent);
1055                         if (btrfs_header_nritems(left) > orig_slot) {
1056                                 path->nodes[level] = left;
1057                                 path->slots[level + 1] -= 1;
1058                                 path->slots[level] = orig_slot;
1059                                 free_extent_buffer(mid);
1060                         } else {
1061                                 orig_slot -=
1062                                         btrfs_header_nritems(left);
1063                                 path->slots[level] = orig_slot;
1064                                 free_extent_buffer(left);
1065                         }
1066                         return 0;
1067                 }
1068                 free_extent_buffer(left);
1069         }
1070         right= read_node_slot(root, parent, pslot + 1);
1071
1072         /*
1073          * then try to empty the right most buffer into the middle
1074          */
1075         if (right) {
1076                 u32 right_nr;
1077                 right_nr = btrfs_header_nritems(right);
1078                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1079                         wret = 1;
1080                 } else {
1081                         ret = btrfs_cow_block(trans, root, right,
1082                                               parent, pslot + 1,
1083                                               &right);
1084                         if (ret)
1085                                 wret = 1;
1086                         else {
1087                                 wret = balance_node_right(trans, root,
1088                                                           right, mid);
1089                         }
1090                 }
1091                 if (wret < 0)
1092                         ret = wret;
1093                 if (wret == 0) {
1094                         struct btrfs_disk_key disk_key;
1095
1096                         btrfs_node_key(right, &disk_key, 0);
1097                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
1098                         btrfs_mark_buffer_dirty(parent);
1099
1100                         if (btrfs_header_nritems(mid) <= orig_slot) {
1101                                 path->nodes[level] = right;
1102                                 path->slots[level + 1] += 1;
1103                                 path->slots[level] = orig_slot -
1104                                         btrfs_header_nritems(mid);
1105                                 free_extent_buffer(mid);
1106                         } else {
1107                                 free_extent_buffer(right);
1108                         }
1109                         return 0;
1110                 }
1111                 free_extent_buffer(right);
1112         }
1113         return 1;
1114 }
1115
1116 /*
1117  * readahead one full node of leaves
1118  */
1119 void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
1120                              int level, int slot, u64 objectid)
1121 {
1122         struct extent_buffer *node;
1123         struct btrfs_disk_key disk_key;
1124         u32 nritems;
1125         u64 search;
1126         u64 lowest_read;
1127         u64 highest_read;
1128         u64 nread = 0;
1129         int direction = path->reada;
1130         struct extent_buffer *eb;
1131         u32 nr;
1132         u32 blocksize;
1133         u32 nscan = 0;
1134
1135         if (level != 1)
1136                 return;
1137
1138         if (!path->nodes[level])
1139                 return;
1140
1141         node = path->nodes[level];
1142         search = btrfs_node_blockptr(node, slot);
1143         blocksize = btrfs_level_size(root, level - 1);
1144         eb = btrfs_find_tree_block(root, search, blocksize);
1145         if (eb) {
1146                 free_extent_buffer(eb);
1147                 return;
1148         }
1149
1150         highest_read = search;
1151         lowest_read = search;
1152
1153         nritems = btrfs_header_nritems(node);
1154         nr = slot;
1155         while(1) {
1156                 if (direction < 0) {
1157                         if (nr == 0)
1158                                 break;
1159                         nr--;
1160                 } else if (direction > 0) {
1161                         nr++;
1162                         if (nr >= nritems)
1163                                 break;
1164                 }
1165                 if (path->reada < 0 && objectid) {
1166                         btrfs_node_key(node, &disk_key, nr);
1167                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1168                                 break;
1169                 }
1170                 search = btrfs_node_blockptr(node, nr);
1171                 if ((search >= lowest_read && search <= highest_read) ||
1172                     (search < lowest_read && lowest_read - search <= 32768) ||
1173                     (search > highest_read && search - highest_read <= 32768)) {
1174                         readahead_tree_block(root, search, blocksize,
1175                                      btrfs_node_ptr_generation(node, nr));
1176                         nread += blocksize;
1177                 }
1178                 nscan++;
1179                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
1180                         break;
1181                 if(nread > (1024 * 1024) || nscan > 128)
1182                         break;
1183
1184                 if (search < lowest_read)
1185                         lowest_read = search;
1186                 if (search > highest_read)
1187                         highest_read = search;
1188         }
1189 }
1190
1191 /*
1192  * look for key in the tree.  path is filled in with nodes along the way
1193  * if key is found, we return zero and you can find the item in the leaf
1194  * level of the path (level 0)
1195  *
1196  * If the key isn't found, the path points to the slot where it should
1197  * be inserted, and 1 is returned.  If there are other errors during the
1198  * search a negative error number is returned.
1199  *
1200  * if ins_len > 0, nodes and leaves will be split as we walk down the
1201  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1202  * possible)
1203  */
1204 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1205                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1206                       ins_len, int cow)
1207 {
1208         struct extent_buffer *b;
1209         int slot;
1210         int ret;
1211         int level;
1212         int should_reada = p->reada;
1213         u8 lowest_level = 0;
1214
1215         lowest_level = p->lowest_level;
1216         WARN_ON(lowest_level && ins_len > 0);
1217         WARN_ON(p->nodes[0] != NULL);
1218         /*
1219         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
1220         */
1221 again:
1222         b = root->node;
1223         extent_buffer_get(b);
1224         while (b) {
1225                 level = btrfs_header_level(b);
1226                 if (cow) {
1227                         int wret;
1228                         wret = btrfs_cow_block(trans, root, b,
1229                                                p->nodes[level + 1],
1230                                                p->slots[level + 1],
1231                                                &b);
1232                         if (wret) {
1233                                 free_extent_buffer(b);
1234                                 return wret;
1235                         }
1236                 }
1237                 BUG_ON(!cow && ins_len);
1238                 if (level != btrfs_header_level(b))
1239                         WARN_ON(1);
1240                 level = btrfs_header_level(b);
1241                 p->nodes[level] = b;
1242                 ret = check_block(root, p, level);
1243                 if (ret)
1244                         return -1;
1245                 ret = bin_search(b, key, level, &slot);
1246                 if (level != 0) {
1247                         if (ret && slot > 0)
1248                                 slot -= 1;
1249                         p->slots[level] = slot;
1250                         if ((p->search_for_split || ins_len > 0) &&
1251                             btrfs_header_nritems(b) >=
1252                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1253                                 int sret = split_node(trans, root, p, level);
1254                                 BUG_ON(sret > 0);
1255                                 if (sret)
1256                                         return sret;
1257                                 b = p->nodes[level];
1258                                 slot = p->slots[level];
1259                         } else if (ins_len < 0) {
1260                                 int sret = balance_level(trans, root, p,
1261                                                          level);
1262                                 if (sret)
1263                                         return sret;
1264                                 b = p->nodes[level];
1265                                 if (!b) {
1266                                         btrfs_release_path(NULL, p);
1267                                         goto again;
1268                                 }
1269                                 slot = p->slots[level];
1270                                 BUG_ON(btrfs_header_nritems(b) == 1);
1271                         }
1272                         /* this is only true while dropping a snapshot */
1273                         if (level == lowest_level)
1274                                 break;
1275
1276                         if (should_reada)
1277                                 reada_for_search(root, p, level, slot,
1278                                                  key->objectid);
1279
1280                         b = read_node_slot(root, b, slot);
1281                         if (!extent_buffer_uptodate(b))
1282                                 return -EIO;
1283                 } else {
1284                         p->slots[level] = slot;
1285                         if (ins_len > 0 &&
1286                             ins_len > btrfs_leaf_free_space(root, b)) {
1287                                 int sret = split_leaf(trans, root, key,
1288                                                       p, ins_len, ret == 0);
1289                                 BUG_ON(sret > 0);
1290                                 if (sret)
1291                                         return sret;
1292                         }
1293                         return ret;
1294                 }
1295         }
1296         return 1;
1297 }
1298
1299 /*
1300  * adjust the pointers going up the tree, starting at level
1301  * making sure the right key of each node is points to 'key'.
1302  * This is used after shifting pointers to the left, so it stops
1303  * fixing up pointers when a given leaf/node is not in slot 0 of the
1304  * higher levels
1305  *
1306  * If this fails to write a tree block, it returns -1, but continues
1307  * fixing up the blocks in ram so the tree is consistent.
1308  */
1309 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1310                           struct btrfs_root *root, struct btrfs_path *path,
1311                           struct btrfs_disk_key *key, int level)
1312 {
1313         int i;
1314         int ret = 0;
1315         struct extent_buffer *t;
1316
1317         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1318                 int tslot = path->slots[i];
1319                 if (!path->nodes[i])
1320                         break;
1321                 t = path->nodes[i];
1322                 btrfs_set_node_key(t, key, tslot);
1323                 btrfs_mark_buffer_dirty(path->nodes[i]);
1324                 if (tslot != 0)
1325                         break;
1326         }
1327         return ret;
1328 }
1329
1330 /*
1331  * update item key.
1332  *
1333  * This function isn't completely safe. It's the caller's responsibility
1334  * that the new key won't break the order
1335  */
1336 int btrfs_set_item_key_safe(struct btrfs_trans_handle *trans,
1337                             struct btrfs_root *root, struct btrfs_path *path,
1338                             struct btrfs_key *new_key)
1339 {
1340         struct btrfs_disk_key disk_key;
1341         struct extent_buffer *eb;
1342         int slot;
1343
1344         eb = path->nodes[0];
1345         slot = path->slots[0];
1346         if (slot > 0) {
1347                 btrfs_item_key(eb, &disk_key, slot - 1);
1348                 if (btrfs_comp_keys(&disk_key, new_key) >= 0)
1349                         return -1;
1350         }
1351         if (slot < btrfs_header_nritems(eb) - 1) {
1352                 btrfs_item_key(eb, &disk_key, slot + 1);
1353                 if (btrfs_comp_keys(&disk_key, new_key) <= 0)
1354                         return -1;
1355         }
1356
1357         btrfs_cpu_key_to_disk(&disk_key, new_key);
1358         btrfs_set_item_key(eb, &disk_key, slot);
1359         btrfs_mark_buffer_dirty(eb);
1360         if (slot == 0)
1361                 fixup_low_keys(trans, root, path, &disk_key, 1);
1362         return 0;
1363 }
1364
1365 /*
1366  * try to push data from one node into the next node left in the
1367  * tree.
1368  *
1369  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1370  * error, and > 0 if there was no room in the left hand block.
1371  */
1372 static int push_node_left(struct btrfs_trans_handle *trans,
1373                           struct btrfs_root *root, struct extent_buffer *dst,
1374                           struct extent_buffer *src, int empty)
1375 {
1376         int push_items = 0;
1377         int src_nritems;
1378         int dst_nritems;
1379         int ret = 0;
1380
1381         src_nritems = btrfs_header_nritems(src);
1382         dst_nritems = btrfs_header_nritems(dst);
1383         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1384         WARN_ON(btrfs_header_generation(src) != trans->transid);
1385         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1386
1387         if (!empty && src_nritems <= 8)
1388                 return 1;
1389
1390         if (push_items <= 0) {
1391                 return 1;
1392         }
1393
1394         if (empty) {
1395                 push_items = min(src_nritems, push_items);
1396                 if (push_items < src_nritems) {
1397                         /* leave at least 8 pointers in the node if
1398                          * we aren't going to empty it
1399                          */
1400                         if (src_nritems - push_items < 8) {
1401                                 if (push_items <= 8)
1402                                         return 1;
1403                                 push_items -= 8;
1404                         }
1405                 }
1406         } else
1407                 push_items = min(src_nritems - 8, push_items);
1408
1409         copy_extent_buffer(dst, src,
1410                            btrfs_node_key_ptr_offset(dst_nritems),
1411                            btrfs_node_key_ptr_offset(0),
1412                            push_items * sizeof(struct btrfs_key_ptr));
1413
1414         if (push_items < src_nritems) {
1415                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1416                                       btrfs_node_key_ptr_offset(push_items),
1417                                       (src_nritems - push_items) *
1418                                       sizeof(struct btrfs_key_ptr));
1419         }
1420         btrfs_set_header_nritems(src, src_nritems - push_items);
1421         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1422         btrfs_mark_buffer_dirty(src);
1423         btrfs_mark_buffer_dirty(dst);
1424
1425         return ret;
1426 }
1427
1428 /*
1429  * try to push data from one node into the next node right in the
1430  * tree.
1431  *
1432  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1433  * error, and > 0 if there was no room in the right hand block.
1434  *
1435  * this will  only push up to 1/2 the contents of the left node over
1436  */
1437 static int balance_node_right(struct btrfs_trans_handle *trans,
1438                               struct btrfs_root *root,
1439                               struct extent_buffer *dst,
1440                               struct extent_buffer *src)
1441 {
1442         int push_items = 0;
1443         int max_push;
1444         int src_nritems;
1445         int dst_nritems;
1446         int ret = 0;
1447
1448         WARN_ON(btrfs_header_generation(src) != trans->transid);
1449         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1450
1451         src_nritems = btrfs_header_nritems(src);
1452         dst_nritems = btrfs_header_nritems(dst);
1453         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1454         if (push_items <= 0) {
1455                 return 1;
1456         }
1457
1458         if (src_nritems < 4) {
1459                 return 1;
1460         }
1461
1462         max_push = src_nritems / 2 + 1;
1463         /* don't try to empty the node */
1464         if (max_push >= src_nritems) {
1465                 return 1;
1466         }
1467
1468         if (max_push < push_items)
1469                 push_items = max_push;
1470
1471         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1472                                       btrfs_node_key_ptr_offset(0),
1473                                       (dst_nritems) *
1474                                       sizeof(struct btrfs_key_ptr));
1475
1476         copy_extent_buffer(dst, src,
1477                            btrfs_node_key_ptr_offset(0),
1478                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1479                            push_items * sizeof(struct btrfs_key_ptr));
1480
1481         btrfs_set_header_nritems(src, src_nritems - push_items);
1482         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1483
1484         btrfs_mark_buffer_dirty(src);
1485         btrfs_mark_buffer_dirty(dst);
1486
1487         return ret;
1488 }
1489
1490 /*
1491  * helper function to insert a new root level in the tree.
1492  * A new node is allocated, and a single item is inserted to
1493  * point to the existing root
1494  *
1495  * returns zero on success or < 0 on failure.
1496  */
1497 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1498                            struct btrfs_root *root,
1499                            struct btrfs_path *path, int level)
1500 {
1501         u64 lower_gen;
1502         struct extent_buffer *lower;
1503         struct extent_buffer *c;
1504         struct extent_buffer *old;
1505         struct btrfs_disk_key lower_key;
1506
1507         BUG_ON(path->nodes[level]);
1508         BUG_ON(path->nodes[level-1] != root->node);
1509
1510         lower = path->nodes[level-1];
1511         if (level == 1)
1512                 btrfs_item_key(lower, &lower_key, 0);
1513         else
1514                 btrfs_node_key(lower, &lower_key, 0);
1515
1516         c = btrfs_alloc_free_block(trans, root, root->nodesize,
1517                                    root->root_key.objectid, &lower_key,
1518                                    level, root->node->start, 0);
1519
1520         if (IS_ERR(c))
1521                 return PTR_ERR(c);
1522
1523         memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
1524         btrfs_set_header_nritems(c, 1);
1525         btrfs_set_header_level(c, level);
1526         btrfs_set_header_bytenr(c, c->start);
1527         btrfs_set_header_generation(c, trans->transid);
1528         btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
1529         btrfs_set_header_owner(c, root->root_key.objectid);
1530
1531         write_extent_buffer(c, root->fs_info->fsid,
1532                             (unsigned long)btrfs_header_fsid(c),
1533                             BTRFS_FSID_SIZE);
1534
1535         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1536                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
1537                             BTRFS_UUID_SIZE);
1538
1539         btrfs_set_node_key(c, &lower_key, 0);
1540         btrfs_set_node_blockptr(c, 0, lower->start);
1541         lower_gen = btrfs_header_generation(lower);
1542         WARN_ON(lower_gen != trans->transid);
1543
1544         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1545
1546         btrfs_mark_buffer_dirty(c);
1547
1548         old = root->node;
1549         root->node = c;
1550
1551         /* the super has an extra ref to root->node */
1552         free_extent_buffer(old);
1553
1554         add_root_to_dirty_list(root);
1555         extent_buffer_get(c);
1556         path->nodes[level] = c;
1557         path->slots[level] = 0;
1558         return 0;
1559 }
1560
1561 /*
1562  * worker function to insert a single pointer in a node.
1563  * the node should have enough room for the pointer already
1564  *
1565  * slot and level indicate where you want the key to go, and
1566  * blocknr is the block the key points to.
1567  *
1568  * returns zero on success and < 0 on any error
1569  */
1570 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1571                       *root, struct btrfs_path *path, struct btrfs_disk_key
1572                       *key, u64 bytenr, int slot, int level)
1573 {
1574         struct extent_buffer *lower;
1575         int nritems;
1576
1577         BUG_ON(!path->nodes[level]);
1578         lower = path->nodes[level];
1579         nritems = btrfs_header_nritems(lower);
1580         if (slot > nritems)
1581                 BUG();
1582         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1583                 BUG();
1584         if (slot != nritems) {
1585                 memmove_extent_buffer(lower,
1586                               btrfs_node_key_ptr_offset(slot + 1),
1587                               btrfs_node_key_ptr_offset(slot),
1588                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1589         }
1590         btrfs_set_node_key(lower, key, slot);
1591         btrfs_set_node_blockptr(lower, slot, bytenr);
1592         WARN_ON(trans->transid == 0);
1593         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1594         btrfs_set_header_nritems(lower, nritems + 1);
1595         btrfs_mark_buffer_dirty(lower);
1596         return 0;
1597 }
1598
1599 /*
1600  * split the node at the specified level in path in two.
1601  * The path is corrected to point to the appropriate node after the split
1602  *
1603  * Before splitting this tries to make some room in the node by pushing
1604  * left and right, if either one works, it returns right away.
1605  *
1606  * returns 0 on success and < 0 on failure
1607  */
1608 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1609                       *root, struct btrfs_path *path, int level)
1610 {
1611         struct extent_buffer *c;
1612         struct extent_buffer *split;
1613         struct btrfs_disk_key disk_key;
1614         int mid;
1615         int ret;
1616         int wret;
1617         u32 c_nritems;
1618
1619         c = path->nodes[level];
1620         WARN_ON(btrfs_header_generation(c) != trans->transid);
1621         if (c == root->node) {
1622                 /* trying to split the root, lets make a new one */
1623                 ret = insert_new_root(trans, root, path, level + 1);
1624                 if (ret)
1625                         return ret;
1626         } else {
1627                 ret = push_nodes_for_insert(trans, root, path, level);
1628                 c = path->nodes[level];
1629                 if (!ret && btrfs_header_nritems(c) <
1630                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
1631                         return 0;
1632                 if (ret < 0)
1633                         return ret;
1634         }
1635
1636         c_nritems = btrfs_header_nritems(c);
1637         mid = (c_nritems + 1) / 2;
1638         btrfs_node_key(c, &disk_key, mid);
1639
1640         split = btrfs_alloc_free_block(trans, root, root->nodesize,
1641                                         root->root_key.objectid,
1642                                         &disk_key, level, c->start, 0);
1643         if (IS_ERR(split))
1644                 return PTR_ERR(split);
1645
1646         memset_extent_buffer(split, 0, 0, sizeof(struct btrfs_header));
1647         btrfs_set_header_level(split, btrfs_header_level(c));
1648         btrfs_set_header_bytenr(split, split->start);
1649         btrfs_set_header_generation(split, trans->transid);
1650         btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
1651         btrfs_set_header_owner(split, root->root_key.objectid);
1652         write_extent_buffer(split, root->fs_info->fsid,
1653                             (unsigned long)btrfs_header_fsid(split),
1654                             BTRFS_FSID_SIZE);
1655         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1656                             (unsigned long)btrfs_header_chunk_tree_uuid(split),
1657                             BTRFS_UUID_SIZE);
1658
1659
1660         copy_extent_buffer(split, c,
1661                            btrfs_node_key_ptr_offset(0),
1662                            btrfs_node_key_ptr_offset(mid),
1663                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1664         btrfs_set_header_nritems(split, c_nritems - mid);
1665         btrfs_set_header_nritems(c, mid);
1666         ret = 0;
1667
1668         btrfs_mark_buffer_dirty(c);
1669         btrfs_mark_buffer_dirty(split);
1670
1671         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1672                           path->slots[level + 1] + 1,
1673                           level + 1);
1674         if (wret)
1675                 ret = wret;
1676
1677         if (path->slots[level] >= mid) {
1678                 path->slots[level] -= mid;
1679                 free_extent_buffer(c);
1680                 path->nodes[level] = split;
1681                 path->slots[level + 1] += 1;
1682         } else {
1683                 free_extent_buffer(split);
1684         }
1685         return ret;
1686 }
1687
1688 /*
1689  * how many bytes are required to store the items in a leaf.  start
1690  * and nr indicate which items in the leaf to check.  This totals up the
1691  * space used both by the item structs and the item data
1692  */
1693 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1694 {
1695         int data_len;
1696         int nritems = btrfs_header_nritems(l);
1697         int end = min(nritems, start + nr) - 1;
1698
1699         if (!nr)
1700                 return 0;
1701         data_len = btrfs_item_end_nr(l, start);
1702         data_len = data_len - btrfs_item_offset_nr(l, end);
1703         data_len += sizeof(struct btrfs_item) * nr;
1704         WARN_ON(data_len < 0);
1705         return data_len;
1706 }
1707
1708 /*
1709  * The space between the end of the leaf items and
1710  * the start of the leaf data.  IOW, how much room
1711  * the leaf has left for both items and data
1712  */
1713 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1714 {
1715         int nritems = btrfs_header_nritems(leaf);
1716         int ret;
1717         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1718         if (ret < 0) {
1719                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1720                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1721                        leaf_space_used(leaf, 0, nritems), nritems);
1722         }
1723         return ret;
1724 }
1725
1726 /*
1727  * push some data in the path leaf to the right, trying to free up at
1728  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1729  *
1730  * returns 1 if the push failed because the other node didn't have enough
1731  * room, 0 if everything worked out and < 0 if there were major errors.
1732  */
1733 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1734                            *root, struct btrfs_path *path, int data_size,
1735                            int empty)
1736 {
1737         struct extent_buffer *left = path->nodes[0];
1738         struct extent_buffer *right;
1739         struct extent_buffer *upper;
1740         struct btrfs_disk_key disk_key;
1741         int slot;
1742         u32 i;
1743         int free_space;
1744         int push_space = 0;
1745         int push_items = 0;
1746         struct btrfs_item *item;
1747         u32 left_nritems;
1748         u32 nr;
1749         u32 right_nritems;
1750         u32 data_end;
1751         u32 this_item_size;
1752         int ret;
1753
1754         slot = path->slots[1];
1755         if (!path->nodes[1]) {
1756                 return 1;
1757         }
1758         upper = path->nodes[1];
1759         if (slot >= btrfs_header_nritems(upper) - 1)
1760                 return 1;
1761
1762         right = read_node_slot(root, upper, slot + 1);
1763         free_space = btrfs_leaf_free_space(root, right);
1764         if (free_space < data_size) {
1765                 free_extent_buffer(right);
1766                 return 1;
1767         }
1768
1769         /* cow and double check */
1770         ret = btrfs_cow_block(trans, root, right, upper,
1771                               slot + 1, &right);
1772         if (ret) {
1773                 free_extent_buffer(right);
1774                 return 1;
1775         }
1776         free_space = btrfs_leaf_free_space(root, right);
1777         if (free_space < data_size) {
1778                 free_extent_buffer(right);
1779                 return 1;
1780         }
1781
1782         left_nritems = btrfs_header_nritems(left);
1783         if (left_nritems == 0) {
1784                 free_extent_buffer(right);
1785                 return 1;
1786         }
1787
1788         if (empty)
1789                 nr = 0;
1790         else
1791                 nr = 1;
1792
1793         i = left_nritems - 1;
1794         while (i >= nr) {
1795                 item = btrfs_item_nr(left, i);
1796
1797                 if (path->slots[0] == i)
1798                         push_space += data_size + sizeof(*item);
1799
1800                 this_item_size = btrfs_item_size(left, item);
1801                 if (this_item_size + sizeof(*item) + push_space > free_space)
1802                         break;
1803                 push_items++;
1804                 push_space += this_item_size + sizeof(*item);
1805                 if (i == 0)
1806                         break;
1807                 i--;
1808         }
1809
1810         if (push_items == 0) {
1811                 free_extent_buffer(right);
1812                 return 1;
1813         }
1814
1815         if (!empty && push_items == left_nritems)
1816                 WARN_ON(1);
1817
1818         /* push left to right */
1819         right_nritems = btrfs_header_nritems(right);
1820
1821         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1822         push_space -= leaf_data_end(root, left);
1823
1824         /* make room in the right data area */
1825         data_end = leaf_data_end(root, right);
1826         memmove_extent_buffer(right,
1827                               btrfs_leaf_data(right) + data_end - push_space,
1828                               btrfs_leaf_data(right) + data_end,
1829                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1830
1831         /* copy from the left data area */
1832         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1833                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1834                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1835                      push_space);
1836
1837         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1838                               btrfs_item_nr_offset(0),
1839                               right_nritems * sizeof(struct btrfs_item));
1840
1841         /* copy the items from left to right */
1842         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1843                    btrfs_item_nr_offset(left_nritems - push_items),
1844                    push_items * sizeof(struct btrfs_item));
1845
1846         /* update the item pointers */
1847         right_nritems += push_items;
1848         btrfs_set_header_nritems(right, right_nritems);
1849         push_space = BTRFS_LEAF_DATA_SIZE(root);
1850         for (i = 0; i < right_nritems; i++) {
1851                 item = btrfs_item_nr(right, i);
1852                 push_space -= btrfs_item_size(right, item);
1853                 btrfs_set_item_offset(right, item, push_space);
1854         }
1855
1856         left_nritems -= push_items;
1857         btrfs_set_header_nritems(left, left_nritems);
1858
1859         if (left_nritems)
1860                 btrfs_mark_buffer_dirty(left);
1861         btrfs_mark_buffer_dirty(right);
1862
1863         btrfs_item_key(right, &disk_key, 0);
1864         btrfs_set_node_key(upper, &disk_key, slot + 1);
1865         btrfs_mark_buffer_dirty(upper);
1866
1867         /* then fixup the leaf pointer in the path */
1868         if (path->slots[0] >= left_nritems) {
1869                 path->slots[0] -= left_nritems;
1870                 free_extent_buffer(path->nodes[0]);
1871                 path->nodes[0] = right;
1872                 path->slots[1] += 1;
1873         } else {
1874                 free_extent_buffer(right);
1875         }
1876         return 0;
1877 }
1878 /*
1879  * push some data in the path leaf to the left, trying to free up at
1880  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1881  */
1882 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1883                           *root, struct btrfs_path *path, int data_size,
1884                           int empty)
1885 {
1886         struct btrfs_disk_key disk_key;
1887         struct extent_buffer *right = path->nodes[0];
1888         struct extent_buffer *left;
1889         int slot;
1890         int i;
1891         int free_space;
1892         int push_space = 0;
1893         int push_items = 0;
1894         struct btrfs_item *item;
1895         u32 old_left_nritems;
1896         u32 right_nritems;
1897         u32 nr;
1898         int ret = 0;
1899         int wret;
1900         u32 this_item_size;
1901         u32 old_left_item_size;
1902
1903         slot = path->slots[1];
1904         if (slot == 0)
1905                 return 1;
1906         if (!path->nodes[1])
1907                 return 1;
1908
1909         right_nritems = btrfs_header_nritems(right);
1910         if (right_nritems == 0) {
1911                 return 1;
1912         }
1913
1914         left = read_node_slot(root, path->nodes[1], slot - 1);
1915         free_space = btrfs_leaf_free_space(root, left);
1916         if (free_space < data_size) {
1917                 free_extent_buffer(left);
1918                 return 1;
1919         }
1920
1921         /* cow and double check */
1922         ret = btrfs_cow_block(trans, root, left,
1923                               path->nodes[1], slot - 1, &left);
1924         if (ret) {
1925                 /* we hit -ENOSPC, but it isn't fatal here */
1926                 free_extent_buffer(left);
1927                 return 1;
1928         }
1929
1930         free_space = btrfs_leaf_free_space(root, left);
1931         if (free_space < data_size) {
1932                 free_extent_buffer(left);
1933                 return 1;
1934         }
1935
1936         if (empty)
1937                 nr = right_nritems;
1938         else
1939                 nr = right_nritems - 1;
1940
1941         for (i = 0; i < nr; i++) {
1942                 item = btrfs_item_nr(right, i);
1943
1944                 if (path->slots[0] == i)
1945                         push_space += data_size + sizeof(*item);
1946
1947                 this_item_size = btrfs_item_size(right, item);
1948                 if (this_item_size + sizeof(*item) + push_space > free_space)
1949                         break;
1950
1951                 push_items++;
1952                 push_space += this_item_size + sizeof(*item);
1953         }
1954
1955         if (push_items == 0) {
1956                 free_extent_buffer(left);
1957                 return 1;
1958         }
1959         if (!empty && push_items == btrfs_header_nritems(right))
1960                 WARN_ON(1);
1961
1962         /* push data from right to left */
1963         copy_extent_buffer(left, right,
1964                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1965                            btrfs_item_nr_offset(0),
1966                            push_items * sizeof(struct btrfs_item));
1967
1968         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1969                      btrfs_item_offset_nr(right, push_items -1);
1970
1971         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1972                      leaf_data_end(root, left) - push_space,
1973                      btrfs_leaf_data(right) +
1974                      btrfs_item_offset_nr(right, push_items - 1),
1975                      push_space);
1976         old_left_nritems = btrfs_header_nritems(left);
1977         BUG_ON(old_left_nritems == 0);
1978
1979         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1980         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1981                 u32 ioff;
1982
1983                 item = btrfs_item_nr(left, i);
1984                 ioff = btrfs_item_offset(left, item);
1985                 btrfs_set_item_offset(left, item,
1986                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1987         }
1988         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1989
1990         /* fixup right node */
1991         if (push_items > right_nritems) {
1992                 printk("push items %d nr %u\n", push_items, right_nritems);
1993                 WARN_ON(1);
1994         }
1995
1996         if (push_items < right_nritems) {
1997                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1998                                                   leaf_data_end(root, right);
1999                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
2000                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
2001                                       btrfs_leaf_data(right) +
2002                                       leaf_data_end(root, right), push_space);
2003
2004                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
2005                               btrfs_item_nr_offset(push_items),
2006                              (btrfs_header_nritems(right) - push_items) *
2007                              sizeof(struct btrfs_item));
2008         }
2009         right_nritems -= push_items;
2010         btrfs_set_header_nritems(right, right_nritems);
2011         push_space = BTRFS_LEAF_DATA_SIZE(root);
2012         for (i = 0; i < right_nritems; i++) {
2013                 item = btrfs_item_nr(right, i);
2014                 push_space = push_space - btrfs_item_size(right, item);
2015                 btrfs_set_item_offset(right, item, push_space);
2016         }
2017
2018         btrfs_mark_buffer_dirty(left);
2019         if (right_nritems)
2020                 btrfs_mark_buffer_dirty(right);
2021
2022         btrfs_item_key(right, &disk_key, 0);
2023         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
2024         if (wret)
2025                 ret = wret;
2026
2027         /* then fixup the leaf pointer in the path */
2028         if (path->slots[0] < push_items) {
2029                 path->slots[0] += old_left_nritems;
2030                 free_extent_buffer(path->nodes[0]);
2031                 path->nodes[0] = left;
2032                 path->slots[1] -= 1;
2033         } else {
2034                 free_extent_buffer(left);
2035                 path->slots[0] -= push_items;
2036         }
2037         BUG_ON(path->slots[0] < 0);
2038         return ret;
2039 }
2040
2041 /*
2042  * split the path's leaf in two, making sure there is at least data_size
2043  * available for the resulting leaf level of the path.
2044  *
2045  * returns 0 if all went well and < 0 on failure.
2046  */
2047 static noinline int copy_for_split(struct btrfs_trans_handle *trans,
2048                                struct btrfs_root *root,
2049                                struct btrfs_path *path,
2050                                struct extent_buffer *l,
2051                                struct extent_buffer *right,
2052                                int slot, int mid, int nritems)
2053 {
2054         int data_copy_size;
2055         int rt_data_off;
2056         int i;
2057         int ret = 0;
2058         int wret;
2059         struct btrfs_disk_key disk_key;
2060
2061         nritems = nritems - mid;
2062         btrfs_set_header_nritems(right, nritems);
2063         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2064
2065         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2066                            btrfs_item_nr_offset(mid),
2067                            nritems * sizeof(struct btrfs_item));
2068
2069         copy_extent_buffer(right, l,
2070                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2071                      data_copy_size, btrfs_leaf_data(l) +
2072                      leaf_data_end(root, l), data_copy_size);
2073
2074         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2075                       btrfs_item_end_nr(l, mid);
2076
2077         for (i = 0; i < nritems; i++) {
2078                 struct btrfs_item *item = btrfs_item_nr(right, i);
2079                 u32 ioff = btrfs_item_offset(right, item);
2080                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2081         }
2082
2083         btrfs_set_header_nritems(l, mid);
2084         ret = 0;
2085         btrfs_item_key(right, &disk_key, 0);
2086         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2087                           path->slots[1] + 1, 1);
2088         if (wret)
2089                 ret = wret;
2090
2091         btrfs_mark_buffer_dirty(right);
2092         btrfs_mark_buffer_dirty(l);
2093         BUG_ON(path->slots[0] != slot);
2094
2095         if (mid <= slot) {
2096                 free_extent_buffer(path->nodes[0]);
2097                 path->nodes[0] = right;
2098                 path->slots[0] -= mid;
2099                 path->slots[1] += 1;
2100         } else {
2101                 free_extent_buffer(right);
2102         }
2103
2104         BUG_ON(path->slots[0] < 0);
2105
2106         return ret;
2107 }
2108
2109 /*
2110  * split the path's leaf in two, making sure there is at least data_size
2111  * available for the resulting leaf level of the path.
2112  *
2113  * returns 0 if all went well and < 0 on failure.
2114  */
2115 static noinline int split_leaf(struct btrfs_trans_handle *trans,
2116                                struct btrfs_root *root,
2117                                struct btrfs_key *ins_key,
2118                                struct btrfs_path *path, int data_size,
2119                                int extend)
2120 {
2121         struct btrfs_disk_key disk_key;
2122         struct extent_buffer *l;
2123         u32 nritems;
2124         int mid;
2125         int slot;
2126         struct extent_buffer *right;
2127         int ret = 0;
2128         int wret;
2129         int split;
2130         int num_doubles = 0;
2131
2132         /* first try to make some room by pushing left and right */
2133         if (data_size && ins_key->type != BTRFS_DIR_ITEM_KEY) {
2134                 wret = push_leaf_right(trans, root, path, data_size, 0);
2135                 if (wret < 0)
2136                         return wret;
2137                 if (wret) {
2138                         wret = push_leaf_left(trans, root, path, data_size, 0);
2139                         if (wret < 0)
2140                                 return wret;
2141                 }
2142                 l = path->nodes[0];
2143
2144                 /* did the pushes work? */
2145                 if (btrfs_leaf_free_space(root, l) >= data_size)
2146                         return 0;
2147         }
2148
2149         if (!path->nodes[1]) {
2150                 ret = insert_new_root(trans, root, path, 1);
2151                 if (ret)
2152                         return ret;
2153         }
2154 again:
2155         split = 1;
2156         l = path->nodes[0];
2157         slot = path->slots[0];
2158         nritems = btrfs_header_nritems(l);
2159         mid = (nritems + 1) / 2;
2160
2161         if (mid <= slot) {
2162                 if (nritems == 1 ||
2163                     leaf_space_used(l, mid, nritems - mid) + data_size >
2164                         BTRFS_LEAF_DATA_SIZE(root)) {
2165                         if (slot >= nritems) {
2166                                 split = 0;
2167                         } else {
2168                                 mid = slot;
2169                                 if (mid != nritems &&
2170                                     leaf_space_used(l, mid, nritems - mid) +
2171                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2172                                         split = 2;
2173                                 }
2174                         }
2175                 }
2176         } else {
2177                 if (leaf_space_used(l, 0, mid) + data_size >
2178                         BTRFS_LEAF_DATA_SIZE(root)) {
2179                         if (!extend && data_size && slot == 0) {
2180                                 split = 0;
2181                         } else if ((extend || !data_size) && slot == 0) {
2182                                 mid = 1;
2183                         } else {
2184                                 mid = slot;
2185                                 if (mid != nritems &&
2186                                     leaf_space_used(l, mid, nritems - mid) +
2187                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2188                                         split = 2 ;
2189                                 }
2190                         }
2191                 }
2192         }
2193
2194         if (split == 0)
2195                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2196         else
2197                 btrfs_item_key(l, &disk_key, mid);
2198
2199         right = btrfs_alloc_free_block(trans, root, root->leafsize,
2200                                         root->root_key.objectid,
2201                                         &disk_key, 0, l->start, 0);
2202         if (IS_ERR(right)) {
2203                 BUG_ON(1);
2204                 return PTR_ERR(right);
2205         }
2206
2207         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2208         btrfs_set_header_bytenr(right, right->start);
2209         btrfs_set_header_generation(right, trans->transid);
2210         btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
2211         btrfs_set_header_owner(right, root->root_key.objectid);
2212         btrfs_set_header_level(right, 0);
2213         write_extent_buffer(right, root->fs_info->fsid,
2214                             (unsigned long)btrfs_header_fsid(right),
2215                             BTRFS_FSID_SIZE);
2216
2217         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2218                             (unsigned long)btrfs_header_chunk_tree_uuid(right),
2219                             BTRFS_UUID_SIZE);
2220
2221         if (split == 0) {
2222                 if (mid <= slot) {
2223                         btrfs_set_header_nritems(right, 0);
2224                         wret = insert_ptr(trans, root, path,
2225                                           &disk_key, right->start,
2226                                           path->slots[1] + 1, 1);
2227                         if (wret)
2228                                 ret = wret;
2229
2230                         free_extent_buffer(path->nodes[0]);
2231                         path->nodes[0] = right;
2232                         path->slots[0] = 0;
2233                         path->slots[1] += 1;
2234                 } else {
2235                         btrfs_set_header_nritems(right, 0);
2236                         wret = insert_ptr(trans, root, path,
2237                                           &disk_key,
2238                                           right->start,
2239                                           path->slots[1], 1);
2240                         if (wret)
2241                                 ret = wret;
2242                         free_extent_buffer(path->nodes[0]);
2243                         path->nodes[0] = right;
2244                         path->slots[0] = 0;
2245                         if (path->slots[1] == 0) {
2246                                 wret = fixup_low_keys(trans, root,
2247                                                 path, &disk_key, 1);
2248                                 if (wret)
2249                                         ret = wret;
2250                         }
2251                 }
2252                 btrfs_mark_buffer_dirty(right);
2253                 return ret;
2254         }
2255
2256         ret = copy_for_split(trans, root, path, l, right, slot, mid, nritems);
2257         BUG_ON(ret);
2258
2259         if (split == 2) {
2260                 BUG_ON(num_doubles != 0);
2261                 num_doubles++;
2262                 goto again;
2263         }
2264
2265         return ret;
2266 }
2267
2268 /*
2269  * This function splits a single item into two items,
2270  * giving 'new_key' to the new item and splitting the
2271  * old one at split_offset (from the start of the item).
2272  *
2273  * The path may be released by this operation.  After
2274  * the split, the path is pointing to the old item.  The
2275  * new item is going to be in the same node as the old one.
2276  *
2277  * Note, the item being split must be smaller enough to live alone on
2278  * a tree block with room for one extra struct btrfs_item
2279  *
2280  * This allows us to split the item in place, keeping a lock on the
2281  * leaf the entire time.
2282  */
2283 int btrfs_split_item(struct btrfs_trans_handle *trans,
2284                      struct btrfs_root *root,
2285                      struct btrfs_path *path,
2286                      struct btrfs_key *new_key,
2287                      unsigned long split_offset)
2288 {
2289         u32 item_size;
2290         struct extent_buffer *leaf;
2291         struct btrfs_key orig_key;
2292         struct btrfs_item *item;
2293         struct btrfs_item *new_item;
2294         int ret = 0;
2295         int slot;
2296         u32 nritems;
2297         u32 orig_offset;
2298         struct btrfs_disk_key disk_key;
2299         char *buf;
2300
2301         leaf = path->nodes[0];
2302         btrfs_item_key_to_cpu(leaf, &orig_key, path->slots[0]);
2303         if (btrfs_leaf_free_space(root, leaf) >= sizeof(struct btrfs_item))
2304                 goto split;
2305
2306         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2307         btrfs_release_path(root, path);
2308
2309         path->search_for_split = 1;
2310
2311         ret = btrfs_search_slot(trans, root, &orig_key, path, 0, 1);
2312         path->search_for_split = 0;
2313
2314         /* if our item isn't there or got smaller, return now */
2315         if (ret != 0 || item_size != btrfs_item_size_nr(path->nodes[0],
2316                                                         path->slots[0])) {
2317                 return -EAGAIN;
2318         }
2319
2320         ret = split_leaf(trans, root, &orig_key, path, 0, 0);
2321         BUG_ON(ret);
2322
2323         BUG_ON(btrfs_leaf_free_space(root, leaf) < sizeof(struct btrfs_item));
2324         leaf = path->nodes[0];
2325
2326 split:
2327         item = btrfs_item_nr(leaf, path->slots[0]);
2328         orig_offset = btrfs_item_offset(leaf, item);
2329         item_size = btrfs_item_size(leaf, item);
2330
2331
2332         buf = kmalloc(item_size, GFP_NOFS);
2333         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
2334                             path->slots[0]), item_size);
2335         slot = path->slots[0] + 1;
2336         leaf = path->nodes[0];
2337
2338         nritems = btrfs_header_nritems(leaf);
2339
2340         if (slot != nritems) {
2341                 /* shift the items */
2342                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
2343                               btrfs_item_nr_offset(slot),
2344                               (nritems - slot) * sizeof(struct btrfs_item));
2345
2346         }
2347
2348         btrfs_cpu_key_to_disk(&disk_key, new_key);
2349         btrfs_set_item_key(leaf, &disk_key, slot);
2350
2351         new_item = btrfs_item_nr(leaf, slot);
2352
2353         btrfs_set_item_offset(leaf, new_item, orig_offset);
2354         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
2355
2356         btrfs_set_item_offset(leaf, item,
2357                               orig_offset + item_size - split_offset);
2358         btrfs_set_item_size(leaf, item, split_offset);
2359
2360         btrfs_set_header_nritems(leaf, nritems + 1);
2361
2362         /* write the data for the start of the original item */
2363         write_extent_buffer(leaf, buf,
2364                             btrfs_item_ptr_offset(leaf, path->slots[0]),
2365                             split_offset);
2366
2367         /* write the data for the new item */
2368         write_extent_buffer(leaf, buf + split_offset,
2369                             btrfs_item_ptr_offset(leaf, slot),
2370                             item_size - split_offset);
2371         btrfs_mark_buffer_dirty(leaf);
2372
2373         ret = 0;
2374         if (btrfs_leaf_free_space(root, leaf) < 0) {
2375                 btrfs_print_leaf(root, leaf);
2376                 BUG();
2377         }
2378         kfree(buf);
2379         return ret;
2380 }
2381
2382 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2383                         struct btrfs_root *root,
2384                         struct btrfs_path *path,
2385                         u32 new_size, int from_end)
2386 {
2387         int ret = 0;
2388         int slot;
2389         struct extent_buffer *leaf;
2390         struct btrfs_item *item;
2391         u32 nritems;
2392         unsigned int data_end;
2393         unsigned int old_data_start;
2394         unsigned int old_size;
2395         unsigned int size_diff;
2396         int i;
2397
2398         leaf = path->nodes[0];
2399         slot = path->slots[0];
2400
2401         old_size = btrfs_item_size_nr(leaf, slot);
2402         if (old_size == new_size)
2403                 return 0;
2404
2405         nritems = btrfs_header_nritems(leaf);
2406         data_end = leaf_data_end(root, leaf);
2407
2408         old_data_start = btrfs_item_offset_nr(leaf, slot);
2409
2410         size_diff = old_size - new_size;
2411
2412         BUG_ON(slot < 0);
2413         BUG_ON(slot >= nritems);
2414
2415         /*
2416          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2417          */
2418         /* first correct the data pointers */
2419         for (i = slot; i < nritems; i++) {
2420                 u32 ioff;
2421                 item = btrfs_item_nr(leaf, i);
2422                 ioff = btrfs_item_offset(leaf, item);
2423                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2424         }
2425
2426         /* shift the data */
2427         if (from_end) {
2428                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2429                               data_end + size_diff, btrfs_leaf_data(leaf) +
2430                               data_end, old_data_start + new_size - data_end);
2431         } else {
2432                 struct btrfs_disk_key disk_key;
2433                 u64 offset;
2434
2435                 btrfs_item_key(leaf, &disk_key, slot);
2436
2437                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2438                         unsigned long ptr;
2439                         struct btrfs_file_extent_item *fi;
2440
2441                         fi = btrfs_item_ptr(leaf, slot,
2442                                             struct btrfs_file_extent_item);
2443                         fi = (struct btrfs_file_extent_item *)(
2444                              (unsigned long)fi - size_diff);
2445
2446                         if (btrfs_file_extent_type(leaf, fi) ==
2447                             BTRFS_FILE_EXTENT_INLINE) {
2448                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2449                                 memmove_extent_buffer(leaf, ptr,
2450                                         (unsigned long)fi,
2451                                         offsetof(struct btrfs_file_extent_item,
2452                                                  disk_bytenr));
2453                         }
2454                 }
2455
2456                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2457                               data_end + size_diff, btrfs_leaf_data(leaf) +
2458                               data_end, old_data_start - data_end);
2459
2460                 offset = btrfs_disk_key_offset(&disk_key);
2461                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2462                 btrfs_set_item_key(leaf, &disk_key, slot);
2463                 if (slot == 0)
2464                         fixup_low_keys(trans, root, path, &disk_key, 1);
2465         }
2466
2467         item = btrfs_item_nr(leaf, slot);
2468         btrfs_set_item_size(leaf, item, new_size);
2469         btrfs_mark_buffer_dirty(leaf);
2470
2471         ret = 0;
2472         if (btrfs_leaf_free_space(root, leaf) < 0) {
2473                 btrfs_print_leaf(root, leaf);
2474                 BUG();
2475         }
2476         return ret;
2477 }
2478
2479 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2480                       struct btrfs_root *root, struct btrfs_path *path,
2481                       u32 data_size)
2482 {
2483         int ret = 0;
2484         int slot;
2485         struct extent_buffer *leaf;
2486         struct btrfs_item *item;
2487         u32 nritems;
2488         unsigned int data_end;
2489         unsigned int old_data;
2490         unsigned int old_size;
2491         int i;
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) < data_size) {
2499                 btrfs_print_leaf(root, leaf);
2500                 BUG();
2501         }
2502         slot = path->slots[0];
2503         old_data = btrfs_item_end_nr(leaf, slot);
2504
2505         BUG_ON(slot < 0);
2506         if (slot >= nritems) {
2507                 btrfs_print_leaf(root, leaf);
2508                 printk("slot %d too large, nritems %d\n", slot, nritems);
2509                 BUG_ON(1);
2510         }
2511
2512         /*
2513          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2514          */
2515         /* first correct the data pointers */
2516         for (i = slot; i < nritems; i++) {
2517                 u32 ioff;
2518                 item = btrfs_item_nr(leaf, i);
2519                 ioff = btrfs_item_offset(leaf, item);
2520                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2521         }
2522
2523         /* shift the data */
2524         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2525                       data_end - data_size, btrfs_leaf_data(leaf) +
2526                       data_end, old_data - data_end);
2527
2528         data_end = old_data;
2529         old_size = btrfs_item_size_nr(leaf, slot);
2530         item = btrfs_item_nr(leaf, slot);
2531         btrfs_set_item_size(leaf, item, old_size + data_size);
2532         btrfs_mark_buffer_dirty(leaf);
2533
2534         ret = 0;
2535         if (btrfs_leaf_free_space(root, leaf) < 0) {
2536                 btrfs_print_leaf(root, leaf);
2537                 BUG();
2538         }
2539         return ret;
2540 }
2541
2542 /*
2543  * Given a key and some data, insert an item into the tree.
2544  * This does all the path init required, making room in the tree if needed.
2545  */
2546 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2547                             struct btrfs_root *root,
2548                             struct btrfs_path *path,
2549                             struct btrfs_key *cpu_key, u32 *data_size,
2550                             int nr)
2551 {
2552         struct extent_buffer *leaf;
2553         struct btrfs_item *item;
2554         int ret = 0;
2555         int slot;
2556         int i;
2557         u32 nritems;
2558         u32 total_size = 0;
2559         u32 total_data = 0;
2560         unsigned int data_end;
2561         struct btrfs_disk_key disk_key;
2562
2563         for (i = 0; i < nr; i++) {
2564                 total_data += data_size[i];
2565         }
2566
2567         /* create a root if there isn't one */
2568         if (!root->node)
2569                 BUG();
2570
2571         total_size = total_data + nr * sizeof(struct btrfs_item);
2572         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2573         if (ret == 0) {
2574                 return -EEXIST;
2575         }
2576         if (ret < 0)
2577                 goto out;
2578
2579         leaf = path->nodes[0];
2580
2581         nritems = btrfs_header_nritems(leaf);
2582         data_end = leaf_data_end(root, leaf);
2583
2584         if (btrfs_leaf_free_space(root, leaf) < total_size) {
2585                 btrfs_print_leaf(root, leaf);
2586                 printk("not enough freespace need %u have %d\n",
2587                        total_size, btrfs_leaf_free_space(root, leaf));
2588                 BUG();
2589         }
2590
2591         slot = path->slots[0];
2592         BUG_ON(slot < 0);
2593
2594         if (slot != nritems) {
2595                 int i;
2596                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2597
2598                 if (old_data < data_end) {
2599                         btrfs_print_leaf(root, leaf);
2600                         printk("slot %d old_data %d data_end %d\n",
2601                                slot, old_data, data_end);
2602                         BUG_ON(1);
2603                 }
2604                 /*
2605                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2606                  */
2607                 /* first correct the data pointers */
2608                 for (i = slot; i < nritems; i++) {
2609                         u32 ioff;
2610
2611                         item = btrfs_item_nr(leaf, i);
2612                         ioff = btrfs_item_offset(leaf, item);
2613                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2614                 }
2615
2616                 /* shift the items */
2617                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2618                               btrfs_item_nr_offset(slot),
2619                               (nritems - slot) * sizeof(struct btrfs_item));
2620
2621                 /* shift the data */
2622                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2623                               data_end - total_data, btrfs_leaf_data(leaf) +
2624                               data_end, old_data - data_end);
2625                 data_end = old_data;
2626         }
2627
2628         /* setup the item for the new data */
2629         for (i = 0; i < nr; i++) {
2630                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2631                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2632                 item = btrfs_item_nr(leaf, slot + i);
2633                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2634                 data_end -= data_size[i];
2635                 btrfs_set_item_size(leaf, item, data_size[i]);
2636         }
2637         btrfs_set_header_nritems(leaf, nritems + nr);
2638         btrfs_mark_buffer_dirty(leaf);
2639
2640         ret = 0;
2641         if (slot == 0) {
2642                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2643                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2644         }
2645
2646         if (btrfs_leaf_free_space(root, leaf) < 0) {
2647                 btrfs_print_leaf(root, leaf);
2648                 BUG();
2649         }
2650
2651 out:
2652         return ret;
2653 }
2654
2655 /*
2656  * Given a key and some data, insert an item into the tree.
2657  * This does all the path init required, making room in the tree if needed.
2658  */
2659 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2660                       *root, struct btrfs_key *cpu_key, void *data, u32
2661                       data_size)
2662 {
2663         int ret = 0;
2664         struct btrfs_path *path;
2665         struct extent_buffer *leaf;
2666         unsigned long ptr;
2667
2668         path = btrfs_alloc_path();
2669         BUG_ON(!path);
2670         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2671         if (!ret) {
2672                 leaf = path->nodes[0];
2673                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2674                 write_extent_buffer(leaf, data, ptr, data_size);
2675                 btrfs_mark_buffer_dirty(leaf);
2676         }
2677         btrfs_free_path(path);
2678         return ret;
2679 }
2680
2681 /*
2682  * delete the pointer from a given node.
2683  *
2684  * If the delete empties a node, the node is removed from the tree,
2685  * continuing all the way the root if required.  The root is converted into
2686  * a leaf if all the nodes are emptied.
2687  */
2688 int btrfs_del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2689                    struct btrfs_path *path, int level, int slot)
2690 {
2691         struct extent_buffer *parent = path->nodes[level];
2692         u32 nritems;
2693         int ret = 0;
2694         int wret;
2695
2696         nritems = btrfs_header_nritems(parent);
2697         if (slot != nritems -1) {
2698                 memmove_extent_buffer(parent,
2699                               btrfs_node_key_ptr_offset(slot),
2700                               btrfs_node_key_ptr_offset(slot + 1),
2701                               sizeof(struct btrfs_key_ptr) *
2702                               (nritems - slot - 1));
2703         }
2704         nritems--;
2705         btrfs_set_header_nritems(parent, nritems);
2706         if (nritems == 0 && parent == root->node) {
2707                 BUG_ON(btrfs_header_level(root->node) != 1);
2708                 /* just turn the root into a leaf and break */
2709                 btrfs_set_header_level(root->node, 0);
2710         } else if (slot == 0) {
2711                 struct btrfs_disk_key disk_key;
2712
2713                 btrfs_node_key(parent, &disk_key, 0);
2714                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2715                 if (wret)
2716                         ret = wret;
2717         }
2718         btrfs_mark_buffer_dirty(parent);
2719         return ret;
2720 }
2721
2722 /*
2723  * a helper function to delete the leaf pointed to by path->slots[1] and
2724  * path->nodes[1].
2725  *
2726  * This deletes the pointer in path->nodes[1] and frees the leaf
2727  * block extent.  zero is returned if it all worked out, < 0 otherwise.
2728  *
2729  * The path must have already been setup for deleting the leaf, including
2730  * all the proper balancing.  path->nodes[1] must be locked.
2731  */
2732 static noinline int btrfs_del_leaf(struct btrfs_trans_handle *trans,
2733                                    struct btrfs_root *root,
2734                                    struct btrfs_path *path,
2735                                    struct extent_buffer *leaf)
2736 {
2737         int ret;
2738
2739         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
2740         ret = btrfs_del_ptr(trans, root, path, 1, path->slots[1]);
2741         if (ret)
2742                 return ret;
2743
2744         ret = btrfs_free_extent(trans, root, leaf->start, leaf->len,
2745                                 0, root->root_key.objectid, 0, 0);
2746         return ret;
2747 }
2748
2749 /*
2750  * delete the item at the leaf level in path.  If that empties
2751  * the leaf, remove it from the tree
2752  */
2753 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2754                     struct btrfs_path *path, int slot, int nr)
2755 {
2756         struct extent_buffer *leaf;
2757         struct btrfs_item *item;
2758         int last_off;
2759         int dsize = 0;
2760         int ret = 0;
2761         int wret;
2762         int i;
2763         u32 nritems;
2764
2765         leaf = path->nodes[0];
2766         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2767
2768         for (i = 0; i < nr; i++)
2769                 dsize += btrfs_item_size_nr(leaf, slot + i);
2770
2771         nritems = btrfs_header_nritems(leaf);
2772
2773         if (slot + nr != nritems) {
2774                 int i;
2775                 int data_end = leaf_data_end(root, leaf);
2776
2777                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2778                               data_end + dsize,
2779                               btrfs_leaf_data(leaf) + data_end,
2780                               last_off - data_end);
2781
2782                 for (i = slot + nr; i < nritems; i++) {
2783                         u32 ioff;
2784
2785                         item = btrfs_item_nr(leaf, i);
2786                         ioff = btrfs_item_offset(leaf, item);
2787                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2788                 }
2789
2790                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2791                               btrfs_item_nr_offset(slot + nr),
2792                               sizeof(struct btrfs_item) *
2793                               (nritems - slot - nr));
2794         }
2795         btrfs_set_header_nritems(leaf, nritems - nr);
2796         nritems -= nr;
2797
2798         /* delete the leaf if we've emptied it */
2799         if (nritems == 0) {
2800                 if (leaf == root->node) {
2801                         btrfs_set_header_level(leaf, 0);
2802                 } else {
2803                         clean_tree_block(trans, root, leaf);
2804                         wait_on_tree_block_writeback(root, leaf);
2805
2806                         wret = btrfs_del_leaf(trans, root, path, leaf);
2807                         BUG_ON(ret);
2808                         if (wret)
2809                                 ret = wret;
2810                 }
2811         } else {
2812                 int used = leaf_space_used(leaf, 0, nritems);
2813                 if (slot == 0) {
2814                         struct btrfs_disk_key disk_key;
2815
2816                         btrfs_item_key(leaf, &disk_key, 0);
2817                         wret = fixup_low_keys(trans, root, path,
2818                                               &disk_key, 1);
2819                         if (wret)
2820                                 ret = wret;
2821                 }
2822
2823                 /* delete the leaf if it is mostly empty */
2824                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2825                         /* push_leaf_left fixes the path.
2826                          * make sure the path still points to our leaf
2827                          * for possible call to del_ptr below
2828                          */
2829                         slot = path->slots[1];
2830                         extent_buffer_get(leaf);
2831
2832                         wret = push_leaf_left(trans, root, path, 1, 1);
2833                         if (wret < 0 && wret != -ENOSPC)
2834                                 ret = wret;
2835
2836                         if (path->nodes[0] == leaf &&
2837                             btrfs_header_nritems(leaf)) {
2838                                 wret = push_leaf_right(trans, root, path, 1, 1);
2839                                 if (wret < 0 && wret != -ENOSPC)
2840                                         ret = wret;
2841                         }
2842
2843                         if (btrfs_header_nritems(leaf) == 0) {
2844                                 clean_tree_block(trans, root, leaf);
2845                                 wait_on_tree_block_writeback(root, leaf);
2846
2847                                 path->slots[1] = slot;
2848                                 ret = btrfs_del_leaf(trans, root, path, leaf);
2849                                 BUG_ON(ret);
2850                                 free_extent_buffer(leaf);
2851
2852                         } else {
2853                                 btrfs_mark_buffer_dirty(leaf);
2854                                 free_extent_buffer(leaf);
2855                         }
2856                 } else {
2857                         btrfs_mark_buffer_dirty(leaf);
2858                 }
2859         }
2860         return ret;
2861 }
2862
2863 /*
2864  * walk up the tree as far as required to find the previous leaf.
2865  * returns 0 if it found something or 1 if there are no lesser leaves.
2866  * returns < 0 on io errors.
2867  */
2868 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2869 {
2870         int slot;
2871         int level = 1;
2872         struct extent_buffer *c;
2873         struct extent_buffer *next = NULL;
2874
2875         while(level < BTRFS_MAX_LEVEL) {
2876                 if (!path->nodes[level])
2877                         return 1;
2878
2879                 slot = path->slots[level];
2880                 c = path->nodes[level];
2881                 if (slot == 0) {
2882                         level++;
2883                         if (level == BTRFS_MAX_LEVEL)
2884                                 return 1;
2885                         continue;
2886                 }
2887                 slot--;
2888
2889                 next = read_node_slot(root, c, slot);
2890                 break;
2891         }
2892         path->slots[level] = slot;
2893         while(1) {
2894                 level--;
2895                 c = path->nodes[level];
2896                 free_extent_buffer(c);
2897                 slot = btrfs_header_nritems(next);
2898                 if (slot != 0)
2899                         slot--;
2900                 path->nodes[level] = next;
2901                 path->slots[level] = slot;
2902                 if (!level)
2903                         break;
2904                 next = read_node_slot(root, next, slot);
2905         }
2906         return 0;
2907 }
2908
2909 /*
2910  * walk up the tree as far as required to find the next leaf.
2911  * returns 0 if it found something or 1 if there are no greater leaves.
2912  * returns < 0 on io errors.
2913  */
2914 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2915 {
2916         int slot;
2917         int level = 1;
2918         struct extent_buffer *c;
2919         struct extent_buffer *next = NULL;
2920
2921         while(level < BTRFS_MAX_LEVEL) {
2922                 if (!path->nodes[level])
2923                         return 1;
2924
2925                 slot = path->slots[level] + 1;
2926                 c = path->nodes[level];
2927                 if (slot >= btrfs_header_nritems(c)) {
2928                         level++;
2929                         if (level == BTRFS_MAX_LEVEL)
2930                                 return 1;
2931                         continue;
2932                 }
2933
2934                 if (path->reada)
2935                         reada_for_search(root, path, level, slot, 0);
2936
2937                 next = read_node_slot(root, c, slot);
2938                 if (!next)
2939                         return -EIO;
2940                 break;
2941         }
2942         path->slots[level] = slot;
2943         while(1) {
2944                 level--;
2945                 c = path->nodes[level];
2946                 free_extent_buffer(c);
2947                 path->nodes[level] = next;
2948                 path->slots[level] = 0;
2949                 if (!level)
2950                         break;
2951                 if (path->reada)
2952                         reada_for_search(root, path, level, 0, 0);
2953                 next = read_node_slot(root, next, 0);
2954                 if (!next)
2955                         return -EIO;
2956         }
2957         return 0;
2958 }
2959
2960 int btrfs_previous_item(struct btrfs_root *root,
2961                         struct btrfs_path *path, u64 min_objectid,
2962                         int type)
2963 {
2964         struct btrfs_key found_key;
2965         struct extent_buffer *leaf;
2966         int ret;
2967
2968         while(1) {
2969                 if (path->slots[0] == 0) {
2970                         ret = btrfs_prev_leaf(root, path);
2971                         if (ret != 0)
2972                                 return ret;
2973                 } else {
2974                         path->slots[0]--;
2975                 }
2976                 leaf = path->nodes[0];
2977                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2978                 if (found_key.type == type)
2979                         return 0;
2980         }
2981         return 1;
2982 }
2983