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         /*
 836          * deal with the case where there is only one pointer in the root
 837          * by promoting the node below to a root
 838          */
 839         if (!parent) {
 840                 struct extent_buffer *child;
 841
 842                 if (btrfs_header_nritems(mid) != 1)
 843                         return 0;
 844
 845                 /* promote the child to a root */
 846                 child = read_node_slot(root, mid, 0);
 847                 BUG_ON(!child);
 848                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
 849                 BUG_ON(ret);
 850
 851                 root->node = child;
 852                 add_root_to_dirty_list(root);
 853                 path->nodes[level] = NULL;
 854                 clean_tree_block(trans, root, mid);
 855                 wait_on_tree_block_writeback(root, mid);
 856                 /* once for the path */
 857                 free_extent_buffer(mid);
 858
 859                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 860                                         0, root->root_key.objectid,
 861                                         level, 1);
 862                 /* once for the root ptr */
 863                 free_extent_buffer(mid);
 864                 return ret;
 865         }
 866         if (btrfs_header_nritems(mid) >
 867             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 868                 return 0;
 869
 870         left = read_node_slot(root, parent, pslot - 1);
 871         if (left) {
 872                 wret = btrfs_cow_block(trans, root, left,
 873                                        parent, pslot - 1, &left);
 874                 if (wret) {
 875                         ret = wret;
 876                         goto enospc;
 877                 }
 878         }
 879         right = read_node_slot(root, parent, pslot + 1);
 880         if (right) {
 881                 wret = btrfs_cow_block(trans, root, right,
 882                                        parent, pslot + 1, &right);
 883                 if (wret) {
 884                         ret = wret;
 885                         goto enospc;
 886                 }
 887         }
 888
 889         /* first, try to make some room in the middle buffer */
 890         if (left) {
 891                 orig_slot += btrfs_header_nritems(left);
 892                 wret = push_node_left(trans, root, left, mid, 1);
 893                 if (wret < 0)
 894                         ret = wret;
 895         }
 896
 897         /*
 898          * then try to empty the right most buffer into the middle
 899          */
 900         if (right) {
 901                 wret = push_node_left(trans, root, mid, right, 1);
 902                 if (wret < 0 && wret != -ENOSPC)
 903                         ret = wret;
 904                 if (btrfs_header_nritems(right) == 0) {
 905                         u64 bytenr = right->start;
 906                         u32 blocksize = right->len;
 907
 908                         clean_tree_block(trans, root, right);
 909                         wait_on_tree_block_writeback(root, right);
 910                         free_extent_buffer(right);
 911                         right = NULL;
 912                         wret = btrfs_del_ptr(trans, root, path,
 913                                              level + 1, pslot + 1);
 914                         if (wret)
 915                                 ret = wret;
 916                         wret = btrfs_free_extent(trans, root, bytenr,
 917                                                  blocksize, 0,
 918                                                  root->root_key.objectid,
 919                                                  level, 0);
 920                         if (wret)
 921                                 ret = wret;
 922                 } else {
 923                         struct btrfs_disk_key right_key;
 924                         btrfs_node_key(right, &right_key, 0);
 925                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 926                         btrfs_mark_buffer_dirty(parent);
 927                 }
 928         }
 929         if (btrfs_header_nritems(mid) == 1) {
 930                 /*
 931                  * we're not allowed to leave a node with one item in the
 932                  * tree during a delete.  A deletion from lower in the tree
 933                  * could try to delete the only pointer in this node.
 934                  * So, pull some keys from the left.
 935                  * There has to be a left pointer at this point because
 936                  * otherwise we would have pulled some pointers from the
 937                  * right
 938                  */
 939                 BUG_ON(!left);
 940                 wret = balance_node_right(trans, root, mid, left);
 941                 if (wret < 0) {
 942                         ret = wret;
 943                         goto enospc;
 944                 }
 945                 if (wret == 1) {
 946                         wret = push_node_left(trans, root, left, mid, 1);
 947                         if (wret < 0)
 948                                 ret = wret;
 949                 }
 950                 BUG_ON(wret == 1);
 951         }
 952         if (btrfs_header_nritems(mid) == 0) {
 953                 /* we've managed to empty the middle node, drop it */
 954                 u64 bytenr = mid->start;
 955                 u32 blocksize = mid->len;
 956                 clean_tree_block(trans, root, mid);
 957                 wait_on_tree_block_writeback(root, mid);
 958                 free_extent_buffer(mid);
 959                 mid = NULL;
 960                 wret = btrfs_del_ptr(trans, root, path, level + 1, pslot);
 961                 if (wret)
 962                         ret = wret;
 963                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 964                                          0, root->root_key.objectid,
 965                                          level, 0);
 966                 if (wret)
 967                         ret = wret;
 968         } else {
 969                 /* update the parent key to reflect our changes */
 970                 struct btrfs_disk_key mid_key;
 971                 btrfs_node_key(mid, &mid_key, 0);
 972                 btrfs_set_node_key(parent, &mid_key, pslot);
 973                 btrfs_mark_buffer_dirty(parent);
 974         }
 975
 976         /* update the path */
 977         if (left) {
 978                 if (btrfs_header_nritems(left) > orig_slot) {
 979                         extent_buffer_get(left);
 980                         path->nodes[level] = left;
 981                         path->slots[level + 1] -= 1;
 982                         path->slots[level] = orig_slot;
 983                         if (mid)
 984                                 free_extent_buffer(mid);
 985                 } else {
 986                         orig_slot -= btrfs_header_nritems(left);
 987                         path->slots[level] = orig_slot;
 988                 }
 989         }
 990         /* double check we haven't messed things up */
 991         check_block(root, path, level);
 992         if (orig_ptr !=
 993             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 994                 BUG();
 995 enospc:
 996         if (right)
 997                 free_extent_buffer(right);
 998         if (left)
 999                 free_extent_buffer(left);
1000         return ret;
1001 }
1002
1003 /* returns zero if the push worked, non-zero otherwise */
1004 static int noinline push_nodes_for_insert(struct btrfs_trans_handle *trans,
1005                                           struct btrfs_root *root,
1006                                           struct btrfs_path *path, int level)
1007 {
1008         struct extent_buffer *right = NULL;
1009         struct extent_buffer *mid;
1010         struct extent_buffer *left = NULL;
1011         struct extent_buffer *parent = NULL;
1012         int ret = 0;
1013         int wret;
1014         int pslot;
1015         int orig_slot = path->slots[level];
1016
1017         if (level == 0)
1018                 return 1;
1019
1020         mid = path->nodes[level];
1021         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1022
1023         if (level < BTRFS_MAX_LEVEL - 1) {
1024                 parent = path->nodes[level + 1];
1025                 pslot = path->slots[level + 1];
1026         }
1027
1028         if (!parent)
1029                 return 1;
1030
1031         left = read_node_slot(root, parent, pslot - 1);
1032
1033         /* first, try to make some room in the middle buffer */
1034         if (left) {
1035                 u32 left_nr;
1036                 left_nr = btrfs_header_nritems(left);
1037                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1038                         wret = 1;
1039                 } else {
1040                         ret = btrfs_cow_block(trans, root, left, parent,
1041                                               pslot - 1, &left);
1042                         if (ret)
1043                                 wret = 1;
1044                         else {
1045                                 wret = push_node_left(trans, root,
1046                                                       left, mid, 0);
1047                         }
1048                 }
1049                 if (wret < 0)
1050                         ret = wret;
1051                 if (wret == 0) {
1052                         struct btrfs_disk_key disk_key;
1053                         orig_slot += left_nr;
1054                         btrfs_node_key(mid, &disk_key, 0);
1055                         btrfs_set_node_key(parent, &disk_key, pslot);
1056                         btrfs_mark_buffer_dirty(parent);
1057                         if (btrfs_header_nritems(left) > orig_slot) {
1058                                 path->nodes[level] = left;
1059                                 path->slots[level + 1] -= 1;
1060                                 path->slots[level] = orig_slot;
1061                                 free_extent_buffer(mid);
1062                         } else {
1063                                 orig_slot -=
1064                                         btrfs_header_nritems(left);
1065                                 path->slots[level] = orig_slot;
1066                                 free_extent_buffer(left);
1067                         }
1068                         return 0;
1069                 }
1070                 free_extent_buffer(left);
1071         }
1072         right= read_node_slot(root, parent, pslot + 1);
1073
1074         /*
1075          * then try to empty the right most buffer into the middle
1076          */
1077         if (right) {
1078                 u32 right_nr;
1079                 right_nr = btrfs_header_nritems(right);
1080                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1081                         wret = 1;
1082                 } else {
1083                         ret = btrfs_cow_block(trans, root, right,
1084                                               parent, pslot + 1,
1085                                               &right);
1086                         if (ret)
1087                                 wret = 1;
1088                         else {
1089                                 wret = balance_node_right(trans, root,
1090                                                           right, mid);
1091                         }
1092                 }
1093                 if (wret < 0)
1094                         ret = wret;
1095                 if (wret == 0) {
1096                         struct btrfs_disk_key disk_key;
1097
1098                         btrfs_node_key(right, &disk_key, 0);
1099                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
1100                         btrfs_mark_buffer_dirty(parent);
1101
1102                         if (btrfs_header_nritems(mid) <= orig_slot) {
1103                                 path->nodes[level] = right;
1104                                 path->slots[level + 1] += 1;
1105                                 path->slots[level] = orig_slot -
1106                                         btrfs_header_nritems(mid);
1107                                 free_extent_buffer(mid);
1108                         } else {
1109                                 free_extent_buffer(right);
1110                         }
1111                         return 0;
1112                 }
1113                 free_extent_buffer(right);
1114         }
1115         return 1;
1116 }
1117
1118 /*
1119  * readahead one full node of leaves
1120  */
1121 void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
1122                              int level, int slot, u64 objectid)
1123 {
1124         struct extent_buffer *node;
1125         struct btrfs_disk_key disk_key;
1126         u32 nritems;
1127         u64 search;
1128         u64 lowest_read;
1129         u64 highest_read;
1130         u64 nread = 0;
1131         int direction = path->reada;
1132         struct extent_buffer *eb;
1133         u32 nr;
1134         u32 blocksize;
1135         u32 nscan = 0;
1136
1137         if (level != 1)
1138                 return;
1139
1140         if (!path->nodes[level])
1141                 return;
1142
1143         node = path->nodes[level];
1144         search = btrfs_node_blockptr(node, slot);
1145         blocksize = btrfs_level_size(root, level - 1);
1146         eb = btrfs_find_tree_block(root, search, blocksize);
1147         if (eb) {
1148                 free_extent_buffer(eb);
1149                 return;
1150         }
1151
1152         highest_read = search;
1153         lowest_read = search;
1154
1155         nritems = btrfs_header_nritems(node);
1156         nr = slot;
1157         while(1) {
1158                 if (direction < 0) {
1159                         if (nr == 0)
1160                                 break;
1161                         nr--;
1162                 } else if (direction > 0) {
1163                         nr++;
1164                         if (nr >= nritems)
1165                                 break;
1166                 }
1167                 if (path->reada < 0 && objectid) {
1168                         btrfs_node_key(node, &disk_key, nr);
1169                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1170                                 break;
1171                 }
1172                 search = btrfs_node_blockptr(node, nr);
1173                 if ((search >= lowest_read && search <= highest_read) ||
1174                     (search < lowest_read && lowest_read - search <= 32768) ||
1175                     (search > highest_read && search - highest_read <= 32768)) {
1176                         readahead_tree_block(root, search, blocksize,
1177                                      btrfs_node_ptr_generation(node, nr));
1178                         nread += blocksize;
1179                 }
1180                 nscan++;
1181                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
1182                         break;
1183                 if(nread > (1024 * 1024) || nscan > 128)
1184                         break;
1185
1186                 if (search < lowest_read)
1187                         lowest_read = search;
1188                 if (search > highest_read)
1189                         highest_read = search;
1190         }
1191 }
1192
1193 /*
1194  * look for key in the tree.  path is filled in with nodes along the way
1195  * if key is found, we return zero and you can find the item in the leaf
1196  * level of the path (level 0)
1197  *
1198  * If the key isn't found, the path points to the slot where it should
1199  * be inserted, and 1 is returned.  If there are other errors during the
1200  * search a negative error number is returned.
1201  *
1202  * if ins_len > 0, nodes and leaves will be split as we walk down the
1203  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1204  * possible)
1205  */
1206 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1207                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1208                       ins_len, int cow)
1209 {
1210         struct extent_buffer *b;
1211         int slot;
1212         int ret;
1213         int level;
1214         int should_reada = p->reada;
1215         u8 lowest_level = 0;
1216
1217         lowest_level = p->lowest_level;
1218         WARN_ON(lowest_level && ins_len > 0);
1219         WARN_ON(p->nodes[0] != NULL);
1220         /*
1221         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
1222         */
1223 again:
1224         b = root->node;
1225         extent_buffer_get(b);
1226         while (b) {
1227                 level = btrfs_header_level(b);
1228                 if (cow) {
1229                         int wret;
1230                         wret = btrfs_cow_block(trans, root, b,
1231                                                p->nodes[level + 1],
1232                                                p->slots[level + 1],
1233                                                &b);
1234                         if (wret) {
1235                                 free_extent_buffer(b);
1236                                 return wret;
1237                         }
1238                 }
1239                 BUG_ON(!cow && ins_len);
1240                 if (level != btrfs_header_level(b))
1241                         WARN_ON(1);
1242                 level = btrfs_header_level(b);
1243                 p->nodes[level] = b;
1244                 ret = check_block(root, p, level);
1245                 if (ret)
1246                         return -1;
1247                 ret = bin_search(b, key, level, &slot);
1248                 if (level != 0) {
1249                         if (ret && slot > 0)
1250                                 slot -= 1;
1251                         p->slots[level] = slot;
1252                         if ((p->search_for_split || ins_len > 0) &&
1253                             btrfs_header_nritems(b) >=
1254                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1255                                 int sret = split_node(trans, root, p, level);
1256                                 BUG_ON(sret > 0);
1257                                 if (sret)
1258                                         return sret;
1259                                 b = p->nodes[level];
1260                                 slot = p->slots[level];
1261                         } else if (ins_len < 0) {
1262                                 int sret = balance_level(trans, root, p,
1263                                                          level);
1264                                 if (sret)
1265                                         return sret;
1266                                 b = p->nodes[level];
1267                                 if (!b) {
1268                                         btrfs_release_path(NULL, p);
1269                                         goto again;
1270                                 }
1271                                 slot = p->slots[level];
1272                                 BUG_ON(btrfs_header_nritems(b) == 1);
1273                         }
1274                         /* this is only true while dropping a snapshot */
1275                         if (level == lowest_level)
1276                                 break;
1277
1278                         if (should_reada)
1279                                 reada_for_search(root, p, level, slot,
1280                                                  key->objectid);
1281
1282                         b = read_node_slot(root, b, slot);
1283                         if (!extent_buffer_uptodate(b))
1284                                 return -EIO;
1285                 } else {
1286                         p->slots[level] = slot;
1287                         if (ins_len > 0 &&
1288                             ins_len > btrfs_leaf_free_space(root, b)) {
1289                                 int sret = split_leaf(trans, root, key,
1290                                                       p, ins_len, ret == 0);
1291                                 BUG_ON(sret > 0);
1292                                 if (sret)
1293                                         return sret;
1294                         }
1295                         return ret;
1296                 }
1297         }
1298         return 1;
1299 }
1300
1301 /*
1302  * adjust the pointers going up the tree, starting at level
1303  * making sure the right key of each node is points to 'key'.
1304  * This is used after shifting pointers to the left, so it stops
1305  * fixing up pointers when a given leaf/node is not in slot 0 of the
1306  * higher levels
1307  *
1308  * If this fails to write a tree block, it returns -1, but continues
1309  * fixing up the blocks in ram so the tree is consistent.
1310  */
1311 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1312                           struct btrfs_root *root, struct btrfs_path *path,
1313                           struct btrfs_disk_key *key, int level)
1314 {
1315         int i;
1316         int ret = 0;
1317         struct extent_buffer *t;
1318
1319         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1320                 int tslot = path->slots[i];
1321                 if (!path->nodes[i])
1322                         break;
1323                 t = path->nodes[i];
1324                 btrfs_set_node_key(t, key, tslot);
1325                 btrfs_mark_buffer_dirty(path->nodes[i]);
1326                 if (tslot != 0)
1327                         break;
1328         }
1329         return ret;
1330 }
1331
1332 /*
1333  * update item key.
1334  *
1335  * This function isn't completely safe. It's the caller's responsibility
1336  * that the new key won't break the order
1337  */
1338 int btrfs_set_item_key_safe(struct btrfs_trans_handle *trans,
1339                             struct btrfs_root *root, struct btrfs_path *path,
1340                             struct btrfs_key *new_key)
1341 {
1342         struct btrfs_disk_key disk_key;
1343         struct extent_buffer *eb;
1344         int slot;
1345
1346         eb = path->nodes[0];
1347         slot = path->slots[0];
1348         if (slot > 0) {
1349                 btrfs_item_key(eb, &disk_key, slot - 1);
1350                 if (btrfs_comp_keys(&disk_key, new_key) >= 0)
1351                         return -1;
1352         }
1353         if (slot < btrfs_header_nritems(eb) - 1) {
1354                 btrfs_item_key(eb, &disk_key, slot + 1);
1355                 if (btrfs_comp_keys(&disk_key, new_key) <= 0)
1356                         return -1;
1357         }
1358
1359         btrfs_cpu_key_to_disk(&disk_key, new_key);
1360         btrfs_set_item_key(eb, &disk_key, slot);
1361         btrfs_mark_buffer_dirty(eb);
1362         if (slot == 0)
1363                 fixup_low_keys(trans, root, path, &disk_key, 1);
1364         return 0;
1365 }
1366
1367 /*
1368  * try to push data from one node into the next node left in the
1369  * tree.
1370  *
1371  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1372  * error, and > 0 if there was no room in the left hand block.
1373  */
1374 static int push_node_left(struct btrfs_trans_handle *trans,
1375                           struct btrfs_root *root, struct extent_buffer *dst,
1376                           struct extent_buffer *src, int empty)
1377 {
1378         int push_items = 0;
1379         int src_nritems;
1380         int dst_nritems;
1381         int ret = 0;
1382
1383         src_nritems = btrfs_header_nritems(src);
1384         dst_nritems = btrfs_header_nritems(dst);
1385         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1386         WARN_ON(btrfs_header_generation(src) != trans->transid);
1387         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1388
1389         if (!empty && src_nritems <= 8)
1390                 return 1;
1391
1392         if (push_items <= 0) {
1393                 return 1;
1394         }
1395
1396         if (empty) {
1397                 push_items = min(src_nritems, push_items);
1398                 if (push_items < src_nritems) {
1399                         /* leave at least 8 pointers in the node if
1400                          * we aren't going to empty it
1401                          */
1402                         if (src_nritems - push_items < 8) {
1403                                 if (push_items <= 8)
1404                                         return 1;
1405                                 push_items -= 8;
1406                         }
1407                 }
1408         } else
1409                 push_items = min(src_nritems - 8, push_items);
1410
1411         copy_extent_buffer(dst, src,
1412                            btrfs_node_key_ptr_offset(dst_nritems),
1413                            btrfs_node_key_ptr_offset(0),
1414                            push_items * sizeof(struct btrfs_key_ptr));
1415
1416         if (push_items < src_nritems) {
1417                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1418                                       btrfs_node_key_ptr_offset(push_items),
1419                                       (src_nritems - push_items) *
1420                                       sizeof(struct btrfs_key_ptr));
1421         }
1422         btrfs_set_header_nritems(src, src_nritems - push_items);
1423         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1424         btrfs_mark_buffer_dirty(src);
1425         btrfs_mark_buffer_dirty(dst);
1426
1427         return ret;
1428 }
1429
1430 /*
1431  * try to push data from one node into the next node right in the
1432  * tree.
1433  *
1434  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1435  * error, and > 0 if there was no room in the right hand block.
1436  *
1437  * this will  only push up to 1/2 the contents of the left node over
1438  */
1439 static int balance_node_right(struct btrfs_trans_handle *trans,
1440                               struct btrfs_root *root,
1441                               struct extent_buffer *dst,
1442                               struct extent_buffer *src)
1443 {
1444         int push_items = 0;
1445         int max_push;
1446         int src_nritems;
1447         int dst_nritems;
1448         int ret = 0;
1449
1450         WARN_ON(btrfs_header_generation(src) != trans->transid);
1451         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1452
1453         src_nritems = btrfs_header_nritems(src);
1454         dst_nritems = btrfs_header_nritems(dst);
1455         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1456         if (push_items <= 0) {
1457                 return 1;
1458         }
1459
1460         if (src_nritems < 4) {
1461                 return 1;
1462         }
1463
1464         max_push = src_nritems / 2 + 1;
1465         /* don't try to empty the node */
1466         if (max_push >= src_nritems) {
1467                 return 1;
1468         }
1469
1470         if (max_push < push_items)
1471                 push_items = max_push;
1472
1473         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1474                                       btrfs_node_key_ptr_offset(0),
1475                                       (dst_nritems) *
1476                                       sizeof(struct btrfs_key_ptr));
1477
1478         copy_extent_buffer(dst, src,
1479                            btrfs_node_key_ptr_offset(0),
1480                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1481                            push_items * sizeof(struct btrfs_key_ptr));
1482
1483         btrfs_set_header_nritems(src, src_nritems - push_items);
1484         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1485
1486         btrfs_mark_buffer_dirty(src);
1487         btrfs_mark_buffer_dirty(dst);
1488
1489         return ret;
1490 }
1491
1492 /*
1493  * helper function to insert a new root level in the tree.
1494  * A new node is allocated, and a single item is inserted to
1495  * point to the existing root
1496  *
1497  * returns zero on success or < 0 on failure.
1498  */
1499 static int noinline insert_new_root(struct btrfs_trans_handle *trans,
1500                            struct btrfs_root *root,
1501                            struct btrfs_path *path, int level)
1502 {
1503         u64 lower_gen;
1504         struct extent_buffer *lower;
1505         struct extent_buffer *c;
1506         struct extent_buffer *old;
1507         struct btrfs_disk_key lower_key;
1508
1509         BUG_ON(path->nodes[level]);
1510         BUG_ON(path->nodes[level-1] != root->node);
1511
1512         lower = path->nodes[level-1];
1513         if (level == 1)
1514                 btrfs_item_key(lower, &lower_key, 0);
1515         else
1516                 btrfs_node_key(lower, &lower_key, 0);
1517
1518         c = btrfs_alloc_free_block(trans, root, root->nodesize,
1519                                    root->root_key.objectid, &lower_key,
1520                                    level, root->node->start, 0);
1521
1522         if (IS_ERR(c))
1523                 return PTR_ERR(c);
1524
1525         memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
1526         btrfs_set_header_nritems(c, 1);
1527         btrfs_set_header_level(c, level);
1528         btrfs_set_header_bytenr(c, c->start);
1529         btrfs_set_header_generation(c, trans->transid);
1530         btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
1531         btrfs_set_header_owner(c, root->root_key.objectid);
1532
1533         write_extent_buffer(c, root->fs_info->fsid,
1534                             (unsigned long)btrfs_header_fsid(c),
1535                             BTRFS_FSID_SIZE);
1536
1537         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
1538                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
1539                             BTRFS_UUID_SIZE);
1540
1541         btrfs_set_node_key(c, &lower_key, 0);
1542         btrfs_set_node_blockptr(c, 0, lower->start);
1543         lower_gen = btrfs_header_generation(lower);
1544         WARN_ON(lower_gen != trans->transid);
1545
1546         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1547
1548         btrfs_mark_buffer_dirty(c);
1549
1550         old = root->node;
1551         root->node = c;
1552
1553         /* the super has an extra ref to root->node */
1554         free_extent_buffer(old);
1555
1556         add_root_to_dirty_list(root);
1557         extent_buffer_get(c);
1558         path->nodes[level] = c;
1559         path->slots[level] = 0;
1560         return 0;
1561 }
1562
1563 /*
1564  * worker function to insert a single pointer in a node.
1565  * the node should have enough room for the pointer already
1566  *
1567  * slot and level indicate where you want the key to go, and
1568  * blocknr is the block the key points to.
1569  *
1570  * returns zero on success and < 0 on any error
1571  */
1572 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1573                       *root, struct btrfs_path *path, struct btrfs_disk_key
1574                       *key, u64 bytenr, int slot, int level)
1575 {
1576         struct extent_buffer *lower;
1577         int nritems;
1578
1579         BUG_ON(!path->nodes[level]);
1580         lower = path->nodes[level];
1581         nritems = btrfs_header_nritems(lower);
1582         if (slot > nritems)
1583                 BUG();
1584         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1585                 BUG();
1586         if (slot != nritems) {
1587                 memmove_extent_buffer(lower,
1588                               btrfs_node_key_ptr_offset(slot + 1),
1589                               btrfs_node_key_ptr_offset(slot),
1590                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1591         }
1592         btrfs_set_node_key(lower, key, slot);
1593         btrfs_set_node_blockptr(lower, slot, bytenr);
1594         WARN_ON(trans->transid == 0);
1595         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1596         btrfs_set_header_nritems(lower, nritems + 1);
1597         btrfs_mark_buffer_dirty(lower);
1598         return 0;
1599 }
1600
1601 /*
1602  * split the node at the specified level in path in two.
1603  * The path is corrected to point to the appropriate node after the split
1604  *
1605  * Before splitting this tries to make some room in the node by pushing
1606  * left and right, if either one works, it returns right away.
1607  *
1608  * returns 0 on success and < 0 on failure
1609  */
1610 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1611                       *root, struct btrfs_path *path, int level)
1612 {
1613         struct extent_buffer *c;
1614         struct extent_buffer *split;
1615         struct btrfs_disk_key disk_key;
1616         int mid;
1617         int ret;
1618         int wret;
1619         u32 c_nritems;
1620
1621         c = path->nodes[level];
1622         WARN_ON(btrfs_header_generation(c) != trans->transid);
1623         if (c == root->node) {
1624                 /* trying to split the root, lets make a new one */
1625                 ret = insert_new_root(trans, root, path, level + 1);
1626                 if (ret)
1627                         return ret;
1628         } else {
1629                 ret = push_nodes_for_insert(trans, root, path, level);
1630                 c = path->nodes[level];
1631                 if (!ret && btrfs_header_nritems(c) <
1632                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
1633                         return 0;
1634                 if (ret < 0)
1635                         return ret;
1636         }
1637
1638         c_nritems = btrfs_header_nritems(c);
1639         mid = (c_nritems + 1) / 2;
1640         btrfs_node_key(c, &disk_key, mid);
1641
1642         split = btrfs_alloc_free_block(trans, root, root->nodesize,
1643                                         root->root_key.objectid,
1644                                         &disk_key, level, c->start, 0);
1645         if (IS_ERR(split))
1646                 return PTR_ERR(split);
1647
1648         memset_extent_buffer(split, 0, 0, sizeof(struct btrfs_header));
1649         btrfs_set_header_level(split, btrfs_header_level(c));
1650         btrfs_set_header_bytenr(split, split->start);
1651         btrfs_set_header_generation(split, trans->transid);
1652         btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
1653         btrfs_set_header_owner(split, root->root_key.objectid);
1654         write_extent_buffer(split, root->fs_info->fsid,
1655                             (unsigned long)btrfs_header_fsid(split),
1656                             BTRFS_FSID_SIZE);
1657         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
1658                             (unsigned long)btrfs_header_chunk_tree_uuid(split),
1659                             BTRFS_UUID_SIZE);
1660
1661
1662         copy_extent_buffer(split, c,
1663                            btrfs_node_key_ptr_offset(0),
1664                            btrfs_node_key_ptr_offset(mid),
1665                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1666         btrfs_set_header_nritems(split, c_nritems - mid);
1667         btrfs_set_header_nritems(c, mid);
1668         ret = 0;
1669
1670         btrfs_mark_buffer_dirty(c);
1671         btrfs_mark_buffer_dirty(split);
1672
1673         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1674                           path->slots[level + 1] + 1,
1675                           level + 1);
1676         if (wret)
1677                 ret = wret;
1678
1679         if (path->slots[level] >= mid) {
1680                 path->slots[level] -= mid;
1681                 free_extent_buffer(c);
1682                 path->nodes[level] = split;
1683                 path->slots[level + 1] += 1;
1684         } else {
1685                 free_extent_buffer(split);
1686         }
1687         return ret;
1688 }
1689
1690 /*
1691  * how many bytes are required to store the items in a leaf.  start
1692  * and nr indicate which items in the leaf to check.  This totals up the
1693  * space used both by the item structs and the item data
1694  */
1695 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1696 {
1697         int data_len;
1698         int nritems = btrfs_header_nritems(l);
1699         int end = min(nritems, start + nr) - 1;
1700
1701         if (!nr)
1702                 return 0;
1703         data_len = btrfs_item_end_nr(l, start);
1704         data_len = data_len - btrfs_item_offset_nr(l, end);
1705         data_len += sizeof(struct btrfs_item) * nr;
1706         WARN_ON(data_len < 0);
1707         return data_len;
1708 }
1709
1710 /*
1711  * The space between the end of the leaf items and
1712  * the start of the leaf data.  IOW, how much room
1713  * the leaf has left for both items and data
1714  */
1715 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1716 {
1717         int nritems = btrfs_header_nritems(leaf);
1718         int ret;
1719         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1720         if (ret < 0) {
1721                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1722                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1723                        leaf_space_used(leaf, 0, nritems), nritems);
1724         }
1725         return ret;
1726 }
1727
1728 /*
1729  * push some data in the path leaf to the right, trying to free up at
1730  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1731  *
1732  * returns 1 if the push failed because the other node didn't have enough
1733  * room, 0 if everything worked out and < 0 if there were major errors.
1734  */
1735 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1736                            *root, struct btrfs_path *path, int data_size,
1737                            int empty)
1738 {
1739         struct extent_buffer *left = path->nodes[0];
1740         struct extent_buffer *right;
1741         struct extent_buffer *upper;
1742         struct btrfs_disk_key disk_key;
1743         int slot;
1744         u32 i;
1745         int free_space;
1746         int push_space = 0;
1747         int push_items = 0;
1748         struct btrfs_item *item;
1749         u32 left_nritems;
1750         u32 nr;
1751         u32 right_nritems;
1752         u32 data_end;
1753         u32 this_item_size;
1754         int ret;
1755
1756         slot = path->slots[1];
1757         if (!path->nodes[1]) {
1758                 return 1;
1759         }
1760         upper = path->nodes[1];
1761         if (slot >= btrfs_header_nritems(upper) - 1)
1762                 return 1;
1763
1764         right = read_node_slot(root, upper, slot + 1);
1765         free_space = btrfs_leaf_free_space(root, right);
1766         if (free_space < data_size) {
1767                 free_extent_buffer(right);
1768                 return 1;
1769         }
1770
1771         /* cow and double check */
1772         ret = btrfs_cow_block(trans, root, right, upper,
1773                               slot + 1, &right);
1774         if (ret) {
1775                 free_extent_buffer(right);
1776                 return 1;
1777         }
1778         free_space = btrfs_leaf_free_space(root, right);
1779         if (free_space < data_size) {
1780                 free_extent_buffer(right);
1781                 return 1;
1782         }
1783
1784         left_nritems = btrfs_header_nritems(left);
1785         if (left_nritems == 0) {
1786                 free_extent_buffer(right);
1787                 return 1;
1788         }
1789
1790         if (empty)
1791                 nr = 0;
1792         else
1793                 nr = 1;
1794
1795         i = left_nritems - 1;
1796         while (i >= nr) {
1797                 item = btrfs_item_nr(left, i);
1798
1799                 if (path->slots[0] == i)
1800                         push_space += data_size + sizeof(*item);
1801
1802                 this_item_size = btrfs_item_size(left, item);
1803                 if (this_item_size + sizeof(*item) + push_space > free_space)
1804                         break;
1805                 push_items++;
1806                 push_space += this_item_size + sizeof(*item);
1807                 if (i == 0)
1808                         break;
1809                 i--;
1810         }
1811
1812         if (push_items == 0) {
1813                 free_extent_buffer(right);
1814                 return 1;
1815         }
1816
1817         if (!empty && push_items == left_nritems)
1818                 WARN_ON(1);
1819
1820         /* push left to right */
1821         right_nritems = btrfs_header_nritems(right);
1822
1823         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1824         push_space -= leaf_data_end(root, left);
1825
1826         /* make room in the right data area */
1827         data_end = leaf_data_end(root, right);
1828         memmove_extent_buffer(right,
1829                               btrfs_leaf_data(right) + data_end - push_space,
1830                               btrfs_leaf_data(right) + data_end,
1831                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1832
1833         /* copy from the left data area */
1834         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1835                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1836                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1837                      push_space);
1838
1839         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1840                               btrfs_item_nr_offset(0),
1841                               right_nritems * sizeof(struct btrfs_item));
1842
1843         /* copy the items from left to right */
1844         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1845                    btrfs_item_nr_offset(left_nritems - push_items),
1846                    push_items * sizeof(struct btrfs_item));
1847
1848         /* update the item pointers */
1849         right_nritems += push_items;
1850         btrfs_set_header_nritems(right, right_nritems);
1851         push_space = BTRFS_LEAF_DATA_SIZE(root);
1852         for (i = 0; i < right_nritems; i++) {
1853                 item = btrfs_item_nr(right, i);
1854                 push_space -= btrfs_item_size(right, item);
1855                 btrfs_set_item_offset(right, item, push_space);
1856         }
1857
1858         left_nritems -= push_items;
1859         btrfs_set_header_nritems(left, left_nritems);
1860
1861         if (left_nritems)
1862                 btrfs_mark_buffer_dirty(left);
1863         btrfs_mark_buffer_dirty(right);
1864
1865         btrfs_item_key(right, &disk_key, 0);
1866         btrfs_set_node_key(upper, &disk_key, slot + 1);
1867         btrfs_mark_buffer_dirty(upper);
1868
1869         /* then fixup the leaf pointer in the path */
1870         if (path->slots[0] >= left_nritems) {
1871                 path->slots[0] -= left_nritems;
1872                 free_extent_buffer(path->nodes[0]);
1873                 path->nodes[0] = right;
1874                 path->slots[1] += 1;
1875         } else {
1876                 free_extent_buffer(right);
1877         }
1878         return 0;
1879 }
1880 /*
1881  * push some data in the path leaf to the left, trying to free up at
1882  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1883  */
1884 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1885                           *root, struct btrfs_path *path, int data_size,
1886                           int empty)
1887 {
1888         struct btrfs_disk_key disk_key;
1889         struct extent_buffer *right = path->nodes[0];
1890         struct extent_buffer *left;
1891         int slot;
1892         int i;
1893         int free_space;
1894         int push_space = 0;
1895         int push_items = 0;
1896         struct btrfs_item *item;
1897         u32 old_left_nritems;
1898         u32 right_nritems;
1899         u32 nr;
1900         int ret = 0;
1901         int wret;
1902         u32 this_item_size;
1903         u32 old_left_item_size;
1904
1905         slot = path->slots[1];
1906         if (slot == 0)
1907                 return 1;
1908         if (!path->nodes[1])
1909                 return 1;
1910
1911         right_nritems = btrfs_header_nritems(right);
1912         if (right_nritems == 0) {
1913                 return 1;
1914         }
1915
1916         left = read_node_slot(root, path->nodes[1], slot - 1);
1917         free_space = btrfs_leaf_free_space(root, left);
1918         if (free_space < data_size) {
1919                 free_extent_buffer(left);
1920                 return 1;
1921         }
1922
1923         /* cow and double check */
1924         ret = btrfs_cow_block(trans, root, left,
1925                               path->nodes[1], slot - 1, &left);
1926         if (ret) {
1927                 /* we hit -ENOSPC, but it isn't fatal here */
1928                 free_extent_buffer(left);
1929                 return 1;
1930         }
1931
1932         free_space = btrfs_leaf_free_space(root, left);
1933         if (free_space < data_size) {
1934                 free_extent_buffer(left);
1935                 return 1;
1936         }
1937
1938         if (empty)
1939                 nr = right_nritems;
1940         else
1941                 nr = right_nritems - 1;
1942
1943         for (i = 0; i < nr; i++) {
1944                 item = btrfs_item_nr(right, i);
1945
1946                 if (path->slots[0] == i)
1947                         push_space += data_size + sizeof(*item);
1948
1949                 this_item_size = btrfs_item_size(right, item);
1950                 if (this_item_size + sizeof(*item) + push_space > free_space)
1951                         break;
1952
1953                 push_items++;
1954                 push_space += this_item_size + sizeof(*item);
1955         }
1956
1957         if (push_items == 0) {
1958                 free_extent_buffer(left);
1959                 return 1;
1960         }
1961         if (!empty && push_items == btrfs_header_nritems(right))
1962                 WARN_ON(1);
1963
1964         /* push data from right to left */
1965         copy_extent_buffer(left, right,
1966                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1967                            btrfs_item_nr_offset(0),
1968                            push_items * sizeof(struct btrfs_item));
1969
1970         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1971                      btrfs_item_offset_nr(right, push_items -1);
1972
1973         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1974                      leaf_data_end(root, left) - push_space,
1975                      btrfs_leaf_data(right) +
1976                      btrfs_item_offset_nr(right, push_items - 1),
1977                      push_space);
1978         old_left_nritems = btrfs_header_nritems(left);
1979         BUG_ON(old_left_nritems == 0);
1980
1981         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1982         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1983                 u32 ioff;
1984
1985                 item = btrfs_item_nr(left, i);
1986                 ioff = btrfs_item_offset(left, item);
1987                 btrfs_set_item_offset(left, item,
1988                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1989         }
1990         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1991
1992         /* fixup right node */
1993         if (push_items > right_nritems) {
1994                 printk("push items %d nr %u\n", push_items, right_nritems);
1995                 WARN_ON(1);
1996         }
1997
1998         if (push_items < right_nritems) {
1999                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
2000                                                   leaf_data_end(root, right);
2001                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
2002                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
2003                                       btrfs_leaf_data(right) +
2004                                       leaf_data_end(root, right), push_space);
2005
2006                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
2007                               btrfs_item_nr_offset(push_items),
2008                              (btrfs_header_nritems(right) - push_items) *
2009                              sizeof(struct btrfs_item));
2010         }
2011         right_nritems -= push_items;
2012         btrfs_set_header_nritems(right, right_nritems);
2013         push_space = BTRFS_LEAF_DATA_SIZE(root);
2014         for (i = 0; i < right_nritems; i++) {
2015                 item = btrfs_item_nr(right, i);
2016                 push_space = push_space - btrfs_item_size(right, item);
2017                 btrfs_set_item_offset(right, item, push_space);
2018         }
2019
2020         btrfs_mark_buffer_dirty(left);
2021         if (right_nritems)
2022                 btrfs_mark_buffer_dirty(right);
2023
2024         btrfs_item_key(right, &disk_key, 0);
2025         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
2026         if (wret)
2027                 ret = wret;
2028
2029         /* then fixup the leaf pointer in the path */
2030         if (path->slots[0] < push_items) {
2031                 path->slots[0] += old_left_nritems;
2032                 free_extent_buffer(path->nodes[0]);
2033                 path->nodes[0] = left;
2034                 path->slots[1] -= 1;
2035         } else {
2036                 free_extent_buffer(left);
2037                 path->slots[0] -= push_items;
2038         }
2039         BUG_ON(path->slots[0] < 0);
2040         return ret;
2041 }
2042
2043 /*
2044  * split the path's leaf in two, making sure there is at least data_size
2045  * available for the resulting leaf level of the path.
2046  *
2047  * returns 0 if all went well and < 0 on failure.
2048  */
2049 static noinline int copy_for_split(struct btrfs_trans_handle *trans,
2050                                struct btrfs_root *root,
2051                                struct btrfs_path *path,
2052                                struct extent_buffer *l,
2053                                struct extent_buffer *right,
2054                                int slot, int mid, int nritems)
2055 {
2056         int data_copy_size;
2057         int rt_data_off;
2058         int i;
2059         int ret = 0;
2060         int wret;
2061         struct btrfs_disk_key disk_key;
2062
2063         nritems = nritems - mid;
2064         btrfs_set_header_nritems(right, nritems);
2065         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2066
2067         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2068                            btrfs_item_nr_offset(mid),
2069                            nritems * sizeof(struct btrfs_item));
2070
2071         copy_extent_buffer(right, l,
2072                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2073                      data_copy_size, btrfs_leaf_data(l) +
2074                      leaf_data_end(root, l), data_copy_size);
2075
2076         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2077                       btrfs_item_end_nr(l, mid);
2078
2079         for (i = 0; i < nritems; i++) {
2080                 struct btrfs_item *item = btrfs_item_nr(right, i);
2081                 u32 ioff = btrfs_item_offset(right, item);
2082                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2083         }
2084
2085         btrfs_set_header_nritems(l, mid);
2086         ret = 0;
2087         btrfs_item_key(right, &disk_key, 0);
2088         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2089                           path->slots[1] + 1, 1);
2090         if (wret)
2091                 ret = wret;
2092
2093         btrfs_mark_buffer_dirty(right);
2094         btrfs_mark_buffer_dirty(l);
2095         BUG_ON(path->slots[0] != slot);
2096
2097         if (mid <= slot) {
2098                 free_extent_buffer(path->nodes[0]);
2099                 path->nodes[0] = right;
2100                 path->slots[0] -= mid;
2101                 path->slots[1] += 1;
2102         } else {
2103                 free_extent_buffer(right);
2104         }
2105
2106         BUG_ON(path->slots[0] < 0);
2107
2108         return ret;
2109 }
2110
2111 /*
2112  * split the path's leaf in two, making sure there is at least data_size
2113  * available for the resulting leaf level of the path.
2114  *
2115  * returns 0 if all went well and < 0 on failure.
2116  */
2117 static noinline int split_leaf(struct btrfs_trans_handle *trans,
2118                                struct btrfs_root *root,
2119                                struct btrfs_key *ins_key,
2120                                struct btrfs_path *path, int data_size,
2121                                int extend)
2122 {
2123         struct btrfs_disk_key disk_key;
2124         struct extent_buffer *l;
2125         u32 nritems;
2126         int mid;
2127         int slot;
2128         struct extent_buffer *right;
2129         int ret = 0;
2130         int wret;
2131         int split;
2132         int num_doubles = 0;
2133
2134         /* first try to make some room by pushing left and right */
2135         if (data_size && ins_key->type != BTRFS_DIR_ITEM_KEY) {
2136                 wret = push_leaf_right(trans, root, path, data_size, 0);
2137                 if (wret < 0)
2138                         return wret;
2139                 if (wret) {
2140                         wret = push_leaf_left(trans, root, path, data_size, 0);
2141                         if (wret < 0)
2142                                 return wret;
2143                 }
2144                 l = path->nodes[0];
2145
2146                 /* did the pushes work? */
2147                 if (btrfs_leaf_free_space(root, l) >= data_size)
2148                         return 0;
2149         }
2150
2151         if (!path->nodes[1]) {
2152                 ret = insert_new_root(trans, root, path, 1);
2153                 if (ret)
2154                         return ret;
2155         }
2156 again:
2157         split = 1;
2158         l = path->nodes[0];
2159         slot = path->slots[0];
2160         nritems = btrfs_header_nritems(l);
2161         mid = (nritems + 1) / 2;
2162
2163         if (mid <= slot) {
2164                 if (nritems == 1 ||
2165                     leaf_space_used(l, mid, nritems - mid) + data_size >
2166                         BTRFS_LEAF_DATA_SIZE(root)) {
2167                         if (slot >= nritems) {
2168                                 split = 0;
2169                         } else {
2170                                 mid = slot;
2171                                 if (mid != nritems &&
2172                                     leaf_space_used(l, mid, nritems - mid) +
2173                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2174                                         split = 2;
2175                                 }
2176                         }
2177                 }
2178         } else {
2179                 if (leaf_space_used(l, 0, mid) + data_size >
2180                         BTRFS_LEAF_DATA_SIZE(root)) {
2181                         if (!extend && data_size && slot == 0) {
2182                                 split = 0;
2183                         } else if ((extend || !data_size) && slot == 0) {
2184                                 mid = 1;
2185                         } else {
2186                                 mid = slot;
2187                                 if (mid != nritems &&
2188                                     leaf_space_used(l, mid, nritems - mid) +
2189                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2190                                         split = 2 ;
2191                                 }
2192                         }
2193                 }
2194         }
2195
2196         if (split == 0)
2197                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2198         else
2199                 btrfs_item_key(l, &disk_key, mid);
2200
2201         right = btrfs_alloc_free_block(trans, root, root->leafsize,
2202                                         root->root_key.objectid,
2203                                         &disk_key, 0, l->start, 0);
2204         if (IS_ERR(right)) {
2205                 BUG_ON(1);
2206                 return PTR_ERR(right);
2207         }
2208
2209         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2210         btrfs_set_header_bytenr(right, right->start);
2211         btrfs_set_header_generation(right, trans->transid);
2212         btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
2213         btrfs_set_header_owner(right, root->root_key.objectid);
2214         btrfs_set_header_level(right, 0);
2215         write_extent_buffer(right, root->fs_info->fsid,
2216                             (unsigned long)btrfs_header_fsid(right),
2217                             BTRFS_FSID_SIZE);
2218
2219         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2220                             (unsigned long)btrfs_header_chunk_tree_uuid(right),
2221                             BTRFS_UUID_SIZE);
2222
2223         if (split == 0) {
2224                 if (mid <= slot) {
2225                         btrfs_set_header_nritems(right, 0);
2226                         wret = insert_ptr(trans, root, path,
2227                                           &disk_key, right->start,
2228                                           path->slots[1] + 1, 1);
2229                         if (wret)
2230                                 ret = wret;
2231
2232                         free_extent_buffer(path->nodes[0]);
2233                         path->nodes[0] = right;
2234                         path->slots[0] = 0;
2235                         path->slots[1] += 1;
2236                 } else {
2237                         btrfs_set_header_nritems(right, 0);
2238                         wret = insert_ptr(trans, root, path,
2239                                           &disk_key,
2240                                           right->start,
2241                                           path->slots[1], 1);
2242                         if (wret)
2243                                 ret = wret;
2244                         free_extent_buffer(path->nodes[0]);
2245                         path->nodes[0] = right;
2246                         path->slots[0] = 0;
2247                         if (path->slots[1] == 0) {
2248                                 wret = fixup_low_keys(trans, root,
2249                                                 path, &disk_key, 1);
2250                                 if (wret)
2251                                         ret = wret;
2252                         }
2253                 }
2254                 btrfs_mark_buffer_dirty(right);
2255                 return ret;
2256         }
2257
2258         ret = copy_for_split(trans, root, path, l, right, slot, mid, nritems);
2259         BUG_ON(ret);
2260
2261         if (split == 2) {
2262                 BUG_ON(num_doubles != 0);
2263                 num_doubles++;
2264                 goto again;
2265         }
2266
2267         return ret;
2268 }
2269
2270 /*
2271  * This function splits a single item into two items,
2272  * giving 'new_key' to the new item and splitting the
2273  * old one at split_offset (from the start of the item).
2274  *
2275  * The path may be released by this operation.  After
2276  * the split, the path is pointing to the old item.  The
2277  * new item is going to be in the same node as the old one.
2278  *
2279  * Note, the item being split must be smaller enough to live alone on
2280  * a tree block with room for one extra struct btrfs_item
2281  *
2282  * This allows us to split the item in place, keeping a lock on the
2283  * leaf the entire time.
2284  */
2285 int btrfs_split_item(struct btrfs_trans_handle *trans,
2286                      struct btrfs_root *root,
2287                      struct btrfs_path *path,
2288                      struct btrfs_key *new_key,
2289                      unsigned long split_offset)
2290 {
2291         u32 item_size;
2292         struct extent_buffer *leaf;
2293         struct btrfs_key orig_key;
2294         struct btrfs_item *item;
2295         struct btrfs_item *new_item;
2296         int ret = 0;
2297         int slot;
2298         u32 nritems;
2299         u32 orig_offset;
2300         struct btrfs_disk_key disk_key;
2301         char *buf;
2302
2303         leaf = path->nodes[0];
2304         btrfs_item_key_to_cpu(leaf, &orig_key, path->slots[0]);
2305         if (btrfs_leaf_free_space(root, leaf) >= sizeof(struct btrfs_item))
2306                 goto split;
2307
2308         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2309         btrfs_release_path(root, path);
2310
2311         path->search_for_split = 1;
2312
2313         ret = btrfs_search_slot(trans, root, &orig_key, path, 0, 1);
2314         path->search_for_split = 0;
2315
2316         /* if our item isn't there or got smaller, return now */
2317         if (ret != 0 || item_size != btrfs_item_size_nr(path->nodes[0],
2318                                                         path->slots[0])) {
2319                 return -EAGAIN;
2320         }
2321
2322         ret = split_leaf(trans, root, &orig_key, path, 0, 0);
2323         BUG_ON(ret);
2324
2325         BUG_ON(btrfs_leaf_free_space(root, leaf) < sizeof(struct btrfs_item));
2326         leaf = path->nodes[0];
2327
2328 split:
2329         item = btrfs_item_nr(leaf, path->slots[0]);
2330         orig_offset = btrfs_item_offset(leaf, item);
2331         item_size = btrfs_item_size(leaf, item);
2332
2333
2334         buf = kmalloc(item_size, GFP_NOFS);
2335         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
2336                             path->slots[0]), item_size);
2337         slot = path->slots[0] + 1;
2338         leaf = path->nodes[0];
2339
2340         nritems = btrfs_header_nritems(leaf);
2341
2342         if (slot != nritems) {
2343                 /* shift the items */
2344                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
2345                               btrfs_item_nr_offset(slot),
2346                               (nritems - slot) * sizeof(struct btrfs_item));
2347
2348         }
2349
2350         btrfs_cpu_key_to_disk(&disk_key, new_key);
2351         btrfs_set_item_key(leaf, &disk_key, slot);
2352
2353         new_item = btrfs_item_nr(leaf, slot);
2354
2355         btrfs_set_item_offset(leaf, new_item, orig_offset);
2356         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
2357
2358         btrfs_set_item_offset(leaf, item,
2359                               orig_offset + item_size - split_offset);
2360         btrfs_set_item_size(leaf, item, split_offset);
2361
2362         btrfs_set_header_nritems(leaf, nritems + 1);
2363
2364         /* write the data for the start of the original item */
2365         write_extent_buffer(leaf, buf,
2366                             btrfs_item_ptr_offset(leaf, path->slots[0]),
2367                             split_offset);
2368
2369         /* write the data for the new item */
2370         write_extent_buffer(leaf, buf + split_offset,
2371                             btrfs_item_ptr_offset(leaf, slot),
2372                             item_size - split_offset);
2373         btrfs_mark_buffer_dirty(leaf);
2374
2375         ret = 0;
2376         if (btrfs_leaf_free_space(root, leaf) < 0) {
2377                 btrfs_print_leaf(root, leaf);
2378                 BUG();
2379         }
2380         kfree(buf);
2381         return ret;
2382 }
2383
2384 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2385                         struct btrfs_root *root,
2386                         struct btrfs_path *path,
2387                         u32 new_size, int from_end)
2388 {
2389         int ret = 0;
2390         int slot;
2391         struct extent_buffer *leaf;
2392         struct btrfs_item *item;
2393         u32 nritems;
2394         unsigned int data_end;
2395         unsigned int old_data_start;
2396         unsigned int old_size;
2397         unsigned int size_diff;
2398         int i;
2399
2400         leaf = path->nodes[0];
2401         slot = path->slots[0];
2402
2403         old_size = btrfs_item_size_nr(leaf, slot);
2404         if (old_size == new_size)
2405                 return 0;
2406
2407         nritems = btrfs_header_nritems(leaf);
2408         data_end = leaf_data_end(root, leaf);
2409
2410         old_data_start = btrfs_item_offset_nr(leaf, slot);
2411
2412         size_diff = old_size - new_size;
2413
2414         BUG_ON(slot < 0);
2415         BUG_ON(slot >= nritems);
2416
2417         /*
2418          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2419          */
2420         /* first correct the data pointers */
2421         for (i = slot; i < nritems; i++) {
2422                 u32 ioff;
2423                 item = btrfs_item_nr(leaf, i);
2424                 ioff = btrfs_item_offset(leaf, item);
2425                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2426         }
2427
2428         /* shift the data */
2429         if (from_end) {
2430                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2431                               data_end + size_diff, btrfs_leaf_data(leaf) +
2432                               data_end, old_data_start + new_size - data_end);
2433         } else {
2434                 struct btrfs_disk_key disk_key;
2435                 u64 offset;
2436
2437                 btrfs_item_key(leaf, &disk_key, slot);
2438
2439                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2440                         unsigned long ptr;
2441                         struct btrfs_file_extent_item *fi;
2442
2443                         fi = btrfs_item_ptr(leaf, slot,
2444                                             struct btrfs_file_extent_item);
2445                         fi = (struct btrfs_file_extent_item *)(
2446                              (unsigned long)fi - size_diff);
2447
2448                         if (btrfs_file_extent_type(leaf, fi) ==
2449                             BTRFS_FILE_EXTENT_INLINE) {
2450                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2451                                 memmove_extent_buffer(leaf, ptr,
2452                                         (unsigned long)fi,
2453                                         offsetof(struct btrfs_file_extent_item,
2454                                                  disk_bytenr));
2455                         }
2456                 }
2457
2458                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2459                               data_end + size_diff, btrfs_leaf_data(leaf) +
2460                               data_end, old_data_start - data_end);
2461
2462                 offset = btrfs_disk_key_offset(&disk_key);
2463                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2464                 btrfs_set_item_key(leaf, &disk_key, slot);
2465                 if (slot == 0)
2466                         fixup_low_keys(trans, root, path, &disk_key, 1);
2467         }
2468
2469         item = btrfs_item_nr(leaf, slot);
2470         btrfs_set_item_size(leaf, item, new_size);
2471         btrfs_mark_buffer_dirty(leaf);
2472
2473         ret = 0;
2474         if (btrfs_leaf_free_space(root, leaf) < 0) {
2475                 btrfs_print_leaf(root, leaf);
2476                 BUG();
2477         }
2478         return ret;
2479 }
2480
2481 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2482                       struct btrfs_root *root, struct btrfs_path *path,
2483                       u32 data_size)
2484 {
2485         int ret = 0;
2486         int slot;
2487         struct extent_buffer *leaf;
2488         struct btrfs_item *item;
2489         u32 nritems;
2490         unsigned int data_end;
2491         unsigned int old_data;
2492         unsigned int old_size;
2493         int i;
2494
2495         leaf = path->nodes[0];
2496
2497         nritems = btrfs_header_nritems(leaf);
2498         data_end = leaf_data_end(root, leaf);
2499
2500         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2501                 btrfs_print_leaf(root, leaf);
2502                 BUG();
2503         }
2504         slot = path->slots[0];
2505         old_data = btrfs_item_end_nr(leaf, slot);
2506
2507         BUG_ON(slot < 0);
2508         if (slot >= nritems) {
2509                 btrfs_print_leaf(root, leaf);
2510                 printk("slot %d too large, nritems %d\n", slot, nritems);
2511                 BUG_ON(1);
2512         }
2513
2514         /*
2515          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2516          */
2517         /* first correct the data pointers */
2518         for (i = slot; i < nritems; i++) {
2519                 u32 ioff;
2520                 item = btrfs_item_nr(leaf, i);
2521                 ioff = btrfs_item_offset(leaf, item);
2522                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2523         }
2524
2525         /* shift the data */
2526         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2527                       data_end - data_size, btrfs_leaf_data(leaf) +
2528                       data_end, old_data - data_end);
2529
2530         data_end = old_data;
2531         old_size = btrfs_item_size_nr(leaf, slot);
2532         item = btrfs_item_nr(leaf, slot);
2533         btrfs_set_item_size(leaf, item, old_size + data_size);
2534         btrfs_mark_buffer_dirty(leaf);
2535
2536         ret = 0;
2537         if (btrfs_leaf_free_space(root, leaf) < 0) {
2538                 btrfs_print_leaf(root, leaf);
2539                 BUG();
2540         }
2541         return ret;
2542 }
2543
2544 /*
2545  * Given a key and some data, insert an item into the tree.
2546  * This does all the path init required, making room in the tree if needed.
2547  */
2548 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
2549                             struct btrfs_root *root,
2550                             struct btrfs_path *path,
2551                             struct btrfs_key *cpu_key, u32 *data_size,
2552                             int nr)
2553 {
2554         struct extent_buffer *leaf;
2555         struct btrfs_item *item;
2556         int ret = 0;
2557         int slot;
2558         int i;
2559         u32 nritems;
2560         u32 total_size = 0;
2561         u32 total_data = 0;
2562         unsigned int data_end;
2563         struct btrfs_disk_key disk_key;
2564
2565         for (i = 0; i < nr; i++) {
2566                 total_data += data_size[i];
2567         }
2568
2569         /* create a root if there isn't one */
2570         if (!root->node)
2571                 BUG();
2572
2573         total_size = total_data + nr * sizeof(struct btrfs_item);
2574         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
2575         if (ret == 0) {
2576                 return -EEXIST;
2577         }
2578         if (ret < 0)
2579                 goto out;
2580
2581         leaf = path->nodes[0];
2582
2583         nritems = btrfs_header_nritems(leaf);
2584         data_end = leaf_data_end(root, leaf);
2585
2586         if (btrfs_leaf_free_space(root, leaf) < total_size) {
2587                 btrfs_print_leaf(root, leaf);
2588                 printk("not enough freespace need %u have %d\n",
2589                        total_size, btrfs_leaf_free_space(root, leaf));
2590                 BUG();
2591         }
2592
2593         slot = path->slots[0];
2594         BUG_ON(slot < 0);
2595
2596         if (slot != nritems) {
2597                 int i;
2598                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2599
2600                 if (old_data < data_end) {
2601                         btrfs_print_leaf(root, leaf);
2602                         printk("slot %d old_data %d data_end %d\n",
2603                                slot, old_data, data_end);
2604                         BUG_ON(1);
2605                 }
2606                 /*
2607                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2608                  */
2609                 /* first correct the data pointers */
2610                 for (i = slot; i < nritems; i++) {
2611                         u32 ioff;
2612
2613                         item = btrfs_item_nr(leaf, i);
2614                         ioff = btrfs_item_offset(leaf, item);
2615                         btrfs_set_item_offset(leaf, item, ioff - total_data);
2616                 }
2617
2618                 /* shift the items */
2619                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
2620                               btrfs_item_nr_offset(slot),
2621                               (nritems - slot) * sizeof(struct btrfs_item));
2622
2623                 /* shift the data */
2624                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2625                               data_end - total_data, btrfs_leaf_data(leaf) +
2626                               data_end, old_data - data_end);
2627                 data_end = old_data;
2628         }
2629
2630         /* setup the item for the new data */
2631         for (i = 0; i < nr; i++) {
2632                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
2633                 btrfs_set_item_key(leaf, &disk_key, slot + i);
2634                 item = btrfs_item_nr(leaf, slot + i);
2635                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
2636                 data_end -= data_size[i];
2637                 btrfs_set_item_size(leaf, item, data_size[i]);
2638         }
2639         btrfs_set_header_nritems(leaf, nritems + nr);
2640         btrfs_mark_buffer_dirty(leaf);
2641
2642         ret = 0;
2643         if (slot == 0) {
2644                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2645                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2646         }
2647
2648         if (btrfs_leaf_free_space(root, leaf) < 0) {
2649                 btrfs_print_leaf(root, leaf);
2650                 BUG();
2651         }
2652
2653 out:
2654         return ret;
2655 }
2656
2657 /*
2658  * Given a key and some data, insert an item into the tree.
2659  * This does all the path init required, making room in the tree if needed.
2660  */
2661 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2662                       *root, struct btrfs_key *cpu_key, void *data, u32
2663                       data_size)
2664 {
2665         int ret = 0;
2666         struct btrfs_path *path;
2667         struct extent_buffer *leaf;
2668         unsigned long ptr;
2669
2670         path = btrfs_alloc_path();
2671         BUG_ON(!path);
2672         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2673         if (!ret) {
2674                 leaf = path->nodes[0];
2675                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2676                 write_extent_buffer(leaf, data, ptr, data_size);
2677                 btrfs_mark_buffer_dirty(leaf);
2678         }
2679         btrfs_free_path(path);
2680         return ret;
2681 }
2682
2683 /*
2684  * delete the pointer from a given node.
2685  *
2686  * If the delete empties a node, the node is removed from the tree,
2687  * continuing all the way the root if required.  The root is converted into
2688  * a leaf if all the nodes are emptied.
2689  */
2690 int btrfs_del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2691                    struct btrfs_path *path, int level, int slot)
2692 {
2693         struct extent_buffer *parent = path->nodes[level];
2694         u32 nritems;
2695         int ret = 0;
2696         int wret;
2697
2698         nritems = btrfs_header_nritems(parent);
2699         if (slot != nritems -1) {
2700                 memmove_extent_buffer(parent,
2701                               btrfs_node_key_ptr_offset(slot),
2702                               btrfs_node_key_ptr_offset(slot + 1),
2703                               sizeof(struct btrfs_key_ptr) *
2704                               (nritems - slot - 1));
2705         }
2706         nritems--;
2707         btrfs_set_header_nritems(parent, nritems);
2708         if (nritems == 0 && parent == root->node) {
2709                 BUG_ON(btrfs_header_level(root->node) != 1);
2710                 /* just turn the root into a leaf and break */
2711                 btrfs_set_header_level(root->node, 0);
2712         } else if (slot == 0) {
2713                 struct btrfs_disk_key disk_key;
2714
2715                 btrfs_node_key(parent, &disk_key, 0);
2716                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2717                 if (wret)
2718                         ret = wret;
2719         }
2720         btrfs_mark_buffer_dirty(parent);
2721         return ret;
2722 }
2723
2724 /*
2725  * a helper function to delete the leaf pointed to by path->slots[1] and
2726  * path->nodes[1].
2727  *
2728  * This deletes the pointer in path->nodes[1] and frees the leaf
2729  * block extent.  zero is returned if it all worked out, < 0 otherwise.
2730  *
2731  * The path must have already been setup for deleting the leaf, including
2732  * all the proper balancing.  path->nodes[1] must be locked.
2733  */
2734 static noinline int btrfs_del_leaf(struct btrfs_trans_handle *trans,
2735                                    struct btrfs_root *root,
2736                                    struct btrfs_path *path,
2737                                    struct extent_buffer *leaf)
2738 {
2739         int ret;
2740
2741         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
2742         ret = btrfs_del_ptr(trans, root, path, 1, path->slots[1]);
2743         if (ret)
2744                 return ret;
2745
2746         ret = btrfs_free_extent(trans, root, leaf->start, leaf->len,
2747                                 0, root->root_key.objectid, 0, 0);
2748         return ret;
2749 }
2750
2751 /*
2752  * delete the item at the leaf level in path.  If that empties
2753  * the leaf, remove it from the tree
2754  */
2755 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2756                     struct btrfs_path *path, int slot, int nr)
2757 {
2758         struct extent_buffer *leaf;
2759         struct btrfs_item *item;
2760         int last_off;
2761         int dsize = 0;
2762         int ret = 0;
2763         int wret;
2764         int i;
2765         u32 nritems;
2766
2767         leaf = path->nodes[0];
2768         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
2769
2770         for (i = 0; i < nr; i++)
2771                 dsize += btrfs_item_size_nr(leaf, slot + i);
2772
2773         nritems = btrfs_header_nritems(leaf);
2774
2775         if (slot + nr != nritems) {
2776                 int i;
2777                 int data_end = leaf_data_end(root, leaf);
2778
2779                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2780                               data_end + dsize,
2781                               btrfs_leaf_data(leaf) + data_end,
2782                               last_off - data_end);
2783
2784                 for (i = slot + nr; i < nritems; i++) {
2785                         u32 ioff;
2786
2787                         item = btrfs_item_nr(leaf, i);
2788                         ioff = btrfs_item_offset(leaf, item);
2789                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2790                 }
2791
2792                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2793                               btrfs_item_nr_offset(slot + nr),
2794                               sizeof(struct btrfs_item) *
2795                               (nritems - slot - nr));
2796         }
2797         btrfs_set_header_nritems(leaf, nritems - nr);
2798         nritems -= nr;
2799
2800         /* delete the leaf if we've emptied it */
2801         if (nritems == 0) {
2802                 if (leaf == root->node) {
2803                         btrfs_set_header_level(leaf, 0);
2804                 } else {
2805                         clean_tree_block(trans, root, leaf);
2806                         wait_on_tree_block_writeback(root, leaf);
2807
2808                         wret = btrfs_del_leaf(trans, root, path, leaf);
2809                         BUG_ON(ret);
2810                         if (wret)
2811                                 ret = wret;
2812                 }
2813         } else {
2814                 int used = leaf_space_used(leaf, 0, nritems);
2815                 if (slot == 0) {
2816                         struct btrfs_disk_key disk_key;
2817
2818                         btrfs_item_key(leaf, &disk_key, 0);
2819                         wret = fixup_low_keys(trans, root, path,
2820                                               &disk_key, 1);
2821                         if (wret)
2822                                 ret = wret;
2823                 }
2824
2825                 /* delete the leaf if it is mostly empty */
2826                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
2827                         /* push_leaf_left fixes the path.
2828                          * make sure the path still points to our leaf
2829                          * for possible call to del_ptr below
2830                          */
2831                         slot = path->slots[1];
2832                         extent_buffer_get(leaf);
2833
2834                         wret = push_leaf_left(trans, root, path, 1, 1);
2835                         if (wret < 0 && wret != -ENOSPC)
2836                                 ret = wret;
2837
2838                         if (path->nodes[0] == leaf &&
2839                             btrfs_header_nritems(leaf)) {
2840                                 wret = push_leaf_right(trans, root, path, 1, 1);
2841                                 if (wret < 0 && wret != -ENOSPC)
2842                                         ret = wret;
2843                         }
2844
2845                         if (btrfs_header_nritems(leaf) == 0) {
2846                                 clean_tree_block(trans, root, leaf);
2847                                 wait_on_tree_block_writeback(root, leaf);
2848
2849                                 path->slots[1] = slot;
2850                                 ret = btrfs_del_leaf(trans, root, path, leaf);
2851                                 BUG_ON(ret);
2852                                 free_extent_buffer(leaf);
2853
2854                         } else {
2855                                 btrfs_mark_buffer_dirty(leaf);
2856                                 free_extent_buffer(leaf);
2857                         }
2858                 } else {
2859                         btrfs_mark_buffer_dirty(leaf);
2860                 }
2861         }
2862         return ret;
2863 }
2864
2865 /*
2866  * walk up the tree as far as required to find the previous leaf.
2867  * returns 0 if it found something or 1 if there are no lesser leaves.
2868  * returns < 0 on io errors.
2869  */
2870 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2871 {
2872         int slot;
2873         int level = 1;
2874         struct extent_buffer *c;
2875         struct extent_buffer *next = NULL;
2876
2877         while(level < BTRFS_MAX_LEVEL) {
2878                 if (!path->nodes[level])
2879                         return 1;
2880
2881                 slot = path->slots[level];
2882                 c = path->nodes[level];
2883                 if (slot == 0) {
2884                         level++;
2885                         if (level == BTRFS_MAX_LEVEL)
2886                                 return 1;
2887                         continue;
2888                 }
2889                 slot--;
2890
2891                 next = read_node_slot(root, c, slot);
2892                 break;
2893         }
2894         path->slots[level] = slot;
2895         while(1) {
2896                 level--;
2897                 c = path->nodes[level];
2898                 free_extent_buffer(c);
2899                 slot = btrfs_header_nritems(next);
2900                 if (slot != 0)
2901                         slot--;
2902                 path->nodes[level] = next;
2903                 path->slots[level] = slot;
2904                 if (!level)
2905                         break;
2906                 next = read_node_slot(root, next, slot);
2907         }
2908         return 0;
2909 }
2910
2911 /*
2912  * walk up the tree as far as required to find the next leaf.
2913  * returns 0 if it found something or 1 if there are no greater leaves.
2914  * returns < 0 on io errors.
2915  */
2916 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2917 {
2918         int slot;
2919         int level = 1;
2920         struct extent_buffer *c;
2921         struct extent_buffer *next = NULL;
2922
2923         while(level < BTRFS_MAX_LEVEL) {
2924                 if (!path->nodes[level])
2925                         return 1;
2926
2927                 slot = path->slots[level] + 1;
2928                 c = path->nodes[level];
2929                 if (slot >= btrfs_header_nritems(c)) {
2930                         level++;
2931                         if (level == BTRFS_MAX_LEVEL)
2932                                 return 1;
2933                         continue;
2934                 }
2935
2936                 if (path->reada)
2937                         reada_for_search(root, path, level, slot, 0);
2938
2939                 next = read_node_slot(root, c, slot);
2940                 if (!next)
2941                         return -EIO;
2942                 break;
2943         }
2944         path->slots[level] = slot;
2945         while(1) {
2946                 level--;
2947                 c = path->nodes[level];
2948                 free_extent_buffer(c);
2949                 path->nodes[level] = next;
2950                 path->slots[level] = 0;
2951                 if (!level)
2952                         break;
2953                 if (path->reada)
2954                         reada_for_search(root, path, level, 0, 0);
2955                 next = read_node_slot(root, next, 0);
2956                 if (!next)
2957                         return -EIO;
2958         }
2959         return 0;
2960 }
2961
2962 int btrfs_previous_item(struct btrfs_root *root,
2963                         struct btrfs_path *path, u64 min_objectid,
2964                         int type)
2965 {
2966         struct btrfs_key found_key;
2967         struct extent_buffer *leaf;
2968         int ret;
2969
2970         while(1) {
2971                 if (path->slots[0] == 0) {
2972                         ret = btrfs_prev_leaf(root, path);
2973                         if (ret != 0)
2974                                 return ret;
2975                 } else {
2976                         path->slots[0]--;
2977                 }
2978                 leaf = path->nodes[0];
2979                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2980                 if (found_key.type == type)
2981                         return 0;
2982         }
2983         return 1;
2984 }
2985