fs/btrfs/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
  19 #include <linux/sched.h>
  20 #include "ctree.h"
  21 #include "disk-io.h"
  22 #include "transaction.h"
  23 #include "print-tree.h"
  24
  25 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  26                       *root, struct btrfs_path *path, int level);
  27 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  28                       *root, struct btrfs_key *ins_key,
  29                       struct btrfs_path *path, int data_size, int extend);
  30 static int push_node_left(struct btrfs_trans_handle *trans,
  31                           struct btrfs_root *root, struct extent_buffer *dst,
  32                           struct extent_buffer *src);
  33 static int balance_node_right(struct btrfs_trans_handle *trans,
  34                               struct btrfs_root *root,
  35                               struct extent_buffer *dst_buf,
  36                               struct extent_buffer *src_buf);
  37 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  38                    struct btrfs_path *path, int level, int slot);
  39
  40 inline void btrfs_init_path(struct btrfs_path *p)
  41 {
  42         memset(p, 0, sizeof(*p));
  43 }
  44
  45 struct btrfs_path *btrfs_alloc_path(void)
  46 {
  47         struct btrfs_path *path;
  48         path = kmem_cache_alloc(btrfs_path_cachep, GFP_NOFS);
  49         if (path) {
  50                 btrfs_init_path(path);
  51                 path->reada = 1;
  52         }
  53         return path;
  54 }
  55
  56 void btrfs_free_path(struct btrfs_path *p)
  57 {
  58         btrfs_release_path(NULL, p);
  59         kmem_cache_free(btrfs_path_cachep, p);
  60 }
  61
  62 void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
  63 {
  64         int i;
  65         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  66                 if (!p->nodes[i])
  67                         break;
  68                 free_extent_buffer(p->nodes[i]);
  69         }
  70         memset(p, 0, sizeof(*p));
  71 }
  72
  73 static int __btrfs_cow_block(struct btrfs_trans_handle *trans,
  74                              struct btrfs_root *root,
  75                              struct extent_buffer *buf,
  76                              struct extent_buffer *parent, int parent_slot,
  77                              struct extent_buffer **cow_ret,
  78                              u64 search_start, u64 empty_size)
  79 {
  80         struct extent_buffer *cow;
  81         int ret = 0;
  82         int different_trans = 0;
  83
  84         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
  85
  86         cow = btrfs_alloc_free_block(trans, root, buf->len,
  87                                      search_start, empty_size);
  88         if (IS_ERR(cow))
  89                 return PTR_ERR(cow);
  90
  91         copy_extent_buffer(cow, buf, 0, 0, cow->len);
  92         btrfs_set_header_bytenr(cow, cow->start);
  93         btrfs_set_header_generation(cow, trans->transid);
  94         btrfs_set_header_owner(cow, root->root_key.objectid);
  95
  96         WARN_ON(btrfs_header_generation(buf) > trans->transid);
  97         if (btrfs_header_generation(buf) != trans->transid) {
  98                 different_trans = 1;
  99                 ret = btrfs_inc_ref(trans, root, buf);
 100                 if (ret)
 101                         return ret;
 102         } else {
 103                 clean_tree_block(trans, root, buf);
 104         }
 105
 106         if (buf == root->node) {
 107                 root->node = cow;
 108                 extent_buffer_get(cow);
 109                 if (buf != root->commit_root) {
 110                         btrfs_free_extent(trans, root, buf->start,
 111                                           buf->len, 1);
 112                 }
 113                 free_extent_buffer(buf);
 114         } else {
 115                 btrfs_set_node_blockptr(parent, parent_slot,
 116                                         cow->start);
 117                 btrfs_mark_buffer_dirty(parent);
 118                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 119                 btrfs_free_extent(trans, root, buf->start, buf->len, 1);
 120         }
 121         free_extent_buffer(buf);
 122         btrfs_mark_buffer_dirty(cow);
 123         *cow_ret = cow;
 124         return 0;
 125 }
 126
 127 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 128                     struct btrfs_root *root, struct extent_buffer *buf,
 129                     struct extent_buffer *parent, int parent_slot,
 130                     struct extent_buffer **cow_ret)
 131 {
 132         u64 search_start;
 133         int ret;
 134         if (trans->transaction != root->fs_info->running_transaction) {
 135                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 136                        root->fs_info->running_transaction->transid);
 137                 WARN_ON(1);
 138         }
 139         if (trans->transid != root->fs_info->generation) {
 140                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 141                        root->fs_info->generation);
 142                 WARN_ON(1);
 143         }
 144         if (btrfs_header_generation(buf) == trans->transid) {
 145                 *cow_ret = buf;
 146                 return 0;
 147         }
 148
 149         search_start = buf->start & ~((u64)BTRFS_BLOCK_GROUP_SIZE - 1);
 150         ret = __btrfs_cow_block(trans, root, buf, parent,
 151                                  parent_slot, cow_ret, search_start, 0);
 152         return ret;
 153 }
 154
 155 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 156 {
 157         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 158                 return 1;
 159         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 160                 return 1;
 161         return 0;
 162 }
 163
 164 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 165                        struct btrfs_root *root, struct extent_buffer *parent,
 166                        int start_slot, int cache_only, u64 *last_ret,
 167                        struct btrfs_key *progress)
 168 {
 169         struct extent_buffer *cur;
 170         struct extent_buffer *tmp;
 171         u64 blocknr;
 172         u64 search_start = *last_ret;
 173         u64 last_block = 0;
 174         u64 other;
 175         u32 parent_nritems;
 176         int end_slot;
 177         int i;
 178         int err = 0;
 179         int parent_level;
 180         int uptodate;
 181         u32 blocksize;
 182
 183         parent_level = btrfs_header_level(parent);
 184         if (cache_only && parent_level != 1)
 185                 return 0;
 186
 187         if (trans->transaction != root->fs_info->running_transaction) {
 188                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 189                        root->fs_info->running_transaction->transid);
 190                 WARN_ON(1);
 191         }
 192         if (trans->transid != root->fs_info->generation) {
 193                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 194                        root->fs_info->generation);
 195                 WARN_ON(1);
 196         }
 197
 198         parent_nritems = btrfs_header_nritems(parent);
 199         blocksize = btrfs_level_size(root, parent_level - 1);
 200         end_slot = parent_nritems;
 201
 202         if (parent_nritems == 1)
 203                 return 0;
 204
 205         for (i = start_slot; i < end_slot; i++) {
 206                 int close = 1;
 207
 208                 if (!parent->map_token) {
 209                         map_extent_buffer(parent,
 210                                         btrfs_node_key_ptr_offset(i),
 211                                         sizeof(struct btrfs_key_ptr),
 212                                         &parent->map_token, &parent->kaddr,
 213                                         &parent->map_start, &parent->map_len,
 214                                         KM_USER1);
 215                 }
 216                 blocknr = btrfs_node_blockptr(parent, i);
 217                 if (last_block == 0)
 218                         last_block = blocknr;
 219
 220                 if (i > 0) {
 221                         other = btrfs_node_blockptr(parent, i - 1);
 222                         close = close_blocks(blocknr, other, blocksize);
 223                 }
 224                 if (close && i < end_slot - 2) {
 225                         other = btrfs_node_blockptr(parent, i + 1);
 226                         close = close_blocks(blocknr, other, blocksize);
 227                 }
 228                 if (close) {
 229                         last_block = blocknr;
 230                         continue;
 231                 }
 232                 if (parent->map_token) {
 233                         unmap_extent_buffer(parent, parent->map_token,
 234                                             KM_USER1);
 235                         parent->map_token = NULL;
 236                 }
 237
 238                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
 239                 if (cur)
 240                         uptodate = btrfs_buffer_uptodate(cur);
 241                 else
 242                         uptodate = 0;
 243                 if (!cur || !uptodate) {
 244                         if (cache_only) {
 245                                 free_extent_buffer(cur);
 246                                 continue;
 247                         }
 248                         if (!cur) {
 249                                 cur = read_tree_block(root, blocknr,
 250                                                          blocksize);
 251                         } else if (!uptodate) {
 252                                 btrfs_read_buffer(cur);
 253                         }
 254                 }
 255                 if (search_start == 0)
 256                         search_start = last_block;
 257
 258                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 259                                         &tmp, search_start,
 260                                         min(16 * blocksize,
 261                                             (end_slot - i) * blocksize));
 262                 if (err) {
 263                         free_extent_buffer(cur);
 264                         break;
 265                 }
 266                 search_start = tmp->start;
 267                 last_block = tmp->start;
 268                 *last_ret = search_start;
 269                 if (parent_level == 1)
 270                         btrfs_clear_buffer_defrag(tmp);
 271                 free_extent_buffer(tmp);
 272         }
 273         if (parent->map_token) {
 274                 unmap_extent_buffer(parent, parent->map_token,
 275                                     KM_USER1);
 276                 parent->map_token = NULL;
 277         }
 278         return err;
 279 }
 280
 281 /*
 282  * The leaf data grows from end-to-front in the node.
 283  * this returns the address of the start of the last item,
 284  * which is the stop of the leaf data stack
 285  */
 286 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 287                                          struct extent_buffer *leaf)
 288 {
 289         u32 nr = btrfs_header_nritems(leaf);
 290         if (nr == 0)
 291                 return BTRFS_LEAF_DATA_SIZE(root);
 292         return btrfs_item_offset_nr(leaf, nr - 1);
 293 }
 294
 295 /*
 296  * compare two keys in a memcmp fashion
 297  */
 298 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 299 {
 300         struct btrfs_key k1;
 301
 302         btrfs_disk_key_to_cpu(&k1, disk);
 303
 304         if (k1.objectid > k2->objectid)
 305                 return 1;
 306         if (k1.objectid < k2->objectid)
 307                 return -1;
 308         if (k1.type > k2->type)
 309                 return 1;
 310         if (k1.type < k2->type)
 311                 return -1;
 312         if (k1.offset > k2->offset)
 313                 return 1;
 314         if (k1.offset < k2->offset)
 315                 return -1;
 316         return 0;
 317 }
 318
 319 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
 320                       int level)
 321 {
 322         struct extent_buffer *parent = NULL;
 323         struct extent_buffer *node = path->nodes[level];
 324         struct btrfs_disk_key parent_key;
 325         struct btrfs_disk_key node_key;
 326         int parent_slot;
 327         int slot;
 328         struct btrfs_key cpukey;
 329         u32 nritems = btrfs_header_nritems(node);
 330
 331         if (path->nodes[level + 1])
 332                 parent = path->nodes[level + 1];
 333
 334         slot = path->slots[level];
 335         BUG_ON(nritems == 0);
 336         if (parent) {
 337                 parent_slot = path->slots[level + 1];
 338                 btrfs_node_key(parent, &parent_key, parent_slot);
 339                 btrfs_node_key(node, &node_key, 0);
 340                 BUG_ON(memcmp(&parent_key, &node_key,
 341                               sizeof(struct btrfs_disk_key)));
 342                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 343                        btrfs_header_bytenr(node));
 344         }
 345         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 346         if (slot != 0) {
 347                 btrfs_node_key_to_cpu(node, &cpukey, slot - 1);
 348                 btrfs_node_key(node, &node_key, slot);
 349                 BUG_ON(comp_keys(&node_key, &cpukey) <= 0);
 350         }
 351         if (slot < nritems - 1) {
 352                 btrfs_node_key_to_cpu(node, &cpukey, slot + 1);
 353                 btrfs_node_key(node, &node_key, slot);
 354                 BUG_ON(comp_keys(&node_key, &cpukey) >= 0);
 355         }
 356         return 0;
 357 }
 358
 359 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 360                       int level)
 361 {
 362         struct extent_buffer *leaf = path->nodes[level];
 363         struct extent_buffer *parent = NULL;
 364         int parent_slot;
 365         struct btrfs_key cpukey;
 366         struct btrfs_disk_key parent_key;
 367         struct btrfs_disk_key leaf_key;
 368         int slot = path->slots[0];
 369
 370         u32 nritems = btrfs_header_nritems(leaf);
 371
 372         if (path->nodes[level + 1])
 373                 parent = path->nodes[level + 1];
 374
 375         if (nritems == 0)
 376                 return 0;
 377
 378         if (parent) {
 379                 parent_slot = path->slots[level + 1];
 380                 btrfs_node_key(parent, &parent_key, parent_slot);
 381                 btrfs_item_key(leaf, &leaf_key, 0);
 382
 383                 BUG_ON(memcmp(&parent_key, &leaf_key,
 384                        sizeof(struct btrfs_disk_key)));
 385                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 386                        btrfs_header_bytenr(leaf));
 387         }
 388 #if 0
 389         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 390                 btrfs_item_key_to_cpu(leaf, &cpukey, i + 1);
 391                 btrfs_item_key(leaf, &leaf_key, i);
 392                 if (comp_keys(&leaf_key, &cpukey) >= 0) {
 393                         btrfs_print_leaf(root, leaf);
 394                         printk("slot %d offset bad key\n", i);
 395                         BUG_ON(1);
 396                 }
 397                 if (btrfs_item_offset_nr(leaf, i) !=
 398                         btrfs_item_end_nr(leaf, i + 1)) {
 399                         btrfs_print_leaf(root, leaf);
 400                         printk("slot %d offset bad\n", i);
 401                         BUG_ON(1);
 402                 }
 403                 if (i == 0) {
 404                         if (btrfs_item_offset_nr(leaf, i) +
 405                                btrfs_item_size_nr(leaf, i) !=
 406                                BTRFS_LEAF_DATA_SIZE(root)) {
 407                                 btrfs_print_leaf(root, leaf);
 408                                 printk("slot %d first offset bad\n", i);
 409                                 BUG_ON(1);
 410                         }
 411                 }
 412         }
 413         if (nritems > 0) {
 414                 if (btrfs_item_size_nr(leaf, nritems - 1) > 4096) {
 415                                 btrfs_print_leaf(root, leaf);
 416                                 printk("slot %d bad size \n", nritems - 1);
 417                                 BUG_ON(1);
 418                 }
 419         }
 420 #endif
 421         if (slot != 0 && slot < nritems - 1) {
 422                 btrfs_item_key(leaf, &leaf_key, slot);
 423                 btrfs_item_key_to_cpu(leaf, &cpukey, slot - 1);
 424                 if (comp_keys(&leaf_key, &cpukey) <= 0) {
 425                         btrfs_print_leaf(root, leaf);
 426                         printk("slot %d offset bad key\n", slot);
 427                         BUG_ON(1);
 428                 }
 429                 if (btrfs_item_offset_nr(leaf, slot - 1) !=
 430                        btrfs_item_end_nr(leaf, slot)) {
 431                         btrfs_print_leaf(root, leaf);
 432                         printk("slot %d offset bad\n", slot);
 433                         BUG_ON(1);
 434                 }
 435         }
 436         if (slot < nritems - 1) {
 437                 btrfs_item_key(leaf, &leaf_key, slot);
 438                 btrfs_item_key_to_cpu(leaf, &cpukey, slot + 1);
 439                 BUG_ON(comp_keys(&leaf_key, &cpukey) >= 0);
 440                 if (btrfs_item_offset_nr(leaf, slot) !=
 441                         btrfs_item_end_nr(leaf, slot + 1)) {
 442                         btrfs_print_leaf(root, leaf);
 443                         printk("slot %d offset bad\n", slot);
 444                         BUG_ON(1);
 445                 }
 446         }
 447         BUG_ON(btrfs_item_offset_nr(leaf, 0) +
 448                btrfs_item_size_nr(leaf, 0) != BTRFS_LEAF_DATA_SIZE(root));
 449         return 0;
 450 }
 451
 452 static int check_block(struct btrfs_root *root, struct btrfs_path *path,
 453                         int level)
 454 {
 455         return 0;
 456 #if 0
 457         struct extent_buffer *buf = path->nodes[level];
 458
 459         if (memcmp_extent_buffer(buf, root->fs_info->fsid,
 460                                  (unsigned long)btrfs_header_fsid(buf),
 461                                  BTRFS_FSID_SIZE)) {
 462                 printk("warning bad block %Lu\n", buf->start);
 463                 return 1;
 464         }
 465 #endif
 466         if (level == 0)
 467                 return check_leaf(root, path, level);
 468         return check_node(root, path, level);
 469 }
 470
 471 /*
 472  * search for key in the extent_buffer.  The items start at offset p,
 473  * and they are item_size apart.  There are 'max' items in p.
 474  *
 475  * the slot in the array is returned via slot, and it points to
 476  * the place where you would insert key if it is not found in
 477  * the array.
 478  *
 479  * slot may point to max if the key is bigger than all of the keys
 480  */
 481 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 482                               int item_size, struct btrfs_key *key,
 483                               int max, int *slot)
 484 {
 485         int low = 0;
 486         int high = max;
 487         int mid;
 488         int ret;
 489         struct btrfs_disk_key *tmp = NULL;
 490         struct btrfs_disk_key unaligned;
 491         unsigned long offset;
 492         char *map_token = NULL;
 493         char *kaddr = NULL;
 494         unsigned long map_start = 0;
 495         unsigned long map_len = 0;
 496         int err;
 497
 498         while(low < high) {
 499                 mid = (low + high) / 2;
 500                 offset = p + mid * item_size;
 501
 502                 if (!map_token || offset < map_start ||
 503                     (offset + sizeof(struct btrfs_disk_key)) >
 504                     map_start + map_len) {
 505                         if (map_token) {
 506                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 507                                 map_token = NULL;
 508                         }
 509                         err = map_extent_buffer(eb, offset,
 510                                                 sizeof(struct btrfs_disk_key),
 511                                                 &map_token, &kaddr,
 512                                                 &map_start, &map_len, KM_USER0);
 513
 514                         if (!err) {
 515                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
 516                                                         map_start);
 517                         } else {
 518                                 read_extent_buffer(eb, &unaligned,
 519                                                    offset, sizeof(unaligned));
 520                                 tmp = &unaligned;
 521                         }
 522
 523                 } else {
 524                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
 525                                                         map_start);
 526                 }
 527                 ret = comp_keys(tmp, key);
 528
 529                 if (ret < 0)
 530                         low = mid + 1;
 531                 else if (ret > 0)
 532                         high = mid;
 533                 else {
 534                         *slot = mid;
 535                         if (map_token)
 536                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 537                         return 0;
 538                 }
 539         }
 540         *slot = low;
 541         if (map_token)
 542                 unmap_extent_buffer(eb, map_token, KM_USER0);
 543         return 1;
 544 }
 545
 546 /*
 547  * simple bin_search frontend that does the right thing for
 548  * leaves vs nodes
 549  */
 550 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 551                       int level, int *slot)
 552 {
 553         if (level == 0) {
 554                 return generic_bin_search(eb,
 555                                           offsetof(struct btrfs_leaf, items),
 556                                           sizeof(struct btrfs_item),
 557                                           key, btrfs_header_nritems(eb),
 558                                           slot);
 559         } else {
 560                 return generic_bin_search(eb,
 561                                           offsetof(struct btrfs_node, ptrs),
 562                                           sizeof(struct btrfs_key_ptr),
 563                                           key, btrfs_header_nritems(eb),
 564                                           slot);
 565         }
 566         return -1;
 567 }
 568
 569 static struct extent_buffer *read_node_slot(struct btrfs_root *root,
 570                                    struct extent_buffer *parent, int slot)
 571 {
 572         if (slot < 0)
 573                 return NULL;
 574         if (slot >= btrfs_header_nritems(parent))
 575                 return NULL;
 576         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 577                        btrfs_level_size(root, btrfs_header_level(parent) - 1));
 578 }
 579
 580 static int balance_level(struct btrfs_trans_handle *trans, struct btrfs_root
 581                          *root, struct btrfs_path *path, int level)
 582 {
 583         struct extent_buffer *right = NULL;
 584         struct extent_buffer *mid;
 585         struct extent_buffer *left = NULL;
 586         struct extent_buffer *parent = NULL;
 587         int ret = 0;
 588         int wret;
 589         int pslot;
 590         int orig_slot = path->slots[level];
 591         int err_on_enospc = 0;
 592         u64 orig_ptr;
 593
 594         if (level == 0)
 595                 return 0;
 596
 597         mid = path->nodes[level];
 598         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 599
 600         if (level < BTRFS_MAX_LEVEL - 1)
 601                 parent = path->nodes[level + 1];
 602         pslot = path->slots[level + 1];
 603
 604         /*
 605          * deal with the case where there is only one pointer in the root
 606          * by promoting the node below to a root
 607          */
 608         if (!parent) {
 609                 struct extent_buffer *child;
 610
 611                 if (btrfs_header_nritems(mid) != 1)
 612                         return 0;
 613
 614                 /* promote the child to a root */
 615                 child = read_node_slot(root, mid, 0);
 616                 BUG_ON(!child);
 617                 root->node = child;
 618                 path->nodes[level] = NULL;
 619                 clean_tree_block(trans, root, mid);
 620                 wait_on_tree_block_writeback(root, mid);
 621                 /* once for the path */
 622                 free_extent_buffer(mid);
 623                 ret = btrfs_free_extent(trans, root, mid->start, mid->len, 1);
 624                 /* once for the root ptr */
 625                 free_extent_buffer(mid);
 626                 return ret;
 627         }
 628         if (btrfs_header_nritems(mid) >
 629             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 630                 return 0;
 631
 632         if (btrfs_header_nritems(mid) < 2)
 633                 err_on_enospc = 1;
 634
 635         left = read_node_slot(root, parent, pslot - 1);
 636         if (left) {
 637                 wret = btrfs_cow_block(trans, root, left,
 638                                        parent, pslot - 1, &left);
 639                 if (wret) {
 640                         ret = wret;
 641                         goto enospc;
 642                 }
 643         }
 644         right = read_node_slot(root, parent, pslot + 1);
 645         if (right) {
 646                 wret = btrfs_cow_block(trans, root, right,
 647                                        parent, pslot + 1, &right);
 648                 if (wret) {
 649                         ret = wret;
 650                         goto enospc;
 651                 }
 652         }
 653
 654         /* first, try to make some room in the middle buffer */
 655         if (left) {
 656                 orig_slot += btrfs_header_nritems(left);
 657                 wret = push_node_left(trans, root, left, mid);
 658                 if (wret < 0)
 659                         ret = wret;
 660                 if (btrfs_header_nritems(mid) < 2)
 661                         err_on_enospc = 1;
 662         }
 663
 664         /*
 665          * then try to empty the right most buffer into the middle
 666          */
 667         if (right) {
 668                 wret = push_node_left(trans, root, mid, right);
 669                 if (wret < 0 && wret != -ENOSPC)
 670                         ret = wret;
 671                 if (btrfs_header_nritems(right) == 0) {
 672                         u64 bytenr = right->start;
 673                         u32 blocksize = right->len;
 674
 675                         clean_tree_block(trans, root, right);
 676                         wait_on_tree_block_writeback(root, right);
 677                         free_extent_buffer(right);
 678                         right = NULL;
 679                         wret = del_ptr(trans, root, path, level + 1, pslot +
 680                                        1);
 681                         if (wret)
 682                                 ret = wret;
 683                         wret = btrfs_free_extent(trans, root, bytenr,
 684                                                  blocksize, 1);
 685                         if (wret)
 686                                 ret = wret;
 687                 } else {
 688                         struct btrfs_disk_key right_key;
 689                         btrfs_node_key(right, &right_key, 0);
 690                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 691                         btrfs_mark_buffer_dirty(parent);
 692                 }
 693         }
 694         if (btrfs_header_nritems(mid) == 1) {
 695                 /*
 696                  * we're not allowed to leave a node with one item in the
 697                  * tree during a delete.  A deletion from lower in the tree
 698                  * could try to delete the only pointer in this node.
 699                  * So, pull some keys from the left.
 700                  * There has to be a left pointer at this point because
 701                  * otherwise we would have pulled some pointers from the
 702                  * right
 703                  */
 704                 BUG_ON(!left);
 705                 wret = balance_node_right(trans, root, mid, left);
 706                 if (wret < 0) {
 707                         ret = wret;
 708                         goto enospc;
 709                 }
 710                 BUG_ON(wret == 1);
 711         }
 712         if (btrfs_header_nritems(mid) == 0) {
 713                 /* we've managed to empty the middle node, drop it */
 714                 u64 bytenr = mid->start;
 715                 u32 blocksize = mid->len;
 716                 clean_tree_block(trans, root, mid);
 717                 wait_on_tree_block_writeback(root, mid);
 718                 free_extent_buffer(mid);
 719                 mid = NULL;
 720                 wret = del_ptr(trans, root, path, level + 1, pslot);
 721                 if (wret)
 722                         ret = wret;
 723                 wret = btrfs_free_extent(trans, root, bytenr, blocksize, 1);
 724                 if (wret)
 725                         ret = wret;
 726         } else {
 727                 /* update the parent key to reflect our changes */
 728                 struct btrfs_disk_key mid_key;
 729                 btrfs_node_key(mid, &mid_key, 0);
 730                 btrfs_set_node_key(parent, &mid_key, pslot);
 731                 btrfs_mark_buffer_dirty(parent);
 732         }
 733
 734         /* update the path */
 735         if (left) {
 736                 if (btrfs_header_nritems(left) > orig_slot) {
 737                         extent_buffer_get(left);
 738                         path->nodes[level] = left;
 739                         path->slots[level + 1] -= 1;
 740                         path->slots[level] = orig_slot;
 741                         if (mid)
 742                                 free_extent_buffer(mid);
 743                 } else {
 744                         orig_slot -= btrfs_header_nritems(left);
 745                         path->slots[level] = orig_slot;
 746                 }
 747         }
 748         /* double check we haven't messed things up */
 749         check_block(root, path, level);
 750         if (orig_ptr !=
 751             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 752                 BUG();
 753 enospc:
 754         if (right)
 755                 free_extent_buffer(right);
 756         if (left)
 757                 free_extent_buffer(left);
 758         return ret;
 759 }
 760
 761 /* returns zero if the push worked, non-zero otherwise */
 762 static int push_nodes_for_insert(struct btrfs_trans_handle *trans,
 763                                 struct btrfs_root *root,
 764                                 struct btrfs_path *path, int level)
 765 {
 766         struct extent_buffer *right = NULL;
 767         struct extent_buffer *mid;
 768         struct extent_buffer *left = NULL;
 769         struct extent_buffer *parent = NULL;
 770         int ret = 0;
 771         int wret;
 772         int pslot;
 773         int orig_slot = path->slots[level];
 774         u64 orig_ptr;
 775
 776         if (level == 0)
 777                 return 1;
 778
 779         mid = path->nodes[level];
 780         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 781
 782         if (level < BTRFS_MAX_LEVEL - 1)
 783                 parent = path->nodes[level + 1];
 784         pslot = path->slots[level + 1];
 785
 786         if (!parent)
 787                 return 1;
 788
 789         left = read_node_slot(root, parent, pslot - 1);
 790
 791         /* first, try to make some room in the middle buffer */
 792         if (left) {
 793                 u32 left_nr;
 794                 left_nr = btrfs_header_nritems(left);
 795                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 796                         wret = 1;
 797                 } else {
 798                         ret = btrfs_cow_block(trans, root, left, parent,
 799                                               pslot - 1, &left);
 800                         if (ret)
 801                                 wret = 1;
 802                         else {
 803                                 wret = push_node_left(trans, root,
 804                                                       left, mid);
 805                         }
 806                 }
 807                 if (wret < 0)
 808                         ret = wret;
 809                 if (wret == 0) {
 810                         struct btrfs_disk_key disk_key;
 811                         orig_slot += left_nr;
 812                         btrfs_node_key(mid, &disk_key, 0);
 813                         btrfs_set_node_key(parent, &disk_key, pslot);
 814                         btrfs_mark_buffer_dirty(parent);
 815                         if (btrfs_header_nritems(left) > orig_slot) {
 816                                 path->nodes[level] = left;
 817                                 path->slots[level + 1] -= 1;
 818                                 path->slots[level] = orig_slot;
 819                                 free_extent_buffer(mid);
 820                         } else {
 821                                 orig_slot -=
 822                                         btrfs_header_nritems(left);
 823                                 path->slots[level] = orig_slot;
 824                                 free_extent_buffer(left);
 825                         }
 826                         return 0;
 827                 }
 828                 free_extent_buffer(left);
 829         }
 830         right= read_node_slot(root, parent, pslot + 1);
 831
 832         /*
 833          * then try to empty the right most buffer into the middle
 834          */
 835         if (right) {
 836                 u32 right_nr;
 837                 right_nr = btrfs_header_nritems(right);
 838                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 839                         wret = 1;
 840                 } else {
 841                         ret = btrfs_cow_block(trans, root, right,
 842                                               parent, pslot + 1,
 843                                               &right);
 844                         if (ret)
 845                                 wret = 1;
 846                         else {
 847                                 wret = balance_node_right(trans, root,
 848                                                           right, mid);
 849                         }
 850                 }
 851                 if (wret < 0)
 852                         ret = wret;
 853                 if (wret == 0) {
 854                         struct btrfs_disk_key disk_key;
 855
 856                         btrfs_node_key(right, &disk_key, 0);
 857                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
 858                         btrfs_mark_buffer_dirty(parent);
 859
 860                         if (btrfs_header_nritems(mid) <= orig_slot) {
 861                                 path->nodes[level] = right;
 862                                 path->slots[level + 1] += 1;
 863                                 path->slots[level] = orig_slot -
 864                                         btrfs_header_nritems(mid);
 865                                 free_extent_buffer(mid);
 866                         } else {
 867                                 free_extent_buffer(right);
 868                         }
 869                         return 0;
 870                 }
 871                 free_extent_buffer(right);
 872         }
 873         return 1;
 874 }
 875
 876 /*
 877  * readahead one full node of leaves
 878  */
 879 static void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
 880                              int level, int slot)
 881 {
 882         struct extent_buffer *node;
 883         u32 nritems;
 884         u64 search;
 885         u64 lowest_read;
 886         u64 highest_read;
 887         u64 nread = 0;
 888         int direction = path->reada;
 889         struct extent_buffer *eb;
 890         u32 nr;
 891         u32 blocksize;
 892         u32 nscan = 0;
 893
 894         if (level != 1)
 895                 return;
 896
 897         if (!path->nodes[level])
 898                 return;
 899
 900         node = path->nodes[level];
 901         search = btrfs_node_blockptr(node, slot);
 902         blocksize = btrfs_level_size(root, level - 1);
 903         eb = btrfs_find_tree_block(root, search, blocksize);
 904         if (eb) {
 905                 free_extent_buffer(eb);
 906                 return;
 907         }
 908
 909         highest_read = search;
 910         lowest_read = search;
 911
 912         nritems = btrfs_header_nritems(node);
 913         nr = slot;
 914         while(1) {
 915                 if (direction < 0) {
 916                         if (nr == 0)
 917                                 break;
 918                         nr--;
 919                 } else if (direction > 0) {
 920                         nr++;
 921                         if (nr >= nritems)
 922                                 break;
 923                 }
 924                 search = btrfs_node_blockptr(node, nr);
 925                 if ((search >= lowest_read && search <= highest_read) ||
 926                     (search < lowest_read && lowest_read - search <= 32768) ||
 927                     (search > highest_read && search - highest_read <= 32768)) {
 928                         readahead_tree_block(root, search, blocksize);
 929                         nread += blocksize;
 930                 }
 931                 nscan++;
 932                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
 933                         break;
 934                 if(nread > (1024 * 1024) || nscan > 128)
 935                         break;
 936
 937                 if (search < lowest_read)
 938                         lowest_read = search;
 939                 if (search > highest_read)
 940                         highest_read = search;
 941         }
 942 }
 943 /*
 944  * look for key in the tree.  path is filled in with nodes along the way
 945  * if key is found, we return zero and you can find the item in the leaf
 946  * level of the path (level 0)
 947  *
 948  * If the key isn't found, the path points to the slot where it should
 949  * be inserted, and 1 is returned.  If there are other errors during the
 950  * search a negative error number is returned.
 951  *
 952  * if ins_len > 0, nodes and leaves will be split as we walk down the
 953  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
 954  * possible)
 955  */
 956 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
 957                       *root, struct btrfs_key *key, struct btrfs_path *p, int
 958                       ins_len, int cow)
 959 {
 960         struct extent_buffer *b;
 961         u64 bytenr;
 962         int slot;
 963         int ret;
 964         int level;
 965         int should_reada = p->reada;
 966         u8 lowest_level = 0;
 967
 968         lowest_level = p->lowest_level;
 969         WARN_ON(lowest_level && ins_len);
 970         WARN_ON(p->nodes[0] != NULL);
 971         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
 972 again:
 973         b = root->node;
 974         extent_buffer_get(b);
 975         while (b) {
 976                 level = btrfs_header_level(b);
 977                 if (cow) {
 978                         int wret;
 979                         wret = btrfs_cow_block(trans, root, b,
 980                                                p->nodes[level + 1],
 981                                                p->slots[level + 1],
 982                                                &b);
 983                         if (wret) {
 984                                 free_extent_buffer(b);
 985                                 return wret;
 986                         }
 987                 }
 988                 BUG_ON(!cow && ins_len);
 989                 if (level != btrfs_header_level(b))
 990                         WARN_ON(1);
 991                 level = btrfs_header_level(b);
 992                 p->nodes[level] = b;
 993                 ret = check_block(root, p, level);
 994                 if (ret)
 995                         return -1;
 996                 ret = bin_search(b, key, level, &slot);
 997                 if (level != 0) {
 998                         if (ret && slot > 0)
 999                                 slot -= 1;
1000                         p->slots[level] = slot;
1001                         if (ins_len > 0 && btrfs_header_nritems(b) >=
1002                             BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1003                                 int sret = split_node(trans, root, p, level);
1004                                 BUG_ON(sret > 0);
1005                                 if (sret)
1006                                         return sret;
1007                                 b = p->nodes[level];
1008                                 slot = p->slots[level];
1009                         } else if (ins_len < 0) {
1010                                 int sret = balance_level(trans, root, p,
1011                                                          level);
1012                                 if (sret)
1013                                         return sret;
1014                                 b = p->nodes[level];
1015                                 if (!b) {
1016                                         btrfs_release_path(NULL, p);
1017                                         goto again;
1018                                 }
1019                                 slot = p->slots[level];
1020                                 BUG_ON(btrfs_header_nritems(b) == 1);
1021                         }
1022                         /* this is only true while dropping a snapshot */
1023                         if (level == lowest_level)
1024                                 break;
1025                         bytenr = btrfs_node_blockptr(b, slot);
1026                         if (should_reada)
1027                                 reada_for_search(root, p, level, slot);
1028                         b = read_tree_block(root, bytenr,
1029                                             btrfs_level_size(root, level - 1));
1030                 } else {
1031                         p->slots[level] = slot;
1032                         if (ins_len > 0 && btrfs_leaf_free_space(root, b) <
1033                             sizeof(struct btrfs_item) + ins_len) {
1034                                 int sret = split_leaf(trans, root, key,
1035                                                       p, ins_len, ret == 0);
1036                                 BUG_ON(sret > 0);
1037                                 if (sret)
1038                                         return sret;
1039                         }
1040                         return ret;
1041                 }
1042         }
1043         return 1;
1044 }
1045
1046 /*
1047  * adjust the pointers going up the tree, starting at level
1048  * making sure the right key of each node is points to 'key'.
1049  * This is used after shifting pointers to the left, so it stops
1050  * fixing up pointers when a given leaf/node is not in slot 0 of the
1051  * higher levels
1052  *
1053  * If this fails to write a tree block, it returns -1, but continues
1054  * fixing up the blocks in ram so the tree is consistent.
1055  */
1056 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1057                           struct btrfs_root *root, struct btrfs_path *path,
1058                           struct btrfs_disk_key *key, int level)
1059 {
1060         int i;
1061         int ret = 0;
1062         struct extent_buffer *t;
1063
1064         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1065                 int tslot = path->slots[i];
1066                 if (!path->nodes[i])
1067                         break;
1068                 t = path->nodes[i];
1069                 btrfs_set_node_key(t, key, tslot);
1070                 btrfs_mark_buffer_dirty(path->nodes[i]);
1071                 if (tslot != 0)
1072                         break;
1073         }
1074         return ret;
1075 }
1076
1077 /*
1078  * try to push data from one node into the next node left in the
1079  * tree.
1080  *
1081  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1082  * error, and > 0 if there was no room in the left hand block.
1083  */
1084 static int push_node_left(struct btrfs_trans_handle *trans, struct btrfs_root
1085                           *root, struct extent_buffer *dst,
1086                           struct extent_buffer *src)
1087 {
1088         int push_items = 0;
1089         int src_nritems;
1090         int dst_nritems;
1091         int ret = 0;
1092
1093         src_nritems = btrfs_header_nritems(src);
1094         dst_nritems = btrfs_header_nritems(dst);
1095         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1096
1097         if (push_items <= 0) {
1098                 return 1;
1099         }
1100
1101         if (src_nritems < push_items)
1102                 push_items = src_nritems;
1103
1104         copy_extent_buffer(dst, src,
1105                            btrfs_node_key_ptr_offset(dst_nritems),
1106                            btrfs_node_key_ptr_offset(0),
1107                            push_items * sizeof(struct btrfs_key_ptr));
1108
1109         if (push_items < src_nritems) {
1110                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1111                                       btrfs_node_key_ptr_offset(push_items),
1112                                       (src_nritems - push_items) *
1113                                       sizeof(struct btrfs_key_ptr));
1114         }
1115         btrfs_set_header_nritems(src, src_nritems - push_items);
1116         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1117         btrfs_mark_buffer_dirty(src);
1118         btrfs_mark_buffer_dirty(dst);
1119         return ret;
1120 }
1121
1122 /*
1123  * try to push data from one node into the next node right in the
1124  * tree.
1125  *
1126  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1127  * error, and > 0 if there was no room in the right hand block.
1128  *
1129  * this will  only push up to 1/2 the contents of the left node over
1130  */
1131 static int balance_node_right(struct btrfs_trans_handle *trans,
1132                               struct btrfs_root *root,
1133                               struct extent_buffer *dst,
1134                               struct extent_buffer *src)
1135 {
1136         int push_items = 0;
1137         int max_push;
1138         int src_nritems;
1139         int dst_nritems;
1140         int ret = 0;
1141
1142         src_nritems = btrfs_header_nritems(src);
1143         dst_nritems = btrfs_header_nritems(dst);
1144         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1145         if (push_items <= 0)
1146                 return 1;
1147
1148         max_push = src_nritems / 2 + 1;
1149         /* don't try to empty the node */
1150         if (max_push >= src_nritems)
1151                 return 1;
1152
1153         if (max_push < push_items)
1154                 push_items = max_push;
1155
1156         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1157                                       btrfs_node_key_ptr_offset(0),
1158                                       (dst_nritems) *
1159                                       sizeof(struct btrfs_key_ptr));
1160
1161         copy_extent_buffer(dst, src,
1162                            btrfs_node_key_ptr_offset(0),
1163                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1164                            push_items * sizeof(struct btrfs_key_ptr));
1165
1166         btrfs_set_header_nritems(src, src_nritems - push_items);
1167         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1168
1169         btrfs_mark_buffer_dirty(src);
1170         btrfs_mark_buffer_dirty(dst);
1171         return ret;
1172 }
1173
1174 /*
1175  * helper function to insert a new root level in the tree.
1176  * A new node is allocated, and a single item is inserted to
1177  * point to the existing root
1178  *
1179  * returns zero on success or < 0 on failure.
1180  */
1181 static int insert_new_root(struct btrfs_trans_handle *trans,
1182                            struct btrfs_root *root,
1183                            struct btrfs_path *path, int level)
1184 {
1185         struct extent_buffer *lower;
1186         struct extent_buffer *c;
1187         struct btrfs_disk_key lower_key;
1188
1189         BUG_ON(path->nodes[level]);
1190         BUG_ON(path->nodes[level-1] != root->node);
1191
1192         c = btrfs_alloc_free_block(trans, root, root->nodesize,
1193                                    root->node->start, 0);
1194         if (IS_ERR(c))
1195                 return PTR_ERR(c);
1196         memset_extent_buffer(c, 0, 0, root->nodesize);
1197         btrfs_set_header_nritems(c, 1);
1198         btrfs_set_header_level(c, level);
1199         btrfs_set_header_bytenr(c, c->start);
1200         btrfs_set_header_generation(c, trans->transid);
1201         btrfs_set_header_owner(c, root->root_key.objectid);
1202         lower = path->nodes[level-1];
1203
1204         write_extent_buffer(c, root->fs_info->fsid,
1205                             (unsigned long)btrfs_header_fsid(c),
1206                             BTRFS_FSID_SIZE);
1207         if (level == 1)
1208                 btrfs_item_key(lower, &lower_key, 0);
1209         else
1210                 btrfs_node_key(lower, &lower_key, 0);
1211         btrfs_set_node_key(c, &lower_key, 0);
1212         btrfs_set_node_blockptr(c, 0, lower->start);
1213
1214         btrfs_mark_buffer_dirty(c);
1215
1216         /* the super has an extra ref to root->node */
1217         free_extent_buffer(root->node);
1218         root->node = c;
1219         extent_buffer_get(c);
1220         path->nodes[level] = c;
1221         path->slots[level] = 0;
1222         return 0;
1223 }
1224
1225 /*
1226  * worker function to insert a single pointer in a node.
1227  * the node should have enough room for the pointer already
1228  *
1229  * slot and level indicate where you want the key to go, and
1230  * blocknr is the block the key points to.
1231  *
1232  * returns zero on success and < 0 on any error
1233  */
1234 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1235                       *root, struct btrfs_path *path, struct btrfs_disk_key
1236                       *key, u64 bytenr, int slot, int level)
1237 {
1238         struct extent_buffer *lower;
1239         int nritems;
1240
1241         BUG_ON(!path->nodes[level]);
1242         lower = path->nodes[level];
1243         nritems = btrfs_header_nritems(lower);
1244         if (slot > nritems)
1245                 BUG();
1246         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1247                 BUG();
1248         if (slot != nritems) {
1249                 memmove_extent_buffer(lower,
1250                               btrfs_node_key_ptr_offset(slot + 1),
1251                               btrfs_node_key_ptr_offset(slot),
1252                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1253         }
1254         btrfs_set_node_key(lower, key, slot);
1255         btrfs_set_node_blockptr(lower, slot, bytenr);
1256         btrfs_set_header_nritems(lower, nritems + 1);
1257         btrfs_mark_buffer_dirty(lower);
1258         return 0;
1259 }
1260
1261 /*
1262  * split the node at the specified level in path in two.
1263  * The path is corrected to point to the appropriate node after the split
1264  *
1265  * Before splitting this tries to make some room in the node by pushing
1266  * left and right, if either one works, it returns right away.
1267  *
1268  * returns 0 on success and < 0 on failure
1269  */
1270 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1271                       *root, struct btrfs_path *path, int level)
1272 {
1273         struct extent_buffer *c;
1274         struct extent_buffer *split;
1275         struct btrfs_disk_key disk_key;
1276         int mid;
1277         int ret;
1278         int wret;
1279         u32 c_nritems;
1280
1281         c = path->nodes[level];
1282         if (c == root->node) {
1283                 /* trying to split the root, lets make a new one */
1284                 ret = insert_new_root(trans, root, path, level + 1);
1285                 if (ret)
1286                         return ret;
1287         } else {
1288                 ret = push_nodes_for_insert(trans, root, path, level);
1289                 c = path->nodes[level];
1290                 if (!ret && btrfs_header_nritems(c) <
1291                     BTRFS_NODEPTRS_PER_BLOCK(root) - 1)
1292                         return 0;
1293                 if (ret < 0)
1294                         return ret;
1295         }
1296
1297         c_nritems = btrfs_header_nritems(c);
1298         split = btrfs_alloc_free_block(trans, root, root->nodesize,
1299                                        c->start, 0);
1300         if (IS_ERR(split))
1301                 return PTR_ERR(split);
1302
1303         btrfs_set_header_flags(split, btrfs_header_flags(c));
1304         btrfs_set_header_level(split, btrfs_header_level(c));
1305         btrfs_set_header_bytenr(split, split->start);
1306         btrfs_set_header_generation(split, trans->transid);
1307         btrfs_set_header_owner(split, root->root_key.objectid);
1308         write_extent_buffer(split, root->fs_info->fsid,
1309                             (unsigned long)btrfs_header_fsid(split),
1310                             BTRFS_FSID_SIZE);
1311
1312         mid = (c_nritems + 1) / 2;
1313
1314         copy_extent_buffer(split, c,
1315                            btrfs_node_key_ptr_offset(0),
1316                            btrfs_node_key_ptr_offset(mid),
1317                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1318         btrfs_set_header_nritems(split, c_nritems - mid);
1319         btrfs_set_header_nritems(c, mid);
1320         ret = 0;
1321
1322         btrfs_mark_buffer_dirty(c);
1323         btrfs_mark_buffer_dirty(split);
1324
1325         btrfs_node_key(split, &disk_key, 0);
1326         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1327                           path->slots[level + 1] + 1,
1328                           level + 1);
1329         if (wret)
1330                 ret = wret;
1331
1332         if (path->slots[level] >= mid) {
1333                 path->slots[level] -= mid;
1334                 free_extent_buffer(c);
1335                 path->nodes[level] = split;
1336                 path->slots[level + 1] += 1;
1337         } else {
1338                 free_extent_buffer(split);
1339         }
1340         return ret;
1341 }
1342
1343 /*
1344  * how many bytes are required to store the items in a leaf.  start
1345  * and nr indicate which items in the leaf to check.  This totals up the
1346  * space used both by the item structs and the item data
1347  */
1348 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1349 {
1350         int data_len;
1351         int nritems = btrfs_header_nritems(l);
1352         int end = min(nritems, start + nr) - 1;
1353
1354         if (!nr)
1355                 return 0;
1356         data_len = btrfs_item_end_nr(l, start);
1357         data_len = data_len - btrfs_item_offset_nr(l, end);
1358         data_len += sizeof(struct btrfs_item) * nr;
1359         WARN_ON(data_len < 0);
1360         return data_len;
1361 }
1362
1363 /*
1364  * The space between the end of the leaf items and
1365  * the start of the leaf data.  IOW, how much room
1366  * the leaf has left for both items and data
1367  */
1368 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1369 {
1370         int nritems = btrfs_header_nritems(leaf);
1371         int ret;
1372         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1373         if (ret < 0) {
1374                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1375                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1376                        leaf_space_used(leaf, 0, nritems), nritems);
1377         }
1378         return ret;
1379 }
1380
1381 /*
1382  * push some data in the path leaf to the right, trying to free up at
1383  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1384  *
1385  * returns 1 if the push failed because the other node didn't have enough
1386  * room, 0 if everything worked out and < 0 if there were major errors.
1387  */
1388 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1389                            *root, struct btrfs_path *path, int data_size)
1390 {
1391         struct extent_buffer *left = path->nodes[0];
1392         struct extent_buffer *right;
1393         struct extent_buffer *upper;
1394         struct btrfs_disk_key disk_key;
1395         int slot;
1396         int i;
1397         int free_space;
1398         int push_space = 0;
1399         int push_items = 0;
1400         struct btrfs_item *item;
1401         u32 left_nritems;
1402         u32 right_nritems;
1403         u32 data_end;
1404         u32 this_item_size;
1405         int ret;
1406
1407         slot = path->slots[1];
1408         if (!path->nodes[1]) {
1409                 return 1;
1410         }
1411         upper = path->nodes[1];
1412         if (slot >= btrfs_header_nritems(upper) - 1)
1413                 return 1;
1414
1415         right = read_tree_block(root, btrfs_node_blockptr(upper, slot + 1),
1416                                 root->leafsize);
1417         free_space = btrfs_leaf_free_space(root, right);
1418         if (free_space < data_size + sizeof(struct btrfs_item)) {
1419                 free_extent_buffer(right);
1420                 return 1;
1421         }
1422
1423         /* cow and double check */
1424         ret = btrfs_cow_block(trans, root, right, upper,
1425                               slot + 1, &right);
1426         if (ret) {
1427                 free_extent_buffer(right);
1428                 return 1;
1429         }
1430         free_space = btrfs_leaf_free_space(root, right);
1431         if (free_space < data_size + sizeof(struct btrfs_item)) {
1432                 free_extent_buffer(right);
1433                 return 1;
1434         }
1435
1436         left_nritems = btrfs_header_nritems(left);
1437         if (left_nritems == 0) {
1438                 free_extent_buffer(right);
1439                 return 1;
1440         }
1441
1442         for (i = left_nritems - 1; i >= 1; i--) {
1443                 item = btrfs_item_nr(left, i);
1444
1445                 if (path->slots[0] == i)
1446                         push_space += data_size + sizeof(*item);
1447
1448                 if (!left->map_token) {
1449                         map_extent_buffer(left, (unsigned long)item,
1450                                         sizeof(struct btrfs_item),
1451                                         &left->map_token, &left->kaddr,
1452                                         &left->map_start, &left->map_len,
1453                                         KM_USER1);
1454                 }
1455
1456                 this_item_size = btrfs_item_size(left, item);
1457                 if (this_item_size + sizeof(*item) + push_space > free_space)
1458                         break;
1459                 push_items++;
1460                 push_space += this_item_size + sizeof(*item);
1461         }
1462         if (left->map_token) {
1463                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1464                 left->map_token = NULL;
1465         }
1466
1467         if (push_items == 0) {
1468                 free_extent_buffer(right);
1469                 return 1;
1470         }
1471
1472         if (push_items == left_nritems)
1473                 WARN_ON(1);
1474
1475         /* push left to right */
1476         right_nritems = btrfs_header_nritems(right);
1477         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1478         push_space -= leaf_data_end(root, left);
1479
1480         /* make room in the right data area */
1481         data_end = leaf_data_end(root, right);
1482         memmove_extent_buffer(right,
1483                               btrfs_leaf_data(right) + data_end - push_space,
1484                               btrfs_leaf_data(right) + data_end,
1485                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1486
1487         /* copy from the left data area */
1488         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1489                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1490                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1491                      push_space);
1492
1493         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1494                               btrfs_item_nr_offset(0),
1495                               right_nritems * sizeof(struct btrfs_item));
1496
1497         /* copy the items from left to right */
1498         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1499                    btrfs_item_nr_offset(left_nritems - push_items),
1500                    push_items * sizeof(struct btrfs_item));
1501
1502         /* update the item pointers */
1503         right_nritems += push_items;
1504         btrfs_set_header_nritems(right, right_nritems);
1505         push_space = BTRFS_LEAF_DATA_SIZE(root);
1506
1507         for (i = 0; i < right_nritems; i++) {
1508                 item = btrfs_item_nr(right, i);
1509                 if (!right->map_token) {
1510                         map_extent_buffer(right, (unsigned long)item,
1511                                         sizeof(struct btrfs_item),
1512                                         &right->map_token, &right->kaddr,
1513                                         &right->map_start, &right->map_len,
1514                                         KM_USER1);
1515                 }
1516                 push_space -= btrfs_item_size(right, item);
1517                 btrfs_set_item_offset(right, item, push_space);
1518         }
1519
1520         if (right->map_token) {
1521                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1522                 right->map_token = NULL;
1523         }
1524         left_nritems -= push_items;
1525         btrfs_set_header_nritems(left, left_nritems);
1526
1527         btrfs_mark_buffer_dirty(left);
1528         btrfs_mark_buffer_dirty(right);
1529
1530         btrfs_item_key(right, &disk_key, 0);
1531         btrfs_set_node_key(upper, &disk_key, slot + 1);
1532         btrfs_mark_buffer_dirty(upper);
1533
1534         /* then fixup the leaf pointer in the path */
1535         if (path->slots[0] >= left_nritems) {
1536                 path->slots[0] -= left_nritems;
1537                 free_extent_buffer(path->nodes[0]);
1538                 path->nodes[0] = right;
1539                 path->slots[1] += 1;
1540         } else {
1541                 free_extent_buffer(right);
1542         }
1543         return 0;
1544 }
1545 /*
1546  * push some data in the path leaf to the left, trying to free up at
1547  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1548  */
1549 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1550                           *root, struct btrfs_path *path, int data_size)
1551 {
1552         struct btrfs_disk_key disk_key;
1553         struct extent_buffer *right = path->nodes[0];
1554         struct extent_buffer *left;
1555         int slot;
1556         int i;
1557         int free_space;
1558         int push_space = 0;
1559         int push_items = 0;
1560         struct btrfs_item *item;
1561         u32 old_left_nritems;
1562         u32 right_nritems;
1563         int ret = 0;
1564         int wret;
1565         u32 this_item_size;
1566         u32 old_left_item_size;
1567
1568         slot = path->slots[1];
1569         if (slot == 0)
1570                 return 1;
1571         if (!path->nodes[1])
1572                 return 1;
1573
1574         right_nritems = btrfs_header_nritems(right);
1575         if (right_nritems == 0) {
1576                 return 1;
1577         }
1578
1579         left = read_tree_block(root, btrfs_node_blockptr(path->nodes[1],
1580                                slot - 1), root->leafsize);
1581         free_space = btrfs_leaf_free_space(root, left);
1582         if (free_space < data_size + sizeof(struct btrfs_item)) {
1583                 free_extent_buffer(left);
1584                 return 1;
1585         }
1586
1587         /* cow and double check */
1588         ret = btrfs_cow_block(trans, root, left,
1589                               path->nodes[1], slot - 1, &left);
1590         if (ret) {
1591                 /* we hit -ENOSPC, but it isn't fatal here */
1592                 free_extent_buffer(left);
1593                 return 1;
1594         }
1595
1596         free_space = btrfs_leaf_free_space(root, left);
1597         if (free_space < data_size + sizeof(struct btrfs_item)) {
1598                 free_extent_buffer(left);
1599                 return 1;
1600         }
1601
1602         for (i = 0; i < right_nritems - 1; i++) {
1603                 item = btrfs_item_nr(right, i);
1604                 if (!right->map_token) {
1605                         map_extent_buffer(right, (unsigned long)item,
1606                                         sizeof(struct btrfs_item),
1607                                         &right->map_token, &right->kaddr,
1608                                         &right->map_start, &right->map_len,
1609                                         KM_USER1);
1610                 }
1611
1612                 if (path->slots[0] == i)
1613                         push_space += data_size + sizeof(*item);
1614
1615                 this_item_size = btrfs_item_size(right, item);
1616                 if (this_item_size + sizeof(*item) + push_space > free_space)
1617                         break;
1618
1619                 push_items++;
1620                 push_space += this_item_size + sizeof(*item);
1621         }
1622
1623         if (right->map_token) {
1624                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1625                 right->map_token = NULL;
1626         }
1627
1628         if (push_items == 0) {
1629                 free_extent_buffer(left);
1630                 return 1;
1631         }
1632         if (push_items == btrfs_header_nritems(right))
1633                 WARN_ON(1);
1634
1635         /* push data from right to left */
1636         copy_extent_buffer(left, right,
1637                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1638                            btrfs_item_nr_offset(0),
1639                            push_items * sizeof(struct btrfs_item));
1640
1641         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1642                      btrfs_item_offset_nr(right, push_items -1);
1643
1644         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1645                      leaf_data_end(root, left) - push_space,
1646                      btrfs_leaf_data(right) +
1647                      btrfs_item_offset_nr(right, push_items - 1),
1648                      push_space);
1649         old_left_nritems = btrfs_header_nritems(left);
1650         BUG_ON(old_left_nritems < 0);
1651
1652         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1653         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1654                 u32 ioff;
1655
1656                 item = btrfs_item_nr(left, i);
1657                 if (!left->map_token) {
1658                         map_extent_buffer(left, (unsigned long)item,
1659                                         sizeof(struct btrfs_item),
1660                                         &left->map_token, &left->kaddr,
1661                                         &left->map_start, &left->map_len,
1662                                         KM_USER1);
1663                 }
1664
1665                 ioff = btrfs_item_offset(left, item);
1666                 btrfs_set_item_offset(left, item,
1667                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1668         }
1669         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1670         if (left->map_token) {
1671                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1672                 left->map_token = NULL;
1673         }
1674
1675         /* fixup right node */
1676         push_space = btrfs_item_offset_nr(right, push_items - 1) -
1677                                           leaf_data_end(root, right);
1678         memmove_extent_buffer(right, btrfs_leaf_data(right) +
1679                               BTRFS_LEAF_DATA_SIZE(root) - push_space,
1680                               btrfs_leaf_data(right) +
1681                               leaf_data_end(root, right), push_space);
1682
1683         memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1684                               btrfs_item_nr_offset(push_items),
1685                              (btrfs_header_nritems(right) - push_items) *
1686                              sizeof(struct btrfs_item));
1687
1688         right_nritems = btrfs_header_nritems(right) - push_items;
1689         btrfs_set_header_nritems(right, right_nritems);
1690         push_space = BTRFS_LEAF_DATA_SIZE(root);
1691
1692         for (i = 0; i < right_nritems; i++) {
1693                 item = btrfs_item_nr(right, i);
1694
1695                 if (!right->map_token) {
1696                         map_extent_buffer(right, (unsigned long)item,
1697                                         sizeof(struct btrfs_item),
1698                                         &right->map_token, &right->kaddr,
1699                                         &right->map_start, &right->map_len,
1700                                         KM_USER1);
1701                 }
1702
1703                 push_space = push_space - btrfs_item_size(right, item);
1704                 btrfs_set_item_offset(right, item, push_space);
1705         }
1706         if (right->map_token) {
1707                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1708                 right->map_token = NULL;
1709         }
1710
1711         btrfs_mark_buffer_dirty(left);
1712         btrfs_mark_buffer_dirty(right);
1713
1714         btrfs_item_key(right, &disk_key, 0);
1715         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1716         if (wret)
1717                 ret = wret;
1718
1719         /* then fixup the leaf pointer in the path */
1720         if (path->slots[0] < push_items) {
1721                 path->slots[0] += old_left_nritems;
1722                 free_extent_buffer(path->nodes[0]);
1723                 path->nodes[0] = left;
1724                 path->slots[1] -= 1;
1725         } else {
1726                 free_extent_buffer(left);
1727                 path->slots[0] -= push_items;
1728         }
1729         BUG_ON(path->slots[0] < 0);
1730         return ret;
1731 }
1732
1733 /*
1734  * split the path's leaf in two, making sure there is at least data_size
1735  * available for the resulting leaf level of the path.
1736  *
1737  * returns 0 if all went well and < 0 on failure.
1738  */
1739 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
1740                       *root, struct btrfs_key *ins_key,
1741                       struct btrfs_path *path, int data_size, int extend)
1742 {
1743         struct extent_buffer *l;
1744         u32 nritems;
1745         int mid;
1746         int slot;
1747         struct extent_buffer *right;
1748         int space_needed = data_size + sizeof(struct btrfs_item);
1749         int data_copy_size;
1750         int rt_data_off;
1751         int i;
1752         int ret = 0;
1753         int wret;
1754         int double_split;
1755         int num_doubles = 0;
1756         struct btrfs_disk_key disk_key;
1757
1758         if (extend)
1759                 space_needed = data_size;
1760
1761         /* first try to make some room by pushing left and right */
1762         if (ins_key->type != BTRFS_DIR_ITEM_KEY) {
1763                 wret = push_leaf_right(trans, root, path, data_size);
1764                 if (wret < 0) {
1765                         return wret;
1766                 }
1767                 if (wret) {
1768                         wret = push_leaf_left(trans, root, path, data_size);
1769                         if (wret < 0)
1770                                 return wret;
1771                 }
1772                 l = path->nodes[0];
1773
1774                 /* did the pushes work? */
1775                 if (btrfs_leaf_free_space(root, l) >= space_needed)
1776                         return 0;
1777         }
1778
1779         if (!path->nodes[1]) {
1780                 ret = insert_new_root(trans, root, path, 1);
1781                 if (ret)
1782                         return ret;
1783         }
1784 again:
1785         double_split = 0;
1786         l = path->nodes[0];
1787         slot = path->slots[0];
1788         nritems = btrfs_header_nritems(l);
1789         mid = (nritems + 1)/ 2;
1790
1791         right = btrfs_alloc_free_block(trans, root, root->leafsize,
1792                                        l->start, 0);
1793         if (IS_ERR(right))
1794                 return PTR_ERR(right);
1795
1796         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
1797         btrfs_set_header_bytenr(right, right->start);
1798         btrfs_set_header_generation(right, trans->transid);
1799         btrfs_set_header_owner(right, root->root_key.objectid);
1800         btrfs_set_header_level(right, 0);
1801         write_extent_buffer(right, root->fs_info->fsid,
1802                             (unsigned long)btrfs_header_fsid(right),
1803                             BTRFS_FSID_SIZE);
1804         if (mid <= slot) {
1805                 if (nritems == 1 ||
1806                     leaf_space_used(l, mid, nritems - mid) + space_needed >
1807                         BTRFS_LEAF_DATA_SIZE(root)) {
1808                         if (slot >= nritems) {
1809                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1810                                 btrfs_set_header_nritems(right, 0);
1811                                 wret = insert_ptr(trans, root, path,
1812                                                   &disk_key, right->start,
1813                                                   path->slots[1] + 1, 1);
1814                                 if (wret)
1815                                         ret = wret;
1816                                 free_extent_buffer(path->nodes[0]);
1817                                 path->nodes[0] = right;
1818                                 path->slots[0] = 0;
1819                                 path->slots[1] += 1;
1820                                 return ret;
1821                         }
1822                         mid = slot;
1823                         if (mid != nritems &&
1824                             leaf_space_used(l, mid, nritems - mid) +
1825                             space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
1826                                 double_split = 1;
1827                         }
1828                 }
1829         } else {
1830                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
1831                         BTRFS_LEAF_DATA_SIZE(root)) {
1832                         if (!extend && slot == 0) {
1833                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1834                                 btrfs_set_header_nritems(right, 0);
1835                                 wret = insert_ptr(trans, root, path,
1836                                                   &disk_key,
1837                                                   right->start,
1838                                                   path->slots[1], 1);
1839                                 if (wret)
1840                                         ret = wret;
1841                                 free_extent_buffer(path->nodes[0]);
1842                                 path->nodes[0] = right;
1843                                 path->slots[0] = 0;
1844                                 if (path->slots[1] == 0) {
1845                                         wret = fixup_low_keys(trans, root,
1846                                                    path, &disk_key, 1);
1847                                         if (wret)
1848                                                 ret = wret;
1849                                 }
1850                                 return ret;
1851                         } else if (extend && slot == 0) {
1852                                 mid = 1;
1853                         } else {
1854                                 mid = slot;
1855                                 if (mid != nritems &&
1856                                     leaf_space_used(l, mid, nritems - mid) +
1857                                     space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
1858                                         double_split = 1;
1859                                 }
1860                         }
1861                 }
1862         }
1863         nritems = nritems - mid;
1864         btrfs_set_header_nritems(right, nritems);
1865         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
1866
1867         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
1868                            btrfs_item_nr_offset(mid),
1869                            nritems * sizeof(struct btrfs_item));
1870
1871         copy_extent_buffer(right, l,
1872                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
1873                      data_copy_size, btrfs_leaf_data(l) +
1874                      leaf_data_end(root, l), data_copy_size);
1875
1876         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
1877                       btrfs_item_end_nr(l, mid);
1878
1879         for (i = 0; i < nritems; i++) {
1880                 struct btrfs_item *item = btrfs_item_nr(right, i);
1881                 u32 ioff;
1882
1883                 if (!right->map_token) {
1884                         map_extent_buffer(right, (unsigned long)item,
1885                                         sizeof(struct btrfs_item),
1886                                         &right->map_token, &right->kaddr,
1887                                         &right->map_start, &right->map_len,
1888                                         KM_USER1);
1889                 }
1890
1891                 ioff = btrfs_item_offset(right, item);
1892                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
1893         }
1894
1895         if (right->map_token) {
1896                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1897                 right->map_token = NULL;
1898         }
1899
1900         btrfs_set_header_nritems(l, mid);
1901         ret = 0;
1902         btrfs_item_key(right, &disk_key, 0);
1903         wret = insert_ptr(trans, root, path, &disk_key, right->start,
1904                           path->slots[1] + 1, 1);
1905         if (wret)
1906                 ret = wret;
1907
1908         btrfs_mark_buffer_dirty(right);
1909         btrfs_mark_buffer_dirty(l);
1910         BUG_ON(path->slots[0] != slot);
1911
1912         if (mid <= slot) {
1913                 free_extent_buffer(path->nodes[0]);
1914                 path->nodes[0] = right;
1915                 path->slots[0] -= mid;
1916                 path->slots[1] += 1;
1917         } else
1918                 free_extent_buffer(right);
1919
1920         BUG_ON(path->slots[0] < 0);
1921
1922         if (double_split) {
1923                 BUG_ON(num_doubles != 0);
1924                 num_doubles++;
1925                 goto again;
1926         }
1927         return ret;
1928 }
1929
1930 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
1931                         struct btrfs_root *root,
1932                         struct btrfs_path *path,
1933                         u32 new_size)
1934 {
1935         int ret = 0;
1936         int slot;
1937         int slot_orig;
1938         struct extent_buffer *leaf;
1939         struct btrfs_item *item;
1940         u32 nritems;
1941         unsigned int data_end;
1942         unsigned int old_data_start;
1943         unsigned int old_size;
1944         unsigned int size_diff;
1945         int i;
1946
1947         slot_orig = path->slots[0];
1948         leaf = path->nodes[0];
1949
1950         nritems = btrfs_header_nritems(leaf);
1951         data_end = leaf_data_end(root, leaf);
1952
1953         slot = path->slots[0];
1954         old_data_start = btrfs_item_offset_nr(leaf, slot);
1955         old_size = btrfs_item_size_nr(leaf, slot); BUG_ON(old_size <= new_size);
1956         size_diff = old_size - new_size;
1957
1958         BUG_ON(slot < 0);
1959         BUG_ON(slot >= nritems);
1960
1961         /*
1962          * item0..itemN ... dataN.offset..dataN.size .. data0.size
1963          */
1964         /* first correct the data pointers */
1965         for (i = slot; i < nritems; i++) {
1966                 u32 ioff;
1967                 item = btrfs_item_nr(leaf, i);
1968
1969                 if (!leaf->map_token) {
1970                         map_extent_buffer(leaf, (unsigned long)item,
1971                                         sizeof(struct btrfs_item),
1972                                         &leaf->map_token, &leaf->kaddr,
1973                                         &leaf->map_start, &leaf->map_len,
1974                                         KM_USER1);
1975                 }
1976
1977                 ioff = btrfs_item_offset(leaf, item);
1978                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
1979         }
1980
1981         if (leaf->map_token) {
1982                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
1983                 leaf->map_token = NULL;
1984         }
1985
1986         /* shift the data */
1987         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
1988                       data_end + size_diff, btrfs_leaf_data(leaf) +
1989                       data_end, old_data_start + new_size - data_end);
1990
1991         item = btrfs_item_nr(leaf, slot);
1992         btrfs_set_item_size(leaf, item, new_size);
1993         btrfs_mark_buffer_dirty(leaf);
1994
1995         ret = 0;
1996         if (btrfs_leaf_free_space(root, leaf) < 0) {
1997                 btrfs_print_leaf(root, leaf);
1998                 BUG();
1999         }
2000         return ret;
2001 }
2002
2003 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2004                       struct btrfs_root *root, struct btrfs_path *path,
2005                       u32 data_size)
2006 {
2007         int ret = 0;
2008         int slot;
2009         int slot_orig;
2010         struct extent_buffer *leaf;
2011         struct btrfs_item *item;
2012         u32 nritems;
2013         unsigned int data_end;
2014         unsigned int old_data;
2015         unsigned int old_size;
2016         int i;
2017
2018         slot_orig = path->slots[0];
2019         leaf = path->nodes[0];
2020
2021         nritems = btrfs_header_nritems(leaf);
2022         data_end = leaf_data_end(root, leaf);
2023
2024         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2025                 btrfs_print_leaf(root, leaf);
2026                 BUG();
2027         }
2028         slot = path->slots[0];
2029         old_data = btrfs_item_end_nr(leaf, slot);
2030
2031         BUG_ON(slot < 0);
2032         if (slot >= nritems) {
2033                 btrfs_print_leaf(root, leaf);
2034                 printk("slot %d too large, nritems %d\n", slot, nritems);
2035                 BUG_ON(1);
2036         }
2037
2038         /*
2039          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2040          */
2041         /* first correct the data pointers */
2042         for (i = slot; i < nritems; i++) {
2043                 u32 ioff;
2044                 item = btrfs_item_nr(leaf, i);
2045
2046                 if (!leaf->map_token) {
2047                         map_extent_buffer(leaf, (unsigned long)item,
2048                                         sizeof(struct btrfs_item),
2049                                         &leaf->map_token, &leaf->kaddr,
2050                                         &leaf->map_start, &leaf->map_len,
2051                                         KM_USER1);
2052                 }
2053                 ioff = btrfs_item_offset(leaf, item);
2054                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2055         }
2056
2057         if (leaf->map_token) {
2058                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2059                 leaf->map_token = NULL;
2060         }
2061
2062         /* shift the data */
2063         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2064                       data_end - data_size, btrfs_leaf_data(leaf) +
2065                       data_end, old_data - data_end);
2066
2067         data_end = old_data;
2068         old_size = btrfs_item_size_nr(leaf, slot);
2069         item = btrfs_item_nr(leaf, slot);
2070         btrfs_set_item_size(leaf, item, old_size + data_size);
2071         btrfs_mark_buffer_dirty(leaf);
2072
2073         ret = 0;
2074         if (btrfs_leaf_free_space(root, leaf) < 0) {
2075                 btrfs_print_leaf(root, leaf);
2076                 BUG();
2077         }
2078         return ret;
2079 }
2080
2081 /*
2082  * Given a key and some data, insert an item into the tree.
2083  * This does all the path init required, making room in the tree if needed.
2084  */
2085 int btrfs_insert_empty_item(struct btrfs_trans_handle *trans,
2086                             struct btrfs_root *root,
2087                             struct btrfs_path *path,
2088                             struct btrfs_key *cpu_key, u32 data_size)
2089 {
2090         struct extent_buffer *leaf;
2091         struct btrfs_item *item;
2092         int ret = 0;
2093         int slot;
2094         int slot_orig;
2095         u32 nritems;
2096         unsigned int data_end;
2097         struct btrfs_disk_key disk_key;
2098
2099         btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2100
2101         /* create a root if there isn't one */
2102         if (!root->node)
2103                 BUG();
2104
2105         ret = btrfs_search_slot(trans, root, cpu_key, path, data_size, 1);
2106         if (ret == 0) {
2107                 return -EEXIST;
2108         }
2109         if (ret < 0)
2110                 goto out;
2111
2112         slot_orig = path->slots[0];
2113         leaf = path->nodes[0];
2114
2115         nritems = btrfs_header_nritems(leaf);
2116         data_end = leaf_data_end(root, leaf);
2117
2118         if (btrfs_leaf_free_space(root, leaf) <
2119             sizeof(struct btrfs_item) + data_size) {
2120                 btrfs_print_leaf(root, leaf);
2121                 printk("not enough freespace need %u have %d\n",
2122                        data_size, btrfs_leaf_free_space(root, leaf));
2123                 BUG();
2124         }
2125
2126         slot = path->slots[0];
2127         BUG_ON(slot < 0);
2128
2129         if (slot != nritems) {
2130                 int i;
2131                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2132
2133                 if (old_data < data_end) {
2134                         btrfs_print_leaf(root, leaf);
2135                         printk("slot %d old_data %d data_end %d\n",
2136                                slot, old_data, data_end);
2137                         BUG_ON(1);
2138                 }
2139                 /*
2140                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2141                  */
2142                 /* first correct the data pointers */
2143                 WARN_ON(leaf->map_token);
2144                 for (i = slot; i < nritems; i++) {
2145                         u32 ioff;
2146
2147                         item = btrfs_item_nr(leaf, i);
2148                         if (!leaf->map_token) {
2149                                 map_extent_buffer(leaf, (unsigned long)item,
2150                                         sizeof(struct btrfs_item),
2151                                         &leaf->map_token, &leaf->kaddr,
2152                                         &leaf->map_start, &leaf->map_len,
2153                                         KM_USER1);
2154                         }
2155
2156                         ioff = btrfs_item_offset(leaf, item);
2157                         btrfs_set_item_offset(leaf, item, ioff - data_size);
2158                 }
2159                 if (leaf->map_token) {
2160                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2161                         leaf->map_token = NULL;
2162                 }
2163
2164                 /* shift the items */
2165                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
2166                               btrfs_item_nr_offset(slot),
2167                               (nritems - slot) * sizeof(struct btrfs_item));
2168
2169                 /* shift the data */
2170                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2171                               data_end - data_size, btrfs_leaf_data(leaf) +
2172                               data_end, old_data - data_end);
2173                 data_end = old_data;
2174         }
2175
2176         /* setup the item for the new data */
2177         btrfs_set_item_key(leaf, &disk_key, slot);
2178         item = btrfs_item_nr(leaf, slot);
2179         btrfs_set_item_offset(leaf, item, data_end - data_size);
2180         btrfs_set_item_size(leaf, item, data_size);
2181         btrfs_set_header_nritems(leaf, nritems + 1);
2182         btrfs_mark_buffer_dirty(leaf);
2183
2184         ret = 0;
2185         if (slot == 0)
2186                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2187
2188         if (btrfs_leaf_free_space(root, leaf) < 0) {
2189                 btrfs_print_leaf(root, leaf);
2190                 BUG();
2191         }
2192 out:
2193         return ret;
2194 }
2195
2196 /*
2197  * Given a key and some data, insert an item into the tree.
2198  * This does all the path init required, making room in the tree if needed.
2199  */
2200 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2201                       *root, struct btrfs_key *cpu_key, void *data, u32
2202                       data_size)
2203 {
2204         int ret = 0;
2205         struct btrfs_path *path;
2206         struct extent_buffer *leaf;
2207         unsigned long ptr;
2208
2209         path = btrfs_alloc_path();
2210         BUG_ON(!path);
2211         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2212         if (!ret) {
2213                 leaf = path->nodes[0];
2214                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2215                 write_extent_buffer(leaf, data, ptr, data_size);
2216                 btrfs_mark_buffer_dirty(leaf);
2217         }
2218         btrfs_free_path(path);
2219         return ret;
2220 }
2221
2222 /*
2223  * delete the pointer from a given node.
2224  *
2225  * If the delete empties a node, the node is removed from the tree,
2226  * continuing all the way the root if required.  The root is converted into
2227  * a leaf if all the nodes are emptied.
2228  */
2229 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2230                    struct btrfs_path *path, int level, int slot)
2231 {
2232         struct extent_buffer *parent = path->nodes[level];
2233         u32 nritems;
2234         int ret = 0;
2235         int wret;
2236
2237         nritems = btrfs_header_nritems(parent);
2238         if (slot != nritems -1) {
2239                 memmove_extent_buffer(parent,
2240                               btrfs_node_key_ptr_offset(slot),
2241                               btrfs_node_key_ptr_offset(slot + 1),
2242                               sizeof(struct btrfs_key_ptr) *
2243                               (nritems - slot - 1));
2244         }
2245         nritems--;
2246         btrfs_set_header_nritems(parent, nritems);
2247         if (nritems == 0 && parent == root->node) {
2248                 BUG_ON(btrfs_header_level(root->node) != 1);
2249                 /* just turn the root into a leaf and break */
2250                 btrfs_set_header_level(root->node, 0);
2251         } else if (slot == 0) {
2252                 struct btrfs_disk_key disk_key;
2253
2254                 btrfs_node_key(parent, &disk_key, 0);
2255                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2256                 if (wret)
2257                         ret = wret;
2258         }
2259         btrfs_mark_buffer_dirty(parent);
2260         return ret;
2261 }
2262
2263 /*
2264  * delete the item at the leaf level in path.  If that empties
2265  * the leaf, remove it from the tree
2266  */
2267 int btrfs_del_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2268                    struct btrfs_path *path)
2269 {
2270         int slot;
2271         struct extent_buffer *leaf;
2272         struct btrfs_item *item;
2273         int doff;
2274         int dsize;
2275         int ret = 0;
2276         int wret;
2277         u32 nritems;
2278
2279         leaf = path->nodes[0];
2280         slot = path->slots[0];
2281         doff = btrfs_item_offset_nr(leaf, slot);
2282         dsize = btrfs_item_size_nr(leaf, slot);
2283         nritems = btrfs_header_nritems(leaf);
2284
2285         if (slot != nritems - 1) {
2286                 int i;
2287                 int data_end = leaf_data_end(root, leaf);
2288
2289                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2290                               data_end + dsize,
2291                               btrfs_leaf_data(leaf) + data_end,
2292                               doff - data_end);
2293
2294                 for (i = slot + 1; i < nritems; i++) {
2295                         u32 ioff;
2296
2297                         item = btrfs_item_nr(leaf, i);
2298                         if (!leaf->map_token) {
2299                                 map_extent_buffer(leaf, (unsigned long)item,
2300                                         sizeof(struct btrfs_item),
2301                                         &leaf->map_token, &leaf->kaddr,
2302                                         &leaf->map_start, &leaf->map_len,
2303                                         KM_USER1);
2304                         }
2305                         ioff = btrfs_item_offset(leaf, item);
2306                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2307                 }
2308
2309                 if (leaf->map_token) {
2310                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2311                         leaf->map_token = NULL;
2312                 }
2313
2314                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2315                               btrfs_item_nr_offset(slot + 1),
2316                               sizeof(struct btrfs_item) *
2317                               (nritems - slot - 1));
2318         }
2319         btrfs_set_header_nritems(leaf, nritems - 1);
2320         nritems--;
2321
2322         /* delete the leaf if we've emptied it */
2323         if (nritems == 0) {
2324                 if (leaf == root->node) {
2325                         btrfs_set_header_level(leaf, 0);
2326                 } else {
2327                         clean_tree_block(trans, root, leaf);
2328                         wait_on_tree_block_writeback(root, leaf);
2329                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
2330                         if (wret)
2331                                 ret = wret;
2332                         wret = btrfs_free_extent(trans, root,
2333                                                  leaf->start, leaf->len, 1);
2334                         if (wret)
2335                                 ret = wret;
2336                 }
2337         } else {
2338                 int used = leaf_space_used(leaf, 0, nritems);
2339                 if (slot == 0) {
2340                         struct btrfs_disk_key disk_key;
2341
2342                         btrfs_item_key(leaf, &disk_key, 0);
2343                         wret = fixup_low_keys(trans, root, path,
2344                                               &disk_key, 1);
2345                         if (wret)
2346                                 ret = wret;
2347                 }
2348
2349                 /* delete the leaf if it is mostly empty */
2350                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 3) {
2351                         /* push_leaf_left fixes the path.
2352                          * make sure the path still points to our leaf
2353                          * for possible call to del_ptr below
2354                          */
2355                         slot = path->slots[1];
2356                         extent_buffer_get(leaf);
2357
2358                         wret = push_leaf_right(trans, root, path, 1);
2359                         if (wret < 0 && wret != -ENOSPC)
2360                                 ret = wret;
2361
2362                         if (path->nodes[0] == leaf &&
2363                             btrfs_header_nritems(leaf)) {
2364                                 wret = push_leaf_left(trans, root, path, 1);
2365                                 if (wret < 0 && wret != -ENOSPC)
2366                                         ret = wret;
2367                         }
2368
2369                         if (btrfs_header_nritems(leaf) == 0) {
2370                                 u64 bytenr = leaf->start;
2371                                 u32 blocksize = leaf->len;
2372
2373                                 clean_tree_block(trans, root, leaf);
2374                                 wait_on_tree_block_writeback(root, leaf);
2375
2376                                 wret = del_ptr(trans, root, path, 1, slot);
2377                                 if (wret)
2378                                         ret = wret;
2379
2380                                 free_extent_buffer(leaf);
2381                                 wret = btrfs_free_extent(trans, root, bytenr,
2382                                                          blocksize, 1);
2383                                 if (wret)
2384                                         ret = wret;
2385                         } else {
2386                                 btrfs_mark_buffer_dirty(leaf);
2387                                 free_extent_buffer(leaf);
2388                         }
2389                 } else {
2390                         btrfs_mark_buffer_dirty(leaf);
2391                 }
2392         }
2393         return ret;
2394 }
2395
2396 /*
2397  * walk up the tree as far as required to find the next leaf.
2398  * returns 0 if it found something or 1 if there are no greater leaves.
2399  * returns < 0 on io errors.
2400  */
2401 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2402 {
2403         int slot;
2404         int level = 1;
2405         u64 bytenr;
2406         struct extent_buffer *c;
2407         struct extent_buffer *next = NULL;
2408
2409         while(level < BTRFS_MAX_LEVEL) {
2410                 if (!path->nodes[level])
2411                         return 1;
2412
2413                 slot = path->slots[level] + 1;
2414                 c = path->nodes[level];
2415                 if (slot >= btrfs_header_nritems(c)) {
2416                         level++;
2417                         continue;
2418                 }
2419
2420                 bytenr = btrfs_node_blockptr(c, slot);
2421                 if (next)
2422                         free_extent_buffer(next);
2423
2424                 if (path->reada)
2425                         reada_for_search(root, path, level, slot);
2426
2427                 next = read_tree_block(root, bytenr,
2428                                        btrfs_level_size(root, level -1));
2429                 break;
2430         }
2431         path->slots[level] = slot;
2432         while(1) {
2433                 level--;
2434                 c = path->nodes[level];
2435                 free_extent_buffer(c);
2436                 path->nodes[level] = next;
2437                 path->slots[level] = 0;
2438                 if (!level)
2439                         break;
2440                 if (path->reada)
2441                         reada_for_search(root, path, level, 0);
2442                 next = read_tree_block(root, btrfs_node_blockptr(next, 0),
2443                                        btrfs_level_size(root, level - 1));
2444         }
2445         return 0;
2446 }