fs/btrfs/ctree.c

   1 #include <linux/module.h>
   2 #include "ctree.h"
   3 #include "disk-io.h"
   4 #include "transaction.h"
   5
   6 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
   7                       *root, struct btrfs_path *path, int level);
   8 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
   9                       *root, struct btrfs_path *path, int data_size);
  10 static int push_node_left(struct btrfs_trans_handle *trans, struct btrfs_root
  11                           *root, struct buffer_head *dst, struct buffer_head
  12                           *src);
  13 static int balance_node_right(struct btrfs_trans_handle *trans, struct
  14                               btrfs_root *root, struct buffer_head *dst_buf,
  15                               struct buffer_head *src_buf);
  16 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  17                    struct btrfs_path *path, int level, int slot);
  18
  19 inline void btrfs_init_path(struct btrfs_path *p)
  20 {
  21         memset(p, 0, sizeof(*p));
  22 }
  23
  24 struct btrfs_path *btrfs_alloc_path(void)
  25 {
  26         struct btrfs_path *path;
  27         path = kmem_cache_alloc(btrfs_path_cachep, GFP_NOFS);
  28         if (path)
  29                 btrfs_init_path(path);
  30         return path;
  31 }
  32
  33 void btrfs_free_path(struct btrfs_path *p)
  34 {
  35         btrfs_release_path(NULL, p);
  36         kmem_cache_free(btrfs_path_cachep, p);
  37 }
  38
  39 void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
  40 {
  41         int i;
  42         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  43                 if (!p->nodes[i])
  44                         break;
  45                 btrfs_block_release(root, p->nodes[i]);
  46         }
  47         memset(p, 0, sizeof(*p));
  48 }
  49
  50 static int btrfs_cow_block(struct btrfs_trans_handle *trans, struct btrfs_root
  51                            *root, struct buffer_head *buf, struct buffer_head
  52                            *parent, int parent_slot, struct buffer_head
  53                            **cow_ret)
  54 {
  55         struct buffer_head *cow;
  56         struct btrfs_node *cow_node;
  57
  58         if (btrfs_header_generation(btrfs_buffer_header(buf)) ==
  59                                     trans->transid) {
  60                 *cow_ret = buf;
  61                 return 0;
  62         }
  63         cow = btrfs_alloc_free_block(trans, root);
  64         cow_node = btrfs_buffer_node(cow);
  65         if (buf->b_size != root->blocksize || cow->b_size != root->blocksize)
  66                 WARN_ON(1);
  67         memcpy(cow_node, btrfs_buffer_node(buf), root->blocksize);
  68         btrfs_set_header_blocknr(&cow_node->header, cow->b_blocknr);
  69         btrfs_set_header_generation(&cow_node->header, trans->transid);
  70         btrfs_inc_ref(trans, root, buf);
  71         if (buf == root->node) {
  72                 root->node = cow;
  73                 get_bh(cow);
  74                 if (buf != root->commit_root) {
  75                         btrfs_free_extent(trans, root, buf->b_blocknr, 1, 1);
  76                 }
  77                 btrfs_block_release(root, buf);
  78         } else {
  79                 btrfs_set_node_blockptr(btrfs_buffer_node(parent), parent_slot,
  80                                         cow->b_blocknr);
  81                 btrfs_mark_buffer_dirty(parent);
  82                 btrfs_free_extent(trans, root, buf->b_blocknr, 1, 1);
  83         }
  84         btrfs_block_release(root, buf);
  85         mark_buffer_dirty(cow);
  86         *cow_ret = cow;
  87         return 0;
  88 }
  89
  90 /*
  91  * The leaf data grows from end-to-front in the node.
  92  * this returns the address of the start of the last item,
  93  * which is the stop of the leaf data stack
  94  */
  95 static inline unsigned int leaf_data_end(struct btrfs_root *root,
  96                                          struct btrfs_leaf *leaf)
  97 {
  98         u32 nr = btrfs_header_nritems(&leaf->header);
  99         if (nr == 0)
 100                 return BTRFS_LEAF_DATA_SIZE(root);
 101         return btrfs_item_offset(leaf->items + nr - 1);
 102 }
 103
 104 /*
 105  * The space between the end of the leaf items and
 106  * the start of the leaf data.  IOW, how much room
 107  * the leaf has left for both items and data
 108  */
 109 int btrfs_leaf_free_space(struct btrfs_root *root, struct btrfs_leaf *leaf)
 110 {
 111         int data_end = leaf_data_end(root, leaf);
 112         int nritems = btrfs_header_nritems(&leaf->header);
 113         char *items_end = (char *)(leaf->items + nritems + 1);
 114         return (char *)(btrfs_leaf_data(leaf) + data_end) - (char *)items_end;
 115 }
 116
 117 /*
 118  * compare two keys in a memcmp fashion
 119  */
 120 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 121 {
 122         struct btrfs_key k1;
 123
 124         btrfs_disk_key_to_cpu(&k1, disk);
 125
 126         if (k1.objectid > k2->objectid)
 127                 return 1;
 128         if (k1.objectid < k2->objectid)
 129                 return -1;
 130         if (k1.offset > k2->offset)
 131                 return 1;
 132         if (k1.offset < k2->offset)
 133                 return -1;
 134         if (k1.flags > k2->flags)
 135                 return 1;
 136         if (k1.flags < k2->flags)
 137                 return -1;
 138         return 0;
 139 }
 140
 141 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
 142                       int level)
 143 {
 144         int i;
 145         struct btrfs_node *parent = NULL;
 146         struct btrfs_node *node = btrfs_buffer_node(path->nodes[level]);
 147         int parent_slot;
 148         u32 nritems = btrfs_header_nritems(&node->header);
 149
 150         if (path->nodes[level + 1])
 151                 parent = btrfs_buffer_node(path->nodes[level + 1]);
 152         parent_slot = path->slots[level + 1];
 153         BUG_ON(nritems == 0);
 154         if (parent) {
 155                 struct btrfs_disk_key *parent_key;
 156                 parent_key = &parent->ptrs[parent_slot].key;
 157                 BUG_ON(memcmp(parent_key, &node->ptrs[0].key,
 158                               sizeof(struct btrfs_disk_key)));
 159                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 160                        btrfs_header_blocknr(&node->header));
 161         }
 162         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 163         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 164                 struct btrfs_key cpukey;
 165                 btrfs_disk_key_to_cpu(&cpukey, &node->ptrs[i + 1].key);
 166                 BUG_ON(comp_keys(&node->ptrs[i].key, &cpukey) >= 0);
 167         }
 168         return 0;
 169 }
 170
 171 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 172                       int level)
 173 {
 174         int i;
 175         struct btrfs_leaf *leaf = btrfs_buffer_leaf(path->nodes[level]);
 176         struct btrfs_node *parent = NULL;
 177         int parent_slot;
 178         u32 nritems = btrfs_header_nritems(&leaf->header);
 179
 180         if (path->nodes[level + 1])
 181                 parent = btrfs_buffer_node(path->nodes[level + 1]);
 182         parent_slot = path->slots[level + 1];
 183         BUG_ON(btrfs_leaf_free_space(root, leaf) < 0);
 184
 185         if (nritems == 0)
 186                 return 0;
 187
 188         if (parent) {
 189                 struct btrfs_disk_key *parent_key;
 190                 parent_key = &parent->ptrs[parent_slot].key;
 191                 BUG_ON(memcmp(parent_key, &leaf->items[0].key,
 192                        sizeof(struct btrfs_disk_key)));
 193                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 194                        btrfs_header_blocknr(&leaf->header));
 195         }
 196         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 197                 struct btrfs_key cpukey;
 198                 btrfs_disk_key_to_cpu(&cpukey, &leaf->items[i + 1].key);
 199                 BUG_ON(comp_keys(&leaf->items[i].key,
 200                                  &cpukey) >= 0);
 201                 BUG_ON(btrfs_item_offset(leaf->items + i) !=
 202                         btrfs_item_end(leaf->items + i + 1));
 203                 if (i == 0) {
 204                         BUG_ON(btrfs_item_offset(leaf->items + i) +
 205                                btrfs_item_size(leaf->items + i) !=
 206                                BTRFS_LEAF_DATA_SIZE(root));
 207                 }
 208         }
 209         return 0;
 210 }
 211
 212 static int check_block(struct btrfs_root *root, struct btrfs_path *path,
 213                         int level)
 214 {
 215         if (level == 0)
 216                 return check_leaf(root, path, level);
 217         return check_node(root, path, level);
 218 }
 219
 220 /*
 221  * search for key in the array p.  items p are item_size apart
 222  * and there are 'max' items in p
 223  * the slot in the array is returned via slot, and it points to
 224  * the place where you would insert key if it is not found in
 225  * the array.
 226  *
 227  * slot may point to max if the key is bigger than all of the keys
 228  */
 229 static int generic_bin_search(char *p, int item_size, struct btrfs_key *key,
 230                        int max, int *slot)
 231 {
 232         int low = 0;
 233         int high = max;
 234         int mid;
 235         int ret;
 236         struct btrfs_disk_key *tmp;
 237
 238         while(low < high) {
 239                 mid = (low + high) / 2;
 240                 tmp = (struct btrfs_disk_key *)(p + mid * item_size);
 241                 ret = comp_keys(tmp, key);
 242
 243                 if (ret < 0)
 244                         low = mid + 1;
 245                 else if (ret > 0)
 246                         high = mid;
 247                 else {
 248                         *slot = mid;
 249                         return 0;
 250                 }
 251         }
 252         *slot = low;
 253         return 1;
 254 }
 255
 256 /*
 257  * simple bin_search frontend that does the right thing for
 258  * leaves vs nodes
 259  */
 260 static int bin_search(struct btrfs_node *c, struct btrfs_key *key, int *slot)
 261 {
 262         if (btrfs_is_leaf(c)) {
 263                 struct btrfs_leaf *l = (struct btrfs_leaf *)c;
 264                 return generic_bin_search((void *)l->items,
 265                                           sizeof(struct btrfs_item),
 266                                           key, btrfs_header_nritems(&c->header),
 267                                           slot);
 268         } else {
 269                 return generic_bin_search((void *)c->ptrs,
 270                                           sizeof(struct btrfs_key_ptr),
 271                                           key, btrfs_header_nritems(&c->header),
 272                                           slot);
 273         }
 274         return -1;
 275 }
 276
 277 static struct buffer_head *read_node_slot(struct btrfs_root *root,
 278                                    struct buffer_head *parent_buf,
 279                                    int slot)
 280 {
 281         struct btrfs_node *node = btrfs_buffer_node(parent_buf);
 282         if (slot < 0)
 283                 return NULL;
 284         if (slot >= btrfs_header_nritems(&node->header))
 285                 return NULL;
 286         return read_tree_block(root, btrfs_node_blockptr(node, slot));
 287 }
 288
 289 static int balance_level(struct btrfs_trans_handle *trans, struct btrfs_root
 290                          *root, struct btrfs_path *path, int level)
 291 {
 292         struct buffer_head *right_buf;
 293         struct buffer_head *mid_buf;
 294         struct buffer_head *left_buf;
 295         struct buffer_head *parent_buf = NULL;
 296         struct btrfs_node *right = NULL;
 297         struct btrfs_node *mid;
 298         struct btrfs_node *left = NULL;
 299         struct btrfs_node *parent = NULL;
 300         int ret = 0;
 301         int wret;
 302         int pslot;
 303         int orig_slot = path->slots[level];
 304         u64 orig_ptr;
 305
 306         if (level == 0)
 307                 return 0;
 308
 309         mid_buf = path->nodes[level];
 310         mid = btrfs_buffer_node(mid_buf);
 311         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 312
 313         if (level < BTRFS_MAX_LEVEL - 1)
 314                 parent_buf = path->nodes[level + 1];
 315         pslot = path->slots[level + 1];
 316
 317         /*
 318          * deal with the case where there is only one pointer in the root
 319          * by promoting the node below to a root
 320          */
 321         if (!parent_buf) {
 322                 struct buffer_head *child;
 323                 u64 blocknr = mid_buf->b_blocknr;
 324
 325                 if (btrfs_header_nritems(&mid->header) != 1)
 326                         return 0;
 327
 328                 /* promote the child to a root */
 329                 child = read_node_slot(root, mid_buf, 0);
 330                 BUG_ON(!child);
 331                 root->node = child;
 332                 path->nodes[level] = NULL;
 333                 clean_tree_block(trans, root, mid_buf);
 334                 wait_on_buffer(mid_buf);
 335                 /* once for the path */
 336                 btrfs_block_release(root, mid_buf);
 337                 /* once for the root ptr */
 338                 btrfs_block_release(root, mid_buf);
 339                 return btrfs_free_extent(trans, root, blocknr, 1, 1);
 340         }
 341         parent = btrfs_buffer_node(parent_buf);
 342
 343         if (btrfs_header_nritems(&mid->header) >
 344             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 345                 return 0;
 346
 347         left_buf = read_node_slot(root, parent_buf, pslot - 1);
 348         right_buf = read_node_slot(root, parent_buf, pslot + 1);
 349
 350         /* first, try to make some room in the middle buffer */
 351         if (left_buf) {
 352                 btrfs_cow_block(trans, root, left_buf, parent_buf, pslot - 1,
 353                                 &left_buf);
 354                 left = btrfs_buffer_node(left_buf);
 355                 orig_slot += btrfs_header_nritems(&left->header);
 356                 wret = push_node_left(trans, root, left_buf, mid_buf);
 357                 if (wret < 0)
 358                         ret = wret;
 359         }
 360
 361         /*
 362          * then try to empty the right most buffer into the middle
 363          */
 364         if (right_buf) {
 365                 btrfs_cow_block(trans, root, right_buf, parent_buf, pslot + 1,
 366                                 &right_buf);
 367                 right = btrfs_buffer_node(right_buf);
 368                 wret = push_node_left(trans, root, mid_buf, right_buf);
 369                 if (wret < 0)
 370                         ret = wret;
 371                 if (btrfs_header_nritems(&right->header) == 0) {
 372                         u64 blocknr = right_buf->b_blocknr;
 373                         clean_tree_block(trans, root, right_buf);
 374                         wait_on_buffer(right_buf);
 375                         btrfs_block_release(root, right_buf);
 376                         right_buf = NULL;
 377                         right = NULL;
 378                         wret = del_ptr(trans, root, path, level + 1, pslot +
 379                                        1);
 380                         if (wret)
 381                                 ret = wret;
 382                         wret = btrfs_free_extent(trans, root, blocknr, 1, 1);
 383                         if (wret)
 384                                 ret = wret;
 385                 } else {
 386                         btrfs_memcpy(root, parent,
 387                                      &parent->ptrs[pslot + 1].key,
 388                                      &right->ptrs[0].key,
 389                                      sizeof(struct btrfs_disk_key));
 390                         btrfs_mark_buffer_dirty(parent_buf);
 391                 }
 392         }
 393         if (btrfs_header_nritems(&mid->header) == 1) {
 394                 /*
 395                  * we're not allowed to leave a node with one item in the
 396                  * tree during a delete.  A deletion from lower in the tree
 397                  * could try to delete the only pointer in this node.
 398                  * So, pull some keys from the left.
 399                  * There has to be a left pointer at this point because
 400                  * otherwise we would have pulled some pointers from the
 401                  * right
 402                  */
 403                 BUG_ON(!left_buf);
 404                 wret = balance_node_right(trans, root, mid_buf, left_buf);
 405                 if (wret < 0)
 406                         ret = wret;
 407                 BUG_ON(wret == 1);
 408         }
 409         if (btrfs_header_nritems(&mid->header) == 0) {
 410                 /* we've managed to empty the middle node, drop it */
 411                 u64 blocknr = mid_buf->b_blocknr;
 412                 clean_tree_block(trans, root, mid_buf);
 413                 wait_on_buffer(mid_buf);
 414                 btrfs_block_release(root, mid_buf);
 415                 mid_buf = NULL;
 416                 mid = NULL;
 417                 wret = del_ptr(trans, root, path, level + 1, pslot);
 418                 if (wret)
 419                         ret = wret;
 420                 wret = btrfs_free_extent(trans, root, blocknr, 1, 1);
 421                 if (wret)
 422                         ret = wret;
 423         } else {
 424                 /* update the parent key to reflect our changes */
 425                 btrfs_memcpy(root, parent,
 426                              &parent->ptrs[pslot].key, &mid->ptrs[0].key,
 427                              sizeof(struct btrfs_disk_key));
 428                 btrfs_mark_buffer_dirty(parent_buf);
 429         }
 430
 431         /* update the path */
 432         if (left_buf) {
 433                 if (btrfs_header_nritems(&left->header) > orig_slot) {
 434                         get_bh(left_buf);
 435                         path->nodes[level] = left_buf;
 436                         path->slots[level + 1] -= 1;
 437                         path->slots[level] = orig_slot;
 438                         if (mid_buf)
 439                                 btrfs_block_release(root, mid_buf);
 440                 } else {
 441                         orig_slot -= btrfs_header_nritems(&left->header);
 442                         path->slots[level] = orig_slot;
 443                 }
 444         }
 445         /* double check we haven't messed things up */
 446         check_block(root, path, level);
 447         if (orig_ptr !=
 448             btrfs_node_blockptr(btrfs_buffer_node(path->nodes[level]),
 449                                 path->slots[level]))
 450                 BUG();
 451
 452         if (right_buf)
 453                 btrfs_block_release(root, right_buf);
 454         if (left_buf)
 455                 btrfs_block_release(root, left_buf);
 456         return ret;
 457 }
 458
 459 /*
 460  * look for key in the tree.  path is filled in with nodes along the way
 461  * if key is found, we return zero and you can find the item in the leaf
 462  * level of the path (level 0)
 463  *
 464  * If the key isn't found, the path points to the slot where it should
 465  * be inserted, and 1 is returned.  If there are other errors during the
 466  * search a negative error number is returned.
 467  *
 468  * if ins_len > 0, nodes and leaves will be split as we walk down the
 469  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
 470  * possible)
 471  */
 472 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
 473                       *root, struct btrfs_key *key, struct btrfs_path *p, int
 474                       ins_len, int cow)
 475 {
 476         struct buffer_head *b;
 477         struct buffer_head *cow_buf;
 478         struct btrfs_node *c;
 479         int slot;
 480         int ret;
 481         int level;
 482
 483         WARN_ON(p->nodes[0] != NULL);
 484         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
 485 again:
 486         b = root->node;
 487         get_bh(b);
 488         while (b) {
 489                 c = btrfs_buffer_node(b);
 490                 level = btrfs_header_level(&c->header);
 491                 if (cow) {
 492                         int wret;
 493                         wret = btrfs_cow_block(trans, root, b,
 494                                                p->nodes[level + 1],
 495                                                p->slots[level + 1],
 496                                                &cow_buf);
 497                         b = cow_buf;
 498                         c = btrfs_buffer_node(b);
 499                 }
 500                 BUG_ON(!cow && ins_len);
 501                 if (level != btrfs_header_level(&c->header))
 502                         WARN_ON(1);
 503                 level = btrfs_header_level(&c->header);
 504                 p->nodes[level] = b;
 505                 ret = check_block(root, p, level);
 506                 if (ret)
 507                         return -1;
 508                 ret = bin_search(c, key, &slot);
 509                 if (!btrfs_is_leaf(c)) {
 510                         if (ret && slot > 0)
 511                                 slot -= 1;
 512                         p->slots[level] = slot;
 513                         if (ins_len > 0 && btrfs_header_nritems(&c->header) ==
 514                             BTRFS_NODEPTRS_PER_BLOCK(root)) {
 515                                 int sret = split_node(trans, root, p, level);
 516                                 BUG_ON(sret > 0);
 517                                 if (sret)
 518                                         return sret;
 519                                 b = p->nodes[level];
 520                                 c = btrfs_buffer_node(b);
 521                                 slot = p->slots[level];
 522                         } else if (ins_len < 0) {
 523                                 int sret = balance_level(trans, root, p,
 524                                                          level);
 525                                 if (sret)
 526                                         return sret;
 527                                 b = p->nodes[level];
 528                                 if (!b)
 529                                         goto again;
 530                                 c = btrfs_buffer_node(b);
 531                                 slot = p->slots[level];
 532                                 BUG_ON(btrfs_header_nritems(&c->header) == 1);
 533                         }
 534                         b = read_tree_block(root, btrfs_node_blockptr(c, slot));
 535                 } else {
 536                         struct btrfs_leaf *l = (struct btrfs_leaf *)c;
 537                         p->slots[level] = slot;
 538                         if (ins_len > 0 && btrfs_leaf_free_space(root, l) <
 539                             sizeof(struct btrfs_item) + ins_len) {
 540                                 int sret = split_leaf(trans, root, p, ins_len);
 541                                 BUG_ON(sret > 0);
 542                                 if (sret)
 543                                         return sret;
 544                         }
 545                         return ret;
 546                 }
 547         }
 548         return 1;
 549 }
 550
 551 /*
 552  * adjust the pointers going up the tree, starting at level
 553  * making sure the right key of each node is points to 'key'.
 554  * This is used after shifting pointers to the left, so it stops
 555  * fixing up pointers when a given leaf/node is not in slot 0 of the
 556  * higher levels
 557  *
 558  * If this fails to write a tree block, it returns -1, but continues
 559  * fixing up the blocks in ram so the tree is consistent.
 560  */
 561 static int fixup_low_keys(struct btrfs_trans_handle *trans, struct btrfs_root
 562                           *root, struct btrfs_path *path, struct btrfs_disk_key
 563                           *key, int level)
 564 {
 565         int i;
 566         int ret = 0;
 567         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
 568                 struct btrfs_node *t;
 569                 int tslot = path->slots[i];
 570                 if (!path->nodes[i])
 571                         break;
 572                 t = btrfs_buffer_node(path->nodes[i]);
 573                 btrfs_memcpy(root, t, &t->ptrs[tslot].key, key, sizeof(*key));
 574                 btrfs_mark_buffer_dirty(path->nodes[i]);
 575                 if (tslot != 0)
 576                         break;
 577         }
 578         return ret;
 579 }
 580
 581 /*
 582  * try to push data from one node into the next node left in the
 583  * tree.
 584  *
 585  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
 586  * error, and > 0 if there was no room in the left hand block.
 587  */
 588 static int push_node_left(struct btrfs_trans_handle *trans, struct btrfs_root
 589                           *root, struct buffer_head *dst_buf, struct
 590                           buffer_head *src_buf)
 591 {
 592         struct btrfs_node *src = btrfs_buffer_node(src_buf);
 593         struct btrfs_node *dst = btrfs_buffer_node(dst_buf);
 594         int push_items = 0;
 595         int src_nritems;
 596         int dst_nritems;
 597         int ret = 0;
 598
 599         src_nritems = btrfs_header_nritems(&src->header);
 600         dst_nritems = btrfs_header_nritems(&dst->header);
 601         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
 602         if (push_items <= 0) {
 603                 return 1;
 604         }
 605
 606         if (src_nritems < push_items)
 607                 push_items = src_nritems;
 608
 609         btrfs_memcpy(root, dst, dst->ptrs + dst_nritems, src->ptrs,
 610                      push_items * sizeof(struct btrfs_key_ptr));
 611         if (push_items < src_nritems) {
 612                 btrfs_memmove(root, src, src->ptrs, src->ptrs + push_items,
 613                         (src_nritems - push_items) *
 614                         sizeof(struct btrfs_key_ptr));
 615         }
 616         btrfs_set_header_nritems(&src->header, src_nritems - push_items);
 617         btrfs_set_header_nritems(&dst->header, dst_nritems + push_items);
 618         btrfs_mark_buffer_dirty(src_buf);
 619         btrfs_mark_buffer_dirty(dst_buf);
 620         return ret;
 621 }
 622
 623 /*
 624  * try to push data from one node into the next node right in the
 625  * tree.
 626  *
 627  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
 628  * error, and > 0 if there was no room in the right hand block.
 629  *
 630  * this will  only push up to 1/2 the contents of the left node over
 631  */
 632 static int balance_node_right(struct btrfs_trans_handle *trans, struct
 633                               btrfs_root *root, struct buffer_head *dst_buf,
 634                               struct buffer_head *src_buf)
 635 {
 636         struct btrfs_node *src = btrfs_buffer_node(src_buf);
 637         struct btrfs_node *dst = btrfs_buffer_node(dst_buf);
 638         int push_items = 0;
 639         int max_push;
 640         int src_nritems;
 641         int dst_nritems;
 642         int ret = 0;
 643
 644         src_nritems = btrfs_header_nritems(&src->header);
 645         dst_nritems = btrfs_header_nritems(&dst->header);
 646         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
 647         if (push_items <= 0) {
 648                 return 1;
 649         }
 650
 651         max_push = src_nritems / 2 + 1;
 652         /* don't try to empty the node */
 653         if (max_push > src_nritems)
 654                 return 1;
 655         if (max_push < push_items)
 656                 push_items = max_push;
 657
 658         btrfs_memmove(root, dst, dst->ptrs + push_items, dst->ptrs,
 659                       dst_nritems * sizeof(struct btrfs_key_ptr));
 660
 661         btrfs_memcpy(root, dst, dst->ptrs,
 662                      src->ptrs + src_nritems - push_items,
 663                      push_items * sizeof(struct btrfs_key_ptr));
 664
 665         btrfs_set_header_nritems(&src->header, src_nritems - push_items);
 666         btrfs_set_header_nritems(&dst->header, dst_nritems + push_items);
 667
 668         btrfs_mark_buffer_dirty(src_buf);
 669         btrfs_mark_buffer_dirty(dst_buf);
 670         return ret;
 671 }
 672
 673 /*
 674  * helper function to insert a new root level in the tree.
 675  * A new node is allocated, and a single item is inserted to
 676  * point to the existing root
 677  *
 678  * returns zero on success or < 0 on failure.
 679  */
 680 static int insert_new_root(struct btrfs_trans_handle *trans, struct btrfs_root
 681                            *root, struct btrfs_path *path, int level)
 682 {
 683         struct buffer_head *t;
 684         struct btrfs_node *lower;
 685         struct btrfs_node *c;
 686         struct btrfs_disk_key *lower_key;
 687
 688         BUG_ON(path->nodes[level]);
 689         BUG_ON(path->nodes[level-1] != root->node);
 690
 691         t = btrfs_alloc_free_block(trans, root);
 692         c = btrfs_buffer_node(t);
 693         memset(c, 0, root->blocksize);
 694         btrfs_set_header_nritems(&c->header, 1);
 695         btrfs_set_header_level(&c->header, level);
 696         btrfs_set_header_blocknr(&c->header, t->b_blocknr);
 697         btrfs_set_header_generation(&c->header, trans->transid);
 698         btrfs_set_header_parentid(&c->header,
 699               btrfs_header_parentid(btrfs_buffer_header(root->node)));
 700         lower = btrfs_buffer_node(path->nodes[level-1]);
 701         if (btrfs_is_leaf(lower))
 702                 lower_key = &((struct btrfs_leaf *)lower)->items[0].key;
 703         else
 704                 lower_key = &lower->ptrs[0].key;
 705         btrfs_memcpy(root, c, &c->ptrs[0].key, lower_key,
 706                      sizeof(struct btrfs_disk_key));
 707         btrfs_set_node_blockptr(c, 0, path->nodes[level - 1]->b_blocknr);
 708
 709         btrfs_mark_buffer_dirty(t);
 710
 711         /* the super has an extra ref to root->node */
 712         btrfs_block_release(root, root->node);
 713         root->node = t;
 714         get_bh(t);
 715         path->nodes[level] = t;
 716         path->slots[level] = 0;
 717         return 0;
 718 }
 719
 720 /*
 721  * worker function to insert a single pointer in a node.
 722  * the node should have enough room for the pointer already
 723  *
 724  * slot and level indicate where you want the key to go, and
 725  * blocknr is the block the key points to.
 726  *
 727  * returns zero on success and < 0 on any error
 728  */
 729 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
 730                       *root, struct btrfs_path *path, struct btrfs_disk_key
 731                       *key, u64 blocknr, int slot, int level)
 732 {
 733         struct btrfs_node *lower;
 734         int nritems;
 735
 736         BUG_ON(!path->nodes[level]);
 737         lower = btrfs_buffer_node(path->nodes[level]);
 738         nritems = btrfs_header_nritems(&lower->header);
 739         if (slot > nritems)
 740                 BUG();
 741         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
 742                 BUG();
 743         if (slot != nritems) {
 744                 btrfs_memmove(root, lower, lower->ptrs + slot + 1,
 745                               lower->ptrs + slot,
 746                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
 747         }
 748         btrfs_memcpy(root, lower, &lower->ptrs[slot].key,
 749                      key, sizeof(struct btrfs_disk_key));
 750         btrfs_set_node_blockptr(lower, slot, blocknr);
 751         btrfs_set_header_nritems(&lower->header, nritems + 1);
 752         btrfs_mark_buffer_dirty(path->nodes[level]);
 753         return 0;
 754 }
 755
 756 /*
 757  * split the node at the specified level in path in two.
 758  * The path is corrected to point to the appropriate node after the split
 759  *
 760  * Before splitting this tries to make some room in the node by pushing
 761  * left and right, if either one works, it returns right away.
 762  *
 763  * returns 0 on success and < 0 on failure
 764  */
 765 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
 766                       *root, struct btrfs_path *path, int level)
 767 {
 768         struct buffer_head *t;
 769         struct btrfs_node *c;
 770         struct buffer_head *split_buffer;
 771         struct btrfs_node *split;
 772         int mid;
 773         int ret;
 774         int wret;
 775         u32 c_nritems;
 776
 777         t = path->nodes[level];
 778         c = btrfs_buffer_node(t);
 779         if (t == root->node) {
 780                 /* trying to split the root, lets make a new one */
 781                 ret = insert_new_root(trans, root, path, level + 1);
 782                 if (ret)
 783                         return ret;
 784         }
 785         c_nritems = btrfs_header_nritems(&c->header);
 786         split_buffer = btrfs_alloc_free_block(trans, root);
 787         split = btrfs_buffer_node(split_buffer);
 788         btrfs_set_header_flags(&split->header, btrfs_header_flags(&c->header));
 789         btrfs_set_header_level(&split->header, btrfs_header_level(&c->header));
 790         btrfs_set_header_blocknr(&split->header, split_buffer->b_blocknr);
 791         btrfs_set_header_generation(&split->header, trans->transid);
 792         btrfs_set_header_parentid(&split->header,
 793               btrfs_header_parentid(btrfs_buffer_header(root->node)));
 794         mid = (c_nritems + 1) / 2;
 795         btrfs_memcpy(root, split, split->ptrs, c->ptrs + mid,
 796                      (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
 797         btrfs_set_header_nritems(&split->header, c_nritems - mid);
 798         btrfs_set_header_nritems(&c->header, mid);
 799         ret = 0;
 800
 801         btrfs_mark_buffer_dirty(t);
 802         btrfs_mark_buffer_dirty(split_buffer);
 803         wret = insert_ptr(trans, root, path, &split->ptrs[0].key,
 804                           split_buffer->b_blocknr, path->slots[level + 1] + 1,
 805                           level + 1);
 806         if (wret)
 807                 ret = wret;
 808
 809         if (path->slots[level] >= mid) {
 810                 path->slots[level] -= mid;
 811                 btrfs_block_release(root, t);
 812                 path->nodes[level] = split_buffer;
 813                 path->slots[level + 1] += 1;
 814         } else {
 815                 btrfs_block_release(root, split_buffer);
 816         }
 817         return ret;
 818 }
 819
 820 /*
 821  * how many bytes are required to store the items in a leaf.  start
 822  * and nr indicate which items in the leaf to check.  This totals up the
 823  * space used both by the item structs and the item data
 824  */
 825 static int leaf_space_used(struct btrfs_leaf *l, int start, int nr)
 826 {
 827         int data_len;
 828         int end = start + nr - 1;
 829
 830         if (!nr)
 831                 return 0;
 832         data_len = btrfs_item_end(l->items + start);
 833         data_len = data_len - btrfs_item_offset(l->items + end);
 834         data_len += sizeof(struct btrfs_item) * nr;
 835         return data_len;
 836 }
 837
 838 /*
 839  * push some data in the path leaf to the right, trying to free up at
 840  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
 841  *
 842  * returns 1 if the push failed because the other node didn't have enough
 843  * room, 0 if everything worked out and < 0 if there were major errors.
 844  */
 845 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
 846                            *root, struct btrfs_path *path, int data_size)
 847 {
 848         struct buffer_head *left_buf = path->nodes[0];
 849         struct btrfs_leaf *left = btrfs_buffer_leaf(left_buf);
 850         struct btrfs_leaf *right;
 851         struct buffer_head *right_buf;
 852         struct buffer_head *upper;
 853         struct btrfs_node *upper_node;
 854         int slot;
 855         int i;
 856         int free_space;
 857         int push_space = 0;
 858         int push_items = 0;
 859         struct btrfs_item *item;
 860         u32 left_nritems;
 861         u32 right_nritems;
 862
 863         slot = path->slots[1];
 864         if (!path->nodes[1]) {
 865                 return 1;
 866         }
 867         upper = path->nodes[1];
 868         upper_node = btrfs_buffer_node(upper);
 869         if (slot >= btrfs_header_nritems(&upper_node->header) - 1) {
 870                 return 1;
 871         }
 872         right_buf = read_tree_block(root,
 873                     btrfs_node_blockptr(btrfs_buffer_node(upper), slot + 1));
 874         right = btrfs_buffer_leaf(right_buf);
 875         free_space = btrfs_leaf_free_space(root, right);
 876         if (free_space < data_size + sizeof(struct btrfs_item)) {
 877                 btrfs_block_release(root, right_buf);
 878                 return 1;
 879         }
 880         /* cow and double check */
 881         btrfs_cow_block(trans, root, right_buf, upper, slot + 1, &right_buf);
 882         right = btrfs_buffer_leaf(right_buf);
 883         free_space = btrfs_leaf_free_space(root, right);
 884         if (free_space < data_size + sizeof(struct btrfs_item)) {
 885                 btrfs_block_release(root, right_buf);
 886                 return 1;
 887         }
 888
 889         left_nritems = btrfs_header_nritems(&left->header);
 890         for (i = left_nritems - 1; i >= 0; i--) {
 891                 item = left->items + i;
 892                 if (path->slots[0] == i)
 893                         push_space += data_size + sizeof(*item);
 894                 if (btrfs_item_size(item) + sizeof(*item) + push_space >
 895                     free_space)
 896                         break;
 897                 push_items++;
 898                 push_space += btrfs_item_size(item) + sizeof(*item);
 899         }
 900         if (push_items == 0) {
 901                 btrfs_block_release(root, right_buf);
 902                 return 1;
 903         }
 904         right_nritems = btrfs_header_nritems(&right->header);
 905         /* push left to right */
 906         push_space = btrfs_item_end(left->items + left_nritems - push_items);
 907         push_space -= leaf_data_end(root, left);
 908         /* make room in the right data area */
 909         btrfs_memmove(root, right, btrfs_leaf_data(right) +
 910                       leaf_data_end(root, right) - push_space,
 911                       btrfs_leaf_data(right) +
 912                       leaf_data_end(root, right), BTRFS_LEAF_DATA_SIZE(root) -
 913                       leaf_data_end(root, right));
 914         /* copy from the left data area */
 915         btrfs_memcpy(root, right, btrfs_leaf_data(right) +
 916                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
 917                      btrfs_leaf_data(left) + leaf_data_end(root, left),
 918                      push_space);
 919         btrfs_memmove(root, right, right->items + push_items, right->items,
 920                 right_nritems * sizeof(struct btrfs_item));
 921         /* copy the items from left to right */
 922         btrfs_memcpy(root, right, right->items, left->items +
 923                      left_nritems - push_items,
 924                      push_items * sizeof(struct btrfs_item));
 925
 926         /* update the item pointers */
 927         right_nritems += push_items;
 928         btrfs_set_header_nritems(&right->header, right_nritems);
 929         push_space = BTRFS_LEAF_DATA_SIZE(root);
 930         for (i = 0; i < right_nritems; i++) {
 931                 btrfs_set_item_offset(right->items + i, push_space -
 932                                       btrfs_item_size(right->items + i));
 933                 push_space = btrfs_item_offset(right->items + i);
 934         }
 935         left_nritems -= push_items;
 936         btrfs_set_header_nritems(&left->header, left_nritems);
 937
 938         btrfs_mark_buffer_dirty(left_buf);
 939         btrfs_mark_buffer_dirty(right_buf);
 940         btrfs_memcpy(root, upper_node, &upper_node->ptrs[slot + 1].key,
 941                 &right->items[0].key, sizeof(struct btrfs_disk_key));
 942         btrfs_mark_buffer_dirty(upper);
 943
 944         /* then fixup the leaf pointer in the path */
 945         if (path->slots[0] >= left_nritems) {
 946                 path->slots[0] -= left_nritems;
 947                 btrfs_block_release(root, path->nodes[0]);
 948                 path->nodes[0] = right_buf;
 949                 path->slots[1] += 1;
 950         } else {
 951                 btrfs_block_release(root, right_buf);
 952         }
 953         return 0;
 954 }
 955 /*
 956  * push some data in the path leaf to the left, trying to free up at
 957  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
 958  */
 959 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
 960                           *root, struct btrfs_path *path, int data_size)
 961 {
 962         struct buffer_head *right_buf = path->nodes[0];
 963         struct btrfs_leaf *right = btrfs_buffer_leaf(right_buf);
 964         struct buffer_head *t;
 965         struct btrfs_leaf *left;
 966         int slot;
 967         int i;
 968         int free_space;
 969         int push_space = 0;
 970         int push_items = 0;
 971         struct btrfs_item *item;
 972         u32 old_left_nritems;
 973         int ret = 0;
 974         int wret;
 975
 976         slot = path->slots[1];
 977         if (slot == 0) {
 978                 return 1;
 979         }
 980         if (!path->nodes[1]) {
 981                 return 1;
 982         }
 983         t = read_tree_block(root,
 984             btrfs_node_blockptr(btrfs_buffer_node(path->nodes[1]), slot - 1));
 985         left = btrfs_buffer_leaf(t);
 986         free_space = btrfs_leaf_free_space(root, left);
 987         if (free_space < data_size + sizeof(struct btrfs_item)) {
 988                 btrfs_block_release(root, t);
 989                 return 1;
 990         }
 991
 992         /* cow and double check */
 993         btrfs_cow_block(trans, root, t, path->nodes[1], slot - 1, &t);
 994         left = btrfs_buffer_leaf(t);
 995         free_space = btrfs_leaf_free_space(root, left);
 996         if (free_space < data_size + sizeof(struct btrfs_item)) {
 997                 btrfs_block_release(root, t);
 998                 return 1;
 999         }
1000
1001         for (i = 0; i < btrfs_header_nritems(&right->header); i++) {
1002                 item = right->items + i;
1003                 if (path->slots[0] == i)
1004                         push_space += data_size + sizeof(*item);
1005                 if (btrfs_item_size(item) + sizeof(*item) + push_space >
1006                     free_space)
1007                         break;
1008                 push_items++;
1009                 push_space += btrfs_item_size(item) + sizeof(*item);
1010         }
1011         if (push_items == 0) {
1012                 btrfs_block_release(root, t);
1013                 return 1;
1014         }
1015         /* push data from right to left */
1016         btrfs_memcpy(root, left, left->items +
1017                      btrfs_header_nritems(&left->header),
1018                      right->items, push_items * sizeof(struct btrfs_item));
1019         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1020                      btrfs_item_offset(right->items + push_items -1);
1021         btrfs_memcpy(root, left, btrfs_leaf_data(left) +
1022                      leaf_data_end(root, left) - push_space,
1023                      btrfs_leaf_data(right) +
1024                      btrfs_item_offset(right->items + push_items - 1),
1025                      push_space);
1026         old_left_nritems = btrfs_header_nritems(&left->header);
1027         BUG_ON(old_left_nritems < 0);
1028
1029         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1030                 u32 ioff = btrfs_item_offset(left->items + i);
1031                 btrfs_set_item_offset(left->items + i, ioff -
1032                                      (BTRFS_LEAF_DATA_SIZE(root) -
1033                                       btrfs_item_offset(left->items +
1034                                                         old_left_nritems - 1)));
1035         }
1036         btrfs_set_header_nritems(&left->header, old_left_nritems + push_items);
1037
1038         /* fixup right node */
1039         push_space = btrfs_item_offset(right->items + push_items - 1) -
1040                      leaf_data_end(root, right);
1041         btrfs_memmove(root, right, btrfs_leaf_data(right) +
1042                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1043                       btrfs_leaf_data(right) +
1044                       leaf_data_end(root, right), push_space);
1045         btrfs_memmove(root, right, right->items, right->items + push_items,
1046                 (btrfs_header_nritems(&right->header) - push_items) *
1047                 sizeof(struct btrfs_item));
1048         btrfs_set_header_nritems(&right->header,
1049                                  btrfs_header_nritems(&right->header) -
1050                                  push_items);
1051         push_space = BTRFS_LEAF_DATA_SIZE(root);
1052
1053         for (i = 0; i < btrfs_header_nritems(&right->header); i++) {
1054                 btrfs_set_item_offset(right->items + i, push_space -
1055                                       btrfs_item_size(right->items + i));
1056                 push_space = btrfs_item_offset(right->items + i);
1057         }
1058
1059         btrfs_mark_buffer_dirty(t);
1060         btrfs_mark_buffer_dirty(right_buf);
1061
1062         wret = fixup_low_keys(trans, root, path, &right->items[0].key, 1);
1063         if (wret)
1064                 ret = wret;
1065
1066         /* then fixup the leaf pointer in the path */
1067         if (path->slots[0] < push_items) {
1068                 path->slots[0] += old_left_nritems;
1069                 btrfs_block_release(root, path->nodes[0]);
1070                 path->nodes[0] = t;
1071                 path->slots[1] -= 1;
1072         } else {
1073                 btrfs_block_release(root, t);
1074                 path->slots[0] -= push_items;
1075         }
1076         BUG_ON(path->slots[0] < 0);
1077         return ret;
1078 }
1079
1080 /*
1081  * split the path's leaf in two, making sure there is at least data_size
1082  * available for the resulting leaf level of the path.
1083  *
1084  * returns 0 if all went well and < 0 on failure.
1085  */
1086 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
1087                       *root, struct btrfs_path *path, int data_size)
1088 {
1089         struct buffer_head *l_buf;
1090         struct btrfs_leaf *l;
1091         u32 nritems;
1092         int mid;
1093         int slot;
1094         struct btrfs_leaf *right;
1095         struct buffer_head *right_buffer;
1096         int space_needed = data_size + sizeof(struct btrfs_item);
1097         int data_copy_size;
1098         int rt_data_off;
1099         int i;
1100         int ret;
1101         int wret;
1102
1103         /* first try to make some room by pushing left and right */
1104         wret = push_leaf_left(trans, root, path, data_size);
1105         if (wret < 0)
1106                 return wret;
1107         if (wret) {
1108                 wret = push_leaf_right(trans, root, path, data_size);
1109                 if (wret < 0)
1110                         return wret;
1111         }
1112         l_buf = path->nodes[0];
1113         l = btrfs_buffer_leaf(l_buf);
1114
1115         /* did the pushes work? */
1116         if (btrfs_leaf_free_space(root, l) >=
1117             sizeof(struct btrfs_item) + data_size)
1118                 return 0;
1119
1120         if (!path->nodes[1]) {
1121                 ret = insert_new_root(trans, root, path, 1);
1122                 if (ret)
1123                         return ret;
1124         }
1125         slot = path->slots[0];
1126         nritems = btrfs_header_nritems(&l->header);
1127         mid = (nritems + 1)/ 2;
1128         right_buffer = btrfs_alloc_free_block(trans, root);
1129         BUG_ON(!right_buffer);
1130         BUG_ON(mid == nritems);
1131         right = btrfs_buffer_leaf(right_buffer);
1132         memset(&right->header, 0, sizeof(right->header));
1133         if (mid <= slot) {
1134                 /* FIXME, just alloc a new leaf here */
1135                 if (leaf_space_used(l, mid, nritems - mid) + space_needed >
1136                         BTRFS_LEAF_DATA_SIZE(root))
1137                         BUG();
1138         } else {
1139                 /* FIXME, just alloc a new leaf here */
1140                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
1141                         BTRFS_LEAF_DATA_SIZE(root))
1142                         BUG();
1143         }
1144         btrfs_set_header_nritems(&right->header, nritems - mid);
1145         btrfs_set_header_blocknr(&right->header, right_buffer->b_blocknr);
1146         btrfs_set_header_generation(&right->header, trans->transid);
1147         btrfs_set_header_level(&right->header, 0);
1148         btrfs_set_header_parentid(&right->header,
1149               btrfs_header_parentid(btrfs_buffer_header(root->node)));
1150         data_copy_size = btrfs_item_end(l->items + mid) -
1151                          leaf_data_end(root, l);
1152         btrfs_memcpy(root, right, right->items, l->items + mid,
1153                      (nritems - mid) * sizeof(struct btrfs_item));
1154         btrfs_memcpy(root, right,
1155                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
1156                      data_copy_size, btrfs_leaf_data(l) +
1157                      leaf_data_end(root, l), data_copy_size);
1158         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
1159                       btrfs_item_end(l->items + mid);
1160
1161         for (i = 0; i < btrfs_header_nritems(&right->header); i++) {
1162                 u32 ioff = btrfs_item_offset(right->items + i);
1163                 btrfs_set_item_offset(right->items + i, ioff + rt_data_off);
1164         }
1165
1166         btrfs_set_header_nritems(&l->header, mid);
1167         ret = 0;
1168         wret = insert_ptr(trans, root, path, &right->items[0].key,
1169                           right_buffer->b_blocknr, path->slots[1] + 1, 1);
1170         if (wret)
1171                 ret = wret;
1172         btrfs_mark_buffer_dirty(right_buffer);
1173         btrfs_mark_buffer_dirty(l_buf);
1174         BUG_ON(path->slots[0] != slot);
1175         if (mid <= slot) {
1176                 btrfs_block_release(root, path->nodes[0]);
1177                 path->nodes[0] = right_buffer;
1178                 path->slots[0] -= mid;
1179                 path->slots[1] += 1;
1180         } else
1181                 btrfs_block_release(root, right_buffer);
1182         BUG_ON(path->slots[0] < 0);
1183         return ret;
1184 }
1185
1186 /*
1187  * Given a key and some data, insert an item into the tree.
1188  * This does all the path init required, making room in the tree if needed.
1189  */
1190 int btrfs_insert_empty_item(struct btrfs_trans_handle *trans, struct btrfs_root
1191                             *root, struct btrfs_path *path, struct btrfs_key
1192                             *cpu_key, u32 data_size)
1193 {
1194         int ret = 0;
1195         int slot;
1196         int slot_orig;
1197         struct btrfs_leaf *leaf;
1198         struct buffer_head *leaf_buf;
1199         u32 nritems;
1200         unsigned int data_end;
1201         struct btrfs_disk_key disk_key;
1202
1203         btrfs_cpu_key_to_disk(&disk_key, cpu_key);
1204
1205         /* create a root if there isn't one */
1206         if (!root->node)
1207                 BUG();
1208         ret = btrfs_search_slot(trans, root, cpu_key, path, data_size, 1);
1209         if (ret == 0) {
1210                 return -EEXIST;
1211         }
1212         if (ret < 0)
1213                 goto out;
1214
1215         slot_orig = path->slots[0];
1216         leaf_buf = path->nodes[0];
1217         leaf = btrfs_buffer_leaf(leaf_buf);
1218
1219         nritems = btrfs_header_nritems(&leaf->header);
1220         data_end = leaf_data_end(root, leaf);
1221
1222         if (btrfs_leaf_free_space(root, leaf) <
1223             sizeof(struct btrfs_item) + data_size)
1224                 BUG();
1225
1226         slot = path->slots[0];
1227         BUG_ON(slot < 0);
1228         if (slot != nritems) {
1229                 int i;
1230                 unsigned int old_data = btrfs_item_end(leaf->items + slot);
1231
1232                 /*
1233                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
1234                  */
1235                 /* first correct the data pointers */
1236                 for (i = slot; i < nritems; i++) {
1237                         u32 ioff = btrfs_item_offset(leaf->items + i);
1238                         btrfs_set_item_offset(leaf->items + i,
1239                                               ioff - data_size);
1240                 }
1241
1242                 /* shift the items */
1243                 btrfs_memmove(root, leaf, leaf->items + slot + 1,
1244                               leaf->items + slot,
1245                               (nritems - slot) * sizeof(struct btrfs_item));
1246
1247                 /* shift the data */
1248                 btrfs_memmove(root, leaf, btrfs_leaf_data(leaf) +
1249                               data_end - data_size, btrfs_leaf_data(leaf) +
1250                               data_end, old_data - data_end);
1251                 data_end = old_data;
1252         }
1253         /* setup the item for the new data */
1254         btrfs_memcpy(root, leaf, &leaf->items[slot].key, &disk_key,
1255                      sizeof(struct btrfs_disk_key));
1256         btrfs_set_item_offset(leaf->items + slot, data_end - data_size);
1257         btrfs_set_item_size(leaf->items + slot, data_size);
1258         btrfs_set_header_nritems(&leaf->header, nritems + 1);
1259         btrfs_mark_buffer_dirty(leaf_buf);
1260
1261         ret = 0;
1262         if (slot == 0)
1263                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
1264
1265         if (btrfs_leaf_free_space(root, leaf) < 0)
1266                 BUG();
1267         check_leaf(root, path, 0);
1268 out:
1269         return ret;
1270 }
1271
1272 /*
1273  * Given a key and some data, insert an item into the tree.
1274  * This does all the path init required, making room in the tree if needed.
1275  */
1276 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
1277                       *root, struct btrfs_key *cpu_key, void *data, u32
1278                       data_size)
1279 {
1280         int ret = 0;
1281         struct btrfs_path *path;
1282         u8 *ptr;
1283
1284         path = btrfs_alloc_path();
1285         BUG_ON(!path);
1286         btrfs_init_path(path);
1287         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
1288         if (!ret) {
1289                 ptr = btrfs_item_ptr(btrfs_buffer_leaf(path->nodes[0]),
1290                                      path->slots[0], u8);
1291                 btrfs_memcpy(root, path->nodes[0]->b_data,
1292                              ptr, data, data_size);
1293                 btrfs_mark_buffer_dirty(path->nodes[0]);
1294         }
1295         btrfs_release_path(root, path);
1296         btrfs_free_path(path);
1297         return ret;
1298 }
1299
1300 /*
1301  * delete the pointer from a given node.
1302  *
1303  * If the delete empties a node, the node is removed from the tree,
1304  * continuing all the way the root if required.  The root is converted into
1305  * a leaf if all the nodes are emptied.
1306  */
1307 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1308                    struct btrfs_path *path, int level, int slot)
1309 {
1310         struct btrfs_node *node;
1311         struct buffer_head *parent = path->nodes[level];
1312         u32 nritems;
1313         int ret = 0;
1314         int wret;
1315
1316         node = btrfs_buffer_node(parent);
1317         nritems = btrfs_header_nritems(&node->header);
1318         if (slot != nritems -1) {
1319                 btrfs_memmove(root, node, node->ptrs + slot,
1320                               node->ptrs + slot + 1,
1321                               sizeof(struct btrfs_key_ptr) *
1322                               (nritems - slot - 1));
1323         }
1324         nritems--;
1325         btrfs_set_header_nritems(&node->header, nritems);
1326         if (nritems == 0 && parent == root->node) {
1327                 struct btrfs_header *header = btrfs_buffer_header(root->node);
1328                 BUG_ON(btrfs_header_level(header) != 1);
1329                 /* just turn the root into a leaf and break */
1330                 btrfs_set_header_level(header, 0);
1331         } else if (slot == 0) {
1332                 wret = fixup_low_keys(trans, root, path, &node->ptrs[0].key,
1333                                       level + 1);
1334                 if (wret)
1335                         ret = wret;
1336         }
1337         btrfs_mark_buffer_dirty(parent);
1338         return ret;
1339 }
1340
1341 /*
1342  * delete the item at the leaf level in path.  If that empties
1343  * the leaf, remove it from the tree
1344  */
1345 int btrfs_del_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1346                    struct btrfs_path *path)
1347 {
1348         int slot;
1349         struct btrfs_leaf *leaf;
1350         struct buffer_head *leaf_buf;
1351         int doff;
1352         int dsize;
1353         int ret = 0;
1354         int wret;
1355         u32 nritems;
1356
1357         leaf_buf = path->nodes[0];
1358         leaf = btrfs_buffer_leaf(leaf_buf);
1359         slot = path->slots[0];
1360         doff = btrfs_item_offset(leaf->items + slot);
1361         dsize = btrfs_item_size(leaf->items + slot);
1362         nritems = btrfs_header_nritems(&leaf->header);
1363
1364         if (slot != nritems - 1) {
1365                 int i;
1366                 int data_end = leaf_data_end(root, leaf);
1367                 btrfs_memmove(root, leaf, btrfs_leaf_data(leaf) +
1368                               data_end + dsize,
1369                               btrfs_leaf_data(leaf) + data_end,
1370                               doff - data_end);
1371                 for (i = slot + 1; i < nritems; i++) {
1372                         u32 ioff = btrfs_item_offset(leaf->items + i);
1373                         btrfs_set_item_offset(leaf->items + i, ioff + dsize);
1374                 }
1375                 btrfs_memmove(root, leaf, leaf->items + slot,
1376                               leaf->items + slot + 1,
1377                               sizeof(struct btrfs_item) *
1378                               (nritems - slot - 1));
1379         }
1380         btrfs_set_header_nritems(&leaf->header, nritems - 1);
1381         nritems--;
1382         /* delete the leaf if we've emptied it */
1383         if (nritems == 0) {
1384                 if (leaf_buf == root->node) {
1385                         btrfs_set_header_level(&leaf->header, 0);
1386                 } else {
1387                         clean_tree_block(trans, root, leaf_buf);
1388                         wait_on_buffer(leaf_buf);
1389                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
1390                         if (wret)
1391                                 ret = wret;
1392                         wret = btrfs_free_extent(trans, root,
1393                                                  leaf_buf->b_blocknr, 1, 1);
1394                         if (wret)
1395                                 ret = wret;
1396                 }
1397         } else {
1398                 int used = leaf_space_used(leaf, 0, nritems);
1399                 if (slot == 0) {
1400                         wret = fixup_low_keys(trans, root, path,
1401                                               &leaf->items[0].key, 1);
1402                         if (wret)
1403                                 ret = wret;
1404                 }
1405
1406                 /* delete the leaf if it is mostly empty */
1407                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 3) {
1408                         /* push_leaf_left fixes the path.
1409                          * make sure the path still points to our leaf
1410                          * for possible call to del_ptr below
1411                          */
1412                         slot = path->slots[1];
1413                         get_bh(leaf_buf);
1414                         wret = push_leaf_left(trans, root, path, 1);
1415                         if (wret < 0)
1416                                 ret = wret;
1417                         if (path->nodes[0] == leaf_buf &&
1418                             btrfs_header_nritems(&leaf->header)) {
1419                                 wret = push_leaf_right(trans, root, path, 1);
1420                                 if (wret < 0)
1421                                         ret = wret;
1422                         }
1423                         if (btrfs_header_nritems(&leaf->header) == 0) {
1424                                 u64 blocknr = leaf_buf->b_blocknr;
1425                                 clean_tree_block(trans, root, leaf_buf);
1426                                 wait_on_buffer(leaf_buf);
1427                                 wret = del_ptr(trans, root, path, 1, slot);
1428                                 if (wret)
1429                                         ret = wret;
1430                                 btrfs_block_release(root, leaf_buf);
1431                                 wret = btrfs_free_extent(trans, root, blocknr,
1432                                                          1, 1);
1433                                 if (wret)
1434                                         ret = wret;
1435                         } else {
1436                                 btrfs_mark_buffer_dirty(leaf_buf);
1437                                 btrfs_block_release(root, leaf_buf);
1438                         }
1439                 } else {
1440                         btrfs_mark_buffer_dirty(leaf_buf);
1441                 }
1442         }
1443         return ret;
1444 }
1445
1446 /*
1447  * walk up the tree as far as required to find the next leaf.
1448  * returns 0 if it found something or 1 if there are no greater leaves.
1449  * returns < 0 on io errors.
1450  */
1451 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
1452 {
1453         int slot;
1454         int level = 1;
1455         u64 blocknr;
1456         struct buffer_head *c;
1457         struct btrfs_node *c_node;
1458         struct buffer_head *next = NULL;
1459
1460         while(level < BTRFS_MAX_LEVEL) {
1461                 if (!path->nodes[level])
1462                         return 1;
1463                 slot = path->slots[level] + 1;
1464                 c = path->nodes[level];
1465                 c_node = btrfs_buffer_node(c);
1466                 if (slot >= btrfs_header_nritems(&c_node->header)) {
1467                         level++;
1468                         continue;
1469                 }
1470                 blocknr = btrfs_node_blockptr(c_node, slot);
1471                 if (next)
1472                         btrfs_block_release(root, next);
1473                 next = read_tree_block(root, blocknr);
1474                 break;
1475         }
1476         path->slots[level] = slot;
1477         while(1) {
1478                 level--;
1479                 c = path->nodes[level];
1480                 btrfs_block_release(root, c);
1481                 path->nodes[level] = next;
1482                 path->slots[level] = 0;
1483                 if (!level)
1484                         break;
1485                 next = read_tree_block(root,
1486                        btrfs_node_blockptr(btrfs_buffer_node(next), 0));
1487         }
1488         return 0;
1489 }