fs/btrfs/ctree.c

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