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