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_key *ins_key,
  10                       struct btrfs_path *path, int data_size);
  11 static int push_node_left(struct btrfs_trans_handle *trans, struct btrfs_root
  12                           *root, struct buffer_head *dst, struct buffer_head
  13                           *src);
  14 static int balance_node_right(struct btrfs_trans_handle *trans, struct
  15                               btrfs_root *root, struct buffer_head *dst_buf,
  16                               struct buffer_head *src_buf);
  17 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  18                    struct btrfs_path *path, int level, int slot);
  19
  20 inline void btrfs_init_path(struct btrfs_path *p)
  21 {
  22         memset(p, 0, sizeof(*p));
  23 }
  24
  25 struct btrfs_path *btrfs_alloc_path(void)
  26 {
  27         struct btrfs_path *path;
  28         path = kmem_cache_alloc(btrfs_path_cachep, GFP_NOFS);
  29         if (path)
  30                 btrfs_init_path(path);
  31         return path;
  32 }
  33
  34 void btrfs_free_path(struct btrfs_path *p)
  35 {
  36         btrfs_release_path(NULL, p);
  37         kmem_cache_free(btrfs_path_cachep, p);
  38 }
  39
  40 void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
  41 {
  42         int i;
  43         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  44                 if (!p->nodes[i])
  45                         break;
  46                 btrfs_block_release(root, p->nodes[i]);
  47         }
  48         memset(p, 0, sizeof(*p));
  49 }
  50
  51 static int btrfs_cow_block(struct btrfs_trans_handle *trans, struct btrfs_root
  52                            *root, struct buffer_head *buf, struct buffer_head
  53                            *parent, int parent_slot, struct buffer_head
  54                            **cow_ret)
  55 {
  56         struct buffer_head *cow;
  57         struct btrfs_node *cow_node;
  58
  59         if (btrfs_header_generation(btrfs_buffer_header(buf)) ==
  60                                     trans->transid) {
  61                 *cow_ret = buf;
  62                 return 0;
  63         }
  64         cow = btrfs_alloc_free_block(trans, root);
  65         cow_node = btrfs_buffer_node(cow);
  66         if (buf->b_size != root->blocksize || cow->b_size != root->blocksize)
  67                 WARN_ON(1);
  68         memcpy(cow_node, btrfs_buffer_node(buf), root->blocksize);
  69         btrfs_set_header_blocknr(&cow_node->header, bh_blocknr(cow));
  70         btrfs_set_header_generation(&cow_node->header, trans->transid);
  71         btrfs_inc_ref(trans, root, buf);
  72         if (buf == root->node) {
  73                 root->node = cow;
  74                 get_bh(cow);
  75                 if (buf != root->commit_root) {
  76                         btrfs_free_extent(trans, root, bh_blocknr(buf), 1, 1);
  77                 }
  78                 btrfs_block_release(root, buf);
  79         } else {
  80                 btrfs_set_node_blockptr(btrfs_buffer_node(parent), parent_slot,
  81                                         bh_blocknr(cow));
  82                 btrfs_mark_buffer_dirty(parent);
  83                 btrfs_free_extent(trans, root, bh_blocknr(buf), 1, 1);
  84         }
  85         btrfs_block_release(root, buf);
  86         mark_buffer_dirty(cow);
  87         *cow_ret = cow;
  88         return 0;
  89 }
  90
  91 /*
  92  * The leaf data grows from end-to-front in the node.
  93  * this returns the address of the start of the last item,
  94  * which is the stop of the leaf data stack
  95  */
  96 static inline unsigned int leaf_data_end(struct btrfs_root *root,
  97                                          struct btrfs_leaf *leaf)
  98 {
  99         u32 nr = btrfs_header_nritems(&leaf->header);
 100         if (nr == 0)
 101                 return BTRFS_LEAF_DATA_SIZE(root);
 102         return btrfs_item_offset(leaf->items + nr - 1);
 103 }
 104
 105 /*
 106  * compare two keys in a memcmp fashion
 107  */
 108 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 109 {
 110         struct btrfs_key k1;
 111
 112         btrfs_disk_key_to_cpu(&k1, disk);
 113
 114         if (k1.objectid > k2->objectid)
 115                 return 1;
 116         if (k1.objectid < k2->objectid)
 117                 return -1;
 118         if (k1.offset > k2->offset)
 119                 return 1;
 120         if (k1.offset < k2->offset)
 121                 return -1;
 122         if (k1.flags > k2->flags)
 123                 return 1;
 124         if (k1.flags < k2->flags)
 125                 return -1;
 126         return 0;
 127 }
 128
 129 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
 130                       int level)
 131 {
 132         int i;
 133         struct btrfs_node *parent = NULL;
 134         struct btrfs_node *node = btrfs_buffer_node(path->nodes[level]);
 135         int parent_slot;
 136         u32 nritems = btrfs_header_nritems(&node->header);
 137
 138         if (path->nodes[level + 1])
 139                 parent = btrfs_buffer_node(path->nodes[level + 1]);
 140         parent_slot = path->slots[level + 1];
 141         BUG_ON(nritems == 0);
 142         if (parent) {
 143                 struct btrfs_disk_key *parent_key;
 144                 parent_key = &parent->ptrs[parent_slot].key;
 145                 BUG_ON(memcmp(parent_key, &node->ptrs[0].key,
 146                               sizeof(struct btrfs_disk_key)));
 147                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 148                        btrfs_header_blocknr(&node->header));
 149         }
 150         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 151         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 152                 struct btrfs_key cpukey;
 153                 btrfs_disk_key_to_cpu(&cpukey, &node->ptrs[i + 1].key);
 154                 BUG_ON(comp_keys(&node->ptrs[i].key, &cpukey) >= 0);
 155         }
 156         return 0;
 157 }
 158
 159 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 160                       int level)
 161 {
 162         int i;
 163         struct btrfs_leaf *leaf = btrfs_buffer_leaf(path->nodes[level]);
 164         struct btrfs_node *parent = NULL;
 165         int parent_slot;
 166         u32 nritems = btrfs_header_nritems(&leaf->header);
 167
 168         if (path->nodes[level + 1])
 169                 parent = btrfs_buffer_node(path->nodes[level + 1]);
 170         parent_slot = path->slots[level + 1];
 171         BUG_ON(btrfs_leaf_free_space(root, leaf) < 0);
 172
 173         if (nritems == 0)
 174                 return 0;
 175
 176         if (parent) {
 177                 struct btrfs_disk_key *parent_key;
 178                 parent_key = &parent->ptrs[parent_slot].key;
 179                 BUG_ON(memcmp(parent_key, &leaf->items[0].key,
 180                        sizeof(struct btrfs_disk_key)));
 181                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 182                        btrfs_header_blocknr(&leaf->header));
 183         }
 184         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 185                 struct btrfs_key cpukey;
 186                 btrfs_disk_key_to_cpu(&cpukey, &leaf->items[i + 1].key);
 187                 BUG_ON(comp_keys(&leaf->items[i].key,
 188                                  &cpukey) >= 0);
 189                 BUG_ON(btrfs_item_offset(leaf->items + i) !=
 190                         btrfs_item_end(leaf->items + i + 1));
 191                 if (i == 0) {
 192                         BUG_ON(btrfs_item_offset(leaf->items + i) +
 193                                btrfs_item_size(leaf->items + i) !=
 194                                BTRFS_LEAF_DATA_SIZE(root));
 195                 }
 196         }
 197         return 0;
 198 }
 199
 200 static int check_block(struct btrfs_root *root, struct btrfs_path *path,
 201                         int level)
 202 {
 203         struct btrfs_node *node = btrfs_buffer_node(path->nodes[level]);
 204         if (memcmp(node->header.fsid, root->fs_info->disk_super->fsid,
 205                    sizeof(node->header.fsid)))
 206                 BUG();
 207         if (level == 0)
 208                 return check_leaf(root, path, level);
 209         return check_node(root, path, level);
 210 }
 211
 212 /*
 213  * search for key in the array p.  items p are item_size apart
 214  * and there are 'max' items in p
 215  * the slot in the array is returned via slot, and it points to
 216  * the place where you would insert key if it is not found in
 217  * the array.
 218  *
 219  * slot may point to max if the key is bigger than all of the keys
 220  */
 221 static int generic_bin_search(char *p, int item_size, struct btrfs_key *key,
 222                        int max, int *slot)
 223 {
 224         int low = 0;
 225         int high = max;
 226         int mid;
 227         int ret;
 228         struct btrfs_disk_key *tmp;
 229
 230         while(low < high) {
 231                 mid = (low + high) / 2;
 232                 tmp = (struct btrfs_disk_key *)(p + mid * item_size);
 233                 ret = comp_keys(tmp, key);
 234
 235                 if (ret < 0)
 236                         low = mid + 1;
 237                 else if (ret > 0)
 238                         high = mid;
 239                 else {
 240                         *slot = mid;
 241                         return 0;
 242                 }
 243         }
 244         *slot = low;
 245         return 1;
 246 }
 247
 248 /*
 249  * simple bin_search frontend that does the right thing for
 250  * leaves vs nodes
 251  */
 252 static int bin_search(struct btrfs_node *c, struct btrfs_key *key, int *slot)
 253 {
 254         if (btrfs_is_leaf(c)) {
 255                 struct btrfs_leaf *l = (struct btrfs_leaf *)c;
 256                 return generic_bin_search((void *)l->items,
 257                                           sizeof(struct btrfs_item),
 258                                           key, btrfs_header_nritems(&c->header),
 259                                           slot);
 260         } else {
 261                 return generic_bin_search((void *)c->ptrs,
 262                                           sizeof(struct btrfs_key_ptr),
 263                                           key, btrfs_header_nritems(&c->header),
 264                                           slot);
 265         }
 266         return -1;
 267 }
 268
 269 static struct buffer_head *read_node_slot(struct btrfs_root *root,
 270                                    struct buffer_head *parent_buf,
 271                                    int slot)
 272 {
 273         struct btrfs_node *node = btrfs_buffer_node(parent_buf);
 274         if (slot < 0)
 275                 return NULL;
 276         if (slot >= btrfs_header_nritems(&node->header))
 277                 return NULL;
 278         return read_tree_block(root, btrfs_node_blockptr(node, slot));
 279 }
 280
 281 static int balance_level(struct btrfs_trans_handle *trans, struct btrfs_root
 282                          *root, struct btrfs_path *path, int level)
 283 {
 284         struct buffer_head *right_buf;
 285         struct buffer_head *mid_buf;
 286         struct buffer_head *left_buf;
 287         struct buffer_head *parent_buf = NULL;
 288         struct btrfs_node *right = NULL;
 289         struct btrfs_node *mid;
 290         struct btrfs_node *left = NULL;
 291         struct btrfs_node *parent = NULL;
 292         int ret = 0;
 293         int wret;
 294         int pslot;
 295         int orig_slot = path->slots[level];
 296         u64 orig_ptr;
 297
 298         if (level == 0)
 299                 return 0;
 300
 301         mid_buf = path->nodes[level];
 302         mid = btrfs_buffer_node(mid_buf);
 303         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 304
 305         if (level < BTRFS_MAX_LEVEL - 1)
 306                 parent_buf = path->nodes[level + 1];
 307         pslot = path->slots[level + 1];
 308
 309         /*
 310          * deal with the case where there is only one pointer in the root
 311          * by promoting the node below to a root
 312          */
 313         if (!parent_buf) {
 314                 struct buffer_head *child;
 315                 u64 blocknr = bh_blocknr(mid_buf);
 316
 317                 if (btrfs_header_nritems(&mid->header) != 1)
 318                         return 0;
 319
 320                 /* promote the child to a root */
 321                 child = read_node_slot(root, mid_buf, 0);
 322                 BUG_ON(!child);
 323                 root->node = child;
 324                 path->nodes[level] = NULL;
 325                 clean_tree_block(trans, root, mid_buf);
 326                 wait_on_buffer(mid_buf);
 327                 /* once for the path */
 328                 btrfs_block_release(root, mid_buf);
 329                 /* once for the root ptr */
 330                 btrfs_block_release(root, mid_buf);
 331                 return btrfs_free_extent(trans, root, blocknr, 1, 1);
 332         }
 333         parent = btrfs_buffer_node(parent_buf);
 334
 335         if (btrfs_header_nritems(&mid->header) >
 336             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 337                 return 0;
 338
 339         left_buf = read_node_slot(root, parent_buf, pslot - 1);
 340         right_buf = read_node_slot(root, parent_buf, pslot + 1);
 341
 342         /* first, try to make some room in the middle buffer */
 343         if (left_buf) {
 344                 btrfs_cow_block(trans, root, left_buf, parent_buf, pslot - 1,
 345                                 &left_buf);
 346                 left = btrfs_buffer_node(left_buf);
 347                 orig_slot += btrfs_header_nritems(&left->header);
 348                 wret = push_node_left(trans, root, left_buf, mid_buf);
 349                 if (wret < 0)
 350                         ret = wret;
 351         }
 352
 353         /*
 354          * then try to empty the right most buffer into the middle
 355          */
 356         if (right_buf) {
 357                 btrfs_cow_block(trans, root, right_buf, parent_buf, pslot + 1,
 358                                 &right_buf);
 359                 right = btrfs_buffer_node(right_buf);
 360                 wret = push_node_left(trans, root, mid_buf, right_buf);
 361                 if (wret < 0)
 362                         ret = wret;
 363                 if (btrfs_header_nritems(&right->header) == 0) {
 364                         u64 blocknr = bh_blocknr(right_buf);
 365                         clean_tree_block(trans, root, right_buf);
 366                         wait_on_buffer(right_buf);
 367                         btrfs_block_release(root, right_buf);
 368                         right_buf = NULL;
 369                         right = NULL;
 370                         wret = del_ptr(trans, root, path, level + 1, pslot +
 371                                        1);
 372                         if (wret)
 373                                 ret = wret;
 374                         wret = btrfs_free_extent(trans, root, blocknr, 1, 1);
 375                         if (wret)
 376                                 ret = wret;
 377                 } else {
 378                         btrfs_memcpy(root, parent,
 379                                      &parent->ptrs[pslot + 1].key,
 380                                      &right->ptrs[0].key,
 381                                      sizeof(struct btrfs_disk_key));
 382                         btrfs_mark_buffer_dirty(parent_buf);
 383                 }
 384         }
 385         if (btrfs_header_nritems(&mid->header) == 1) {
 386                 /*
 387                  * we're not allowed to leave a node with one item in the
 388                  * tree during a delete.  A deletion from lower in the tree
 389                  * could try to delete the only pointer in this node.
 390                  * So, pull some keys from the left.
 391                  * There has to be a left pointer at this point because
 392                  * otherwise we would have pulled some pointers from the
 393                  * right
 394                  */
 395                 BUG_ON(!left_buf);
 396                 wret = balance_node_right(trans, root, mid_buf, left_buf);
 397                 if (wret < 0)
 398                         ret = wret;
 399                 BUG_ON(wret == 1);
 400         }
 401         if (btrfs_header_nritems(&mid->header) == 0) {
 402                 /* we've managed to empty the middle node, drop it */
 403                 u64 blocknr = bh_blocknr(mid_buf);
 404                 clean_tree_block(trans, root, mid_buf);
 405                 wait_on_buffer(mid_buf);
 406                 btrfs_block_release(root, mid_buf);
 407                 mid_buf = NULL;
 408                 mid = NULL;
 409                 wret = del_ptr(trans, root, path, level + 1, pslot);
 410                 if (wret)
 411                         ret = wret;
 412                 wret = btrfs_free_extent(trans, root, blocknr, 1, 1);
 413                 if (wret)
 414                         ret = wret;
 415         } else {
 416                 /* update the parent key to reflect our changes */
 417                 btrfs_memcpy(root, parent,
 418                              &parent->ptrs[pslot].key, &mid->ptrs[0].key,
 419                              sizeof(struct btrfs_disk_key));
 420                 btrfs_mark_buffer_dirty(parent_buf);
 421         }
 422
 423         /* update the path */
 424         if (left_buf) {
 425                 if (btrfs_header_nritems(&left->header) > orig_slot) {
 426                         get_bh(left_buf);
 427                         path->nodes[level] = left_buf;
 428                         path->slots[level + 1] -= 1;
 429                         path->slots[level] = orig_slot;
 430                         if (mid_buf)
 431                                 btrfs_block_release(root, mid_buf);
 432                 } else {
 433                         orig_slot -= btrfs_header_nritems(&left->header);
 434                         path->slots[level] = orig_slot;
 435                 }
 436         }
 437         /* double check we haven't messed things up */
 438         check_block(root, path, level);
 439         if (orig_ptr !=
 440             btrfs_node_blockptr(btrfs_buffer_node(path->nodes[level]),
 441                                 path->slots[level]))
 442                 BUG();
 443
 444         if (right_buf)
 445                 btrfs_block_release(root, right_buf);
 446         if (left_buf)
 447                 btrfs_block_release(root, left_buf);
 448         return ret;
 449 }
 450
 451 /*
 452  * look for key in the tree.  path is filled in with nodes along the way
 453  * if key is found, we return zero and you can find the item in the leaf
 454  * level of the path (level 0)
 455  *
 456  * If the key isn't found, the path points to the slot where it should
 457  * be inserted, and 1 is returned.  If there are other errors during the
 458  * search a negative error number is returned.
 459  *
 460  * if ins_len > 0, nodes and leaves will be split as we walk down the
 461  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
 462  * possible)
 463  */
 464 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
 465                       *root, struct btrfs_key *key, struct btrfs_path *p, int
 466                       ins_len, int cow)
 467 {
 468         struct buffer_head *b;
 469         struct buffer_head *cow_buf;
 470         struct btrfs_node *c;
 471         int slot;
 472         int ret;
 473         int level;
 474
 475         WARN_ON(p->nodes[0] != NULL);
 476         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
 477 again:
 478         b = root->node;
 479         get_bh(b);
 480         while (b) {
 481                 c = btrfs_buffer_node(b);
 482                 level = btrfs_header_level(&c->header);
 483                 if (cow) {
 484                         int wret;
 485                         wret = btrfs_cow_block(trans, root, b,
 486                                                p->nodes[level + 1],
 487                                                p->slots[level + 1],
 488                                                &cow_buf);
 489                         b = cow_buf;
 490                         c = btrfs_buffer_node(b);
 491                 }
 492                 BUG_ON(!cow && ins_len);
 493                 if (level != btrfs_header_level(&c->header))
 494                         WARN_ON(1);
 495                 level = btrfs_header_level(&c->header);
 496                 p->nodes[level] = b;
 497                 ret = check_block(root, p, level);
 498                 if (ret)
 499                         return -1;
 500                 ret = bin_search(c, key, &slot);
 501                 if (!btrfs_is_leaf(c)) {
 502                         if (ret && slot > 0)
 503                                 slot -= 1;
 504                         p->slots[level] = slot;
 505                         if (ins_len > 0 && btrfs_header_nritems(&c->header) >=
 506                             BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 507                                 int sret = split_node(trans, root, p, level);
 508                                 BUG_ON(sret > 0);
 509                                 if (sret)
 510                                         return sret;
 511                                 b = p->nodes[level];
 512                                 c = btrfs_buffer_node(b);
 513                                 slot = p->slots[level];
 514                         } else if (ins_len < 0) {
 515                                 int sret = balance_level(trans, root, p,
 516                                                          level);
 517                                 if (sret)
 518                                         return sret;
 519                                 b = p->nodes[level];
 520                                 if (!b)
 521                                         goto again;
 522                                 c = btrfs_buffer_node(b);
 523                                 slot = p->slots[level];
 524                                 BUG_ON(btrfs_header_nritems(&c->header) == 1);
 525                         }
 526                         b = read_tree_block(root, btrfs_node_blockptr(c, slot));
 527                 } else {
 528                         struct btrfs_leaf *l = (struct btrfs_leaf *)c;
 529                         p->slots[level] = slot;
 530                         if (ins_len > 0 && btrfs_leaf_free_space(root, l) <
 531                             sizeof(struct btrfs_item) + ins_len) {
 532                                 int sret = split_leaf(trans, root, key,
 533                                                       p, ins_len);
 534                                 BUG_ON(sret > 0);
 535                                 if (sret)
 536                                         return sret;
 537                         }
 538                         return ret;
 539                 }
 540         }
 541         return 1;
 542 }
 543
 544 /*
 545  * adjust the pointers going up the tree, starting at level
 546  * making sure the right key of each node is points to 'key'.
 547  * This is used after shifting pointers to the left, so it stops
 548  * fixing up pointers when a given leaf/node is not in slot 0 of the
 549  * higher levels
 550  *
 551  * If this fails to write a tree block, it returns -1, but continues
 552  * fixing up the blocks in ram so the tree is consistent.
 553  */
 554 static int fixup_low_keys(struct btrfs_trans_handle *trans, struct btrfs_root
 555                           *root, struct btrfs_path *path, struct btrfs_disk_key
 556                           *key, int level)
 557 {
 558         int i;
 559         int ret = 0;
 560         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
 561                 struct btrfs_node *t;
 562                 int tslot = path->slots[i];
 563                 if (!path->nodes[i])
 564                         break;
 565                 t = btrfs_buffer_node(path->nodes[i]);
 566                 btrfs_memcpy(root, t, &t->ptrs[tslot].key, key, sizeof(*key));
 567                 btrfs_mark_buffer_dirty(path->nodes[i]);
 568                 if (tslot != 0)
 569                         break;
 570         }
 571         return ret;
 572 }
 573
 574 /*
 575  * try to push data from one node into the next node left in the
 576  * tree.
 577  *
 578  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
 579  * error, and > 0 if there was no room in the left hand block.
 580  */
 581 static int push_node_left(struct btrfs_trans_handle *trans, struct btrfs_root
 582                           *root, struct buffer_head *dst_buf, struct
 583                           buffer_head *src_buf)
 584 {
 585         struct btrfs_node *src = btrfs_buffer_node(src_buf);
 586         struct btrfs_node *dst = btrfs_buffer_node(dst_buf);
 587         int push_items = 0;
 588         int src_nritems;
 589         int dst_nritems;
 590         int ret = 0;
 591
 592         src_nritems = btrfs_header_nritems(&src->header);
 593         dst_nritems = btrfs_header_nritems(&dst->header);
 594         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
 595         if (push_items <= 0) {
 596                 return 1;
 597         }
 598
 599         if (src_nritems < push_items)
 600                 push_items = src_nritems;
 601
 602         btrfs_memcpy(root, dst, dst->ptrs + dst_nritems, src->ptrs,
 603                      push_items * sizeof(struct btrfs_key_ptr));
 604         if (push_items < src_nritems) {
 605                 btrfs_memmove(root, src, src->ptrs, src->ptrs + push_items,
 606                         (src_nritems - push_items) *
 607                         sizeof(struct btrfs_key_ptr));
 608         }
 609         btrfs_set_header_nritems(&src->header, src_nritems - push_items);
 610         btrfs_set_header_nritems(&dst->header, dst_nritems + push_items);
 611         btrfs_mark_buffer_dirty(src_buf);
 612         btrfs_mark_buffer_dirty(dst_buf);
 613         return ret;
 614 }
 615
 616 /*
 617  * try to push data from one node into the next node right in the
 618  * tree.
 619  *
 620  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
 621  * error, and > 0 if there was no room in the right hand block.
 622  *
 623  * this will  only push up to 1/2 the contents of the left node over
 624  */
 625 static int balance_node_right(struct btrfs_trans_handle *trans, struct
 626                               btrfs_root *root, struct buffer_head *dst_buf,
 627                               struct buffer_head *src_buf)
 628 {
 629         struct btrfs_node *src = btrfs_buffer_node(src_buf);
 630         struct btrfs_node *dst = btrfs_buffer_node(dst_buf);
 631         int push_items = 0;
 632         int max_push;
 633         int src_nritems;
 634         int dst_nritems;
 635         int ret = 0;
 636
 637         src_nritems = btrfs_header_nritems(&src->header);
 638         dst_nritems = btrfs_header_nritems(&dst->header);
 639         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
 640         if (push_items <= 0) {
 641                 return 1;
 642         }
 643
 644         max_push = src_nritems / 2 + 1;
 645         /* don't try to empty the node */
 646         if (max_push > src_nritems)
 647                 return 1;
 648         if (max_push < push_items)
 649                 push_items = max_push;
 650
 651         btrfs_memmove(root, dst, dst->ptrs + push_items, dst->ptrs,
 652                       dst_nritems * sizeof(struct btrfs_key_ptr));
 653
 654         btrfs_memcpy(root, dst, dst->ptrs,
 655                      src->ptrs + src_nritems - push_items,
 656                      push_items * sizeof(struct btrfs_key_ptr));
 657
 658         btrfs_set_header_nritems(&src->header, src_nritems - push_items);
 659         btrfs_set_header_nritems(&dst->header, dst_nritems + push_items);
 660
 661         btrfs_mark_buffer_dirty(src_buf);
 662         btrfs_mark_buffer_dirty(dst_buf);
 663         return ret;
 664 }
 665
 666 /*
 667  * helper function to insert a new root level in the tree.
 668  * A new node is allocated, and a single item is inserted to
 669  * point to the existing root
 670  *
 671  * returns zero on success or < 0 on failure.
 672  */
 673 static int insert_new_root(struct btrfs_trans_handle *trans, struct btrfs_root
 674                            *root, struct btrfs_path *path, int level)
 675 {
 676         struct buffer_head *t;
 677         struct btrfs_node *lower;
 678         struct btrfs_node *c;
 679         struct btrfs_disk_key *lower_key;
 680
 681         BUG_ON(path->nodes[level]);
 682         BUG_ON(path->nodes[level-1] != root->node);
 683
 684         t = btrfs_alloc_free_block(trans, root);
 685         c = btrfs_buffer_node(t);
 686         memset(c, 0, root->blocksize);
 687         btrfs_set_header_nritems(&c->header, 1);
 688         btrfs_set_header_level(&c->header, level);
 689         btrfs_set_header_blocknr(&c->header, bh_blocknr(t));
 690         btrfs_set_header_generation(&c->header, trans->transid);
 691         lower = btrfs_buffer_node(path->nodes[level-1]);
 692         memcpy(c->header.fsid, root->fs_info->disk_super->fsid,
 693                sizeof(c->header.fsid));
 694         if (btrfs_is_leaf(lower))
 695                 lower_key = &((struct btrfs_leaf *)lower)->items[0].key;
 696         else
 697                 lower_key = &lower->ptrs[0].key;
 698         btrfs_memcpy(root, c, &c->ptrs[0].key, lower_key,
 699                      sizeof(struct btrfs_disk_key));
 700         btrfs_set_node_blockptr(c, 0, bh_blocknr(path->nodes[level - 1]));
 701
 702         btrfs_mark_buffer_dirty(t);
 703
 704         /* the super has an extra ref to root->node */
 705         btrfs_block_release(root, root->node);
 706         root->node = t;
 707         get_bh(t);
 708         path->nodes[level] = t;
 709         path->slots[level] = 0;
 710         return 0;
 711 }
 712
 713 /*
 714  * worker function to insert a single pointer in a node.
 715  * the node should have enough room for the pointer already
 716  *
 717  * slot and level indicate where you want the key to go, and
 718  * blocknr is the block the key points to.
 719  *
 720  * returns zero on success and < 0 on any error
 721  */
 722 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
 723                       *root, struct btrfs_path *path, struct btrfs_disk_key
 724                       *key, u64 blocknr, int slot, int level)
 725 {
 726         struct btrfs_node *lower;
 727         int nritems;
 728
 729         BUG_ON(!path->nodes[level]);
 730         lower = btrfs_buffer_node(path->nodes[level]);
 731         nritems = btrfs_header_nritems(&lower->header);
 732         if (slot > nritems)
 733                 BUG();
 734         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
 735                 BUG();
 736         if (slot != nritems) {
 737                 btrfs_memmove(root, lower, lower->ptrs + slot + 1,
 738                               lower->ptrs + slot,
 739                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
 740         }
 741         btrfs_memcpy(root, lower, &lower->ptrs[slot].key,
 742                      key, sizeof(struct btrfs_disk_key));
 743         btrfs_set_node_blockptr(lower, slot, blocknr);
 744         btrfs_set_header_nritems(&lower->header, nritems + 1);
 745         btrfs_mark_buffer_dirty(path->nodes[level]);
 746         return 0;
 747 }
 748
 749 /*
 750  * split the node at the specified level in path in two.
 751  * The path is corrected to point to the appropriate node after the split
 752  *
 753  * Before splitting this tries to make some room in the node by pushing
 754  * left and right, if either one works, it returns right away.
 755  *
 756  * returns 0 on success and < 0 on failure
 757  */
 758 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
 759                       *root, struct btrfs_path *path, int level)
 760 {
 761         struct buffer_head *t;
 762         struct btrfs_node *c;
 763         struct buffer_head *split_buffer;
 764         struct btrfs_node *split;
 765         int mid;
 766         int ret;
 767         int wret;
 768         u32 c_nritems;
 769
 770         t = path->nodes[level];
 771         c = btrfs_buffer_node(t);
 772         if (t == root->node) {
 773                 /* trying to split the root, lets make a new one */
 774                 ret = insert_new_root(trans, root, path, level + 1);
 775                 if (ret)
 776                         return ret;
 777         }
 778         c_nritems = btrfs_header_nritems(&c->header);
 779         split_buffer = btrfs_alloc_free_block(trans, root);
 780         split = btrfs_buffer_node(split_buffer);
 781         btrfs_set_header_flags(&split->header, btrfs_header_flags(&c->header));
 782         btrfs_set_header_level(&split->header, btrfs_header_level(&c->header));
 783         btrfs_set_header_blocknr(&split->header, bh_blocknr(split_buffer));
 784         btrfs_set_header_generation(&split->header, trans->transid);
 785         memcpy(split->header.fsid, root->fs_info->disk_super->fsid,
 786                sizeof(split->header.fsid));
 787         mid = (c_nritems + 1) / 2;
 788         btrfs_memcpy(root, split, split->ptrs, c->ptrs + mid,
 789                      (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
 790         btrfs_set_header_nritems(&split->header, c_nritems - mid);
 791         btrfs_set_header_nritems(&c->header, mid);
 792         ret = 0;
 793
 794         btrfs_mark_buffer_dirty(t);
 795         btrfs_mark_buffer_dirty(split_buffer);
 796         wret = insert_ptr(trans, root, path, &split->ptrs[0].key,
 797                           bh_blocknr(split_buffer), path->slots[level + 1] + 1,
 798                           level + 1);
 799         if (wret)
 800                 ret = wret;
 801
 802         if (path->slots[level] >= mid) {
 803                 path->slots[level] -= mid;
 804                 btrfs_block_release(root, t);
 805                 path->nodes[level] = split_buffer;
 806                 path->slots[level + 1] += 1;
 807         } else {
 808                 btrfs_block_release(root, split_buffer);
 809         }
 810         return ret;
 811 }
 812
 813 /*
 814  * how many bytes are required to store the items in a leaf.  start
 815  * and nr indicate which items in the leaf to check.  This totals up the
 816  * space used both by the item structs and the item data
 817  */
 818 static int leaf_space_used(struct btrfs_leaf *l, int start, int nr)
 819 {
 820         int data_len;
 821         int nritems = btrfs_header_nritems(&l->header);
 822         int end = min(nritems, start + nr) - 1;
 823
 824         if (!nr)
 825                 return 0;
 826         data_len = btrfs_item_end(l->items + start);
 827         data_len = data_len - btrfs_item_offset(l->items + end);
 828         data_len += sizeof(struct btrfs_item) * nr;
 829         WARN_ON(data_len < 0);
 830         return data_len;
 831 }
 832
 833 /*
 834  * The space between the end of the leaf items and
 835  * the start of the leaf data.  IOW, how much room
 836  * the leaf has left for both items and data
 837  */
 838 int btrfs_leaf_free_space(struct btrfs_root *root, struct btrfs_leaf *leaf)
 839 {
 840         int nritems = btrfs_header_nritems(&leaf->header);
 841         return BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
 842 }
 843
 844 /*
 845  * push some data in the path leaf to the right, trying to free up at
 846  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
 847  *
 848  * returns 1 if the push failed because the other node didn't have enough
 849  * room, 0 if everything worked out and < 0 if there were major errors.
 850  */
 851 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
 852                            *root, struct btrfs_path *path, int data_size)
 853 {
 854         struct buffer_head *left_buf = path->nodes[0];
 855         struct btrfs_leaf *left = btrfs_buffer_leaf(left_buf);
 856         struct btrfs_leaf *right;
 857         struct buffer_head *right_buf;
 858         struct buffer_head *upper;
 859         struct btrfs_node *upper_node;
 860         int slot;
 861         int i;
 862         int free_space;
 863         int push_space = 0;
 864         int push_items = 0;
 865         struct btrfs_item *item;
 866         u32 left_nritems;
 867         u32 right_nritems;
 868
 869         slot = path->slots[1];
 870         if (!path->nodes[1]) {
 871                 return 1;
 872         }
 873         upper = path->nodes[1];
 874         upper_node = btrfs_buffer_node(upper);
 875         if (slot >= btrfs_header_nritems(&upper_node->header) - 1) {
 876                 return 1;
 877         }
 878         right_buf = read_tree_block(root,
 879                     btrfs_node_blockptr(btrfs_buffer_node(upper), slot + 1));
 880         right = btrfs_buffer_leaf(right_buf);
 881         free_space = btrfs_leaf_free_space(root, right);
 882         if (free_space < data_size + sizeof(struct btrfs_item)) {
 883                 btrfs_block_release(root, right_buf);
 884                 return 1;
 885         }
 886         /* cow and double check */
 887         btrfs_cow_block(trans, root, right_buf, upper, slot + 1, &right_buf);
 888         right = btrfs_buffer_leaf(right_buf);
 889         free_space = btrfs_leaf_free_space(root, right);
 890         if (free_space < data_size + sizeof(struct btrfs_item)) {
 891                 btrfs_block_release(root, right_buf);
 892                 return 1;
 893         }
 894
 895         left_nritems = btrfs_header_nritems(&left->header);
 896         for (i = left_nritems - 1; i >= 0; i--) {
 897                 item = left->items + i;
 898                 if (path->slots[0] == i)
 899                         push_space += data_size + sizeof(*item);
 900                 if (btrfs_item_size(item) + sizeof(*item) + push_space >
 901                     free_space)
 902                         break;
 903                 push_items++;
 904                 push_space += btrfs_item_size(item) + sizeof(*item);
 905         }
 906         if (push_items == 0) {
 907                 btrfs_block_release(root, right_buf);
 908                 return 1;
 909         }
 910         right_nritems = btrfs_header_nritems(&right->header);
 911         /* push left to right */
 912         push_space = btrfs_item_end(left->items + left_nritems - push_items);
 913         push_space -= leaf_data_end(root, left);
 914         /* make room in the right data area */
 915         btrfs_memmove(root, right, btrfs_leaf_data(right) +
 916                       leaf_data_end(root, right) - push_space,
 917                       btrfs_leaf_data(right) +
 918                       leaf_data_end(root, right), BTRFS_LEAF_DATA_SIZE(root) -
 919                       leaf_data_end(root, right));
 920         /* copy from the left data area */
 921         btrfs_memcpy(root, right, btrfs_leaf_data(right) +
 922                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
 923                      btrfs_leaf_data(left) + leaf_data_end(root, left),
 924                      push_space);
 925         btrfs_memmove(root, right, right->items + push_items, right->items,
 926                 right_nritems * sizeof(struct btrfs_item));
 927         /* copy the items from left to right */
 928         btrfs_memcpy(root, right, right->items, left->items +
 929                      left_nritems - push_items,
 930                      push_items * sizeof(struct btrfs_item));
 931
 932         /* update the item pointers */
 933         right_nritems += push_items;
 934         btrfs_set_header_nritems(&right->header, right_nritems);
 935         push_space = BTRFS_LEAF_DATA_SIZE(root);
 936         for (i = 0; i < right_nritems; i++) {
 937                 btrfs_set_item_offset(right->items + i, push_space -
 938                                       btrfs_item_size(right->items + i));
 939                 push_space = btrfs_item_offset(right->items + i);
 940         }
 941         left_nritems -= push_items;
 942         btrfs_set_header_nritems(&left->header, left_nritems);
 943
 944         btrfs_mark_buffer_dirty(left_buf);
 945         btrfs_mark_buffer_dirty(right_buf);
 946         btrfs_memcpy(root, upper_node, &upper_node->ptrs[slot + 1].key,
 947                 &right->items[0].key, sizeof(struct btrfs_disk_key));
 948         btrfs_mark_buffer_dirty(upper);
 949
 950         /* then fixup the leaf pointer in the path */
 951         if (path->slots[0] >= left_nritems) {
 952                 path->slots[0] -= left_nritems;
 953                 btrfs_block_release(root, path->nodes[0]);
 954                 path->nodes[0] = right_buf;
 955                 path->slots[1] += 1;
 956         } else {
 957                 btrfs_block_release(root, right_buf);
 958         }
 959         return 0;
 960 }
 961 /*
 962  * push some data in the path leaf to the left, trying to free up at
 963  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
 964  */
 965 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
 966                           *root, struct btrfs_path *path, int data_size)
 967 {
 968         struct buffer_head *right_buf = path->nodes[0];
 969         struct btrfs_leaf *right = btrfs_buffer_leaf(right_buf);
 970         struct buffer_head *t;
 971         struct btrfs_leaf *left;
 972         int slot;
 973         int i;
 974         int free_space;
 975         int push_space = 0;
 976         int push_items = 0;
 977         struct btrfs_item *item;
 978         u32 old_left_nritems;
 979         int ret = 0;
 980         int wret;
 981
 982         slot = path->slots[1];
 983         if (slot == 0) {
 984                 return 1;
 985         }
 986         if (!path->nodes[1]) {
 987                 return 1;
 988         }
 989         t = read_tree_block(root,
 990             btrfs_node_blockptr(btrfs_buffer_node(path->nodes[1]), slot - 1));
 991         left = btrfs_buffer_leaf(t);
 992         free_space = btrfs_leaf_free_space(root, left);
 993         if (free_space < data_size + sizeof(struct btrfs_item)) {
 994                 btrfs_block_release(root, t);
 995                 return 1;
 996         }
 997
 998         /* cow and double check */
 999         btrfs_cow_block(trans, root, t, path->nodes[1], slot - 1, &t);
1000         left = btrfs_buffer_leaf(t);
1001         free_space = btrfs_leaf_free_space(root, left);
1002         if (free_space < data_size + sizeof(struct btrfs_item)) {
1003                 btrfs_block_release(root, t);
1004                 return 1;
1005         }
1006
1007         for (i = 0; i < btrfs_header_nritems(&right->header); i++) {
1008                 item = right->items + i;
1009                 if (path->slots[0] == i)
1010                         push_space += data_size + sizeof(*item);
1011                 if (btrfs_item_size(item) + sizeof(*item) + push_space >
1012                     free_space)
1013                         break;
1014                 push_items++;
1015                 push_space += btrfs_item_size(item) + sizeof(*item);
1016         }
1017         if (push_items == 0) {
1018                 btrfs_block_release(root, t);
1019                 return 1;
1020         }
1021         /* push data from right to left */
1022         btrfs_memcpy(root, left, left->items +
1023                      btrfs_header_nritems(&left->header),
1024                      right->items, push_items * sizeof(struct btrfs_item));
1025         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1026                      btrfs_item_offset(right->items + push_items -1);
1027         btrfs_memcpy(root, left, btrfs_leaf_data(left) +
1028                      leaf_data_end(root, left) - push_space,
1029                      btrfs_leaf_data(right) +
1030                      btrfs_item_offset(right->items + push_items - 1),
1031                      push_space);
1032         old_left_nritems = btrfs_header_nritems(&left->header);
1033         BUG_ON(old_left_nritems < 0);
1034
1035         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1036                 u32 ioff = btrfs_item_offset(left->items + i);
1037                 btrfs_set_item_offset(left->items + i, ioff -
1038                                      (BTRFS_LEAF_DATA_SIZE(root) -
1039                                       btrfs_item_offset(left->items +
1040                                                         old_left_nritems - 1)));
1041         }
1042         btrfs_set_header_nritems(&left->header, old_left_nritems + push_items);
1043
1044         /* fixup right node */
1045         push_space = btrfs_item_offset(right->items + push_items - 1) -
1046                      leaf_data_end(root, right);
1047         btrfs_memmove(root, right, btrfs_leaf_data(right) +
1048                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1049                       btrfs_leaf_data(right) +
1050                       leaf_data_end(root, right), push_space);
1051         btrfs_memmove(root, right, right->items, right->items + push_items,
1052                 (btrfs_header_nritems(&right->header) - push_items) *
1053                 sizeof(struct btrfs_item));
1054         btrfs_set_header_nritems(&right->header,
1055                                  btrfs_header_nritems(&right->header) -
1056                                  push_items);
1057         push_space = BTRFS_LEAF_DATA_SIZE(root);
1058
1059         for (i = 0; i < btrfs_header_nritems(&right->header); i++) {
1060                 btrfs_set_item_offset(right->items + i, push_space -
1061                                       btrfs_item_size(right->items + i));
1062                 push_space = btrfs_item_offset(right->items + i);
1063         }
1064
1065         btrfs_mark_buffer_dirty(t);
1066         btrfs_mark_buffer_dirty(right_buf);
1067
1068         wret = fixup_low_keys(trans, root, path, &right->items[0].key, 1);
1069         if (wret)
1070                 ret = wret;
1071
1072         /* then fixup the leaf pointer in the path */
1073         if (path->slots[0] < push_items) {
1074                 path->slots[0] += old_left_nritems;
1075                 btrfs_block_release(root, path->nodes[0]);
1076                 path->nodes[0] = t;
1077                 path->slots[1] -= 1;
1078         } else {
1079                 btrfs_block_release(root, t);
1080                 path->slots[0] -= push_items;
1081         }
1082         BUG_ON(path->slots[0] < 0);
1083         return ret;
1084 }
1085
1086 /*
1087  * split the path's leaf in two, making sure there is at least data_size
1088  * available for the resulting leaf level of the path.
1089  *
1090  * returns 0 if all went well and < 0 on failure.
1091  */
1092 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
1093                       *root, struct btrfs_key *ins_key,
1094                       struct btrfs_path *path, int data_size)
1095 {
1096         struct buffer_head *l_buf;
1097         struct btrfs_leaf *l;
1098         u32 nritems;
1099         int mid;
1100         int slot;
1101         struct btrfs_leaf *right;
1102         struct buffer_head *right_buffer;
1103         int space_needed = data_size + sizeof(struct btrfs_item);
1104         int data_copy_size;
1105         int rt_data_off;
1106         int i;
1107         int ret = 0;
1108         int wret;
1109         int double_split = 0;
1110         struct btrfs_disk_key disk_key;
1111
1112         /* first try to make some room by pushing left and right */
1113         wret = push_leaf_left(trans, root, path, data_size);
1114         if (wret < 0)
1115                 return wret;
1116         if (wret) {
1117                 wret = push_leaf_right(trans, root, path, data_size);
1118                 if (wret < 0)
1119                         return wret;
1120         }
1121         l_buf = path->nodes[0];
1122         l = btrfs_buffer_leaf(l_buf);
1123
1124         /* did the pushes work? */
1125         if (btrfs_leaf_free_space(root, l) >=
1126             sizeof(struct btrfs_item) + data_size)
1127                 return 0;
1128
1129         if (!path->nodes[1]) {
1130                 ret = insert_new_root(trans, root, path, 1);
1131                 if (ret)
1132                         return ret;
1133         }
1134         slot = path->slots[0];
1135         nritems = btrfs_header_nritems(&l->header);
1136         mid = (nritems + 1)/ 2;
1137         right_buffer = btrfs_alloc_free_block(trans, root);
1138         BUG_ON(!right_buffer);
1139         right = btrfs_buffer_leaf(right_buffer);
1140         memset(&right->header, 0, sizeof(right->header));
1141         btrfs_set_header_blocknr(&right->header, bh_blocknr(right_buffer));
1142         btrfs_set_header_generation(&right->header, trans->transid);
1143         btrfs_set_header_level(&right->header, 0);
1144         memcpy(right->header.fsid, root->fs_info->disk_super->fsid,
1145                sizeof(right->header.fsid));
1146         if (mid <= slot) {
1147                 if (nritems == 1 ||
1148                     leaf_space_used(l, mid, nritems - mid) + space_needed >
1149                         BTRFS_LEAF_DATA_SIZE(root)) {
1150                         if (slot >= nritems) {
1151                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1152                                 btrfs_set_header_nritems(&right->header, 0);
1153                                 wret = insert_ptr(trans, root, path,
1154                                                   &disk_key,
1155                                                   bh_blocknr(right_buffer),
1156                                                   path->slots[1] + 1, 1);
1157                                 if (wret)
1158                                         ret = wret;
1159                                 btrfs_block_release(root, path->nodes[0]);
1160                                 path->nodes[0] = right_buffer;
1161                                 path->slots[0] = 0;
1162                                 path->slots[1] += 1;
1163                                 return ret;
1164                         }
1165                         mid = slot;
1166                         double_split = 1;
1167                 }
1168         } else {
1169                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
1170                         BTRFS_LEAF_DATA_SIZE(root)) {
1171                         if (slot == 0) {
1172                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1173                                 btrfs_set_header_nritems(&right->header, 0);
1174                                 wret = insert_ptr(trans, root, path,
1175                                                   &disk_key,
1176                                                   bh_blocknr(right_buffer),
1177                                                   path->slots[1] - 1, 1);
1178                                 if (wret)
1179                                         ret = wret;
1180                                 btrfs_block_release(root, path->nodes[0]);
1181                                 path->nodes[0] = right_buffer;
1182                                 path->slots[0] = 0;
1183                                 path->slots[1] -= 1;
1184                                 return ret;
1185                         }
1186                         mid = slot;
1187                         double_split = 1;
1188                 }
1189         }
1190         btrfs_set_header_nritems(&right->header, nritems - mid);
1191         data_copy_size = btrfs_item_end(l->items + mid) -
1192                          leaf_data_end(root, l);
1193         btrfs_memcpy(root, right, right->items, l->items + mid,
1194                      (nritems - mid) * sizeof(struct btrfs_item));
1195         btrfs_memcpy(root, right,
1196                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
1197                      data_copy_size, btrfs_leaf_data(l) +
1198                      leaf_data_end(root, l), data_copy_size);
1199         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
1200                       btrfs_item_end(l->items + mid);
1201
1202         for (i = 0; i < btrfs_header_nritems(&right->header); i++) {
1203                 u32 ioff = btrfs_item_offset(right->items + i);
1204                 btrfs_set_item_offset(right->items + i, ioff + rt_data_off);
1205         }
1206
1207         btrfs_set_header_nritems(&l->header, mid);
1208         ret = 0;
1209         wret = insert_ptr(trans, root, path, &right->items[0].key,
1210                           bh_blocknr(right_buffer), path->slots[1] + 1, 1);
1211         if (wret)
1212                 ret = wret;
1213         btrfs_mark_buffer_dirty(right_buffer);
1214         btrfs_mark_buffer_dirty(l_buf);
1215         BUG_ON(path->slots[0] != slot);
1216         if (mid <= slot) {
1217                 btrfs_block_release(root, path->nodes[0]);
1218                 path->nodes[0] = right_buffer;
1219                 path->slots[0] -= mid;
1220                 path->slots[1] += 1;
1221         } else
1222                 btrfs_block_release(root, right_buffer);
1223         BUG_ON(path->slots[0] < 0);
1224
1225         if (!double_split)
1226                 return ret;
1227         right_buffer = btrfs_alloc_free_block(trans, root);
1228         BUG_ON(!right_buffer);
1229         right = btrfs_buffer_leaf(right_buffer);
1230         memset(&right->header, 0, sizeof(right->header));
1231         btrfs_set_header_blocknr(&right->header, bh_blocknr(right_buffer));
1232         btrfs_set_header_generation(&right->header, trans->transid);
1233         btrfs_set_header_level(&right->header, 0);
1234         memcpy(right->header.fsid, root->fs_info->disk_super->fsid,
1235                sizeof(right->header.fsid));
1236         btrfs_cpu_key_to_disk(&disk_key, ins_key);
1237         btrfs_set_header_nritems(&right->header, 0);
1238         wret = insert_ptr(trans, root, path,
1239                           &disk_key,
1240                           bh_blocknr(right_buffer),
1241                           path->slots[1], 1);
1242         if (wret)
1243                 ret = wret;
1244         btrfs_block_release(root, path->nodes[0]);
1245         path->nodes[0] = right_buffer;
1246         path->slots[0] = 0;
1247         check_node(root, path, 1);
1248         check_leaf(root, path, 0);
1249         return ret;
1250 }
1251
1252 int btrfs_extend_item(struct btrfs_trans_handle *trans, struct btrfs_root
1253                       *root, struct btrfs_path *path, u32 data_size)
1254 {
1255         int ret = 0;
1256         int slot;
1257         int slot_orig;
1258         struct btrfs_leaf *leaf;
1259         struct buffer_head *leaf_buf;
1260         u32 nritems;
1261         unsigned int data_end;
1262         unsigned int old_data;
1263         unsigned int old_size;
1264         int i;
1265
1266         slot_orig = path->slots[0];
1267         leaf_buf = path->nodes[0];
1268         leaf = btrfs_buffer_leaf(leaf_buf);
1269
1270         nritems = btrfs_header_nritems(&leaf->header);
1271         data_end = leaf_data_end(root, leaf);
1272
1273         if (btrfs_leaf_free_space(root, leaf) < data_size)
1274                 BUG();
1275         slot = path->slots[0];
1276         old_data = btrfs_item_end(leaf->items + slot);
1277
1278         BUG_ON(slot < 0);
1279         BUG_ON(slot >= nritems);
1280
1281         /*
1282          * item0..itemN ... dataN.offset..dataN.size .. data0.size
1283          */
1284         /* first correct the data pointers */
1285         for (i = slot; i < nritems; i++) {
1286                 u32 ioff = btrfs_item_offset(leaf->items + i);
1287                 btrfs_set_item_offset(leaf->items + i,
1288                                       ioff - data_size);
1289         }
1290         /* shift the data */
1291         btrfs_memmove(root, leaf, btrfs_leaf_data(leaf) +
1292                       data_end - data_size, btrfs_leaf_data(leaf) +
1293                       data_end, old_data - data_end);
1294         data_end = old_data;
1295         old_size = btrfs_item_size(leaf->items + slot);
1296         btrfs_set_item_size(leaf->items + slot, old_size + data_size);
1297         btrfs_mark_buffer_dirty(leaf_buf);
1298
1299         ret = 0;
1300         if (btrfs_leaf_free_space(root, leaf) < 0)
1301                 BUG();
1302         check_leaf(root, path, 0);
1303         return ret;
1304 }
1305
1306 /*
1307  * Given a key and some data, insert an item into the tree.
1308  * This does all the path init required, making room in the tree if needed.
1309  */
1310 int btrfs_insert_empty_item(struct btrfs_trans_handle *trans, struct btrfs_root
1311                             *root, struct btrfs_path *path, struct btrfs_key
1312                             *cpu_key, u32 data_size)
1313 {
1314         int ret = 0;
1315         int slot;
1316         int slot_orig;
1317         struct btrfs_leaf *leaf;
1318         struct buffer_head *leaf_buf;
1319         u32 nritems;
1320         unsigned int data_end;
1321         struct btrfs_disk_key disk_key;
1322
1323         btrfs_cpu_key_to_disk(&disk_key, cpu_key);
1324
1325         /* create a root if there isn't one */
1326         if (!root->node)
1327                 BUG();
1328         ret = btrfs_search_slot(trans, root, cpu_key, path, data_size, 1);
1329         if (ret == 0) {
1330                 return -EEXIST;
1331         }
1332         if (ret < 0)
1333                 goto out;
1334
1335         slot_orig = path->slots[0];
1336         leaf_buf = path->nodes[0];
1337         leaf = btrfs_buffer_leaf(leaf_buf);
1338
1339         nritems = btrfs_header_nritems(&leaf->header);
1340         data_end = leaf_data_end(root, leaf);
1341
1342         if (btrfs_leaf_free_space(root, leaf) <
1343             sizeof(struct btrfs_item) + data_size) {
1344                 BUG();
1345         }
1346         slot = path->slots[0];
1347         BUG_ON(slot < 0);
1348         if (slot != nritems) {
1349                 int i;
1350                 unsigned int old_data = btrfs_item_end(leaf->items + slot);
1351
1352                 /*
1353                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
1354                  */
1355                 /* first correct the data pointers */
1356                 for (i = slot; i < nritems; i++) {
1357                         u32 ioff = btrfs_item_offset(leaf->items + i);
1358                         btrfs_set_item_offset(leaf->items + i,
1359                                               ioff - data_size);
1360                 }
1361
1362                 /* shift the items */
1363                 btrfs_memmove(root, leaf, leaf->items + slot + 1,
1364                               leaf->items + slot,
1365                               (nritems - slot) * sizeof(struct btrfs_item));
1366
1367                 /* shift the data */
1368                 btrfs_memmove(root, leaf, btrfs_leaf_data(leaf) +
1369                               data_end - data_size, btrfs_leaf_data(leaf) +
1370                               data_end, old_data - data_end);
1371                 data_end = old_data;
1372         }
1373         /* setup the item for the new data */
1374         btrfs_memcpy(root, leaf, &leaf->items[slot].key, &disk_key,
1375                      sizeof(struct btrfs_disk_key));
1376         btrfs_set_item_offset(leaf->items + slot, data_end - data_size);
1377         btrfs_set_item_size(leaf->items + slot, data_size);
1378         btrfs_set_header_nritems(&leaf->header, nritems + 1);
1379         btrfs_mark_buffer_dirty(leaf_buf);
1380
1381         ret = 0;
1382         if (slot == 0)
1383                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
1384
1385         if (btrfs_leaf_free_space(root, leaf) < 0)
1386                 BUG();
1387         check_leaf(root, path, 0);
1388 out:
1389         return ret;
1390 }
1391
1392 /*
1393  * Given a key and some data, insert an item into the tree.
1394  * This does all the path init required, making room in the tree if needed.
1395  */
1396 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
1397                       *root, struct btrfs_key *cpu_key, void *data, u32
1398                       data_size)
1399 {
1400         int ret = 0;
1401         struct btrfs_path *path;
1402         u8 *ptr;
1403
1404         path = btrfs_alloc_path();
1405         BUG_ON(!path);
1406         btrfs_init_path(path);
1407         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
1408         if (!ret) {
1409                 ptr = btrfs_item_ptr(btrfs_buffer_leaf(path->nodes[0]),
1410                                      path->slots[0], u8);
1411                 btrfs_memcpy(root, path->nodes[0]->b_data,
1412                              ptr, data, data_size);
1413                 btrfs_mark_buffer_dirty(path->nodes[0]);
1414         }
1415         btrfs_release_path(root, path);
1416         btrfs_free_path(path);
1417         return ret;
1418 }
1419
1420 /*
1421  * delete the pointer from a given node.
1422  *
1423  * If the delete empties a node, the node is removed from the tree,
1424  * continuing all the way the root if required.  The root is converted into
1425  * a leaf if all the nodes are emptied.
1426  */
1427 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1428                    struct btrfs_path *path, int level, int slot)
1429 {
1430         struct btrfs_node *node;
1431         struct buffer_head *parent = path->nodes[level];
1432         u32 nritems;
1433         int ret = 0;
1434         int wret;
1435
1436         node = btrfs_buffer_node(parent);
1437         nritems = btrfs_header_nritems(&node->header);
1438         if (slot != nritems -1) {
1439                 btrfs_memmove(root, node, node->ptrs + slot,
1440                               node->ptrs + slot + 1,
1441                               sizeof(struct btrfs_key_ptr) *
1442                               (nritems - slot - 1));
1443         }
1444         nritems--;
1445         btrfs_set_header_nritems(&node->header, nritems);
1446         if (nritems == 0 && parent == root->node) {
1447                 struct btrfs_header *header = btrfs_buffer_header(root->node);
1448                 BUG_ON(btrfs_header_level(header) != 1);
1449                 /* just turn the root into a leaf and break */
1450                 btrfs_set_header_level(header, 0);
1451         } else if (slot == 0) {
1452                 wret = fixup_low_keys(trans, root, path, &node->ptrs[0].key,
1453                                       level + 1);
1454                 if (wret)
1455                         ret = wret;
1456         }
1457         btrfs_mark_buffer_dirty(parent);
1458         return ret;
1459 }
1460
1461 /*
1462  * delete the item at the leaf level in path.  If that empties
1463  * the leaf, remove it from the tree
1464  */
1465 int btrfs_del_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1466                    struct btrfs_path *path)
1467 {
1468         int slot;
1469         struct btrfs_leaf *leaf;
1470         struct buffer_head *leaf_buf;
1471         int doff;
1472         int dsize;
1473         int ret = 0;
1474         int wret;
1475         u32 nritems;
1476
1477         leaf_buf = path->nodes[0];
1478         leaf = btrfs_buffer_leaf(leaf_buf);
1479         slot = path->slots[0];
1480         doff = btrfs_item_offset(leaf->items + slot);
1481         dsize = btrfs_item_size(leaf->items + slot);
1482         nritems = btrfs_header_nritems(&leaf->header);
1483
1484         if (slot != nritems - 1) {
1485                 int i;
1486                 int data_end = leaf_data_end(root, leaf);
1487                 btrfs_memmove(root, leaf, btrfs_leaf_data(leaf) +
1488                               data_end + dsize,
1489                               btrfs_leaf_data(leaf) + data_end,
1490                               doff - data_end);
1491                 for (i = slot + 1; i < nritems; i++) {
1492                         u32 ioff = btrfs_item_offset(leaf->items + i);
1493                         btrfs_set_item_offset(leaf->items + i, ioff + dsize);
1494                 }
1495                 btrfs_memmove(root, leaf, leaf->items + slot,
1496                               leaf->items + slot + 1,
1497                               sizeof(struct btrfs_item) *
1498                               (nritems - slot - 1));
1499         }
1500         btrfs_set_header_nritems(&leaf->header, nritems - 1);
1501         nritems--;
1502         /* delete the leaf if we've emptied it */
1503         if (nritems == 0) {
1504                 if (leaf_buf == root->node) {
1505                         btrfs_set_header_level(&leaf->header, 0);
1506                 } else {
1507                         clean_tree_block(trans, root, leaf_buf);
1508                         wait_on_buffer(leaf_buf);
1509                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
1510                         if (wret)
1511                                 ret = wret;
1512                         wret = btrfs_free_extent(trans, root,
1513                                                  bh_blocknr(leaf_buf), 1, 1);
1514                         if (wret)
1515                                 ret = wret;
1516                 }
1517         } else {
1518                 int used = leaf_space_used(leaf, 0, nritems);
1519                 if (slot == 0) {
1520                         wret = fixup_low_keys(trans, root, path,
1521                                               &leaf->items[0].key, 1);
1522                         if (wret)
1523                                 ret = wret;
1524                 }
1525
1526                 /* delete the leaf if it is mostly empty */
1527                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 3) {
1528                         /* push_leaf_left fixes the path.
1529                          * make sure the path still points to our leaf
1530                          * for possible call to del_ptr below
1531                          */
1532                         slot = path->slots[1];
1533                         get_bh(leaf_buf);
1534                         wret = push_leaf_left(trans, root, path, 1);
1535                         if (wret < 0)
1536                                 ret = wret;
1537                         if (path->nodes[0] == leaf_buf &&
1538                             btrfs_header_nritems(&leaf->header)) {
1539                                 wret = push_leaf_right(trans, root, path, 1);
1540                                 if (wret < 0)
1541                                         ret = wret;
1542                         }
1543                         if (btrfs_header_nritems(&leaf->header) == 0) {
1544                                 u64 blocknr = bh_blocknr(leaf_buf);
1545                                 clean_tree_block(trans, root, leaf_buf);
1546                                 wait_on_buffer(leaf_buf);
1547                                 wret = del_ptr(trans, root, path, 1, slot);
1548                                 if (wret)
1549                                         ret = wret;
1550                                 btrfs_block_release(root, leaf_buf);
1551                                 wret = btrfs_free_extent(trans, root, blocknr,
1552                                                          1, 1);
1553                                 if (wret)
1554                                         ret = wret;
1555                         } else {
1556                                 btrfs_mark_buffer_dirty(leaf_buf);
1557                                 btrfs_block_release(root, leaf_buf);
1558                         }
1559                 } else {
1560                         btrfs_mark_buffer_dirty(leaf_buf);
1561                 }
1562         }
1563         return ret;
1564 }
1565
1566 /*
1567  * walk up the tree as far as required to find the next leaf.
1568  * returns 0 if it found something or 1 if there are no greater leaves.
1569  * returns < 0 on io errors.
1570  */
1571 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
1572 {
1573         int slot;
1574         int level = 1;
1575         u64 blocknr;
1576         struct buffer_head *c;
1577         struct btrfs_node *c_node;
1578         struct buffer_head *next = NULL;
1579
1580         while(level < BTRFS_MAX_LEVEL) {
1581                 if (!path->nodes[level])
1582                         return 1;
1583                 slot = path->slots[level] + 1;
1584                 c = path->nodes[level];
1585                 c_node = btrfs_buffer_node(c);
1586                 if (slot >= btrfs_header_nritems(&c_node->header)) {
1587                         level++;
1588                         continue;
1589                 }
1590                 blocknr = btrfs_node_blockptr(c_node, slot);
1591                 if (next)
1592                         btrfs_block_release(root, next);
1593                 next = read_tree_block(root, blocknr);
1594                 break;
1595         }
1596         path->slots[level] = slot;
1597         while(1) {
1598                 level--;
1599                 c = path->nodes[level];
1600                 btrfs_block_release(root, c);
1601                 path->nodes[level] = next;
1602                 path->slots[level] = 0;
1603                 if (!level)
1604                         break;
1605                 next = read_tree_block(root,
1606                        btrfs_node_blockptr(btrfs_buffer_node(next), 0));
1607         }
1608         return 0;
1609 }