merge leaves before split
[btrfs-progs-unstable.git] / ctree.c
blobdc620186e501cdaa206f6e6dc27d15131fd5bf8e
1 #include <stdio.h>
2 #include <stdlib.h>
3 #include "kerncompat.h"
4 #include "radix-tree.h"
5 #include "ctree.h"
6 #include "disk-io.h"
7 #include "print-tree.h"
9 static int split_node(struct btrfs_root *root, struct btrfs_path *path,
10 int level);
11 static int split_leaf(struct btrfs_root *root, struct btrfs_path *path,
12 int data_size);
13 static int push_node_left(struct btrfs_root *root, struct btrfs_buffer *dst,
14 struct btrfs_buffer *src);
15 static int balance_node_right(struct btrfs_root *root,
16 struct btrfs_buffer *dst_buf,
17 struct btrfs_buffer *src_buf);
18 static int del_ptr(struct btrfs_root *root, struct btrfs_path *path, int level,
19 int slot);
21 inline void btrfs_init_path(struct btrfs_path *p)
23 memset(p, 0, sizeof(*p));
26 void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
28 int i;
29 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
30 if (!p->nodes[i])
31 break;
32 btrfs_block_release(root, p->nodes[i]);
34 memset(p, 0, sizeof(*p));
37 static int btrfs_cow_block(struct btrfs_root *root,
38 struct btrfs_buffer *buf,
39 struct btrfs_buffer *parent,
40 int parent_slot,
41 struct btrfs_buffer **cow_ret)
43 struct btrfs_buffer *cow;
45 if (!list_empty(&buf->dirty)) {
46 *cow_ret = buf;
47 return 0;
49 cow = btrfs_alloc_free_block(root);
50 memcpy(&cow->node, &buf->node, sizeof(buf->node));
51 btrfs_set_header_blocknr(&cow->node.header, cow->blocknr);
52 *cow_ret = cow;
53 btrfs_inc_ref(root, buf);
54 if (buf == root->node) {
55 root->node = cow;
56 cow->count++;
57 if (buf != root->commit_root)
58 btrfs_free_extent(root, buf->blocknr, 1);
59 btrfs_block_release(root, buf);
60 } else {
61 btrfs_set_node_blockptr(&parent->node, parent_slot,
62 cow->blocknr);
63 BUG_ON(list_empty(&parent->dirty));
64 btrfs_free_extent(root, buf->blocknr, 1);
66 btrfs_block_release(root, buf);
67 return 0;
71 * The leaf data grows from end-to-front in the node.
72 * this returns the address of the start of the last item,
73 * which is the stop of the leaf data stack
75 static inline unsigned int leaf_data_end(struct btrfs_leaf *leaf)
77 u32 nr = btrfs_header_nritems(&leaf->header);
78 if (nr == 0)
79 return sizeof(leaf->data);
80 return btrfs_item_offset(leaf->items + nr - 1);
84 * The space between the end of the leaf items and
85 * the start of the leaf data. IOW, how much room
86 * the leaf has left for both items and data
88 int btrfs_leaf_free_space(struct btrfs_leaf *leaf)
90 int data_end = leaf_data_end(leaf);
91 int nritems = btrfs_header_nritems(&leaf->header);
92 char *items_end = (char *)(leaf->items + nritems + 1);
93 return (char *)(leaf->data + data_end) - (char *)items_end;
97 * compare two keys in a memcmp fashion
99 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
101 struct btrfs_key k1;
103 btrfs_disk_key_to_cpu(&k1, disk);
105 if (k1.objectid > k2->objectid)
106 return 1;
107 if (k1.objectid < k2->objectid)
108 return -1;
109 if (k1.flags > k2->flags)
110 return 1;
111 if (k1.flags < k2->flags)
112 return -1;
113 if (k1.offset > k2->offset)
114 return 1;
115 if (k1.offset < k2->offset)
116 return -1;
117 return 0;
120 static int check_node(struct btrfs_path *path, int level)
122 int i;
123 struct btrfs_node *parent = NULL;
124 struct btrfs_node *node = &path->nodes[level]->node;
125 int parent_slot;
126 u32 nritems = btrfs_header_nritems(&node->header);
128 if (path->nodes[level + 1])
129 parent = &path->nodes[level + 1]->node;
130 parent_slot = path->slots[level + 1];
131 BUG_ON(nritems == 0);
132 if (parent) {
133 struct btrfs_disk_key *parent_key;
134 parent_key = &parent->keys[parent_slot];
135 BUG_ON(memcmp(parent_key, node->keys,
136 sizeof(struct btrfs_disk_key)));
137 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
138 btrfs_header_blocknr(&node->header));
140 BUG_ON(nritems > NODEPTRS_PER_BLOCK);
141 for (i = 0; nritems > 1 && i < nritems - 2; i++) {
142 struct btrfs_key cpukey;
143 btrfs_disk_key_to_cpu(&cpukey, &node->keys[i + 1]);
144 BUG_ON(comp_keys(&node->keys[i], &cpukey) >= 0);
146 return 0;
149 static int check_leaf(struct btrfs_path *path, int level)
151 int i;
152 struct btrfs_leaf *leaf = &path->nodes[level]->leaf;
153 struct btrfs_node *parent = NULL;
154 int parent_slot;
155 u32 nritems = btrfs_header_nritems(&leaf->header);
157 if (path->nodes[level + 1])
158 parent = &path->nodes[level + 1]->node;
159 parent_slot = path->slots[level + 1];
160 BUG_ON(btrfs_leaf_free_space(leaf) < 0);
162 if (nritems == 0)
163 return 0;
165 if (parent) {
166 struct btrfs_disk_key *parent_key;
167 parent_key = &parent->keys[parent_slot];
168 BUG_ON(memcmp(parent_key, &leaf->items[0].key,
169 sizeof(struct btrfs_disk_key)));
170 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
171 btrfs_header_blocknr(&leaf->header));
173 for (i = 0; nritems > 1 && i < nritems - 2; i++) {
174 struct btrfs_key cpukey;
175 btrfs_disk_key_to_cpu(&cpukey, &leaf->items[i + 1].key);
176 BUG_ON(comp_keys(&leaf->items[i].key,
177 &cpukey) >= 0);
178 BUG_ON(btrfs_item_offset(leaf->items + i) !=
179 btrfs_item_end(leaf->items + i + 1));
180 if (i == 0) {
181 BUG_ON(btrfs_item_offset(leaf->items + i) +
182 btrfs_item_size(leaf->items + i) !=
183 LEAF_DATA_SIZE);
186 return 0;
189 static int check_block(struct btrfs_path *path, int level)
191 if (level == 0)
192 return check_leaf(path, level);
193 return check_node(path, level);
197 * search for key in the array p. items p are item_size apart
198 * and there are 'max' items in p
199 * the slot in the array is returned via slot, and it points to
200 * the place where you would insert key if it is not found in
201 * the array.
203 * slot may point to max if the key is bigger than all of the keys
205 static int generic_bin_search(char *p, int item_size, struct btrfs_key *key,
206 int max, int *slot)
208 int low = 0;
209 int high = max;
210 int mid;
211 int ret;
212 struct btrfs_disk_key *tmp;
214 while(low < high) {
215 mid = (low + high) / 2;
216 tmp = (struct btrfs_disk_key *)(p + mid * item_size);
217 ret = comp_keys(tmp, key);
219 if (ret < 0)
220 low = mid + 1;
221 else if (ret > 0)
222 high = mid;
223 else {
224 *slot = mid;
225 return 0;
228 *slot = low;
229 return 1;
233 * simple bin_search frontend that does the right thing for
234 * leaves vs nodes
236 static int bin_search(struct btrfs_node *c, struct btrfs_key *key, int *slot)
238 if (btrfs_is_leaf(c)) {
239 struct btrfs_leaf *l = (struct btrfs_leaf *)c;
240 return generic_bin_search((void *)l->items,
241 sizeof(struct btrfs_item),
242 key, btrfs_header_nritems(&c->header),
243 slot);
244 } else {
245 return generic_bin_search((void *)c->keys,
246 sizeof(struct btrfs_disk_key),
247 key, btrfs_header_nritems(&c->header),
248 slot);
250 return -1;
253 static struct btrfs_buffer *read_node_slot(struct btrfs_root *root,
254 struct btrfs_buffer *parent_buf,
255 int slot)
257 struct btrfs_node *node = &parent_buf->node;
258 if (slot < 0)
259 return NULL;
260 if (slot >= btrfs_header_nritems(&node->header))
261 return NULL;
262 return read_tree_block(root, btrfs_node_blockptr(node, slot));
265 static int balance_level(struct btrfs_root *root, struct btrfs_path *path,
266 int level)
268 struct btrfs_buffer *right_buf;
269 struct btrfs_buffer *mid_buf;
270 struct btrfs_buffer *left_buf;
271 struct btrfs_buffer *parent_buf = NULL;
272 struct btrfs_node *right = NULL;
273 struct btrfs_node *mid;
274 struct btrfs_node *left = NULL;
275 struct btrfs_node *parent = NULL;
276 int ret = 0;
277 int wret;
278 int pslot;
279 int orig_slot = path->slots[level];
280 u64 orig_ptr;
282 if (level == 0)
283 return 0;
285 mid_buf = path->nodes[level];
286 mid = &mid_buf->node;
287 orig_ptr = btrfs_node_blockptr(mid, orig_slot);
289 if (level < BTRFS_MAX_LEVEL - 1)
290 parent_buf = path->nodes[level + 1];
291 pslot = path->slots[level + 1];
293 if (!parent_buf) {
294 struct btrfs_buffer *child;
295 u64 blocknr = mid_buf->blocknr;
297 if (btrfs_header_nritems(&mid->header) != 1)
298 return 0;
300 /* promote the child to a root */
301 child = read_node_slot(root, mid_buf, 0);
302 BUG_ON(!child);
303 root->node = child;
304 path->nodes[level] = NULL;
305 /* once for the path */
306 btrfs_block_release(root, mid_buf);
307 /* once for the root ptr */
308 btrfs_block_release(root, mid_buf);
309 clean_tree_block(root, mid_buf);
310 return btrfs_free_extent(root, blocknr, 1);
312 parent = &parent_buf->node;
314 if (btrfs_header_nritems(&mid->header) > NODEPTRS_PER_BLOCK / 4)
315 return 0;
317 left_buf = read_node_slot(root, parent_buf, pslot - 1);
318 right_buf = read_node_slot(root, parent_buf, pslot + 1);
320 /* first, try to make some room in the middle buffer */
321 if (left_buf) {
322 btrfs_cow_block(root, left_buf, parent_buf,
323 pslot - 1, &left_buf);
324 left = &left_buf->node;
325 orig_slot += btrfs_header_nritems(&left->header);
326 wret = push_node_left(root, left_buf, mid_buf);
327 if (wret < 0)
328 ret = wret;
332 * then try to empty the right most buffer into the middle
334 if (right_buf) {
335 btrfs_cow_block(root, right_buf, parent_buf,
336 pslot + 1, &right_buf);
337 right = &right_buf->node;
338 wret = push_node_left(root, mid_buf, right_buf);
339 if (wret < 0)
340 ret = wret;
341 if (btrfs_header_nritems(&right->header) == 0) {
342 u64 blocknr = right_buf->blocknr;
343 btrfs_block_release(root, right_buf);
344 clean_tree_block(root, right_buf);
345 right_buf = NULL;
346 right = NULL;
347 wret = del_ptr(root, path, level + 1, pslot + 1);
348 if (wret)
349 ret = wret;
350 wret = btrfs_free_extent(root, blocknr, 1);
351 if (wret)
352 ret = wret;
353 } else {
354 memcpy(parent->keys + pslot + 1, right->keys,
355 sizeof(struct btrfs_disk_key));
356 BUG_ON(list_empty(&parent_buf->dirty));
359 if (btrfs_header_nritems(&mid->header) == 1) {
361 * we're not allowed to leave a node with one item in the
362 * tree during a delete. A deletion from lower in the tree
363 * could try to delete the only pointer in this node.
364 * So, pull some keys from the left.
365 * There has to be a left pointer at this point because
366 * otherwise we would have pulled some pointers from the
367 * right
369 BUG_ON(!left_buf);
370 wret = balance_node_right(root, mid_buf, left_buf);
371 if (wret < 0)
372 ret = wret;
373 BUG_ON(wret == 1);
375 if (btrfs_header_nritems(&mid->header) == 0) {
376 /* we've managed to empty the middle node, drop it */
377 u64 blocknr = mid_buf->blocknr;
378 btrfs_block_release(root, mid_buf);
379 clean_tree_block(root, mid_buf);
380 mid_buf = NULL;
381 mid = NULL;
382 wret = del_ptr(root, path, level + 1, pslot);
383 if (wret)
384 ret = wret;
385 wret = btrfs_free_extent(root, blocknr, 1);
386 if (wret)
387 ret = wret;
388 } else {
389 /* update the parent key to reflect our changes */
390 memcpy(parent->keys + pslot, mid->keys,
391 sizeof(struct btrfs_disk_key));
392 BUG_ON(list_empty(&parent_buf->dirty));
395 /* update the path */
396 if (left_buf) {
397 if (btrfs_header_nritems(&left->header) > orig_slot) {
398 left_buf->count++; // released below
399 path->nodes[level] = left_buf;
400 path->slots[level + 1] -= 1;
401 path->slots[level] = orig_slot;
402 if (mid_buf)
403 btrfs_block_release(root, mid_buf);
404 } else {
405 orig_slot -= btrfs_header_nritems(&left->header);
406 path->slots[level] = orig_slot;
409 /* double check we haven't messed things up */
410 check_block(path, level);
411 if (orig_ptr != btrfs_node_blockptr(&path->nodes[level]->node,
412 path->slots[level]))
413 BUG();
415 if (right_buf)
416 btrfs_block_release(root, right_buf);
417 if (left_buf)
418 btrfs_block_release(root, left_buf);
419 return ret;
423 * look for key in the tree. path is filled in with nodes along the way
424 * if key is found, we return zero and you can find the item in the leaf
425 * level of the path (level 0)
427 * If the key isn't found, the path points to the slot where it should
428 * be inserted, and 1 is returned. If there are other errors during the
429 * search a negative error number is returned.
431 * if ins_len > 0, nodes and leaves will be split as we walk down the
432 * tree. if ins_len < 0, nodes will be merged as we walk down the tree (if
433 * possible)
435 int btrfs_search_slot(struct btrfs_root *root, struct btrfs_key *key,
436 struct btrfs_path *p, int ins_len, int cow)
438 struct btrfs_buffer *b;
439 struct btrfs_buffer *cow_buf;
440 struct btrfs_node *c;
441 int slot;
442 int ret;
443 int level;
445 again:
446 b = root->node;
447 b->count++;
448 while (b) {
449 level = btrfs_header_level(&b->node.header);
450 if (cow) {
451 int wret;
452 wret = btrfs_cow_block(root, b, p->nodes[level + 1],
453 p->slots[level + 1], &cow_buf);
454 b = cow_buf;
456 BUG_ON(!cow && ins_len);
457 c = &b->node;
458 p->nodes[level] = b;
459 ret = check_block(p, level);
460 if (ret)
461 return -1;
462 ret = bin_search(c, key, &slot);
463 if (!btrfs_is_leaf(c)) {
464 if (ret && slot > 0)
465 slot -= 1;
466 p->slots[level] = slot;
467 if (ins_len > 0 && btrfs_header_nritems(&c->header) ==
468 NODEPTRS_PER_BLOCK) {
469 int sret = split_node(root, p, level);
470 BUG_ON(sret > 0);
471 if (sret)
472 return sret;
473 b = p->nodes[level];
474 c = &b->node;
475 slot = p->slots[level];
476 } else if (ins_len < 0) {
477 int sret = balance_level(root, p, level);
478 if (sret)
479 return sret;
480 b = p->nodes[level];
481 if (!b)
482 goto again;
483 c = &b->node;
484 slot = p->slots[level];
485 BUG_ON(btrfs_header_nritems(&c->header) == 1);
487 b = read_tree_block(root, btrfs_node_blockptr(c, slot));
488 } else {
489 struct btrfs_leaf *l = (struct btrfs_leaf *)c;
490 p->slots[level] = slot;
491 if (ins_len > 0 && btrfs_leaf_free_space(l) <
492 sizeof(struct btrfs_item) + ins_len) {
493 int sret = split_leaf(root, p, ins_len);
494 BUG_ON(sret > 0);
495 if (sret)
496 return sret;
498 BUG_ON(root->node->count == 1);
499 return ret;
502 BUG_ON(root->node->count == 1);
503 return 1;
507 * adjust the pointers going up the tree, starting at level
508 * making sure the right key of each node is points to 'key'.
509 * This is used after shifting pointers to the left, so it stops
510 * fixing up pointers when a given leaf/node is not in slot 0 of the
511 * higher levels
513 * If this fails to write a tree block, it returns -1, but continues
514 * fixing up the blocks in ram so the tree is consistent.
516 static int fixup_low_keys(struct btrfs_root *root,
517 struct btrfs_path *path, struct btrfs_disk_key *key,
518 int level)
520 int i;
521 int ret = 0;
522 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
523 struct btrfs_node *t;
524 int tslot = path->slots[i];
525 if (!path->nodes[i])
526 break;
527 t = &path->nodes[i]->node;
528 memcpy(t->keys + tslot, key, sizeof(*key));
529 BUG_ON(list_empty(&path->nodes[i]->dirty));
530 if (tslot != 0)
531 break;
533 return ret;
537 * try to push data from one node into the next node left in the
538 * tree.
540 * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
541 * error, and > 0 if there was no room in the left hand block.
543 static int push_node_left(struct btrfs_root *root, struct btrfs_buffer *dst_buf,
544 struct btrfs_buffer *src_buf)
546 struct btrfs_node *src = &src_buf->node;
547 struct btrfs_node *dst = &dst_buf->node;
548 int push_items = 0;
549 int src_nritems;
550 int dst_nritems;
551 int ret = 0;
553 src_nritems = btrfs_header_nritems(&src->header);
554 dst_nritems = btrfs_header_nritems(&dst->header);
555 push_items = NODEPTRS_PER_BLOCK - dst_nritems;
556 if (push_items <= 0) {
557 return 1;
560 if (src_nritems < push_items)
561 push_items = src_nritems;
563 memcpy(dst->keys + dst_nritems, src->keys,
564 push_items * sizeof(struct btrfs_disk_key));
565 memcpy(dst->blockptrs + dst_nritems, src->blockptrs,
566 push_items * sizeof(u64));
567 if (push_items < src_nritems) {
568 memmove(src->keys, src->keys + push_items,
569 (src_nritems - push_items) *
570 sizeof(struct btrfs_disk_key));
571 memmove(src->blockptrs, src->blockptrs + push_items,
572 (src_nritems - push_items) * sizeof(u64));
574 btrfs_set_header_nritems(&src->header, src_nritems - push_items);
575 btrfs_set_header_nritems(&dst->header, dst_nritems + push_items);
576 BUG_ON(list_empty(&src_buf->dirty));
577 BUG_ON(list_empty(&dst_buf->dirty));
578 return ret;
582 * try to push data from one node into the next node right in the
583 * tree.
585 * returns 0 if some ptrs were pushed, < 0 if there was some horrible
586 * error, and > 0 if there was no room in the right hand block.
588 * this will only push up to 1/2 the contents of the left node over
590 static int balance_node_right(struct btrfs_root *root,
591 struct btrfs_buffer *dst_buf,
592 struct btrfs_buffer *src_buf)
594 struct btrfs_node *src = &src_buf->node;
595 struct btrfs_node *dst = &dst_buf->node;
596 int push_items = 0;
597 int max_push;
598 int src_nritems;
599 int dst_nritems;
600 int ret = 0;
602 src_nritems = btrfs_header_nritems(&src->header);
603 dst_nritems = btrfs_header_nritems(&dst->header);
604 push_items = NODEPTRS_PER_BLOCK - dst_nritems;
605 if (push_items <= 0) {
606 return 1;
609 max_push = src_nritems / 2 + 1;
610 /* don't try to empty the node */
611 if (max_push > src_nritems)
612 return 1;
613 if (max_push < push_items)
614 push_items = max_push;
616 memmove(dst->keys + push_items, dst->keys,
617 dst_nritems * sizeof(struct btrfs_disk_key));
618 memmove(dst->blockptrs + push_items, dst->blockptrs,
619 dst_nritems * sizeof(u64));
620 memcpy(dst->keys, src->keys + src_nritems - push_items,
621 push_items * sizeof(struct btrfs_disk_key));
622 memcpy(dst->blockptrs, src->blockptrs + src_nritems - push_items,
623 push_items * sizeof(u64));
625 btrfs_set_header_nritems(&src->header, src_nritems - push_items);
626 btrfs_set_header_nritems(&dst->header, dst_nritems + push_items);
628 BUG_ON(list_empty(&src_buf->dirty));
629 BUG_ON(list_empty(&dst_buf->dirty));
630 return ret;
634 * helper function to insert a new root level in the tree.
635 * A new node is allocated, and a single item is inserted to
636 * point to the existing root
638 * returns zero on success or < 0 on failure.
640 static int insert_new_root(struct btrfs_root *root,
641 struct btrfs_path *path, int level)
643 struct btrfs_buffer *t;
644 struct btrfs_node *lower;
645 struct btrfs_node *c;
646 struct btrfs_disk_key *lower_key;
648 BUG_ON(path->nodes[level]);
649 BUG_ON(path->nodes[level-1] != root->node);
651 t = btrfs_alloc_free_block(root);
652 c = &t->node;
653 memset(c, 0, sizeof(c));
654 btrfs_set_header_nritems(&c->header, 1);
655 btrfs_set_header_level(&c->header, level);
656 btrfs_set_header_blocknr(&c->header, t->blocknr);
657 btrfs_set_header_parentid(&c->header,
658 btrfs_header_parentid(&root->node->node.header));
659 lower = &path->nodes[level-1]->node;
660 if (btrfs_is_leaf(lower))
661 lower_key = &((struct btrfs_leaf *)lower)->items[0].key;
662 else
663 lower_key = lower->keys;
664 memcpy(c->keys, lower_key, sizeof(struct btrfs_disk_key));
665 btrfs_set_node_blockptr(c, 0, path->nodes[level - 1]->blocknr);
666 /* the super has an extra ref to root->node */
667 btrfs_block_release(root, root->node);
668 root->node = t;
669 t->count++;
670 path->nodes[level] = t;
671 path->slots[level] = 0;
672 return 0;
676 * worker function to insert a single pointer in a node.
677 * the node should have enough room for the pointer already
679 * slot and level indicate where you want the key to go, and
680 * blocknr is the block the key points to.
682 * returns zero on success and < 0 on any error
684 static int insert_ptr(struct btrfs_root *root,
685 struct btrfs_path *path, struct btrfs_disk_key *key,
686 u64 blocknr, int slot, int level)
688 struct btrfs_node *lower;
689 int nritems;
691 BUG_ON(!path->nodes[level]);
692 lower = &path->nodes[level]->node;
693 nritems = btrfs_header_nritems(&lower->header);
694 if (slot > nritems)
695 BUG();
696 if (nritems == NODEPTRS_PER_BLOCK)
697 BUG();
698 if (slot != nritems) {
699 memmove(lower->keys + slot + 1, lower->keys + slot,
700 (nritems - slot) * sizeof(struct btrfs_disk_key));
701 memmove(lower->blockptrs + slot + 1, lower->blockptrs + slot,
702 (nritems - slot) * sizeof(u64));
704 memcpy(lower->keys + slot, key, sizeof(struct btrfs_disk_key));
705 btrfs_set_node_blockptr(lower, slot, blocknr);
706 btrfs_set_header_nritems(&lower->header, nritems + 1);
707 if (lower->keys[1].objectid == 0)
708 BUG();
709 BUG_ON(list_empty(&path->nodes[level]->dirty));
710 return 0;
714 * split the node at the specified level in path in two.
715 * The path is corrected to point to the appropriate node after the split
717 * Before splitting this tries to make some room in the node by pushing
718 * left and right, if either one works, it returns right away.
720 * returns 0 on success and < 0 on failure
722 static int split_node(struct btrfs_root *root, struct btrfs_path *path,
723 int level)
725 struct btrfs_buffer *t;
726 struct btrfs_node *c;
727 struct btrfs_buffer *split_buffer;
728 struct btrfs_node *split;
729 int mid;
730 int ret;
731 int wret;
732 u32 c_nritems;
734 t = path->nodes[level];
735 c = &t->node;
736 if (t == root->node) {
737 /* trying to split the root, lets make a new one */
738 ret = insert_new_root(root, path, level + 1);
739 if (ret)
740 return ret;
742 c_nritems = btrfs_header_nritems(&c->header);
743 split_buffer = btrfs_alloc_free_block(root);
744 split = &split_buffer->node;
745 btrfs_set_header_flags(&split->header, btrfs_header_flags(&c->header));
746 btrfs_set_header_blocknr(&split->header, split_buffer->blocknr);
747 btrfs_set_header_parentid(&split->header,
748 btrfs_header_parentid(&root->node->node.header));
749 mid = (c_nritems + 1) / 2;
750 memcpy(split->keys, c->keys + mid,
751 (c_nritems - mid) * sizeof(struct btrfs_disk_key));
752 memcpy(split->blockptrs, c->blockptrs + mid,
753 (c_nritems - mid) * sizeof(u64));
754 btrfs_set_header_nritems(&split->header, c_nritems - mid);
755 btrfs_set_header_nritems(&c->header, mid);
756 ret = 0;
758 BUG_ON(list_empty(&t->dirty));
759 wret = insert_ptr(root, path, split->keys, split_buffer->blocknr,
760 path->slots[level + 1] + 1, level + 1);
761 if (wret)
762 ret = wret;
764 if (path->slots[level] >= mid) {
765 path->slots[level] -= mid;
766 btrfs_block_release(root, t);
767 path->nodes[level] = split_buffer;
768 path->slots[level + 1] += 1;
769 } else {
770 btrfs_block_release(root, split_buffer);
772 return ret;
776 * how many bytes are required to store the items in a leaf. start
777 * and nr indicate which items in the leaf to check. This totals up the
778 * space used both by the item structs and the item data
780 static int leaf_space_used(struct btrfs_leaf *l, int start, int nr)
782 int data_len;
783 int end = start + nr - 1;
785 if (!nr)
786 return 0;
787 data_len = btrfs_item_end(l->items + start);
788 data_len = data_len - btrfs_item_offset(l->items + end);
789 data_len += sizeof(struct btrfs_item) * nr;
790 return data_len;
794 * push some data in the path leaf to the right, trying to free up at
795 * least data_size bytes. returns zero if the push worked, nonzero otherwise
797 * returns 1 if the push failed because the other node didn't have enough
798 * room, 0 if everything worked out and < 0 if there were major errors.
800 static int push_leaf_right(struct btrfs_root *root, struct btrfs_path *path,
801 int data_size)
803 struct btrfs_buffer *left_buf = path->nodes[0];
804 struct btrfs_leaf *left = &left_buf->leaf;
805 struct btrfs_leaf *right;
806 struct btrfs_buffer *right_buf;
807 struct btrfs_buffer *upper;
808 int slot;
809 int i;
810 int free_space;
811 int push_space = 0;
812 int push_items = 0;
813 struct btrfs_item *item;
814 u32 left_nritems;
815 u32 right_nritems;
817 slot = path->slots[1];
818 if (!path->nodes[1]) {
819 return 1;
821 upper = path->nodes[1];
822 if (slot >= btrfs_header_nritems(&upper->node.header) - 1) {
823 return 1;
825 right_buf = read_tree_block(root, btrfs_node_blockptr(&upper->node,
826 slot + 1));
827 right = &right_buf->leaf;
828 free_space = btrfs_leaf_free_space(right);
829 if (free_space < data_size + sizeof(struct btrfs_item)) {
830 btrfs_block_release(root, right_buf);
831 return 1;
833 /* cow and double check */
834 btrfs_cow_block(root, right_buf, upper, slot + 1, &right_buf);
835 right = &right_buf->leaf;
836 free_space = btrfs_leaf_free_space(right);
837 if (free_space < data_size + sizeof(struct btrfs_item)) {
838 btrfs_block_release(root, right_buf);
839 return 1;
842 left_nritems = btrfs_header_nritems(&left->header);
843 for (i = left_nritems - 1; i >= 0; i--) {
844 item = left->items + i;
845 if (path->slots[0] == i)
846 push_space += data_size + sizeof(*item);
847 if (btrfs_item_size(item) + sizeof(*item) + push_space >
848 free_space)
849 break;
850 push_items++;
851 push_space += btrfs_item_size(item) + sizeof(*item);
853 if (push_items == 0) {
854 btrfs_block_release(root, right_buf);
855 return 1;
857 right_nritems = btrfs_header_nritems(&right->header);
858 /* push left to right */
859 push_space = btrfs_item_end(left->items + left_nritems - push_items);
860 push_space -= leaf_data_end(left);
861 /* make room in the right data area */
862 memmove(right->data + leaf_data_end(right) - push_space,
863 right->data + leaf_data_end(right),
864 LEAF_DATA_SIZE - leaf_data_end(right));
865 /* copy from the left data area */
866 memcpy(right->data + LEAF_DATA_SIZE - push_space,
867 left->data + leaf_data_end(left),
868 push_space);
869 memmove(right->items + push_items, right->items,
870 right_nritems * sizeof(struct btrfs_item));
871 /* copy the items from left to right */
872 memcpy(right->items, left->items + left_nritems - push_items,
873 push_items * sizeof(struct btrfs_item));
875 /* update the item pointers */
876 right_nritems += push_items;
877 btrfs_set_header_nritems(&right->header, right_nritems);
878 push_space = LEAF_DATA_SIZE;
879 for (i = 0; i < right_nritems; i++) {
880 btrfs_set_item_offset(right->items + i, push_space -
881 btrfs_item_size(right->items + i));
882 push_space = btrfs_item_offset(right->items + i);
884 left_nritems -= push_items;
885 btrfs_set_header_nritems(&left->header, left_nritems);
887 BUG_ON(list_empty(&left_buf->dirty));
888 BUG_ON(list_empty(&right_buf->dirty));
889 memcpy(upper->node.keys + slot + 1,
890 &right->items[0].key, sizeof(struct btrfs_disk_key));
891 BUG_ON(list_empty(&upper->dirty));
893 /* then fixup the leaf pointer in the path */
894 if (path->slots[0] >= left_nritems) {
895 path->slots[0] -= left_nritems;
896 btrfs_block_release(root, path->nodes[0]);
897 path->nodes[0] = right_buf;
898 path->slots[1] += 1;
899 } else {
900 btrfs_block_release(root, right_buf);
902 return 0;
905 * push some data in the path leaf to the left, trying to free up at
906 * least data_size bytes. returns zero if the push worked, nonzero otherwise
908 static int push_leaf_left(struct btrfs_root *root, struct btrfs_path *path,
909 int data_size)
911 struct btrfs_buffer *right_buf = path->nodes[0];
912 struct btrfs_leaf *right = &right_buf->leaf;
913 struct btrfs_buffer *t;
914 struct btrfs_leaf *left;
915 int slot;
916 int i;
917 int free_space;
918 int push_space = 0;
919 int push_items = 0;
920 struct btrfs_item *item;
921 u32 old_left_nritems;
922 int ret = 0;
923 int wret;
925 slot = path->slots[1];
926 if (slot == 0) {
927 return 1;
929 if (!path->nodes[1]) {
930 return 1;
932 t = read_tree_block(root, btrfs_node_blockptr(&path->nodes[1]->node,
933 slot - 1));
934 left = &t->leaf;
935 free_space = btrfs_leaf_free_space(left);
936 if (free_space < data_size + sizeof(struct btrfs_item)) {
937 btrfs_block_release(root, t);
938 return 1;
941 /* cow and double check */
942 btrfs_cow_block(root, t, path->nodes[1], slot - 1, &t);
943 left = &t->leaf;
944 free_space = btrfs_leaf_free_space(left);
945 if (free_space < data_size + sizeof(struct btrfs_item)) {
946 btrfs_block_release(root, t);
947 return 1;
950 for (i = 0; i < btrfs_header_nritems(&right->header); i++) {
951 item = right->items + i;
952 if (path->slots[0] == i)
953 push_space += data_size + sizeof(*item);
954 if (btrfs_item_size(item) + sizeof(*item) + push_space >
955 free_space)
956 break;
957 push_items++;
958 push_space += btrfs_item_size(item) + sizeof(*item);
960 if (push_items == 0) {
961 btrfs_block_release(root, t);
962 return 1;
964 /* push data from right to left */
965 memcpy(left->items + btrfs_header_nritems(&left->header),
966 right->items, push_items * sizeof(struct btrfs_item));
967 push_space = LEAF_DATA_SIZE -
968 btrfs_item_offset(right->items + push_items -1);
969 memcpy(left->data + leaf_data_end(left) - push_space,
970 right->data + btrfs_item_offset(right->items + push_items - 1),
971 push_space);
972 old_left_nritems = btrfs_header_nritems(&left->header);
973 BUG_ON(old_left_nritems < 0);
975 for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
976 u16 ioff = btrfs_item_offset(left->items + i);
977 btrfs_set_item_offset(left->items + i, ioff - (LEAF_DATA_SIZE -
978 btrfs_item_offset(left->items +
979 old_left_nritems - 1)));
981 btrfs_set_header_nritems(&left->header, old_left_nritems + push_items);
983 /* fixup right node */
984 push_space = btrfs_item_offset(right->items + push_items - 1) -
985 leaf_data_end(right);
986 memmove(right->data + LEAF_DATA_SIZE - push_space, right->data +
987 leaf_data_end(right), push_space);
988 memmove(right->items, right->items + push_items,
989 (btrfs_header_nritems(&right->header) - push_items) *
990 sizeof(struct btrfs_item));
991 btrfs_set_header_nritems(&right->header,
992 btrfs_header_nritems(&right->header) -
993 push_items);
994 push_space = LEAF_DATA_SIZE;
996 for (i = 0; i < btrfs_header_nritems(&right->header); i++) {
997 btrfs_set_item_offset(right->items + i, push_space -
998 btrfs_item_size(right->items + i));
999 push_space = btrfs_item_offset(right->items + i);
1002 BUG_ON(list_empty(&t->dirty));
1003 BUG_ON(list_empty(&right_buf->dirty));
1005 wret = fixup_low_keys(root, path, &right->items[0].key, 1);
1006 if (wret)
1007 ret = wret;
1009 /* then fixup the leaf pointer in the path */
1010 if (path->slots[0] < push_items) {
1011 path->slots[0] += old_left_nritems;
1012 btrfs_block_release(root, path->nodes[0]);
1013 path->nodes[0] = t;
1014 path->slots[1] -= 1;
1015 } else {
1016 btrfs_block_release(root, t);
1017 path->slots[0] -= push_items;
1019 BUG_ON(path->slots[0] < 0);
1020 return ret;
1024 * split the path's leaf in two, making sure there is at least data_size
1025 * available for the resulting leaf level of the path.
1027 * returns 0 if all went well and < 0 on failure.
1029 static int split_leaf(struct btrfs_root *root, struct btrfs_path *path,
1030 int data_size)
1032 struct btrfs_buffer *l_buf;
1033 struct btrfs_leaf *l;
1034 u32 nritems;
1035 int mid;
1036 int slot;
1037 struct btrfs_leaf *right;
1038 struct btrfs_buffer *right_buffer;
1039 int space_needed = data_size + sizeof(struct btrfs_item);
1040 int data_copy_size;
1041 int rt_data_off;
1042 int i;
1043 int ret;
1044 int wret;
1046 wret = push_leaf_left(root, path, data_size);
1047 if (wret < 0)
1048 return wret;
1049 if (wret) {
1050 wret = push_leaf_right(root, path, data_size);
1051 if (wret < 0)
1052 return wret;
1055 l_buf = path->nodes[0];
1056 l = &l_buf->leaf;
1058 /* did the pushes work? */
1059 if (btrfs_leaf_free_space(l) >= sizeof(struct btrfs_item) + data_size)
1060 return 0;
1062 if (!path->nodes[1]) {
1063 ret = insert_new_root(root, path, 1);
1064 if (ret)
1065 return ret;
1067 slot = path->slots[0];
1068 nritems = btrfs_header_nritems(&l->header);
1069 mid = (nritems + 1)/ 2;
1070 right_buffer = btrfs_alloc_free_block(root);
1071 BUG_ON(!right_buffer);
1072 BUG_ON(mid == nritems);
1073 right = &right_buffer->leaf;
1074 memset(right, 0, sizeof(*right));
1075 if (mid <= slot) {
1076 /* FIXME, just alloc a new leaf here */
1077 if (leaf_space_used(l, mid, nritems - mid) + space_needed >
1078 LEAF_DATA_SIZE)
1079 BUG();
1080 } else {
1081 /* FIXME, just alloc a new leaf here */
1082 if (leaf_space_used(l, 0, mid + 1) + space_needed >
1083 LEAF_DATA_SIZE)
1084 BUG();
1086 btrfs_set_header_nritems(&right->header, nritems - mid);
1087 btrfs_set_header_blocknr(&right->header, right_buffer->blocknr);
1088 btrfs_set_header_level(&right->header, 0);
1089 btrfs_set_header_parentid(&right->header,
1090 btrfs_header_parentid(&root->node->node.header));
1091 data_copy_size = btrfs_item_end(l->items + mid) - leaf_data_end(l);
1092 memcpy(right->items, l->items + mid,
1093 (nritems - mid) * sizeof(struct btrfs_item));
1094 memcpy(right->data + LEAF_DATA_SIZE - data_copy_size,
1095 l->data + leaf_data_end(l), data_copy_size);
1096 rt_data_off = LEAF_DATA_SIZE - btrfs_item_end(l->items + mid);
1098 for (i = 0; i < btrfs_header_nritems(&right->header); i++) {
1099 u16 ioff = btrfs_item_offset(right->items + i);
1100 btrfs_set_item_offset(right->items + i, ioff + rt_data_off);
1103 btrfs_set_header_nritems(&l->header, mid);
1104 ret = 0;
1105 wret = insert_ptr(root, path, &right->items[0].key,
1106 right_buffer->blocknr, path->slots[1] + 1, 1);
1107 if (wret)
1108 ret = wret;
1109 BUG_ON(list_empty(&right_buffer->dirty));
1110 BUG_ON(list_empty(&l_buf->dirty));
1111 BUG_ON(path->slots[0] != slot);
1112 if (mid <= slot) {
1113 btrfs_block_release(root, path->nodes[0]);
1114 path->nodes[0] = right_buffer;
1115 path->slots[0] -= mid;
1116 path->slots[1] += 1;
1117 } else
1118 btrfs_block_release(root, right_buffer);
1119 BUG_ON(path->slots[0] < 0);
1120 return ret;
1124 * Given a key and some data, insert an item into the tree.
1125 * This does all the path init required, making room in the tree if needed.
1127 int btrfs_insert_item(struct btrfs_root *root, struct btrfs_key *cpu_key,
1128 void *data, int data_size)
1130 int ret = 0;
1131 int slot;
1132 int slot_orig;
1133 struct btrfs_leaf *leaf;
1134 struct btrfs_buffer *leaf_buf;
1135 u32 nritems;
1136 unsigned int data_end;
1137 struct btrfs_path path;
1138 struct btrfs_disk_key disk_key;
1140 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
1142 /* create a root if there isn't one */
1143 if (!root->node)
1144 BUG();
1145 btrfs_init_path(&path);
1146 ret = btrfs_search_slot(root, cpu_key, &path, data_size, 1);
1147 if (ret == 0) {
1148 btrfs_release_path(root, &path);
1149 return -EEXIST;
1151 if (ret < 0)
1152 goto out;
1154 slot_orig = path.slots[0];
1155 leaf_buf = path.nodes[0];
1156 leaf = &leaf_buf->leaf;
1158 nritems = btrfs_header_nritems(&leaf->header);
1159 data_end = leaf_data_end(leaf);
1161 if (btrfs_leaf_free_space(leaf) <
1162 sizeof(struct btrfs_item) + data_size)
1163 BUG();
1165 slot = path.slots[0];
1166 BUG_ON(slot < 0);
1167 if (slot != nritems) {
1168 int i;
1169 unsigned int old_data = btrfs_item_end(leaf->items + slot);
1172 * item0..itemN ... dataN.offset..dataN.size .. data0.size
1174 /* first correct the data pointers */
1175 for (i = slot; i < nritems; i++) {
1176 u16 ioff = btrfs_item_offset(leaf->items + i);
1177 btrfs_set_item_offset(leaf->items + i,
1178 ioff - data_size);
1181 /* shift the items */
1182 memmove(leaf->items + slot + 1, leaf->items + slot,
1183 (nritems - slot) * sizeof(struct btrfs_item));
1185 /* shift the data */
1186 memmove(leaf->data + data_end - data_size, leaf->data +
1187 data_end, old_data - data_end);
1188 data_end = old_data;
1190 /* copy the new data in */
1191 memcpy(&leaf->items[slot].key, &disk_key,
1192 sizeof(struct btrfs_disk_key));
1193 btrfs_set_item_offset(leaf->items + slot, data_end - data_size);
1194 btrfs_set_item_size(leaf->items + slot, data_size);
1195 memcpy(leaf->data + data_end - data_size, data, data_size);
1196 btrfs_set_header_nritems(&leaf->header, nritems + 1);
1198 ret = 0;
1199 if (slot == 0)
1200 ret = fixup_low_keys(root, &path, &disk_key, 1);
1202 BUG_ON(list_empty(&leaf_buf->dirty));
1203 if (btrfs_leaf_free_space(leaf) < 0)
1204 BUG();
1205 check_leaf(&path, 0);
1206 out:
1207 btrfs_release_path(root, &path);
1208 return ret;
1212 * delete the pointer from a given node.
1214 * If the delete empties a node, the node is removed from the tree,
1215 * continuing all the way the root if required. The root is converted into
1216 * a leaf if all the nodes are emptied.
1218 static int del_ptr(struct btrfs_root *root, struct btrfs_path *path, int level,
1219 int slot)
1221 struct btrfs_node *node;
1222 struct btrfs_buffer *parent = path->nodes[level];
1223 u32 nritems;
1224 int ret = 0;
1225 int wret;
1227 node = &parent->node;
1228 nritems = btrfs_header_nritems(&node->header);
1229 if (slot != nritems -1) {
1230 memmove(node->keys + slot, node->keys + slot + 1,
1231 sizeof(struct btrfs_disk_key) * (nritems - slot - 1));
1232 memmove(node->blockptrs + slot,
1233 node->blockptrs + slot + 1,
1234 sizeof(u64) * (nritems - slot - 1));
1236 nritems--;
1237 btrfs_set_header_nritems(&node->header, nritems);
1238 if (nritems == 0 && parent == root->node) {
1239 BUG_ON(btrfs_header_level(&root->node->node.header) != 1);
1240 /* just turn the root into a leaf and break */
1241 btrfs_set_header_level(&root->node->node.header, 0);
1242 } else if (slot == 0) {
1243 wret = fixup_low_keys(root, path, node->keys, level + 1);
1244 if (wret)
1245 ret = wret;
1247 BUG_ON(list_empty(&parent->dirty));
1248 return ret;
1252 * delete the item at the leaf level in path. If that empties
1253 * the leaf, remove it from the tree
1255 int btrfs_del_item(struct btrfs_root *root, struct btrfs_path *path)
1257 int slot;
1258 struct btrfs_leaf *leaf;
1259 struct btrfs_buffer *leaf_buf;
1260 int doff;
1261 int dsize;
1262 int ret = 0;
1263 int wret;
1264 u32 nritems;
1266 leaf_buf = path->nodes[0];
1267 leaf = &leaf_buf->leaf;
1268 slot = path->slots[0];
1269 doff = btrfs_item_offset(leaf->items + slot);
1270 dsize = btrfs_item_size(leaf->items + slot);
1271 nritems = btrfs_header_nritems(&leaf->header);
1273 if (slot != nritems - 1) {
1274 int i;
1275 int data_end = leaf_data_end(leaf);
1276 memmove(leaf->data + data_end + dsize,
1277 leaf->data + data_end,
1278 doff - data_end);
1279 for (i = slot + 1; i < nritems; i++) {
1280 u16 ioff = btrfs_item_offset(leaf->items + i);
1281 btrfs_set_item_offset(leaf->items + i, ioff + dsize);
1283 memmove(leaf->items + slot, leaf->items + slot + 1,
1284 sizeof(struct btrfs_item) *
1285 (nritems - slot - 1));
1287 btrfs_set_header_nritems(&leaf->header, nritems - 1);
1288 nritems--;
1289 /* delete the leaf if we've emptied it */
1290 if (nritems == 0) {
1291 if (leaf_buf == root->node) {
1292 btrfs_set_header_level(&leaf->header, 0);
1293 BUG_ON(list_empty(&leaf_buf->dirty));
1294 } else {
1295 clean_tree_block(root, leaf_buf);
1296 wret = del_ptr(root, path, 1, path->slots[1]);
1297 if (wret)
1298 ret = wret;
1299 wret = btrfs_free_extent(root, leaf_buf->blocknr, 1);
1300 if (wret)
1301 ret = wret;
1303 } else {
1304 int used = leaf_space_used(leaf, 0, nritems);
1305 if (slot == 0) {
1306 wret = fixup_low_keys(root, path,
1307 &leaf->items[0].key, 1);
1308 if (wret)
1309 ret = wret;
1311 BUG_ON(list_empty(&leaf_buf->dirty));
1313 /* delete the leaf if it is mostly empty */
1314 if (used < LEAF_DATA_SIZE / 3) {
1315 /* push_leaf_left fixes the path.
1316 * make sure the path still points to our leaf
1317 * for possible call to del_ptr below
1319 slot = path->slots[1];
1320 leaf_buf->count++;
1321 wret = push_leaf_left(root, path, 1);
1322 if (wret < 0)
1323 ret = wret;
1324 if (path->nodes[0] == leaf_buf &&
1325 btrfs_header_nritems(&leaf->header)) {
1326 wret = push_leaf_right(root, path, 1);
1327 if (wret < 0)
1328 ret = wret;
1330 if (btrfs_header_nritems(&leaf->header) == 0) {
1331 u64 blocknr = leaf_buf->blocknr;
1332 clean_tree_block(root, leaf_buf);
1333 wret = del_ptr(root, path, 1, slot);
1334 if (wret)
1335 ret = wret;
1336 btrfs_block_release(root, leaf_buf);
1337 wret = btrfs_free_extent(root, blocknr, 1);
1338 if (wret)
1339 ret = wret;
1340 } else {
1341 btrfs_block_release(root, leaf_buf);
1345 return ret;
1349 * walk up the tree as far as required to find the next leaf.
1350 * returns 0 if it found something or 1 if there are no greater leaves.
1351 * returns < 0 on io errors.
1353 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
1355 int slot;
1356 int level = 1;
1357 u64 blocknr;
1358 struct btrfs_buffer *c;
1359 struct btrfs_buffer *next = NULL;
1361 while(level < BTRFS_MAX_LEVEL) {
1362 if (!path->nodes[level])
1363 return 1;
1364 slot = path->slots[level] + 1;
1365 c = path->nodes[level];
1366 if (slot >= btrfs_header_nritems(&c->node.header)) {
1367 level++;
1368 continue;
1370 blocknr = btrfs_node_blockptr(&c->node, slot);
1371 if (next)
1372 btrfs_block_release(root, next);
1373 next = read_tree_block(root, blocknr);
1374 break;
1376 path->slots[level] = slot;
1377 while(1) {
1378 level--;
1379 c = path->nodes[level];
1380 btrfs_block_release(root, c);
1381 path->nodes[level] = next;
1382 path->slots[level] = 0;
1383 if (!level)
1384 break;
1385 next = read_tree_block(root,
1386 btrfs_node_blockptr(&next->node, 0));
1388 return 0;