xref: /openbmc/linux/fs/btrfs/ctree.c (revision ed2ff2cba766dfe7976a0113f667c9a0a50dff02)
1be0e5c09SChris Mason #include <stdio.h>
2be0e5c09SChris Mason #include <stdlib.h>
3be0e5c09SChris Mason #include "kerncompat.h"
4eb60ceacSChris Mason #include "radix-tree.h"
5eb60ceacSChris Mason #include "ctree.h"
6eb60ceacSChris Mason #include "disk-io.h"
75de08d7dSChris Mason #include "print-tree.h"
89a8dd150SChris Mason 
9aa5d6bedSChris Mason static int split_node(struct ctree_root *root, struct ctree_path *path,
10aa5d6bedSChris Mason 		      int level);
11aa5d6bedSChris Mason static int split_leaf(struct ctree_root *root, struct ctree_path *path,
12aa5d6bedSChris Mason 		      int data_size);
13bb803951SChris Mason static int push_node_left(struct ctree_root *root, struct tree_buffer *dst,
14bb803951SChris Mason 			  struct tree_buffer *src);
1579f95c82SChris Mason static int balance_node_right(struct ctree_root *root,
1679f95c82SChris Mason 			      struct tree_buffer *dst_buf,
1779f95c82SChris Mason 			      struct tree_buffer *src_buf);
18bb803951SChris Mason static int del_ptr(struct ctree_root *root, struct ctree_path *path, int level,
19bb803951SChris Mason 		   int slot);
20d97e63b6SChris Mason 
215de08d7dSChris Mason inline void init_path(struct ctree_path *p)
22be0e5c09SChris Mason {
23be0e5c09SChris Mason 	memset(p, 0, sizeof(*p));
24be0e5c09SChris Mason }
25be0e5c09SChris Mason 
265de08d7dSChris Mason void release_path(struct ctree_root *root, struct ctree_path *p)
27eb60ceacSChris Mason {
28eb60ceacSChris Mason 	int i;
29eb60ceacSChris Mason 	for (i = 0; i < MAX_LEVEL; i++) {
30eb60ceacSChris Mason 		if (!p->nodes[i])
31eb60ceacSChris Mason 			break;
32eb60ceacSChris Mason 		tree_block_release(root, p->nodes[i]);
33eb60ceacSChris Mason 	}
34aa5d6bedSChris Mason 	memset(p, 0, sizeof(*p));
35eb60ceacSChris Mason }
36eb60ceacSChris Mason 
3774123bd7SChris Mason /*
3874123bd7SChris Mason  * The leaf data grows from end-to-front in the node.
3974123bd7SChris Mason  * this returns the address of the start of the last item,
4074123bd7SChris Mason  * which is the stop of the leaf data stack
4174123bd7SChris Mason  */
42be0e5c09SChris Mason static inline unsigned int leaf_data_end(struct leaf *leaf)
43be0e5c09SChris Mason {
44be0e5c09SChris Mason 	unsigned int nr = leaf->header.nritems;
45be0e5c09SChris Mason 	if (nr == 0)
46d97e63b6SChris Mason 		return sizeof(leaf->data);
47be0e5c09SChris Mason 	return leaf->items[nr-1].offset;
48be0e5c09SChris Mason }
49be0e5c09SChris Mason 
5074123bd7SChris Mason /*
5174123bd7SChris Mason  * The space between the end of the leaf items and
5274123bd7SChris Mason  * the start of the leaf data.  IOW, how much room
5374123bd7SChris Mason  * the leaf has left for both items and data
5474123bd7SChris Mason  */
555de08d7dSChris Mason int leaf_free_space(struct leaf *leaf)
56be0e5c09SChris Mason {
57be0e5c09SChris Mason 	int data_end = leaf_data_end(leaf);
58be0e5c09SChris Mason 	int nritems = leaf->header.nritems;
59be0e5c09SChris Mason 	char *items_end = (char *)(leaf->items + nritems + 1);
60be0e5c09SChris Mason 	return (char *)(leaf->data + data_end) - (char *)items_end;
61be0e5c09SChris Mason }
62be0e5c09SChris Mason 
6374123bd7SChris Mason /*
6474123bd7SChris Mason  * compare two keys in a memcmp fashion
6574123bd7SChris Mason  */
66be0e5c09SChris Mason int comp_keys(struct key *k1, struct key *k2)
67be0e5c09SChris Mason {
68be0e5c09SChris Mason 	if (k1->objectid > k2->objectid)
69be0e5c09SChris Mason 		return 1;
70be0e5c09SChris Mason 	if (k1->objectid < k2->objectid)
71be0e5c09SChris Mason 		return -1;
72be0e5c09SChris Mason 	if (k1->flags > k2->flags)
73be0e5c09SChris Mason 		return 1;
74be0e5c09SChris Mason 	if (k1->flags < k2->flags)
75be0e5c09SChris Mason 		return -1;
76be0e5c09SChris Mason 	if (k1->offset > k2->offset)
77be0e5c09SChris Mason 		return 1;
78be0e5c09SChris Mason 	if (k1->offset < k2->offset)
79be0e5c09SChris Mason 		return -1;
80be0e5c09SChris Mason 	return 0;
81be0e5c09SChris Mason }
8274123bd7SChris Mason 
83aa5d6bedSChris Mason int check_node(struct ctree_path *path, int level)
84aa5d6bedSChris Mason {
85aa5d6bedSChris Mason 	int i;
86aa5d6bedSChris Mason 	struct node *parent = NULL;
87aa5d6bedSChris Mason 	struct node *node = &path->nodes[level]->node;
88aa5d6bedSChris Mason 	int parent_slot;
89aa5d6bedSChris Mason 
90aa5d6bedSChris Mason 	if (path->nodes[level + 1])
91aa5d6bedSChris Mason 		parent = &path->nodes[level + 1]->node;
92aa5d6bedSChris Mason 	parent_slot = path->slots[level + 1];
93aa5d6bedSChris Mason 	if (parent && node->header.nritems > 0) {
94aa5d6bedSChris Mason 		struct key *parent_key;
95aa5d6bedSChris Mason 		parent_key = &parent->keys[parent_slot];
96aa5d6bedSChris Mason 		BUG_ON(memcmp(parent_key, node->keys, sizeof(struct key)));
97aa5d6bedSChris Mason 		BUG_ON(parent->blockptrs[parent_slot] != node->header.blocknr);
98aa5d6bedSChris Mason 	}
99aa5d6bedSChris Mason 	BUG_ON(node->header.nritems > NODEPTRS_PER_BLOCK);
100aa5d6bedSChris Mason 	for (i = 0; i < node->header.nritems - 2; i++) {
101aa5d6bedSChris Mason 		BUG_ON(comp_keys(&node->keys[i], &node->keys[i+1]) >= 0);
102aa5d6bedSChris Mason 	}
103aa5d6bedSChris Mason 	return 0;
104aa5d6bedSChris Mason }
105aa5d6bedSChris Mason 
106aa5d6bedSChris Mason int check_leaf(struct ctree_path *path, int level)
107aa5d6bedSChris Mason {
108aa5d6bedSChris Mason 	int i;
109aa5d6bedSChris Mason 	struct leaf *leaf = &path->nodes[level]->leaf;
110aa5d6bedSChris Mason 	struct node *parent = NULL;
111aa5d6bedSChris Mason 	int parent_slot;
112aa5d6bedSChris Mason 
113aa5d6bedSChris Mason 	if (path->nodes[level + 1])
114aa5d6bedSChris Mason 		parent = &path->nodes[level + 1]->node;
115aa5d6bedSChris Mason 	parent_slot = path->slots[level + 1];
116aa5d6bedSChris Mason 	if (parent && leaf->header.nritems > 0) {
117aa5d6bedSChris Mason 		struct key *parent_key;
118aa5d6bedSChris Mason 		parent_key = &parent->keys[parent_slot];
119aa5d6bedSChris Mason 		BUG_ON(memcmp(parent_key, &leaf->items[0].key,
120aa5d6bedSChris Mason 		       sizeof(struct key)));
121aa5d6bedSChris Mason 		BUG_ON(parent->blockptrs[parent_slot] != leaf->header.blocknr);
122aa5d6bedSChris Mason 	}
123aa5d6bedSChris Mason 	for (i = 0; i < leaf->header.nritems - 2; i++) {
124aa5d6bedSChris Mason 		BUG_ON(comp_keys(&leaf->items[i].key,
125aa5d6bedSChris Mason 		                 &leaf->items[i+1].key) >= 0);
126aa5d6bedSChris Mason 		BUG_ON(leaf->items[i].offset != leaf->items[i + 1].offset +
127aa5d6bedSChris Mason 		    leaf->items[i + 1].size);
128aa5d6bedSChris Mason 		if (i == 0) {
129aa5d6bedSChris Mason 			BUG_ON(leaf->items[i].offset + leaf->items[i].size !=
130aa5d6bedSChris Mason 				LEAF_DATA_SIZE);
131aa5d6bedSChris Mason 		}
132aa5d6bedSChris Mason 	}
133aa5d6bedSChris Mason 	BUG_ON(leaf_free_space(leaf) < 0);
134aa5d6bedSChris Mason 	return 0;
135aa5d6bedSChris Mason }
136aa5d6bedSChris Mason 
137aa5d6bedSChris Mason int check_block(struct ctree_path *path, int level)
138aa5d6bedSChris Mason {
139aa5d6bedSChris Mason 	if (level == 0)
140aa5d6bedSChris Mason 		return check_leaf(path, level);
141aa5d6bedSChris Mason 	return check_node(path, level);
142aa5d6bedSChris Mason }
143aa5d6bedSChris Mason 
14474123bd7SChris Mason /*
14574123bd7SChris Mason  * search for key in the array p.  items p are item_size apart
14674123bd7SChris Mason  * and there are 'max' items in p
14774123bd7SChris Mason  * the slot in the array is returned via slot, and it points to
14874123bd7SChris Mason  * the place where you would insert key if it is not found in
14974123bd7SChris Mason  * the array.
15074123bd7SChris Mason  *
15174123bd7SChris Mason  * slot may point to max if the key is bigger than all of the keys
15274123bd7SChris Mason  */
153be0e5c09SChris Mason int generic_bin_search(char *p, int item_size, struct key *key,
154be0e5c09SChris Mason 		       int max, int *slot)
155be0e5c09SChris Mason {
156be0e5c09SChris Mason 	int low = 0;
157be0e5c09SChris Mason 	int high = max;
158be0e5c09SChris Mason 	int mid;
159be0e5c09SChris Mason 	int ret;
160be0e5c09SChris Mason 	struct key *tmp;
161be0e5c09SChris Mason 
162be0e5c09SChris Mason 	while(low < high) {
163be0e5c09SChris Mason 		mid = (low + high) / 2;
164be0e5c09SChris Mason 		tmp = (struct key *)(p + mid * item_size);
165be0e5c09SChris Mason 		ret = comp_keys(tmp, key);
166be0e5c09SChris Mason 
167be0e5c09SChris Mason 		if (ret < 0)
168be0e5c09SChris Mason 			low = mid + 1;
169be0e5c09SChris Mason 		else if (ret > 0)
170be0e5c09SChris Mason 			high = mid;
171be0e5c09SChris Mason 		else {
172be0e5c09SChris Mason 			*slot = mid;
173be0e5c09SChris Mason 			return 0;
174be0e5c09SChris Mason 		}
175be0e5c09SChris Mason 	}
176be0e5c09SChris Mason 	*slot = low;
177be0e5c09SChris Mason 	return 1;
178be0e5c09SChris Mason }
179be0e5c09SChris Mason 
18097571fd0SChris Mason /*
18197571fd0SChris Mason  * simple bin_search frontend that does the right thing for
18297571fd0SChris Mason  * leaves vs nodes
18397571fd0SChris Mason  */
184be0e5c09SChris Mason int bin_search(struct node *c, struct key *key, int *slot)
185be0e5c09SChris Mason {
186be0e5c09SChris Mason 	if (is_leaf(c->header.flags)) {
187be0e5c09SChris Mason 		struct leaf *l = (struct leaf *)c;
188be0e5c09SChris Mason 		return generic_bin_search((void *)l->items, sizeof(struct item),
189be0e5c09SChris Mason 					  key, c->header.nritems, slot);
190be0e5c09SChris Mason 	} else {
191be0e5c09SChris Mason 		return generic_bin_search((void *)c->keys, sizeof(struct key),
192be0e5c09SChris Mason 					  key, c->header.nritems, slot);
193be0e5c09SChris Mason 	}
194be0e5c09SChris Mason 	return -1;
195be0e5c09SChris Mason }
196be0e5c09SChris Mason 
197bb803951SChris Mason struct tree_buffer *read_node_slot(struct ctree_root *root,
198bb803951SChris Mason 				   struct tree_buffer *parent_buf,
199bb803951SChris Mason 				   int slot)
200bb803951SChris Mason {
201bb803951SChris Mason 	struct node *node = &parent_buf->node;
202bb803951SChris Mason 	if (slot < 0)
203bb803951SChris Mason 		return NULL;
204bb803951SChris Mason 	if (slot >= node->header.nritems)
205bb803951SChris Mason 		return NULL;
206bb803951SChris Mason 	return read_tree_block(root, node->blockptrs[slot]);
207bb803951SChris Mason }
208bb803951SChris Mason 
209bb803951SChris Mason static int balance_level(struct ctree_root *root, struct ctree_path *path,
210bb803951SChris Mason 			int level)
211bb803951SChris Mason {
212bb803951SChris Mason 	struct tree_buffer *right_buf;
213bb803951SChris Mason 	struct tree_buffer *mid_buf;
214bb803951SChris Mason 	struct tree_buffer *left_buf;
215bb803951SChris Mason 	struct tree_buffer *parent_buf = NULL;
216bb803951SChris Mason 	struct node *right = NULL;
217bb803951SChris Mason 	struct node *mid;
218bb803951SChris Mason 	struct node *left = NULL;
219bb803951SChris Mason 	struct node *parent = NULL;
220bb803951SChris Mason 	int ret = 0;
221bb803951SChris Mason 	int wret;
222bb803951SChris Mason 	int pslot;
223bb803951SChris Mason 	int orig_slot = path->slots[level];
22479f95c82SChris Mason 	u64 orig_ptr;
225bb803951SChris Mason 
226bb803951SChris Mason 	if (level == 0)
227bb803951SChris Mason 		return 0;
228bb803951SChris Mason 
229bb803951SChris Mason 	mid_buf = path->nodes[level];
230bb803951SChris Mason 	mid = &mid_buf->node;
23179f95c82SChris Mason 	orig_ptr = mid->blockptrs[orig_slot];
23279f95c82SChris Mason 
233bb803951SChris Mason 	if (level < MAX_LEVEL - 1)
234bb803951SChris Mason 		parent_buf = path->nodes[level + 1];
235bb803951SChris Mason 	pslot = path->slots[level + 1];
236bb803951SChris Mason 
237bb803951SChris Mason 	if (!parent_buf) {
238bb803951SChris Mason 		struct tree_buffer *child;
239bb803951SChris Mason 		u64 blocknr = mid_buf->blocknr;
240bb803951SChris Mason 
241bb803951SChris Mason 		if (mid->header.nritems != 1)
242bb803951SChris Mason 			return 0;
243bb803951SChris Mason 
244bb803951SChris Mason 		/* promote the child to a root */
245bb803951SChris Mason 		child = read_node_slot(root, mid_buf, 0);
246bb803951SChris Mason 		BUG_ON(!child);
247bb803951SChris Mason 		root->node = child;
248bb803951SChris Mason 		path->nodes[level] = NULL;
249bb803951SChris Mason 		/* once for the path */
250bb803951SChris Mason 		tree_block_release(root, mid_buf);
251bb803951SChris Mason 		/* once for the root ptr */
252bb803951SChris Mason 		tree_block_release(root, mid_buf);
253ed2ff2cbSChris Mason 		clean_tree_block(root, mid_buf);
254bb803951SChris Mason 		return free_extent(root, blocknr, 1);
255bb803951SChris Mason 	}
256bb803951SChris Mason 	parent = &parent_buf->node;
257bb803951SChris Mason 
258bb803951SChris Mason 	if (mid->header.nritems > NODEPTRS_PER_BLOCK / 4)
259bb803951SChris Mason 		return 0;
260bb803951SChris Mason 
261bb803951SChris Mason 	left_buf = read_node_slot(root, parent_buf, pslot - 1);
262bb803951SChris Mason 	right_buf = read_node_slot(root, parent_buf, pslot + 1);
26379f95c82SChris Mason 
26479f95c82SChris Mason 	/* first, try to make some room in the middle buffer */
265bb803951SChris Mason 	if (left_buf) {
266bb803951SChris Mason 		left = &left_buf->node;
267bb803951SChris Mason 		orig_slot += left->header.nritems;
26879f95c82SChris Mason 		wret = push_node_left(root, left_buf, mid_buf);
26979f95c82SChris Mason 		if (wret < 0)
27079f95c82SChris Mason 			ret = wret;
271bb803951SChris Mason 	}
27279f95c82SChris Mason 
27379f95c82SChris Mason 	/*
27479f95c82SChris Mason 	 * then try to empty the right most buffer into the middle
27579f95c82SChris Mason 	 */
276bb803951SChris Mason 	if (right_buf) {
27779f95c82SChris Mason 		right = &right_buf->node;
27879f95c82SChris Mason 		wret = push_node_left(root, mid_buf, right_buf);
27979f95c82SChris Mason 		if (wret < 0)
28079f95c82SChris Mason 			ret = wret;
281bb803951SChris Mason 		if (right->header.nritems == 0) {
282bb803951SChris Mason 			u64 blocknr = right_buf->blocknr;
283bb803951SChris Mason 			tree_block_release(root, right_buf);
284ed2ff2cbSChris Mason 			clean_tree_block(root, right_buf);
285bb803951SChris Mason 			right_buf = NULL;
286bb803951SChris Mason 			right = NULL;
287bb803951SChris Mason 			wret = del_ptr(root, path, level + 1, pslot + 1);
288bb803951SChris Mason 			if (wret)
289bb803951SChris Mason 				ret = wret;
290bb803951SChris Mason 			wret = free_extent(root, blocknr, 1);
291bb803951SChris Mason 			if (wret)
292bb803951SChris Mason 				ret = wret;
293bb803951SChris Mason 		} else {
294bb803951SChris Mason 			memcpy(parent->keys + pslot + 1, right->keys,
295bb803951SChris Mason 				sizeof(struct key));
296ed2ff2cbSChris Mason 			wret = dirty_tree_block(root, parent_buf);
29779f95c82SChris Mason 			if (wret)
29879f95c82SChris Mason 				ret = wret;
299bb803951SChris Mason 		}
300bb803951SChris Mason 	}
30179f95c82SChris Mason 	if (mid->header.nritems == 1) {
30279f95c82SChris Mason 		/*
30379f95c82SChris Mason 		 * we're not allowed to leave a node with one item in the
30479f95c82SChris Mason 		 * tree during a delete.  A deletion from lower in the tree
30579f95c82SChris Mason 		 * could try to delete the only pointer in this node.
30679f95c82SChris Mason 		 * So, pull some keys from the left.
30779f95c82SChris Mason 		 * There has to be a left pointer at this point because
30879f95c82SChris Mason 		 * otherwise we would have pulled some pointers from the
30979f95c82SChris Mason 		 * right
31079f95c82SChris Mason 		 */
31179f95c82SChris Mason 		BUG_ON(!left_buf);
31279f95c82SChris Mason 		wret = balance_node_right(root, mid_buf, left_buf);
31379f95c82SChris Mason 		if (wret < 0)
31479f95c82SChris Mason 			ret = wret;
31579f95c82SChris Mason 		BUG_ON(wret == 1);
31679f95c82SChris Mason 	}
317bb803951SChris Mason 	if (mid->header.nritems == 0) {
31879f95c82SChris Mason 		/* we've managed to empty the middle node, drop it */
319bb803951SChris Mason 		u64 blocknr = mid_buf->blocknr;
320bb803951SChris Mason 		tree_block_release(root, mid_buf);
321ed2ff2cbSChris Mason 		clean_tree_block(root, mid_buf);
322bb803951SChris Mason 		mid_buf = NULL;
323bb803951SChris Mason 		mid = NULL;
324bb803951SChris Mason 		wret = del_ptr(root, path, level + 1, pslot);
325bb803951SChris Mason 		if (wret)
326bb803951SChris Mason 			ret = wret;
327bb803951SChris Mason 		wret = free_extent(root, blocknr, 1);
328bb803951SChris Mason 		if (wret)
329bb803951SChris Mason 			ret = wret;
33079f95c82SChris Mason 	} else {
33179f95c82SChris Mason 		/* update the parent key to reflect our changes */
332bb803951SChris Mason 		memcpy(parent->keys + pslot, mid->keys, sizeof(struct key));
333ed2ff2cbSChris Mason 		wret = dirty_tree_block(root, parent_buf);
33479f95c82SChris Mason 		if (wret)
33579f95c82SChris Mason 			ret = wret;
33679f95c82SChris Mason 	}
337bb803951SChris Mason 
33879f95c82SChris Mason 	/* update the path */
339bb803951SChris Mason 	if (left_buf) {
34079f95c82SChris Mason 		if (left->header.nritems > orig_slot) {
341bb803951SChris Mason 			left_buf->count++; // released below
342bb803951SChris Mason 			path->nodes[level] = left_buf;
343bb803951SChris Mason 			path->slots[level + 1] -= 1;
344bb803951SChris Mason 			path->slots[level] = orig_slot;
345bb803951SChris Mason 			if (mid_buf)
346bb803951SChris Mason 				tree_block_release(root, mid_buf);
347bb803951SChris Mason 		} else {
348bb803951SChris Mason 			orig_slot -= left->header.nritems;
349bb803951SChris Mason 			path->slots[level] = orig_slot;
350bb803951SChris Mason 		}
351bb803951SChris Mason 	}
35279f95c82SChris Mason 	/* double check we haven't messed things up */
35379f95c82SChris Mason 	check_block(path, level);
35479f95c82SChris Mason 	if (orig_ptr != path->nodes[level]->node.blockptrs[path->slots[level]])
35579f95c82SChris Mason 		BUG();
356bb803951SChris Mason 
357bb803951SChris Mason 	if (right_buf)
358bb803951SChris Mason 		tree_block_release(root, right_buf);
359bb803951SChris Mason 	if (left_buf)
360bb803951SChris Mason 		tree_block_release(root, left_buf);
361bb803951SChris Mason 	return ret;
362bb803951SChris Mason }
363bb803951SChris Mason 
36474123bd7SChris Mason /*
36574123bd7SChris Mason  * look for key in the tree.  path is filled in with nodes along the way
36674123bd7SChris Mason  * if key is found, we return zero and you can find the item in the leaf
36774123bd7SChris Mason  * level of the path (level 0)
36874123bd7SChris Mason  *
36974123bd7SChris Mason  * If the key isn't found, the path points to the slot where it should
370aa5d6bedSChris Mason  * be inserted, and 1 is returned.  If there are other errors during the
371aa5d6bedSChris Mason  * search a negative error number is returned.
37297571fd0SChris Mason  *
37397571fd0SChris Mason  * if ins_len > 0, nodes and leaves will be split as we walk down the
37497571fd0SChris Mason  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
37597571fd0SChris Mason  * possible)
37674123bd7SChris Mason  */
3775de08d7dSChris Mason int search_slot(struct ctree_root *root, struct key *key,
3785de08d7dSChris Mason 		struct ctree_path *p, int ins_len)
379be0e5c09SChris Mason {
380bb803951SChris Mason 	struct tree_buffer *b;
381eb60ceacSChris Mason 	struct node *c;
382be0e5c09SChris Mason 	int slot;
383be0e5c09SChris Mason 	int ret;
384be0e5c09SChris Mason 	int level;
3855c680ed6SChris Mason 
386bb803951SChris Mason again:
387bb803951SChris Mason 	b = root->node;
388eb60ceacSChris Mason 	b->count++;
389eb60ceacSChris Mason 	while (b) {
390eb60ceacSChris Mason 		c = &b->node;
391be0e5c09SChris Mason 		level = node_level(c->header.flags);
392eb60ceacSChris Mason 		p->nodes[level] = b;
393aa5d6bedSChris Mason 		ret = check_block(p, level);
394aa5d6bedSChris Mason 		if (ret)
395aa5d6bedSChris Mason 			return -1;
396be0e5c09SChris Mason 		ret = bin_search(c, key, &slot);
397be0e5c09SChris Mason 		if (!is_leaf(c->header.flags)) {
398be0e5c09SChris Mason 			if (ret && slot > 0)
399be0e5c09SChris Mason 				slot -= 1;
400be0e5c09SChris Mason 			p->slots[level] = slot;
4015de08d7dSChris Mason 			if (ins_len > 0 &&
4025de08d7dSChris Mason 			    c->header.nritems == NODEPTRS_PER_BLOCK) {
4035c680ed6SChris Mason 				int sret = split_node(root, p, level);
4045c680ed6SChris Mason 				BUG_ON(sret > 0);
4055c680ed6SChris Mason 				if (sret)
4065c680ed6SChris Mason 					return sret;
4075c680ed6SChris Mason 				b = p->nodes[level];
4085c680ed6SChris Mason 				c = &b->node;
4095c680ed6SChris Mason 				slot = p->slots[level];
410bb803951SChris Mason 			} else if (ins_len < 0) {
411bb803951SChris Mason 				int sret = balance_level(root, p, level);
412bb803951SChris Mason 				if (sret)
413bb803951SChris Mason 					return sret;
414bb803951SChris Mason 				b = p->nodes[level];
415bb803951SChris Mason 				if (!b)
416bb803951SChris Mason 					goto again;
417bb803951SChris Mason 				c = &b->node;
418bb803951SChris Mason 				slot = p->slots[level];
41979f95c82SChris Mason 				BUG_ON(c->header.nritems == 1);
4205c680ed6SChris Mason 			}
421eb60ceacSChris Mason 			b = read_tree_block(root, c->blockptrs[slot]);
422be0e5c09SChris Mason 		} else {
4235c680ed6SChris Mason 			struct leaf *l = (struct leaf *)c;
424be0e5c09SChris Mason 			p->slots[level] = slot;
4255de08d7dSChris Mason 			if (ins_len > 0 && leaf_free_space(l) <
4265de08d7dSChris Mason 			    sizeof(struct item) + ins_len) {
4275c680ed6SChris Mason 				int sret = split_leaf(root, p, ins_len);
4285c680ed6SChris Mason 				BUG_ON(sret > 0);
4295c680ed6SChris Mason 				if (sret)
4305c680ed6SChris Mason 					return sret;
4315c680ed6SChris Mason 			}
432bb803951SChris Mason 			BUG_ON(root->node->count == 1);
433be0e5c09SChris Mason 			return ret;
434be0e5c09SChris Mason 		}
435be0e5c09SChris Mason 	}
436bb803951SChris Mason 	BUG_ON(root->node->count == 1);
437aa5d6bedSChris Mason 	return 1;
438be0e5c09SChris Mason }
439be0e5c09SChris Mason 
44074123bd7SChris Mason /*
44174123bd7SChris Mason  * adjust the pointers going up the tree, starting at level
44274123bd7SChris Mason  * making sure the right key of each node is points to 'key'.
44374123bd7SChris Mason  * This is used after shifting pointers to the left, so it stops
44474123bd7SChris Mason  * fixing up pointers when a given leaf/node is not in slot 0 of the
44574123bd7SChris Mason  * higher levels
446aa5d6bedSChris Mason  *
447aa5d6bedSChris Mason  * If this fails to write a tree block, it returns -1, but continues
448aa5d6bedSChris Mason  * fixing up the blocks in ram so the tree is consistent.
44974123bd7SChris Mason  */
450aa5d6bedSChris Mason static int fixup_low_keys(struct ctree_root *root,
451eb60ceacSChris Mason 			   struct ctree_path *path, struct key *key,
452be0e5c09SChris Mason 			   int level)
453be0e5c09SChris Mason {
454be0e5c09SChris Mason 	int i;
455aa5d6bedSChris Mason 	int ret = 0;
456aa5d6bedSChris Mason 	int wret;
457be0e5c09SChris Mason 	for (i = level; i < MAX_LEVEL; i++) {
458eb60ceacSChris Mason 		struct node *t;
459be0e5c09SChris Mason 		int tslot = path->slots[i];
460eb60ceacSChris Mason 		if (!path->nodes[i])
461be0e5c09SChris Mason 			break;
462eb60ceacSChris Mason 		t = &path->nodes[i]->node;
463be0e5c09SChris Mason 		memcpy(t->keys + tslot, key, sizeof(*key));
464ed2ff2cbSChris Mason 		wret = dirty_tree_block(root, path->nodes[i]);
465aa5d6bedSChris Mason 		if (wret)
466aa5d6bedSChris Mason 			ret = wret;
467be0e5c09SChris Mason 		if (tslot != 0)
468be0e5c09SChris Mason 			break;
469be0e5c09SChris Mason 	}
470aa5d6bedSChris Mason 	return ret;
471be0e5c09SChris Mason }
472be0e5c09SChris Mason 
47374123bd7SChris Mason /*
47474123bd7SChris Mason  * try to push data from one node into the next node left in the
47579f95c82SChris Mason  * tree.
476aa5d6bedSChris Mason  *
477aa5d6bedSChris Mason  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
478aa5d6bedSChris Mason  * error, and > 0 if there was no room in the left hand block.
47974123bd7SChris Mason  */
480bb803951SChris Mason static int push_node_left(struct ctree_root *root, struct tree_buffer *dst_buf,
481bb803951SChris Mason 			  struct tree_buffer *src_buf)
482be0e5c09SChris Mason {
483bb803951SChris Mason 	struct node *src = &src_buf->node;
484bb803951SChris Mason 	struct node *dst = &dst_buf->node;
485be0e5c09SChris Mason 	int push_items = 0;
486bb803951SChris Mason 	int src_nritems;
487bb803951SChris Mason 	int dst_nritems;
488aa5d6bedSChris Mason 	int ret = 0;
489aa5d6bedSChris Mason 	int wret;
490be0e5c09SChris Mason 
491be0e5c09SChris Mason 	src_nritems = src->header.nritems;
492bb803951SChris Mason 	dst_nritems = dst->header.nritems;
493bb803951SChris Mason 	push_items = NODEPTRS_PER_BLOCK - dst_nritems;
494eb60ceacSChris Mason 	if (push_items <= 0) {
495be0e5c09SChris Mason 		return 1;
496eb60ceacSChris Mason 	}
497be0e5c09SChris Mason 
498be0e5c09SChris Mason 	if (src_nritems < push_items)
499be0e5c09SChris Mason 		push_items = src_nritems;
50079f95c82SChris Mason 
501bb803951SChris Mason 	memcpy(dst->keys + dst_nritems, src->keys,
502be0e5c09SChris Mason 		push_items * sizeof(struct key));
503bb803951SChris Mason 	memcpy(dst->blockptrs + dst_nritems, src->blockptrs,
504be0e5c09SChris Mason 		push_items * sizeof(u64));
505bb803951SChris Mason 	if (push_items < src_nritems) {
506bb803951SChris Mason 		memmove(src->keys, src->keys + push_items,
507bb803951SChris Mason 			(src_nritems - push_items) * sizeof(struct key));
508bb803951SChris Mason 		memmove(src->blockptrs, src->blockptrs + push_items,
509bb803951SChris Mason 			(src_nritems - push_items) * sizeof(u64));
510bb803951SChris Mason 	}
511be0e5c09SChris Mason 	src->header.nritems -= push_items;
512be0e5c09SChris Mason 	dst->header.nritems += push_items;
513be0e5c09SChris Mason 
514ed2ff2cbSChris Mason 	wret = dirty_tree_block(root, src_buf);
515bb803951SChris Mason 	if (wret < 0)
516bb803951SChris Mason 		ret = wret;
517eb60ceacSChris Mason 
518ed2ff2cbSChris Mason 	wret = dirty_tree_block(root, dst_buf);
519bb803951SChris Mason 	if (wret < 0)
520bb803951SChris Mason 		ret = wret;
521bb803951SChris Mason 	return ret;
522be0e5c09SChris Mason }
523be0e5c09SChris Mason 
52497571fd0SChris Mason /*
52579f95c82SChris Mason  * try to push data from one node into the next node right in the
52679f95c82SChris Mason  * tree.
52779f95c82SChris Mason  *
52879f95c82SChris Mason  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
52979f95c82SChris Mason  * error, and > 0 if there was no room in the right hand block.
53079f95c82SChris Mason  *
53179f95c82SChris Mason  * this will  only push up to 1/2 the contents of the left node over
53279f95c82SChris Mason  */
53379f95c82SChris Mason static int balance_node_right(struct ctree_root *root,
53479f95c82SChris Mason 			      struct tree_buffer *dst_buf,
53579f95c82SChris Mason 			      struct tree_buffer *src_buf)
53679f95c82SChris Mason {
53779f95c82SChris Mason 	struct node *src = &src_buf->node;
53879f95c82SChris Mason 	struct node *dst = &dst_buf->node;
53979f95c82SChris Mason 	int push_items = 0;
54079f95c82SChris Mason 	int max_push;
54179f95c82SChris Mason 	int src_nritems;
54279f95c82SChris Mason 	int dst_nritems;
54379f95c82SChris Mason 	int ret = 0;
54479f95c82SChris Mason 	int wret;
54579f95c82SChris Mason 
54679f95c82SChris Mason 	src_nritems = src->header.nritems;
54779f95c82SChris Mason 	dst_nritems = dst->header.nritems;
54879f95c82SChris Mason 	push_items = NODEPTRS_PER_BLOCK - dst_nritems;
54979f95c82SChris Mason 	if (push_items <= 0) {
55079f95c82SChris Mason 		return 1;
55179f95c82SChris Mason 	}
55279f95c82SChris Mason 
55379f95c82SChris Mason 	max_push = src_nritems / 2 + 1;
55479f95c82SChris Mason 	/* don't try to empty the node */
55579f95c82SChris Mason 	if (max_push > src_nritems)
55679f95c82SChris Mason 		return 1;
55779f95c82SChris Mason 	if (max_push < push_items)
55879f95c82SChris Mason 		push_items = max_push;
55979f95c82SChris Mason 
56079f95c82SChris Mason 	memmove(dst->keys + push_items, dst->keys,
56179f95c82SChris Mason 		dst_nritems * sizeof(struct key));
56279f95c82SChris Mason 	memmove(dst->blockptrs + push_items, dst->blockptrs,
56379f95c82SChris Mason 		dst_nritems * sizeof(u64));
56479f95c82SChris Mason 	memcpy(dst->keys, src->keys + src_nritems - push_items,
56579f95c82SChris Mason 		push_items * sizeof(struct key));
56679f95c82SChris Mason 	memcpy(dst->blockptrs, src->blockptrs + src_nritems - push_items,
56779f95c82SChris Mason 		push_items * sizeof(u64));
56879f95c82SChris Mason 
56979f95c82SChris Mason 	src->header.nritems -= push_items;
57079f95c82SChris Mason 	dst->header.nritems += push_items;
57179f95c82SChris Mason 
572ed2ff2cbSChris Mason 	wret = dirty_tree_block(root, src_buf);
57379f95c82SChris Mason 	if (wret < 0)
57479f95c82SChris Mason 		ret = wret;
57579f95c82SChris Mason 
576ed2ff2cbSChris Mason 	wret = dirty_tree_block(root, dst_buf);
57779f95c82SChris Mason 	if (wret < 0)
57879f95c82SChris Mason 		ret = wret;
57979f95c82SChris Mason 	return ret;
58079f95c82SChris Mason }
58179f95c82SChris Mason 
58279f95c82SChris Mason /*
58397571fd0SChris Mason  * helper function to insert a new root level in the tree.
58497571fd0SChris Mason  * A new node is allocated, and a single item is inserted to
58597571fd0SChris Mason  * point to the existing root
586aa5d6bedSChris Mason  *
587aa5d6bedSChris Mason  * returns zero on success or < 0 on failure.
58897571fd0SChris Mason  */
5895de08d7dSChris Mason static int insert_new_root(struct ctree_root *root,
5905de08d7dSChris Mason 			   struct ctree_path *path, int level)
59174123bd7SChris Mason {
59274123bd7SChris Mason 	struct tree_buffer *t;
5935c680ed6SChris Mason 	struct node *lower;
5945c680ed6SChris Mason 	struct node *c;
5955c680ed6SChris Mason 	struct key *lower_key;
5965c680ed6SChris Mason 
5975c680ed6SChris Mason 	BUG_ON(path->nodes[level]);
5985c680ed6SChris Mason 	BUG_ON(path->nodes[level-1] != root->node);
5995c680ed6SChris Mason 
60074123bd7SChris Mason 	t = alloc_free_block(root);
60174123bd7SChris Mason 	c = &t->node;
60274123bd7SChris Mason 	memset(c, 0, sizeof(c));
6035c680ed6SChris Mason 	c->header.nritems = 1;
60474123bd7SChris Mason 	c->header.flags = node_level(level);
60574123bd7SChris Mason 	c->header.blocknr = t->blocknr;
606cfaa7295SChris Mason 	c->header.parentid = root->node->node.header.parentid;
60774123bd7SChris Mason 	lower = &path->nodes[level-1]->node;
60874123bd7SChris Mason 	if (is_leaf(lower->header.flags))
60974123bd7SChris Mason 		lower_key = &((struct leaf *)lower)->items[0].key;
61074123bd7SChris Mason 	else
61174123bd7SChris Mason 		lower_key = lower->keys;
61274123bd7SChris Mason 	memcpy(c->keys, lower_key, sizeof(struct key));
61374123bd7SChris Mason 	c->blockptrs[0] = path->nodes[level-1]->blocknr;
614cfaa7295SChris Mason 	/* the super has an extra ref to root->node */
61574123bd7SChris Mason 	tree_block_release(root, root->node);
61674123bd7SChris Mason 	root->node = t;
61774123bd7SChris Mason 	t->count++;
618ed2ff2cbSChris Mason 	dirty_tree_block(root, t);
61974123bd7SChris Mason 	path->nodes[level] = t;
62074123bd7SChris Mason 	path->slots[level] = 0;
62174123bd7SChris Mason 	return 0;
62274123bd7SChris Mason }
6235c680ed6SChris Mason 
6245c680ed6SChris Mason /*
6255c680ed6SChris Mason  * worker function to insert a single pointer in a node.
6265c680ed6SChris Mason  * the node should have enough room for the pointer already
62797571fd0SChris Mason  *
6285c680ed6SChris Mason  * slot and level indicate where you want the key to go, and
6295c680ed6SChris Mason  * blocknr is the block the key points to.
630aa5d6bedSChris Mason  *
631aa5d6bedSChris Mason  * returns zero on success and < 0 on any error
6325c680ed6SChris Mason  */
633aa5d6bedSChris Mason static int insert_ptr(struct ctree_root *root,
6345c680ed6SChris Mason 		struct ctree_path *path, struct key *key,
6355c680ed6SChris Mason 		u64 blocknr, int slot, int level)
6365c680ed6SChris Mason {
6375c680ed6SChris Mason 	struct node *lower;
6385c680ed6SChris Mason 	int nritems;
6395c680ed6SChris Mason 
6405c680ed6SChris Mason 	BUG_ON(!path->nodes[level]);
64174123bd7SChris Mason 	lower = &path->nodes[level]->node;
64274123bd7SChris Mason 	nritems = lower->header.nritems;
64374123bd7SChris Mason 	if (slot > nritems)
64474123bd7SChris Mason 		BUG();
64574123bd7SChris Mason 	if (nritems == NODEPTRS_PER_BLOCK)
64674123bd7SChris Mason 		BUG();
64774123bd7SChris Mason 	if (slot != nritems) {
64874123bd7SChris Mason 		memmove(lower->keys + slot + 1, lower->keys + slot,
64974123bd7SChris Mason 			(nritems - slot) * sizeof(struct key));
65074123bd7SChris Mason 		memmove(lower->blockptrs + slot + 1, lower->blockptrs + slot,
65174123bd7SChris Mason 			(nritems - slot) * sizeof(u64));
65274123bd7SChris Mason 	}
65374123bd7SChris Mason 	memcpy(lower->keys + slot, key, sizeof(struct key));
65474123bd7SChris Mason 	lower->blockptrs[slot] = blocknr;
65574123bd7SChris Mason 	lower->header.nritems++;
65674123bd7SChris Mason 	if (lower->keys[1].objectid == 0)
65774123bd7SChris Mason 			BUG();
658ed2ff2cbSChris Mason 	dirty_tree_block(root, path->nodes[level]);
65974123bd7SChris Mason 	return 0;
66074123bd7SChris Mason }
66174123bd7SChris Mason 
66297571fd0SChris Mason /*
66397571fd0SChris Mason  * split the node at the specified level in path in two.
66497571fd0SChris Mason  * The path is corrected to point to the appropriate node after the split
66597571fd0SChris Mason  *
66697571fd0SChris Mason  * Before splitting this tries to make some room in the node by pushing
66797571fd0SChris Mason  * left and right, if either one works, it returns right away.
668aa5d6bedSChris Mason  *
669aa5d6bedSChris Mason  * returns 0 on success and < 0 on failure
67097571fd0SChris Mason  */
671aa5d6bedSChris Mason static int split_node(struct ctree_root *root, struct ctree_path *path,
672aa5d6bedSChris Mason 		      int level)
673be0e5c09SChris Mason {
6745c680ed6SChris Mason 	struct tree_buffer *t;
6755c680ed6SChris Mason 	struct node *c;
6765c680ed6SChris Mason 	struct tree_buffer *split_buffer;
6775c680ed6SChris Mason 	struct node *split;
678be0e5c09SChris Mason 	int mid;
6795c680ed6SChris Mason 	int ret;
680aa5d6bedSChris Mason 	int wret;
681be0e5c09SChris Mason 
6825c680ed6SChris Mason 	t = path->nodes[level];
683eb60ceacSChris Mason 	c = &t->node;
6845c680ed6SChris Mason 	if (t == root->node) {
6855c680ed6SChris Mason 		/* trying to split the root, lets make a new one */
6865c680ed6SChris Mason 		ret = insert_new_root(root, path, level + 1);
6875c680ed6SChris Mason 		if (ret)
6885c680ed6SChris Mason 			return ret;
6895c680ed6SChris Mason 	}
6905c680ed6SChris Mason 	split_buffer = alloc_free_block(root);
6915c680ed6SChris Mason 	split = &split_buffer->node;
6925c680ed6SChris Mason 	split->header.flags = c->header.flags;
6935c680ed6SChris Mason 	split->header.blocknr = split_buffer->blocknr;
6945c680ed6SChris Mason 	split->header.parentid = root->node->node.header.parentid;
695be0e5c09SChris Mason 	mid = (c->header.nritems + 1) / 2;
6965c680ed6SChris Mason 	memcpy(split->keys, c->keys + mid,
697be0e5c09SChris Mason 		(c->header.nritems - mid) * sizeof(struct key));
6985c680ed6SChris Mason 	memcpy(split->blockptrs, c->blockptrs + mid,
699be0e5c09SChris Mason 		(c->header.nritems - mid) * sizeof(u64));
7005c680ed6SChris Mason 	split->header.nritems = c->header.nritems - mid;
701be0e5c09SChris Mason 	c->header.nritems = mid;
702aa5d6bedSChris Mason 	ret = 0;
703aa5d6bedSChris Mason 
704ed2ff2cbSChris Mason 	wret = dirty_tree_block(root, t);
705aa5d6bedSChris Mason 	if (wret)
706aa5d6bedSChris Mason 		ret = wret;
707ed2ff2cbSChris Mason 	wret = dirty_tree_block(root, split_buffer);
708aa5d6bedSChris Mason 	if (wret)
709aa5d6bedSChris Mason 		ret = wret;
710aa5d6bedSChris Mason 	wret = insert_ptr(root, path, split->keys, split_buffer->blocknr,
7115c680ed6SChris Mason 			  path->slots[level + 1] + 1, level + 1);
712aa5d6bedSChris Mason 	if (wret)
713aa5d6bedSChris Mason 		ret = wret;
714aa5d6bedSChris Mason 
7155de08d7dSChris Mason 	if (path->slots[level] >= mid) {
7165c680ed6SChris Mason 		path->slots[level] -= mid;
7175c680ed6SChris Mason 		tree_block_release(root, t);
7185c680ed6SChris Mason 		path->nodes[level] = split_buffer;
7195c680ed6SChris Mason 		path->slots[level + 1] += 1;
720eb60ceacSChris Mason 	} else {
7215c680ed6SChris Mason 		tree_block_release(root, split_buffer);
722be0e5c09SChris Mason 	}
723aa5d6bedSChris Mason 	return ret;
724be0e5c09SChris Mason }
725be0e5c09SChris Mason 
72674123bd7SChris Mason /*
72774123bd7SChris Mason  * how many bytes are required to store the items in a leaf.  start
72874123bd7SChris Mason  * and nr indicate which items in the leaf to check.  This totals up the
72974123bd7SChris Mason  * space used both by the item structs and the item data
73074123bd7SChris Mason  */
731aa5d6bedSChris Mason static int leaf_space_used(struct leaf *l, int start, int nr)
732be0e5c09SChris Mason {
733be0e5c09SChris Mason 	int data_len;
734be0e5c09SChris Mason 	int end = start + nr - 1;
735be0e5c09SChris Mason 
736be0e5c09SChris Mason 	if (!nr)
737be0e5c09SChris Mason 		return 0;
738be0e5c09SChris Mason 	data_len = l->items[start].offset + l->items[start].size;
739be0e5c09SChris Mason 	data_len = data_len - l->items[end].offset;
740be0e5c09SChris Mason 	data_len += sizeof(struct item) * nr;
741be0e5c09SChris Mason 	return data_len;
742be0e5c09SChris Mason }
743be0e5c09SChris Mason 
74474123bd7SChris Mason /*
74500ec4c51SChris Mason  * push some data in the path leaf to the right, trying to free up at
74600ec4c51SChris Mason  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
747aa5d6bedSChris Mason  *
748aa5d6bedSChris Mason  * returns 1 if the push failed because the other node didn't have enough
749aa5d6bedSChris Mason  * room, 0 if everything worked out and < 0 if there were major errors.
75000ec4c51SChris Mason  */
751aa5d6bedSChris Mason static int push_leaf_right(struct ctree_root *root, struct ctree_path *path,
75200ec4c51SChris Mason 			   int data_size)
75300ec4c51SChris Mason {
75400ec4c51SChris Mason 	struct tree_buffer *left_buf = path->nodes[0];
75500ec4c51SChris Mason 	struct leaf *left = &left_buf->leaf;
75600ec4c51SChris Mason 	struct leaf *right;
75700ec4c51SChris Mason 	struct tree_buffer *right_buf;
75800ec4c51SChris Mason 	struct tree_buffer *upper;
75900ec4c51SChris Mason 	int slot;
76000ec4c51SChris Mason 	int i;
76100ec4c51SChris Mason 	int free_space;
76200ec4c51SChris Mason 	int push_space = 0;
76300ec4c51SChris Mason 	int push_items = 0;
76400ec4c51SChris Mason 	struct item *item;
76500ec4c51SChris Mason 
76600ec4c51SChris Mason 	slot = path->slots[1];
76700ec4c51SChris Mason 	if (!path->nodes[1]) {
76800ec4c51SChris Mason 		return 1;
76900ec4c51SChris Mason 	}
77000ec4c51SChris Mason 	upper = path->nodes[1];
77100ec4c51SChris Mason 	if (slot >= upper->node.header.nritems - 1) {
77200ec4c51SChris Mason 		return 1;
77300ec4c51SChris Mason 	}
77400ec4c51SChris Mason 	right_buf = read_tree_block(root, upper->node.blockptrs[slot + 1]);
77500ec4c51SChris Mason 	right = &right_buf->leaf;
77600ec4c51SChris Mason 	free_space = leaf_free_space(right);
77700ec4c51SChris Mason 	if (free_space < data_size + sizeof(struct item)) {
77800ec4c51SChris Mason 		tree_block_release(root, right_buf);
77900ec4c51SChris Mason 		return 1;
78000ec4c51SChris Mason 	}
78100ec4c51SChris Mason 	for (i = left->header.nritems - 1; i >= 0; i--) {
78200ec4c51SChris Mason 		item = left->items + i;
78300ec4c51SChris Mason 		if (path->slots[0] == i)
78400ec4c51SChris Mason 			push_space += data_size + sizeof(*item);
78500ec4c51SChris Mason 		if (item->size + sizeof(*item) + push_space > free_space)
78600ec4c51SChris Mason 			break;
78700ec4c51SChris Mason 		push_items++;
78800ec4c51SChris Mason 		push_space += item->size + sizeof(*item);
78900ec4c51SChris Mason 	}
79000ec4c51SChris Mason 	if (push_items == 0) {
79100ec4c51SChris Mason 		tree_block_release(root, right_buf);
79200ec4c51SChris Mason 		return 1;
79300ec4c51SChris Mason 	}
79400ec4c51SChris Mason 	/* push left to right */
79500ec4c51SChris Mason 	push_space = left->items[left->header.nritems - push_items].offset +
79600ec4c51SChris Mason 		     left->items[left->header.nritems - push_items].size;
79700ec4c51SChris Mason 	push_space -= leaf_data_end(left);
79800ec4c51SChris Mason 	/* make room in the right data area */
79900ec4c51SChris Mason 	memmove(right->data + leaf_data_end(right) - push_space,
80000ec4c51SChris Mason 		right->data + leaf_data_end(right),
80100ec4c51SChris Mason 		LEAF_DATA_SIZE - leaf_data_end(right));
80200ec4c51SChris Mason 	/* copy from the left data area */
80300ec4c51SChris Mason 	memcpy(right->data + LEAF_DATA_SIZE - push_space,
80400ec4c51SChris Mason 		left->data + leaf_data_end(left),
80500ec4c51SChris Mason 		push_space);
80600ec4c51SChris Mason 	memmove(right->items + push_items, right->items,
80700ec4c51SChris Mason 		right->header.nritems * sizeof(struct item));
80800ec4c51SChris Mason 	/* copy the items from left to right */
80900ec4c51SChris Mason 	memcpy(right->items, left->items + left->header.nritems - push_items,
81000ec4c51SChris Mason 		push_items * sizeof(struct item));
81100ec4c51SChris Mason 
81200ec4c51SChris Mason 	/* update the item pointers */
81300ec4c51SChris Mason 	right->header.nritems += push_items;
81400ec4c51SChris Mason 	push_space = LEAF_DATA_SIZE;
81500ec4c51SChris Mason 	for (i = 0; i < right->header.nritems; i++) {
81600ec4c51SChris Mason 		right->items[i].offset = push_space - right->items[i].size;
81700ec4c51SChris Mason 		push_space = right->items[i].offset;
81800ec4c51SChris Mason 	}
81900ec4c51SChris Mason 	left->header.nritems -= push_items;
82000ec4c51SChris Mason 
821ed2ff2cbSChris Mason 	dirty_tree_block(root, left_buf);
822ed2ff2cbSChris Mason 	dirty_tree_block(root, right_buf);
82300ec4c51SChris Mason 	memcpy(upper->node.keys + slot + 1,
82400ec4c51SChris Mason 		&right->items[0].key, sizeof(struct key));
825ed2ff2cbSChris Mason 	dirty_tree_block(root, upper);
82600ec4c51SChris Mason 	/* then fixup the leaf pointer in the path */
82700ec4c51SChris Mason 	if (path->slots[0] >= left->header.nritems) {
82800ec4c51SChris Mason 		path->slots[0] -= left->header.nritems;
82900ec4c51SChris Mason 		tree_block_release(root, path->nodes[0]);
83000ec4c51SChris Mason 		path->nodes[0] = right_buf;
83100ec4c51SChris Mason 		path->slots[1] += 1;
83200ec4c51SChris Mason 	} else {
83300ec4c51SChris Mason 		tree_block_release(root, right_buf);
83400ec4c51SChris Mason 	}
83500ec4c51SChris Mason 	return 0;
83600ec4c51SChris Mason }
83700ec4c51SChris Mason /*
83874123bd7SChris Mason  * push some data in the path leaf to the left, trying to free up at
83974123bd7SChris Mason  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
84074123bd7SChris Mason  */
841aa5d6bedSChris Mason static int push_leaf_left(struct ctree_root *root, struct ctree_path *path,
842be0e5c09SChris Mason 			  int data_size)
843be0e5c09SChris Mason {
844eb60ceacSChris Mason 	struct tree_buffer *right_buf = path->nodes[0];
845eb60ceacSChris Mason 	struct leaf *right = &right_buf->leaf;
846eb60ceacSChris Mason 	struct tree_buffer *t;
847be0e5c09SChris Mason 	struct leaf *left;
848be0e5c09SChris Mason 	int slot;
849be0e5c09SChris Mason 	int i;
850be0e5c09SChris Mason 	int free_space;
851be0e5c09SChris Mason 	int push_space = 0;
852be0e5c09SChris Mason 	int push_items = 0;
853be0e5c09SChris Mason 	struct item *item;
854be0e5c09SChris Mason 	int old_left_nritems;
855aa5d6bedSChris Mason 	int ret = 0;
856aa5d6bedSChris Mason 	int wret;
857be0e5c09SChris Mason 
858be0e5c09SChris Mason 	slot = path->slots[1];
859be0e5c09SChris Mason 	if (slot == 0) {
860be0e5c09SChris Mason 		return 1;
861be0e5c09SChris Mason 	}
862be0e5c09SChris Mason 	if (!path->nodes[1]) {
863be0e5c09SChris Mason 		return 1;
864be0e5c09SChris Mason 	}
865eb60ceacSChris Mason 	t = read_tree_block(root, path->nodes[1]->node.blockptrs[slot - 1]);
866eb60ceacSChris Mason 	left = &t->leaf;
867be0e5c09SChris Mason 	free_space = leaf_free_space(left);
868be0e5c09SChris Mason 	if (free_space < data_size + sizeof(struct item)) {
869eb60ceacSChris Mason 		tree_block_release(root, t);
870be0e5c09SChris Mason 		return 1;
871be0e5c09SChris Mason 	}
872be0e5c09SChris Mason 	for (i = 0; i < right->header.nritems; i++) {
873be0e5c09SChris Mason 		item = right->items + i;
874be0e5c09SChris Mason 		if (path->slots[0] == i)
875be0e5c09SChris Mason 			push_space += data_size + sizeof(*item);
876be0e5c09SChris Mason 		if (item->size + sizeof(*item) + push_space > free_space)
877be0e5c09SChris Mason 			break;
878be0e5c09SChris Mason 		push_items++;
879be0e5c09SChris Mason 		push_space += item->size + sizeof(*item);
880be0e5c09SChris Mason 	}
881be0e5c09SChris Mason 	if (push_items == 0) {
882eb60ceacSChris Mason 		tree_block_release(root, t);
883be0e5c09SChris Mason 		return 1;
884be0e5c09SChris Mason 	}
885be0e5c09SChris Mason 	/* push data from right to left */
886be0e5c09SChris Mason 	memcpy(left->items + left->header.nritems,
887be0e5c09SChris Mason 		right->items, push_items * sizeof(struct item));
888be0e5c09SChris Mason 	push_space = LEAF_DATA_SIZE - right->items[push_items -1].offset;
889be0e5c09SChris Mason 	memcpy(left->data + leaf_data_end(left) - push_space,
890be0e5c09SChris Mason 		right->data + right->items[push_items - 1].offset,
891be0e5c09SChris Mason 		push_space);
892be0e5c09SChris Mason 	old_left_nritems = left->header.nritems;
893eb60ceacSChris Mason 	BUG_ON(old_left_nritems < 0);
894eb60ceacSChris Mason 
895be0e5c09SChris Mason 	for(i = old_left_nritems; i < old_left_nritems + push_items; i++) {
896be0e5c09SChris Mason 		left->items[i].offset -= LEAF_DATA_SIZE -
897be0e5c09SChris Mason 			left->items[old_left_nritems -1].offset;
898be0e5c09SChris Mason 	}
899be0e5c09SChris Mason 	left->header.nritems += push_items;
900be0e5c09SChris Mason 
901be0e5c09SChris Mason 	/* fixup right node */
902be0e5c09SChris Mason 	push_space = right->items[push_items-1].offset - leaf_data_end(right);
903be0e5c09SChris Mason 	memmove(right->data + LEAF_DATA_SIZE - push_space, right->data +
904be0e5c09SChris Mason 		leaf_data_end(right), push_space);
905be0e5c09SChris Mason 	memmove(right->items, right->items + push_items,
906be0e5c09SChris Mason 		(right->header.nritems - push_items) * sizeof(struct item));
907be0e5c09SChris Mason 	right->header.nritems -= push_items;
908be0e5c09SChris Mason 	push_space = LEAF_DATA_SIZE;
909eb60ceacSChris Mason 
910be0e5c09SChris Mason 	for (i = 0; i < right->header.nritems; i++) {
911be0e5c09SChris Mason 		right->items[i].offset = push_space - right->items[i].size;
912be0e5c09SChris Mason 		push_space = right->items[i].offset;
913be0e5c09SChris Mason 	}
914eb60ceacSChris Mason 
915ed2ff2cbSChris Mason 	wret = dirty_tree_block(root, t);
916aa5d6bedSChris Mason 	if (wret)
917aa5d6bedSChris Mason 		ret = wret;
918ed2ff2cbSChris Mason 	wret = dirty_tree_block(root, right_buf);
919aa5d6bedSChris Mason 	if (wret)
920aa5d6bedSChris Mason 		ret = wret;
921eb60ceacSChris Mason 
922aa5d6bedSChris Mason 	wret = fixup_low_keys(root, path, &right->items[0].key, 1);
923aa5d6bedSChris Mason 	if (wret)
924aa5d6bedSChris Mason 		ret = wret;
925be0e5c09SChris Mason 
926be0e5c09SChris Mason 	/* then fixup the leaf pointer in the path */
927be0e5c09SChris Mason 	if (path->slots[0] < push_items) {
928be0e5c09SChris Mason 		path->slots[0] += old_left_nritems;
929eb60ceacSChris Mason 		tree_block_release(root, path->nodes[0]);
930eb60ceacSChris Mason 		path->nodes[0] = t;
931be0e5c09SChris Mason 		path->slots[1] -= 1;
932be0e5c09SChris Mason 	} else {
933eb60ceacSChris Mason 		tree_block_release(root, t);
934be0e5c09SChris Mason 		path->slots[0] -= push_items;
935be0e5c09SChris Mason 	}
936eb60ceacSChris Mason 	BUG_ON(path->slots[0] < 0);
937aa5d6bedSChris Mason 	return ret;
938be0e5c09SChris Mason }
939be0e5c09SChris Mason 
94074123bd7SChris Mason /*
94174123bd7SChris Mason  * split the path's leaf in two, making sure there is at least data_size
94274123bd7SChris Mason  * available for the resulting leaf level of the path.
943aa5d6bedSChris Mason  *
944aa5d6bedSChris Mason  * returns 0 if all went well and < 0 on failure.
94574123bd7SChris Mason  */
946aa5d6bedSChris Mason static int split_leaf(struct ctree_root *root, struct ctree_path *path,
947aa5d6bedSChris Mason 		      int data_size)
948be0e5c09SChris Mason {
949aa5d6bedSChris Mason 	struct tree_buffer *l_buf;
950aa5d6bedSChris Mason 	struct leaf *l;
951eb60ceacSChris Mason 	int nritems;
952eb60ceacSChris Mason 	int mid;
953eb60ceacSChris Mason 	int slot;
954be0e5c09SChris Mason 	struct leaf *right;
955eb60ceacSChris Mason 	struct tree_buffer *right_buffer;
956be0e5c09SChris Mason 	int space_needed = data_size + sizeof(struct item);
957be0e5c09SChris Mason 	int data_copy_size;
958be0e5c09SChris Mason 	int rt_data_off;
959be0e5c09SChris Mason 	int i;
960be0e5c09SChris Mason 	int ret;
961aa5d6bedSChris Mason 	int wret;
962be0e5c09SChris Mason 
963aa5d6bedSChris Mason 	wret = push_leaf_left(root, path, data_size);
964aa5d6bedSChris Mason 	if (wret < 0)
965aa5d6bedSChris Mason 		return wret;
966aa5d6bedSChris Mason 	if (wret) {
967aa5d6bedSChris Mason 		wret = push_leaf_right(root, path, data_size);
968aa5d6bedSChris Mason 		if (wret < 0)
969aa5d6bedSChris Mason 			return wret;
970aa5d6bedSChris Mason 	}
971eb60ceacSChris Mason 	l_buf = path->nodes[0];
972eb60ceacSChris Mason 	l = &l_buf->leaf;
973aa5d6bedSChris Mason 
974aa5d6bedSChris Mason 	/* did the pushes work? */
975eb60ceacSChris Mason 	if (leaf_free_space(l) >= sizeof(struct item) + data_size)
976be0e5c09SChris Mason 		return 0;
977aa5d6bedSChris Mason 
9785c680ed6SChris Mason 	if (!path->nodes[1]) {
9795c680ed6SChris Mason 		ret = insert_new_root(root, path, 1);
9805c680ed6SChris Mason 		if (ret)
9815c680ed6SChris Mason 			return ret;
9825c680ed6SChris Mason 	}
983eb60ceacSChris Mason 	slot = path->slots[0];
984eb60ceacSChris Mason 	nritems = l->header.nritems;
985eb60ceacSChris Mason 	mid = (nritems + 1)/ 2;
986eb60ceacSChris Mason 
987eb60ceacSChris Mason 	right_buffer = alloc_free_block(root);
988eb60ceacSChris Mason 	BUG_ON(!right_buffer);
989eb60ceacSChris Mason 	BUG_ON(mid == nritems);
990eb60ceacSChris Mason 	right = &right_buffer->leaf;
991be0e5c09SChris Mason 	memset(right, 0, sizeof(*right));
992be0e5c09SChris Mason 	if (mid <= slot) {
99397571fd0SChris Mason 		/* FIXME, just alloc a new leaf here */
994be0e5c09SChris Mason 		if (leaf_space_used(l, mid, nritems - mid) + space_needed >
995be0e5c09SChris Mason 			LEAF_DATA_SIZE)
996be0e5c09SChris Mason 			BUG();
997be0e5c09SChris Mason 	} else {
99897571fd0SChris Mason 		/* FIXME, just alloc a new leaf here */
999be0e5c09SChris Mason 		if (leaf_space_used(l, 0, mid + 1) + space_needed >
1000be0e5c09SChris Mason 			LEAF_DATA_SIZE)
1001be0e5c09SChris Mason 			BUG();
1002be0e5c09SChris Mason 	}
1003be0e5c09SChris Mason 	right->header.nritems = nritems - mid;
1004eb60ceacSChris Mason 	right->header.blocknr = right_buffer->blocknr;
1005eb60ceacSChris Mason 	right->header.flags = node_level(0);
1006cfaa7295SChris Mason 	right->header.parentid = root->node->node.header.parentid;
1007be0e5c09SChris Mason 	data_copy_size = l->items[mid].offset + l->items[mid].size -
1008be0e5c09SChris Mason 			 leaf_data_end(l);
1009be0e5c09SChris Mason 	memcpy(right->items, l->items + mid,
1010be0e5c09SChris Mason 	       (nritems - mid) * sizeof(struct item));
1011be0e5c09SChris Mason 	memcpy(right->data + LEAF_DATA_SIZE - data_copy_size,
1012be0e5c09SChris Mason 	       l->data + leaf_data_end(l), data_copy_size);
1013be0e5c09SChris Mason 	rt_data_off = LEAF_DATA_SIZE -
1014be0e5c09SChris Mason 		     (l->items[mid].offset + l->items[mid].size);
101574123bd7SChris Mason 
101674123bd7SChris Mason 	for (i = 0; i < right->header.nritems; i++)
1017be0e5c09SChris Mason 		right->items[i].offset += rt_data_off;
101874123bd7SChris Mason 
1019be0e5c09SChris Mason 	l->header.nritems = mid;
1020aa5d6bedSChris Mason 	ret = 0;
1021aa5d6bedSChris Mason 	wret = insert_ptr(root, path, &right->items[0].key,
10225c680ed6SChris Mason 			  right_buffer->blocknr, path->slots[1] + 1, 1);
1023aa5d6bedSChris Mason 	if (wret)
1024aa5d6bedSChris Mason 		ret = wret;
1025ed2ff2cbSChris Mason 	wret = dirty_tree_block(root, right_buffer);
1026aa5d6bedSChris Mason 	if (wret)
1027aa5d6bedSChris Mason 		ret = wret;
1028ed2ff2cbSChris Mason 	wret = dirty_tree_block(root, l_buf);
1029aa5d6bedSChris Mason 	if (wret)
1030aa5d6bedSChris Mason 		ret = wret;
1031eb60ceacSChris Mason 
1032eb60ceacSChris Mason 	BUG_ON(path->slots[0] != slot);
1033be0e5c09SChris Mason 	if (mid <= slot) {
1034eb60ceacSChris Mason 		tree_block_release(root, path->nodes[0]);
1035eb60ceacSChris Mason 		path->nodes[0] = right_buffer;
1036be0e5c09SChris Mason 		path->slots[0] -= mid;
1037be0e5c09SChris Mason 		path->slots[1] += 1;
1038eb60ceacSChris Mason 	} else
1039eb60ceacSChris Mason 		tree_block_release(root, right_buffer);
1040eb60ceacSChris Mason 	BUG_ON(path->slots[0] < 0);
1041be0e5c09SChris Mason 	return ret;
1042be0e5c09SChris Mason }
1043be0e5c09SChris Mason 
104474123bd7SChris Mason /*
104574123bd7SChris Mason  * Given a key and some data, insert an item into the tree.
104674123bd7SChris Mason  * This does all the path init required, making room in the tree if needed.
104774123bd7SChris Mason  */
1048be0e5c09SChris Mason int insert_item(struct ctree_root *root, struct key *key,
1049be0e5c09SChris Mason 			  void *data, int data_size)
1050be0e5c09SChris Mason {
1051aa5d6bedSChris Mason 	int ret = 0;
1052aa5d6bedSChris Mason 	int wret;
1053be0e5c09SChris Mason 	int slot;
1054eb60ceacSChris Mason 	int slot_orig;
1055be0e5c09SChris Mason 	struct leaf *leaf;
1056eb60ceacSChris Mason 	struct tree_buffer *leaf_buf;
1057be0e5c09SChris Mason 	unsigned int nritems;
1058be0e5c09SChris Mason 	unsigned int data_end;
1059be0e5c09SChris Mason 	struct ctree_path path;
1060be0e5c09SChris Mason 
106174123bd7SChris Mason 	/* create a root if there isn't one */
10625c680ed6SChris Mason 	if (!root->node)
1063cfaa7295SChris Mason 		BUG();
1064be0e5c09SChris Mason 	init_path(&path);
10655c680ed6SChris Mason 	ret = search_slot(root, key, &path, data_size);
1066eb60ceacSChris Mason 	if (ret == 0) {
1067eb60ceacSChris Mason 		release_path(root, &path);
1068ed2ff2cbSChris Mason 		ret = -EEXIST;
1069ed2ff2cbSChris Mason 		wret = commit_transaction(root);
1070ed2ff2cbSChris Mason 		if (wret)
1071ed2ff2cbSChris Mason 			ret = wret;
1072aa5d6bedSChris Mason 		return ret;
1073aa5d6bedSChris Mason 	}
1074ed2ff2cbSChris Mason 	if (ret < 0)
1075ed2ff2cbSChris Mason 		goto out;
1076be0e5c09SChris Mason 
1077eb60ceacSChris Mason 	slot_orig = path.slots[0];
1078eb60ceacSChris Mason 	leaf_buf = path.nodes[0];
1079eb60ceacSChris Mason 	leaf = &leaf_buf->leaf;
108074123bd7SChris Mason 
1081be0e5c09SChris Mason 	nritems = leaf->header.nritems;
1082be0e5c09SChris Mason 	data_end = leaf_data_end(leaf);
1083eb60ceacSChris Mason 
1084be0e5c09SChris Mason 	if (leaf_free_space(leaf) <  sizeof(struct item) + data_size)
1085be0e5c09SChris Mason 		BUG();
1086be0e5c09SChris Mason 
1087be0e5c09SChris Mason 	slot = path.slots[0];
1088eb60ceacSChris Mason 	BUG_ON(slot < 0);
1089be0e5c09SChris Mason 	if (slot != nritems) {
1090be0e5c09SChris Mason 		int i;
1091be0e5c09SChris Mason 		unsigned int old_data = leaf->items[slot].offset +
1092be0e5c09SChris Mason 					leaf->items[slot].size;
1093be0e5c09SChris Mason 
1094be0e5c09SChris Mason 		/*
1095be0e5c09SChris Mason 		 * item0..itemN ... dataN.offset..dataN.size .. data0.size
1096be0e5c09SChris Mason 		 */
1097be0e5c09SChris Mason 		/* first correct the data pointers */
1098be0e5c09SChris Mason 		for (i = slot; i < nritems; i++)
1099be0e5c09SChris Mason 			leaf->items[i].offset -= data_size;
1100be0e5c09SChris Mason 
1101be0e5c09SChris Mason 		/* shift the items */
1102be0e5c09SChris Mason 		memmove(leaf->items + slot + 1, leaf->items + slot,
1103be0e5c09SChris Mason 		        (nritems - slot) * sizeof(struct item));
1104be0e5c09SChris Mason 
1105be0e5c09SChris Mason 		/* shift the data */
1106be0e5c09SChris Mason 		memmove(leaf->data + data_end - data_size, leaf->data +
1107be0e5c09SChris Mason 		        data_end, old_data - data_end);
1108be0e5c09SChris Mason 		data_end = old_data;
1109be0e5c09SChris Mason 	}
111074123bd7SChris Mason 	/* copy the new data in */
1111be0e5c09SChris Mason 	memcpy(&leaf->items[slot].key, key, sizeof(struct key));
1112be0e5c09SChris Mason 	leaf->items[slot].offset = data_end - data_size;
1113be0e5c09SChris Mason 	leaf->items[slot].size = data_size;
1114be0e5c09SChris Mason 	memcpy(leaf->data + data_end - data_size, data, data_size);
1115be0e5c09SChris Mason 	leaf->header.nritems += 1;
1116aa5d6bedSChris Mason 
1117aa5d6bedSChris Mason 	ret = 0;
11188e19f2cdSChris Mason 	if (slot == 0)
1119aa5d6bedSChris Mason 		ret = fixup_low_keys(root, &path, key, 1);
1120aa5d6bedSChris Mason 
1121ed2ff2cbSChris Mason 	wret = dirty_tree_block(root, leaf_buf);
1122aa5d6bedSChris Mason 	if (wret)
1123aa5d6bedSChris Mason 		ret = wret;
1124aa5d6bedSChris Mason 
1125be0e5c09SChris Mason 	if (leaf_free_space(leaf) < 0)
1126be0e5c09SChris Mason 		BUG();
1127bb803951SChris Mason 	check_leaf(&path, 0);
1128ed2ff2cbSChris Mason out:
1129eb60ceacSChris Mason 	release_path(root, &path);
1130ed2ff2cbSChris Mason 	wret = commit_transaction(root);
1131ed2ff2cbSChris Mason 	if (wret)
1132ed2ff2cbSChris Mason 		ret = wret;
1133aa5d6bedSChris Mason 	return ret;
1134be0e5c09SChris Mason }
1135be0e5c09SChris Mason 
113674123bd7SChris Mason /*
11375de08d7dSChris Mason  * delete the pointer from a given node.
113874123bd7SChris Mason  *
113974123bd7SChris Mason  * If the delete empties a node, the node is removed from the tree,
114074123bd7SChris Mason  * continuing all the way the root if required.  The root is converted into
114174123bd7SChris Mason  * a leaf if all the nodes are emptied.
114274123bd7SChris Mason  */
1143bb803951SChris Mason static int del_ptr(struct ctree_root *root, struct ctree_path *path, int level,
1144bb803951SChris Mason 		   int slot)
1145be0e5c09SChris Mason {
1146be0e5c09SChris Mason 	struct node *node;
1147bb803951SChris Mason 	struct tree_buffer *parent = path->nodes[level];
1148be0e5c09SChris Mason 	int nritems;
1149aa5d6bedSChris Mason 	int ret = 0;
1150bb803951SChris Mason 	int wret;
1151be0e5c09SChris Mason 
1152bb803951SChris Mason 	node = &parent->node;
1153be0e5c09SChris Mason 	nritems = node->header.nritems;
1154be0e5c09SChris Mason 
1155be0e5c09SChris Mason 	if (slot != nritems -1) {
1156be0e5c09SChris Mason 		memmove(node->keys + slot, node->keys + slot + 1,
1157be0e5c09SChris Mason 			sizeof(struct key) * (nritems - slot - 1));
1158be0e5c09SChris Mason 		memmove(node->blockptrs + slot,
1159be0e5c09SChris Mason 			node->blockptrs + slot + 1,
1160be0e5c09SChris Mason 			sizeof(u64) * (nritems - slot - 1));
1161be0e5c09SChris Mason 	}
1162be0e5c09SChris Mason 	node->header.nritems--;
1163bb803951SChris Mason 	if (node->header.nritems == 0 && parent == root->node) {
1164bb803951SChris Mason 		BUG_ON(node_level(root->node->node.header.flags) != 1);
1165eb60ceacSChris Mason 		/* just turn the root into a leaf and break */
1166eb60ceacSChris Mason 		root->node->node.header.flags = node_level(0);
1167bb803951SChris Mason 	} else if (slot == 0) {
1168bb803951SChris Mason 		wret = fixup_low_keys(root, path, node->keys, level + 1);
11690f70abe2SChris Mason 		if (wret)
11700f70abe2SChris Mason 			ret = wret;
1171be0e5c09SChris Mason 	}
1172ed2ff2cbSChris Mason 	wret = dirty_tree_block(root, parent);
1173bb803951SChris Mason 	if (wret)
1174bb803951SChris Mason 		ret = wret;
1175aa5d6bedSChris Mason 	return ret;
1176be0e5c09SChris Mason }
1177be0e5c09SChris Mason 
117874123bd7SChris Mason /*
117974123bd7SChris Mason  * delete the item at the leaf level in path.  If that empties
118074123bd7SChris Mason  * the leaf, remove it from the tree
118174123bd7SChris Mason  */
11824920c9acSChris Mason int del_item(struct ctree_root *root, struct ctree_path *path)
1183be0e5c09SChris Mason {
1184be0e5c09SChris Mason 	int slot;
1185be0e5c09SChris Mason 	struct leaf *leaf;
1186eb60ceacSChris Mason 	struct tree_buffer *leaf_buf;
1187be0e5c09SChris Mason 	int doff;
1188be0e5c09SChris Mason 	int dsize;
1189aa5d6bedSChris Mason 	int ret = 0;
1190aa5d6bedSChris Mason 	int wret;
1191be0e5c09SChris Mason 
1192eb60ceacSChris Mason 	leaf_buf = path->nodes[0];
1193eb60ceacSChris Mason 	leaf = &leaf_buf->leaf;
11944920c9acSChris Mason 	slot = path->slots[0];
1195be0e5c09SChris Mason 	doff = leaf->items[slot].offset;
1196be0e5c09SChris Mason 	dsize = leaf->items[slot].size;
1197be0e5c09SChris Mason 
1198be0e5c09SChris Mason 	if (slot != leaf->header.nritems - 1) {
1199be0e5c09SChris Mason 		int i;
1200be0e5c09SChris Mason 		int data_end = leaf_data_end(leaf);
1201be0e5c09SChris Mason 		memmove(leaf->data + data_end + dsize,
1202be0e5c09SChris Mason 			leaf->data + data_end,
1203be0e5c09SChris Mason 			doff - data_end);
1204be0e5c09SChris Mason 		for (i = slot + 1; i < leaf->header.nritems; i++)
1205be0e5c09SChris Mason 			leaf->items[i].offset += dsize;
1206be0e5c09SChris Mason 		memmove(leaf->items + slot, leaf->items + slot + 1,
1207be0e5c09SChris Mason 			sizeof(struct item) *
1208be0e5c09SChris Mason 			(leaf->header.nritems - slot - 1));
1209be0e5c09SChris Mason 	}
1210be0e5c09SChris Mason 	leaf->header.nritems -= 1;
121174123bd7SChris Mason 	/* delete the leaf if we've emptied it */
1212be0e5c09SChris Mason 	if (leaf->header.nritems == 0) {
1213eb60ceacSChris Mason 		if (leaf_buf == root->node) {
1214eb60ceacSChris Mason 			leaf->header.flags = node_level(0);
1215ed2ff2cbSChris Mason 			dirty_tree_block(root, leaf_buf);
12169a8dd150SChris Mason 		} else {
1217ed2ff2cbSChris Mason 			clean_tree_block(root, leaf_buf);
1218bb803951SChris Mason 			wret = del_ptr(root, path, 1, path->slots[1]);
1219aa5d6bedSChris Mason 			if (wret)
1220aa5d6bedSChris Mason 				ret = wret;
12210f70abe2SChris Mason 			wret = free_extent(root, leaf_buf->blocknr, 1);
12220f70abe2SChris Mason 			if (wret)
12230f70abe2SChris Mason 				ret = wret;
12249a8dd150SChris Mason 		}
1225be0e5c09SChris Mason 	} else {
12265de08d7dSChris Mason 		int used = leaf_space_used(leaf, 0, leaf->header.nritems);
1227aa5d6bedSChris Mason 		if (slot == 0) {
1228aa5d6bedSChris Mason 			wret = fixup_low_keys(root, path,
1229aa5d6bedSChris Mason 						   &leaf->items[0].key, 1);
1230aa5d6bedSChris Mason 			if (wret)
1231aa5d6bedSChris Mason 				ret = wret;
1232aa5d6bedSChris Mason 		}
1233ed2ff2cbSChris Mason 		wret = dirty_tree_block(root, leaf_buf);
1234aa5d6bedSChris Mason 		if (wret)
1235aa5d6bedSChris Mason 			ret = wret;
1236aa5d6bedSChris Mason 
123774123bd7SChris Mason 		/* delete the leaf if it is mostly empty */
12385de08d7dSChris Mason 		if (used < LEAF_DATA_SIZE / 3) {
1239be0e5c09SChris Mason 			/* push_leaf_left fixes the path.
1240be0e5c09SChris Mason 			 * make sure the path still points to our leaf
1241be0e5c09SChris Mason 			 * for possible call to del_ptr below
1242be0e5c09SChris Mason 			 */
12434920c9acSChris Mason 			slot = path->slots[1];
1244eb60ceacSChris Mason 			leaf_buf->count++;
1245aa5d6bedSChris Mason 			wret = push_leaf_left(root, path, 1);
1246aa5d6bedSChris Mason 			if (wret < 0)
1247aa5d6bedSChris Mason 				ret = wret;
1248aa5d6bedSChris Mason 			if (leaf->header.nritems) {
1249aa5d6bedSChris Mason 				wret = push_leaf_right(root, path, 1);
1250aa5d6bedSChris Mason 				if (wret < 0)
1251aa5d6bedSChris Mason 					ret = wret;
1252aa5d6bedSChris Mason 			}
1253be0e5c09SChris Mason 			if (leaf->header.nritems == 0) {
12545de08d7dSChris Mason 				u64 blocknr = leaf_buf->blocknr;
1255ed2ff2cbSChris Mason 				clean_tree_block(root, leaf_buf);
1256bb803951SChris Mason 				wret = del_ptr(root, path, 1, slot);
1257aa5d6bedSChris Mason 				if (wret)
1258aa5d6bedSChris Mason 					ret = wret;
12595de08d7dSChris Mason 				tree_block_release(root, leaf_buf);
12600f70abe2SChris Mason 				wret = free_extent(root, blocknr, 1);
12610f70abe2SChris Mason 				if (wret)
12620f70abe2SChris Mason 					ret = wret;
12635de08d7dSChris Mason 			} else {
1264eb60ceacSChris Mason 				tree_block_release(root, leaf_buf);
1265be0e5c09SChris Mason 			}
1266be0e5c09SChris Mason 		}
12679a8dd150SChris Mason 	}
1268ed2ff2cbSChris Mason 	wret = commit_transaction(root);
1269ed2ff2cbSChris Mason 	if (wret)
1270ed2ff2cbSChris Mason 		ret = wret;
1271aa5d6bedSChris Mason 	return ret;
12729a8dd150SChris Mason }
12739a8dd150SChris Mason 
127497571fd0SChris Mason /*
127597571fd0SChris Mason  * walk up the tree as far as required to find the next leaf.
12760f70abe2SChris Mason  * returns 0 if it found something or 1 if there are no greater leaves.
12770f70abe2SChris Mason  * returns < 0 on io errors.
127897571fd0SChris Mason  */
1279d97e63b6SChris Mason int next_leaf(struct ctree_root *root, struct ctree_path *path)
1280d97e63b6SChris Mason {
1281d97e63b6SChris Mason 	int slot;
1282d97e63b6SChris Mason 	int level = 1;
1283d97e63b6SChris Mason 	u64 blocknr;
1284d97e63b6SChris Mason 	struct tree_buffer *c;
1285cfaa7295SChris Mason 	struct tree_buffer *next = NULL;
1286d97e63b6SChris Mason 
1287d97e63b6SChris Mason 	while(level < MAX_LEVEL) {
1288d97e63b6SChris Mason 		if (!path->nodes[level])
12890f70abe2SChris Mason 			return 1;
1290d97e63b6SChris Mason 		slot = path->slots[level] + 1;
1291d97e63b6SChris Mason 		c = path->nodes[level];
1292d97e63b6SChris Mason 		if (slot >= c->node.header.nritems) {
1293d97e63b6SChris Mason 			level++;
1294d97e63b6SChris Mason 			continue;
1295d97e63b6SChris Mason 		}
1296d97e63b6SChris Mason 		blocknr = c->node.blockptrs[slot];
1297cfaa7295SChris Mason 		if (next)
1298cfaa7295SChris Mason 			tree_block_release(root, next);
1299d97e63b6SChris Mason 		next = read_tree_block(root, blocknr);
1300d97e63b6SChris Mason 		break;
1301d97e63b6SChris Mason 	}
1302d97e63b6SChris Mason 	path->slots[level] = slot;
1303d97e63b6SChris Mason 	while(1) {
1304d97e63b6SChris Mason 		level--;
1305d97e63b6SChris Mason 		c = path->nodes[level];
1306d97e63b6SChris Mason 		tree_block_release(root, c);
1307d97e63b6SChris Mason 		path->nodes[level] = next;
1308d97e63b6SChris Mason 		path->slots[level] = 0;
1309d97e63b6SChris Mason 		if (!level)
1310d97e63b6SChris Mason 			break;
1311d97e63b6SChris Mason 		next = read_tree_block(root, next->node.blockptrs[0]);
1312d97e63b6SChris Mason 	}
1313d97e63b6SChris Mason 	return 0;
1314d97e63b6SChris Mason }
1315d97e63b6SChris Mason 
1316