/*
* Prepares for removal from one level of the hierarchy. The caller must
* call delete_at() to remove the entry at index.
*/
static int remove_raw(struct shadow_spine *s, struct dm_btree_info *info,
struct dm_btree_value_type *vt, dm_block_t root,
uint64_t key, unsigned *index)
{
int i = *index, r;
struct btree_node *n;
for (;;) {
r = shadow_step(s, root, vt);
if (r < 0)
break;
/*
* We have to patch up the parent node, ugly, but I don't
* see a way to do this automatically as part of the spine
* op.
*/
if (shadow_has_parent(s)) {
__le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
memcpy(value_ptr(dm_block_data(shadow_parent(s)), i),
&location, sizeof(__le64));
}
n = dm_block_data(shadow_current(s));
if (le32_to_cpu(n->header.flags) & LEAF_NODE)
return do_leaf(n, key, index);
r = rebalance_children(s, info, vt, key);
if (r)
break;
n = dm_block_data(shadow_current(s));
if (le32_to_cpu(n->header.flags) & LEAF_NODE)
return do_leaf(n, key, index);
i = lower_bound(n, key);
/*
* We know the key is present, or else
* rebalance_children would have returned
* -ENODATA
*/
root = value64(n, i);
}
return r;
}
static int rebalance2(struct shadow_spine *s, struct dm_btree_info *info,
struct dm_btree_value_type *vt, unsigned left_index)
{
int r;
struct btree_node *parent;
struct child left, right;
parent = dm_block_data(shadow_current(s));
r = init_child(info, vt, parent, left_index, &left);
if (r)
return r;
r = init_child(info, vt, parent, left_index + 1, &right);
if (r) {
exit_child(info, &left);
return r;
}
__rebalance2(info, parent, &left, &right);
r = exit_child(info, &left);
if (r) {
exit_child(info, &right);
return r;
}
return exit_child(info, &right);
}
static int rebalance_children(struct shadow_spine *s,
struct dm_btree_info *info,
struct dm_btree_value_type *vt, uint64_t key)
{
int i, r, has_left_sibling, has_right_sibling;
uint32_t child_entries;
struct btree_node *n;
n = dm_block_data(shadow_current(s));
if (le32_to_cpu(n->header.nr_entries) == 1) {
struct dm_block *child;
dm_block_t b = value64(n, 0);
r = dm_tm_read_lock(info->tm, b, &btree_node_validator, &child);
if (r)
return r;
memcpy(n, dm_block_data(child),
dm_bm_block_size(dm_tm_get_bm(info->tm)));
r = dm_tm_unlock(info->tm, child);
if (r)
return r;
dm_tm_dec(info->tm, dm_block_location(child));
return 0;
}
i = lower_bound(n, key);
if (i < 0)
return -ENODATA;
r = get_nr_entries(info->tm, value64(n, i), &child_entries);
if (r)
return r;
has_left_sibling = i > 0;
has_right_sibling = i < (le32_to_cpu(n->header.nr_entries) - 1);
if (!has_left_sibling)
r = rebalance2(s, info, vt, i);
else if (!has_right_sibling)
r = rebalance2(s, info, vt, i - 1);
else
r = rebalance3(s, info, vt, i - 1);
return r;
}
static int rebalance3(struct shadow_spine *s, struct dm_btree_info *info,
struct dm_btree_value_type *vt, unsigned left_index)
{
int r;
struct btree_node *parent = dm_block_data(shadow_current(s));
struct child left, center, right;
/*
* FIXME: fill out an array?
*/
r = init_child(info, vt, parent, left_index, &left);
if (r)
return r;
r = init_child(info, vt, parent, left_index + 1, ¢er);
if (r) {
exit_child(info, &left);
return r;
}
r = init_child(info, vt, parent, left_index + 2, &right);
if (r) {
exit_child(info, &left);
exit_child(info, ¢er);
return r;
}
__rebalance3(info, parent, &left, ¢er, &right);
r = exit_child(info, &left);
if (r) {
exit_child(info, ¢er);
exit_child(info, &right);
return r;
}
r = exit_child(info, ¢er);
if (r) {
exit_child(info, &right);
return r;
}
r = exit_child(info, &right);
if (r)
return r;
return 0;
}
int dm_btree_remove(struct dm_btree_info *info, dm_block_t root,
uint64_t *keys, dm_block_t *new_root)
{
unsigned level, last_level = info->levels - 1;
int index = 0, r = 0;
struct shadow_spine spine;
struct btree_node *n;
struct dm_btree_value_type le64_vt;
init_le64_type(info->tm, &le64_vt);
init_shadow_spine(&spine, info);
for (level = 0; level < info->levels; level++) {
r = remove_raw(&spine, info,
(level == last_level ?
&info->value_type : &le64_vt),
root, keys[level], (unsigned *)&index);
if (r < 0)
break;
n = dm_block_data(shadow_current(&spine));
if (level != last_level) {
root = value64(n, index);
continue;
}
BUG_ON(index < 0 || index >= le32_to_cpu(n->header.nr_entries));
if (info->value_type.dec)
info->value_type.dec(info->value_type.context,
value_ptr(n, index));
delete_at(n, index);
}
*new_root = shadow_root(&spine);
exit_shadow_spine(&spine);
return r;
}
static int btree_insert_raw(struct shadow_spine *s, dm_block_t root,
struct dm_btree_value_type *vt,
uint64_t key, unsigned *index)
{
int r, i = *index, top = 1;
struct btree_node *node;
for (;;) {
r = shadow_step(s, root, vt);
if (r < 0)
return r;
node = dm_block_data(shadow_current(s));
/*
* We have to patch up the parent node, ugly, but I don't
* see a way to do this automatically as part of the spine
* op.
*/
if (shadow_has_parent(s) && i >= 0) { /* FIXME: second clause unness. */
__le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
__dm_bless_for_disk(&location);
memcpy_disk(value_ptr(dm_block_data(shadow_parent(s)), i),
&location, sizeof(__le64));
}
node = dm_block_data(shadow_current(s));
if (node->header.nr_entries == node->header.max_entries) {
if (top)
r = btree_split_beneath(s, key);
else
r = btree_split_sibling(s, root, i, key);
if (r < 0)
return r;
}
node = dm_block_data(shadow_current(s));
i = lower_bound(node, key);
if (le32_to_cpu(node->header.flags) & LEAF_NODE)
break;
if (i < 0) {
/* change the bounds on the lowest key */
node->keys[0] = cpu_to_le64(key);
i = 0;
}
root = value64(node, i);
top = 0;
}
if (i < 0 || le64_to_cpu(node->keys[i]) != key)
i++;
*index = i;
return 0;
}
/*
* Splits a node by creating two new children beneath the given node.
*
* Before:
* +----------+
* | A ++++++ |
* +----------+
*
*
* After:
* +------------+
* | A (shadow) |
* +------------+
* | |
* +------+ +----+
* | |
* v v
* +-------+ +-------+
* | B +++ | | C +++ |
* +-------+ +-------+
*/
static int btree_split_beneath(struct shadow_spine *s, uint64_t key)
{
int r;
size_t size;
unsigned nr_left, nr_right;
struct dm_block *left, *right, *new_parent;
struct btree_node *pn, *ln, *rn;
__le64 val;
new_parent = shadow_current(s);
r = new_block(s->info, &left);
if (r < 0)
return r;
r = new_block(s->info, &right);
if (r < 0) {
/* FIXME: put left */
return r;
}
pn = dm_block_data(new_parent);
ln = dm_block_data(left);
rn = dm_block_data(right);
nr_left = le32_to_cpu(pn->header.nr_entries) / 2;
nr_right = le32_to_cpu(pn->header.nr_entries) - nr_left;
ln->header.flags = pn->header.flags;
ln->header.nr_entries = cpu_to_le32(nr_left);
ln->header.max_entries = pn->header.max_entries;
ln->header.value_size = pn->header.value_size;
rn->header.flags = pn->header.flags;
rn->header.nr_entries = cpu_to_le32(nr_right);
rn->header.max_entries = pn->header.max_entries;
rn->header.value_size = pn->header.value_size;
memcpy(ln->keys, pn->keys, nr_left * sizeof(pn->keys[0]));
memcpy(rn->keys, pn->keys + nr_left, nr_right * sizeof(pn->keys[0]));
size = le32_to_cpu(pn->header.flags) & INTERNAL_NODE ?
sizeof(__le64) : s->info->value_type.size;
memcpy(value_ptr(ln, 0), value_ptr(pn, 0), nr_left * size);
memcpy(value_ptr(rn, 0), value_ptr(pn, nr_left),
nr_right * size);
/* new_parent should just point to l and r now */
pn->header.flags = cpu_to_le32(INTERNAL_NODE);
pn->header.nr_entries = cpu_to_le32(2);
pn->header.max_entries = cpu_to_le32(
calc_max_entries(sizeof(__le64),
dm_bm_block_size(
dm_tm_get_bm(s->info->tm))));
pn->header.value_size = cpu_to_le32(sizeof(__le64));
val = cpu_to_le64(dm_block_location(left));
__dm_bless_for_disk(&val);
pn->keys[0] = ln->keys[0];
memcpy_disk(value_ptr(pn, 0), &val, sizeof(__le64));
val = cpu_to_le64(dm_block_location(right));
__dm_bless_for_disk(&val);
pn->keys[1] = rn->keys[0];
memcpy_disk(value_ptr(pn, 1), &val, sizeof(__le64));
/*
* rejig the spine. This is ugly, since it knows too
* much about the spine
*/
if (s->nodes[0] != new_parent) {
unlock_block(s->info, s->nodes[0]);
s->nodes[0] = new_parent;
}
if (key < le64_to_cpu(rn->keys[0])) {
unlock_block(s->info, right);
s->nodes[1] = left;
} else {
unlock_block(s->info, left);
s->nodes[1] = right;
}
s->count = 2;
return 0;
}
/*
* Splits a node by creating a sibling node and shifting half the nodes
* contents across. Assumes there is a parent node, and it has room for
* another child.
*
* Before:
* +--------+
* | Parent |
* +--------+
* |
* v
* +----------+
* | A ++++++ |
* +----------+
*
*
* After:
* +--------+
* | Parent |
* +--------+
* | |
* v +------+
* +---------+ |
* | A* +++ | v
* +---------+ +-------+
* | B +++ |
* +-------+
*
* Where A* is a shadow of A.
*/
static int btree_split_sibling(struct shadow_spine *s, dm_block_t root,
unsigned parent_index, uint64_t key)
{
int r;
size_t size;
unsigned nr_left, nr_right;
struct dm_block *left, *right, *parent;
struct btree_node *ln, *rn, *pn;
__le64 location;
left = shadow_current(s);
r = new_block(s->info, &right);
if (r < 0)
return r;
ln = dm_block_data(left);
rn = dm_block_data(right);
nr_left = le32_to_cpu(ln->header.nr_entries) / 2;
nr_right = le32_to_cpu(ln->header.nr_entries) - nr_left;
ln->header.nr_entries = cpu_to_le32(nr_left);
rn->header.flags = ln->header.flags;
rn->header.nr_entries = cpu_to_le32(nr_right);
rn->header.max_entries = ln->header.max_entries;
rn->header.value_size = ln->header.value_size;
memcpy(rn->keys, ln->keys + nr_left, nr_right * sizeof(rn->keys[0]));
size = le32_to_cpu(ln->header.flags) & INTERNAL_NODE ?
sizeof(uint64_t) : s->info->value_type.size;
memcpy(value_ptr(rn, 0), value_ptr(ln, nr_left),
size * nr_right);
/*
* Patch up the parent
*/
parent = shadow_parent(s);
pn = dm_block_data(parent);
location = cpu_to_le64(dm_block_location(left));
__dm_bless_for_disk(&location);
memcpy_disk(value_ptr(pn, parent_index),
&location, sizeof(__le64));
location = cpu_to_le64(dm_block_location(right));
__dm_bless_for_disk(&location);
r = insert_at(sizeof(__le64), pn, parent_index + 1,
le64_to_cpu(rn->keys[0]), &location);
if (r)
return r;
if (key < le64_to_cpu(rn->keys[0])) {
unlock_block(s->info, right);
s->nodes[1] = left;
} else {
unlock_block(s->info, left);
s->nodes[1] = right;
}
return 0;
}
