[btree] Implement btree::remove()

persistent-data/data-structures/btree-remove.tcc

@@ -0,0 +1,373 @@
+// This file is part of the thin-provisioning-tools source.
+//
+// thin-provisioning-tools is free software: you can redistribute it
+// and/or modify it under the terms of the GNU General Public License
+// as published by the Free Software Foundation, either version 3 of
+// the License, or (at your option) any later version.
+//
+// thin-provisioning-tools is distributed in the hope that it will be
+// useful, but WITHOUT ANY WARRANTY; without even the implied warranty
+// of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+//
+// You should have received a copy of the GNU General Public License along
+// with thin-provisioning-tools.  If not, see
+// <http://www.gnu.org/licenses/>.
+
+namespace persistent_data {
+	template <unsigned Levels, typename ValueTraits>
+	btree_detail::shadow_child
+	btree<Levels, ValueTraits>::
+	create_shadow_child(internal_node &parent,
+			    unsigned index)
+	{
+		block_address b = parent.value_at(index);
+
+		pair<write_ref, bool> p = tm_.shadow(b, validator_);
+		write_ref &wr = p.first;
+		btree_detail::node_type type;
+
+		node_ref<block_traits> n = to_node<block_traits>(wr);
+		if (n.get_type() == btree_detail::INTERNAL) {
+			type = btree_detail::INTERNAL;
+			if (p.second)
+				n.inc_children(internal_rc_);
+		} else {
+			type = btree_detail::LEAF;
+			if (p.second) {
+				node_ref<ValueTraits> leaf = to_node<ValueTraits>(wr);
+				leaf.inc_children(rc_);
+			}
+		}
+
+		parent.set_value(index, wr.get_location());
+
+		return btree_detail::shadow_child(wr, type);
+	}
+
+	template <unsigned Levels, typename ValueTraits>
+	void
+	btree<Levels, ValueTraits>::
+	remove(key const &key)
+	{
+		using namespace btree_detail;
+
+		block_address block = root_;
+		unsigned index = 0;
+		shadow_spine spine(tm_, validator_);
+		bool need_remove = true;
+
+		for (unsigned level = 0; level < Levels - 1; ++level) {
+			need_remove = remove_location<block_traits>(spine, block,
+								    key[level], &index,
+								    internal_rc_);
+			if (!need_remove)
+				break;
+
+			internal_node n = spine.get_node<block_traits>();
+			block = n.value_at(index);
+		}
+
+		if (need_remove) {
+			need_remove = remove_location<ValueTraits>(spine, block,
+								   key[Levels - 1], &index,
+								   rc_);
+			if (need_remove) {
+				leaf_node leaf = spine.get_node<ValueTraits>();
+				leaf.delete_at(index);
+			}
+		}
+
+		root_ = spine.get_root();
+	}
+
+	template <unsigned Levels, typename _>
+	template <typename ValueTraits, typename RC>
+	bool
+	btree<Levels, _>::
+	remove_location(btree_detail::shadow_spine &spine,
+			block_address block,
+			uint64_t key,
+			unsigned *index,
+			RC &leaf_rc)
+	{
+		using namespace btree_detail;
+
+		unsigned i = 0;
+		bool r = false;
+
+		for (;;) {
+			r = spine.step(block);
+
+			// patch up the parent to point to the new shadow
+			if (spine.has_parent()) {
+				internal_node p = spine.get_parent();
+				p.set_value(i, spine.get_block());
+			}
+
+			internal_node n = spine.get_node<block_traits>();
+			if (n.get_type() == btree_detail::LEAF) {
+				node_ref<ValueTraits> leaf = spine.get_node<ValueTraits>();
+				boost::optional<unsigned> idx = leaf.exact_search(key);
+				if (!idx)
+					return false;
+				*index = *idx;
+				return true;
+			}
+
+			r = rebalance_children<ValueTraits>(spine, key);
+			if (!r)
+				break;
+
+			n = spine.get_node<block_traits>();
+			if (n.get_type() == btree_detail::LEAF) {
+				node_ref<ValueTraits> leaf = spine.get_node<ValueTraits>();
+				boost::optional<unsigned> idx = leaf.exact_search(key);
+				if (!idx)
+					return false;
+				*index = *idx;
+				return true;
+			}
+
+			i = n.lower_bound(key);
+			block = n.value_at(i);
+		}
+
+		return r;
+	}
+
+	template <unsigned Levels, typename _>
+	template <typename ValueTraits>
+	bool
+	btree<Levels, _>::
+	rebalance_children(btree_detail::shadow_spine &spine, uint64_t key)
+	{
+		internal_node n = spine.get_node<block_traits>();
+
+		if (n.get_nr_entries() == 1) {
+			block_address b = n.value_at(0);
+			read_ref child = tm_.read_lock(b, validator_);
+
+			// FIXME: is it safe?
+			::memcpy(n.raw(), child.data(), read_ref::BLOCK_SIZE);
+
+			tm_.get_sm()->dec(child.get_location());
+			return true;
+		}
+
+		int i = n.lower_bound(key);
+		if (i < 0)
+			return false;
+
+		bool has_left_sibling = i > 0;
+		bool has_right_sibling = static_cast<unsigned>(i) < (n.get_nr_entries() - 1);
+
+		if (!has_left_sibling)
+			rebalance2<ValueTraits>(spine, i);
+		else if (!has_right_sibling)
+			rebalance2<ValueTraits>(spine, i - 1);
+		else
+			rebalance3<ValueTraits>(spine, i - 1);
+
+		return true;
+	}
+
+	template <unsigned Levels, typename _>
+	template <typename ValueTraits>
+	void
+	btree<Levels, _>::
+	rebalance2(btree_detail::shadow_spine &spine, unsigned left_index)
+	{
+		internal_node parent = spine.get_node<block_traits>();
+		shadow_child left = create_shadow_child(parent, left_index);
+		shadow_child right = create_shadow_child(parent, left_index + 1);
+
+		// FIXME: ugly
+		if (left.get_type() == btree_detail::INTERNAL) {
+			internal_node l = left.get_node<block_traits>();
+			internal_node r = right.get_node<block_traits>();
+			__rebalance2(parent, l, r, left_index);
+		} else {
+			node_ref<ValueTraits> l = left.get_node<ValueTraits>();
+			node_ref<ValueTraits> r = right.get_node<ValueTraits>();
+			__rebalance2(parent, l, r, left_index);
+		}
+	}
+
+	template <unsigned Levels, typename _>
+	template <typename ValueTraits>
+	void
+	btree<Levels, _>::
+	__rebalance2(internal_node &parent,
+		     node_ref<ValueTraits> &left,
+		     node_ref<ValueTraits> &right,
+		     unsigned left_index)
+	{
+		unsigned nr_left = left.get_nr_entries();
+		unsigned nr_right = right.get_nr_entries();
+		unsigned right_index = left_index + 1;
+
+		unsigned threshold = 2 * (left.merge_threshold() + 1);
+		if (nr_left + nr_right < threshold) {
+			// Merge the right child into the left
+			left.copy_entries_to_left(right, nr_right);
+			left.set_nr_entries(nr_left + nr_right);
+			parent.delete_at(right_index);
+			tm_.get_sm()->dec(right.get_location());
+		} else {
+			// Rebalance
+			unsigned target_left = (nr_left + nr_right) / 2;
+			left.move_entries(right, nr_left - target_left);
+			parent.set_key(right_index, right.key_at(0));
+		}
+	}
+
+	template <unsigned Levels, typename _>
+	template <typename ValueTraits>
+	void
+	btree<Levels, _>::
+	rebalance3(btree_detail::shadow_spine &spine, unsigned left_index)
+	{
+		internal_node parent = spine.get_node<block_traits>();
+		shadow_child left = create_shadow_child(parent, left_index);
+		shadow_child center = create_shadow_child(parent, left_index + 1);
+		shadow_child right = create_shadow_child(parent, left_index + 2);
+
+		// FIXME: ugly
+		if (left.get_type() == btree_detail::INTERNAL) {
+			internal_node l = left.get_node<block_traits>();
+			internal_node c = center.get_node<block_traits>();
+			internal_node r = right.get_node<block_traits>();
+			__rebalance3(parent, l, c, r, left_index);
+		} else {
+			node_ref<ValueTraits> l = left.get_node<ValueTraits>();
+			node_ref<ValueTraits> c = center.get_node<ValueTraits>();
+			node_ref<ValueTraits> r = right.get_node<ValueTraits>();
+			__rebalance3(parent, l, c, r, left_index);
+		}
+	}
+
+	template <unsigned Levels, typename _>
+	template <typename ValueTraits>
+	void
+	btree<Levels, _>::
+	__rebalance3(internal_node &parent,
+		     node_ref<ValueTraits> &left,
+		     node_ref<ValueTraits> &center,
+		     node_ref<ValueTraits> &right,
+		     unsigned left_index)
+	{
+		unsigned nr_left = left.get_nr_entries();
+		unsigned nr_center = center.get_nr_entries();
+		unsigned nr_right = right.get_nr_entries();
+
+		unsigned threshold = left.merge_threshold() * 4 + 1;
+
+		if ((nr_left + nr_center + nr_right) < threshold)
+			delete_center_node(parent, left, center, right, left_index);
+		else
+			redistribute3(parent, left, center, right, left_index);
+	}
+
+	template <unsigned Levels, typename _>
+	template <typename ValueTraits>
+	void
+	btree<Levels, _>::
+	delete_center_node(internal_node &parent,
+			   node_ref<ValueTraits> &left,
+			   node_ref<ValueTraits> &center,
+			   node_ref<ValueTraits> &right,
+			   unsigned left_index)
+	{
+		unsigned center_index = left_index + 1;
+		unsigned right_index = left_index + 2;
+
+		unsigned max_entries = left.get_max_entries();
+		unsigned nr_left = left.get_nr_entries();
+		unsigned nr_center = center.get_nr_entries();
+		unsigned nr_right = right.get_nr_entries();
+		unsigned shift = std::min(max_entries - nr_left, nr_center);
+
+		if (nr_left + shift > max_entries)
+			throw std::runtime_error("too many entries");
+
+		left.copy_entries_to_left(center, shift);
+		left.set_nr_entries(nr_left + shift);
+
+		if (shift != nr_center) {
+			shift = nr_center - shift;
+			if ((nr_right + shift) > max_entries)
+				throw std::runtime_error("too many entries");
+			right.shift_entries_right(shift);
+			center.copy_entries_to_right(right, shift);
+			right.set_nr_entries(nr_right + shift);
+		}
+		parent.set_key(right_index, right.key_at(0));
+
+		parent.delete_at(center_index);
+		--right_index;
+
+		tm_.get_sm()->dec(center.get_location());
+		__rebalance2(parent, left, right, left_index);
+	}
+
+	template <unsigned Levels, typename _>
+	template <typename ValueTraits>
+	void
+	btree<Levels, _>::
+	redistribute3(internal_node &parent,
+		      node_ref<ValueTraits> &left,
+		      node_ref<ValueTraits> &center,
+		      node_ref<ValueTraits> &right,
+		      unsigned left_index)
+	{
+		unsigned center_index = left_index + 1;
+		unsigned right_index = left_index + 2;
+
+		unsigned nr_left = left.get_nr_entries();
+		unsigned nr_center = center.get_nr_entries();
+		unsigned nr_right = right.get_nr_entries();
+
+		unsigned max_entries = left.get_max_entries();
+		unsigned total = nr_left + nr_center + nr_right;
+		unsigned target_right = total / 3;
+		unsigned remainder = (target_right * 3) != total;
+		unsigned target_left = target_right + remainder;
+
+		if (target_left > max_entries || target_right > max_entries)
+			throw std::runtime_error("too many entries");
+
+		if (nr_left < nr_right) {
+			int s = nr_left - target_left;
+
+			// FIXME: signed & unsigned comparison
+			if (s < 0 && nr_center < static_cast<unsigned>(-s)) {
+				// not enough in central node
+				left.move_entries(center, -nr_center);
+				s += nr_center;
+				left.move_entries(right, s);
+				nr_right += s;
+			} else
+				left.move_entries(center, s);
+
+			center.move_entries(right, target_right - nr_right);
+
+		} else {
+			int s = target_right - nr_right;
+
+			if (s > 0 && nr_center < static_cast<unsigned>(s)) {
+				// not enough in central node
+				center.move_entries(right, nr_center);
+				s -= nr_center;
+				left.move_entries(right, s);
+				nr_left -= s;
+			} else
+				center.move_entries(right, s);
+
+			left.move_entries(center, nr_left - target_left);
+		}
+
+		parent.set_key(center_index, center.key_at(0));
+		parent.set_key(right_index, right.key_at(0));
+	}
+};

persistent-data/data-structures/btree.h

@@ -110,12 +110,34 @@ namespace persistent_data {
 					  uint64_t key,
 					  typename ValueTraits::value_type const &v);
 
+			// Decrements the nr_entries field
+			void delete_at(unsigned i);
+
 			// Copies entries from another node, appends them
 			// to the back of this node.  Adjusts nr_entries.
 			void copy_entries(node_ref const &rhs,
 					  unsigned begin,
 					  unsigned end);
 
+			// Moves entries between the sibling node,
+			// and maintains the key ordering.
+			// The nr_entreis of both nodes are adjusted.
+			void move_entries(node_ref &rhs,
+					  int count);
+
+			// Copies entries from the beginning of rhs to the end of lhs,
+			// or copies entries from the end of lhs to the beginning of rhs.
+			// The nr_entries is not adjusted.
+			void copy_entries_to_left(node_ref const &rhs, unsigned count);
+			void copy_entries_to_right(node_ref &rhs, unsigned count) const;
+
+			// Shifts entries to left or right.
+			// The nr_entries is not adjusted.
+			void shift_entries_left(unsigned shift);
+			void shift_entries_right(unsigned shift);
+
+			unsigned merge_threshold() const;
+
 			// Various searches
 			int bsearch(uint64_t key, int want_hi) const;
 			boost::optional<unsigned> exact_search(uint64_t key) const;
@@ -259,6 +281,26 @@ namespace persistent_data {
 		        maybe_block root_;
 		};
 
+		class shadow_child {
+		public:
+			shadow_child(block_manager::write_ref &wr, node_type type)
+				: wr_(wr), type_(type) {
+			}
+
+			node_type get_type() const {
+				return type_;
+			}
+
+			template <typename ValueTraits>
+			node_ref<ValueTraits> get_node() {
+				return to_node<ValueTraits>(wr_);
+			}
+
+		private:
+			block_manager::write_ref wr_;
+			node_type type_;
+		};
+
 		// Used to keep a record of a nested btree's position.
 		typedef std::vector<uint64_t> btree_path;
 
@@ -399,6 +441,14 @@ namespace persistent_data {
 				int *index,
 				RC &leaf_rc);
 
+		template <typename ValueTraits2, typename RC>
+		bool
+		remove_location(btree_detail::shadow_spine &spine,
+				block_address block,
+				uint64_t key,
+				unsigned *index,
+				RC &leaf_rc);
+
 		void walk_tree(visitor &visitor,
 			       btree_detail::node_location const &loc,
 			       block_address b) const;
@@ -411,6 +461,53 @@ namespace persistent_data {
 		void inc_children(btree_detail::shadow_spine &spine,
 				  RefCounter &leaf_rc);
 
+		btree_detail::shadow_child
+		create_shadow_child(internal_node &parent,
+				    unsigned index);
+
+		template <typename ValueTraits2>
+		bool rebalance_children(btree_detail::shadow_spine &spine,
+					uint64_t key);
+
+		template <typename ValueTraits2>
+		void rebalance2(btree_detail::shadow_spine &spine,
+				unsigned left_index);
+
+		template <typename ValueTraits2>
+		void rebalance3(btree_detail::shadow_spine &spine,
+				unsigned left_index);
+
+		template <typename ValueTraits2>
+		void
+		__rebalance2(internal_node &parent,
+			     btree_detail::node_ref<ValueTraits2> &left,
+			     btree_detail::node_ref<ValueTraits2> &right,
+			     unsigned left_index);
+
+		template <typename ValueTraits2>
+		void
+		__rebalance3(internal_node &parent,
+			     btree_detail::node_ref<ValueTraits2> &left,
+			     btree_detail::node_ref<ValueTraits2> &center,
+			     btree_detail::node_ref<ValueTraits2> &right,
+			     unsigned left_index);
+
+		template <typename ValueTraits2>
+		void
+		delete_center_node(internal_node &parent,
+				   btree_detail::node_ref<ValueTraits2> &left,
+				   btree_detail::node_ref<ValueTraits2> &center,
+				   btree_detail::node_ref<ValueTraits2> &right,
+				   unsigned left_index);
+
+		template <typename ValueTraits2>
+		void
+		redistribute3(internal_node &parent,
+			      btree_detail::node_ref<ValueTraits2> &left,
+			      btree_detail::node_ref<ValueTraits2> &center,
+			      btree_detail::node_ref<ValueTraits2> &right,
+			      unsigned left_index);
+
 		transaction_manager &tm_;
 		bool destroy_;
 		block_address root_;
@@ -421,6 +518,7 @@ namespace persistent_data {
 };
 
 #include "btree.tcc"
+#include "btree-remove.tcc"
 
 //----------------------------------------------------------------
 

persistent-data/data-structures/btree.tcc

@@ -293,6 +293,23 @@ namespace persistent_data {
 		set_value(i, v);
 	}
 
+	template <typename ValueTraits>
+	void
+	node_ref<ValueTraits>::delete_at(unsigned i)
+	{
+		unsigned nr_entries = get_nr_entries();
+		if (i >= nr_entries)
+			throw runtime_error("key index out of bounds");
+		unsigned nr_to_copy = nr_entries - (i + 1);
+
+		if (nr_to_copy) {
+			::memmove(key_ptr(i), key_ptr(i + 1), sizeof(uint64_t) * nr_to_copy);
+			::memmove(value_ptr(i), value_ptr(i + 1), sizeof(typename ValueTraits::disk_type) * nr_to_copy);
+		}
+
+		set_nr_entries(nr_entries - 1);
+	}
+
 	template <typename ValueTraits>
 	void
 	node_ref<ValueTraits>::copy_entries(node_ref const &rhs,
@@ -309,6 +326,90 @@ namespace persistent_data {
 		set_nr_entries(n + count);
 	}
 
+	template <typename ValueTraits>
+	void
+	node_ref<ValueTraits>::move_entries(node_ref<ValueTraits> &rhs,
+					    int count)
+	{
+		if (!count)
+			return;
+
+		unsigned nr_left = get_nr_entries();
+		unsigned nr_right = rhs.get_nr_entries();
+		unsigned max_entries = get_max_entries();
+
+		if (nr_left - count > max_entries || nr_right - count > max_entries)
+			throw runtime_error("too many entries");
+
+		if (count > 0) {
+			rhs.shift_entries_right(count);
+			copy_entries_to_right(rhs, count);
+		} else {
+			copy_entries_to_left(rhs, -count);
+			rhs.shift_entries_left(-count);
+		}
+
+		set_nr_entries(nr_left - count);
+		rhs.set_nr_entries(nr_right + count);
+	}
+
+	template <typename ValueTraits>
+	void
+	node_ref<ValueTraits>::copy_entries_to_left(node_ref const &rhs, unsigned count)
+	{
+		unsigned n = get_nr_entries();
+		if ((n + count) > get_max_entries())
+			throw runtime_error("too many entries");
+
+		::memcpy(key_ptr(n), rhs.key_ptr(0), sizeof(uint64_t) * count);
+		::memcpy(value_ptr(n), rhs.value_ptr(0), sizeof(typename ValueTraits::disk_type) * count);
+	}
+
+	template <typename ValueTraits>
+	void
+	node_ref<ValueTraits>::copy_entries_to_right(node_ref &rhs, unsigned count) const
+	{
+		unsigned n = rhs.get_nr_entries();
+		if ((n + count) > get_max_entries())
+			throw runtime_error("too many entries");
+
+		unsigned nr_left = get_nr_entries();
+		::memcpy(rhs.key_ptr(0), key_ptr(nr_left - count), sizeof(uint64_t) * count);
+		::memcpy(rhs.value_ptr(0), value_ptr(nr_left - count), sizeof(typename ValueTraits::disk_type) * count);
+	}
+
+	template <typename ValueTraits>
+	void
+	node_ref<ValueTraits>::shift_entries_left(unsigned shift)
+	{
+		unsigned n = get_nr_entries();
+		if (shift > n)
+			throw runtime_error("too many entries");
+
+		unsigned nr_shifted = n - shift;
+		::memmove(key_ptr(0), key_ptr(shift), sizeof(uint64_t) * nr_shifted);
+		::memmove(value_ptr(0), value_ptr(shift), sizeof(typename ValueTraits::disk_type) * nr_shifted);
+	}
+
+	template <typename ValueTraits>
+	void
+	node_ref<ValueTraits>::shift_entries_right(unsigned shift)
+	{
+		unsigned n = get_nr_entries();
+		if (n + shift > get_max_entries())
+			throw runtime_error("too many entries");
+
+		::memmove(key_ptr(shift), key_ptr(0), sizeof(uint64_t) * n);
+		::memmove(value_ptr(shift), value_ptr(0), sizeof(typename ValueTraits::disk_type) * n);
+	}
+
+	template <typename ValueTraits>
+	unsigned
+	node_ref<ValueTraits>::merge_threshold() const
+	{
+		return get_max_entries() / 3;
+	}
+
 	template <typename ValueTraits>
 	int
 	node_ref<ValueTraits>::bsearch(uint64_t key, int want_hi) const
@@ -601,13 +702,6 @@ namespace persistent_data {
 		return need_insert;
 	}
 
-	template <unsigned Levels, typename ValueTraits>
-	void
-	btree<Levels, ValueTraits>::remove(key const &key)
-	{
-		using namespace btree_detail;
-	}
-
 	template <unsigned Levels, typename ValueTraits>
 	block_address
 	btree<Levels, ValueTraits>::get_root() const