thin-provisioning-tools/btree.h

375 lines
9.4 KiB
C
Raw Normal View History

#ifndef BTREE_H
#define BTREE_H
2011-08-22 15:12:13 +05:30
#include "endian_utils.h"
#include "transaction_manager.h"
2011-07-15 19:51:28 +05:30
#include <boost/noncopyable.hpp>
#include <boost/optional.hpp>
#include <list>
//----------------------------------------------------------------
namespace persistent_data {
2011-07-15 19:51:28 +05:30
2011-07-22 20:39:56 +05:30
template <typename ValueType>
class NoOpRefCounter {
public:
void inc(ValueType const &v) {}
void dec(ValueType const &v) {}
};
2011-07-15 19:51:28 +05:30
struct uint64_traits {
typedef base::__le64 disk_type;
typedef uint64_t value_type;
2011-07-22 20:39:56 +05:30
typedef NoOpRefCounter<uint64_t> ref_counter;
2011-07-15 19:51:28 +05:30
static void unpack(disk_type const &disk, value_type &value) {
value = base::to_cpu<uint64_t>(disk);
}
static void pack(value_type const &value, disk_type &disk) {
disk = base::to_disk<base::__le64>(value);
}
};
namespace btree_detail {
using namespace base;
using namespace std;
using namespace boost;
//------------------------------------------------
// On disk data layout for btree nodes
enum node_flags {
INTERNAL_NODE = 1,
LEAF_NODE = 1 << 1
};
struct node_header {
__le32 csum;
__le32 flags;
__le64 blocknr; /* which block this node is supposed to live in */
__le32 nr_entries;
__le32 max_entries;
2011-08-23 16:25:37 +05:30
__le32 value_size;
__le32 padding;
2011-07-15 19:51:28 +05:30
} __attribute__((packed));
struct disk_node {
struct node_header header;
__le64 keys[0];
} __attribute__((packed));
enum node_type {
INTERNAL,
LEAF
};
//------------------------------------------------
// Class that acts as an interface over the raw little endian btree
// node data.
template <typename ValueTraits>
2011-07-15 19:51:28 +05:30
class node_ref {
public:
2011-08-22 18:44:10 +05:30
explicit node_ref(block_address b, disk_node *raw);
block_address get_location() const {
return location_;
}
2011-07-15 19:51:28 +05:30
2011-08-23 16:25:37 +05:30
block_address get_block_nr() const;
2011-07-15 19:51:28 +05:30
node_type get_type() const;
void set_type(node_type t);
unsigned get_nr_entries() const;
void set_nr_entries(unsigned n);
unsigned get_max_entries() const;
void set_max_entries(unsigned n);
2011-07-22 20:39:56 +05:30
// FIXME: remove this, and get the constructor to do it.
2011-07-15 19:51:28 +05:30
void set_max_entries(); // calculates the max for you.
2011-08-23 16:25:37 +05:30
size_t get_value_size() const;
2011-07-15 19:51:28 +05:30
uint64_t key_at(unsigned i) const;
void set_key(unsigned i, uint64_t k);
typename ValueTraits::value_type value_at(unsigned i) const;
void set_value(unsigned i,
typename ValueTraits::value_type const &v);
// Increments the nr_entries field
void insert_at(unsigned i,
uint64_t key,
typename ValueTraits::value_type const &v);
// Does not increment nr_entries
void overwrite_at(unsigned i,
uint64_t key,
typename ValueTraits::value_type const &v);
// 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);
// Various searches
int bsearch(uint64_t key, int want_hi) const;
optional<unsigned> exact_search(uint64_t key) const;
int lower_bound(uint64_t key) const;
2011-07-22 20:39:56 +05:30
template <typename RefCounter>
void inc_children(RefCounter &rc);
// FIXME: remove
void *raw() {
return raw_;
}
2011-07-15 19:51:28 +05:30
private:
static unsigned calc_max_entries(void);
void *key_ptr(unsigned i) const;
void *value_ptr(unsigned i) const;
2011-08-22 18:44:10 +05:30
block_address location_;
2011-07-15 19:51:28 +05:30
disk_node *raw_;
};
//------------------------------------------------
//
template <typename ValueTraits>
node_ref<ValueTraits>
to_node(typename block_manager<>::read_ref &b)
2011-07-15 19:51:28 +05:30
{
// FIXME: this should return a const read_ref somehow.
return node_ref<ValueTraits>(
2011-08-22 18:44:10 +05:30
b.get_location(),
2011-07-15 19:51:28 +05:30
reinterpret_cast<disk_node *>(
const_cast<unsigned char *>(b.data())));
}
template <typename ValueTraits>
node_ref<ValueTraits>
to_node(typename block_manager<>::write_ref &b)
2011-07-15 19:51:28 +05:30
{
return node_ref<ValueTraits>(
2011-08-22 18:44:10 +05:30
b.get_location(),
2011-07-15 19:51:28 +05:30
reinterpret_cast<disk_node *>(
const_cast<unsigned char *>(b.data())));
}
class ro_spine : private noncopyable {
public:
ro_spine(typename transaction_manager::ptr tm)
2011-07-15 19:51:28 +05:30
: tm_(tm) {
}
void step(block_address b) {
spine_.push_back(tm_->read_lock(b));
if (spine_.size() > 2)
spine_.pop_front();
}
template <typename ValueTraits>
node_ref<ValueTraits> get_node() {
return to_node<ValueTraits>(spine_.back());
2011-07-15 19:51:28 +05:30
}
private:
typename transaction_manager::ptr tm_;
std::list<typename block_manager<>::read_ref> spine_;
2011-07-15 19:51:28 +05:30
};
class shadow_spine : private noncopyable {
public:
typedef typename transaction_manager::read_ref read_ref;
typedef typename transaction_manager::write_ref write_ref;
shadow_spine(typename transaction_manager::ptr tm)
2011-07-15 19:51:28 +05:30
: tm_(tm) {
}
// true if the children of the shadow need incrementing
bool step(block_address b) {
pair<write_ref, bool> p = tm_->shadow(b);
2011-07-15 19:51:28 +05:30
try {
step(p.first);
} catch (...) {
tm_->get_sm()->dec(p.first.get_location());
throw;
}
return p.second;
}
void step(typename transaction_manager::write_ref b) {
2011-07-15 19:51:28 +05:30
spine_.push_back(b);
if (spine_.size() == 1)
root_ = spine_.front().get_location();
else if (spine_.size() > 2)
spine_.pop_front();
}
void pop() {
spine_.pop_back();
}
template <typename ValueTraits>
node_ref<ValueTraits> get_node() {
return to_node<ValueTraits>(spine_.back());
2011-07-15 19:51:28 +05:30
}
block_address get_block() const {
return spine_.back().get_location();
}
bool has_parent() const {
return spine_.size() > 1;
}
node_ref<uint64_traits> get_parent() {
2011-07-15 19:51:28 +05:30
if (spine_.size() < 2)
throw std::runtime_error("no parent");
return to_node<uint64_traits>(spine_.front());
2011-07-15 19:51:28 +05:30
}
2011-07-22 20:39:56 +05:30
block_address get_parent_location() const {
return spine_.front().get_location();
}
2011-07-15 19:51:28 +05:30
block_address get_root() const {
return root_;
}
private:
typename transaction_manager::ptr tm_;
std::list<typename block_manager<>::write_ref> spine_;
2011-07-15 19:51:28 +05:30
block_address root_;
};
2011-07-22 20:39:56 +05:30
// FIXME: make a member of btree
template <typename ValueTraits>
2011-07-15 19:51:28 +05:30
optional<typename ValueTraits::value_type>
lookup_raw(ro_spine &spine, block_address block, uint64_t key) {
2011-07-15 19:51:28 +05:30
using namespace boost;
typedef typename ValueTraits::value_type leaf_type;
for (;;) {
spine.step(block);
node_ref<ValueTraits> leaf = spine.template get_node<ValueTraits>();
2011-07-15 19:51:28 +05:30
optional<unsigned> mi = leaf.exact_search(key);
2011-07-15 19:51:28 +05:30
if (!mi)
return optional<leaf_type>();
if (leaf.get_type() == btree_detail::LEAF)
return optional<leaf_type>(leaf.value_at(*mi));
node_ref<uint64_traits> internal = spine.template get_node<uint64_traits>();
2011-07-15 19:51:28 +05:30
block = internal.value_at(*mi);
}
}
}
template <unsigned Levels, typename ValueTraits>
class btree {
public:
typedef boost::shared_ptr<btree<Levels, ValueTraits> > ptr;
2011-07-15 19:51:28 +05:30
typedef uint64_t key[Levels];
typedef typename ValueTraits::value_type value_type;
typedef boost::optional<value_type> maybe_value;
typedef boost::optional<std::pair<unsigned, value_type> > maybe_pair;
typedef typename block_manager<>::read_ref read_ref;
typedef typename block_manager<>::write_ref write_ref;
typedef typename btree_detail::node_ref<ValueTraits> leaf_node;
typedef typename btree_detail::node_ref<uint64_traits> internal_node;
btree(typename persistent_data::transaction_manager::ptr tm,
2011-07-22 20:39:56 +05:30
typename ValueTraits::ref_counter rc);
btree(typename transaction_manager::ptr tm,
2011-07-22 20:39:56 +05:30
block_address root,
typename ValueTraits::ref_counter rc);
~btree();
maybe_value lookup(key const &key) const;
maybe_pair lookup_le(key const &key) const;
maybe_pair lookup_ge(key const &key) const;
void insert(key const &key, typename ValueTraits::value_type const &value);
void remove(key const &key);
void set_root(block_address root);
block_address get_root() const;
ptr clone() const;
// free the on disk btree when the destructor is called
void destroy();
// Derive a class from this base class if you need to
// inspect the individual nodes that make up a btree.
class visitor {
public:
virtual ~visitor() {}
typedef boost::shared_ptr<visitor> ptr;
// The bool return values indicate whether the walk
// should be continued into sub trees of the node (true == continue).
virtual bool visit_internal(unsigned level, bool is_root, internal_node const &n) = 0;
virtual bool visit_internal_leaf(unsigned level, bool is_root, internal_node const &n) = 0;
virtual bool visit_leaf(unsigned level, bool is_root, leaf_node const &n) = 0;
};
// Walks the tree in depth first order
2011-08-25 20:57:58 +05:30
void visit(typename visitor::ptr visitor) const;
private:
2011-07-15 19:51:28 +05:30
template <typename ValueTraits2>
void split_node(btree_detail::shadow_spine &spine,
2011-07-15 19:51:28 +05:30
block_address parent_index,
uint64_t key,
bool top);
2011-06-27 15:15:30 +05:30
2011-07-15 19:51:28 +05:30
template <typename ValueTraits2>
void split_beneath(btree_detail::shadow_spine &spine, uint64_t key);
2011-06-27 15:15:30 +05:30
2011-07-15 19:51:28 +05:30
template <typename ValueTraits2>
void split_sibling(btree_detail::shadow_spine &spine,
2011-07-15 19:51:28 +05:30
block_address parent_index,
uint64_t key);
2011-06-27 15:15:30 +05:30
2011-07-15 19:51:28 +05:30
template <typename ValueTraits2>
bool
insert_location(btree_detail::shadow_spine &spine,
2011-07-15 19:51:28 +05:30
block_address block,
uint64_t key,
int *index);
void walk_tree(typename visitor::ptr visitor,
unsigned level, bool is_root,
2011-08-25 20:57:58 +05:30
block_address b) const;
typename persistent_data::transaction_manager::ptr tm_;
2011-07-15 19:51:28 +05:30
bool destroy_;
block_address root_;
2011-07-22 20:39:56 +05:30
NoOpRefCounter<uint64_t> internal_rc_;
typename ValueTraits::ref_counter rc_;
2011-06-27 15:15:30 +05:30
};
};
2011-06-27 15:15:30 +05:30
#include "btree.tcc"
//----------------------------------------------------------------
#endif