boost::pool 库速记

Boost::pool

使用示例

#include <iostream>
#include <vector>
#include <list>
#include <boost/pool/object_pool.hpp>
#include <boost/pool/pool_alloc.hpp>

#include <boost/timer/timer.hpp>

using namespace std;
using namespace boost;

const int MAXLENGTH = 100000;

class A
{
public:
    A()
    {
        cout << "Construct: " << endl;
    }
    A ( int a )
    {
        cout << "Construct: " << a << endl;
    }
    ~A()
    {
        cout << "Destruct" << endl;
    }
};
function<void ( void ) > pool_sample = []()
{
    cout << "==============================\n";
    boost::object_pool<A> p;
    A *ptr = p.construct ( 1 );
    p.destroy ( ptr );
};

function<void ( void ) > pool_sample_1 = []()
{
    cout << "==============================\n";
    boost::object_pool<A> p;
    A *ptr = p.malloc();
    cout << "malloc doesn't invoke constructor and destructor.\n";
    ptr = new ( ptr ) A ( 1 );
    ptr->~A();
    p.free ( ptr );
};

auto test_pool_alloc = []()
{
    cout << "==============================\n";
    vector<int, pool_allocator<int>> vec1;
    vector<int> vec2;
    {
        cout << "USE pool_allocator:\n";
        boost::timer::auto_cpu_timer t1;
        for ( int i = 0; i < MAXLENGTH; ++i )
        {
            vec1.push_back ( i );
            vec1.pop_back();
        }
    }
    {
        cout << "USE STL allocator:\n";
        boost::timer::auto_cpu_timer t2;
        for ( int i = 0; i < MAXLENGTH; ++i )
        {
            vec2.push_back ( i );
            vec2.pop_back();
        }
    }
};

auto test_fast_pool_alloc = []()
{
    cout << "==============================\n";
    list<int, fast_pool_allocator<int>> vec1;
    list<int> vec2;
    {
        cout << "USE fast_pool_allocator:\n";
        boost::timer::auto_cpu_timer t1;
        for ( int i = 0; i < MAXLENGTH; ++i )
        {
            vec1.push_back ( i );
            vec1.pop_back();
        }
    }
    {
        cout << "USE STL allocator:\n";
        boost::timer::auto_cpu_timer t2;
        for ( int i = 0; i < MAXLENGTH; ++i )
        {
            vec2.push_back ( i );
            vec2.pop_back();
        }
    }
};

int main()
{
    pool_sample();
    pool_sample_1();
    test_pool_alloc();
    test_fast_pool_alloc();
    system ( "pause" );
}

boost::pool 的实现原理

pool去按照一定的增长规则，从操作系统申请一大块内存，称为block，源码中用PODptr表示。

这个PODptr结构将block分为三块，第一块是大块数据区，第二块只有sizeof(void*) 个字节，即指针大小，保存下一个PODptr的指针，第三块保存下一PODptr的长度。最后一个PODptr指针为空。

PODptr的数据区被simple_segregated_storage格式化为许多个小块，称为chunk。一个chunk的大小是定义boost::object_pool时决定的，即 sizeof(T)>sizeof(void)?sizeof(T):sizeof(void)。任意一个chunk未被占用时，使用其前sizeof(void*)个字节作为一个指针指向下一个未被占用的chunk。是的，单向链表。而从pool::malloc，就执行单向链表的删除节点操作,每次都返回首个chunk，因此未进行重新申请block前，malloc都是O(1)。

pool::free(ptr)操作就是找到ptr属于哪个PODptr，然后把ptr添加到单向链表头。

pool::ordered_free(ptr)找到ptr属于哪个PODptr，然后通过插入排序把ptr添加到单向链表。

部分源码

/*
该函数是simple_segregated_storage的成员函数。第一次看到一下懵逼了，不知其何用意。难道不就是得到 *ptr 的功能吗？！
事实是，对于一个void*是不能dereference的。因为*ptr你将得到一个void类型，C++不允许void类型。
*/
static void * & nextof(void * const ptr)
{
      return *(static_cast<void **>(ptr));
}

simple_segregated_storage

//segregate会把给的一个sz大小的内存块,拆分为每个partition_sz大小的多个chunk单元，
//每个chunk的前4字节指向下一个chunk(作为链表的next)，而最后一个chunk头指向end。

template <typename SizeType>
void * simple_segregated_storage<SizeType>::segregate(
    void * const block,
    const size_type sz,
    const size_type partition_sz,
    void * const end)
{
  //找到最后一个chunk
  char * old = static_cast<char *>(block)
      + ((sz - partition_sz) / partition_sz) * partition_sz;

  nextof(old) = end;//把最后一个chunk指向end

  if (old == block)
    return block;//如果这块内存只有一个chunk就返回
  //格式化其他的chunk，使每个chunk的前4字节指向下一个chunk
  for (char * iter = old - partition_sz; iter != block;
      old = iter, iter -= partition_sz)
    nextof(iter) = old;

  nextof(block) = old;

  return block;
}

//添加一个block时，会把这该块分解成chunk，添加到链表的头部。因为无序,所以复杂度O(1)
void add_block(void * const block,
        const size_type nsz, const size_type npartition_sz)
    {
      first = segregate(block, nsz, npartition_sz, first);
    }
//通过find_prev找到这个内存块对应的位置，然后添加进去。复杂度O(n)
void add_ordered_block(void * const block,
        const size_type nsz, const size_type npartition_sz)
    {
      void * const loc = find_prev(block);
      if (loc == 0)
        add_block(block, nsz, npartition_sz);
      else
        nextof(loc) = segregate(block, nsz, npartition_sz, nextof(loc));
    }

//这个没什么好说的，通过比较地址，找到ptr在当前block中的位置，类似插入排序。
template <typename SizeType>
void * simple_segregated_storage<SizeType>::find_prev(void * const ptr)
{
  if (first == 0 || std::greater<void *>()(first, ptr))
    return 0;

  void * iter = first;
  while (true)
  {
    if (nextof(iter) == 0 || std::greater<void *>()(nextof(iter), ptr))
      return iter;

    iter = nextof(iter);
  }
}
//simple_segregated_storage成员变量。 链表头指针。
void * first;

下段代码从simple_segregated_storage链表中获取内存：

template <typename SizeType>
void * simple_segregated_storage<SizeType>::malloc_n(const size_type n,
    const size_type partition_size)
{
  if(n == 0)
    return 0;
  void * start = &first;
  void * iter;
  do
  {
    if (nextof(start) == 0)
      return 0;
    //try_malloc_n会从start开始(不算start)向后申请n个partition_size大小的chunk，返回最后一个chunk的指针
    iter = try_malloc_n(start, n, partition_size);
  } while (iter == 0);
  //此处返回内存chunk头
  void * const ret = nextof(start);
  //此处是经典的单向链表移除其中一个节点的操作。把该内存的前面chunk头指向该内存尾部chunk头指向的内存。即把该部分排除出链表。
  nextof(start) = nextof(iter);
  return ret;
}

//start会指向满足条件(连续的n个partition_size大小的chunk内存)的chunk头部，返回最后一个chunk指针。
template <typename SizeType>
void * simple_segregated_storage<SizeType>::try_malloc_n(
    void * & start, size_type n, const size_type partition_size)
{
  void * iter = nextof(start);
  //start后面的块是否是连续的n块partition_size大小的内存
  while (--n != 0)
  {
    void * next = nextof(iter);
    //如果next != static_cast<char *>(iter) + partition_size，说明下一块chunk被占用或是到了大块内存(block)的尾部。
    if (next != static_cast<char *>(iter) + partition_size)
    {
      // next == 0 (end-of-list) or non-contiguous chunk found
      start = iter;
      return 0;
    }
    iter = next;
  }
  return iter;
}

class PODptr

如上图，类PODptr指示了一个block结构，这个block大小不一定相同，但都由 chunk data+ next ptr + next block size三部分组成。

• chunk data部分被构造成一个simple_segregated_storage，切分为多个chunk，是一块连续的内存
• next ptr 指向下一个block结构，next block size指出了下一个block结构的大小。
• 也就是说，多个PODptr结构组成一个链表，而PODptr内部由simple_segregated_storage分成一个顺序表。
• PODptr的大小不固定，增长方式见void * pool<UserAllocator>::malloc_need_resize().
• 初始化的每个chunk都指向下一个chunk

class pool

//pool 从simple_segregated_storage派生
template <typename UserAllocator>
class pool: protected simple_segregated_storage < typename UserAllocator::size_type >;

//返回父类指针以便调用父类函数，其实就是类型转换
simple_segregated_storage<size_type> & store()
{ //! \returns pointer to store.
      return *this;
}

在调用pool::malloc只申请一个chunk时，如果有足够空间，使用父类指针调用malloc返回内存，否则就重新申请一个大block。代码简单，就不贴了。

下面代码是申请n个连续的chunk。如果没有连续的n个内存就需要重新分配内存了。分配好的内存，通过add_ordered_block添加到chunks的有序链表，并通过地址大小把刚申请的block放到PODptr链表的排序位置。

template <typename UserAllocator>
void * pool<UserAllocator>::ordered_malloc(const size_type n)
{ //! Gets address of a chunk n, allocating new memory if not already available.
  //! \returns Address of chunk n if allocated ok.
  //! \returns 0 if not enough memory for n chunks.

  const size_type partition_size = alloc_size();
  const size_type total_req_size = n * requested_size;
  const size_type num_chunks = total_req_size / partition_size +
      ((total_req_size % partition_size) ? true : false);

  void * ret = store().malloc_n(num_chunks, partition_size);

#ifdef BOOST_POOL_INSTRUMENT
  std::cout << "Allocating " << n << " chunks from pool of size " << partition_size << std::endl;
#endif
  if ((ret != 0) || (n == 0))
    return ret;

#ifdef BOOST_POOL_INSTRUMENT
  std::cout << "Cache miss, allocating another chunk...\n";
#endif

  // Not enough memory in our storages; make a new storage,
  BOOST_USING_STD_MAX();

  //计算下次申请内存的大小，基本就是乘以2.integer::static_lcm是求最小公倍数。
  next_size = max BOOST_PREVENT_MACRO_SUBSTITUTION(next_size, num_chunks);
  size_type POD_size = static_cast<size_type>(next_size * partition_size +
      integer::static_lcm<sizeof(size_type), sizeof(void *)>::value + sizeof(size_type));
  char * ptr = (UserAllocator::malloc)(POD_size);
  if (ptr == 0)
  {
     if(num_chunks < next_size)
     {
        // Try again with just enough memory to do the job, or at least whatever we
        // allocated last time:
        next_size >>= 1;
        next_size = max BOOST_PREVENT_MACRO_SUBSTITUTION(next_size, num_chunks);
        POD_size = static_cast<size_type>(next_size * partition_size +
            integer::static_lcm<sizeof(size_type), sizeof(void *)>::value + sizeof(size_type));
        ptr = (UserAllocator::malloc)(POD_size);
     }
     if(ptr == 0)
       return 0;
  }
  const details::PODptr<size_type> node(ptr, POD_size);

  // Split up block so we can use what wasn't requested.
  if (next_size > num_chunks)
    store().add_ordered_block(node.begin() + num_chunks * partition_size,
        node.element_size() - num_chunks * partition_size, partition_size);

  BOOST_USING_STD_MIN();
  if(!max_size)
    next_size <<= 1;
  else if( next_size*partition_size/requested_size < max_size)
    next_size = min BOOST_PREVENT_MACRO_SUBSTITUTION(next_size << 1, max_size*requested_size/ partition_size);

  //  insert it into the list,
  //   handle border case.
  //对大块block进行排序
  if (!list.valid() || std::greater<void *>()(list.begin(), node.begin()))
  {
    node.next(list);
    list = node;
  }
  else
  {
    details::PODptr<size_type> prev = list;

    while (true)
    {
      // if we're about to hit the end, or if we've found where "node" goes.
      if (prev.next_ptr() == 0
          || std::greater<void *>()(prev.next_ptr(), node.begin()))
        break;

      prev = prev.next();
    }

    node.next(prev.next());
    prev.next(node);
  }

  //  and return it.
  return node.begin();
}

下面代码是释放未被占用的块。(一个block任何一个chunk被占用就不会释放)

template <typename UserAllocator>
bool pool<UserAllocator>::release_memory()
{ //! pool must be ordered. Frees every memory block that doesn't have any allocated chunks.
  //! \returns true if at least one memory block was freed.

  // ret is the return value: it will be set to true when we actually call
  //  UserAllocator::free(..)
  bool ret = false;

  // This is a current & previous iterator pair over the memory block list
  details::PODptr<size_type> ptr = list;
  details::PODptr<size_type> prev;

  // This is a current & previous iterator pair over the free memory chunk list
  //  Note that "prev_free" in this case does NOT point to the previous memory
  //  chunk in the free list, but rather the last free memory chunk before the
  //  current block.
  void * free_p = this->first;
  void * prev_free_p = 0;

  const size_type partition_size = alloc_size();

  // Search through all the all the allocated memory blocks
  while (ptr.valid())
  {
    // At this point:
    //  ptr points to a valid memory block
    //  free_p points to either:
    //    0 if there are no more free chunks
    //    the first free chunk in this or some next memory block
    //  prev_free_p points to either:
    //    the last free chunk in some previous memory block
    //    0 if there is no such free chunk
    //  prev is either:
    //    the PODptr whose next() is ptr
    //    !valid() if there is no such PODptr

    // If there are no more free memory chunks, then every remaining
    //  block is allocated out to its fullest capacity, and we can't
    //  release any more memory
    if (free_p == 0)
      break;

    // We have to check all the chunks.  If they are *all* free (i.e., present
    //  in the free list), then we can free the block.
    bool all_chunks_free = true;

    // Iterate 'i' through all chunks in the memory block
    // if free starts in the memory block, be careful to keep it there
    void * saved_free = free_p;
    for (char * i = ptr.begin(); i != ptr.end(); i += partition_size)
    {
      // If this chunk is not free
      if (i != free_p)
      {
        // We won't be able to free this block
        all_chunks_free = false;

        // free_p might have travelled outside ptr
        free_p = saved_free;
        // Abort searching the chunks; we won't be able to free this
        //  block because a chunk is not free.
        break;
      }

      // We do not increment prev_free_p because we are in the same block
      free_p = nextof(free_p);
    }

    // post: if the memory block has any chunks, free_p points to one of them
    // otherwise, our assertions above are still valid

    const details::PODptr<size_type> next = ptr.next();

    if (!all_chunks_free)
    {
      if (is_from(free_p, ptr.begin(), ptr.element_size()))
      {
        std::less<void *> lt;
        void * const end = ptr.end();
        do
        {
          prev_free_p = free_p;
          free_p = nextof(free_p);
        } while (free_p && lt(free_p, end));
      }
      // This invariant is now restored:
      //     free_p points to the first free chunk in some next memory block, or
      //       0 if there is no such chunk.
      //     prev_free_p points to the last free chunk in this memory block.

      // We are just about to advance ptr.  Maintain the invariant:
      // prev is the PODptr whose next() is ptr, or !valid()
      // if there is no such PODptr
      prev = ptr;
    }
    else
    {
      // All chunks from this block are free

      // Remove block from list
      if (prev.valid())
        prev.next(next);
      else
        list = next;

      // Remove all entries in the free list from this block
      //关键点在这里，释放了一个block之后，会把上一个chunk头修改。
      if (prev_free_p != 0)
        nextof(prev_free_p) = free_p;
      else
        this->first = free_p;

      // And release memory
      (UserAllocator::free)(ptr.begin());
      ret = true;
    }

    // Increment ptr
    ptr = next;
  }

  next_size = start_size;
  return ret;
}

pool总结

pool的实现基本就是利用simple_segregated_storage内部实现的维护chunk的链表来实现内存管理的。simple_segregated_storage可以说是pool的核心。pool内部一共维护了两个链表：

• simple_segregated_storage内部的chunk链表。分配单个chunk时，直接从这个链表拿一个chunk，复杂度O(1)。
• pool内部有个成员变量details::PODptr<size_type> list;用来维护一个大块内存block的链表。可以知道，一个block内部是连续的，但block之间可以认为是不连续的内存。这个链表相当于一个内存地址索引，主要是为了提高查找效率：对于有序排列的内存池，归还内存时，用来快速判断是属于哪个块的。如果没有这个链表，就需要挨个chunk去判断地址大小。

class object_pool

class object_pool: protected pool<UserAllocator>;

object_pool继承自pool，但和pool的区别是，pool用于申请固定大小的内存，而object_pool用于申请固定类型的内存，并会调用构造函数和析构函数。主要的函数就两个：

调用构造函数，用到了一个placement new的方式，老生常谈。

唯一需要注意的是construct和destroy调用的malloc和free，都是调用的 ordered_malloc 和 ordered_free。

elem``ent_type * construct(Arg1&, ... ArgN&){...}
element_type * construct()
{
  element_type * const ret = (malloc)();
  if (ret == 0)
    return ret;
  try { new (ret) element_type(); }
  catch (...) { (free)(ret); throw; }
  return ret;
}
element_type * malloc BOOST_PREVENT_MACRO_SUBSTITUTION()
{
      return static_cast<element_type *>(store().ordered_malloc());
}

destroy显式调用析构函数去析构，然后把内存还给链表维护。

void destroy(element_type * const chunk)
{
  chunk->~T();
  (free)(chunk);
}
void free BOOST_PREVENT_MACRO_SUBSTITUTION(element_type * const chunk)
{
  store().ordered_free(chunk);
}

class singleton_pool

单例内存池的实现，值得注意的有如下几点：

• 单线程使用单例时(保证无同步问题)，可以通过定义宏BOOST_POOL_NO_MT来取消同步的损耗。

#if !defined(BOOST_HAS_THREADS) || defined(BOOST_NO_MT) || defined(BOOST_POOL_NO_MT)
  typedef null_mutex default_mutex;

• 单例内存池的单例实现如下，通过内部类object_creator调用private函数get_pool()，通过create_object.do_nothing();来保证在main之前实例化静态对象static object_creator create_object;

class singleton_pool
{
public:
    ...
private:
   typedef boost::aligned_storage<sizeof(pool_type), boost::alignment_of<pool_type>::value> storage_type;
   static storage_type storage;

   static pool_type& get_pool()
   {
      static bool f = false;
      if(!f)
      {
         // This code *must* be called before main() starts,
         // and when only one thread is executing.
         f = true;
         new (&storage) pool_type;
      }

      // The following line does nothing else than force the instantiation
      //  of singleton<T>::create_object, whose constructor is
      //  called before main() begins.
      create_object.do_nothing();

      return *static_cast<pool_type*>(static_cast<void*>(&storage));
   }

   struct object_creator
   {
      object_creator()
      {  // This constructor does nothing more than ensure that instance()
         //  is called before main() begins, thus creating the static
         //  T object before multithreading race issues can come up.
         singleton_pool<Tag, RequestedSize, UserAllocator, Mutex, NextSize, MaxSize>::get_pool();
      }
      inline void do_nothing() const
      {
      }
   };
   static object_creator create_object;
};

总结

• 适用范围:频繁申请释放相同大小的内存，如需要频繁的创建同一个类的对象。
• 优点:可以防止内存碎片、极快，避免频繁申请内存的调用.

boost::pool 的源代码一共就几个文件，简洁明了，读起来也不很难。由于代码时间远早于现代C++(C++11之后)成型，兼容编译器的代码建议忽略。因为重要的是其设计思想：如何通过自构两个链表来提升内存管理效率的。

数据结构很简单。适用场景比较狭窄，跟GC没法比。