double buffer 实现

1. 背景

字典或者模型加载经常会用到双buffer这样读多写少的数据结构。这里介绍一个从百度的brpc拿出来的一个双buffer实现。

2. 痛点

经常会遇到的是一个读远多于写的数据结构:大部分时候,所有线程从一个不变的server列表中选取一台server。如果server列表真是“不变的”,那么选取server的过程就不用加锁,我们可以写更复杂的分流算法。一个方法是用读写锁,但当读临界区不是特别大时(毫秒级),读写锁并不比mutex快,而实用的分流算法不可能到毫秒级,否则开销也太大了。另一个方法是双缓冲,很多检索端用类似的方法实现无锁的查找过程,它大概这么工作:

  • 数据分前台和后台。
  • 检索线程只读前台,不用加锁。
  • 只有一个写线程:修改后台数据,切换前后台,睡眠一段时间,以确保老前台(新后台)不再被检索线程访问。

这个方法的问题在于它假定睡眠一段时间后就能避免和前台读线程发生竞争,这个时间一般是若干秒。由于多次写之间有间隔,这儿的写往往是批量写入,睡眠时正好用于积累数据增量。

但这套机制对“server列表”不太好用:总不能插入一个server就得等几秒钟才能插入下一个吧,即使我们用批量插入,这个"冷却"间隔多少会让用户觉得疑惑:短了担心安全性,长了觉得没有必要。我们能尽量降低这个时间并使其安全么?

3. 实现方式

我们需要写以某种形式和读同步,但读之间相互没竞争。一种解法是,读拿一把thread-local锁,写需要拿到所有的thread-local锁。具体过程如下:

  • 数据分前台和后台。
  • 读拿到自己所在线程的thread-local锁,执行查询逻辑后释放锁。
  • 同时只有一个写:修改后台数据,切换前后台,挨个获得所有thread-local锁并立刻释放,结束后再改一遍新后台(老前台)。

我们来分析下这个方法的基本原理:

  • 当一个读正在发生时,它会拿着所在线程的thread-local锁,这把锁会挡住同时进行的写,从而保证前台数据不会被修改。
  • 在大部分时候thread-local锁都没有竞争,对性能影响很小。
  • 逐个获取thread-local锁并立刻释放是为了确保对应的读线程看到了切换后的新前台。如果所有的读线程都看到了新前台,写线程便可以安全地修改老前台(新后台)了。

其他特点:

  • 不同的读之间没有竞争,高度并发。
  • 如果没有写,读总是能无竞争地获取和释放thread-local锁,一般小于25ns,对延时基本无影响。如果有写,由于其临界区极小(拿到立刻释放),读在大部分时候仍能快速地获得锁,少数时候释放锁时可能有唤醒写线程的代价。由于写本身就是少数情况,读整体上几乎不会碰到竞争锁。

完成这些功能的数据结构是DoublyBufferedData<>,我们常简称为DBD。brpc中的所有load balancer都使用了这个数据结构,使不同线程在分流时几乎不会互斥。而其他rpc实现往往使用了全局锁,这使得它们无法写出复杂的分流算法:否则分流代码将会成为竞争热点。

3. 应用场景

这个结构有广泛的应用场景:

  • reload词典。大部分时候词典都是只读的,不同线程同时查询时不应互斥。
  • 可替换的全局callback。像butil/logging.cpp支持配置全局LogSink以重定向日志,这个LogSink就是一个带状态的callback。如果只是简单的全局变量,在替换后我们无法直接删除LogSink,因为可能还有都写线程在用。用DBD可以解决这个问题。

4. 代码实现

代码是如下:

#include 
#include 
#include 
#include 
#include 
#include 
#include 
#include 
#include 
#include 
#include 
#include 

namespace common {

template 
std::lock_guard::type> get_lock_guard();

#define SCOPED_LOCK(ref_of_lock)                    \
  decltype(get_lock_guard()) \
      scoped_locker_dummy_at_line_##__LINE__(ref_of_lock)

template 
void delete_object(void* arg) {
  delete static_cast(arg);
}

// This data structure makes Read() almost lock-free by making Modify()
// *much* slower. It's very suitable for implementing LoadBalancers which
// have a lot of concurrent read-only ops from many threads and occasional
// modifications of data. As a side effect, this data structure can store
// a thread-local data for user.
//
// Read(): begin with a thread-local mutex locked then read the foreground
// instance which will not be changed before the mutex is unlocked. Since the
// mutex is only locked by Modify() with an empty critical section, the
// function is almost lock-free.
//
// Modify(): Modify background instance which is not used by any Read(), flip
// foreground and background, lock thread-local mutexes one by one to make
// sure all existing Read() finish and later Read() see new foreground,
// then modify background(foreground before flip) again.

class Void {};

template 
class DoublyBufferedData {
  class Wrapper;

 public:
  class ScopedPtr {
    friend class DoublyBufferedData;

   public:
    ScopedPtr() : _data(NULL), _w(NULL) {}
    ~ScopedPtr() {
      if (_w) {
        _w->EndRead();
      }
    }

    ScopedPtr(const ScopedPtr&) = delete;
    ScopedPtr& operator=(const ScopedPtr&) = delete;

    const T* get() const { return _data; }
    const T& operator*() const { return *_data; }
    const T* operator->() const { return _data; }
    TLS& tls() { return _w->user_tls(); }

   private:
    const T* _data;
    Wrapper* _w;
  };

  DoublyBufferedData();
  ~DoublyBufferedData();

  // Put foreground instance into ptr. The instance will not be changed until
  // ptr is destructed.
  // This function is not blocked by Read() and Modify() in other threads.
  // Returns 0 on success, -1 otherwise.
  int Read(ScopedPtr* ptr);

  // Modify background and foreground instances. fn(T&, ...) will be called
  // twice. Modify() from different threads are exclusive from each other.
  // NOTE: Call same series of fn to different equivalent instances should
  // result in equivalent instances, otherwise foreground and background
  // instance will be inconsistent.
  template 
  size_t Modify(Fn& fn);
  template 
  size_t Modify(Fn& fn, const Arg1&);
  template 
  size_t Modify(Fn& fn, const Arg1&, const Arg2&);

  // fn(T& background, const T& foreground, ...) will be called to background
  // and foreground instances respectively.
  template 
  size_t ModifyWithForeground(Fn& fn);
  template 
  size_t ModifyWithForeground(Fn& fn, const Arg1&);
  template 
  size_t ModifyWithForeground(Fn& fn, const Arg1&, const Arg2&);

 private:
  template 
  struct WithFG0 {
    WithFG0(Fn& fn, T* data) : _fn(fn), _data(data) {}
    size_t operator()(T& bg) { return _fn(bg, (const T&)_data[&bg == _data]); }

   private:
    Fn& _fn;
    T* _data;
  };

  template 
  struct WithFG1 {
    WithFG1(Fn& fn, T* data, const Arg1& arg1)
        : _fn(fn), _data(data), _arg1(arg1) {}
    size_t operator()(T& bg) {
      return _fn(bg, (const T&)_data[&bg == _data], _arg1);
    }

   private:
    Fn& _fn;
    T* _data;
    const Arg1& _arg1;
  };

  template 
  struct WithFG2 {
    WithFG2(Fn& fn, T* data, const Arg1& arg1, const Arg2& arg2)
        : _fn(fn), _data(data), _arg1(arg1), _arg2(arg2) {}
    size_t operator()(T& bg) {
      return _fn(bg, (const T&)_data[&bg == _data], _arg1, _arg2);
    }

   private:
    Fn& _fn;
    T* _data;
    const Arg1& _arg1;
    const Arg2& _arg2;
  };

  template 
  struct Closure1 {
    Closure1(Fn& fn, const Arg1& arg1) : _fn(fn), _arg1(arg1) {}
    size_t operator()(T& bg) { return _fn(bg, _arg1); }

   private:
    Fn& _fn;
    const Arg1& _arg1;
  };

  template 
  struct Closure2 {
    Closure2(Fn& fn, const Arg1& arg1, const Arg2& arg2)
        : _fn(fn), _arg1(arg1), _arg2(arg2) {}
    size_t operator()(T& bg) { return _fn(bg, _arg1, _arg2); }

   private:
    Fn& _fn;
    const Arg1& _arg1;
    const Arg2& _arg2;
  };

  const T* UnsafeRead() const {
    return _data + _index.load(std::memory_order_acquire);
  }
  Wrapper* AddWrapper();
  void RemoveWrapper(Wrapper*);

  // Foreground and background void.
  T _data[2];

  // Index of foreground instance.
  std::atomic _index;

  // Key to access thread-local wrappers.
  bool _created_key;
  pthread_key_t _wrapper_key;

  // All thread-local instances.
  std::vector _wrappers;

  // Sequence access to _wrappers.
  std::mutex _wrappers_mutex;

  // Sequence modifications.
  std::mutex _modify_mutex;
};

static const pthread_key_t INVALID_PTHREAD_KEY = (pthread_key_t)-1;

template 
class DoublyBufferedDataWrapperBase {
 public:
  TLS& user_tls() { return _user_tls; }

 protected:
  TLS _user_tls;
};

template 
class DoublyBufferedDataWrapperBase {};

template 
class DoublyBufferedData::Wrapper
    : public DoublyBufferedDataWrapperBase {
  friend class DoublyBufferedData;

 public:
  explicit Wrapper(DoublyBufferedData* c) : _control(c) {}

  ~Wrapper() {
    if (_control != NULL) {
      _control->RemoveWrapper(this);
    }
  }

  // _mutex will be locked by the calling pthread and DoublyBufferedData.
  // Most of the time, no modifications are done, so the mutex is
  // uncontended and fast.
  inline void BeginRead() { _mutex.lock(); }

  inline void EndRead() { _mutex.unlock(); }

  inline void WaitReadDone() { SCOPED_LOCK(_mutex); }

 private:
  DoublyBufferedData* _control;
  std::mutex _mutex;
};

// Called when thread initializes thread-local wrapper.
template 
typename DoublyBufferedData::Wrapper*
DoublyBufferedData::AddWrapper() {
  std::unique_ptr w(new (std::nothrow) Wrapper(this));
  if (NULL == w) {
    return NULL;
  }
  try {
    SCOPED_LOCK(_wrappers_mutex);
    _wrappers.push_back(w.get());
  } catch (std::exception& e) {
    return NULL;
  }
  return w.release();
}

// Called when thread quits.
template 
void DoublyBufferedData::RemoveWrapper(
    typename DoublyBufferedData::Wrapper* w) {
  if (NULL == w) {
    return;
  }
  SCOPED_LOCK(_wrappers_mutex);
  for (size_t i = 0; i < _wrappers.size(); ++i) {
    if (_wrappers[i] == w) {
      _wrappers[i] = _wrappers.back();
      _wrappers.pop_back();
      return;
    }
  }
}

template 
DoublyBufferedData::DoublyBufferedData()
    : _index(0), _created_key(false), _wrapper_key(0) {
  _wrappers.reserve(64);
  const int rc = pthread_key_create(&_wrapper_key, delete_object);
  if (rc != 0) {
    std::cout << "Fail to pthread_key_create: " << rc << std::endl;
  } else {
    _created_key = true;
  }
  // Initialize _data for some POD types. This is essential for pointer
  // types because they should be Read() as NULL before any Modify().
  if (std::is_integral::value || std::is_floating_point::value ||
      std::is_pointer::value || std::is_member_function_pointer::value) {
    _data[0] = T();
    _data[1] = T();
  }
}

template 
DoublyBufferedData::~DoublyBufferedData() {
  // User is responsible for synchronizations between Read()/Modify() and
  // this function.
  if (_created_key) {
    pthread_key_delete(_wrapper_key);
  }

  {
    SCOPED_LOCK(_wrappers_mutex);
    for (size_t i = 0; i < _wrappers.size(); ++i) {
      _wrappers[i]->_control = NULL;  // hack: disable removal.
      delete _wrappers[i];
    }
    _wrappers.clear();
  }
}

template 
int DoublyBufferedData::Read(
    typename DoublyBufferedData::ScopedPtr* ptr) {
  if (__builtin_expect((bool)(!_created_key), false)) {
    return -1;
  }
  Wrapper* w = static_cast(pthread_getspecific(_wrapper_key));
  if (__builtin_expect((bool)(w != NULL), true)) {
    w->BeginRead();
    ptr->_data = UnsafeRead();
    ptr->_w = w;
    return 0;
  }
  w = AddWrapper();
  if (__builtin_expect((bool)(w != NULL), true)) {
    const int rc = pthread_setspecific(_wrapper_key, w);
    if (rc == 0) {
      w->BeginRead();
      ptr->_data = UnsafeRead();
      ptr->_w = w;
      return 0;
    }
  }
  return -1;
}

template 
template 
size_t DoublyBufferedData::Modify(Fn& fn) {
  // _modify_mutex sequences modifications. Using a separate mutex rather
  // than _wrappers_mutex is to avoid blocking threads calling
  // AddWrapper() or RemoveWrapper() too long. Most of the time, modifications
  // are done by one thread, contention should be negligible.
  SCOPED_LOCK(_modify_mutex);
  int bg_index = !_index.load(std::memory_order_relaxed);
  // background instance is not accessed by other threads, being safe to
  // modify.
  const size_t ret = fn(_data[bg_index]);
  if (ret != 0) {
    std::cout << "ret:" << ret << std::endl;
    return 0;
  }

  // Publish, flip background and foreground.
  // The release fence matches with the acquire fence in UnsafeRead() to
  // make readers which just begin to read the new foreground instance see
  // all changes made in fn.
  _index.store(bg_index, std::memory_order_release);
  bg_index = !bg_index;

  // Wait until all threads finishes current reading. When they begin next
  // read, they should see updated _index.
  {
    SCOPED_LOCK(_wrappers_mutex);
    for (size_t i = 0; i < _wrappers.size(); ++i) {
      _wrappers[i]->WaitReadDone();
    }
  }

  const size_t ret2 = fn(_data[bg_index]);
  std::cout << "index=" << _index.load(std::memory_order_relaxed) << std::endl;
  return ret2;
}

template 
template 
size_t DoublyBufferedData::Modify(Fn& fn, const Arg1& arg1) {
  Closure1 c(fn, arg1);
  return Modify(c);
}

template 
template 
size_t DoublyBufferedData::Modify(Fn& fn, const Arg1& arg1,
                                          const Arg2& arg2) {
  Closure2 c(fn, arg1, arg2);
  return Modify(c);
}

template 
template 
size_t DoublyBufferedData::ModifyWithForeground(Fn& fn) {
  WithFG0 c(fn, _data);
  return Modify(c);
}

template 
template 
size_t DoublyBufferedData::ModifyWithForeground(Fn& fn,
                                                        const Arg1& arg1) {
  WithFG1 c(fn, _data, arg1);
  return Modify(c);
}

template 
template 
size_t DoublyBufferedData::ModifyWithForeground(Fn& fn,
                                                        const Arg1& arg1,
                                                        const Arg2& arg2) {
  WithFG2 c(fn, _data, arg1, arg2);
  return Modify(c);
}

}  // namespace common

struct TestData {
  std::set list;
};

class TestDoubleBuffer {
 public:
  bool Read() {
    common::DoublyBufferedData::ScopedPtr s;
    if (_reload_data.Read(&s) != 0) {
      return false;
    }

    for (auto& num : s->list) {
      std::cout << num << std::endl;
    }

    std::cout << "read end." << std::endl;

    return true;
  }

  bool Add(const int num) { return _reload_data.Modify(AddData, num); }

  bool Remove(const int num) { return _reload_data.Modify(RemoveData, num); }

  bool Reset(const std::set& num_set) {
    return _reload_data.Modify(ResetData, num_set);
  }

  bool Clear() { return _reload_data.Modify(ClearData); }

 private:
  static bool AddData(TestData& data, const int num);
  static bool RemoveData(TestData& data, const int num);
  static bool ResetData(TestData& data, const std::set& num_set);
  static bool ClearData(TestData& data);

 private:
  common::DoublyBufferedData _reload_data;
};

bool TestDoubleBuffer::AddData(TestData& data, const int num) {
  if (data.list.find(num) != data.list.end()) {
    return true;
  } else {
    data.list.insert(num);
  }
  return true;
}

bool TestDoubleBuffer::RemoveData(TestData& data, const int num) {
  auto it = data.list.find(num);
  if (it != data.list.end()) {
    data.list.erase(it);
  }
  return true;
}

bool TestDoubleBuffer::ResetData(TestData& data, const std::set& num_set) {
  data.list = num_set;
  return true;
}

bool TestDoubleBuffer::ClearData(TestData& data) {
  data.list.clear();
  return true;
}

int main() {
  std::set num_set;
  num_set.insert(3);
  num_set.insert(4);

  TestDoubleBuffer test;
  test.Add(1);
  test.Add(2);
  test.Read();
  test.Remove(1);
  test.Read();
  test.Reset(num_set);
  test.Read();

  return 0;
}

你可能感兴趣的:(微服务)