Hotspot VM的Parallel GC 中 GC线程是如何创建与工作的

• 1.源码中WorkGang, GangWorker结构关系

SharedHeap的结构如下，在结构中含有成员FlexibleWorkGang* _workers;

class SharedHeap : public CollectedHeap {
  friend class VMStructs;
  friend class VM_GC_Operation;
  friend class VM_CGC_Operation;
private:
  // For claiming strong_roots tasks.
  SubTasksDone* _process_strong_tasks;
protected:
  ……
  // If we're doing parallel GC, use this gang of threads.
  FlexibleWorkGang* _workers;
  ……
}; 

FlexibleWorkGang的继承关系如下

class FlexibleWorkGang: public WorkGang {
  ……
};
class WorkGang: public AbstractWorkGang {
  ……
  Void run_task(AbstractGangTask* task);
  ……
};
class AbstractWorkGang: public CHeapObj {
  // Here's the public interface to this class.
public:
  // Constructor and destructor.
  AbstractWorkGang(const char* name, bool are_GC_task_threads,
                   bool are_ConcurrentGC_threads);
  ~AbstractWorkGang();
  // Run a task, returns when the task is done (or terminated).
  virtual void run_task(AbstractGangTask* task) = 0;
  // Stop and terminate all workers.
  virtual void stop();
public:
  // Debugging.
  const char* name() const;
protected:
  // Initialize only instance data.
  const bool _are_GC_task_threads;
  const bool _are_ConcurrentGC_threads;
  // Printing support.
  const char* _name;
  // The monitor which protects these data,
  // and notifies of changes in it.
  Monitor*  _monitor;
  // The count of the number of workers in the gang.
  int _total_workers;
  // Whether the workers should terminate.
  bool _terminate;
  // The array of worker threads for this gang.
  // This is only needed for cleaning up.
  GangWorker** _gang_workers;
  // The task for this gang.
  AbstractGangTask* _task;
  // A sequence number for the current task.
  int _sequence_number;
  // The number of started workers.
  int _started_workers;
  // The number of finished workers.
  int _finished_workers;
}; 

GangWorker的结构如下

class GangWorker: public WorkerThread {
public:
  // Constructors and destructor.
  GangWorker(AbstractWorkGang* gang, uint id);

  // The only real method: run a task for the gang.
  virtual void run();
  // Predicate for Thread
  virtual bool is_GC_task_thread() const;
  virtual bool is_ConcurrentGC_thread() const;
  // Printing
  void print_on(outputStream* st) const;
  virtual void print() const { print_on(tty); }
protected:
  AbstractWorkGang* _gang;

  virtual void initialize();
  virtual void loop();

public:
  AbstractWorkGang* gang() const { return _gang; }
}; 

 

• 2. GC线程创建

GangWorker就是GC线程，那么它是如何创建起来的呢？
当虚拟机启动的时候，会进行一些初始化操作，我们看一看是通过怎样的路径创建GangWorker的。Trace信息如下所示

#0  0x00002adb4964bace in FlexibleWorkGang (this=0x48,
    name=0xb00000000 <Address 0xb00000000 out of bounds>, workers=10971,
    are_GC_task_threads=8, are_ConcurrentGC_threads=false)
    at /home/chengtao/hotspot20/src/share/vm/utilities/workgroup.hpp:280
#1  0x00002adb4998e403 in SharedHeap (this=0x5903a430, policy_=0x5903a2f0)
    at /home/chengtao/hotspot20/src/share/vm/memory/sharedHeap.cpp:76
#2  0x00002adb49731331 in GenCollectedHeap (this=0x5903a430, policy=0x5903a2f0)
    at /home/chengtao/hotspot20/src/share/vm/memory/genCollectedHeap.cpp:76
#3  0x00002adb49a23f95 in Universe::initialize_heap ()
    at /home/chengtao/hotspot20/src/share/vm/memory/universe.cpp:921
#4  0x00002adb49a243f8 in universe_init ()
    at /home/chengtao/hotspot20/src/share/vm/memory/universe.cpp:781
#5  0x00002adb49767a52 in init_globals ()
    at /home/chengtao/hotspot20/src/share/vm/runtime/init.cpp:98
#6  0x00002adb49a0e56a in Threads::create_vm (args=0x41964080,
    canTryAgain=0x4196405b)
    at /home/chengtao/hotspot20/src/share/vm/runtime/thread.cpp:3092
#7  0x00002adb497a181c in JNI_CreateJavaVM (vm=0x419640e0, penv=0x419640e8,
    args=0x41964080)
    at /home/chengtao/hotspot20/src/share/vm/prims/jni.cpp:3344
#8  0x00000000400035f8 in InitializeJVM ()
#9  0x000000004000206e in JavaMain ()

在WorkGang进行完成构造以后，会进行初始化，调用_workers->initialize_workers();在这个函数中，进行创建GangWorker，具体如下

//创建保存GangWork的指针数组
_gang_workers = NEW_C_HEAP_ARRAY(GangWorker*, total_workers());

 

// 创建_total_workers个GangWorker，实际上_total_workers是与ParallelGCThreads相等的，ParallelGCThreads这个值是根据cpu core计算出来的一个值
//每创建一个GangWorker就会立即启动这个线程
for (int worker = 0; worker < total_workers(); worker += 1) {
    GangWorker* new_worker = allocate_worker(worker);
    assert(new_worker != NULL, "Failed to allocate GangWorker");
    _gang_workers[worker] = new_worker;
    if (new_worker == NULL || !os::create_thread(new_worker, worker_type)) {
      vm_exit_out_of_memory(0, "Cannot create worker GC thread. Out of system resources.");
      return false;
    }
    if (!DisableStartThread) {
      os::start_thread(new_worker);
    }
  }

到此，Parallel GC线程就被创建完了，并启动。

• 3. Parallel GC线程是如何工作的

我们来看一下GangWorker的一些函数，可以看到，当GC线程被创建以后，首先会进行一些初始化工作，然后就进入到了loop函数中。仔细分析这个loop函数，可以发现，刚开始GC线程是进入wait状态的，等待被唤醒然后进行GC，接下来我们看一看GC线程是如何被唤醒的。

void GangWorker::run() {
  initialize();
  loop();
}
void GangWorker::initialize() {
  this->initialize_thread_local_storage();
  assert(_gang != NULL, "No gang to run in");
  os::set_priority(this, NearMaxPriority);
  if (TraceWorkGang) {
    tty->print_cr("Running gang worker for gang %s id %d",
                  gang()->name(), id());
  }
  // The VM thread should not execute here because MutexLocker's are used
  // as (opposed to MutexLockerEx's).
  assert(!Thread::current()->is_VM_thread(), "VM thread should not be part"
         " of a work gang");
}

void GangWorker::loop() {
  int previous_sequence_number = 0;
  Monitor* gang_monitor = gang()->monitor();
  for ( ; /* !terminate() */; ) {
    WorkData data;
    int part;  // Initialized below.
    {
      // Grab the gang mutex.
      MutexLocker ml(gang_monitor);
      // Wait for something to do.
      // Polling outside the while { wait } avoids missed notifies
      // in the outer loop.
      gang()->internal_worker_poll(&data);
      if (TraceWorkGang) {
        tty->print("Polled outside for work in gang %s worker %d",
                   gang()->name(), id());
        tty->print("  terminate: %s",
                   data.terminate() ? "true" : "false");
        tty->print("  sequence: %d (prev: %d)",
                   data.sequence_number(), previous_sequence_number);
        if (data.task() != NULL) {
          tty->print("  task: %s", data.task()->name());
        } else {
          tty->print("  task: NULL");
        }
        tty->cr();
      }
      for ( ; /* break or return */; ) {
        // Terminate if requested.
        if (data.terminate()) {
          gang()->internal_note_finish();
          gang_monitor->notify_all();
          return;
        }
        // Check for new work.
        if ((data.task() != NULL) &&
            (data.sequence_number() != previous_sequence_number)) {
          gang()->internal_note_start();
          gang_monitor->notify_all();
          part = gang()->started_workers() - 1;
          break;
        }
        // Nothing to do.
        gang_monitor->wait(/* no_safepoint_check */ true);
        gang()->internal_worker_poll(&data);
        if (TraceWorkGang) {
          tty->print("Polled inside for work in gang %s worker %d",
                     gang()->name(), id());
          tty->print("  terminate: %s",
                     data.terminate() ? "true" : "false");
          tty->print("  sequence: %d (prev: %d)",
                     data.sequence_number(), previous_sequence_number);
          if (data.task() != NULL) {
            tty->print("  task: %s", data.task()->name());
          } else {
            tty->print("  task: NULL");
          }
          tty->cr();
        }
      }
      // Drop gang mutex.
    }
    if (TraceWorkGang) {
      tty->print("Work for work gang %s id %d task %s part %d",
                 gang()->name(), id(), data.task()->name(), part);
    }
    assert(data.task() != NULL, "Got null task");
    data.task()->work(part);
    {
      if (TraceWorkGang) {
        tty->print("Finish for work gang %s id %d task %s part %d",
                   gang()->name(), id(), data.task()->name(), part);
      }
      // Grab the gang mutex.
      MutexLocker ml(gang_monitor);
      gang()->internal_note_finish();
      // Tell the gang you are done.
      gang_monitor->notify_all();
      // Drop the gang mutex.
    }
    previous_sequence_number = data.sequence_number();
  }
}

当需要进行GC的时候，以ParNew为例，会进入到ParNewGeneration::collect这个函数，这个函数中有这样一段代码

  if (n_workers > 1) {
    GenCollectedHeap::StrongRootsScope srs(gch);
    workers->run_task(&tsk);
  } else {
    GenCollectedHeap::StrongRootsScope srs(gch);
    tsk.work(0);
  }

  进行Parallel GC的时候，就会进入到workers->run_task(&tsk)这个函数中了。函数原型就是void WorkGang::run_task(AbstractGangTask* task)；
在这个函数中有这样一段代码

  // Initialize.
  _task = task;
  _sequence_number += 1;
  _started_workers = 0;
  _finished_workers = 0;
  // Tell the workers to get to work.
  monitor()->notify_all();
  // Wait for them to be finished
  while (finished_workers() < total_workers()) {
    if (TraceWorkGang) {
      tty->print_cr("Waiting in work gang %s: %d/%d finished sequence %d",
                    name(), finished_workers(), total_workers(),
                    _sequence_number);
    }
    monitor()->wait(/* no_safepoint_check */ true);
  }
  _task = NULL;

    从代码中我们可以发现，monitor()->notify_all();就可以唤醒GC线程了，然后进入到wait状态，等待所有GC线程完成GC任务结束，然后退出。

• 4. 总结

总结起来很简单，就是初始时，WorkGang创建ParallelGCThreads个GC线程(GangWorker)，这些GC线程处于wait状态。当进行Paralel GC的时候，WorkGang就会唤醒wait的GC线程，进行Parallel GC。WorkGang等待Parallel GC完成后返回。