g1-gc-init 1

_posts/2024-08-10-g1-gc-init.md

@@ -5,7 +5,7 @@ date:   2024-08-10 11:00:00 +0200
 tags: [GC, G1]
 ---
 
-[前文](https://yoa1226.github.io/2024/08/05/g1-gc-important-component.html) 介绍了 G1 非常重要多的重要概念，并介绍了相关的源码，这些概念在后续文章中会重复的提到。了解完成上述概念以后，本文将开始介绍堆初始化相关的内容。
+[前文](https://yoa1226.github.io/2024/08/05/g1-gc-important-component.html) 介绍了 G1 非常重要多的重要概念，并介绍了相关的源码，这些概念在后续文章中会重复提到。本文将开始介绍堆初始化相关的内容，介绍 G1 中重要的组件。
 
 ## 启动 JVM
 
@@ -95,7 +95,7 @@ jint Threads::create_vm(JavaVMInitArgs* args, bool* canTryAgain) {
   //omit
 }
 ```
-### Java 堆创建
+## Java 堆创建
 
 接着看找到 JVM 堆创建的代码
 
@@ -112,7 +112,7 @@ jint Universe::initialize_heap() {
 G1CollectedHeap::initialize();
 ```
 
-#### GC 类型选择
+### GC 类型选择
 
 未指定 GC 的情况下，JVM 通过 `select_gc_ergonomically` 函数在服务端模式默认选择 G1。
 
@@ -138,8 +138,259 @@ void GCConfig::select_gc_ergonomically() {
   } else { /*omit*/ }
 }
 ```
+
+### G1CollectedHeap()
+
+G1 堆属性很多，仅枚举比较重要的，具体逻辑使用时再次介绍。
+
+```cpp
+// Handle G1 NUMA support. 与 G1 numa 优化有关，
+//详情见 JEP 345: NUMA-Aware Memory Allocation for G1
+G1NUMA* _numa;
+
+// The sequence of all heap regions in the heap.
+//负责对堆中所有 region 进行管理
+G1HeapRegionManager _hrm;
+
+// Manages all allocations with regions except humongous object allocations.
+// 分配器
+G1Allocator* _allocator;
+
+// The young region list.
+G1EdenRegions _eden;
+G1SurvivorRegions _survivor;
+
+//cset
+G1CollectionSet _collection_set;
+
+//年轻代共享的记忆集
+G1CardSet _young_regions_cardset;
+//收集策略，回收过程保存一下统计信息，用于预测回收时间
+G1Policy* _policy;
+
+//GC 线程数
+ParallelGCThreads;
+//全局任务队列
+_task_queues
+```
+
 ## G1 堆初始化
 
-## G1 堆组件
+`initialize()` 初始化内部组件，此处只是枚举重要属性，特别重要的会在单独的小节介绍。
+
+```cpp
+//全局卡表，前面的文章介绍过
+G1CardTable* ct = new G1CardTable(heap_rs.region());
+
+//写屏障相关操作
+G1BarrierSet* bs = new G1BarrierSet(ct);
+
+//遍历记忆集相关
+_rem_set = new G1RemSet(this, _card_table);
+
+//GC 工作线程
+  _workers = new WorkerThreads("GC Thread", ParallelGCThreads);
+
+//并发标记相关
+_cm = new G1ConcurrentMark(this, bitmap_storage);
+_cm_thread = _cm->cm_thread();
+
+//GC 策略相关初始化
+policy()->init(this, &_collection_set);
+
+//处理 dirty card 线程
+jint ecode = initialize_concurrent_refinement();
+
+//负责协调异步任务的执行和后台的维护操作
+ecode = initialize_service_thread();
+_service_thread->register_task(_periodic_gc_task);
+_service_thread->register_task(_free_arena_memory_task);
+
+```
+## G1 组件
+
+### WorkerThreads
+
+处理特定任务的线程池，用于 GC 收集和 并发标记。`_workers` 是实际的线程数据。
+
+```cpp
+_workers = new WorkerThreads("GC Thread", ParallelGCThreads);
+_concurrent_workers = new WorkerThreads("G1 Conc", _max_concurrent_workers)
+
+uint WorkerThreads::set_active_workers(uint num_workers) {
+  while (_created_workers < num_workers) {
+    WorkerThread* const worker = create_worker(_created_workers);
+    _workers[_created_workers] = worker;
+  }
+}
+```
+
+`WorkerThread` 被创建完成以后就会执行 `run` 方法，阻塞在 `_start_semaphore` 等待被唤醒。
+
+当任务执行完成则换醒分发任务的线程。
+
+```cpp
+void WorkerThread::run() {
+  while (true) { _dispatcher->worker_run_task(); }
+}
+
+void WorkerTaskDispatcher::worker_run_task() {
+  _start_semaphore.wait();
+
+  // Get and set worker id.
+  const uint worker_id = Atomic::fetch_then_add(&_started, 1u);
+  WorkerThread::set_worker_id(worker_id);
+
+  // Run task.
+  GCIdMark gc_id_mark(_task->gc_id()); //保存到工作线程，这里就是在日志中看到第几轮 GC 编号。
+
+  _task->work(worker_id);
+  const uint not_finished = Atomic::sub(&_not_finished, 1u);
+  if (not_finished == 0) { _end_semaphore.signal(); /*任务执行完成，唤醒分发任务的线程。*/ }
+}
+```
+
+`coordinator_distribute_task` 负责任务的分发，唤醒任务线程，并且等待任务执行完成。
+
+```cpp
+void WorkerTaskDispatcher::coordinator_distribute_task(WorkerTask* task, uint num_workers) {
+  _task = task;
+  _not_finished = num_workers;
+
+  _start_semaphore.signal(num_workers);//唤醒任务线程
+  _end_semaphore.wait(); //等待任务执行完成
+
+  _task = nullptr;
+  _started = 0;
+}
+```
+### WorkerTask
+
+`WorkerTask` 是顶级接口，几乎所有的 GC 任务都继承它，并且重写 `work` 方法。
+
+后续阅读源码时，可直接查看 `work` 方法的逻辑。
+
+```cpp
+class WorkerTask : public CHeapObj<mtInternal> {
+  const char* _name;
+  const uint _gc_id;
+  virtual void work(uint worker_id) = 0;
+}
+```
+
+### G1ConcurrentMark
+
+此类主要用于发标记阶段。
+
+```cpp
+G1ConcurrentMarkThread* _cm_thread;     // The thread doing the work
+// Concurrent marking support structures
+G1CMBitMap   _mark_bitmap
+
+G1CMTask**              _tasks;       // Task queue array (max_worker_id length)
+G1CMTaskQueueSet*       _task_queues; // Task queue se
+
+WorkerThreads* _concurrent_workers;
+uint      _num_concurrent_workers; // The number of marking worker threads we're using
+uint      _max_concurrent_workers; // Maximum number of marking worker thread>
+```
+
+_mark_bitmap：存储标记结果的 bitMap，1 bit 标记 8 bytes，它额外消耗堆内存的 1.5%（1/64）。
+_concurrent_workers：工作线程前文已经说过。
+
+#### G1ConcurrentMarkThread
+
+负责并发标记的整体调度、任务分发
+
+```cpp
+//继承树 oncurrentGCThread<-NamedThread<-NonJavaThread<-Thread<-ThreadShadow
+class G1ConcurrentMarkThread: public ConcurrentGCThread {
+    void run_service();
+}
+class ConcurrentGCThread: public NamedThread {
+  void ConcurrentGCThread::run() { run_service(); }
+}
+
+void G1ConcurrentMarkThread::run_service() {
+
+  while (wait_for_next_cycle()) { //等待被唤醒
+    concurrent_cycle_start();
+    concurrent_mark_cycle_do();
+    concurrent_cycle_end(_state == FullMark && !_cm->has_aborted());
+  }
+}
+//唤醒代码
+//-> do_collection_pause_at_safepoint_helper()
+//->start_concurrent_cycle()->CGC_lock->notify()
+```
+
+### G1ConcurrentRefine
+
+此类负责处理异步处理 `dirty card`，`do_refinement_step` 是入口函数。`_dcqs` 是 `dirty card` 集合。
+
+```cpp
+class G1ConcurrentRefine : public CHeapObj<mtGC> {
+  G1DirtyCardQueueSet& _dcqs
+  G1ConcurrentRefineThreadControl _thread_control;
+}
+
+//->G1ConcurrentRefineThreadControl::initialize
+//->create_refinement_thread
+//-> G1ConcurrentRefineThread::create(_cr, worker_id);
+//crt = ->new (std::nothrow) G1PrimaryConcurrentRefineThread(cr);
+//-> crt->create_and_start()
+
+class G1ConcurrentRefineThread: public ConcurrentGCThread {
+  void G1ConcurrentRefineThread::run_service() {
+
+  while (wait_for_completed_buffers()) {
+    SuspendibleThreadSetJoiner sts_join;
+    G1ConcurrentRefineStats active_stats_start = _refinement_stats;
+    while (!should_terminate()) {
+      if (sts_join.should_yield()) {
+      } else if (maybe_deactivate()) { break; }
+       else {
+        do_refinement_step(); //处理 dirty card
+  } } } } }
+```
+
+### G1ServiceThread
+
+此类负责协调异步任务的执行和后台的维护操作
+
+```cpp
+class G1ServiceThread: public ConcurrentGCThread {
+    G1ServiceTaskQueue _task_queue;
+
+    void G1ServiceThread::run_service() {
+        while (true) {
+        G1ServiceTask* task = wait_for_task();
+        run_task(task);
+    } } }
+//注册任务
+_service_thread->register_task(_periodic_gc_task);
+_service_thread->register_task(_free_arena_memory_task);
+```
+
+## 总结
+
+本文从 JVM 启动到 G1 堆，介绍了 G1 重要组件，G1 重要非常多，余下的后续使用时再介绍。
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
 
-## 总结