JVM G1 源码分析（七）- Full GC

1. 简介

当晋升失败、疏散失败、大对象分配失败、Evac失败时，有可能触发Full GC，在JDK10之前，Full GC是串行的，JEP 307: Parallel Full GC for G1之后引入了并行Full GC。本文主要介绍并行Full GC机制。

2. 源码分析

Full GC的入口在g1CollectedHeap.cpp的G1CollectedHeap::do_full_collection

bool G1CollectedHeap::do_full_collection(bool explicit_gc,
                                         bool clear_all_soft_refs) {
  assert_at_safepoint_on_vm_thread();

  if (GCLocker::check_active_before_gc()) {
    // Full GC was not completed.
    return false;
  }

  const bool do_clear_all_soft_refs = clear_all_soft_refs ||
      soft_ref_policy()->should_clear_all_soft_refs();

  G1FullCollector collector(this, explicit_gc, do_clear_all_soft_refs);
  GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause(), true);

  collector.prepare_collection();
  collector.collect();
  collector.complete_collection();

  // Full collection was successfully completed.
  return true;
}

• 准备回收，prepare_collection
• 回收，collect
• 回收后处理，complete_collection

2.1 准备阶段

void G1FullCollector::prepare_collection() {
  _heap->g1_policy()->record_full_collection_start();

  _heap->print_heap_before_gc();
  _heap->print_heap_regions();

  _heap->abort_concurrent_cycle();
  _heap->verify_before_full_collection(scope()->is_explicit_gc());

  _heap->gc_prologue(true);
  _heap->prepare_heap_for_full_collection();

  reference_processor()->enable_discovery();
  reference_processor()->setup_policy(scope()->should_clear_soft_refs());

  // When collecting the permanent generation Method*s may be moving,
  // so we either have to flush all bcp data or convert it into bci.
  CodeCache::gc_prologue();

  // We should save the marks of the currently locked biased monitors.
  // The marking doesn't preserve the marks of biased objects.
  BiasedLocking::preserve_marks();

  // Clear and activate derived pointer collection.
  clear_and_activate_derived_pointers();
}

• Full GC应当清理软引用
• 由于Full GC过程中，永久代（元空间）中的方法可能被移动，需要保存bcp字节码指针数据或者转化为bci字节码索引
• 保存轻量级锁和重量级锁的对象头
• 清理和处理对象的派生关系

2.2 回收阶段

void G1FullCollector::collect() {
  phase1_mark_live_objects();
  verify_after_marking();

  // Don't add any more derived pointers during later phases
  deactivate_derived_pointers();

  phase2_prepare_compaction();

  phase3_adjust_pointers();

  phase4_do_compaction();
}

• phase1 并行标记对象
• phase2 并行准备压缩
• phase3 并行调整指针
• phase4 并行压缩

2.2.1 并行标记

从GC roots出发，递归标记所有的活跃对象。

void G1FullCollector::phase1_mark_live_objects() {
  // Recursively traverse all live objects and mark them.
  GCTraceTime(Info, gc, phases) info("Phase 1: Mark live objects", scope()->timer());

  // Do the actual marking.
  G1FullGCMarkTask marking_task(this);
  run_task(&marking_task);

  // Process references discovered during marking.
  G1FullGCReferenceProcessingExecutor reference_processing(this);
  reference_processing.execute(scope()->timer(), scope()->tracer());

  // Weak oops cleanup.
  {
    GCTraceTime(Debug, gc, phases) debug("Phase 1: Weak Processing", scope()->timer());
    WeakProcessor::weak_oops_do(_heap->workers(), &_is_alive, &do_nothing_cl, 1);
  }

  // Class unloading and cleanup.
  if (ClassUnloading) {
    GCTraceTime(Debug, gc, phases) debug("Phase 1: Class Unloading and Cleanup", scope()->timer());
    // Unload classes and purge the SystemDictionary.
    bool purged_class = SystemDictionary::do_unloading(scope()->timer());
    _heap->complete_cleaning(&_is_alive, purged_class);
  } else {
    GCTraceTime(Debug, gc, phases) debug("Phase 1: String and Symbol Tables Cleanup", scope()->timer());
    // If no class unloading just clean out strings.
    _heap->partial_cleaning(&_is_alive, true, G1StringDedup::is_enabled());
  }

  scope()->tracer()->report_object_count_after_gc(&_is_alive);
}

• 标记对象，具体逻辑在G1FullGCMarkTask中
• 清理弱引用
• 卸载类的元数据（complete_cleaning）或仅清理字符串（partial_cleaning）
• 清理字符串会清理StringTable和字符串去重（JEP 192: String Deduplication in G1）

G1FullGCMarkTask

void G1FullGCMarkTask::work(uint worker_id) {
  Ticks start = Ticks::now();
  ResourceMark rm;
  G1FullGCMarker* marker = collector()->marker(worker_id);
  MarkingCodeBlobClosure code_closure(marker->mark_closure(), !CodeBlobToOopClosure::FixRelocations);

  if (ClassUnloading) {
    _root_processor.process_strong_roots(
        marker->mark_closure(),
        marker->cld_closure(),
        &code_closure);
  } else {
    _root_processor.process_all_roots_no_string_table(
        marker->mark_closure(),
        marker->cld_closure(),
        &code_closure);
  }

  // Mark stack is populated, now process and drain it.
  marker->complete_marking(collector()->oop_queue_set(), collector()->array_queue_set(), _terminator.terminator());

  // This is the point where the entire marking should have completed.
  assert(marker->oop_stack()->is_empty(), "Marking should have completed");
  assert(marker->objarray_stack()->is_empty(), "Array marking should have completed");
  log_task("Marking task", worker_id, start);
}

void G1RootProcessor::process_strong_roots(OopClosure* oops,
                                           CLDClosure* clds,
                                           CodeBlobClosure* blobs) {
  StrongRootsClosures closures(oops, clds, blobs);

  process_java_roots(&closures, NULL, 0);
  process_vm_roots(&closures, NULL, 0);

  _process_strong_tasks.all_tasks_completed(n_workers());
}

void G1RootProcessor::process_all_roots(OopClosure* oops,
                                        CLDClosure* clds,
                                        CodeBlobClosure* blobs,
                                        bool process_string_table) {
  AllRootsClosures closures(oops, clds);

  process_java_roots(&closures, NULL, 0);
  process_vm_roots(&closures, NULL, 0);

  if (process_string_table) {
    process_string_table_roots(&closures, NULL, 0);
  }
  process_code_cache_roots(blobs, NULL, 0);

  _process_strong_tasks.all_tasks_completed(n_workers());
}

void G1RootProcessor::process_all_roots(OopClosure* oops,
                                        CLDClosure* clds,
                                        CodeBlobClosure* blobs) {
  process_all_roots(oops, clds, blobs, true);
}

void G1RootProcessor::process_all_roots_no_string_table(OopClosure* oops,
                                                        CLDClosure* clds,
                                                        CodeBlobClosure* blobs) {
  assert(!ClassUnloading, "Should only be used when class unloading is disabled");
  process_all_roots(oops, clds, blobs, false);
}

• 如果允许卸载类的元数据，则调用process_strong_roots；否则调用process_all_roots_no_string_table
• process_strong_roots的GC roots仅强根
• process_all_roots_no_string_table的GC roots包括弱根、强根，但是不含StringTable
• 遍历标记栈中的所有对象

2.2.2 准备压缩

计算每个活跃对象应该在什么位置，即计算对象压缩后的新位置指针并写入对象头。

void G1FullCollector::phase2_prepare_compaction() {
  GCTraceTime(Info, gc, phases) info("Phase 2: Prepare for compaction", scope()->timer());
  G1FullGCPrepareTask task(this);
  run_task(&task);

  // To avoid OOM when there is memory left.
  if (!task.has_freed_regions()) {
    task.prepare_serial_compaction();
  }
}

• 调用G1FullGCPrepareTask准备压缩
• 如果任务没有空闲Region，则调用prepare_serial_compaction串行合并所有线程的最后一个分区，以避免OOM

G1FullGCPrepareTask

void G1FullGCPrepareTask::work(uint worker_id) {
  Ticks start = Ticks::now();
  G1FullGCCompactionPoint* compaction_point = collector()->compaction_point(worker_id);
  G1CalculatePointersClosure closure(collector()->mark_bitmap(), compaction_point);
  G1CollectedHeap::heap()->heap_region_par_iterate_from_start(&closure, &_hrclaimer);

  // Update humongous region sets
  closure.update_sets();
  compaction_point->update();

  // Check if any regions was freed by this worker and store in task.
  if (closure.freed_regions()) {
    set_freed_regions();
  }
  log_task("Prepare compaction task", worker_id, start);
}

• 压缩对象具体逻辑在G1FullGCCompactionPoint中实现，执行完成后，对象头存储了对象的新地址
• 如果是大对象分区，且对象已经都死亡，则直接释放分区

2.2.3 调整指针

在上一步计算出所有活跃对象的新位置后，需要修改引用到新地址。

void G1FullCollector::phase3_adjust_pointers() {
  // Adjust the pointers to reflect the new locations
  GCTraceTime(Info, gc, phases) info("Phase 3: Adjust pointers", scope()->timer());

  G1FullGCAdjustTask task(this);
  run_task(&task);
}

void G1FullGCAdjustTask::work(uint worker_id) {
  Ticks start = Ticks::now();
  ResourceMark rm;

  // Adjust preserved marks first since they are not balanced.
  G1FullGCMarker* marker = collector()->marker(worker_id);
  marker->preserved_stack()->adjust_during_full_gc();

  // Adjust the weak roots.

  if (Atomic::add(1u, &_references_done) == 1u) { // First incr claims task.
    G1CollectedHeap::heap()->ref_processor_stw()->weak_oops_do(&_adjust);
  }

  AlwaysTrueClosure always_alive;
  _weak_proc_task.work(worker_id, &always_alive, &_adjust);

  CLDToOopClosure adjust_cld(&_adjust, ClassLoaderData::_claim_strong);
  CodeBlobToOopClosure adjust_code(&_adjust, CodeBlobToOopClosure::FixRelocations);
  _root_processor.process_all_roots(
      &_adjust,
      &adjust_cld,
      &adjust_code);

  // Adjust string dedup if enabled.
  if (G1StringDedup::is_enabled()) {
    G1StringDedup::parallel_unlink(&_adjust_string_dedup, worker_id);
  }

  // Now adjust pointers region by region
  G1AdjustRegionClosure blk(collector()->mark_bitmap(), worker_id);
  G1CollectedHeap::heap()->heap_region_par_iterate_from_worker_offset(&blk, &_hrclaimer, worker_id);
  log_task("Adjust task", worker_id, start);
}

• 调整之前保存的轻量级锁和重量级锁对象的引用地址
• 调整弱根
• 调整全部根对象
• 处理字符串去重逻辑
• 一个region一个region的调整引用地址

2.2.4 移动对象

对象的新地址和引用都已经更新，现在需要把对象移动到新位置

void G1FullCollector::phase4_do_compaction() {
  // Compact the heap using the compaction queues created in phase 2.
  GCTraceTime(Info, gc, phases) info("Phase 4: Compact heap", scope()->timer());
  G1FullGCCompactTask task(this);
  run_task(&task);

  // Serial compact to avoid OOM when very few free regions.
  if (serial_compaction_point()->has_regions()) {
    task.serial_compaction();
  }
}

• 具体压缩对象逻辑在G1FullGCCompactTask
• 如果phase2计算位置中使用了串行处理，则移动对象时也要使用串行处理移动每任务最后一个分区的对象

G1FullGCCompactTask

void G1FullGCCompactTask::work(uint worker_id) {
  Ticks start = Ticks::now();
  GrowableArray<HeapRegion*>* compaction_queue = collector()->compaction_point(worker_id)->regions();
  for (GrowableArrayIterator<HeapRegion*> it = compaction_queue->begin();
       it != compaction_queue->end();
       ++it) {
    compact_region(*it);
  }

  G1ResetHumongousClosure hc(collector()->mark_bitmap());
  G1CollectedHeap::heap()->heap_region_par_iterate_from_worker_offset(&hc, &_claimer, worker_id);
  log_task("Compaction task", worker_id, start);
}

size_t G1FullGCCompactTask::G1CompactRegionClosure::apply(oop obj) {
  size_t size = obj->size();
  HeapWord* destination = (HeapWord*)obj->forwardee();
  if (destination == NULL) {
    // Object not moving
    return size;
  }

  // copy object and reinit its mark
  HeapWord* obj_addr = (HeapWord*) obj;
  assert(obj_addr != destination, "everything in this pass should be moving");
  Copy::aligned_conjoint_words(obj_addr, destination, size);
  oop(destination)->init_mark_raw();
  assert(oop(destination)->klass() != NULL, "should have a class");

  return size;
}

• 迭代处理每个Region
• 调用闭包G1CompactRegionClosure的apply函数移动对象到Region头部
• 如果Region中的全部对象都已清理，则回收该Region

3. 引用

jdk12源代码[https://hg.openjdk.java.net/jdk/jdk12]
JEP-307[http://openjdk.java.net/jeps/307]
JEP-192[http://openjdk.java.net/jeps/192]