G1的RSET

　　本文的一些图片和源码 来自于 https://www.jianshu.com/p/870abddaba41 

　　　　　　　　　　　　　　　　https://blog.csdn.net/baiye_xing/article/details/73743395

　　前一篇介绍了SATB，SATB属于pre write barrier。而今天要介绍的remembered set是一种post write barrier.

　　RSet

　　　每个Region初始化时，会初始化一个remembered set（已记忆集合），这个翻译有点拗口，以下简称RSet，该集合用来记录并跟踪其它Region指向该Region中对象的引用，每个Region默认按照512b划分成多个Card，所以RSet需要记录的东西应该是 xx Region的 xx Card。

　　RSet记录了其他Region中的对象引用本Region中对象的关系，属于points-into结构（谁引用了我的对象）。而Card Table则是一种points-out（我引用了谁的对象）的结构，每个Card 覆盖一定范围的Heap（一般为512Bytes）。G1的RSet是在Card Table的基础上实现的：每个Region会记录下别的Region有指向自己的指针，并标记这些指针分别在哪些Card的范围内。 这个RSet其实是一个Hash Table，Key是别的Region的起始地址，Value是一个集合，里面的元素是Card Table的Index。

　　下图表示了RSet、Card和Region的关系

　　

　　上图中有三个Region，每个Region被分成了多个Card，在不同Region中的Card会相互引用，Region1中的Card中的对象引用了Region2中的Card中的对象，蓝色实线表示的就是points-out的关系，而在Region2的RSet中，记录了Region1的Card，即红色虚线表示的关系，这就是points-into。

　　RSet实现过程

　　为了维护这些RSet，如果每次给引用类型的字段赋值都要更新RSet，这带来的额外开销实在太大，G1中采用post-write barrier和concurrent refinement threads实现了RSet的更新。

//假设对象young和old分别在不同的Region中
Object young = new Object();
old.p = young;

　　只有old引用young才需要记录rset，所以例子才会这么举。

　　在赋值动作的前后，JVM插入一个pre-write barrier和post-write barrier，其中post-write barrier的最终动作如下：

　　

　　

• 找到该字段所在的位置(Card)，并设置为dirty_card
• 如果当前是应用线程，每个Java线程有一个dirty card queue，把该card插入队列
• 除了每个线程自带的dirty card queue，还有一个全局共享的queue　　

　　

赋值动作到此结束，接下来的RSet更新操作交由多个ConcurrentG1RefineThread并发完成，每当全局队列集合超过一定阈值后，ConcurrentG1RefineThread会取出若干个队列，遍历每个队列中记录的card，并进行处理，位于G1RemSet::refine_card方法，大概实现逻辑如下：
1、根据card的地址，计算出card所在的Region
2、如果Region不存在，或者Region是Young区，或者该Region在回收集合中，则不进行处理
3、最终使用闭合函数G1UpdateRSOrPushRefOopClosure::do_oop_nv()的处理该card中的对象

inline void G1UpdateRSOrPushRefOopClosure::do_oop_nv(T* p) {
  oop obj = oopDesc::load_decode_heap_oop(p);
  if (obj == NULL) {
    return;
  }
#ifdef ASSERT
  // can't do because of races
  // assert(obj == NULL || obj->is_oop(), "expected an oop");

  // Do the safe subset of is_oop
#ifdef CHECK_UNHANDLED_OOPS
  oopDesc* o = obj.obj();
#else
  oopDesc* o = obj;
#endif // CHECK_UNHANDLED_OOPS
  assert((intptr_t)o % MinObjAlignmentInBytes == 0, "not oop aligned");
  assert(Universe::heap()->is_in_reserved(obj), "must be in heap");
#endif // ASSERT

  assert(_from != NULL, "from region must be non-NULL");
  assert(_from->is_in_reserved(p), "p is not in from");

  HeapRegion* to = _g1->heap_region_containing(obj);
  if (_from == to) {
    // Normally this closure should only be called with cross-region references.
    // But since Java threads are manipulating the references concurrently and we
    // reload the values things may have changed.
    return;
  }
  // The _record_refs_into_cset flag is true during the RSet
  // updating part of an evacuation pause. It is false at all
  // other times:
  //  * rebuilding the remembered sets after a full GC
  //  * during concurrent refinement.
  //  * updating the remembered sets of regions in the collection
  //    set in the event of an evacuation failure (when deferred
  //    updates are enabled).

  if (_record_refs_into_cset && to->in_collection_set()) {
    // We are recording references that point into the collection
    // set and this particular reference does exactly that...
    // If the referenced object has already been forwarded
    // to itself, we are handling an evacuation failure and
    // we have already visited/tried to copy this object
    // there is no need to retry.
    if (!self_forwarded(obj)) {
      assert(_push_ref_cl != NULL, "should not be null");
      // Push the reference in the refs queue of the G1ParScanThreadState
      // instance for this worker thread.
      _push_ref_cl->do_oop(p);
     }

    // Deferred updates to the CSet are either discarded (in the normal case),
    // or processed (if an evacuation failure occurs) at the end
    // of the collection.
    // See G1RemSet::cleanup_after_oops_into_collection_set_do().
  } else {
    // We either don't care about pushing references that point into the
    // collection set (i.e. we're not during an evacuation pause) _or_
    // the reference doesn't point into the collection set. Either way
    // we add the reference directly to the RSet of the region containing
    // the referenced object.
    assert(to->rem_set() != NULL, "Need per-region 'into' remsets.");
    to->rem_set()->add_reference(p, _worker_i);
  }
}　　

　　其中_from是持有引用的对象所在的Region，to是引用对象所在的Region，通过add_reference方法加入到RSet中，更细节的实现在OtherRegionsTable::add_reference方法中

　　

void OtherRegionsTable::add_reference(OopOrNarrowOopStar from, int tid) {
  uint cur_hrm_ind = hr()->hrm_index();

  if (G1TraceHeapRegionRememberedSet) {
    gclog_or_tty->print_cr("ORT::add_reference_work(" PTR_FORMAT "->" PTR_FORMAT ").",
                                                    from,
                                                    UseCompressedOops
                                                    ? (void *)oopDesc::load_decode_heap_oop((narrowOop*)from)
                                                    : (void *)oopDesc::load_decode_heap_oop((oop*)from));
  }

  int from_card = (int)(uintptr_t(from) >> CardTableModRefBS::card_shift);

  if (G1TraceHeapRegionRememberedSet) {
    gclog_or_tty->print_cr("Table for [" PTR_FORMAT "...): card %d (cache = "INT32_FORMAT")",
                  hr()->bottom(), from_card,
                  FromCardCache::at((uint)tid, cur_hrm_ind));
  }

  if (FromCardCache::contains_or_replace((uint)tid, cur_hrm_ind, from_card)) {
    if (G1TraceHeapRegionRememberedSet) {
      gclog_or_tty->print_cr("  from-card cache hit.");
    }
    assert(contains_reference(from), "We just added it!");
    return;
  }

  // Note that this may be a continued H region.
  HeapRegion* from_hr = _g1h->heap_region_containing_raw(from);
  RegionIdx_t from_hrm_ind = (RegionIdx_t) from_hr->hrm_index();

  // If the region is already coarsened, return.
  if (_coarse_map.at(from_hrm_ind)) {
    if (G1TraceHeapRegionRememberedSet) {
      gclog_or_tty->print_cr("  coarse map hit.");
    }
    assert(contains_reference(from), "We just added it!");
    return;
  }

  // Otherwise find a per-region table to add it to.
  size_t ind = from_hrm_ind & _mod_max_fine_entries_mask;
  PerRegionTable* prt = find_region_table(ind, from_hr);
  if (prt == NULL) {
    MutexLockerEx x(_m, Mutex::_no_safepoint_check_flag);
    // Confirm that it's really not there...
    prt = find_region_table(ind, from_hr);
    if (prt == NULL) {

      uintptr_t from_hr_bot_card_index =
        uintptr_t(from_hr->bottom())
          >> CardTableModRefBS::card_shift;
      CardIdx_t card_index = from_card - from_hr_bot_card_index;
      assert(0 <= card_index && (size_t)card_index < HeapRegion::CardsPerRegion,
             "Must be in range.");
      if (G1HRRSUseSparseTable &&
          _sparse_table.add_card(from_hrm_ind, card_index)) {
        if (G1RecordHRRSOops) {
          HeapRegionRemSet::record(hr(), from);
          if (G1TraceHeapRegionRememberedSet) {
            gclog_or_tty->print("   Added card " PTR_FORMAT " to region "
                                "[" PTR_FORMAT "...) for ref " PTR_FORMAT ".\n",
                                align_size_down(uintptr_t(from),
                                                CardTableModRefBS::card_size),
                                hr()->bottom(), from);
          }
        }
        if (G1TraceHeapRegionRememberedSet) {
          gclog_or_tty->print_cr("   added card to sparse table.");
        }
        assert(contains_reference_locked(from), "We just added it!");
        return;
      } else {
        if (G1TraceHeapRegionRememberedSet) {
          gclog_or_tty->print_cr("   [tid %d] sparse table entry "
                        "overflow(f: %d, t: %u)",
                        tid, from_hrm_ind, cur_hrm_ind);
        }
      }

      if (_n_fine_entries == _max_fine_entries) {
        prt = delete_region_table();
        // There is no need to clear the links to the 'all' list here:
        // prt will be reused immediately, i.e. remain in the 'all' list.
        prt->init(from_hr, false /* clear_links_to_all_list */);
      } else {
        prt = PerRegionTable::alloc(from_hr);
        link_to_all(prt);
      }

      PerRegionTable* first_prt = _fine_grain_regions[ind];
      prt->set_collision_list_next(first_prt);
      _fine_grain_regions[ind] = prt;
      _n_fine_entries++;

      if (G1HRRSUseSparseTable) {
        // Transfer from sparse to fine-grain.
        SparsePRTEntry *sprt_entry = _sparse_table.get_entry(from_hrm_ind);
        assert(sprt_entry != NULL, "There should have been an entry");
        for (int i = 0; i < SparsePRTEntry::cards_num(); i++) {
          CardIdx_t c = sprt_entry->card(i);
          if (c != SparsePRTEntry::NullEntry) {
            prt->add_card(c);
          }
        }
        // Now we can delete the sparse entry.
        bool res = _sparse_table.delete_entry(from_hrm_ind);
        assert(res, "It should have been there.");
      }
    }
    assert(prt != NULL && prt->hr() == from_hr, "consequence");
  }
  // Note that we can't assert "prt->hr() == from_hr", because of the
  // possibility of concurrent reuse.  But see head comment of
  // OtherRegionsTable for why this is OK.
  assert(prt != NULL, "Inv");

  prt->add_reference(from);

  if (G1RecordHRRSOops) {
    HeapRegionRemSet::record(hr(), from);
    if (G1TraceHeapRegionRememberedSet) {
      gclog_or_tty->print("Added card " PTR_FORMAT " to region "
                          "[" PTR_FORMAT "...) for ref " PTR_FORMAT ".\n",
                          align_size_down(uintptr_t(from),
                                          CardTableModRefBS::card_size),
                          hr()->bottom(), from);
    }
  }
  assert(contains_reference(from), "We just added it!");
}

　　代码真的多，不过我们可以看重点看红色部分。 int from_card 拿到card的index。_sparse_table 保存的是region的index作为key，card的index作为value

  HeapRegion* from_hr = _g1h->heap_region_containing_raw(from);
  RegionIdx_t from_hrm_ind = (RegionIdx_t) from_hr->hrm_index();

　　还有 就是 record方法

void HeapRegionRemSet::record(HeapRegion* hr, OopOrNarrowOopStar f) {
  if (_recorded_oops == NULL) {
    assert(_n_recorded == 0
           && _recorded_cards == NULL
           && _recorded_regions == NULL,
           "Inv");
    _recorded_oops    = NEW_C_HEAP_ARRAY(OopOrNarrowOopStar, MaxRecorded, mtGC);
    _recorded_cards   = NEW_C_HEAP_ARRAY(HeapWord*,          MaxRecorded, mtGC);
    _recorded_regions = NEW_C_HEAP_ARRAY(HeapRegion*,        MaxRecorded, mtGC);
  }
  if (_n_recorded == MaxRecorded) {
    gclog_or_tty->print_cr("Filled up 'recorded' (%d).", MaxRecorded);
  } else {
    _recorded_cards[_n_recorded] =
      (HeapWord*)align_size_down(uintptr_t(f),
                                 CardTableModRefBS::card_size);
    _recorded_oops[_n_recorded] = f;
    _recorded_regions[_n_recorded] = hr;
    _n_recorded++;
  }
}

　　看看不看不知道，一看吓一跳啊。之前看过太多的文章说RSET是一个map结构，key是对方的region指针，value是card的index。这么说其实不能算错，但是你看源码实际是通过两个数组实现的。

　　上面的话我说错了，的确也是有map结构的，_sparse_table.add_card(from_hrm_ind, card_index))。不过为啥还要在record再保存一遍。　　

　　现在我们是彻底知道RSET的存储结构了，这样对RSET的理解就会更深了。

RSet有什么好处？

进行垃圾回收时，如果Region1有根对象A引用了Region2的对象B，显然对象B是活的，如果没有Rset，就需要扫描整个Region1或者其它Region，才能确定对象B是活跃的，有了Rset可以避免对整个堆进行扫描。RSet究竟是怎么辅助GC的呢？在做YGC的时候，只需要选定young generation region的RSet作为根集，这些RSet记录了old->young的跨代引用，避免了扫描整个old generation。 而mixed gc的时候，old generation中记录了old->old的RSet，young->old的引用由扫描全部young generation region得到，这样也不用扫描全部old generation region。所以RSet的引入大大减少了GC的工作量。

RSet有什么风险？

通过对RSet实现过程的研究，我们得知应用线程只负责把更新字段所在的Card插入到dirty card queue中，然后由后台线程refinement threads负责RSet的更新操作，如果应用线程插入速度过快，refinement threads来不及处理，那么应用线程将接管RSet更新的任务，这是必须要避免的。

refinement threads线程数量可以通过-XX:G1ConcRefinementThreads或-XX:ParallelGCThreads参数设置

　　不知道大家注意到没有，无论是SATB还是RSET都有一个缓存结构，SATB会缓存旧的引用，RSET缓存指向新引用的card的地址。而且还有一个共同的特点，就是这个缓存结构都是在当前的线程里的，

　　之前我一直理解是这两个queue是在region里的，其实是当前的应用线程里的。