JVM G1源码分析——慢速分配

当不能进行快速分配，就进入到慢速分配。实际上在TLAB中也有可能进入到慢速分配，就是我们前面提到的attempt_allocation，前面已经解释过。这里的慢速分配是指在TLAB中经过努力分配还不能成功，再次进入慢速分配，我们来看一下这个更慢的慢速分配：·attempt_allocation尝试进行对象分配，如果成功则返回。值得注意的是在attempt_allocation里面可能会进行垃圾回收，这里的垃圾回收是指增量的垃圾回收，主要是新生代或者混合收集。

• 如果大对象在attempt_allocation_humongous，直接分配的老生代。
• 如果分配不成功，则进行GC垃圾回收，注意这里的回收主要是Full GC，然后再分配。因为这里是分配的最后一步，所以进行几次不同的垃圾回收和尝试。主要代码在satisfy_failed_allocation中。
• 最终成功分配或者失败达到一定次数，则分配失败。

慢速分配代码如下所示：

HeapWord* G1CollectedHeap::mem_allocate(size_t word_size,
                              bool*  gc_overhead_limit_was_exceeded) {
  assert_heap_not_locked_and_not_at_safepoint();

  // 先去分配内存，如果分配成功返回，否则去GC，循环执行该过程
  // Loop until the allocation is satisfied, or unsatisfied after GC.
  for (uint try_count = 1, gclocker_retry_count = 0; /* we'll return */; try_count += 1) 
  {
    uint gc_count_before;

    HeapWord* result = NULL;
    if (!isHumongous(word_size)) {
      result = attempt_allocation(word_size, &gc_count_before, &gclocker_retry_count);
    } else {
      // 分配大对象内存
      result = attempt_allocation_humongous(word_size, &gc_count_before, &gclocker_retry_count);
    }
    if (result != NULL) {
      return result;
    }

    // Create the garbage collection operation...
    VM_G1CollectForAllocation op(gc_count_before, word_size);
    op.set_allocation_context(AllocationContext::current());

    // ...and get the VM thread to execute it.
    VMThread::execute(&op);

    if (op.prologue_succeeded() && op.pause_succeeded()) {
      // If the operation was successful we'll return the result even
      // if it is NULL. If the allocation attempt failed immediately
      // after a Full GC, it's unlikely we'll be able to allocate now.
      HeapWord* result = op.result();
      if (result != NULL && !isHumongous(word_size)) {
        // Allocations that take place on VM operations do not do any
        // card dirtying and we have to do it here. We only have to do
        // this for non-humongous allocations, though.
        dirty_young_block(result, word_size);
      }
      return result;
    } else {
      if (gclocker_retry_count > GCLockerRetryAllocationCount) {
        return NULL;
      }
      assert(op.result() == NULL,
             "the result should be NULL if the VM op did not succeed");
    }

    // Give a warning if we seem to be looping forever.
    if ((QueuedAllocationWarningCount > 0) &&
        (try_count % QueuedAllocationWarningCount == 0)) {
      warning("G1CollectedHeap::mem_allocate retries %d times", try_count);
    }
  }

  ShouldNotReachHere();
  return NULL;
}

大对象分配和TLAB中的慢速分配基本类似。唯一的区别就是对象大小不同。步骤主要：

• 尝试垃圾回收，这里主要是增量回收，同时启动并发标记。
• 尝试开始分配对象，对于大对象分为两类，一类是大于HeapRegionSize的一半，但是小于HeapRegionSize，即一个完整的堆分区可以保存，则直接从空闲列表直接拿一个堆分区，或者分配一个新的堆分区。如果是连续对象，则需要多个堆分区，思路同上，但是处理的时候需要加锁。
• 如果失败再次尝试垃圾回收，之后再分配。
• 最终成功分配或者失败达到一定次数，则分配失败。

大对象分配代码如下所示：

HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size,
                                                        uint* gc_count_before_ret,
                                                        uint* gclocker_retry_count_ret) {
  // The structure of this method has a lot of similarities to
  // attempt_allocation_slow(). The reason these two were not merged
  // into a single one is that such a method would require several "if
  // allocation is not humongous do this, otherwise do that"
  // conditional paths which would obscure its flow. In fact, an early
  // version of this code did use a unified method which was harder to
  // follow and, as a result, it had subtle bugs that were hard to
  // track down. So keeping these two methods separate allows each to
  // be more readable. It will be good to keep these two in sync as
  // much as possible.

  assert_heap_not_locked_and_not_at_safepoint();
  assert(isHumongous(word_size), "attempt_allocation_humongous() "
         "should only be called for humongous allocations");

  // Humongous objects can exhaust the heap quickly, so we should check if we
  // need to start a marking cycle at each humongous object allocation. We do
  // the check before we do the actual allocation. The reason for doing it
  // before the allocation is that we avoid having to keep track of the newly
  // allocated memory while we do a GC.
  // 开始尝试垃圾回收
  if (g1_policy()->need_to_start_conc_mark("concurrent humongous allocation",
                                           word_size)) {
    collect(GCCause::_g1_humongous_allocation);
  }

  // We will loop until a) we manage to successfully perform the
  // allocation or b) we successfully schedule a collection which
  // fails to perform the allocation. b) is the only case when we'll
  // return NULL.
  HeapWord* result = NULL;
  for (int try_count = 1; /* we'll return */; try_count += 1) {
    bool should_try_gc;
    uint gc_count_before;

    {
      MutexLockerEx x(Heap_lock);

      // Given that humongous objects are not allocated in young
      // regions, we'll first try to do the allocation without doing a
      // collection hoping that there's enough space in the heap.
      result = humongous_obj_allocate(word_size, AllocationContext::current());
      if (result != NULL) {
        return result;
      }

      if (GC_locker::is_active_and_needs_gc()) {
        should_try_gc = false;
      } else {
         // The GCLocker may not be active but the GCLocker initiated
        // GC may not yet have been performed (GCLocker::needs_gc()
        // returns true). In this case we do not try this GC and
        // wait until the GCLocker initiated GC is performed, and
        // then retry the allocation.
        if (GC_locker::needs_gc()) {
          should_try_gc = false;
        } else {
          // Read the GC count while still holding the Heap_lock.
          gc_count_before = total_collections();
          should_try_gc = true;
        }
      }
    }

    if (should_try_gc) {
      // If we failed to allocate the humongous object, we should try to
      // do a collection pause (if we're allowed) in case it reclaims
      // enough space for the allocation to succeed after the pause.

      bool succeeded;
      result = do_collection_pause(word_size, gc_count_before, &succeeded,
                                   GCCause::_g1_humongous_allocation);
      if (result != NULL) {
        assert(succeeded, "only way to get back a non-NULL result");
        return result;
      }

      if (succeeded) {
        // If we get here we successfully scheduled a collection which
        // failed to allocate. No point in trying to allocate
        // further. We'll just return NULL.
        MutexLockerEx x(Heap_lock);
        *gc_count_before_ret = total_collections();
        return NULL;
      }
    } else {
      if (*gclocker_retry_count_ret > GCLockerRetryAllocationCount) {
        MutexLockerEx x(Heap_lock);
        *gc_count_before_ret = total_collections();
        return NULL;
      }
      // The GCLocker is either active or the GCLocker initiated
      // GC has not yet been performed. Stall until it is and
      // then retry the allocation.
      GC_locker::stall_until_clear();
      (*gclocker_retry_count_ret) += 1;
    }

    // We can reach here if we were unsuccessful in scheduling a
    // collection (because another thread beat us to it) or if we were
    // stalled due to the GC locker. In either can we should retry the
    // allocation attempt in case another thread successfully
    // performed a collection and reclaimed enough space.  Give a
    // warning if we seem to be looping forever.

    if ((QueuedAllocationWarningCount > 0) &&
        (try_count % QueuedAllocationWarningCount == 0)) {
      warning("G1CollectedHeap::attempt_allocation_humongous() "
              "retries %d times", try_count);
    }
  }

  ShouldNotReachHere();
  return NULL;
}

void G1CollectedHeap::collect(GCCause::Cause cause) {
  assert_heap_not_locked();

  uint gc_count_before;
  uint old_marking_count_before;
  uint full_gc_count_before;
  bool retry_gc;

  do {
    retry_gc = false;

    {
      MutexLocker ml(Heap_lock);

      // Read the GC count while holding the Heap_lock
      gc_count_before = total_collections();
      full_gc_count_before = total_full_collections();
      old_marking_count_before = _old_marking_cycles_started;
    }

    if (should_do_concurrent_full_gc(cause)) {
      // Schedule an initial-mark evacuation pause that will start a
      // concurrent cycle. We're setting word_size to 0 which means that
      // we are not requesting a post-GC allocation.
      VM_G1IncCollectionPause op(gc_count_before,
                                 0,     /* word_size */
                                 true,  /* should_initiate_conc_mark */
                                 g1_policy()->max_pause_time_ms(),
                                 cause);
      op.set_allocation_context(AllocationContext::current());

      VMThread::execute(&op);
      if (!op.pause_succeeded()) {
        if (old_marking_count_before == _old_marking_cycles_started) {
          retry_gc = op.should_retry_gc();
        } else {
          // A Full GC happened while we were trying to schedule the
          // initial-mark GC. No point in starting a new cycle given
          // that the whole heap was collected anyway.
        }

        if (retry_gc) {
          if (GC_locker::is_active_and_needs_gc()) {
            GC_locker::stall_until_clear();
          }
        }
      }
    } else if (GC_locker::should_discard(cause, gc_count_before)) {
      // Return to be consistent with VMOp failure due to another
      // collection slipping in after our gc_count but before our
      // request is processed.  _gc_locker collections upgraded by
      // GCLockerInvokesConcurrent are handled above and never discarded.
      return;
    } else {
      if (cause == GCCause::_gc_locker || cause == GCCause::_wb_young_gc
          DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) {

        // Schedule a standard evacuation pause. We're setting word_size
        // to 0 which means that we are not requesting a post-GC allocation.
        VM_G1IncCollectionPause op(gc_count_before,
                                   0,     /* word_size */
                                   false, /* should_initiate_conc_mark */
                                   g1_policy()->max_pause_time_ms(),
                                   cause);
        VMThread::execute(&op);
      } else {
        // Schedule a Full GC.
        VM_G1CollectFull op(gc_count_before, full_gc_count_before, cause);
        VMThread::execute(&op);
      }
    }
  } while (retry_gc);
}

// If could fit into free regions w/o expansion, try.
// Otherwise, if can expand, do so.
// Otherwise, if using ex regions might help, try with ex given back.
HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size, AllocationContext_t context) {
  assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);

  verify_region_sets_optional();

  uint first = G1_NO_HRM_INDEX;
  uint obj_regions = (uint)(align_size_up_(word_size, HeapRegion::GrainWords) / HeapRegion::GrainWords);

  if (obj_regions == 1) {
    // Only one region to allocate, try to use a fast path by directly allocating
    // from the free lists. Do not try to expand here, we will potentially do that
    // later.
    HeapRegion* hr = new_region(word_size, true /* is_old */, false /* do_expand */);
    if (hr != NULL) {
      first = hr->hrm_index();
    }
  } else {
    // We can't allocate humongous regions spanning more than one region while
    // cleanupComplete() is running, since some of the regions we find to be
    // empty might not yet be added to the free list. It is not straightforward
    // to know in which list they are on so that we can remove them. We only
    // need to do this if we need to allocate more than one region to satisfy the
    // current humongous allocation request. If we are only allocating one region
    // we use the one-region region allocation code (see above), that already
    // potentially waits for regions from the secondary free list.
    wait_while_free_regions_coming();
    append_secondary_free_list_if_not_empty_with_lock();

    // Policy: Try only empty regions (i.e. already committed first). Maybe we
    // are lucky enough to find some.
    first = _hrm.find_contiguous_only_empty(obj_regions);
    if (first != G1_NO_HRM_INDEX) {
      _hrm.allocate_free_regions_starting_at(first, obj_regions);
    }
  }

  if (first == G1_NO_HRM_INDEX) {
    // Policy: We could not find enough regions for the humongous object in the
    // free list. Look through the heap to find a mix of free and uncommitted regions.
    // If so, try expansion.
    first = _hrm.find_contiguous_empty_or_unavailable(obj_regions);
    if (first != G1_NO_HRM_INDEX) {
      // We found something. Make sure these regions are committed, i.e. expand
      // the heap. Alternatively we could do a defragmentation GC.
      ergo_verbose1(ErgoHeapSizing,
                    "attempt heap expansion",
                    ergo_format_reason("humongous allocation request failed")
                    ergo_format_byte("allocation request"),
                    word_size * HeapWordSize);

      _hrm.expand_at(first, obj_regions);
      g1_policy()->record_new_heap_size(num_regions());

#ifdef ASSERT
      for (uint i = first; i < first + obj_regions; ++i) {
        HeapRegion* hr = region_at(i);
        assert(hr->is_free(), "sanity");
        assert(hr->is_empty(), "sanity");
        assert(is_on_master_free_list(hr), "sanity");
      }
#endif
      _hrm.allocate_free_regions_starting_at(first, obj_regions);
    } else {
      // Policy: Potentially trigger a defragmentation GC.
    }
  }

  HeapWord* result = NULL;
  if (first != G1_NO_HRM_INDEX) {
    result = humongous_obj_allocate_initialize_regions(first, obj_regions,
                                                       word_size, context);
    assert(result != NULL, "it should always return a valid result");

    // A successful humongous object allocation changes the used space
    // information of the old generation so we need to recalculate the
    // sizes and update the jstat counters here.
    g1mm()->update_sizes();
  }

  verify_region_sets_optional();

  return result;
}