由dubbo服务禁用system.gc而引起的思考

我一直都有一个疑问，丰巢业务服务的生产环境jvm参数设置是禁止system.gc的，也就是开启设置：-XX:+DisableExplicitGC，但是生产环境却从来没有出现过堆外内存溢出的情况。说明一下，丰巢使用了阿里开源的dubbo，而dubbo底层通信默认情况下使用了3.2.5.Final版本的netty，而我们对于netty的常规认知里，netty一定是使用了堆外内存，并且堆外内存在禁止了system.gc这个函数调用的话，在服务没有主动回收分配的堆外内存的情况下，一定会出现堆外内存的泄露。带着这个问题，刚好前天晚上有些时间，研究了一下3.2.5版本的netty源码，又是在科兴科兴园等馒头妈妈时候，发现了秘密之所在，我只能说，科兴科学园真是我的宝地啊。

 

涉及到的netty类：NioWorker、HeapChannelBufferFactory、BigEndianHeapChannelBuffer、SocketReceiveBufferPool

 

核心的秘密在SocketReceiveBufferPool中

 1 final class SocketReceiveBufferPool {
 2
 3     private static final int POOL_SIZE = 8;
 4
 5     @SuppressWarnings("unchecked")
 6     private final SoftReference<ByteBuffer>[] pool = new SoftReference[POOL_SIZE];
 7
 8     SocketReceiveBufferPool() {
 9         super();
10     }
11
12     final ByteBuffer acquire(int size) {
13         final SoftReference<ByteBuffer>[] pool = this.pool;
14         for (int i = 0; i < POOL_SIZE; i ++) {
15             SoftReference<ByteBuffer> ref = pool[i];
16             if (ref == null) {
17                 continue;
18             }
19
20             ByteBuffer buf = ref.get();
21             if (buf == null) {
22                 pool[i] = null;
23                 continue;
24             }
25
26             if (buf.capacity() < size) {
27                 continue;
28             }
29
30             pool[i] = null;
31
32             buf.clear();
33             return buf;
34         }
35
36         ByteBuffer buf = ByteBuffer.allocateDirect(normalizeCapacity(size));
37         buf.clear();
38         return buf;
39     }
40
41     final void release(ByteBuffer buffer) {
42         final SoftReference<ByteBuffer>[] pool = this.pool;
43         for (int i = 0; i < POOL_SIZE; i ++) {
44             SoftReference<ByteBuffer> ref = pool[i];
45             if (ref == null || ref.get() == null) {
46                 pool[i] = new SoftReference<ByteBuffer>(buffer);
47                 return;
48             }
49         }
50
51         // pool is full - replace one
52         final int capacity = buffer.capacity();
53         for (int i = 0; i< POOL_SIZE; i ++) {
54             SoftReference<ByteBuffer> ref = pool[i];
55             ByteBuffer pooled = ref.get();
56             if (pooled == null) {
57                 pool[i] = null;
58                 continue;
59             }
60
61             if (pooled.capacity() < capacity) {
62                 pool[i] = new SoftReference<ByteBuffer>(buffer);
63                 return;
64             }
65         }
66     }
67
68     private static final int normalizeCapacity(int capacity) {
69         // Normalize to multiple of 1024
70         int q = capacity >>> 10;
71         int r = capacity & 1023;
72         if (r != 0) {
73             q ++;
74         }
75         return q << 10;
76     }
77 }

SocketReceiveBufferPool中维护了一个SoftReference<ByteBuffer>类型的数组，关于java的SoftReference，大家可以自行搜索。其实就是在此类中维护了一个directbuffer的内存池，此部分的内存是可以重复利用的。那么问题来了，如果我们把netty用于接收网络信息的directbuffer直接传给dubbo的业务代码，那么这个内存池的作用是什么呢，内存如何被release回内存池？带着这个疑问，继续分析调用了SocketReceiveBufferPool的NioWorker代码。

 1     private boolean read(SelectionKey k) {
 2         final SocketChannel ch = (SocketChannel) k.channel();
 3         final NioSocketChannel channel = (NioSocketChannel) k.attachment();
 4
 5         final ReceiveBufferSizePredictor predictor =
 6             channel.getConfig().getReceiveBufferSizePredictor();
 7         final int predictedRecvBufSize = predictor.nextReceiveBufferSize();
 8
 9         int ret = 0;
10         int readBytes = 0;
11         boolean failure = true;
12
13         ByteBuffer bb = recvBufferPool.acquire(predictedRecvBufSize);
14         15         try {
16             while ((ret = ch.read(bb)) > 0) {
17                 readBytes += ret;
18                 if (!bb.hasRemaining()) {
19                     break;
20                 }
21             }
22             failure = false;
23         } catch (ClosedChannelException e) {
24             // Can happen, and does not need a user attention.
25         } catch (Throwable t) {
26             fireExceptionCaught(channel, t);
27         }
28
29         if (readBytes > 0) {
30             bb.flip();
31
32             final ChannelBufferFactory bufferFactory =
33                 channel.getConfig().getBufferFactory();
34             final ChannelBuffer buffer = bufferFactory.getBuffer(readBytes);
35             buffer.setBytes(0, bb);
36             buffer.writerIndex(readBytes);
37             //if(buffer instanceof BigEndianHeapChannelBuffer){
38             //    logger2.info("buffer instanceof BigEndianHeapChannelBuffer.");
39             //}
40             recvBufferPool.release(bb);
41
42             // Update the predi||\\|||||
43             predictor.previousReceiveBufferSize(readBytes);
44
45             // Fire the event.
46             fireMessageReceived(channel, buffer);
47         } else {
48             recvBufferPool.release(bb);
49         }
50
51         if (ret < 0 || failure) {
52             k.cancel(); // Some JDK implementations run into an infinite loop without this.
53             close(channel, succeededFuture(channel));
54             return false;
55         }
56
57         return true;
58     }

在代码里发现了netty会再创造一个chanelbuffer对象，然后将directbuffer里的内容复制到chanelbuffer里面，而这个chanelbuffer对象实际上是一个堆内内存，然后netty会真对这块内存进行解码及返回给上层调用服务等，也就是说没有直接将directbuffer返回给dubbo服务，这样也就解释了，我们在提供dubbo服务的jvm里，禁止掉了system.gc的情况下，没有发生过堆外内存泄漏的原因。后面我会找时间详细的分析一下netty4和kafka使用directbuffer的情况。