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

final class SocketReceiveBufferPool {

    private static final int POOL_SIZE = 8;

    @SuppressWarnings("unchecked")
    private final SoftReference<ByteBuffer>[] pool = new SoftReference[POOL_SIZE];

    SocketReceiveBufferPool() {
        super();
    }

    final ByteBuffer acquire(int size) {
        final SoftReference<ByteBuffer>[] pool = this.pool;
        for (int i = 0; i < POOL_SIZE; i ++) {
            SoftReference<ByteBuffer> ref = pool[i];
            if (ref == null) {
                continue;
            }

            ByteBuffer buf = ref.get();
            if (buf == null) {
                pool[i] = null;
                continue;
            }

            if (buf.capacity() < size) {
                continue;
            }

            pool[i] = null;

            buf.clear();
            return buf;
        }

        ByteBuffer buf = ByteBuffer.allocateDirect(normalizeCapacity(size));
        buf.clear();
        return buf;
    }

    final void release(ByteBuffer buffer) {
        final SoftReference<ByteBuffer>[] pool = this.pool;
        for (int i = 0; i < POOL_SIZE; i ++) {
            SoftReference<ByteBuffer> ref = pool[i];
            if (ref == null || ref.get() == null) {
                pool[i] = new SoftReference<ByteBuffer>(buffer);
                return;
            }
        }

        // pool is full - replace one
        final int capacity = buffer.capacity();
        for (int i = 0; i< POOL_SIZE; i ++) {
            SoftReference<ByteBuffer> ref = pool[i];
            ByteBuffer pooled = ref.get();
            if (pooled == null) {
                pool[i] = null;
                continue;
            }

            if (pooled.capacity() < capacity) {
                pool[i] = new SoftReference<ByteBuffer>(buffer);
                return;
            }
        }
    }

    private static final int normalizeCapacity(int capacity) {
        // Normalize to multiple of 1024
        int q = capacity >>> 10;
        int r = capacity & 1023;
        if (r != 0) {
            q ++;
        }
        return q << 10;
    }
}

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。

    private boolean read(SelectionKey k) {
        final SocketChannel ch = (SocketChannel) k.channel();
        final NioSocketChannel channel = (NioSocketChannel) k.attachment();

        final ReceiveBufferSizePredictor predictor =
            channel.getConfig().getReceiveBufferSizePredictor();
        final int predictedRecvBufSize = predictor.nextReceiveBufferSize();

        int ret = 0;
        int readBytes = 0;
        boolean failure = true;

        ByteBuffer bb = recvBufferPool.acquire(predictedRecvBufSize);
        15         try {
            while ((ret = ch.read(bb)) > 0) {
                readBytes += ret;
                if (!bb.hasRemaining()) {
                    break;
                }
            }
            failure = false;
        } catch (ClosedChannelException e) {
            // Can happen, and does not need a user attention.
        } catch (Throwable t) {
            fireExceptionCaught(channel, t);
        }

        if (readBytes > 0) {
            bb.flip();

            final ChannelBufferFactory bufferFactory =
                channel.getConfig().getBufferFactory();
            final ChannelBuffer buffer = bufferFactory.getBuffer(readBytes);
            buffer.setBytes(0, bb);
            buffer.writerIndex(readBytes);
            //if(buffer instanceof BigEndianHeapChannelBuffer){
            //    logger2.info("buffer instanceof BigEndianHeapChannelBuffer.");
            //}
            recvBufferPool.release(bb);

            // Update the predi||\\|||||
            predictor.previousReceiveBufferSize(readBytes);

            // Fire the event.
            fireMessageReceived(channel, buffer);
        } else {
            recvBufferPool.release(bb);
        }

        if (ret < 0 || failure) {
            k.cancel(); // Some JDK implementations run into an infinite loop without this.
            close(channel, succeededFuture(channel));
            return false;
        }

        return true;
    }

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。