概述
最近线上的日志处理服务偶尔会出现Out Of Memory的问题,从Exception的call stack中顺藤摸瓜,最终定位到是thrift反序列化的问题。
发现问题
先交代一下问题现场:
- thirft版本: 0.5.0,很久远的版本,但是公司统一使用的版本;
- 反序列化使用的协议:TCompactProtocol协议;
-
出错的call stack:
Exception in thread "pool-10-thread-1" java.lang.RuntimeException: java.lang.OutOfMemoryError: Requested array size exceeds VM limit
at com.lmax.disruptor.FatalExceptionHandler.handleEventException(FatalExceptionHandler.java:45)
at com.lmax.disruptor.BatchEventProcessor.run(BatchEventProcessor.java:152)
at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
at java.lang.Thread.run(Thread.java:745)
Caused by: java.lang.OutOfMemoryError: Requested array size exceeds VM limit
at org.apache.thrift.protocol.TCompactProtocol.readBinary(TCompactProtocol.java:651)
at org.apache.thrift.protocol.TCompactProtocol.readString(TCompactProtocol.java:626)
at com.xiaomi.data.spec.log.push.XmPushMessageInfo.read(XmPushMessageInfo.java:2779)
寻找原因
从上面的call stack可以看出,问题出在thrift TCompactProtocol类的651行,看看这一行干了些什么:
/**
* Read a byte[] of a known length from the wire.
*/
private byte[] readBinary(int length) throws TException {
if (length == 0) return new byte[0];
byte[] buf = new byte[length]; // TCompactProtocol第651行
trans_.readAll(buf, 0, length);
return buf;
}
从上面代码可以看到,TCompactProtocol类的651行申请了一个长度为length的byte数组,而此时可用内存已经不足以分配这么大的空间,所以报了java.lang.OutOfMemoryError错误,导致程序异常退出。
这个length是怎么来的呢?还是从call stack寻找答案,看看TCompactProtocol类的626行:
/**
* Reads a byte[] (via readBinary), and then UTF-8 decodes it.
*/
public String readString() throws TException {
int length = readVarint32();
if (length == 0) {
return "";
}
try {
if (trans_.getBytesRemainingInBuffer() >= length) {
String str = new String(trans_.getBuffer(), trans_.getBufferPosition(), length, "UTF-8");
trans_.consumeBuffer(length);
return str;
} else {
return new String(readBinary(length), "UTF-8"); // 626行
}
} catch (UnsupportedEncodingException e) {
throw new TException("UTF-8 not supported!");
}
}
从上面的代码可以看到,length是通过 readVarint32
这个函数读到的一个int型数字,然后thrift使用这个数字来申请内存。
这种方式在正常情况下是没有问题的,但是如果源binary数据被写坏了,或者网络传输过程中出现了差错,就有可能导致 readVarint32
读到的是一个非常大的数字(可能达到10多亿),这种情况下申请内存必然会OOM。
解决问题
问题原因找到了,但是怎么解决呢?
下面是一个比较简单的解决方案:
每次读取到length之后都做一下长度的check,如果这个长度超过一定的长度,则直接抛出异常,不要再申请内存。
thrift中需要check读取到的langth的地方有以下几个地方(如果使用其他Protocol也类似):
/**
* Read a map header off the wire. If the size is zero, skip reading the key
* and value type. This means that 0-length maps will yield TMaps without the
* "correct" types.
*/
public TMap readMapBegin() throws TException {
int size = readVarint32(); //此处需要check size
byte keyAndValueType = size == 0 ? 0 : readByte();
return new TMap(getTType((byte)(keyAndValueType >> 4)), getTType((byte)(keyAndValueType & 0xf)), size);
}
/**
* Read a list header off the wire. If the list size is 0-14, the size will
* be packed into the element type header. If it's a longer list, the 4 MSB
* of the element type header will be 0xF, and a varint will follow with the
* true size.
*/
public TList readListBegin() throws TException {
byte size_and_type = readByte();
int size = (size_and_type >> 4) & 0x0f;
if (size == 15) {
size = readVarint32(); // 此处需要check size
}
byte type = getTType(size_and_type);
return new TList(type, size);
}
/**
* Reads a byte[] (via readBinary), and then UTF-8 decodes it.
*/
public String readString() throws TException {
int length = readVarint32(); // 此处需要check length
if (length == 0) {
return "";
}
try {
if (trans_.getBytesRemainingInBuffer() >= length) {
String str = new String(trans_.getBuffer(), trans_.getBufferPosition(), length, "UTF-8");
trans_.consumeBuffer(length);
return str;
} else {
return new String(readBinary(length), "UTF-8");
}
} catch (UnsupportedEncodingException e) {
throw new TException("UTF-8 not supported!");
}
}
以 readMapBegin
举例,可以这样修改:
/**
* Read a map header off the wire. If the size is zero, skip reading the key
* and value type. This means that 0-length maps will yield TMaps without the
* "correct" types.
*/
public TMap readMapBegin() throws TException {
int size = readVarint32();
if (size > trans_.getBytesRemainingInBuffer() || size > MAX_THRIFT_MAP_SIZE) {
throw new TPushProtocolException(TProtocolException.SIZE_LIMIT, "Thrift map size " + size + " out of range, remaining size = " + trans_.getBytesRemainingInBuffer());
}
byte keyAndValueType = size == 0 ? 0 : readByte();
return new TMap(getTType((byte)(keyAndValueType >> 4)), getTType((byte)(keyAndValueType & 0xf)), size);
}
其中的 MAX_THRIFT_MAP_SIZE
是一个常量,一个自定义的thrift map的最大size,此处是10000。
后记
看了一下thrift 0.9.3版本的源码,这个版本中已经加上了类似的check逻辑。
http://outofmemory.cn/java/thrift-desearialize-outOfMemory