前些天有位朋友找到我,说他们的系统出现了CPU 100%的情况,让我帮忙看一下怎么回事?dump也拿到了,本想着这种情况让他多抓几个,既然有了就拿现有的分析吧。
既然说是 100%,作为调试者得拿数据说话,可以使用 !tp
来观测一下。
0:000:x86> !tp
CPU utilization: 100%
Worker Thread: Total: 382 Running: 382 Idle: 0 MaxLimit: 8191 MinLimit: 8
Work Request in Queue: 8694
Unknown Function: 6f62b650 Context: 4a36bbbc
Unknown Function: 6f62b650 Context: 4a36e1d4
Unknown Function: 6f62b650 Context: 4a372384
Unknown Function: 6f62b650 Context: 239adfec
Unknown Function: 6f62b650 Context: 4a374994
Unknown Function: 6f62b650 Context: 239b9e14
Unknown Function: 6f62b650 Context: 2399fd9c
...
从卦中看,不得了,CPU 100% 之外,所有的线程池线程全部被打满,人生自古最忌满,半贫半富半自安。同时线程池队列还累计了8694个任务待处理,说明这时候的线程池已经全面沦陷,要想找到这个答案,需要用 ~*e !clrstack
命令观察每一个线程此时正在做什么,输出如下:
0:000:x86> ~*e !clrstack
OS Thread Id: 0x22f4 (429)
Child SP IP Call Site
4bc1e060 0000002b [GCFrame: 4bc1e060]
4bc1e110 0000002b [HelperMethodFrame_1OBJ: 4bc1e110] System.Threading.Monitor.ObjWait(Boolean, Int32, System.Object)
4bc1e19c 24aad7da System.Threading.Monitor.Wait(System.Object, Int32, Boolean)
4bc1e1ac 2376f0d6 ServiceStack.Redis.PooledRedisClientManager.GetClient()
4bc1e1dc 2420bbc6 xxx.Service.CacheService.GetClient()
...
4bc1e234 24206fbe xxxBLL.GetxxxCount(System.Collections.Generic.Dictionary`2<System.String,System.Object>)
4bc1e3e0 216e25f9 DynamicClass.lambda_method(System.Runtime.CompilerServices.Closure, System.Web.Mvc.ControllerBase, System.Object[])
4bc1e3f0 238b86b7 System.Web.Mvc.ActionMethodDispatcher.Execute(System.Web.Mvc.ControllerBase, System.Object[])
...
4bc1eee0 2353d448 System.Web.Hosting.PipelineRuntime.ProcessRequestNotification(IntPtr, IntPtr, IntPtr, Int32)
4bc1efb8 00a9e3c2 [ContextTransitionFrame: 4bc1efb8]
从卦中可以看到当前有 371个线程在 PooledRedisClientManager.GetClient
中的 Wait 上出不来,那为什么出不来呢?
要想找到这个答案,只能从源代码中观察,简化后的代码如下:
public IRedisClient GetClient()
{
lock (writeClients)
{
AssertValidReadWritePool();
RedisClient inActiveWriteClient;
while ((inActiveWriteClient = GetInActiveWriteClient()) == null)
{
if (!Monitor.Wait(writeClients, PoolTimeout.Value))
{
throw new TimeoutException("Redis Timeout expired. The timeout period elapsed prior to obtaining a connection from the pool. This may have occurred because all pooled connections were in use.");
}
}
}
}
private RedisClient GetInActiveWriteClient()
{
int num = WritePoolIndex % writeClients.Length;
for (int i = 0; i < ReadWriteHosts.Count; i++)
{
int num2 = (num + i) % ReadWriteHosts.Count;
RedisEndPoint redisEndPoint = ReadWriteHosts[num2];
for (int j = num2; j < writeClients.Length; j += ReadWriteHosts.Count)
{
if (writeClients[j] != null && !writeClients[j].Active && !writeClients[j].HadExceptions)
{
return writeClients[j];
}
}
}
return null;
}
仔细阅读卦中代码,之所以进入Wait主要是因为 GetInActiveWriteClient()
方法返回 null 所致,从异常信息看也知道此时是因为 writeClients 池已满,那这个池是不是满了呢?可以把 writeClients 数组挖出来,使用 !dso
命令。
0:429:x86> !dso
OS Thread Id: 0x22f4 (429)
ESP/REG Object Name
...
4BC1E0D0 0ea38d18 ServiceStack.Redis.RedisClient[]
4BC1E100 0ea38bb0 ServiceStack.Redis.PooledRedisClientManager
...
0:429:x86> !da 0ea38d18
Name: ServiceStack.Redis.RedisClient[]
MethodTable: 237af1c0
EEClass: 0129a224
Size: 52(0x34) bytes
Array: Rank 1, Number of elements 10, Type CLASS
Element Methodtable: 237ae954
[0] 0ea38dd4
[1] 0a9f9f58
[2] 0296e468
[3] 0c9786a0
[4] 0a9fe768
[5] 04a21f24
[6] 0aa0d758
[7] 10946d90
[8] 04a8c8b0
[9] 02a2a2a0
0:429:x86> !DumpObj /d 0ea38dd4
Name: ServiceStack.Redis.RedisClient
MethodTable: 237ae954
EEClass: 2375d154
Size: 152(0x98) bytes
File: C:\Windows\xxx\ServiceStack.Redis.dll
Fields:
...
0129aa48 4000169 7d System.Boolean 1 instance 1 <Active>k__BackingField
...
从卦中看 writeClients 池只有10个大小,并且都是 Active=1
,所以返回 null 就不足为奇了。
要想找到这个答案,需要看下上层的 xxxBLL.GetxxxCount
方法是如何调用的,为了保护隐私,就多模糊一点。
从图中可以看到,问题出在用 foreach 去不断的迭代 ServiceStack.Redis 导致 writeClient 池耗尽,导致大量的请求在不断的阻塞,不要忘了这里有371个线程在争抢哦,真是大忌。
接下来顺带洞察下这个 foreach 要 foreach 多少次? 继续用 !dso 去挖。
0:429:x86> !DumpObj /d 077cec20
Name: System.Collections.Generic.List`1[[xxxInfo, xxx]]
MethodTable: 241ad794
EEClass: 0193166c
Size: 24(0x18) bytes
File: C:\Windows\Microsoft.Net\assembly\GAC_32\mscorlib\v4.0_4.0.0.0__b77a5c561934e089\mscorlib.dll
Fields:
MT Field Offset Type VT Attr Value Name
01860eec 4001891 4 System.__Canon[] 0 instance 077e0048 _items
0129c9b0 4001892 c System.Int32 1 instance 307 _size
0129c9b0 4001893 10 System.Int32 1 instance 307 _version
01296780 4001894 8 System.Object 0 instance 00000000 _syncRoot
01860eec 4001895 4 System.__Canon[] 0 static <no information>
从卦中看当前需要循环307次,也就再次验证了池耗尽的说法,我知道心细的朋友肯定会说,卡死这个我认,但能导致 CPU爆高 我就不能理解了,其实你仔细阅读源码就能理解了,这是经典的 锁护送(lock convoy)
现象,因为满足如下两个条件。
知道了前因后果,解决起来就比较简单了,三种做法:
这次生产事故分析还是非常有意思的,一个看似阻塞的问题也会引发CPU爆高,超出了一些人的认知吧,对,其实它就是经典的 lock convoy
现象,大家有任何dump问题可以找我,一如既往的免费分析。