current position:Home>Remember a CPU burst analysis of an entertainment chat streaming platform in .NET

Remember a CPU burst analysis of an entertainment chat streaming platform in .NET

2022-10-27 03:32:35dotNET cross-platform

一:背景

1.讲故事

前段时间有位朋友加微信,说他的程序直接 CPU=100%,每次只能手工介入重启,让我帮忙看下到底怎么回事,哈哈,这种CPU打满的事故,程序员压力会非常大, 我让朋友在 CPU 高的时候抓 2 个 dump 下来,然后发给我分析.

二:WinDbg 分析

1. CPU 真的被打满了吗?

为了防止南辕北辙,一定要用 !tp 命令去验证下是不是真的 CPU 爆高.

0:000> !tp
CPU utilization: 100%
Worker Thread: Total: 21 Running: 7 Idle: 0 MaxLimit: 32767 MinLimit: 4
Work Request in Queue: 3
    AsyncTimerCallbackCompletion [email protected]
    AsyncTimerCallbackCompletion [email protected]
    AsyncTimerCallbackCompletion [email protected]
--------------------------------------
Number of Timers: 0
--------------------------------------
Completion Port Thread:Total: 18 Free: 9 MaxFree: 8 CurrentLimit: 18 MaxLimit: 1000 MinLimit: 4

从卦中看确实 100%,太牛了,而且 WorkRequest 还有任务堆积现象,确认无疑后,接下来看下是谁引发的?

2. 谁导致的 CPU 爆高

根据惯例首先怀疑是不是 GC 触发所致,可以用 !t 查看下线程列表,观察下有没有 GC 字样.

:000> !t
ThreadCount:      53
UnstartedThread:  0
BackgroundThread: 53
PendingThread:    0
DeadThread:       0
Hosted Runtime:   no
                                                                                                        Lock  
       ID OSID ThreadOBJ           State GC Mode     GC Alloc Context                  Domain           Count Apt Exception
   4    1 1240 00000000021cdf30    2a220 Preemptive  0000000000000000:0000000000000000 00000000021d94c0 0     MTA 
  23    2 4db4 00000000041cdaa0    2b220 Preemptive  0000000000000000:0000000000000000 00000000021d94c0 0     MTA (Finalizer) 
  ...
  65  156 22f4 000000000b1a3f60  8029220 Preemptive  00000004527751F0:0000000452775EE8 00000000021d94c0 0     MTA (Threadpool Completion Port) 
  66  205 2ef8 000000000b1a1080  8029220 Preemptive  0000000157641DE0:00000001576435B0 00000000021d94c0 0     MTA (Threadpool Completion Port) 
  ...

从卦中看没有 GC 字样,也表明这个程序并不是由 GC 触发所致,接下来该怎么排查呢?一般来说 CPU 的爆高是由线程抬起来的,所以接下来就是看下 CPU 的档次和各个线程栈,看有没有什么新线索, 可以使用 ~*e !clrstack .

0:000> !cpuid
CP  F/M/S  Manufacturer     MHz
 0  6,79,1  <unavailable>   2299
 1  6,79,1  <unavailable>   2299
 2  6,79,1  <unavailable>   2299
 3  6,79,1  <unavailable>   2299

0:000> ~*e !clrstack 
OS Thread Id: 0x2cc4 (68)
        Child SP               IP Call Site
000000000c14e758 00007ffadeb86e4a [GCFrame: 000000000c14e758] 
000000000c14e840 00007ffadeb86e4a [GCFrame: 000000000c14e840] 
000000000c14e878 00007ffadeb86e4a [HelperMethodFrame: 000000000c14e878] System.Threading.Monitor.Enter(System.Object)
000000000c14e970 00007ffaceb40491 System.Net.ConnectionGroup.Disassociate(System.Net.Connection) [f:\dd\NDP\fx\src\net\System\Net\_ConnectionGroup.cs @ 148]
000000000c14e9d0 00007ffaceb3fc93 System.Net.Connection.PrepareCloseConnectionSocket(System.Net.ConnectionReturnResult ByRef) [f:\dd\NDP\fx\src\net\System\Net\_Connection.cs @ 3048]
000000000c14eaa0 00007ffacf139bfb System.Net.Connection.HandleError(Boolean, Boolean, System.Net.WebExceptionStatus, System.Net.ConnectionReturnResult ByRef) [f:\dd\NDP\fx\src\net\System\Net\_Connection.cs @ 3119]
000000000c14eb00 00007ffacebc4118 System.Net.Connection.ReadComplete(Int32, System.Net.WebExceptionStatus) [f:\dd\NDP\fx\src\net\System\Net\_Connection.cs @ 3387]
000000000c14eb80 00007ffacebe3dc5 System.Net.LazyAsyncResult.Complete(IntPtr) [f:\dd\NDP\fx\src\net\System\Net\_LazyAsyncResult.cs @ 415]
000000000c14ebe0 00007ffacebe3d07 System.Net.LazyAsyncResult.ProtectedInvokeCallback(System.Object, IntPtr) [f:\dd\NDP\fx\src\net\System\Net\_LazyAsyncResult.cs @ 368]
000000000c14ec20 00007ffacf3a476f System.Net.Security._SslStream.StartFrameBody(Int32, Byte[], Int32, Int32, System.Net.AsyncProtocolRequest)
000000000c14ec80 00007ffacebb3ed8 System.Net.Security._SslStream.ReadHeaderCallback(System.Net.AsyncProtocolRequest) [f:\dd\NDP\fx\src\net\System\Net\SecureProtocols\_SslStream.cs @ 1007]
000000000c14ece0 00007ffacebae5ee System.Net.AsyncProtocolRequest.CompleteRequest(Int32) [f:\dd\NDP\fx\src\net\System\Net\SecureProtocols\_HelperAsyncResults.cs @ 142]
000000000c14ed10 00007ffacf3a3567 System.Net.FixedSizeReader.CheckCompletionBeforeNextRead(Int32)
000000000c14ed40 00007ffacebae507 System.Net.FixedSizeReader.ReadCallback(System.IAsyncResult) [f:\dd\NDP\fx\src\net\System\Net\SecureProtocols\_FixedSizeReader.cs @ 148]
000000000c14ed90 00007ffacebe3dc5 System.Net.LazyAsyncResult.Complete(IntPtr) [f:\dd\NDP\fx\src\net\System\Net\_LazyAsyncResult.cs @ 415]
000000000c14edf0 00007ffadcbe3a63 System.Threading.ExecutionContext.RunInternal(System.Threading.ExecutionContext, System.Threading.ContextCallback, System.Object, Boolean) [f:\dd\ndp\clr\src\BCL\system\threading\executioncontext.cs @ 954]
000000000c14eec0 00007ffadcbe38f4 System.Threading.ExecutionContext.Run(System.Threading.ExecutionContext, System.Threading.ContextCallback, System.Object, Boolean) [f:\dd\ndp\clr\src\BCL\system\threading\executioncontext.cs @ 902]
000000000c14eef0 00007ffadcbe38c2 System.Threading.ExecutionContext.Run(System.Threading.ExecutionContext, System.Threading.ContextCallback, System.Object) [f:\dd\ndp\clr\src\BCL\system\threading\executioncontext.cs @ 891]
000000000c14ef40 00007ffaceba60cf System.Net.ContextAwareResult.Complete(IntPtr) [f:\dd\NDP\fx\src\net\System\Net\_ContextAwareResult.cs @ 463]
000000000c14ef90 00007ffacebe3d07 System.Net.LazyAsyncResult.ProtectedInvokeCallback(System.Object, IntPtr) [f:\dd\NDP\fx\src\net\System\Net\_LazyAsyncResult.cs @ 368]
000000000c14efd0 00007ffaceba5e2f System.Net.Sockets.BaseOverlappedAsyncResult.CompletionPortCallback(UInt32, UInt32, System.Threading.NativeOverlapped*) [f:\dd\NDP\fx\src\net\System\Net\Sockets\_BaseOverlappedAsyncResult.cs @ 399]
000000000c14f040 00007ffadcc2ffef System.Threading._IOCompletionCallback.PerformIOCompletionCallback(UInt32, UInt32, System.Threading.NativeOverlapped*) [f:\dd\ndp\clr\src\BCL\system\threading\overlapped.cs @ 135]
000000000c14f1f0 00007ffade9a6d93 [GCFrame: 000000000c14f1f0] 

OS Thread Id: 0x5d6c (70)
        Child SP               IP Call Site
000000000943e920 00007ffacf139bfb System.Net.Connection.HandleError(Boolean, Boolean, System.Net.WebExceptionStatus, System.Net.ConnectionReturnResult ByRef) 

OS Thread Id: 0x4ad4 (75)
        Child SP               IP Call Site
...
000000000c94e5a0 00007ffacf139bfb System.Net.Connection.HandleError(Boolean, Boolean, System.Net.WebExceptionStatus, System.Net.ConnectionReturnResult ByRef)
...

OS Thread Id: 0x1d70 (80)
        Child SP               IP Call Site
...
000000000d24e3a0 00007ffacf139bfb System.Net.Connection.HandleError(Boolean, Boolean, System.Net.WebExceptionStatus, System.Net.ConnectionReturnResult ByRef) [f:\dd\NDP\fx\src\net\System\Net\_Connection.cs @ 3119]
...

从线程栈上看,这个 CPU 是 4个核,刚好对应着 4 个 HandleError 报错,看样子是什么网络出问题了,接下来切到 80 号线程看一下有没有什么异常类.

0:000> ~80s
clr!AwareLock::Contention+0x194:
00007ffa`deb86e40 4883e801        sub     rax,1
0:080> !mdso
Thread 80:
Location          Object            Type
------------------------------------------------------------
000000000d24e098  000000015765e028  System.Net.WebException
000000000d24e0f8  0000000340b07110  System.Collections.ArrayList
000000000d24e110  000000015765e2b8  System.Net.HttpWebRequest[]
000000000d24e1c0  0000000340b070b8  System.Net.ConnectionGroup
000000000d24e258  0000000144a79678  System.Net.Connection

0:080> !mdt 000000015765e028
000000015765e028 (System.Net.WebException)
    _className:NULL (System.String)
    _exceptionMethod:NULL (System.Reflection.MethodBase)
    _exceptionMethodString:NULL (System.String)
    _message:000000015765df70 (System.String) Length=77, String="The underlying connection was closed: The connection was closed unexpectedly."
    ...

果然看到了 System.Net.WebException, 从异常信息看貌似是 连接关闭了,到这里我就有了一个大胆的猜测,是不是高频的异常输出导致的 CPU 爆高呢?为了验证,可以到托管堆上找下 WebException 的个数.

0:080> !dumpheap -stat
Statistics:
              MT    Count    TotalSize Class Name
...
00007ffacecc38b0    13315      2343440 System.Net.WebException
00007ffadcdf6570    11369      1909992 System.IO.IOException
00007ffadcdf5fb8    13380      2247840 System.ObjectDisposedException
...

看到这么多异常还是挺吓人的,刚好朋友抓了两个dump可以做个比较.

0:048> !dumpheap -stat
Statistics:
              MT    Count    TotalSize Class Name
00007ffacecc38b0    26745      4707120 System.Net.WebException
00007ffadcdf6570    26722      4489296 System.IO.IOException
00007ffadcdf5fb8    28745      4829160 System.ObjectDisposedException

可以看到,2 min 之内异常增加了合计 4w 左右,这就验证了程序确实是疯狂的抛异常,在 Windows 平台上不管是硬件异常还是软件异常都由 Windows SEH 异常处理框架统一处理,会出现用户态和内核态的切换,这样疯狂的抛出,必然会导致 CPU 爆高,终于找到原因了,接下来就是寻找诱发因素.

3. 异常是谁诱发的

再回头看 HandleError 函数的调用栈都是底层的库函数,从线程栈的 PerformIOCompletionCallback 函数来看是 IO线程 诱发的,能被 IO 线程兜到是因为这是做了异步处理,既然是 异步,自然 OverlappedData 也会非常多.

0:080> !gchandles -stat
Statistics:
              MT    Count    TotalSize Class Name
00007ffadc6f7b98    14511      1625232 System.Threading.OverlappedData
Total 17550 objects

Handles:
    Strong Handles:       426
    Pinned Handles:       23
    Async Pinned Handles: 14511
    Ref Count Handles:    24
    Weak Long Handles:    2430
    Weak Short Handles:   132
    SizedRef Handles:     4

说明此时有大概 1.5w 的异步请求待回头,请求量还是蛮大的,但还是没找到异常的用户代码,只能找下到底是谁发起了什么请求.

0:080> !mdso
Thread 80:
Location          Object            Type
------------------------------------------------------------
...
000000000d24e488  0000000358c57918  System.Net.HttpWebRequest
000000000d24e2e8  00000001407b5b40  System.String  "net_io_readfailure"
...

0:080> !mdt -r:2 0000000358c57918
0000000358c57918 (System.Net.HttpWebRequest)
    _Uri:0000000358c57210 (System.Uri)
        m_String:00000002407ee430 (System.String) Length=98, String="https://api.xxxx/peer_messages"
        ....

可以看到请求的连接是 https://api.xxxx/peer_messages,是一个第三方的API接口, 由于底层的连接关闭,导致了最后 net_io_readfailure.

把所有的信息整合一下就是:

当请求量大了之后,访问 https://api.xxxx/peer_messages 会出问题,对方关闭了底层连接,导致客户端这边请求出现了大量 IO 回调异常:IOException: Unable to read data from the transport connection: The connection was closed.,2min之间多达合计 4w 的异常抛出,进而引发 CPU 爆高,将信息告诉了朋友,让朋友重点关注下 https://api.xxxx/peer_messages 这个连接.

三:总结

这次生产事故主要是由于高峰期请求量过大,由于某种原因 Socket 连接关闭,导致了大量的异步回调异常.

解决方法在调用端做好限流,据朋友说减少了不必要的 https://api.xxxx/peer_messages 调用,目前没有出现 CPU 爆高现象.

copyright notice
author[dotNET cross-platform],Please bring the original link to reprint, thank you.
https://en.bfun.fun/2022/300/202210270329025399.html

Random recommended