一:背景
1. 讲故事
前段时间有位朋友微信上找到我,说它的程序出现了卡死,让我帮忙看下是怎么回事? 说来也奇怪,那段时间求助卡死类的dump特别多,被迫训练了一下对这类问题的洞察力 😄😄😄,再次声明一下,我分析 dump 是免费的,没有某软高额的分析费用,你要问我图什么,图技术的精进。
回到正题,卡死类的问题分析入口点在于主线程此时在做什么,导致它不能处理自己的任务队列的原因是各种各样的,接下来上 windbg 分析一下。
二:WinDbg 分析
1. 主线程此时在做什么
看主线程很简单,直接一个 k
命令搞定。
:000> k
# ChildEBP RetAddr
00 00f3ec2c 74ef6439 ntdll!NtWaitForSingleObject+0xc
01 00f3ec2c 70293370 KERNELBASE!WaitForSingleObjectEx+0x99
02 00f3ec5c 702933cc clr!CLREventWaitHelper2+0x33
03 00f3ecac 70293319 clr!CLREventWaitHelper+0x2a
04 00f3ece4 7037bbcf clr!CLREventBase::WaitEx+0x14b
05 00f3ecfc 70374f08 clr!WKS::GCHeap::WaitUntilGCComplete+0x35
06 00f3ed44 7032c2a2 clr!Thread::RareDisablePreemptiveGC+0x20b
07 00f3edc8 6d453712 clr!JIT_RareDisableHelper+0x22
08 00f3ee4c 6d4530d1 System_Windows_Forms_ni!System.Windows.Forms.Application+ComponentManager.System.Windows.Forms.UnsafeNativeMethods.IMsoComponentManager.FPushMessageLoop(IntPtr, Int32, Int32)$##600550A+0x48e
09 00f3eea0 6d452f23 System_Windows_Forms_ni!System.Windows.Forms.Application+ThreadContext.RunMessageLoopInner(Int32, System.Windows.Forms.ApplicationContext)$##6005539+0x175
a 00f3eecc 6d42b83d System_Windows_Forms_ni!System.Windows.Forms.Application+ThreadContext.RunMessageLoop(Int32, System.Windows.Forms.ApplicationContext)$##6005538+0x4f
0b 00f3eee4 02d90c83 System_Windows_Forms_ni!System.Windows.Forms.Application.Run(System.Windows.Forms.Form)$##60005FA+0x35
...
从卦上信息看,主线程正在等待(WaitUntilGCComplete) GC 处理完成,接下来看下是哪一个线程触发了 GC
操作,并看下它正在 GC 三阶段的哪一步?
:000> !t
ThreadCount: 33
UnstartedThread: 3
BackgroundThread: 29
PendingThread: 3
DeadThread:
Hosted Runtime: no
Lock
ID OSID ThreadOBJ State GC Mode GC Alloc Context Domain Count Apt Exception
1 17a0 0113ecd0 26020 Preemptive 00000000:00000000 01139490 STA
2 2 2074 0114e0e0 2b220 Preemptive 00000000:00000000 01139490 MTA (Finalizer)
5 3 27f4 011ddaa8 102a220 Preemptive 00000000:00000000 01139490 MTA (Threadpool Worker)
8 4 2568 05bf0b90 1029220 Cooperative 00000000:00000000 01139490 1 MTA (GC) (Threadpool Worker)
:000> ~8s
eax=00000000 ebx=05bf0b90 ecx=00000000 edx=00000000 esi=00000000 edi=00000f6c
eip=77e5aa5c esp=098ce03c ebp=098ce0ac iopl= nv up ei pl nz na pe nc
cs=0023 ss=002b ds=002b es=002b fs=0053 gs=002b efl=00000206
ntdll!NtWaitForSingleObject+0xc:
77e5aa5c c20c00 ret Ch
:008> k
# ChildEBP RetAddr
00 098ce0ac 74ef6439 ntdll!NtWaitForSingleObject+0xc
01 098ce0ac 70293370 KERNELBASE!WaitForSingleObjectEx+0x99
02 098ce0dc 702933cc clr!CLREventWaitHelper2+0x33
03 098ce12c 70293319 clr!CLREventWaitHelper+0x2a
04 098ce164 7029333b clr!CLREventBase::WaitEx+0x14b
05 098ce17c 7032a9b7 clr!CLREventBase::Wait+0x1a
06 098ce1fc 7032a9ef clr!`anonymous namespace'::CreateSuspendableThread+0x165
07 098ce378 704322bd clr!GCToEEInterface::CreateThread+0x15c
08 098ce3a8 704332e5 clr!WKS::gc_heap::prepare_bgc_thread+0x36
09 098ce3a8 70378585 clr!WKS::gc_heap::garbage_collect+0x215
a 098ce3c8 7037868a clr!WKS::GCHeap::GarbageCollectGeneration+0x1bd
0b 098ce3ec 70378703 clr!WKS::gc_heap::trigger_gc_for_alloc+0x1f
c 098ce424 70396f6a clr!WKS::gc_heap::try_allocate_more_space+0x152
d 098ce46c 70397311 clr!WKS::gc_heap::allocate_large_object+0x51
e 098ce478 702fde2d clr!WKS::GCHeap::AllocLHeap+0x11
0f 098ce494 7063f1af clr!AllocLHeap+0x4b
10 098ce4e8 7028f598 clr!FastAllocatePrimitiveArray+0xa7
11 098ce58c 0e2d3fb6 clr!JIT_NewArr1+0x126
12 098cf26c 6f1f1b5a 0xe2d3fb6
13 098cf274 6f188604 mscorlib_ni!System.Runtime.CompilerServices.AsyncMethodBuilderCore+MoveNextRunner.InvokeMoveNext(System.Object)$##6006FCB+0x1a
...
从卦中的 clr!GCToEEInterface::CreateThread
方法看,如果你没有这方面的分析经验,或许你已经无能为力了,以我的经验来说,既然卡死,那这里的 CreateThread
自然无法创建成功,创建不成功的原因在于 进程加载锁 LdrpLoaderLock
被别人持有了,导致它获取不到这把锁而一直茫然等待。
2. 谁持有了加载锁
要想找到谁持有了加载锁,可以在 ntdll 中搜 LdrpLoaderLock
字段,接下来使用 x 命令搜索。
:008> x ntdll!LdrpLoaderLock
77f053b8 ntdll!LdrpLoaderLock = <no type information>
:008> !cs 77f053b8
-----------------------------------------
Critical section = 0x77f053b8 (ntdll!LdrpLoaderLock+0x0)
DebugInfo = 0x77f0573c
LOCKED
LockCount = 0x0
WaiterWoken = No
OwningThread = 0x00001314
RecursionCount = 0x1
LockSemaphore = 0x0
SpinCount = 0x04000000
从卦中的 OwningThread = 0x00001314
迹象看,当前是线程 1314
持有了加载锁,看样子是要一生一世的持有。。。。 难怪程序不卡死。。。。 接下来切到这个线程看看到底咋这么痴情?
:043> ~~[1314]s
eax=00000000 ebx=0589f784 ecx=00000000 edx=00000000 esi=00000000 edi=0589f770
eip=77e5c6bc esp=0589f72c ebp=0589f748 iopl= nv up ei pl nz na po nc
cs=0023 ss=002b ds=002b es=002b fs=0053 gs=002b efl=00000202
ntdll!NtWaitForAlertByThreadId+0xc:
77e5c6bc c20800 ret 8
:043> kb
# ChildEBP RetAddr Args to Child
00 0589f748 77e1336b ad114e8 00000000 00000000 ntdll!NtWaitForAlertByThreadId+0xc
01 0589f748 77e50630 00000000 00000000 ffffffff ntdll!RtlpWaitOnAddressWithTimeout+0x33
02 0589f78c 77e13299 00000004 00000000 00000000 ntdll!RtlpWaitOnAddress+0xa5
03 0589f7c8 77e2ed76 00000000 ad114e0 ad114e4 ntdll!RtlpWaitOnCriticalSection+0xaf
04 0589f7f0 77e2ec99 0589f814 ac5d1b4 ad114e4 ntdll!RtlpEnterCriticalSectionContended+0xd6
05 0589f7f8 ac5d1b4 ad114e4 00000003 ac50000 ntdll!RtlEnterCriticalSection+0x49
06 0589f814 ac5a890 00001314 acddd2a ac50000 fwlibe1!xxx+0x13e3
07 0589f83c 77e5a9f6 ac50000 00000003 00000000 fwlibe1!xxx+0x290
...
从卦中看,当前是 fwlibe1
这个非托管dll 准备获取 临界区
时卡死,接下来可以用 !cs
把 临界区变量 给提取出来看看到底谁在持有这个 临界区变量。
:043> !cs ad114e4
-----------------------------------------
Critical section = 0x0ad114e4 (fwlibe1!xxxx+0x3A1C2)
DebugInfo = 0x0ff43360
LOCKED
LockCount = 0x2
WaiterWoken = No
OwningThread = 0x00002378
RecursionCount = 0x1
LockSemaphore = 0xFFFFFFFF
SpinCount = 0x020007d0
从卦中可以清晰的看到,当前临界区变量被 2378
号线程持有,接下来切到这个线程看看到底啥情况。
:043> ~~[2378]s
eax=00000036 ebx=00002378 ecx=00000400 edx=00000157 esi=ad114fc edi=ad114e4
eip=ac5cf88 esp=caddf18 ebp=cade7fc iopl= nv up ei pl nz na po nc
cs=0023 ss=002b ds=002b es=002b fs=0053 gs=002b efl=00000202
fwlibe1!xxx+0x11b7:
ac5cf88 8d45f0 lea eax,[ebp-10h]
:028> k
# ChildEBP RetAddr
WARNING: Stack unwind information not available. Following frames may be wrong.
00 cade7fc ac5a8df fwlibe1!xxx+0x11b7
01 cade938 ac5abae fwlibe1!xxx+0x22
02 cadea20 09417ccd fwlibe1!xxx+0x17
03 cadebfc 09412c11 0x9417ccd
04 caded2c 6f1f1b5a 0x9412c11
05 caded34 6f188604 mscorlib_ni!System.Runtime.CompilerServices.AsyncMethodBuilderCore+MoveNextRunner.InvokeMoveNext(System.Object)$##6006FCB+0x1a
...
:028> !clrstack
OS Thread Id: 0x2378 (28)
Child SP IP Call Site
cade9b4 ac5cf88 [InlinedCallFrame: cade9b4]
cade998 0941a0e9 *** WARNING: Unable to verify checksum for System.Data.ni.dll
DomainBoundILStubClass.IL_STUB_PInvoke(System.Object, UInt16, Int32, UInt16 ByRef)
cade9b4 09417ccd [InlinedCallFrame: cade9b4] Focas.xxx(System.Object, UInt16, Int32, UInt16 ByRef)
从卦中看,托管层调用了 Focas.xxx
函数进入了 fwlibe1.dll
,后有一句 WARNING: Stack unwind information not available. Following frames may be wrong.
,可能是展开有一定的问题,不过排查到这里,应该就是这个 Focas.xxx
的问题了。
3. 真相大白
接下来将我的排查结果跟朋友反馈了下,朋友排查下来说是近改了这里面的连接方式所致,修改之后就搞定了。
三:总结
这个案例告诉我们:只要代码还能跑,就不要动它,它要是不动了,你得要做好动的准备。