一：背景
1. 讲故事
前两天给训练营里的一位学员分析了一个dump，学员因为弄了一整天也没找到祸根，被我一下子弄出来了，极度想看看我是怎么分析的？由于在微信上不能一言两语表尽，干脆写一篇文章出来详细的讲讲吧，哈哈，训练营里的学员得有求必应哈。。。话不多说，我们一起探索下这个程序的崩溃之路吧。
二：WinDbg分析
1. 为什么会崩溃
这个比较简单，因为默认会自动定位到崩溃的线程，使用 .ecxr 切到崩溃前的上下文即可，然后使用 k 观察崩溃点，参考如下：
0:044> .ecxr
rax=0000000000000000 rbx=00000217a6c78f90 rcx=0000000000000000
rdx=0000000000000000 rsi=000000242c46fdd0 rdi=00000217ff1594a0
rip=00007ffdb35eb699 rsp=000000242c46fb40 rbp=000000242c46fc70
r8=00007ffdb35eb699 r9=00000217ea61fc08 r10=00000000e0434352
r11=0000000000000001 r12=00000217a6c78f90 r13=0000000000000000
r14=0000000000000000 r15=0000000000000000
iopl=0 nv up ei pl nz na pe nc
cs=0033 ss=002b ds=002b es=002b fs=0053 gs=002b efl=00000202
KERNELBASE!RaiseException+0x69:
00007ffd`b35eb699 0f1f440000 nop dword ptr [rax+rax]
0:044> k
*** Stack trace for last set context - .thread/.cxr resets it
# Child-SP RetAddr Call Site
00 00000024`2c46fb40 00007ffd`51bf8c6f KERNELBASE!RaiseException+0x69
01 00000024`2c46fc20 00007ffd`51890108 clr!UMThunkStubRareDisableWorker+0x3f
02 00000024`2c46fc50 00007ffc`dc826748 clr!UMThunkStub+0x128
03 00000024`2c46fce0 00007ffc`dc827406 MVCAMSDK_X64+0x6748
04 00000024`2c46fd80 00007ffc`dc9d21eb MVCAMSDK_X64!CameraGrabber_GetCameraDevInfo+0xb96
05 00000024`2c46fe50 00007ffc`dc9d227f MVCAMSDK_X64!CameraGigeEnumerateDevice+0x167f2b
06 00000024`2c46fe80 00007ffd`b5a07374 MVCAMSDK_X64!CameraGigeEnumerateDevice+0x167fbf
07 00000024`2c46feb0 00007ffd`b5ebcc91 kernel32!BaseThreadInitThunk+0x14
08 00000024`2c46fee0 00000000`00000000 ntdll!RtlUserThreadStart+0x21
从卦中可以看到程序是在clr 的 UMThunkStubRareDisableWorker 函数中故意抛出的异常，接下来观察下该函数代码。
extern "C" VOID __stdcall UMThunkStubRareDisableWorker(Thread* pThread, UMEntryThunk* pUMEntryThunk, Frame* pFrame)
{
if (!CanRunManagedCode())
{
pThread->m_fPreemptiveGCDisabled = 0;
COMPlusThrowBoot(E_PROCESS_SHUTDOWN_REENTRY);
}
}
BOOL CanRunManagedCode(BOOL fCannotRunIsUserError, HINSTANCE hInst)
{
if (g_fForbidEnterEE == TRUE)
return FALSE;
if ((g_fEEShutDown & ShutDown_Finalize2) && !GCHeap::GetGCHeap()->IsCurrentThreadFinalizer())
return FALSE;
if (g_pPreallocatedOutOfMemoryException == NULL)
return FALSE;
return TRUE;
}
从卦中的代码看是因为 CanRunManagedCode() 返回false导致的，到内存中观察这几个变量即可。
0:044> dp clr!g_fForbidEnterEE L1
00007ffd`5201d548 00000000`00000000
0:044> dp clr!g_fEEShutDown L1
00007ffd`520140b0 00000000`00000007
0:044> dp clr!g_pPreallocatedOutOfMemoryException L1
00007ffd`52013e08 00000217`e4f513d8
看到卦中的几个变量大概就知道了，原来当前程序正在等待 终结器线程 执行完毕的 EEShutDown 状态。
2. 真的处于关闭状态吗
刚才只是通过变量观察，接下来我们观察线程栈来进一步求证，分别切到主线程和终结器线程观察调用栈，参考如下：
0:000> ~0s;k;~2s;k
win32u!NtUserMsgWaitForMultipleObjectsEx+0x14:
00007ffd`b3dfa104 c3 ret
# Child-SP RetAddr Call Site
00 00000024`2773f578 00007ffd`b3f40c8e win32u!NtUserMsgWaitForMultipleObjectsEx+0x14
...
08 00000024`2773f9c0 00007ffd`518828ff clr!CLREventBase::WaitEx+0xab
09 00000024`2773fa30 00007ffd`51882842 clr!WaitForEndOfShutdown_OneIteration+0xb3
0a 00000024`2773fac0 00007ffd`5171f703 clr!WaitForEndOfShutdown+0x1a
0b 00000024`2773faf0 00007ffd`518243f9 clr!EEShutDown+0xd3
0c 00000024`2773fb40 00007ffd`51823f34 clr!HandleExitProcessHelper+0x29
0d 00000024`2773fb70 00007ffd`51823e14 clr!_CorExeMainInternal+0xf8
0e 00000024`2773fc00 00007ffd`96d7d6ea clr!CorExeMain+0x14
0f 00000024`2773fc40 00007ffd`96c3ac42 mscoreei!CorExeMain+0xfa
10 00000024`2773fca0 00007ffd`b5a07374 mscoree!CorExeMain_Exported+0x72
11 00000024`2773fcd0 00007ffd`b5ebcc91 kernel32!BaseThreadInitThunk+0x14
12 00000024`2773fd00 00000000`00000000 ntdll!RtlUserThreadStart+0x21
paddle_inference_c!paddle::CreatePaddlePredictor<paddle::AnalysisConfig,3>+0x180be:
00007ffc`d39a1a9e 488b01 mov rax,qword ptr [rcx] ds:00000217`89df37b0=00007ffcd7791ff0
# Child-SP RetAddr Call Site
00 00000024`27eff080 00007ffc`d39a1a8f paddle_inference_c!paddle::CreatePaddlePredictor<paddle::AnalysisConfig,3>+0x180be
01 00000024`27eff0b0 00007ffc`d39a1a8f paddle_inference_c!paddle::CreatePaddlePredictor<paddle::AnalysisConfig,3>+0x180af
02 00000024`27eff0e0 00007ffc`d39a1a8f paddle_inference_c!paddle::CreatePaddlePredictor<paddle::AnalysisConfig,3>+0x180af
...
0e 00000024`27eff3b0 00007ffc`f20cbbde paddle_inference_c!PD_PredictorDestroy+0x39
0f 00000024`27eff3e0 00007ffc`f943796a 0x00007ffc`f20cbbde
10 00000024`27eff4b0 00007ffc`f9437930 Sdcb_PaddleInference!Sdcb.PaddleInference.PaddlePredictor.Dispose+0x1a
11 00000024`27eff4e0 00007ffd`518915e6 Sdcb_PaddleInference!Sdcb.PaddleInference.PaddlePredictor.Finalize+0x10
12 00000024`27eff510 00007ffd`5173b6c6 clr!FastCallFinalizeWorker+0x6
13 00000024`27eff540 00007ffd`5173bd73 clr!FastCallFinalize+0x5a
14 00000024`27eff580 00007ffd`5173bcae clr!MethodTable::CallFinalizer+0xb7
15 00000024`27eff5c0 00007ffd`5173b8e7 clr!CallFinalizer+0x5e
...
1e 00000024`27effa50 00000000`00000000 ntdll!RtlUserThreadStart+0x21
其实到这里逻辑全部清楚了，用户点击了关闭程序，但此时终结器线程正在执行，所以执行引擎需要等待它执行完，在这期间是禁止除终结器线程之外的任何线程再对托管方法的调用，一旦有就会抛出异常，画个简图如下：
3. call托管方法在哪里
在错误的时间出现了一个对的人，这是一种意难平，接下来把这个对的人给找出来，即 UMThunkStubRareDisableWorker方法的第二个参数pUMEntryThunk，因为这里面的 m_pManagedTarget 即托管方法的入口点，模型如下：
class UMEntryThunk
{
private:
// The start of the managed code
const BYTE* m_pManagedTarget;
// This is used for profiling.
PTR_MethodDesc m_pMD;
}
那如何挖呢？观察UMThunkStubRareDisableWorker方法汇编代码即可。
0:044> k
*** Stack trace for last set context - .thread/.cxr resets it
# Child-SP RetAddr Call Site
00 00000024`2c46fb40 00007ffd`51bf8c6f KERNELBASE!RaiseException+0x69
01 00000024`2c46fc20 00007ffd`51890108 clr!UMThunkStubRareDisableWorker+0x3f
02 00000024`2c46fc50 00007ffc`dc826748 clr!UMThunkStub+0x128
03 00000024`2c46fce0 00007ffc`dc827406 MVCAMSDK_X64+0x6748
04 00000024`2c46fd80 00007ffc`dc9d21eb MVCAMSDK_X64!CameraGrabber_GetCameraDevInfo+0xb96
05 00000024`2c46fe50 00007ffc`dc9d227f MVCAMSDK_X64!CameraGigeEnumerateDevice+0x167f2b
06 00000024`2c46fe80 00007ffd`b5a07374 MVCAMSDK_X64!CameraGigeEnumerateDevice+0x167fbf
07 00000024`2c46feb0 00007ffd`b5ebcc91 kernel32!BaseThreadInitThunk+0x14
08 00000024`2c46fee0 00000000`00000000 ntdll!RtlUserThreadStart+0x21
0:044> uf 00007ffd`51890108
00007ffd`5188ffe0 4154 push r12
00007ffd`5188ffe2 55 push rbp
00007ffd`5188ffe3 4883ec78 sub rsp,78h
00007ffd`5188ffe7 488d6c2420 lea rbp,[rsp+20h]
00007ffd`5188ffec c6454000 mov byte ptr [rbp+40h],0
00007ffd`5188fff0 e8abe0ffff call clr!GetThread (00007ffd`5188e0a0)
00007ffd`5188fff5 4885c0 test rax,rax
00007ffd`5188fff8 746f je clr!UMThunkStub+0x89 (00007ffd`51890069) Branch
...
00007ffd`518900f9 4c895548 mov qword ptr [rbp+48h],r10
00007ffd`518900fd 498bcc mov rcx,r12
00007ffd`51890100 498bd2 mov rdx,r10
00007ffd`51890103 e8288b3600 call clr!UMThunkStubRareDisableWorker (00007ffd`51bf8c30)
根据x64调用协定，参数是放在 r10 寄存器中，而r10刚好放在 rbp+48h 中，所以挖出这个栈位置即可，简单计算为。
0:044> ? 00000024`2c46fce0-0x8-0x8-0x8-0x78+0x20
Evaluate expression: 155361672304 = 00000024`2c46fc70
0:044> dp 00000024`2c46fc70+0x48 L1
00000024`2c46fcb8 00000217`ff1594a0
0:044> dp 00000217`ff1594a0 L1
00000217`ff1594a0 00007ffc`f2f49380
0:044> !U 00007ffc`f2f49380
Unmanaged code
00007ffc`f2f49380 e98bbd0500 jmp 00007ffc`f2fa5110
00007ffc`f2f49385 5f pop rdi
00007ffc`f2f49386 0f ???
00007ffc`f2f49387 25e8e35194 and eax,9451E3E8h
00007ffc`f2f4938c 5e pop rsi
00007ffc`f2f4938d 5e pop rsi
00007ffc`f2f4938e 1124e8 adc dword ptr [rax+rbp*8],esp
00007ffc`f2f49391 db5194 fist dword ptr [rcx-6Ch]
00007ffc`f2f49394 5e pop rsi
00007ffc`f2f49395 5e pop rsi
0:044> !ip2md 00007ffc`f2fa5110
MethodDesc: 00007ffcf33bd9e8
Method Name: baslerCamera.MVCam.CameraGrabberFrameCallback(IntPtr, IntPtr, baslerCamera.tSdkFrameHead ByRef, IntPtr)
Class: 00007ffcf33c1c98
MethodTable: 00007ffcf33be828
mdToken: 0000000006000067
Module: 00007ffcf2352f60
IsJitted: yes
CodeAddr: 00007ffcf2fa5110
Transparency: Critical
从 baslerCamera 来看是一个工业相机，接下来赶紧寻找源代码，截图如下：
图中的 m_FrameCallback 是一个delegate函数，并通过 CameraGrabber_SetRGBCallback 方法传给了C++，所以在关闭程序的时候一定要先释放掉这个 calllback 来和C++撇清关系，否则就会有灾难的降临。
三：总结
在我的dump分析之旅中曾遇到过一次相似的生产故障，本篇文章主要还是对训练营里这位朋友的有求必应吧，不过说实话，多分析几个像这样的dump，会极大的提升你的高级调试能力。
