Possible .NET JIT call parameter lifetime bug?

19

I've been chasing down the cause of an intermittent crash in one of our .NET services due to an internal error in the .NET Runtime (exit code 0x80131506). The service in question doesn't perform any of the kinds of operation that are usually to blame for such errors (unsafe code, PInvoke, etc.). I've tried disabling concurrent GC as described in KB2679415, as well as switching to server GC, but the intermittent crashes persist. The issue manifests on .NET 4.7.2 and earlier versions, when compiled in debug mode.

The service makes extensive use of an old version of NHibernate (2.0.1), and when I've examined crash dumps in the debugger, there is always NHibernate code in the callstack when the error occurs, though NHibernate itself is all managed code, so should not be capable of causing this kind of crash.

I have managed to reproduce the crash under a debugger and with the GC Stress Log and Heap Verification enabled, and whilst it seems to point to an issue within the JIT/GC, I'm not sure I'm interpreting the output correctly.

Looking at the thread on which the crash occurs, on this occasion it's occurring at clr!JIT_Stelem_Ref:

clr!JIT_Stelem_Ref+0x18: cmp     r9,qword ptr [r8] ds:aaaaaaaa`aaaaaaaa=????????????????

In this case the string of 0xaas appears to be the result of having HeapVerify enabled, which causes the GC fill collected memory regions, presumably for easier identification, and suggests that somehow we still have a reference to the old location of a collected/relocated object.

Tracking back in the stack, there are plenty of 0xaaaaaaaaaaaaaaaa entries, however these cease appearing at the method that was at the top of the call stack when the most recent GC occurred, which in this case was NHibernate.Loader.Loader.GetRow() according to the GC stress log for the most recent GC on this thread:

(Note: I've reversed the order of the logged lines from SOS' !dumplog output for easier reading):

2404 12445.672380360 : `GC`GCROOTS`         Starting scan of Thread 000000001EF4DED0 ID = 20 {
2404 12445.672380963 : `GCROOTS`            Scanning ExplicitFrame 000000001E6ED3B8 AssocMethod = 0000000000000000 frameVTable = 000007FEF365B640 (clr!RedirectedThreadFrame::`vftable')
2404 12445.672386397 : `GCROOTS`            Scanning Frameless method 000007FE93F43460 (NHibernate.Loader.Loader.GetRow(System.Data.IDataReader, NHibernate.Persister.Entity.ILoadable[], NHibernate.Engine.EntityKey[], System.Object, NHibernate.Engine.EntityKey, NHibernate.LockMode[], System.Collections.IList, NHibernate.Engine.ISessionImplementor)) ControlPC = 000007FE945E3095
2404 12445.672388208 : `GC`GCROOTS`             GC Root 000000001E6ED4C0 RELOCATED 000000003B1A7708 -> 000000003AC89F08  MT = 000007FE93DDF5C8 (...)
2404 12445.672388510 : `GC`GCROOTS`             GC Root 000000001E6ED4D8 RELOCATED 000000003B1A73A0 -> 000000003AC89D00  MT = 000007FEF1FD6EA8 (System.Object[])
2404 12445.672388510 : `GC`GCROOTS`             GC Root 000000001E6ED4E8 RELOCATED 000000003B1A7358 -> 000000003AC89CB8  MT = 000007FE9491D7C8 (NHibernate.Engine.EntityKey)
2404 12445.672388510 : `GC`GCROOTS`             GC Root 000000001E6ED4F8 RELOCATED 000000003B1A73A0 -> 000000003AC89D00  MT = 000007FEF1FD6EA8 (System.Object[])

The stack area for this method is as follows:

00000000`1e6ed470 000000003b1a7358 ✕
00000000`1e6ed478 000000000291e3d0 
00000000`1e6ed480 0000000000000000 
00000000`1e6ed488 0000000000000000 
00000000`1e6ed490 000000000662a900 
00000000`1e6ed498 0000000006523c80 
00000000`1e6ed4a0 0000000000000000 
00000000`1e6ed4a8 0000000000000000 
00000000`1e6ed4b0 0000000000000000 
00000000`1e6ed4b8 0000000000000000 
00000000`1e6ed4c0 000000003ac89f08 ✔
00000000`1e6ed4c8 0000000000000000 
00000000`1e6ed4d0 0000000006524248 
00000000`1e6ed4d8 000000003ac89d00 ✔
00000000`1e6ed4e0 0000000000000000 
00000000`1e6ed4e8 000000003ac89cb8 ✔
00000000`1e6ed4f0 0000000000000000 
00000000`1e6ed4f8 000000003ac89d00 ✔
00000000`1e6ed500 0000000100000000 
00000000`1e6ed508 0000000c0000000b 
00000000`1e6ed510 0000000006621660 
00000000`1e6ed518 000000001e6ed690 
00000000`1e6ed520 000000001e6ed6a0

I have indicated the 4 entries mentioned in the GC stress log as relocated, which have been correctly updated with their new addresses, however the first stack entry (000000003b1a7358 - an NHibernate.Engine.EntityKey), whilst it is one of the relocated objects has not been updated with the new address. This would of course be completely normal if this was no longer going to be used, however it is in fact about to be passed as a parameter to a call to NHibernate.Loader.Loader.InstanceNotYetLoaded().

InstanceNotYetLoaded() takes 9 parameters (plus this), and I have marked where each of these is loaded into the stack/register on the following assembly listing. I have also included the relevant output from SOS' !gcinfo as it relates to each of the parameters on the stack:

Param Address              Instruction                        GC Info
      000007fe`945e3071    mov     r9,qword ptr [rbp-38h]
  P4> 000007fe`945e3075    mov     qword ptr [rsp+20h],r9
      000007fe`945e307a    mov     r9d,dword ptr [rbp-18h]    +sp+20
      000007fe`945e307e    mov     rcx,qword ptr [rbp+40h]
      000007fe`945e3082    cmp     r9,qword ptr [rcx+8]
      000007fe`945e3086    jb      000007fe`945e308d
      000007fe`945e3088    call    clr!JIT_RngChkFail
      000007fe`945e308d    lea     rcx,[rcx+r9*8+10h]         -sp+20
      000007fe`945e3092    mov     r9,qword ptr [rcx]
-- GC Occurred Here --
  P5> 000007fe`945e3095    mov     qword ptr [rsp+28h],r9
      000007fe`945e309a    mov     r9,qword ptr [rbp+38h]     +sp+28
  P6> 000007fe`945e309e    mov     qword ptr [rsp+30h],r9
      000007fe`945e30a3    mov     r9,qword ptr [rbp+30h]     +sp+30
  P7> 000007fe`945e30a7    mov     qword ptr [rsp+38h],r9
      000007fe`945e30ac    mov     r9,qword ptr [rbp+48h]     +sp+38
  P8> 000007fe`945e30b0    mov     qword ptr [rsp+40h],r9
      000007fe`945e30b5    mov     r9,qword ptr [rbp+50h]     +sp+40
  P9> 000007fe`945e30b9    mov     qword ptr [rsp+48h],r9
      000007fe`945e30be    mov     r9d,dword ptr [rbp-18h]    +sp+48
      000007fe`945e30c2    mov     rcx,qword ptr [rbp+20h]
      000007fe`945e30c6    cmp     r9,qword ptr [rcx+8]
      000007fe`945e30ca    jb      000007fe`945e30d1
      000007fe`945e30cc    call    clr!JIT_RngChkFail
      000007fe`945e30d1    lea     rcx,[rcx+r9*8+10h]         -sp+48 -sp+40 -sp+38 -sp+30 -sp+28
  P3> 000007fe`945e30d6    mov     r9,qword ptr [rcx]
this> 000007fe`945e30d9    mov     rcx,qword ptr [rbp+10h]
  P1> 000007fe`945e30dd    mov     rdx,qword ptr [rbp+18h]
  P2> 000007fe`945e30e1    mov     r8d,dword ptr [rbp-18h]
      000007fe`945e30e5    call    InstanceNotYetLoaded(...)

The GC just prior to the crash occurred at 000007fe945e3095, which is after parameter 4 is loaded onto the stack (at 000007fe945e3075), but also after this stack entry has become dead (at 000007fe945e308d) according to the GC Info, which would explain why the GC relocate phase didn't update this reference.

It also looks like the GC Info for parameters 5-9 is also incorrectly marking them as dead too early, and perhaps tellingly in both cases they are being marked dead immediately after what looks like an array index range check.

This to me looks like a JIT bug with the lifetimes of these stack parameters being incorrectly tracked. Is this analysis correct, if so where best to report if. If it's not a JIT bug, what am I missing that could explain these unexpected crashes on purely managed code?

Edit:

I believe the following snippet will reproduce the issue, at least as far as generating the bad GC info in debug mode.

public void Repro(int p1, object p2, object p3, object p4, object[] p5)
{
    // Incorrect GC Info generated for this call
    ReproHelper(p1, p2, p3, p4, p5[p1]);
}

public void ReproHelper(int p1, object p2, object p3, object p4, object p5)
{
    Console.WriteLine(p1);
    Console.WriteLine(p2);
    Console.WriteLine(p3);
    Console.WriteLine(p4);
    Console.WriteLine(p5);
}

In essence, there must be a method call to a method that:

  • Requires at least 2 parameters to be passed on the stack (i.e. at least 5 parameters for an instance method).
  • The 2nd parameter passed on the stack (parameter 5) must be the result of an array access.

When these criteria are met, the 4th parameter is loaded onto the stack for the call, and the stack entry is correctly marked as containing a reference. However in determining the value for parameter 5, an array index range check is made, and after this occurs the stack entry for parameter 4 is marked as dead.

If a GC occurs after the range check but before the actual call takes place, and the GC results in the object that was being passed as parameter 4 being relocated, when the method resumes, the call will pass the old (invalid) address to parameter 4, rather than the new one.

c#
.net
garbage-collection
jit
asked on Stack Overflow Aug 23, 2018 by Iridium • edited Aug 24, 2018 by Iridium

2 Answers

5

While it is not an answer to the problem, I treat it as such because I confirm it is a bug and should be fixed by the .NET team.

When running your snippet on .NET Framework 4.7.1 (clrjit.dll version 4.7.2xxx), proper GCInfo was generated (and +sp+20 is in fact written to only just before ReproHelper call):

00007ffb`99450630 55              push    rbp
00007ffb`99450631 4883ec40        sub     rsp,40h
00000003 is a safepoint: 
00007ffb`99450635 488d6c2440      lea     rbp,[rsp+40h]
00007ffb`9945063a 33c0            xor     eax,eax
00007ffb`9945063c 488945f8        mov     qword ptr [rbp-8],rax
00007ffb`99450640 48894d10        mov     qword ptr [rbp+10h],rcx
00007ffb`99450644 895518          mov     dword ptr [rbp+18h],edx
00007ffb`99450647 4c894520        mov     qword ptr [rbp+20h],r8
00007ffb`9945064b 4c894d28        mov     qword ptr [rbp+28h],r9
interruptible
+rbp+28 +rbp+20 +rbp+10 +rbp-8
00007ffb`9945064f 833d3a3fefff00  cmp     dword ptr [00007ffb`99344590],0
00007ffb`99450656 7405            je      00007ffb`9945065d
00007ffb`99450658 e863eaab5f      call    clr!JIT_DbgIsJustMyCode (00007ffb`f8f0f0c0)
00007ffb`9945065d 90              nop
00007ffb`9945065e 8b5518          mov     edx,dword ptr [rbp+18h]
00007ffb`99450661 4c8b4538        mov     r8,qword ptr [rbp+38h]
+r8
00007ffb`99450665 413b5008        cmp     edx,dword ptr [r8+8]
00007ffb`99450669 7205            jb      00007ffb`99450670
-rbp-8
00007ffb`9945066b e8f015ac5f      call    clr!JIT_RngChkFail (00007ffb`f8f11c60)
-r8
00007ffb`99450670 488b5538        mov     rdx,qword ptr [rbp+38h]
+rdx
00007ffb`99450674 448b4518        mov     r8d,dword ptr [rbp+18h]
00007ffb`99450678 4d63c0          movsxd  r8,r8d
00007ffb`9945067b 4a8b54c210      mov     rdx,qword ptr [rdx+r8*8+10h]
00007ffb`99450680 488955f8        mov     qword ptr [rbp-8],rdx
+rbp-8
00007ffb`99450684 488b55f8        mov     rdx,qword ptr [rbp-8]
00007ffb`99450688 4889542428      mov     qword ptr [rsp+28h],rdx
+sp+28
00007ffb`9945068d 8b5518          mov     edx,dword ptr [rbp+18h]
-rdx
00007ffb`99450690 4c8b4520        mov     r8,qword ptr [rbp+20h]
+r8
00007ffb`99450694 4c8b4d28        mov     r9,qword ptr [rbp+28h]
+r9
00007ffb`99450698 488b4d30        mov     rcx,qword ptr [rbp+30h]
+rcx
00007ffb`9945069c 48894c2420      mov     qword ptr [rsp+20h],rcx
+sp+20
00007ffb`994506a1 488b4d10        mov     rcx,qword ptr [rbp+10h]
-rbp-8

But after upgrading to .NET Framework 4.7.2 (clrjit.dll version 4.7.3062), it is no longer correct (+sp+20 is written to before array index range check, properly set but out of a sudden unset afterwards, while still used in the ReproHelper call):

00007ffe`62290630 55              push    rbp
00007ffe`62290631 4883ec30        sub     rsp,30h
00007ffe`62290635 488d6c2430      lea     rbp,[rsp+30h]
00000007 is a safepoint: 
00007ffe`6229063a 48894d10        mov     qword ptr [rbp+10h],rcx
00007ffe`6229063e 895518          mov     dword ptr [rbp+18h],edx
00007ffe`62290641 4c894520        mov     qword ptr [rbp+20h],r8
00007ffe`62290645 4c894d28        mov     qword ptr [rbp+28h],r9
interruptible
+rbp+28 +rbp+20 +rbp+10
00007ffe`62290649 833d483fefff00  cmp     dword ptr [00007ffe`62184598],0
00007ffe`62290650 7405            je      00007ffe`62290657
00007ffe`62290652 e869f7aa5f      call    clr!TranslateSecurityAttributes+0x857b0 (00007ffe`c1d3fdc0) (JitHelp: CORINFO_HELP_DBG_IS_JUST_MY_CODE)
00007ffe`62290657 90              nop
00007ffe`62290658 488b4d30        mov     rcx,qword ptr [rbp+30h]    
+rcx
00007ffe`6229065c 48894c2420      mov     qword ptr [rsp+20h],rcx
+sp+20
00007ffe`62290661 8b4d18          mov     ecx,dword ptr [rbp+18h]
-rcx
00007ffe`62290664 488b5538        mov     rdx,qword ptr [rbp+38h]
+rdx
00007ffe`62290668 483b4a08        cmp     rcx,qword ptr [rdx+8]
00007ffe`6229066c 7205            jb      00007ffe`62290673
00007ffe`6229066e e8ed22ab5f      call    clr!TranslateSecurityAttributes+0x88350 (00007ffe`c1d42960) (JitHelp: CORINFO_HELP_RNGCHKFAIL)
-sp+20
00007ffe`62290673 488d54ca10      lea     rdx,[rdx+rcx*8+10h]
-rdx +rdx(interior)
00007ffe`62290678 488b0a          mov     rcx,qword ptr [rdx]
+rcx
00007ffe`6229067b 48894c2428      mov     qword ptr [rsp+28h],rcx
+sp+28
00007ffe`62290680 488b4d10        mov     rcx,qword ptr [rbp+10h]
00007ffe`62290684 8b5518          mov     edx,dword ptr [rbp+18h]
-rdx(interior)
00007ffe`62290687 4c8b4520        mov     r8,qword ptr [rbp+20h]
+r8
00007ffe`6229068b 4c8b4d28        mov     r9,qword ptr [rbp+28h]
+r9
00007ffe`6229068f e804faffff      call    00007ffe`62290098 (GCInfoBug.Bug.ReproHelper(Int32, System.Object, System.Object, System.Object, System.Object), mdToken: 0000000006000004)
-sp+28 -r9 -r8 -rcx
00007ffe`62290694 90              nop
00007ffe`62290695 90              nop
not interruptible
-rbp+28 -rbp+20 -rbp+10
00007ffe`62290696 488d6500        lea     rsp,[rbp]
00007ffe`6229069a 5d              pop     rbp
00007ffe`6229069b c3              ret
answered on Stack Overflow Aug 24, 2018 by Konrad Kokosa
3

Thanks you for providing such a detailed investigation along with this bug report.

Microsoft will be fixing this issue in 4.7.2 an upcoming servicing release.

answered on Stack Overflow Nov 26, 2018 by Brian Sullivan • edited Nov 26, 2018 by Brian Sullivan

User contributions licensed under CC BY-SA 3.0