[llvm-dev] [ORC JIT][MLIR] GDBRegistrationListener "second attempt to perform debug registration" assert

[Straw, Adam D via llvm-dev]

Hi all,
Attention:  Lang Hames

I am developing the nGraph MLIR<https://github.com/NervanaSystems/ngraph/tree/master/src/contrib/mlir> implementation and hitting the following assert while running nGraph unit tests:

  assert(ObjectBufferMap.find(K) == ObjectBufferMap.end() &&
         "Second attempt to perform debug registration.");

Here is a permalink<https://github.com/llvm/llvm-project/blob/3d5360a4398bfa6878f94ca9ac55bc568692c765/llvm/lib/ExecutionEngine/GDBRegistrationListener.cpp#L163> for that line of code on GitHub.  The failure only occurs when running multiple unit tests back-to-back.  Inevitably, if I rerun the failing unit test, it passes.  The failure also tends to move around with a different unit tests failing on successive runs.

I am able to hit the failure in GDB.  Here is a partial backtrace:

#3  0x00007fffec4cf412 in __GI___assert_fail (assertion=assertion at entry=0x7ffff1d145e0 "ObjectBufferMap.find(K) == ObjectBufferMap.end() && \"Second attempt to perform debug registration.\"",
    file=file at entry=0x7ffff1d144d8 "/localdisk/adstraw/ngraph/build/mlir_project/llvm-project/llvm/lib/ExecutionEngine/GDBRegistrationListener.cpp", line=line at entry=164,
    function=function at entry=0x7ffff1d15520 <(anonymous namespace)::GDBJITRegistrationListener::notifyObjectLoaded(unsigned long, llvm::object::ObjectFile const&, llvm::RuntimeDyld::LoadedObjectInfo const&)::__PRETTY_FUNCTION__> "virtual void {anonymous}::GDBJITRegistrationListener::notifyObjectLoaded(llvm::JITEventListener::ObjectKey, const llvm::object::ObjectFile&, const llvm::RuntimeDyld::LoadedObjectInfo&)")
    at assert.c:101
#4  0x00007ffff01ed4d3 in (anonymous namespace)::GDBJITRegistrationListener::notifyObjectLoaded (this=0x5555591a9270, K=93825060893584, Obj=..., L=...)
    at /localdisk/adstraw/ngraph/build/mlir_project/llvm-project/llvm/lib/ExecutionEngine/GDBRegistrationListener.cpp:163
#5  0x00007ffff01b0eca in std::function<void (unsigned long, llvm::object::ObjectFile const&, llvm::RuntimeDyld::LoadedObjectInfo const&)>::operator()(unsigned long, llvm::object::ObjectFile const&, llvm::RuntimeDyld::LoadedObjectInfo const&) const (__args#2=..., __args#1=..., __args#0=<optimized out>, this=0x5555594e09b8) at /usr/include/c++/7/bits/std_function.h:706
#6  llvm::orc::RTDyldObjectLinkingLayer::onObjLoad (this=<optimized out>, K=3, R=..., Obj=..., MemMgr=0x5555591c4160, LoadedObjInfo=std::unique_ptr<llvm::RuntimeDyld::LoadedObjectInfo> = {...},
    Resolved=std::map with 8 elements = {...}, InternalSymbols=std::set with 0 elements)
    at /localdisk/adstraw/ngraph/build/mlir_project/llvm-project/llvm/lib/ExecutionEngine/Orc/RTDyldObjectLinkingLayer.cpp:267

A little debug context:


  *   Frame #6 with K=3 we are about to execute the `NotifyLoaded` callback
  *   Frame #5 we execute the `NotifyLoaded` callback which apparently transforms K from 3 to 93825060893584 (0x5555596CF390) which looks like a pointer.  I don't know the code well enough to grok what's going on here.  I'm not even sure where this callback resides, but it could be here<https://github.com/llvm/llvm-project/blob/c66f89005f6d23b6885d8f93f33ff27dc60ce7dd/llvm/lib/ExecutionEngine/Orc/RTDyldObjectLinkingLayer.cpp#L312> in RTDyldObjectLinkingLater.cpp where we are using the address of the `MemMgr` as K.
  *   Frame #4 now in `notifyObjectLoaded` with K=93825060893584 which will go on to assert in frame #3 due to the fact that that key (K) already exists in the map

All of this leads me to believe there is some sort of race.  Theory:  with two unit tests A and B running back to back...


  1.  Unit test A allocates and deallocates a `MemMgr` (or whatever other object we are using for a key) at address X
  2.  Unit test A calls ` notifyFreeingObject` to free the object at K=X but gets stuck (does not lock) the ` JITDebugLock` mutex here<https://github.com/llvm/llvm-project/blob/3d5360a4398bfa6878f94ca9ac55bc568692c765/llvm/lib/ExecutionEngine/GDBRegistrationListener.cpp#L181> for whatever reason
  3.  Meanwhile unit test B allocates a `MemMgr` (or whatever other object we are using for a key) at newly freed address X
  4.  Unit test B calls `notifyObjectLoaded` to register the object at K=X and locks the ` JITDebugLock` here<https://github.com/llvm/llvm-project/blob/3d5360a4398bfa6878f94ca9ac55bc568692c765/llvm/lib/ExecutionEngine/GDBRegistrationListener.cpp#L162> with unit test A still waiting in step #2 above
  5.  This goes on to assert as the key was not erased (step #2) before it was added (step #4)

I could use a little help to debug the error further.  Curiously, we do hit this issue in continuous integration (CI) --- only when running on development systems "by hand".  I am investigating the differences (perhaps single vs. multi threaded?) as to why this might be the case.

Note that I am using a slightly old version of MLIR master, I'm at this commit:

commit 2f8b4545f4960778e37114c024073d208751ca89
Author: Adam Straw <adam.d.straw at intel.com<mailto:adam.d.straw at intel.com>>
Date:   Tue Apr 14 22:49:18 2020 +0300

    [mlir] Fix assert on signed integer type in EDSC

    Integer type in Std dialect is signless so we should be checking
    for signless integer type instead of signed integer type in EDSC.

    Differential Revision: https://reviews.llvm.org/D78144

Thanks in advance for the help.

Adam

Note:  I am out of office this Fri and the following Mon but otherwise should be prompt with email replies.


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20200520/1799c9a2/attachment.html>