Fix a deadlock in NonGC + Profiler API #90847
Conversation
ghost
assigned 
|
/azp run runtime-coreclr outerloop |
|
Azure Pipelines successfully started running 1 pipeline(s). |
EgorBo
force-pushed
the
fix-deadlock-nongc
branch
from
64b9fd0 to
cec6257
Compare
|
/azp run runtime-coreclr outerloop |
|
Azure Pipelines successfully started running 1 pipeline(s). |
Co-authored-by: Jan Kotas <[email protected]>
|
/azp run runtime-coreclr outerloop |
|
Azure Pipelines successfully started running 1 pipeline(s). |
|
/azp run runtime-coreclr outerloop |
|
Azure Pipelines successfully started running 1 pipeline(s). |
Co-authored-by: Jan Kotas <[email protected]>
src/coreclr/vm/frozenobjectheap.cpp
Outdated
| @@ -221,11 +266,11 @@ Object* FrozenObjectSegment::GetNextObject(Object* obj) const | |||
| { | |||
| // Input must not be null and should be within the segment | |||
| _ASSERT(obj != nullptr); | |||
| _ASSERT((uint8_t*)obj >= m_pStart + sizeof(ObjHeader) && (uint8_t*)obj < m_pCurrent); | |||
| _ASSERT((uint8_t*)obj >= m_pStart + sizeof(ObjHeader) && (uint8_t*)obj < m_pCurrentRegistered); | |||
|
|
|||
| // FOH doesn't support objects with non-DATA_ALIGNMENT alignment yet. | |||
There was a problem hiding this comment.
Do we need to set m_NumComponents for arrays as part of TryAllocateObject?
We are setting it too late and we can end up enumerating arrays without m_NumComponents set that is not going to end wel..
There was a problem hiding this comment.
Good point! It also allowed to simplify the PublishObject logic a bit. The final API might be simplified a bit with C++ template to allow use of capturing lambdas for simplicity but that needed a bit more changes
|
/azp run runtime-coreclr outerloop |
|
Azure Pipelines successfully started running 1 pipeline(s). |
Thanks for the help! I wish I could easily spot all possible race conditions/corner cases just like you 🙂 |
Co-authored-by: Jan Kotas <[email protected]>
|
/backport to release/8.0 |
|
Started backporting to release/8.0: https://github.com/dotnet/runtime/actions/runs/5979164788 |
Fixes #90830
Quick explanation how's the dead-lock happening:
Thread1:
Someone (typically, JIT) tries to allocate an object on NonGC heap.
FrozenObjectHeapManager(FOHM) acquires its lock and calls GC's APIRegisterNewSegment. That API internally can hit a case when a GC is happening so it has to wait for GC to complete.Thread2 (GC's):
GC is executing a callback (e.g.
GarbageCollectionFinishedor*Started) and Profiler uses that callback to enumerate objects on NonGC heap viaICorProfilerInfo14::GetNonGCHeapBoundsthus, it also tries to acquire FOHM's lock (to be able to safely enumerate the objects). Thus, GC's thread (Thread2) is wating for FOHM's lock to release (it's taken by Thread1) while Thread1 is waiting for GC to finish.The fix is #90830 (comment)