Uh oh!
There was an error while loading. Please reload this page.
[release/10.0] Move coreclr EH second pass to native code - #120263
Conversation
There was a problem hiding this comment.
Pull Request Overview
This PR moves the CoreCLR exception handling second pass from managed C# code to native C++ code to fix GC holes that occurred when the garbage collector was triggered between finally handler and catch handler calls during the second pass. This change addresses a customer-reported access violation crash in RavenDB that occurred intermittently due to these GC holes.
Key changes:
- Remove managed second pass dispatch code and move logic to native implementation
- Eliminate complex GC reference tracking infrastructure that was needed to handle GC during managed EH code
- Simplify stack walking by removing special handling for managed exception handling frames
Reviewed Changes
Copilot reviewed 15 out of 15 changed files in this pull request and generated 3 comments.
Show a summary per file
| File | Description |
|---|---|
| src/coreclr/vm/stackwalk.h | Remove GC reference tracking fields and methods for funclet handling |
| src/coreclr/vm/stackwalk.cpp | Remove complex GC reference tracking logic during stack walking |
| src/coreclr/vm/qcallentrypoints.cpp | Remove QCall entry points for managed funclet calls |
| src/coreclr/vm/metasig.h | Remove method signature for managed unwind and intercept |
| src/coreclr/vm/gcenv.ee.common.cpp | Remove saved funclet info GC reference reporting |
| src/coreclr/vm/exinfo.h | Add new fields for catch handler info, remove saved funclet tracking |
| src/coreclr/vm/exinfo.cpp | Remove saved funclet initialization |
| src/coreclr/vm/exceptionhandlingqcalls.h | Remove QCall declarations for managed funclet calls |
| src/coreclr/vm/exceptionhandling.h | Remove second pass funclet caller marker |
| src/coreclr/vm/exceptionhandling.cpp | Add native second pass implementation, convert funclet callers to native |
| src/coreclr/vm/excep.cpp | Add second pass calls and remove old intercept unwind code |
| src/coreclr/vm/corelib.h | Remove managed unwind and intercept method definition |
| src/coreclr/nativeaot/Runtime.Base/src/System/Runtime/ExceptionHandling.cs | Add conditional compilation for CoreCLR vs NativeAOT paths |
| src/coreclr/System.Private.CoreLib/src/System/Runtime/ExceptionServices/InternalCalls.cs | Remove managed funclet call declarations |
| src/coreclr/System.Private.CoreLib/src/System/Runtime/ExceptionServices/AsmOffsets.cs | Update structure sizes and add new ExInfo field offsets |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
b8ea280 to
bf4fa55CompareThere were some GC holes discovered caused by the fact that GC can be triggered during 2nd pass of EH in-between calls to finally handlers and catch handler. After considering options, moving the 2nd pass to native code seems the most reasonable solution.
bf4fa55 to
7fc9c86Comparejanvorli
commented
Oct 1, 2025
The interpreter test is failing because it would need another PR from main to work correctly with this changed EH. The reason is that resume after catch in the .NET 10 branch expects that the RhThrowEx and the like are on the stack. I've simplified / fixed that in main. |
c666bc0 to
288916dCompare
Backport of #119863 to release/10.0
Customer Impact
There were some GC holes discovered caused by the fact that GC can be triggered during 2nd pass of EH in-between calls to finally handlers and catch handler. After considering options, moving the 2nd pass to native code seems the most reasonable solution.
The issue was reported by the RavenDB folks in #110769. It causes access violation crash on their system on Windows once in one or two days on average.
Regression
Introduced in .NET 8 by #88034 - the new exception handling, but got reported in .NET 9 since that's where the new EH was enabled by default.
It took almost 9 months to figure out the culprit due to the very narrow window in which it can occur and due to the fact that GC holes in general don't necessarily result in an observed bad behavior. That's why it was fixed this late in the release cycle.
Testing
Risk
Low, the change is 1:1 port of the 2nd pass of the EH from C# to C++ and then removal of a number of extra handling in the stack frame iterator and GC reference enumeration that was only necessary to support the cases when GC could have been happening in-between calls to finallys. That's no longer possible with the fix.