[release/9.0] Fix edge cases in Tarjan GC bridge (Android)

[filipnavara]

Backport of #112825, #112970, #113044 and #113703 to release/9.0-staging

/cc @BrzVlad @vitek-karas

Customer Impact

• Customer reported
• Found internally

Two customer issues on Android were traced back to bugs in the MonoVM Tarjan GC bridge responsible for bridging .NET and Java garbage collectors.

In the first issue (#115611) the runtime failed to report edges to the Android GC bridge code, resulting in alive objects collected in the Java runtime and their .NET peers throwing ObjectDisposedException. The particular pattern could have been triggered by code using await on the UI thread where the continuation was scheduled to run through synchronization context on the same thread.

In the second issue (#106410) the runtime crashed on an assertion for object graphs of certain shape. The cause of the crash was an incorrect assumption about duplicate edges in a compressed object graph and how later stages of the algorithm process them. When the later stages of the processing deduplicated the edges it may have resulted in an unforseen change to some of the graph properties leading to a logic error in the algorithm.

Regression

• Yes
• No

Both of these bugs existed in .NET MonoVM / Android code since the inception. Legacy Xamarin.Android offered 3 different bridge algorithms (old, new, and tarjan) with the new bridge being the default for a while. The default was changed to tarjan before MonoVM was merged into .NET.

Testing

Test cases reproducing both issues were presented and verified against the fix. Additionally, a new testing code was introduced to specifically test the problematic patterns as part of .NET runtime testing infrastructure.

Risk

Medium - the changes simply make reporting the graph more accurate. Even though we have high confidence in the changes, we recognize it does touch a sensitive area.

[jeffschwMSFT]

lgtm. we will take for consideration in 9.0.x

[akoeplinger]

@filipnavara there's a failure in the new test on wasm:

    GC/Features/Bridge/BridgeTester/BridgeTester.sh [FAIL]
      
      Return code:      1
      Raw output file:      /root/helix/work/workitem/uploads/Reports/GC.Features/Bridge/BridgeTester/BridgeTester.output.txt
      Raw output:
      BEGIN EXECUTION
      MSBuild version 17.12.27+e0b90a9a8 for .NET
        AppDir: /root/helix/work/workitem/e/GC/Features/Bridge/BridgeTester/WasmApp/
        TestBinDir: /root/helix/work/workitem/e/GC/Features/Bridge/BridgeTester
        ArtifactsBinDir: 
        Generated app bundle at /root/helix/work/workitem/e/GC/Features/Bridge/BridgeTester/WasmApp/
      Warning: unknown flag --expose_wasm.
      Try --help for options
      Incoming arguments: --run BridgeTester.dll
      Application arguments: --run BridgeTester.dll
      console.info: Initializing dotnet version 9.0.5 commit hash badb953d260984e8897ad6a65194e5bf1f27a078
      System.ArgumentNullException: Value cannot be null. (Parameter 'path1')
         at System.ArgumentNullException.Throw(String paramName) in /_/src/libraries/System.Private.CoreLib/src/System/ArgumentNullException.cs:line 97
         at System.ArgumentNullException.ThrowIfNull(Object argument, String paramName) in /_/src/libraries/System.Private.CoreLib/src/System/ArgumentNullException.cs:line 59
         at System.IO.Path.Combine(String path1, String path2) in /_/src/libraries/System.Private.CoreLib/src/System/IO/Path.cs:line 346
         at BridgeTester.RunTests()
         at __GeneratedMainWrapper.Main()

Not sure why it doesn't happen in main

[filipnavara]

Would it make sense to disable the test on WASM to unblock us? WASM doesn't use the GC bridge anyway.

(I am currently in middle of moving our office so I don't know when I will be able to address it.)

[akoeplinger]

Sure, I can handle it.

[carlossanlop]

Friendly reminder that code complete for the may release is next Monday April 14th. If you want this change included in that release, please merge the PR before EOD Monday.

[akoeplinger]

/ba-g failures in the wasm tests are unrelated, it doesn't even compile this code.

[simonrozsival]

I can no longer reproduce the original bug reported by a customer in #115611 using their repro project with the latest changes.