macOS: Sporadic druntime integration test 'gc' (forkgc2) failure

[kinke]

Testing forkgc2
/Users/runner/work/1/build/runtime/druntime-test-gc-debug/forkgc2 
gnumake: *** [/Users/runner/work/1/build/runtime/druntime-test-gc-debug/forkgc2.done] Abort trap: 6

Happened twice in a row yesterday with Azure image 20210626.1 after the tiny Travis change. Was green with 20210620.1. The release variant of the test passed both times, so it might be a timing issue.

[kinke]

The same image apparently still worked both times with GitHub Actions for the LLVM 7 job (the test isn't run for the LLVM 10 job because excluded with BUILD_SHARED_LIBS=ON...).

[MoonlightSentinel]

For reference, I've seen several spurious failures for that test happening on different CI machines. E.g.:

Testing forkgc timelimit -t 10 ./generated/osx/release/64/forkgc [parent] Creating garbage... [parent] Collecting garbage... [parent] Forking... [child] In fork. [child] Creating garbage... [child] Collecting garbage... [child] Exiting fork. [parent] Creating garbage... [parent] Collecting garbage... [parent] Forking... [parent] Waiting for fork (PID 48354). [parent] Fork 48354 exited (0). Testing forkgc2 timelimit -t 10 ./generated/osx/release/64/forkgc2 forkgc2(48359,0x102145000) malloc: *** error for object 0x7fb2c9500150: double free *** set a breakpoint in malloc_error_break to debug make[2]: *** [generated/osx/release/64/forkgc2.done] Error 134 make[1]: *** [test/gc/.run] Error 2 make: *** [unittest-release] Error 2

[kinke]

Ah, double free, definitely interesting.

on different CI machines

Do you recall if it was Mac only or also happening on other OSs?

[MoonlightSentinel]

Only found this recent failure (also on Mac) but i doubt that it was restricted to Mac. Not sure though, this error occurred rarely over the last year(s).

[kinke]

A retry attempt a few hours ago has worked, so it's apparently sporadic. Maybe just bad luck on that day, or the probability of failure has increased for some reason...