misc/cgo/test: Test9400 in TestCrossPackageTests hangs on ppc64le

[pmur]

While trying to reproduce some sporadic failures on seen on the ppc64le CI machines, I came across a hang in Test9400 on a POWER10 machine with GOMAXPROCS=2.

What version of Go are you using (go version)?

$ go version
go version devel go1.18-e8cda0a6c9 Fri Nov 19 00:05:59 2021 +0000 linux/ppc64le

What did you do?

cd $GOPATH/misc/cgo/test
go test -buildmode=pie -ldflags=-linkmode=internal -tags=internal,internal_pie -c
export GOMAXPROCS=2
while [ 1 ]; do ./test.test -test.v -test.short ; done

Eventually this hangs on TestCrossPackageTests running Test9400.

This kind of looks like it may be an untimely GC starting while issue9400.RewindAndSetgid is spinning on a condition.

I inserted a call to runtime.GC prior to the above call, and was unable to reproduce the hang.

[gopherbot]

Change https://golang.org/cl/365836 mentions this issue: misc/cgo/test: reduce likeliness of hang in Test9400

[gopherbot]

Change https://golang.org/cl/367874 mentions this issue: misc/cgo/test: further reduce likeliness of hang in Test9400

[mknyszek]

Update on the root cause here: the fixes for #49680, #49695, and #45867 all assumed that SetGCPercent(-1) doesn't block until the GC's mark phase is done, but it actually does. The cause of those failures is that at the beginning of the sweep phase, the GC does try to preempt every P once, and this may run concurrently with test code that has non-preemptible goroutines that create the potential for a deadlock.

[gopherbot]

Change https://golang.org/cl/369815 mentions this issue: runtime: fix comments on the behavior of SetGCPercent