The great witness test introduced by @bsdphk and reworked relatively
recently by @Dridi reported a hypothetical lock cycle here.

This was not a real issue, because the scheduling code would not re-enter
the probe code, yet we might still drop the probe lock while
scheduling.