On Mon, Oct 19, 2020 at 01:32:00PM +0200, Vitaly Kuznetsov wrote: > Sean Christopherson <sean.j.christopherson@xxxxxxxxx> writes: > >> > diff --git a/x86/unittests.cfg b/x86/unittests.cfg > >> > index 872d679..c72a659 100644 > >> > --- a/x86/unittests.cfg > >> > +++ b/x86/unittests.cfg > >> > @@ -41,7 +41,7 @@ file = apic.flat > >> > smp = 2 > >> > extra_params = -cpu qemu64,+x2apic,+tsc-deadline > >> > arch = x86_64 > >> > -timeout = 30 > >> > +timeout = 240 > >> > > >> > [ioapic] > >> > file = ioapic.flat > >> > >> AFAIR the default timeout for tests where timeout it not set explicitly > >> is 90s so don't you need to also modify it for other tests like > >> 'apic-split', 'ioapic', 'ioapic-split', ... ? > >> > >> I was thinking about introducing a 'timeout multiplier' or something to > >> run_tests.sh for running in slow (read: nested) environments, doing that > >> would allow us to keep reasonably small timeouts by default. This is > >> somewhat important as tests tend to hang and waiting for 4 minutes every > >> time is not great. > > > > I would much prefer to go in the other direction and make tests like APIC not > > do so many loops (in a nested environment?). The port80 test in particular is > > an absolute waste of time. > > I don't think these two suggestions are opposite. Yes, making tests run fast > is good, however, some of the tests are doomed to be slow. E.g. running > VMX testsuite while nested (leaving aside the question about who needs > three level nesting) is always going to be much slower than on bare metal. Ya, I was specifically referring to tests that arbitrarily choose a high loop count, without any real/documented justification for running millions of loops. > > E.g. does running 1M loops in test_multiple_nmi() really add value versus > > say 10k or 100k loops? > > Oddly enough, I vaguely remember this particular test hanging > *sometimes* after a few thousand loops but I don't remember any > details. Thousands still ain't millions :-D. IMO, the unit tests should sit between a smoke test and a long running, intensive stress test, i.e. the default config shouldn't be trying to find literal one-in-a-million bugs on every run.
