On Tue, Aug 15, 2017 at 05:54:34PM +0200, Stefan Fritsch wrote:
> On Tue, 15 Aug 2017, Mike Larkin wrote:
>
> > On Tue, Aug 15, 2017 at 08:46:59AM -0700, Mike Larkin wrote:
> > > On Tue, Aug 15, 2017 at 05:39:29PM +0200, Stefan Fritsch wrote:
> > > > I have got a report that openbsd panics on boot with qemu -cpu
> > > > Opteron_G3
> > > > (but Opteron_G2 works).
> > > >
> > > > kernel: protection fault trap, code=0
> > > > Stopped at amd64_errata_setmsr+0x14: rdmsr
> > > > ddb{0}> >> OpenBSD/amd64 BOOT 3.33
> > > > boot>
> > > >
> > > > Qemu does not implement all the secret errata MSRs. Does it make sense
> > > > to
> > > > simply skip all errata processing if we detect a hypervisor?
> > > >
> > > >
> > > > diff --git a/sys/arch/amd64/amd64/identcpu.c
> > > > b/sys/arch/amd64/amd64/identcpu.c
> > > > index a448b885ba7..371c0c8ff48 100644
> > > > --- a/sys/arch/amd64/amd64/identcpu.c
> > > > +++ b/sys/arch/amd64/amd64/identcpu.c
> > > > @@ -708,7 +708,7 @@ identifycpu(struct cpu_info *ci)
> > > > }
> > > > #endif
> > > >
> > > > - if (!strcmp(cpu_vendor, "AuthenticAMD"))
> > > > + if (!strcmp(cpu_vendor, "AuthenticAMD") &&
> > > > !ISSET(cpu_ecxfeature, CPUIDECX_HV))
> > > > amd64_errata(ci);
> > > >
> > > > if (CPU_IS_PRIMARY(ci) && !strcmp(cpu_vendor, "CentaurHauls")) {
> > > >
> > >
> > > I think this is an upstream bug. If they claim to emulate a G3 Opteron,
> > > they
> > > should be faithfully emulating it.
>
> True. But I could imagine that this is difficult if there is no hint that
> a specific MSR exists until AMD releases an erratum that uses it as a
> work-around. At the very least, there will be old qemu/kvm versions around
> that don't implement it, yet. Not sure if that is the problem, here,
> though.
>
> > Also, is this only an error in TCG?
>
> The panic above is from qemu+kvm+linux. Not sure what happens with TCG.
I'm still not convinced we should be working around this. It can be fixed by
a 1 character change to the user's launch script/profile (G3->2) and unless
there is evidence we are applying the errata incorrectly, we should press
upstream to fix it. "But what about old kvm" doesn't really make a convincing
argument. We could be saying "but what about VMware workstation from 1999",
too. Should we be working around their bugs also?
-ml
