On Fri, Nov 20, 2020 at 03:08:42PM +0100, Mark Kettenis wrote:
> > Date: Fri, 20 Nov 2020 07:41:20 -0500
> > From: Bryan Steele <bry...@gmail.com>
> >
> > On Fri, Nov 20, 2020 at 09:26:08AM +0100, Otto Moerbeek wrote:
> > > Hi,
> > >
> > > I got a new Ryzen machine, dmesg below. What I'm observing might be a
> > > issue with hw.setperf.
> > >
> > > On startsup it shows:
> > >
> > > hw.cpuspeed=3800
> > > hw.setperf=100
> > >
> > > If I lower hw.setperf to zero, the new state is reflect immediately in
> > > hw.cpuspeed:
> > >
> > > hw.cpuspeed=2200
> > > hw.setperf=0
> > >
> > > And also sha256 -t becomes slower as expected.
> > >
> > > But If I raise hw.setperf to 100 I'm seeing:
> > >
> > > hw.cpuspeed=2200
> > > hw.setperf=100
> > >
> > > and sha256 -t is still slow. Only after some time passes (lets say a
> > > couple of tens of seconds) it does show:
> > >
> > > hw.cpuspeed=3800
> > > hw.setperf=100
> > >
> > > and sha256 -t is fast again.
> > >
> > > This behaviour is different from my old machine, where setting
> > > hs.setperf was reflected in hs.cpuspeed immediately both ways
> > >
> > > Any clue?
> > >
> > > -Otto
> >
> > Hey Otto,
> >
> > Nice machine! :-)
> >
> > I've seen this "sticking" issue before (as have others), but haven't
> > been able to narrow it down unfortunately. I'm not sure if it's a
> > bug in the k1x-pstate.c code I wrote, it's some undocumented new
> > behaviour on newer Ryzen CPUs, or if a MI setperf change happened
> > at some point that's unhandled..
> >
> > At least on a desktop I'd suggest to leaved apmd(8) and not do any
> > manual hw.setperf tweaking, you should have adequate cooling and the
> > BIOS will automatically adjust the CPU fan to keep it so. I believe
> > it will also allow it to more quickly move into CPB boost frequencies
> > if left at P-state L0 (but don't quote me on that).
>
> I would expect this machine to use the acpucpu(4) setperf
> implementation. Figuring out if that is indeed the case would
> probably be step 1 in debugging this.
I didn't realize there was a setperf implementation in acpicpu(4),
k1x-pstate depends on acpicpu(4) to to gather PSS information, but
otherwise writes the MSRs out itself rather than calling any ACPI
methods.
In identifycpu() we're just matching on family, e.g:
if (ci->ci_family >= 0x10)
setperf_setup = k1x_init;
The Intel SpeedStep case is below and matches based on a CPUID flag,
so I don't see when the acpucpu implementation would ever be chosen
at on either..
-Bryan.
