Re: [PATCH] x86/msr: do not warn on writes to OC_MAILBOX

[Jason A. Donenfeld]

On Tue, Sep 8, 2020 at 7:10 PM Srinivas Pandruvada
<srinivas.pandruvada@xxxxxxxxxxxxxxx> wrote:
>
> On Mon, 2020-09-07 at 12:06 +0200, Borislav Petkov wrote:
> > + Srinivas.
> > + kitsunyan.
> >
> > On Mon, Sep 07, 2020 at 11:48:43AM +0200, Jason A. Donenfeld wrote:
> > > Popular tools, like intel-undervolt, use MSR 0x150 to control the
> > > CPU
> > > voltage offset. In fact, evidently the intel_turbo_max_3 driver in-
> > > tree
> > > also uses this MSR. So, teach the kernel's MSR list about this, so
> > > that
> > > intel-undervolt and other such tools don't spew warnings to dmesg,
> > > while
> > > unifying the constant used throughout the kernel.
> > >
>
> [...]
>
> > > -   if (reg == MSR_IA32_ENERGY_PERF_BIAS)
> > > +   switch (reg) {
> > > +   case MSR_IA32_ENERGY_PERF_BIAS:
> There is already sysfs interface for it.
>
> > > +   case MSR_IA32_OC_MAILBOX:
> > >             return 0;
> > > +   }
> > >
>
> [...]
>
> > Actually, we added the filtering to catch exactly such misuses and,
> > lemme check what is the proper word now... /me checks, aha, adding
> > new
> > MSRs to the "passlist" is the wrong thing to do.
> >
> > Srinivas, can you pls convert this in-tree driver to use a proper
> > sysfs
> > interface for that mailbox MSR and also work with the intel-undervolt
> > author - I hope I have the right person CCed from the git repo on
> > github
> > - to come up with a proper interface so that we can drop this MSR use
> > too.
>
> Overclocking is not architectural I/F and is supported by some special
> CPU skews. I can't find any public document to specify the commands
> which can be used via this OC mailbox. I have to check internally to
> see if there is any. To add a proper sysfs interface we have to make
> sure that we are not allowing some random commands to hardware and
> crash the system.

Well you can definitely crash the system this way -- undervolting can
result in all sorts of nice glitching. You might be able to even
programmatically undervolt to compromise the kernel in clever ways (a
lockdown bypass, I guess, but who cares).

That's why I initially suggested this was pretty squarely in the realm
of hobbyists and should just be added to that whitelist.