kvm selftest 'msr' fails on some skylake cpus

kvm selftest 'msr' fails on some skylake cpus

[Maxim Levitsky]

Hi!

This kvm unit test tests that all reserved bits of the MSR_IA32_FLUSH_CMD #GP, but apparently
on some systems this test fails.

For example I reproduced this on:

model name	: Intel(R) Xeon(R) CPU E3-1260L v5 @ 2.90GHz
stepping	: 3
microcode	: 0xf0


As I see in the 'vmx_vcpu_after_set_cpuid', we passthough this msr to the guest AS IS,
thus the unit test tests the microcode.

So I suspect that the test actually caught a harmless microcode bug.

What do you think we should do to workaround this? Maybe disable this check on
affected cpus or turn it into a warning because MSR_IA32_FLUSH_CMD reserved bits
test doesn't test KVM?

Best regards,
	Maxim Levitsky

Re: kvm selftest 'msr' fails on some skylake cpus

[Sean Christopherson]

On Thu, Jun 13, 2024, Maxim Levitsky wrote:
> Hi!
> 
> This kvm unit test tests that all reserved bits of the MSR_IA32_FLUSH_CMD #GP, but apparently
> on some systems this test fails.
> 
> For example I reproduced this on:
> 
> model name	: Intel(R) Xeon(R) CPU E3-1260L v5 @ 2.90GHz
> stepping	: 3
> microcode	: 0xf0
> 
> 
> As I see in the 'vmx_vcpu_after_set_cpuid', we passthough this msr to the guest AS IS,
> thus the unit test tests the microcode.
> 
> So I suspect that the test actually caught a harmless microcode bug.

Yeah, we encountered the same thing and came to the same conclusion.

> What do you think we should do to workaround this? Maybe disable this check on
> affected cpus or turn it into a warning because MSR_IA32_FLUSH_CMD reserved bits
> test doesn't test KVM?

Ya, Mingwei posted a patch[*] to force KVM to emulate the faulting accesses, which
more or less does exactly that, but preserves a bit of KVM coverage.  I'll get a
KUT pull request sent to Paolo today, I've got a sizeable number of changes ready.

[*] https://lore.kernel.org/all/20240417232906.3057638-3-mizhang@google.com

Re: kvm selftest 'msr' fails on some skylake cpus

[Maxim Levitsky]

On Fri, 2024-06-14 at 07:26 -0700, Sean Christopherson wrote:
> On Thu, Jun 13, 2024, Maxim Levitsky wrote:
> > Hi!
> > 
> > This kvm unit test tests that all reserved bits of the MSR_IA32_FLUSH_CMD #GP, but apparently
> > on some systems this test fails.
> > 
> > For example I reproduced this on:
> > 
> > model name	: Intel(R) Xeon(R) CPU E3-1260L v5 @ 2.90GHz
> > stepping	: 3
> > microcode	: 0xf0
> > 
> > 
> > As I see in the 'vmx_vcpu_after_set_cpuid', we passthough this msr to the guest AS IS,
> > thus the unit test tests the microcode.
> > 
> > So I suspect that the test actually caught a harmless microcode bug.
> 
> Yeah, we encountered the same thing and came to the same conclusion.
> 
> > What do you think we should do to workaround this? Maybe disable this check on
> > affected cpus or turn it into a warning because MSR_IA32_FLUSH_CMD reserved bits
> > test doesn't test KVM?
> 
> Ya, Mingwei posted a patch[*] to force KVM to emulate the faulting accesses, which
> more or less does exactly that, but preserves a bit of KVM coverage.  I'll get a
> KUT pull request sent to Paolo today, I've got a sizeable number of changes ready.
> 
> [*] https://lore.kernel.org/all/20240417232906.3057638-3-mizhang@google.com
> 

This works for me.
Thanks,

Best regards,
	Maxim Levitsky

Re: kvm selftest 'msr' fails on some skylake cpus

[Jim Mattson]

On Mon, Jun 17, 2024 at 11:29 AM Maxim Levitsky <mlevitsk@redhat.com> wrote:
>
> On Fri, 2024-06-14 at 07:26 -0700, Sean Christopherson wrote:
> > On Thu, Jun 13, 2024, Maxim Levitsky wrote:
> > > Hi!
> > >
> > > This kvm unit test tests that all reserved bits of the MSR_IA32_FLUSH_CMD #GP, but apparently
> > > on some systems this test fails.
> > >
> > > For example I reproduced this on:
> > >
> > > model name  : Intel(R) Xeon(R) CPU E3-1260L v5 @ 2.90GHz
> > > stepping    : 3
> > > microcode   : 0xf0
> > >
> > >
> > > As I see in the 'vmx_vcpu_after_set_cpuid', we passthough this msr to the guest AS IS,
> > > thus the unit test tests the microcode.
> > >
> > > So I suspect that the test actually caught a harmless microcode bug.
> >
> > Yeah, we encountered the same thing and came to the same conclusion.
> >
> > > What do you think we should do to workaround this? Maybe disable this check on
> > > affected cpus or turn it into a warning because MSR_IA32_FLUSH_CMD reserved bits
> > > test doesn't test KVM?
> >
> > Ya, Mingwei posted a patch[*] to force KVM to emulate the faulting accesses, which
> > more or less does exactly that, but preserves a bit of KVM coverage.  I'll get a
> > KUT pull request sent to Paolo today, I've got a sizeable number of changes ready.
> >
> > [*] https://lore.kernel.org/all/20240417232906.3057638-3-mizhang@google.com
> >
>
> This works for me.
> Thanks,
>
> Best regards,
>         Maxim Levitsky

Sean and/or Paolo,

I'm still waiting for this to show up. :)

Re: kvm selftest 'msr' fails on some skylake cpus

[Sean Christopherson]

On Tue, Jun 25, 2024, Jim Mattson wrote:
> On Mon, Jun 17, 2024 at 11:29 AM Maxim Levitsky <mlevitsk@redhat.com> wrote:
> >
> > On Fri, 2024-06-14 at 07:26 -0700, Sean Christopherson wrote:
> > > On Thu, Jun 13, 2024, Maxim Levitsky wrote:
> > > > Hi!
> > > >
> > > > This kvm unit test tests that all reserved bits of the MSR_IA32_FLUSH_CMD #GP, but apparently
> > > > on some systems this test fails.
> > > >
> > > > For example I reproduced this on:
> > > >
> > > > model name  : Intel(R) Xeon(R) CPU E3-1260L v5 @ 2.90GHz
> > > > stepping    : 3
> > > > microcode   : 0xf0
> > > >
> > > >
> > > > As I see in the 'vmx_vcpu_after_set_cpuid', we passthough this msr to the guest AS IS,
> > > > thus the unit test tests the microcode.
> > > >
> > > > So I suspect that the test actually caught a harmless microcode bug.
> > >
> > > Yeah, we encountered the same thing and came to the same conclusion.
> > >
> > > > What do you think we should do to workaround this? Maybe disable this check on
> > > > affected cpus or turn it into a warning because MSR_IA32_FLUSH_CMD reserved bits
> > > > test doesn't test KVM?
> > >
> > > Ya, Mingwei posted a patch[*] to force KVM to emulate the faulting accesses, which
> > > more or less does exactly that, but preserves a bit of KVM coverage.  I'll get a
> > > KUT pull request sent to Paolo today, I've got a sizeable number of changes ready.
> > >
> > > [*] https://lore.kernel.org/all/20240417232906.3057638-3-mizhang@google.com
> > >
> >
> > This works for me.
> > Thanks,
> >
> > Best regards,
> >         Maxim Levitsky
> 
> Sean and/or Paolo,
> 
> I'm still waiting for this to show up. :)

13a12056be1e4939dadbd0d2cfb65df400832905 x86: msr: testing MSR_IA32_FLUSH_CMD reserved bits only in KVM emulation
51b87946279cf6e9248ceacf0f27833b6ebeef5e x86: Add FEP support on read/write register instructions

/drop mic