Re: FSGSBASE causing panic on 5.9-rc1

[Jim Mattson]

On Thu, Aug 20, 2020 at 11:38 AM Jim Mattson <jmattson@xxxxxxxxxx> wrote:
>
> On Thu, Aug 20, 2020 at 11:34 AM Tom Lendacky <thomas.lendacky@xxxxxxx> wrote:
> >
> > On 8/20/20 11:30 AM, Tom Lendacky wrote:
> > > On 8/20/20 11:17 AM, Tom Lendacky wrote:
> > >> On 8/20/20 10:55 AM, Andy Lutomirski wrote:
> > >>> On Thu, Aug 20, 2020 at 8:21 AM Tom Lendacky <thomas.lendacky@xxxxxxx>
> > >>> wrote:
> > >>>>
> > >>>> On 8/20/20 10:10 AM, Sean Christopherson wrote:
> > >>>>> On Wed, Aug 19, 2020 at 05:21:33PM -0700, Andy Lutomirski wrote:
> > >>>>>> On Wed, Aug 19, 2020 at 2:25 PM Andy Lutomirski <luto@xxxxxxxxxx>
> > >>>>>> wrote:
> > >>>>>>>
> > >>>>>>> On Wed, Aug 19, 2020 at 11:19 AM Tom Lendacky
> > >>>>>>> <thomas.lendacky@xxxxxxx> wrote:
> > >>>>>>>>
> > >>>>>>>> On 8/19/20 1:07 PM, Tom Lendacky wrote:
> > >>>>>>>>> It looks like the FSGSBASE support is crashing my second
> > >>>>>>>>> generation EPYC
> > >>>>>>>>> system. I was able to bisect it to:
> > >>>>>>>>>
> > >>>>>>>>> b745cfba44c1 ("x86/cpu: Enable FSGSBASE on 64bit by default and
> > >>>>>>>>> add a chicken bit")
> > >>>>>>>>>
> > >>>>>>>>> The panic only happens when using KVM. Doing kernel builds or stress
> > >>>>>>>>> on bare-metal appears fine. But if I fire up, in this case, a
> > >>>>>>>>> 64-vCPU
> > >>>>>>>>> guest and do a kernel build within the guest, I get the following:
> > >>>>>>>>
> > >>>>>>>> I should clarify that this panic is on the bare-metal system, not
> > >>>>>>>> in the
> > >>>>>>>> guest. And that specifying nofsgsbase on the bare-metal command
> > >>>>>>>> line fixes
> > >>>>>>>> the issue.
> > >>>>>>>
> > >>>>>>> I certainly see some oddities:
> > >>>>>>>
> > >>>>>>> We have this code:
> > >>>>>>>
> > >>>>>>> static void svm_vcpu_put(struct kvm_vcpu *vcpu)
> > >>>>>>> {
> > >>>>>>>           struct vcpu_svm *svm = to_svm(vcpu);
> > >>>>>>>           int i;
> > >>>>>>>
> > >>>>>>>           avic_vcpu_put(vcpu);
> > >>>>>>>
> > >>>>>>>           ++vcpu->stat.host_state_reload;
> > >>>>>>>           kvm_load_ldt(svm->host.ldt);
> > >>>>>>> #ifdef CONFIG_X86_64
> > >>>>>>>           loadsegment(fs, svm->host.fs);
> > >>>>>>>           wrmsrl(MSR_KERNEL_GS_BASE, current->thread.gsbase);
> > >>>>>
> > >>>>> Pretty sure current->thread.gsbase can be stale, i.e. this needs:
> > >>>>>
> > >>>>>        current_save_fsgs();
> > >>>>
> > >>>> I did try adding current_save_fsgs() in svm_vcpu_load(), saving the
> > >>>> current->thread.gsbase value to a new variable in the svm struct. I then
> > >>>> used that variable in the wrmsrl below, but it still crashed.
> > >>>
> > >>> Can you try bisecting all the way back to:
> > >>>
> > >>> commit dd649bd0b3aa012740059b1ba31ecad28a408f7f
> > >>> Author: Andy Lutomirski <luto@xxxxxxxxxx>
> > >>> Date:   Thu May 28 16:13:48 2020 -0400
> > >>>
> > >>>      x86/cpu: Add 'unsafe_fsgsbase' to enable CR4.FSGSBASE
> > >>>
> > >>> and adding the unsafe_fsgsbase command line option while you bisect.
> > >>
> > >> I'll give that a try.
> >
> > Bisecting with unsafe_fsgsbase identified:
> >
> > c82965f9e530 ("x86/entry/64: Handle FSGSBASE enabled paranoid entry/exit")
> >
> > But I'm thinking that could be because it starts using GET_PERCPU_BASE,
> > which on Rome would use RDPID. So is SVM restoring TSC_AUX_MSR too late?
> > That would explain why I don't see the issue on Naples, which doesn't
> > support RDPID.
>
> It looks to me like SVM loads the guest TSC_AUX from vcpu_load to
> vcpu_put, with this comment:
>
> /* This assumes that the kernel never uses MSR_TSC_AUX */
> if (static_cpu_has(X86_FEATURE_RDTSCP))
>         wrmsrl(MSR_TSC_AUX, svm->tsc_aux);

Correction: It never restores TSC_AUX, AFAICT.