Re: FSGSBASE causing panic on 5.9-rc1

[Tom Lendacky <thomas.lendacky@amd.com>]

On 8/20/20 5:34 PM, Sean Christopherson wrote:
> On Thu, Aug 20, 2020 at 03:07:10PM -0700, Andy Lutomirski wrote:
>> On Thu, Aug 20, 2020 at 3:05 PM Sean Christopherson
>> <sean.j.christopherson@intel.com> wrote:
>>>
>>> On Thu, Aug 20, 2020 at 01:36:46PM -0700, Andy Lutomirski wrote:
>>>>
>>>>
>>>>> On Aug 20, 2020, at 1:15 PM, Tom Lendacky <thomas.lendacky@amd.com> wrote:
>>>>>
>>>>> On 8/20/20 3:07 PM, Dave Hansen wrote:
>>>>>> On 8/20/20 12:05 PM, Tom Lendacky wrote:
>>>>>>>> I added a quick hack to save TSC_AUX to a new variable in the SVM
>>>>>>>> struct and then restore it right after VMEXIT (just after where GS is
>>>>>>>> restored in svm_vcpu_enter_exit()) and my guest is no longer crashing.
>>>>>>>
>>>>>>> Sorry, I mean my host is no longer crashing.
>>>>>> Just to make sure I've got this:
>>>>>> 1. Older CPUs didn't have X86_FEATURE_RDPID
>>>>>> 2. FSGSBASE patches started using RDPID in the NMI entry path when
>>>>>>     supported *AND* FSGSBASE was enabled
>>>>>> 3. There was a latent SVM bug which did not restore the RDPID data
>>>>>>     before NMIs were reenabled after VMEXIT
>>>>>> 4. If an NMI comes in the window between VMEXIT and the
>>>>>>     wrmsr(TSC_AUX)... boom
>>>>>
>>>>> Right, which means that the setting of TSC_AUX to the guest value needs to be moved, too.
>>>>>
>>>>
>>>> Depending on how much of a perf hit this is, we could also skip using RDPID
>>>> in the paranoid path on SVM-capable CPUs.
>>>
>>> Doesn't this affect VMX as well?  KVM+VMX doesn't restore TSC_AUX until the
>>> kernel returns to userspace.  I don't see anything that prevents the NMI
>>> RDPID path from affecting Intel CPUs.
>>>
>>> Assuming that's the case, I would strongly prefer this be handled in the
>>> paranoid path.  NMIs are unblocked immediately on VMX VM-Exit, which means
>>> using the MSR load lists in the VMCS, and I hate those with a vengeance.
>>>
>>> Perf overhead on VMX would be 8-10% for VM-Exits that would normally stay
>>> in KVM's run loop, e.g. ~125 cycles for the WMRSR, ~1300-1500 cycles to
>>> handle the most common VM-Exits.  It'd be even higher overhead for the
>>> VMX preemption timer, which is handled without even enabling IRQs and is
>>> a hot path as it's used to emulate the TSC deadline timer for the guest.
>>
>> I'm fine with that -- let's get rid of RDPID unconditionally in the
>> paranoid path.  Want to send a patch that also adds as comment
>> explaining why we're not using RDPID?
> 
> Sure, though I won't object if Tom beats me to the punch :-)

I can do it, but won't be able to get to it until sometime tomorrow.

Thanks,
Tom

>