Re: [RFC PATCH v2 02/11] KVM: x86: Advertise CPUID.7.2.EDX and RRSBA_CTRL support

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Thu, May 18, 2023 at 05:50:17PM +0800, Xiaoyao Li wrote:
>On 5/16/2023 5:09 PM, Chao Gao wrote:
>> On Tue, May 16, 2023 at 03:03:15PM +0800, Xiaoyao Li wrote:
>> > On 5/16/2023 11:01 AM, Chao Gao wrote:
>> > > On Tue, May 16, 2023 at 10:22:22AM +0800, Xiaoyao Li wrote:
>> > > > > > I think we need to fix this bug at first.
>> > > > > 
>> > > > > I have no idea how to fix the "bug" without intercepting the MSR. The
>> > > > > performance penalty makes me think intercepting the MSR is not a viable
>> > > > > solution.
>> > > > 
>> > > > I thought correctness always takes higher priority over performance.
>> > > 
>> > > It is generally true. however, there are situations where we should make
>> > > trade-offs between correctness and other factors (like performance):
>> > > 
>> > > E.g., instructions without control bits, to be 100% compliant with CPU
>> > > spec, in theory, VMMs can trap/decode every instruction and inject #UD
>> > > if a guest tries to use some instructions it shouldn't.
>> > 
>> > This is the virtualization hole. IMHO, they are different things.
>> 
>> what are the differences between?
>> 1. Executing some unsupported instructions should cause #UD. But this is allowed
>>     in a KVM guest.
>> 2. Setting some reserved bits in SPEC_CTRL MSR should cause #GP. But this is
>>     allowed in a KVM guest.
>
>The difference is that for virtualization hole, there is no way but intercept
>and decode every instruction if we want the correctness. It's a disaster.
>
>But for MSR virtualization, we do have an option and we don't need to trap
>every instruction. MSR interception is the designated mechanism to correctly
>and elegantly virtualize the MSR.

The gains in this two cases are similar: some operations in guest are prevented.
But the costs on performance are not. So, how do you draw the line when we can
sacrafice correctness for performance?

>
>> > 
>> > Pass through MSR_IA32_SPEC_CTRL was introduced in commit d28b387fb74d
>> > ("KVM/VMX: Allow direct access to MSR_IA32_SPEC_CTRL"). At that time there
>> > was only a few bits defined, and the changelog called out that
>> > 
>> >   No attempt is made to handle STIBP here, intentionally. Filtering
>> >   STIBP may be added in a future patch, which may require trapping all
>> >   writes if we don't want to pass it through directly to the guest.
>> > 
>> > Per my undesrstanding, it implied that we need to re-visit it when more bits
>> > added instead of following the pass-through design siliently.
>> 
>> I don't object to re-visiting the design. My point is that to prevent guests from
>> setting RRSBA_CTRL/BHI_CTRL when they are not advertised isn't a strong
>> justfication for intercepting the MSR. STIBP and other bits (except IBRS) have
>> the same problem. And the gain of fixing this is too small.
>> 
>> If passing through the SPEC_CTRL MSR to guests might cause security issues, I
>> would agree to intercept accesses to the MSR.
>
>I never buy it. How to interpret the security? If the user wants to hide one
>feature from guest but KVM allows it when KVM does have a reasonable way to
>hide it. Does it violate the security?

I would say no. By "security", I mean guest becomes vulnerable to some issues or
guest attacks host or guest can access unauthorized data.

I tried to say that if the value of intercepting IA32_SPEC_CTRL outweighs the
perform penalty, it makes sense to do that. I guess the decision has been made
when enabling STIBP and it means people care more about performance than
preventing guests from setting STIBP.



[Index of Archives]     [KVM ARM]     [KVM ia64]     [KVM ppc]     [Virtualization Tools]     [Spice Development]     [Libvirt]     [Libvirt Users]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite Questions]     [Linux Kernel]     [Linux SCSI]     [XFree86]

  Powered by Linux