Re: [PATCH 0/2] KVM: kvm-coco-queue: Support protected TSC

[Marcelo Tosatti <mtosatti@redhat.com>]

On Sun, Oct 27, 2024 at 10:06:17PM +0800, Xiaoyao Li wrote:
> On 10/26/2024 12:24 AM, Marcelo Tosatti wrote:
> > On Mon, Oct 14, 2024 at 08:17:19PM +0530, Nikunj A. Dadhania wrote:
> > > Hi Isaku,
> > > 
> > > On 10/12/2024 1:25 PM, Isaku Yamahata wrote:
> > > > This patch series is for the kvm-coco-queue branch.  The change for TDX KVM is
> > > > included at the last.  The test is done by create TDX vCPU and run, get TSC
> > > > offset via vCPU device attributes and compare it with the TDX TSC OFFSET
> > > > metadata.  Because the test requires the TDX KVM and TDX KVM kselftests, don't
> > > > include it in this patch series.
> > > > 
> > > > 
> > > > Background
> > > > ----------
> > > > X86 confidential computing technology defines protected guest TSC so that the
> > > > VMM can't change the TSC offset/multiplier once vCPU is initialized and the
> > > > guest can trust TSC.  The SEV-SNP defines Secure TSC as optional.  TDX mandates
> > > > it.  The TDX module determines the TSC offset/multiplier.  The VMM has to
> > > > retrieve them.
> > > > 
> > > > On the other hand, the x86 KVM common logic tries to guess or adjust the TSC
> > > > offset/multiplier for better guest TSC and TSC interrupt latency at KVM vCPU
> > > > creation (kvm_arch_vcpu_postcreate()), vCPU migration over pCPU
> > > > (kvm_arch_vcpu_load()), vCPU TSC device attributes (kvm_arch_tsc_set_attr()) and
> > > > guest/host writing to TSC or TSC adjust MSR (kvm_set_msr_common()).
> > > > 
> > > > 
> > > > Problem
> > > > -------
> > > > The current x86 KVM implementation conflicts with protected TSC because the
> > > > VMM can't change the TSC offset/multiplier.  Disable or ignore the KVM
> > > > logic to change/adjust the TSC offset/multiplier somehow.
> > > > 
> > > > Because KVM emulates the TSC timer or the TSC deadline timer with the TSC
> > > > offset/multiplier, the TSC timer interrupts are injected to the guest at the
> > > > wrong time if the KVM TSC offset is different from what the TDX module
> > > > determined.
> > > > 
> > > > Originally the issue was found by cyclic test of rt-test [1] as the latency in
> > > > TDX case is worse than VMX value + TDX SEAMCALL overhead.  It turned out that
> > > > the KVM TSC offset is different from what the TDX module determines.
> > > 
> > > Can you provide what is the exact command line to reproduce this problem ?
> > 
> > Nikunj,
> > 
> > Run cyclictest, on an isolated CPU, in a VM. For the maximum latency
> > metric, rather than 50us, one gets 500us at times.
> > 
> > > Any links to this reported issue ?
> > 
> > This was not posted publically. But its not hard to reproduce.
> > 
> > > > Solution
> > > > --------
> > > > The solution is to keep the KVM TSC offset/multiplier the same as the value of
> > > > the TDX module somehow.  Possible solutions are as follows.
> > > > - Skip the logic
> > > >    Ignore (or don't call related functions) the request to change the TSC
> > > >    offset/multiplier.
> > > >    Pros
> > > >    - Logically clean.  This is similar to the guest_protected case.
> > > >    Cons
> > > >    - Needs to identify the call sites.
> > > > 
> > > > - Revert the change at the hooks after TSC adjustment
> > > >    x86 KVM defines the vendor hooks when the TSC offset/multiplier are
> > > >    changed.  The callback can revert the change.
> > > >    Pros
> > > >    - We don't need to care about the logic to change the TSC offset/multiplier.
> > > >    Cons:
> > > >    - Hacky to revert the KVM x86 common code logic.
> > > > 
> > > > Choose the first one.  With this patch series, SEV-SNP secure TSC can be
> > > > supported.
> > > 
> > > I am not sure how will this help SNP Secure TSC, as the GUEST_TSC_OFFSET and
> > > GUEST_TSC_SCALE are only available to the guest.
> > 
> > Nikunj,
> > 
> > FYI:
> > 
> > SEV-SNP processors (at least the one below) do not seem affected by this problem.
> 
> Did you apply Secure TSC patches of (guest kernel, KVM and QEMU) manualy?
> because none of them are merged. 

Yes. cyclictest latency, on a system configured with tuned
realtime-virtual-host/realtime-virtual-guest tuned profiles,
goes from 30us to 50us.

> Otherwise, I think SNP guest is still using
> KVM emulated TSC.

Not in the case the test was made.