Re: [PATCH v17 092/116] KVM: TDX: Handle TDX PV HLT hypercall

From: Sean Christopherson
Date: Tue Jan 09 2024 - 11:21:42 EST


On Mon, Jan 08, 2024, Chao Gao wrote:
> On Fri, Jan 05, 2024 at 03:05:12PM -0800, Sean Christopherson wrote:
> >On Tue, Nov 07, 2023, isaku.yamahata@xxxxxxxxx wrote:
> >> From: Isaku Yamahata <isaku.yamahata@xxxxxxxxx>
> >>
> >> Wire up TDX PV HLT hypercall to the KVM backend function.
> >>
> >> Signed-off-by: Isaku Yamahata <isaku.yamahata@xxxxxxxxx>
> >> ---
> >> arch/x86/kvm/vmx/tdx.c | 42 +++++++++++++++++++++++++++++++++++++++++-
> >> arch/x86/kvm/vmx/tdx.h | 3 +++
> >> 2 files changed, 44 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/arch/x86/kvm/vmx/tdx.c b/arch/x86/kvm/vmx/tdx.c
> >> index 3a1fe74b95c3..4e48989d364f 100644
> >> --- a/arch/x86/kvm/vmx/tdx.c
> >> +++ b/arch/x86/kvm/vmx/tdx.c
> >> @@ -662,7 +662,32 @@ void tdx_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
> >>
> >> bool tdx_protected_apic_has_interrupt(struct kvm_vcpu *vcpu)
> >> {
> >> - return pi_has_pending_interrupt(vcpu);
> >> + bool ret = pi_has_pending_interrupt(vcpu);
> >> + struct vcpu_tdx *tdx = to_tdx(vcpu);
> >> +
> >> + if (ret || vcpu->arch.mp_state != KVM_MP_STATE_HALTED)
> >> + return true;
> >> +
> >> + if (tdx->interrupt_disabled_hlt)
> >> + return false;
> >> +
> >> + /*
> >> + * This is for the case where the virtual interrupt is recognized,
> >> + * i.e. set in vmcs.RVI, between the STI and "HLT". KVM doesn't have
> >> + * access to RVI and the interrupt is no longer in the PID (because it
> >> + * was "recognized". It doesn't get delivered in the guest because the
> >> + * TDCALL completes before interrupts are enabled.
> >> + *
> >> + * TDX modules sets RVI while in an STI interrupt shadow.
> >> + * - TDExit(typically TDG.VP.VMCALL<HLT>) from the guest to TDX module.
> >> + * The interrupt shadow at this point is gone.
> >> + * - It knows that there is an interrupt that can be delivered
> >> + * (RVI > PPR && EFLAGS.IF=1, the other conditions of 29.2.2 don't
> >> + * matter)
> >> + * - It forwards the TDExit nevertheless, to a clueless hypervisor that
> >> + * has no way to glean either RVI or PPR.
> >
> >WTF. Seriously, what in the absolute hell is going on. I reported this internally
> >four ***YEARS*** ago. This is not some obscure theoretical edge case, this is core
> >functionality and it's completely broken garbage.
> >
> >NAK. Hard NAK. Fix the TDX module, full stop.
> >
> >Even worse, TDX 1.5 apparently _already_ has the necessary logic for dealing with
> >interrupts that are pending in RVI when handling NESTED VM-Enter. Really!?!?!
> >Y'all went and added nested virtualization support of some kind, but can't find
> >the time to get the basics right?
>
> We actually fixed the TDX module. See 11.9.5. Pending Virtual Interrupt
> Delivery Indication in TDX module 1.5 spec [1]
>
> The host VMM can detect whether a virtual interrupt is pending delivery to a
> VCPU in the Virtual APIC page, using TDH.VP.RD to read the VCPU_STATE_DETAILS
> TDVPS field.
>
> The typical use case is when the guest TD VCPU indicates to the host VMM, using
> TDG.VP.VMCALL, that it has no work to do and can be halted. The guest TD is
> expected to pass an “interrupt blocked” flag. The guest TD is expected to set
> this flag to 0 if and only if RFLAGS.IF is 1 or the TDCALL instruction that
> invokes TDG.VP.VMCALL immediately follows an STI instruction. If the “interrupt
> blocked” flag is 0, the host VMM can determine whether to re-schedule the guest
> TD VCPU based on VCPU_STATE_DETAILS.
>
> Isaku, this patch didn't read VCPU_STATE_DETAILS. Maybe you missed something
> during rebase? Regarding buggy_hlt_workaround, do you aim to avoid reading
> VCPU_STATE_DETAILS as much as possible (because reading it via SEAMCALL is
> costly, ~3-4K cycles)?

*sigh* Why only earth doesn't the TDX module simply compute VMXIP on TDVMCALL?
It's literally one bit and one extra VMREAD. There are plenty of register bits
available, and I highly doubt ~20 cycles in the TDVMCALL path will be noticeable,
let alone problematic. Such functionality could even be added on top in a TDX
module update, and Intel could even bill it as a performance optimization.

Eating 4k cycles in the HLT path isn't the end of the world, but it's far from
optimal and it's just so incredibly wasteful. I wouldn't be surprised if the
latency is measurable for certain workloads, which will lead to guests using
idle=poll and/or other games being played in the guest.

And AFAICT, the TDX module doesn't support HLT passthrough, so fully dedicated
CPUs can't even mitigate the pain that way.

Anyways, regarding the "workaround", my NAK stands. It has bad tradeoffs of its
own, e.g. will result in spurious wakeups, and can't possibly work for VMs with
passthrough devices. Not to mention that the implementation has several races
and false positives.