Re: Re: [RFC] KVM: x86: SVM: don't expose PV_SEND_IPI feature with AVIC

[Wanpeng Li]

On Tue, 16 Nov 2021 at 16:56, Chao Gao <chao.gao@xxxxxxxxx> wrote:
>
> On Tue, Nov 16, 2021 at 10:56:25AM +0800, zhenwei pi wrote:
> >
> >
> >On 11/16/21 10:48 AM, Wanpeng Li wrote:
> >> On Mon, 8 Nov 2021 at 22:09, Maxim Levitsky <mlevitsk@xxxxxxxxxx> wrote:
> >> >
> >> > On Mon, 2021-11-08 at 11:30 +0100, Paolo Bonzini wrote:
> >> > > On 11/8/21 10:59, Kele Huang wrote:
> >> > > > Currently, AVIC is disabled if x2apic feature is exposed to guest
> >> > > > or in-kernel PIT is in re-injection mode.
> >> > > >
> >> > > > We can enable AVIC with options:
> >> > > >
> >> > > >     Kmod args:
> >> > > >     modprobe kvm_amd avic=1 nested=0 npt=1
> >> > > >     QEMU args:
> >> > > >     ... -cpu host,-x2apic -global kvm-pit.lost_tick_policy=discard ...
> >> > > >
> >> > > > When LAPIC works in xapic mode, both AVIC and PV_SEND_IPI feature
> >> > > > can accelerate IPI operations for guest. However, the relationship
> >> > > > between AVIC and PV_SEND_IPI feature is not sorted out.
> >> > > >
> >> > > > In logical, AVIC accelerates most of frequently IPI operations
> >> > > > without VMM intervention, while the re-hooking of apic->send_IPI_xxx
> >> > > > from PV_SEND_IPI feature masks out it. People can get confused
> >> > > > if AVIC is enabled while getting lots of hypercall kvm_exits
> >> > > > from IPI.
> >> > > >
> >> > > > In performance, benchmark tool
> >> > > > https://lore.kernel.org/kvm/20171219085010.4081-1-ynorov@xxxxxxxxxxxxxxxxxx/
> >> > > > shows below results:
> >> > > >
> >> > > >     Test env:
> >> > > >     CPU: AMD EPYC 7742 64-Core Processor
> >> > > >     2 vCPUs pinned 1:1
> >> > > >     idle=poll
> >> > > >
> >> > > >     Test result (average ns per IPI of lots of running):
> >> > > >     PV_SEND_IPI      : 1860
> >> > > >     AVIC             : 1390
> >> > > >
> >> > > > Besides, disscussions in https://lkml.org/lkml/2021/10/20/423
> >> > > > do have some solid performance test results to this.
> >> > > >
> >> > > > This patch fixes this by masking out PV_SEND_IPI feature when
> >> > > > AVIC is enabled in setting up of guest vCPUs' CPUID.
> >> > > >
> >> > > > Signed-off-by: Kele Huang <huangkele@xxxxxxxxxxxxx>
> >> > >
> >> > > AVIC can change across migration.  I think we should instead use a new
> >> > > KVM_HINTS_* bit (KVM_HINTS_ACCELERATED_LAPIC or something like that).
> >> > > The KVM_HINTS_* bits are intended to be changeable across migration,
> >> > > even though we don't have for now anything equivalent to the Hyper-V
> >> > > reenlightenment interrupt.
> >> >
> >> > Note that the same issue exists with HyperV. It also has PV APIC,
> >> > which is harmful when AVIC is enabled (that is guest uses it instead
> >> > of using AVIC, negating AVIC benefits).
> >> >
> >> > Also note that Intel recently posted IPI virtualizaion, which
> >> > will make this issue relevant to APICv too soon.
> >>
> >> The recently posted Intel IPI virtualization will accelerate unicast
> >> ipi but not broadcast ipis, AMD AVIC accelerates unicast ipi well but
> >> accelerates broadcast ipis worse than pv ipis. Could we just handle
> >> unicast ipi here?
> >>
> >>      Wanpeng
> >>
> >Depend on the number of target vCPUs, broadcast IPIs gets unstable
> >performance on AVIC, and usually worse than PV Send IPI.
> >So agree with Wanpeng's point, is it possible to separate single IPI and
> >broadcast IPI on a hardware acceleration platform?
>
> Actually, this is how kernel works in x2apic mode: use PV interface
> (hypercall) to send multi-cast IPIs and write ICR MSR directly to send
> unicast IPIs.
>
> But if guest works in xapic mode, both unicast and multi-cast are issued
> via PV interface. It is a side-effect introduced by commit aaffcfd1e82d.
>
> how about just correcting the logic for xapic:
>
> From 13447b221252b64cd85ed1329f7d917afa54efc8 Mon Sep 17 00:00:00 2001
> From: Jiaqing Zhao <jiaqing.zhao@xxxxxxxxx>
> Date: Fri, 9 Apr 2021 13:53:39 +0800
> Subject: [PATCH 1/2] x86/apic/flat: Add specific send IPI logic
>
> Currently, apic_flat.send_IPI() uses default_send_IPI_single(), which
> is a wrapper of apic->send_IPI_mask(). Since commit aaffcfd1e82d
> ("KVM: X86: Implement PV IPIs in linux guest"), KVM PV IPI driver will
> override apic->send_IPI_mask(), and may cause unwated side effects.
>
> This patch removes such side effects by creating a specific send_IPI
> method.

This looks reasonable to me.

    Wanpeng