Re: [PATCH V3 2/2] target/i386/kvm: Add Hyper-V direct tlb flush support

[Roman Kagan <rkagan@virtuozzo.com>]

On Tue, Oct 22, 2019 at 07:04:11PM +0200, Paolo Bonzini wrote:
> On 16/10/19 15:07, lantianyu1986@gmail.com wrote:

Somehow this patch never got through to me so I'll reply here.

> > From: Tianyu Lan <Tianyu.Lan@microsoft.com>
> > 
> > Hyper-V direct tlb flush targets KVM on Hyper-V guest.
> > Enable direct TLB flush for its guests meaning that TLB
> > flush hypercalls are handled by Level 0 hypervisor (Hyper-V)
> > bypassing KVM in Level 1. Due to the different ABI for hypercall
> > parameters between Hyper-V and KVM, KVM capabilities should be
> > hidden when enable Hyper-V direct tlb flush otherwise KVM
> > hypercalls may be intercepted by Hyper-V. Add new parameter
> > "hv-direct-tlbflush". Check expose_kvm and Hyper-V tlb flush
> > capability status before enabling the feature.
> > 
> > Signed-off-by: Tianyu Lan <Tianyu.Lan@microsoft.com>
> > ---
> > Change sicne v2:
> >        - Update new feature description and name.
> >        - Change failure print log.
> > 
> > Change since v1:
> >        - Add direct tlb flush's Hyper-V property and use
> >        hv_cpuid_check_and_set() to check the dependency of tlbflush
> >        feature.
> >        - Make new feature work with Hyper-V passthrough mode.
> > ---
> >  docs/hyperv.txt   | 10 ++++++++++
> >  target/i386/cpu.c |  2 ++
> >  target/i386/cpu.h |  1 +
> >  target/i386/kvm.c | 24 ++++++++++++++++++++++++
> >  4 files changed, 37 insertions(+)
> > 
> > diff --git a/docs/hyperv.txt b/docs/hyperv.txt
> > index 8fdf25c829..140a5c7e44 100644
> > --- a/docs/hyperv.txt
> > +++ b/docs/hyperv.txt
> > @@ -184,6 +184,16 @@ enabled.
> >  
> >  Requires: hv-vpindex, hv-synic, hv-time, hv-stimer
> >  
> > +3.18. hv-direct-tlbflush
> > +=======================
> > +Enable direct TLB flush for KVM when it is running as a nested
> > +hypervisor on top Hyper-V. When enabled, TLB flush hypercalls from L2
> > +guests are being passed through to L0 (Hyper-V) for handling. Due to ABI
> > +differences between Hyper-V and KVM hypercalls, L2 guests will not be
> > +able to issue KVM hypercalls (as those could be mishanled by L0
> > +Hyper-V), this requires KVM hypervisor signature to be hidden.
> > +
> > +Requires: hv-tlbflush, -kvm
> >  
> >  4. Development features
> >  ========================
> > diff --git a/target/i386/cpu.c b/target/i386/cpu.c
> > index 44f1bbdcac..7bc7fee512 100644
> > --- a/target/i386/cpu.c
> > +++ b/target/i386/cpu.c
> > @@ -6156,6 +6156,8 @@ static Property x86_cpu_properties[] = {
> >                        HYPERV_FEAT_IPI, 0),
> >      DEFINE_PROP_BIT64("hv-stimer-direct", X86CPU, hyperv_features,
> >                        HYPERV_FEAT_STIMER_DIRECT, 0),
> > +    DEFINE_PROP_BIT64("hv-direct-tlbflush", X86CPU, hyperv_features,
> > +                      HYPERV_FEAT_DIRECT_TLBFLUSH, 0),
> >      DEFINE_PROP_BOOL("hv-passthrough", X86CPU, hyperv_passthrough, false),
> >  
> >      DEFINE_PROP_BOOL("check", X86CPU, check_cpuid, true),
> > diff --git a/target/i386/cpu.h b/target/i386/cpu.h
> > index eaa5395aa5..3cb105f7d6 100644
> > --- a/target/i386/cpu.h
> > +++ b/target/i386/cpu.h
> > @@ -907,6 +907,7 @@ typedef uint64_t FeatureWordArray[FEATURE_WORDS];
> >  #define HYPERV_FEAT_EVMCS               12
> >  #define HYPERV_FEAT_IPI                 13
> >  #define HYPERV_FEAT_STIMER_DIRECT       14
> > +#define HYPERV_FEAT_DIRECT_TLBFLUSH     15
> >  
> >  #ifndef HYPERV_SPINLOCK_NEVER_RETRY
> >  #define HYPERV_SPINLOCK_NEVER_RETRY             0xFFFFFFFF
> > diff --git a/target/i386/kvm.c b/target/i386/kvm.c
> > index 11b9c854b5..043b66ab22 100644
> > --- a/target/i386/kvm.c
> > +++ b/target/i386/kvm.c
> > @@ -900,6 +900,10 @@ static struct {
> >          },
> >          .dependencies = BIT(HYPERV_FEAT_STIMER)
> >      },
> > +    [HYPERV_FEAT_DIRECT_TLBFLUSH] = {
> > +        .desc = "direct paravirtualized TLB flush (hv-direct-tlbflush)",
> > +        .dependencies = BIT(HYPERV_FEAT_TLBFLUSH)
> > +    },
> >  };
> >  
> >  static struct kvm_cpuid2 *try_get_hv_cpuid(CPUState *cs, int max)
> > @@ -1224,6 +1228,7 @@ static int hyperv_handle_properties(CPUState *cs,
> >      r |= hv_cpuid_check_and_set(cs, cpuid, HYPERV_FEAT_EVMCS);
> >      r |= hv_cpuid_check_and_set(cs, cpuid, HYPERV_FEAT_IPI);
> >      r |= hv_cpuid_check_and_set(cs, cpuid, HYPERV_FEAT_STIMER_DIRECT);
> > +    r |= hv_cpuid_check_and_set(cs, cpuid, HYPERV_FEAT_DIRECT_TLBFLUSH);

AFAICS this will turn HYPERV_FEAT_DIRECT_TLBFLUSH on if
hyperv_passthrough is on, so ...

> >  
> >      /* Additional dependencies not covered by kvm_hyperv_properties[] */
> >      if (hyperv_feat_enabled(cpu, HYPERV_FEAT_SYNIC) &&
> > @@ -1243,6 +1248,25 @@ static int hyperv_handle_properties(CPUState *cs,
> >          goto free;
> >      }
> >  
> > +    if (hyperv_feat_enabled(cpu, HYPERV_FEAT_DIRECT_TLBFLUSH) ||
> > +        cpu->hyperv_passthrough) {

... the test for ->hyperv_passthrough is redundant, and ...

> > +        if (!cpu->expose_kvm) {
> > +            r = kvm_vcpu_enable_cap(cs, KVM_CAP_HYPERV_DIRECT_TLBFLUSH, 0, 0);
> > +            if (hyperv_feat_enabled(cpu, HYPERV_FEAT_DIRECT_TLBFLUSH) && r) {

... , more importantly, this will abort QEMU if
HYPERV_FEAT_DIRECT_TLBFLUSH wasn't requested explicitly, but was
activated by ->hyperv_passthrough, and setting the capability failed.  I
think the meaning of hyperv_passthrough is "enable all hyperv features
supported by the KVM", so in this case it looks more correct to just
clear the feature bit and go ahead.

> > +                fprintf(stderr,
> > +                    "Hyper-V %s is not supported by kernel\n",
> > +                    kvm_hyperv_properties[HYPERV_FEAT_DIRECT_TLBFLUSH].desc);
> > +                return -ENOSYS;
> > +            }
> > +        } else if (!cpu->hyperv_passthrough) {
> > +            fprintf(stderr,
> > +                "Hyper-V %s requires KVM hypervisor signature "
> > +                "to be hidden (-kvm).\n",
> > +                kvm_hyperv_properties[HYPERV_FEAT_DIRECT_TLBFLUSH].desc);
> > +            return -ENOSYS;
> > +        }

You reach here if ->expose_kvm && ->hyperv_passthrough, and no
capability is activated, and you go ahead with the feature bit set.
This doesn't look right either.

So in general it should probably look like

    if (hyperv_feat_enabled(HYPERV_FEAT_DIRECT_TLBFLUSH)) {
        if (kvm_vcpu_enable_cap(KVM_CAP_HYPERV_DIRECT_TLBFLUSH)) {
            if (!cpu->hyperv_passthrough) {
                ... report feature unsupported by kernel ...
                return -ENOSYS;
            }
            cpu->hyperv_features &= ~BIT(HYPERV_FEAT_DIRECT_TLBFLUSH);
        } else if (cpu->expose_kvm) {
            ... report conflict ...
            return -ENOSYS;
        }
    }

[Yes, hyperv_passthrough hurts, but you've been warned ;)]

Thanks,
Roman.