Re: [RFC PATCH 13/13] KVM: nSVM: Stop bombing the TLB on nested transitions

[Yosry Ahmed <yosry.ahmed@linux.dev>]

On Fri, Feb 28, 2025 at 09:21:54PM -0500, Maxim Levitsky wrote:
> On Wed, 2025-02-05 at 18:24 +0000, Yosry Ahmed wrote:
> > Now that nested TLB flushes are properly tracked with a well-maintained
> > separate ASID for L2 and proper handling of L1's TLB flush requests,
> > drop the unconditional flushes and syncs on nested transitions.
> > 
> > On a Milan machine, an L1 and L2 guests were booted, both with a single
> > vCPU, and pinned to a single physical CPU to maximize TLB collisions. In
> > this setup, the cpuid_rate microbenchmark [1] showed the following
> > changes with this patch:
> > 
> > +--------+--------+-------------------+----------------------+
> > > L0     | L1     | cpuid_rate (base) | cpuid_rate (patched) |
> > +========+========+===================+======================+
> > > NPT    | NPT    | 256621            | 301113 (+17.3%)      |
> > > NPT    | Shadow | 180017            | 203347 (+12.96%)     |
> > > Shadow | Shadow | 177006            | 189150 (+6.86%)      |
> > +--------+--------+-------------------+----------------------+
> > 
> > [1]https://lore.kernel.org/kvm/20231109180646.2963718-1-khorenko@virtuozzo.com/
> > 
> > Signed-off-by: Yosry Ahmed <yosry.ahmed@linux.dev>
> > ---
> >  arch/x86/kvm/svm/nested.c | 7 -------
> >  1 file changed, 7 deletions(-)
> > 
> > diff --git a/arch/x86/kvm/svm/nested.c b/arch/x86/kvm/svm/nested.c
> > index 8e40ff21f7353..45a187d4c23d1 100644
> > --- a/arch/x86/kvm/svm/nested.c
> > +++ b/arch/x86/kvm/svm/nested.c
> > @@ -512,9 +512,6 @@ static void nested_svm_entry_tlb_flush(struct kvm_vcpu *vcpu)
> >  		svm->nested.last_asid = svm->nested.ctl.asid;
> >  		kvm_make_request(KVM_REQ_TLB_FLUSH_GUEST, vcpu);
> >  	}
> > -	/* TODO: optimize unconditional TLB flush/MMU sync */
> > -	kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
> > -	kvm_make_request(KVM_REQ_TLB_FLUSH_CURRENT, vcpu);
> >  }
> >  
> >  static void nested_svm_exit_tlb_flush(struct kvm_vcpu *vcpu)
> > @@ -530,10 +527,6 @@ static void nested_svm_exit_tlb_flush(struct kvm_vcpu *vcpu)
> >  	 */
> >  	if (svm->nested.ctl.tlb_ctl == TLB_CONTROL_FLUSH_ALL_ASID)
> >  		kvm_make_request(KVM_REQ_TLB_FLUSH_GUEST, vcpu);
> > -
> > -	/* TODO: optimize unconditional TLB flush/MMU sync */
> > -	kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
> > -	kvm_make_request(KVM_REQ_TLB_FLUSH_CURRENT, vcpu);
> >  }
> >  
> >  /*
> 
> 
> Assuming that all previous patches are correct this one should work as well.
> 
> However only a very heavy stress testing, including hyperv, windows guests
> of various types, etc can give me confidence that there is no some ugly bug lurking
> somewhere.

I tried booting an L2 and running some workloads like netperf in there.
I also tried booting an L3.

I am planning to try and run some testing with a windows L2 guest. I am
assuming this exercises the hyper-V emulation in L1, which could be
interesting.

I am not sure if I will be able to test more scenarios though,
especially Windows as an L1 (and something else as an L2).

Let me know if you have something specific in mind.

> 
> TLB management can be very tricky, so I can't be 100% sure that I haven't missed something.
> 
> Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>

Thanks!