Re: [PATCH v2] KVM:SVM: Flush cache only on CPUs running SEV guest

[Tom Lendacky]

On 3/12/24 09:45, Tom Lendacky wrote:
> On 3/6/24 20:14, Zheyun Shen wrote:
> > On AMD CPUs without ensuring cache consistency, each memory page
> > reclamation in an SEV guest triggers a call to wbinvd_on_all_cpus(),
> > thereby affecting the performance of other programs on the host.
> >
> > Typically, an AMD server may have 128 cores or more, while the SEV guest
> > might only utilize 8 of these cores. Meanwhile, host can use qemu-affinity
> > to bind these 8 vCPUs to specific physical CPUs.
> >
> > Therefore, keeping a record of the physical core numbers each time a vCPU
> > runs can help avoid flushing the cache for all CPUs every time.
> >
> > Since the usage of sev_flush_asids() isn't tied to a single VM, we just
> > replace all wbinvd_on_all_cpus() with sev_do_wbinvd() except for that
> > in sev_flush_asids().
> >
> > Signed-off-by: Zheyun Shen <szy0127@xxxxxxxxxxx>
>
> I'm unable to launch my SEV or SEV-ES guests with this patch (haven't 
> tried an SEV-SNP guest, yet). Qemu segfaults at launch.
>
> I'll try to dig into what is happening, but not sure when I'll be able to 
> do that at the moment.

Looks like it's the use of get_cpu() without an associated put_cpu() when 
setting the cpumask. I think what you really want to use is just the cpu 
parameter that is passed into pre_sev_run().


>

> >   void sev_free_vcpu(struct kvm_vcpu *vcpu)
> > @@ -2648,6 +2666,7 @@ void pre_sev_run(struct vcpu_svm *svm, int cpu)
> >       sd->sev_vmcbs[asid] = svm->vmcb;
> >       svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
> >       vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
> > +    cpumask_set_cpu(get_cpu(), sev_get_wbinvd_dirty_mask(svm->vcpu.kvm));

Just use 'cpu' here....  ^

When I do that I'm able to boot an SEV and SEV-ES guest.

I'll wait for the next version, in case there are other changes, before 
running through our CI for more thorough testing than just a single boot.

Thanks,
Tom

> >   }
> >   #define GHCB_SCRATCH_AREA_LIMIT        (16ULL * PAGE_SIZE)