[PATCH v3 5/5] RISC-V: Add SBI HSM extension in KVM

[Atish Patra]

On Mon, 2021-10-11 at 14:32 +0000, Sean Christopherson wrote:
> On Mon, Oct 11, 2021, Atish Patra wrote:
> > On Fri, 2021-10-08 at 15:02 +0000, Sean Christopherson wrote:
> > > On Thu, Oct 07, 2021, Atish Patra wrote:
> > > > +       preempt_disable();
> > > > +       loaded = (vcpu->cpu != -1);
> > > > +       if (loaded)
> > > > +               kvm_arch_vcpu_put(vcpu);
> > > 
> > > Oof.  Looks like this pattern was taken from arm64. 
> > 
> > Yes. This part is similar to arm64 because the same race condition
> > can
> > happen in riscv due to save/restore of CSRs during reset.
> > 
> > 
> > >  Is there really no better approach to handling this?  I don't
> > > see anything
> > >  in kvm_riscv_reset_vcpu() that will obviously break if the vCPU
> > > is
> > >  loaded.  If the goal is purely to effect a CSR reset via
> > >  kvm_arch_vcpu_load(), then why not just factor out a helper to
> > > do exactly
> > >  that?
> 
> What about the question here?

Are you suggesting to factor the csr reset part to a different function
?

> 
> > > 
> > > >  
> > > >         memcpy(csr, reset_csr, sizeof(*csr));
> > > >  
> > > > @@ -144,6 +151,11 @@ static void kvm_riscv_reset_vcpu(struct
> > > > kvm_vcpu *vcpu)
> > > >  
> > > >         WRITE_ONCE(vcpu->arch.irqs_pending, 0);
> > > >         WRITE_ONCE(vcpu->arch.irqs_pending_mask, 0);
> > > > +
> > > > +       /* Reset the guest CSRs for hotplug usecase */
> > > > +       if (loaded)
> > > > +               kvm_arch_vcpu_load(vcpu, smp_processor_id());
> > > 
> > > If the preempt shenanigans really have to stay, at least use
> > > get_cpu()/put_cpu().
> > > 
> > 
> > Is there any specific advantage to that ? get_cpu/put_cpu are just
> > macros which calls preempt_disable/preempt_enable.
> > 
> > The only advantage of get_cpu is that it returns the current cpu. 
> > vcpu_load function uses get_cpu because it requires the current cpu
> > id.
> > 
> > However, we don't need that in this case. I am not against changing
> > it
> > to get_cpu/put_cpu. Just wanted to understand the reasoning behind
> > your
> > suggestion.
> 
> It would make the code a bit self-documenting, because AFAICT it
> doesn't truly
> care about being preempted, it cares about keeping the vCPU on the
> correct pCPU.


Sure. I will change it to get_cpu/put_cpu interface.

> 
> > > > +       preempt_enable();
> > > >  }
> > > >  
> > > >  int kvm_arch_vcpu_precreate(struct kvm *kvm, unsigned int id)
> > > > @@ -180,6 +192,13 @@ int kvm_arch_vcpu_create(struct kvm_vcpu
> > > > *vcpu)
> > > >  
> > > >  void kvm_arch_vcpu_postcreate(struct kvm_vcpu *vcpu)
> > > >  {
> > > > +       /**
> > > > +        * vcpu with id 0 is the designated boot cpu.
> > > > +        * Keep all vcpus with non-zero cpu id in power-off
> > > > state
> > > > so that they
> > > > +        * can brought to online using SBI HSM extension.
> > > > +        */
> > > > +       if (vcpu->vcpu_idx != 0)
> > > > +               kvm_riscv_vcpu_power_off(vcpu);
> > > 
> > > Why do this in postcreate?
> > > 
> > 
> > Because we need to absolutely sure that the vcpu is created. It is
> > cleaner in this way rather than doing this here at the end of
> > kvm_arch_vcpu_create. create_vcpu can also fail after
> > kvm_arch_vcpu_create returns.
> 
> But kvm_riscv_vcpu_power_off() doesn't doesn't anything outside of
> the vCPU.  It
> clears vcpu->arch.power_off, makes a request, and kicks the vCPU. 
> None of that
> has side effects to anything else in KVM.  If the vCPU isn't created
> successfully,
> it gets deleted and nothing ever sees that state change.

I am assuming that you are suggesting to add this logic at the end of
the kvm_arch_vcpu_create() instead of kvm_arch_vcpu_postcreate().

vcpu_idx is assigned after kvm_arch_vcpu_create() returns in the
kvm_vm_ioctl_create_vcpu. kvm_arch_vcpu_postcreate() is the arch hookup
after vcpu_idx is assigned.

-- 
Regards,
Atish