Re: [RFC PATCH v2 07/19] RISC-V: KVM: Implement KVM_GET_ONE_REG/KVM_SET_ONE_REG ioctls

From: Paolo Bonzini
Date: Mon Aug 05 2019 - 03:10:37 EST

On 05/08/19 08:55, Anup Patel wrote:
> On Fri, Aug 2, 2019 at 2:33 PM Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
>> On 02/08/19 09:47, Anup Patel wrote:
>>> + if (reg_num == KVM_REG_RISCV_CSR_REG(sip))
>>> + kvm_riscv_vcpu_flush_interrupts(vcpu, false);
>> Not updating the vsip CSR here can cause an interrupt to be lost, if the
>> next call to kvm_riscv_vcpu_flush_interrupts finds a zero mask.
> Thanks for catching this issue. I will address it in v3.
> If we think more on similar lines then we also need to handle the case
> where Guest VCPU had pending interrupts and we suddenly stopped it
> for Guest migration. In this case, we would eventually use SET_ONE_REG
> ioctl on destination Host which should set vsip_shadow instead of vsip so
> that we force update HW after resuming Guest VCPU on destination host.

I think it's simpler than that.

vcpu->vsip_shadow is just the current value of CSR_VSIP so that you do
not need to update it unconditionally on every vmentry. That is,
kvm_vcpu_arch_load should do

csr_write(CSR_VSIP, vcpu->arch.guest_csr.vsip);
vcpu->vsip_shadow = vcpu->arch.guest_csr.vsip;

while every other write can go through kvm_riscv_update_vsip. But
vsip_shadow is completely disconnected from SET_ONE_REG; SET_ONE_REG can
just write vcpu->arch.guest_csr.vsip and clear irqs_pending_mask, the
next entry will write CSR_VSIP and vsip_shadow if needed.

In fact, instead of placing it in kvm_vcpu, vsip_shadow could be a
percpu variable; on hardware_enable you write 0 to both vsip_shadow and
CSR_VSIP, and then kvm_arch_vcpu_load does not have to touch CSR_VSIP at
all (only kvm_riscv_vcpu_flush_interrupts). I think this makes the
purpose of vsip_shadow even clearer, so I highly suggest doing that.

>> You could add a new field vcpu->vsip_shadow that is updated every time
>> CSR_VSIP is written (including kvm_arch_vcpu_load) with a function like
>> void kvm_riscv_update_vsip(struct kvm_vcpu *vcpu)
>> {
>> if (vcpu->vsip_shadow != vcpu->arch.guest_csr.vsip) {
>> csr_write(CSR_VSIP, vcpu->arch.guest_csr.vsip);
>> vcpu->vsip_shadow = vcpu->arch.guest_csr.vsip;
>> }
>> }
>> And just call this unconditionally from kvm_vcpu_ioctl_run. The cost is
>> just a memory load per VS-mode entry, it should hardly be measurable.
> I think we can do this at start of kvm_riscv_vcpu_flush_interrupts() as well.

Did you mean at the end? (That is, after modifying
vcpu->arch.guest_csr.vsip based on mask and val). With the above switch
to percpu, the only write of CSR_VSIP and vsip_shadow should be in
kvm_riscv_vcpu_flush_interrupts, which in turn is only called from