Re: [RFC PATCH v3 2/2] KVM: s390: Extend the USER_SIGP capability

[David Hildenbrand <david@redhat.com>]

On 10.11.21 21:33, Eric Farman wrote:
> With commit 2444b352c3ac ("KVM: s390: forward most SIGP orders to user
> space") we have a capability that allows the "fast" SIGP orders (as
> defined by the Programming Notes for the SIGNAL PROCESSOR instruction in
> the Principles of Operation) to be handled in-kernel, while all others are
> sent to userspace for processing.
> 
> This works fine but it creates a situation when, for example, a SIGP SENSE
> might return CC1 (STATUS STORED, and status bits indicating the vcpu is
> stopped), when in actuality userspace is still processing a SIGP STOP AND
> STORE STATUS order, and the vcpu is not yet actually stopped. Thus, the
> SIGP SENSE should actually be returning CC2 (busy) instead of CC1.
> 
> To fix this, add another CPU capability, dependent on the USER_SIGP one,
> and two associated IOCTLs. One IOCTL will be used by userspace to mark a
> vcpu "busy" processing a SIGP order, and cause concurrent orders handled
> in-kernel to be returned with CC2 (busy). Another IOCTL will be used by
> userspace to mark the SIGP "finished", and the vcpu free to process
> additional orders.
> 

This looks much cleaner to me, thanks!

[...]

> diff --git a/arch/s390/kvm/kvm-s390.h b/arch/s390/kvm/kvm-s390.h
> index c07a050d757d..54371cede485 100644
> --- a/arch/s390/kvm/kvm-s390.h
> +++ b/arch/s390/kvm/kvm-s390.h
> @@ -82,6 +82,22 @@ static inline int is_vcpu_idle(struct kvm_vcpu *vcpu)
>  	return test_bit(vcpu->vcpu_idx, vcpu->kvm->arch.idle_mask);
>  }
>  
> +static inline bool kvm_s390_vcpu_is_sigp_busy(struct kvm_vcpu *vcpu)
> +{
> +	return (atomic_read(&vcpu->arch.sigp_busy) == 1);

You can drop ()

> +}
> +
> +static inline bool kvm_s390_vcpu_set_sigp_busy(struct kvm_vcpu *vcpu)
> +{
> +	/* Return zero for success, or -EBUSY if another vcpu won */
> +	return (atomic_cmpxchg(&vcpu->arch.sigp_busy, 0, 1) == 0) ? 0 : -EBUSY;

You can drop () as well.

We might not need the -EBUSY semantics after all. User space can just
track if it was set, because it's in charge of setting it.

> +}
> +
> +static inline void kvm_s390_vcpu_clear_sigp_busy(struct kvm_vcpu *vcpu)
> +{
> +	atomic_set(&vcpu->arch.sigp_busy, 0);
> +}
> +
>  static inline int kvm_is_ucontrol(struct kvm *kvm)
>  {
>  #ifdef CONFIG_KVM_S390_UCONTROL
> diff --git a/arch/s390/kvm/sigp.c b/arch/s390/kvm/sigp.c
> index 5ad3fb4619f1..a37496ea6dfa 100644
> --- a/arch/s390/kvm/sigp.c
> +++ b/arch/s390/kvm/sigp.c
> @@ -276,6 +276,10 @@ static int handle_sigp_dst(struct kvm_vcpu *vcpu, u8 order_code,
>  	if (!dst_vcpu)
>  		return SIGP_CC_NOT_OPERATIONAL;
>  
> +	if (kvm_s390_vcpu_is_sigp_busy(dst_vcpu)) {
> +		return SIGP_CC_BUSY;
> +	}

You can drop {}

> +
>  	switch (order_code) {
>  	case SIGP_SENSE:
>  		vcpu->stat.instruction_sigp_sense++;
> @@ -411,6 +415,12 @@ int kvm_s390_handle_sigp(struct kvm_vcpu *vcpu)
>  	if (handle_sigp_order_in_user_space(vcpu, order_code, cpu_addr))
>  		return -EOPNOTSUPP;
>  
> +	/* Check the current vcpu, if it was a target from another vcpu */
> +	if (kvm_s390_vcpu_is_sigp_busy(vcpu)) {
> +		kvm_s390_set_psw_cc(vcpu, SIGP_CC_BUSY);
> +		return 0;
> +	}


I don't think we need this. I think the above (checking the target of a
SIGP order) is sufficient. Or which situation do you have in mind?



I do wonder if we want to make this a kvm_arch_vcpu_ioctl() instead,
essentially just providing a KVM_S390_SET_SIGP_BUSY *and* providing the
order. "order == 0" sets it to !busy. Not that we would need the value
right now, but who knows for what we might reuse that interface in the
future.

Thanks!

-- 
Thanks,

David / dhildenb