[v2,09/11] KVM: riscv: selftests: Make check_supported arch specific

  check_supported() was used to verify whether a feature/extension was
supported in a guest in the get-reg-list test. Currently this info
can be retrieved through the KVM_CAP_ARM_* API in aarch64, but in
riscv, this info was only exposed through the KVM_GET_ONE_REG on
KVM_REG_RISCV_ISA_EXT pseudo registers.

Signed-off-by: Haibo Xu <haibo1.xu@intel.com>
---
 tools/testing/selftests/kvm/get-reg-list.c | 32 +++++++++++-----------
 1 file changed, 16 insertions(+), 16 deletions(-)
  

On Thu, May 25, 2023 at 03:38:33PM +0800, Haibo Xu wrote:
> check_supported() was used to verify whether a feature/extension was
> supported in a guest in the get-reg-list test. Currently this info
> can be retrieved through the KVM_CAP_ARM_* API in aarch64, but in
> riscv, this info was only exposed through the KVM_GET_ONE_REG on
> KVM_REG_RISCV_ISA_EXT pseudo registers.
> 
> Signed-off-by: Haibo Xu <haibo1.xu@intel.com>
> ---
>  tools/testing/selftests/kvm/get-reg-list.c | 32 +++++++++++-----------
>  1 file changed, 16 insertions(+), 16 deletions(-)
> 
> diff --git a/tools/testing/selftests/kvm/get-reg-list.c b/tools/testing/selftests/kvm/get-reg-list.c
> index f6ad7991a812..f1fc113e9719 100644
> --- a/tools/testing/selftests/kvm/get-reg-list.c
> +++ b/tools/testing/selftests/kvm/get-reg-list.c
> @@ -99,6 +99,20 @@ void __weak print_reg(const char *prefix, __u64 id)
>  }
>  
>  #ifdef __aarch64__
> +static void check_supported(struct vcpu_reg_list *c)
> +{
> +	struct vcpu_reg_sublist *s;
> +
> +	for_each_sublist(c, s) {
> +		if (!s->capability)
> +			continue;

I was going to say that making this function aarch64 shouldn't be
necessary, since riscv leaves capability set to zero and this function
doesn't do anything, but then looking ahead I see riscv is abusing
capability by putting isa extensions in it. IMO, capability should
only be set to KVM_CAP_* values. Since riscv doesn't use it, then it
should be left zero.

If we're going to abuse something, then I'd rather abuse the 'feature'
member, but since it's only an int (not an unsigned long), then let's
just add an 'unsigned long extension' member.

Then, the finalize_vcpu() call can be moved back to run_test(), from
aarch64's vcpu_config_get_vcpu(). Both aarch64 and riscv will call it
right after vcpu_config_get_vcpu() and the riscv version of it will
do what your current riscv check_supported() is doing, using the
new 'extension' member instead of 'capability'.

And this patch gets dropped.

Thanks,
drew

> +
> +		__TEST_REQUIRE(kvm_has_cap(s->capability),
> +			       "%s: %s not available, skipping tests\n",
> +			       config_name(c), s->name);
> +	}
> +}
> +
>  static void prepare_vcpu_init(struct vcpu_reg_list *c, struct kvm_vcpu_init *init)
>  {
>  	struct vcpu_reg_sublist *s;
> @@ -126,6 +140,8 @@ static struct kvm_vcpu *vcpu_config_get_vcpu(struct vcpu_reg_list *c, struct kvm
>  	struct kvm_vcpu_init init = { .target = -1, };
>  	struct kvm_vcpu *vcpu;
>  
> +	check_supported(c);
> +
>  	prepare_vcpu_init(c, &init);
>  	vcpu = __vm_vcpu_add(vm, 0);
>  	aarch64_vcpu_setup(vcpu, &init);
> @@ -140,20 +156,6 @@ static struct kvm_vcpu *vcpu_config_get_vcpu(struct vcpu_reg_list *c, struct kvm
>  }
>  #endif
>  
> -static void check_supported(struct vcpu_reg_list *c)
> -{
> -	struct vcpu_reg_sublist *s;
> -
> -	for_each_sublist(c, s) {
> -		if (!s->capability)
> -			continue;
> -
> -		__TEST_REQUIRE(kvm_has_cap(s->capability),
> -			       "%s: %s not available, skipping tests\n",
> -			       config_name(c), s->name);
> -	}
> -}
> -
>  static bool print_list;
>  static bool print_filtered;
>  
> @@ -165,8 +167,6 @@ static void run_test(struct vcpu_reg_list *c)
>  	struct kvm_vm *vm;
>  	struct vcpu_reg_sublist *s;
>  
> -	check_supported(c);
> -
>  	vm = vm_create_barebones();
>  	vcpu = vcpu_config_get_vcpu(c, vm);
>  
> -- 
> 2.34.1
>
  

On Fri, May 26, 2023 at 12:40 AM Andrew Jones <ajones@ventanamicro.com> wrote:
>
> On Thu, May 25, 2023 at 03:38:33PM +0800, Haibo Xu wrote:
> > check_supported() was used to verify whether a feature/extension was
> > supported in a guest in the get-reg-list test. Currently this info
> > can be retrieved through the KVM_CAP_ARM_* API in aarch64, but in
> > riscv, this info was only exposed through the KVM_GET_ONE_REG on
> > KVM_REG_RISCV_ISA_EXT pseudo registers.
> >
> > Signed-off-by: Haibo Xu <haibo1.xu@intel.com>
> > ---
> >  tools/testing/selftests/kvm/get-reg-list.c | 32 +++++++++++-----------
> >  1 file changed, 16 insertions(+), 16 deletions(-)
> >
> > diff --git a/tools/testing/selftests/kvm/get-reg-list.c b/tools/testing/selftests/kvm/get-reg-list.c
> > index f6ad7991a812..f1fc113e9719 100644
> > --- a/tools/testing/selftests/kvm/get-reg-list.c
> > +++ b/tools/testing/selftests/kvm/get-reg-list.c
> > @@ -99,6 +99,20 @@ void __weak print_reg(const char *prefix, __u64 id)
> >  }
> >
> >  #ifdef __aarch64__
> > +static void check_supported(struct vcpu_reg_list *c)
> > +{
> > +     struct vcpu_reg_sublist *s;
> > +
> > +     for_each_sublist(c, s) {
> > +             if (!s->capability)
> > +                     continue;
>
> I was going to say that making this function aarch64 shouldn't be
> necessary, since riscv leaves capability set to zero and this function
> doesn't do anything, but then looking ahead I see riscv is abusing
> capability by putting isa extensions in it. IMO, capability should
> only be set to KVM_CAP_* values. Since riscv doesn't use it, then it
> should be left zero.
>
> If we're going to abuse something, then I'd rather abuse the 'feature'
> member, but since it's only an int (not an unsigned long), then let's
> just add an 'unsigned long extension' member.
>

Good idea!

For the new 'extension' member in riscv, I think its use case should be
identical to the 'feature' member in aarch64(KVM_RISCV_ISA_EXT_F
was similar to KVM_ARM_VCPU_SVE)? If so, I think we can just reuse
the 'feature' member since the data type was not a big deal.

> Then, the finalize_vcpu() call can be moved back to run_test(), from
> aarch64's vcpu_config_get_vcpu(). Both aarch64 and riscv will call it
> right after vcpu_config_get_vcpu() and the riscv version of it will
> do what your current riscv check_supported() is doing, using the
> new 'extension' member instead of 'capability'.
>
> And this patch gets dropped.
>
> Thanks,
> drew
>
> > +
> > +             __TEST_REQUIRE(kvm_has_cap(s->capability),
> > +                            "%s: %s not available, skipping tests\n",
> > +                            config_name(c), s->name);
> > +     }
> > +}
> > +
> >  static void prepare_vcpu_init(struct vcpu_reg_list *c, struct kvm_vcpu_init *init)
> >  {
> >       struct vcpu_reg_sublist *s;
> > @@ -126,6 +140,8 @@ static struct kvm_vcpu *vcpu_config_get_vcpu(struct vcpu_reg_list *c, struct kvm
> >       struct kvm_vcpu_init init = { .target = -1, };
> >       struct kvm_vcpu *vcpu;
> >
> > +     check_supported(c);
> > +
> >       prepare_vcpu_init(c, &init);
> >       vcpu = __vm_vcpu_add(vm, 0);
> >       aarch64_vcpu_setup(vcpu, &init);
> > @@ -140,20 +156,6 @@ static struct kvm_vcpu *vcpu_config_get_vcpu(struct vcpu_reg_list *c, struct kvm
> >  }
> >  #endif
> >
> > -static void check_supported(struct vcpu_reg_list *c)
> > -{
> > -     struct vcpu_reg_sublist *s;
> > -
> > -     for_each_sublist(c, s) {
> > -             if (!s->capability)
> > -                     continue;
> > -
> > -             __TEST_REQUIRE(kvm_has_cap(s->capability),
> > -                            "%s: %s not available, skipping tests\n",
> > -                            config_name(c), s->name);
> > -     }
> > -}
> > -
> >  static bool print_list;
> >  static bool print_filtered;
> >
> > @@ -165,8 +167,6 @@ static void run_test(struct vcpu_reg_list *c)
> >       struct kvm_vm *vm;
> >       struct vcpu_reg_sublist *s;
> >
> > -     check_supported(c);
> > -
> >       vm = vm_create_barebones();
> >       vcpu = vcpu_config_get_vcpu(c, vm);
> >
> > --
> > 2.34.1
> >
  

On Fri, May 26, 2023 at 03:50:32PM +0800, Haibo Xu wrote:
> On Fri, May 26, 2023 at 12:40 AM Andrew Jones <ajones@ventanamicro.com> wrote:
> >
> > On Thu, May 25, 2023 at 03:38:33PM +0800, Haibo Xu wrote:
> > > check_supported() was used to verify whether a feature/extension was
> > > supported in a guest in the get-reg-list test. Currently this info
> > > can be retrieved through the KVM_CAP_ARM_* API in aarch64, but in
> > > riscv, this info was only exposed through the KVM_GET_ONE_REG on
> > > KVM_REG_RISCV_ISA_EXT pseudo registers.
> > >
> > > Signed-off-by: Haibo Xu <haibo1.xu@intel.com>
> > > ---
> > >  tools/testing/selftests/kvm/get-reg-list.c | 32 +++++++++++-----------
> > >  1 file changed, 16 insertions(+), 16 deletions(-)
> > >
> > > diff --git a/tools/testing/selftests/kvm/get-reg-list.c b/tools/testing/selftests/kvm/get-reg-list.c
> > > index f6ad7991a812..f1fc113e9719 100644
> > > --- a/tools/testing/selftests/kvm/get-reg-list.c
> > > +++ b/tools/testing/selftests/kvm/get-reg-list.c
> > > @@ -99,6 +99,20 @@ void __weak print_reg(const char *prefix, __u64 id)
> > >  }
> > >
> > >  #ifdef __aarch64__
> > > +static void check_supported(struct vcpu_reg_list *c)
> > > +{
> > > +     struct vcpu_reg_sublist *s;
> > > +
> > > +     for_each_sublist(c, s) {
> > > +             if (!s->capability)
> > > +                     continue;
> >
> > I was going to say that making this function aarch64 shouldn't be
> > necessary, since riscv leaves capability set to zero and this function
> > doesn't do anything, but then looking ahead I see riscv is abusing
> > capability by putting isa extensions in it. IMO, capability should
> > only be set to KVM_CAP_* values. Since riscv doesn't use it, then it
> > should be left zero.
> >
> > If we're going to abuse something, then I'd rather abuse the 'feature'
> > member, but since it's only an int (not an unsigned long), then let's
> > just add an 'unsigned long extension' member.
> >
> 
> Good idea!
> 
> For the new 'extension' member in riscv, I think its use case should be
> identical to the 'feature' member in aarch64(KVM_RISCV_ISA_EXT_F
> was similar to KVM_ARM_VCPU_SVE)? If so, I think we can just reuse
> the 'feature' member since the data type was not a big deal.

You're right. An int is fine for the isa extension index, which is all we
need to represent.

Thanks,
drew
  

On Fri, May 26, 2023 at 4:44 PM Andrew Jones <ajones@ventanamicro.com> wrote:
>
> On Fri, May 26, 2023 at 03:50:32PM +0800, Haibo Xu wrote:
> > On Fri, May 26, 2023 at 12:40 AM Andrew Jones <ajones@ventanamicro.com> wrote:
> > >
> > > On Thu, May 25, 2023 at 03:38:33PM +0800, Haibo Xu wrote:
> > > > check_supported() was used to verify whether a feature/extension was
> > > > supported in a guest in the get-reg-list test. Currently this info
> > > > can be retrieved through the KVM_CAP_ARM_* API in aarch64, but in
> > > > riscv, this info was only exposed through the KVM_GET_ONE_REG on
> > > > KVM_REG_RISCV_ISA_EXT pseudo registers.
> > > >
> > > > Signed-off-by: Haibo Xu <haibo1.xu@intel.com>
> > > > ---
> > > >  tools/testing/selftests/kvm/get-reg-list.c | 32 +++++++++++-----------
> > > >  1 file changed, 16 insertions(+), 16 deletions(-)
> > > >
> > > > diff --git a/tools/testing/selftests/kvm/get-reg-list.c b/tools/testing/selftests/kvm/get-reg-list.c
> > > > index f6ad7991a812..f1fc113e9719 100644
> > > > --- a/tools/testing/selftests/kvm/get-reg-list.c
> > > > +++ b/tools/testing/selftests/kvm/get-reg-list.c
> > > > @@ -99,6 +99,20 @@ void __weak print_reg(const char *prefix, __u64 id)
> > > >  }
> > > >
> > > >  #ifdef __aarch64__
> > > > +static void check_supported(struct vcpu_reg_list *c)
> > > > +{
> > > > +     struct vcpu_reg_sublist *s;
> > > > +
> > > > +     for_each_sublist(c, s) {
> > > > +             if (!s->capability)
> > > > +                     continue;
> > >
> > > I was going to say that making this function aarch64 shouldn't be
> > > necessary, since riscv leaves capability set to zero and this function
> > > doesn't do anything, but then looking ahead I see riscv is abusing
> > > capability by putting isa extensions in it. IMO, capability should
> > > only be set to KVM_CAP_* values. Since riscv doesn't use it, then it
> > > should be left zero.
> > >
> > > If we're going to abuse something, then I'd rather abuse the 'feature'
> > > member, but since it's only an int (not an unsigned long), then let's
> > > just add an 'unsigned long extension' member.
> > >
> >
> > Good idea!
> >
> > For the new 'extension' member in riscv, I think its use case should be
> > identical to the 'feature' member in aarch64(KVM_RISCV_ISA_EXT_F
> > was similar to KVM_ARM_VCPU_SVE)? If so, I think we can just reuse
> > the 'feature' member since the data type was not a big deal.
>
> You're right. An int is fine for the isa extension index, which is all we
> need to represent.
>
> Thanks,
> drew

Thanks for the suggestion! I will include the change in v3 soon.