[RFC,v7,03/64] KVM: SVM: Advertise private memory support to KVM

From: Nikunj A Dadhania <nikunj@amd.com>

KVM should use private memory for guests that have upm_mode flag set.

Add a kvm_x86_ops hook for determining UPM support that accounts for
this situation by only enabling UPM test mode in the case of non-SEV
guests.

Signed-off-by: Nikunj A Dadhania <nikunj@amd.com>
[mdr: add x86 hook for determining restricted/private memory support]
Signed-off-by: Michael Roth <michael.roth@amd.com>
---
 arch/x86/include/asm/kvm-x86-ops.h |  1 +
 arch/x86/include/asm/kvm_host.h    |  1 +
 arch/x86/kvm/svm/svm.c             | 10 ++++++++++
 arch/x86/kvm/x86.c                 |  8 ++++++++
 4 files changed, 20 insertions(+)

On Wed, Dec 14, 2022 at 01:39:55PM -0600, Michael Roth wrote:
> +       bool (*private_mem_enabled)(struct kvm *kvm);

This looks like a function returning boolean to me. IOW, you can
simplify this to:

diff --git a/arch/x86/include/asm/kvm-x86-ops.h b/arch/x86/include/asm/kvm-x86-ops.h
index 82ba4a564e58..4449aeff0dff 100644
--- a/arch/x86/include/asm/kvm-x86-ops.h
+++ b/arch/x86/include/asm/kvm-x86-ops.h
@@ -129,6 +129,7 @@ KVM_X86_OP(msr_filter_changed)
 KVM_X86_OP(complete_emulated_msr)
 KVM_X86_OP(vcpu_deliver_sipi_vector)
 KVM_X86_OP_OPTIONAL_RET0(vcpu_get_apicv_inhibit_reasons);
+KVM_X86_OP_OPTIONAL_RET0(private_mem_enabled);
 
 #undef KVM_X86_OP
 #undef KVM_X86_OP_OPTIONAL
diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index 1da0474edb2d..1b4b89ddeb55 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -1574,6 +1574,7 @@ struct kvm_x86_ops {
 
 	void (*load_mmu_pgd)(struct kvm_vcpu *vcpu, hpa_t root_hpa,
 			     int root_level);
+	bool (*private_mem_enabled)(struct kvm *kvm);
 
 	bool (*has_wbinvd_exit)(void);
 
diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
index ce362e88a567..73b780fa4653 100644
--- a/arch/x86/kvm/svm/svm.c
+++ b/arch/x86/kvm/svm/svm.c
@@ -4680,6 +4680,14 @@ static int svm_vm_init(struct kvm *kvm)
 	return 0;
 }
 
+static bool svm_private_mem_enabled(struct kvm *kvm)
+{
+	if (sev_guest(kvm))
+		return kvm->arch.upm_mode;
+
+	return IS_ENABLED(CONFIG_HAVE_KVM_PRIVATE_MEM_TESTING);
+}
+
 static struct kvm_x86_ops svm_x86_ops __initdata = {
 	.name = "kvm_amd",
 
@@ -4760,6 +4768,8 @@ static struct kvm_x86_ops svm_x86_ops __initdata = {
 
 	.vcpu_after_set_cpuid = svm_vcpu_after_set_cpuid,
 
+	.private_mem_enabled = svm_private_mem_enabled,
+
 	.has_wbinvd_exit = svm_has_wbinvd_exit,
 
 	.get_l2_tsc_offset = svm_get_l2_tsc_offset,
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 823646d601db..9a1ca59d36a4 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -12556,6 +12556,11 @@ void __user * __x86_set_memory_region(struct kvm *kvm, int id, gpa_t gpa,
 }
 EXPORT_SYMBOL_GPL(__x86_set_memory_region);
 
+bool kvm_arch_has_private_mem(struct kvm *kvm)
+{
+	return static_call(kvm_x86_private_mem_enabled)(kvm);
+}
+
 void kvm_arch_pre_destroy_vm(struct kvm *kvm)
 {
 	kvm_mmu_pre_destroy_vm(kvm);

On Wed, Jan 04, 2023 at 08:14:19PM -0600, Michael Roth wrote:
> Maybe that's not actually enforced, by it seems awkward to try to use a
> bool return instead. At least for KVM_X86_OP_OPTIONAL_RET0().

I don't see there being a problem/restriction for bool functions, see

5be2226f417d ("KVM: x86: allow defining return-0 static calls")

and __static_call_return0() returns a long which, if you wanna interpret as
bool, works too as "false".

I still need to disassemble and single-step through a static_call to see what
all that magic does in detail, to be sure.

> However, we could just use KVM_X86_OP() to declare it so we can cleanly
> use a function that returns bool, and then we just need to do:
> 
>   bool kvm_arch_has_private_mem(struct kvm *kvm)
>   {
>           if (kvm_x86_ops.private_mem_enabled)
>                   return static_call(kvm_x86_private_mem_enabled)(kvm);

That would be defeating the whole purpose of static calls, AFAICT, as you're
testing the pointer. Might as well leave it be a normal function pointer then.

> On a separate topic though, at a high level, this hook is basically a way
> for platform-specific code to tell generic KVM code that private memslots
> are supported by overriding the kvm_arch_has_private_mem() weak
> reference. In this case the AMD platform is using using kvm->arch.upm_mode
> flag to convey that, which is in turn set by the
> KVM_CAP_UNMAPPED_PRIVATE_MEMORY introduced in this series.
> 
> But if, as I suggested in response to your PATCH 2 comments, we drop
> KVM_CAP_UNAMMPED_PRIVATE_MEMORY in favor of
> KVM_SET_SUPPORTED_MEMORY_ATTRIBUTES ioctl to enable "UPM mode" in SEV/SNP
> code, then we need to rethink things a bit, since KVM_SET_MEMORY_ATTRIBUTES
> in-part relies on kvm_arch_has_private_mem() to determine what flags are
> supported, whereas SEV/SNP code would be using what was set by
> KVM_SET_MEMORY_ATTRIBUTES to determine the return value in
> kvm_arch_has_private_mem().
> 
> So, for AMD, the return value of kvm_arch_has_private_mem() needs to rely
> on something else. Maybe the logic can just be:
> 
>   bool svm_private_mem_enabled(struct kvm *kvm)
>   {
>     return sev_enabled(kvm) || sev_snp_enabled(kvm)

I haven't followed the whole discussion in detail but this means that SEV/SNP
*means* UPM. I.e., no SEV/SNP without UPM, correct? I guess that's the final
thing you guys decided to do ...

Thx.

On Thu, Jan 05, 2023 at 12:17:41PM -0600, Michael Roth wrote:
> In the case of SEV, it would still be up to userspace whether or not it
> actually wants to make use of UPM functionality like KVM_SET_MEMORY_ATTRIBUTES
> and private memslots. Otherwise, to maintain backward-compatibility, 
> userspace can do things as it has always done and continue running SEV without
> relying on private memslots/KVM_SET_MEMORY_ATTRIBUTES or any of the new ioctls.
> 
> For SNP however it is required that userspace uses/implements UPM
> functionality.

Makes sense to me.

On Wed, Jan 04, 2023 at 08:14:19PM -0600, Michael Roth wrote:
> On Fri, Dec 23, 2022 at 05:56:50PM +0100, Borislav Petkov wrote:
> > On Wed, Dec 14, 2022 at 01:39:55PM -0600, Michael Roth wrote:
> > > +       bool (*private_mem_enabled)(struct kvm *kvm);
> > 
> > This looks like a function returning boolean to me. IOW, you can
> > simplify this to:
> 
> The semantics and existing uses of KVM_X86_OP_OPTIONAL_RET0() gave me the
> impression it needed to return an integer value, since by default if a
> platform doesn't implement the op it would "return 0", and so could
> still be called unconditionally.
> 
> Maybe that's not actually enforced, by it seems awkward to try to use a
> bool return instead. At least for KVM_X86_OP_OPTIONAL_RET0().
> 
> However, we could just use KVM_X86_OP() to declare it so we can cleanly
> use a function that returns bool, and then we just need to do:
> 
>   bool kvm_arch_has_private_mem(struct kvm *kvm)
>   {
>           if (kvm_x86_ops.private_mem_enabled)
>                   return static_call(kvm_x86_private_mem_enabled)(kvm);

I guess this is missing:

        return false;

>   }
>     
> instead of relying on default return value. So I'll take that approach
> and adopt your other suggested changes.
> 
> ...
> 
> On a separate topic though, at a high level, this hook is basically a way
> for platform-specific code to tell generic KVM code that private memslots
> are supported by overriding the kvm_arch_has_private_mem() weak
> reference. In this case the AMD platform is using using kvm->arch.upm_mode
> flag to convey that, which is in turn set by the
> KVM_CAP_UNMAPPED_PRIVATE_MEMORY introduced in this series.
> 
> But if, as I suggested in response to your PATCH 2 comments, we drop
> KVM_CAP_UNAMMPED_PRIVATE_MEMORY in favor of
> KVM_SET_SUPPORTED_MEMORY_ATTRIBUTES ioctl to enable "UPM mode" in SEV/SNP
> code, then we need to rethink things a bit, since KVM_SET_MEMORY_ATTRIBUTES
> in-part relies on kvm_arch_has_private_mem() to determine what flags are
> supported, whereas SEV/SNP code would be using what was set by
> KVM_SET_MEMORY_ATTRIBUTES to determine the return value in
> kvm_arch_has_private_mem().

Does this mean that internal calls to  kvm_vm_set_region_attr() will
cease to exist, and it will rely for user space to use the ioctl
properly instead?

> So, for AMD, the return value of kvm_arch_has_private_mem() needs to rely
> on something else. Maybe the logic can just be:
> 
>   bool svm_private_mem_enabled(struct kvm *kvm)
>   {
>     return sev_enabled(kvm) || sev_snp_enabled(kvm)
>   }
> 
> (at least in the context of this patchset where UPM support is added for
> both SEV and SNP).
> 
> So I'll plan to make that change as well.
> 
> -Mike
> 
> > 
> > diff --git a/arch/x86/include/asm/kvm-x86-ops.h b/arch/x86/include/asm/kvm-x86-ops.h
> > index 82ba4a564e58..4449aeff0dff 100644
> > --- a/arch/x86/include/asm/kvm-x86-ops.h
> > +++ b/arch/x86/include/asm/kvm-x86-ops.h
> > @@ -129,6 +129,7 @@ KVM_X86_OP(msr_filter_changed)
> >  KVM_X86_OP(complete_emulated_msr)
> >  KVM_X86_OP(vcpu_deliver_sipi_vector)
> >  KVM_X86_OP_OPTIONAL_RET0(vcpu_get_apicv_inhibit_reasons);
> > +KVM_X86_OP_OPTIONAL_RET0(private_mem_enabled);
> >  
> >  #undef KVM_X86_OP
> >  #undef KVM_X86_OP_OPTIONAL
> > diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
> > index 1da0474edb2d..1b4b89ddeb55 100644
> > --- a/arch/x86/include/asm/kvm_host.h
> > +++ b/arch/x86/include/asm/kvm_host.h
> > @@ -1574,6 +1574,7 @@ struct kvm_x86_ops {
> >  
> >  	void (*load_mmu_pgd)(struct kvm_vcpu *vcpu, hpa_t root_hpa,
> >  			     int root_level);
> > +	bool (*private_mem_enabled)(struct kvm *kvm);
> >  
> >  	bool (*has_wbinvd_exit)(void);
> >  
> > diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
> > index ce362e88a567..73b780fa4653 100644
> > --- a/arch/x86/kvm/svm/svm.c
> > +++ b/arch/x86/kvm/svm/svm.c
> > @@ -4680,6 +4680,14 @@ static int svm_vm_init(struct kvm *kvm)
> >  	return 0;
> >  }
> >  
> > +static bool svm_private_mem_enabled(struct kvm *kvm)
> > +{
> > +	if (sev_guest(kvm))
> > +		return kvm->arch.upm_mode;
> > +
> > +	return IS_ENABLED(CONFIG_HAVE_KVM_PRIVATE_MEM_TESTING);
> > +}
> > +
> >  static struct kvm_x86_ops svm_x86_ops __initdata = {
> >  	.name = "kvm_amd",
> >  
> > @@ -4760,6 +4768,8 @@ static struct kvm_x86_ops svm_x86_ops __initdata = {
> >  
> >  	.vcpu_after_set_cpuid = svm_vcpu_after_set_cpuid,
> >  
> > +	.private_mem_enabled = svm_private_mem_enabled,
> > +
> >  	.has_wbinvd_exit = svm_has_wbinvd_exit,
> >  
> >  	.get_l2_tsc_offset = svm_get_l2_tsc_offset,
> > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > index 823646d601db..9a1ca59d36a4 100644
> > --- a/arch/x86/kvm/x86.c
> > +++ b/arch/x86/kvm/x86.c
> > @@ -12556,6 +12556,11 @@ void __user * __x86_set_memory_region(struct kvm *kvm, int id, gpa_t gpa,
> >  }
> >  EXPORT_SYMBOL_GPL(__x86_set_memory_region);
> >  
> > +bool kvm_arch_has_private_mem(struct kvm *kvm)
> > +{
> > +	return static_call(kvm_x86_private_mem_enabled)(kvm);
> > +}
> > +
> >  void kvm_arch_pre_destroy_vm(struct kvm *kvm)
> >  {
> >  	kvm_mmu_pre_destroy_vm(kvm);
> > 
> > -- 
> > Regards/Gruss,
> >     Boris.
> > 
> > https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fpeople.kernel.org%2Ftglx%2Fnotes-about-netiquette&data=05%7C01%7Cmichael.roth%40amd.com%7C319e89ce555a46eace4d08dae506b51a%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C638074114318137471%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=aG11K7va1BhemwlKCKKdcIXEwXGUzImYL%2BZ9%2FQ7XToI%3D&reserved=0

BR, Jarkko