RE: [PATCH] KVM: PPC: Book3S HV: Add a capability for enabling secure guests

[Ram Pai <linuxram@us.ibm.com>]

On Fri, Mar 20, 2020 at 10:17:13AM +1100, Paul Mackerras wrote:
> On Thu, Mar 19, 2020 at 12:41:08PM -0700, Ram Pai wrote:
> > On Thu, Mar 19, 2020 at 03:33:01PM +1100, Paul Mackerras wrote:
> [snip]
> > > --- a/arch/powerpc/kvm/powerpc.c
> > > +++ b/arch/powerpc/kvm/powerpc.c
> > > @@ -670,6 +670,11 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
> > >  		     (hv_enabled && cpu_has_feature(CPU_FTR_P9_TM_HV_ASSIST));
> > >  		break;
> > >  #endif
> > > +#if defined(CONFIG_KVM_BOOK3S_HV_POSSIBLE) && defined(CONFIG_PPC_UV)
> > > +	case KVM_CAP_PPC_SECURE_GUEST:
> > > +		r = hv_enabled && !!firmware_has_feature(FW_FEATURE_ULTRAVISOR);
> > 
> > We also need to check if the kvmppc_uvmem_init() has been successfully
> > called and initialized.
> > 
> > 	r = hv_enabled && !!firmware_has_feature(FW_FEATURE_ULTRAVISOR)
> > 		&& kvmppc_uvmem_bitmap;
> 
> Well I can't do that exactly because kvmppc_uvmem_bitmap is in a
> different module (the kvm_hv module, whereas this code is in the kvm
> module), and I wouldn't want to depend on kvmppc_uvmem_bitmap, since
> that's an internal implementation detail.

yes. checking for kvmppc_uvmem_bitmap depends on internal implementation
detail. Its also a loose approximation.  There has to be something
better which can tell, if everything needed to support secure guests, is
available and initialized.

> 
> The firmware_has_feature(FW_FEATURE_ULTRAVISOR) test ultimately
> depends on there being a device tree node with "ibm,ultravisor" in its
> compatible property (see early_init_dt_scan_ultravisor()).  So that
> means there is an ultravisor there.  The cases where that test would
> pass but kvmppc_uvmem_bitmap == NULL would be those where the device
> tree nodes are present but not right, or where the host is so short of
> memory that it couldn't allocate the kvmppc_uvmem_bitmap.  If you
> think those cases are worth worrying about then I will have to devise
> a way to do the test without depending on any symbols from the kvm-hv
> module.

the cases, where incorrect behavior can happen; if we do not have this additional
check, are --

a) zero secure memory in the system.
b) "kvmppc_uvmem" memory region is not defined.
c) the memory region fails to map.
d) kvmppc_uvmem_bitmap allocation failed.

All these are possible to varying level of certainity.

I do not know we should be concerned about these possibilities.
But if we do, than will a patch like this help? compile tested.

------------------
diff --git a/arch/powerpc/include/asm/kvm_book3s_uvmem.h b/arch/powerpc/include/asm/kvm_book3s_uvmem.h
index 5a9834e..643c497 100644
--- a/arch/powerpc/include/asm/kvm_book3s_uvmem.h
+++ b/arch/powerpc/include/asm/kvm_book3s_uvmem.h
@@ -4,6 +4,7 @@
 
 #ifdef CONFIG_PPC_UV
 int kvmppc_uvmem_init(void);
+int kvmppc_uv_enabled(void);
 void kvmppc_uvmem_free(void);
 int kvmppc_uvmem_slot_init(struct kvm *kvm, const struct kvm_memory_slot *slot);
 void kvmppc_uvmem_slot_free(struct kvm *kvm,
@@ -28,6 +29,11 @@ static inline int kvmppc_uvmem_init(void)
 	return 0;
 }
 
+static inline int kvmppc_uv_enabled(void)
+{
+	return 0;
+}
+
 static inline void kvmppc_uvmem_free(void) { }
 
 static inline int
diff --git a/arch/powerpc/kvm/book3s_hv_uvmem.c b/arch/powerpc/kvm/book3s_hv_uvmem.c
index 79b1202..3331ac5 100644
--- a/arch/powerpc/kvm/book3s_hv_uvmem.c
+++ b/arch/powerpc/kvm/book3s_hv_uvmem.c
@@ -804,6 +804,11 @@ int kvmppc_uvmem_init(void)
 	return ret;
 }
 
+int kvmppc_uv_enabled(void)
+{
+	return !kvmppc_uvmem_bitmap;
+}
+
 void kvmppc_uvmem_free(void)
 {
 	memunmap_pages(&kvmppc_uvmem_pgmap);
------------------

> 
> Paul.

-- 
Ram Pai