KVM: x86/mmu: Reject direct bits in gpa passed to KVM_PRE_FAULT_MEMORY

[PATCH] KVM: x86/mmu: Reject direct bits in gpa passed to KVM_PRE_FAULT_MEMORY

Posted by Paolo Bonzini 4 months ago

Only let userspace pass the same addresses that were used in KVM_SET_USER_MEMORY_REGION
(or KVM_SET_USER_MEMORY_REGION2); gpas in the the upper half of the address space
are an implementation detail of TDX and KVM.

Extracted from a patch by Sean Christopherson <seanjc@google.com>.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
 arch/x86/kvm/mmu/mmu.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
index a4040578b537..4e06e2e89a8f 100644
--- a/arch/x86/kvm/mmu/mmu.c
+++ b/arch/x86/kvm/mmu/mmu.c
@@ -4903,6 +4903,9 @@ long kvm_arch_vcpu_pre_fault_memory(struct kvm_vcpu *vcpu,
 	if (!vcpu->kvm->arch.pre_fault_allowed)
 		return -EOPNOTSUPP;
 
+	if (kvm_is_gfn_alias(vcpu->kvm, gpa_to_gfn(range->gpa)))
+		return -EINVAL;
+
 	/*
 	 * reload is efficient when called repeatedly, so we can do it on
 	 * every iteration.
-- 
2.43.5

Re: [PATCH] KVM: x86/mmu: Reject direct bits in gpa passed to KVM_PRE_FAULT_MEMORY

Posted by Yan Zhao 3 months, 4 weeks ago

On Thu, Jun 12, 2025 at 12:49:43AM -0400, Paolo Bonzini wrote:
> Only let userspace pass the same addresses that were used in KVM_SET_USER_MEMORY_REGION
> (or KVM_SET_USER_MEMORY_REGION2); gpas in the the upper half of the address space
> are an implementation detail of TDX and KVM.
> 
> Extracted from a patch by Sean Christopherson <seanjc@google.com>.
> 
> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
> ---
>  arch/x86/kvm/mmu/mmu.c | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
> index a4040578b537..4e06e2e89a8f 100644
> --- a/arch/x86/kvm/mmu/mmu.c
> +++ b/arch/x86/kvm/mmu/mmu.c
> @@ -4903,6 +4903,9 @@ long kvm_arch_vcpu_pre_fault_memory(struct kvm_vcpu *vcpu,
>  	if (!vcpu->kvm->arch.pre_fault_allowed)
>  		return -EOPNOTSUPP;
>  
> +	if (kvm_is_gfn_alias(vcpu->kvm, gpa_to_gfn(range->gpa)))
> +		return -EINVAL;
> +
Do we need the same check in kvm_vm_ioctl_set_mem_attributes()?

>  	/*
>  	 * reload is efficient when called repeatedly, so we can do it on
>  	 * every iteration.
> -- 
> 2.43.5
>

Re: [PATCH] KVM: x86/mmu: Reject direct bits in gpa passed to KVM_PRE_FAULT_MEMORY

Posted by Sean Christopherson 3 months, 4 weeks ago

On Fri, Jun 13, 2025, Yan Zhao wrote:
> On Thu, Jun 12, 2025 at 12:49:43AM -0400, Paolo Bonzini wrote:
> > Only let userspace pass the same addresses that were used in KVM_SET_USER_MEMORY_REGION
> > (or KVM_SET_USER_MEMORY_REGION2); gpas in the the upper half of the address space
> > are an implementation detail of TDX and KVM.
> > 
> > Extracted from a patch by Sean Christopherson <seanjc@google.com>.
> > 
> > Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
> > ---
> >  arch/x86/kvm/mmu/mmu.c | 3 +++
> >  1 file changed, 3 insertions(+)
> > 
> > diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
> > index a4040578b537..4e06e2e89a8f 100644
> > --- a/arch/x86/kvm/mmu/mmu.c
> > +++ b/arch/x86/kvm/mmu/mmu.c
> > @@ -4903,6 +4903,9 @@ long kvm_arch_vcpu_pre_fault_memory(struct kvm_vcpu *vcpu,
> >  	if (!vcpu->kvm->arch.pre_fault_allowed)
> >  		return -EOPNOTSUPP;
> >  
> > +	if (kvm_is_gfn_alias(vcpu->kvm, gpa_to_gfn(range->gpa)))
> > +		return -EINVAL;
> > +
> Do we need the same check in kvm_vm_ioctl_set_mem_attributes()?

Yeah, we probably should disallow aliases there.  It should be benign?  Because
memslots disallow aliases, and so the aliased gfn entries in the xarray will never
actually be consumed.

Re: [PATCH] KVM: x86/mmu: Reject direct bits in gpa passed to KVM_PRE_FAULT_MEMORY

Posted by Yan Zhao 3 months, 3 weeks ago

On Fri, Jun 13, 2025 at 01:16:38PM -0700, Sean Christopherson wrote:
> On Fri, Jun 13, 2025, Yan Zhao wrote:
> > On Thu, Jun 12, 2025 at 12:49:43AM -0400, Paolo Bonzini wrote:
> > > Only let userspace pass the same addresses that were used in KVM_SET_USER_MEMORY_REGION
> > > (or KVM_SET_USER_MEMORY_REGION2); gpas in the the upper half of the address space
> > > are an implementation detail of TDX and KVM.
> > > 
> > > Extracted from a patch by Sean Christopherson <seanjc@google.com>.
> > > 
> > > Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
> > > ---
> > >  arch/x86/kvm/mmu/mmu.c | 3 +++
> > >  1 file changed, 3 insertions(+)
> > > 
> > > diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
> > > index a4040578b537..4e06e2e89a8f 100644
> > > --- a/arch/x86/kvm/mmu/mmu.c
> > > +++ b/arch/x86/kvm/mmu/mmu.c
> > > @@ -4903,6 +4903,9 @@ long kvm_arch_vcpu_pre_fault_memory(struct kvm_vcpu *vcpu,
> > >  	if (!vcpu->kvm->arch.pre_fault_allowed)
> > >  		return -EOPNOTSUPP;
> > >  
> > > +	if (kvm_is_gfn_alias(vcpu->kvm, gpa_to_gfn(range->gpa)))
> > > +		return -EINVAL;
> > > +
> > Do we need the same check in kvm_vm_ioctl_set_mem_attributes()?
> 
> Yeah, we probably should disallow aliases there.  It should be benign?  Because
> memslots disallow aliases, and so the aliased gfn entries in the xarray will never
> actually be consumed.
Yes, it's benign after this patch.
Userspace may find that setting attribute for an aliased gfn has no effect.
Though there's  a "kvm_mem_is_private(vcpu->kvm, gpa_to_gfn(cr2_or_gpa)" in
kvm_mmu_page_fault(), it's only for KVM_X86_SW_PROTECTED_VM. So it's benign,
just odd :)

Re: [PATCH] KVM: x86/mmu: Reject direct bits in gpa passed to KVM_PRE_FAULT_MEMORY

Posted by Xiaoyao Li 4 months ago

On 6/12/2025 12:49 PM, Paolo Bonzini wrote:
> Only let userspace pass the same addresses that were used in KVM_SET_USER_MEMORY_REGION
> (or KVM_SET_USER_MEMORY_REGION2); gpas in the the upper half of the address space
> are an implementation detail of TDX and KVM.
> 
> Extracted from a patch by Sean Christopherson <seanjc@google.com>.
> 
> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
> ---
>   arch/x86/kvm/mmu/mmu.c | 3 +++
>   1 file changed, 3 insertions(+)
> 
> diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
> index a4040578b537..4e06e2e89a8f 100644
> --- a/arch/x86/kvm/mmu/mmu.c
> +++ b/arch/x86/kvm/mmu/mmu.c
> @@ -4903,6 +4903,9 @@ long kvm_arch_vcpu_pre_fault_memory(struct kvm_vcpu *vcpu,
>   	if (!vcpu->kvm->arch.pre_fault_allowed)
>   		return -EOPNOTSUPP;
>   
> +	if (kvm_is_gfn_alias(vcpu->kvm, gpa_to_gfn(range->gpa)))
> +		return -EINVAL;

Do we need to worry about the case (range->gpa + range->size) becomes alias?

>   	/*
>   	 * reload is efficient when called repeatedly, so we can do it on
>   	 * every iteration.

Re: [PATCH] KVM: x86/mmu: Reject direct bits in gpa passed to KVM_PRE_FAULT_MEMORY

Posted by Paolo Bonzini 4 months ago

On Thu, Jun 12, 2025 at 7:29 AM Xiaoyao Li <xiaoyao.li@intel.com> wrote:
>
> On 6/12/2025 12:49 PM, Paolo Bonzini wrote:
> > Only let userspace pass the same addresses that were used in KVM_SET_USER_MEMORY_REGION
> > (or KVM_SET_USER_MEMORY_REGION2); gpas in the the upper half of the address space
> > are an implementation detail of TDX and KVM.
> >
> > Extracted from a patch by Sean Christopherson <seanjc@google.com>.
> >
> > Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
> > ---
> >   arch/x86/kvm/mmu/mmu.c | 3 +++
> >   1 file changed, 3 insertions(+)
> >
> > diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
> > index a4040578b537..4e06e2e89a8f 100644
> > --- a/arch/x86/kvm/mmu/mmu.c
> > +++ b/arch/x86/kvm/mmu/mmu.c
> > @@ -4903,6 +4903,9 @@ long kvm_arch_vcpu_pre_fault_memory(struct kvm_vcpu *vcpu,
> >       if (!vcpu->kvm->arch.pre_fault_allowed)
> >               return -EOPNOTSUPP;
> >
> > +     if (kvm_is_gfn_alias(vcpu->kvm, gpa_to_gfn(range->gpa)))
> > +             return -EINVAL;
>
> Do we need to worry about the case (range->gpa + range->size) becomes alias?

No, because the function only processes a single page and everything
in the non-aliased part of the address space *can* be prefaulted.

KVM's generic kvm_vcpu_pre_fault_memory() call will see the EINVAL on
a later invocation and will stop processing the part of the request
that is has the shared/direct bit set.

Paolo

>
> >       /*
> >        * reload is efficient when called repeatedly, so we can do it on
> >        * every iteration.
>

Re: [PATCH] KVM: x86/mmu: Reject direct bits in gpa passed to KVM_PRE_FAULT_MEMORY

Posted by Xiaoyao Li 4 months ago

On 6/12/2025 1:34 PM, Paolo Bonzini wrote:
> On Thu, Jun 12, 2025 at 7:29 AM Xiaoyao Li <xiaoyao.li@intel.com> wrote:
>>
>> On 6/12/2025 12:49 PM, Paolo Bonzini wrote:
>>> Only let userspace pass the same addresses that were used in KVM_SET_USER_MEMORY_REGION
>>> (or KVM_SET_USER_MEMORY_REGION2); gpas in the the upper half of the address space
>>> are an implementation detail of TDX and KVM.
>>>
>>> Extracted from a patch by Sean Christopherson <seanjc@google.com>.
>>>
>>> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
>>> ---
>>>    arch/x86/kvm/mmu/mmu.c | 3 +++
>>>    1 file changed, 3 insertions(+)
>>>
>>> diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
>>> index a4040578b537..4e06e2e89a8f 100644
>>> --- a/arch/x86/kvm/mmu/mmu.c
>>> +++ b/arch/x86/kvm/mmu/mmu.c
>>> @@ -4903,6 +4903,9 @@ long kvm_arch_vcpu_pre_fault_memory(struct kvm_vcpu *vcpu,
>>>        if (!vcpu->kvm->arch.pre_fault_allowed)
>>>                return -EOPNOTSUPP;
>>>
>>> +     if (kvm_is_gfn_alias(vcpu->kvm, gpa_to_gfn(range->gpa)))
>>> +             return -EINVAL;
>>
>> Do we need to worry about the case (range->gpa + range->size) becomes alias?
> 
> No, because the function only processes a single page and everything
> in the non-aliased part of the address space *can* be prefaulted.
> 
> KVM's generic kvm_vcpu_pre_fault_memory() call will see the EINVAL on
> a later invocation and will stop processing the part of the request
> that is has the shared/direct bit set.

reasonable..

Reviewed-by: Xiaoyao Li <xiaoyao.li@intel.com>

> Paolo
> 
>>
>>>        /*
>>>         * reload is efficient when called repeatedly, so we can do it on
>>>         * every iteration.
>>
> 
>