[PATCH] KVM: VMX: Use wrapper macro ~RMODE_GUEST_OWNED_EFLAGS_BITS directly

[PATCH] KVM: VMX: Use wrapper macro ~RMODE_GUEST_OWNED_EFLAGS_BITS directly

[linmiaohe]

From: Miaohe Lin <linmiaohe@huawei.com>

(X86_EFLAGS_IOPL | X86_EFLAGS_VM) indicates the eflag bits that can not be
owned by realmode guest, i.e. ~RMODE_GUEST_OWNED_EFLAGS_BITS. Use wrapper
macro directly to make it clear and also improve readability.

Signed-off-by: Miaohe Lin <linmiaohe@huawei.com>
---
 arch/x86/kvm/vmx/vmx.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
index 743b81642ce2..9571f8dea016 100644
--- a/arch/x86/kvm/vmx/vmx.c
+++ b/arch/x86/kvm/vmx/vmx.c
@@ -1466,7 +1466,7 @@ void vmx_set_rflags(struct kvm_vcpu *vcpu, unsigned long rflags)
 	vmx->rflags = rflags;
 	if (vmx->rmode.vm86_active) {
 		vmx->rmode.save_rflags = rflags;
-		rflags |= X86_EFLAGS_IOPL | X86_EFLAGS_VM;
+		rflags |= ~RMODE_GUEST_OWNED_EFLAGS_BITS;
 	}
 	vmcs_writel(GUEST_RFLAGS, rflags);
 
@@ -2797,7 +2797,7 @@ static void enter_rmode(struct kvm_vcpu *vcpu)
 	flags = vmcs_readl(GUEST_RFLAGS);
 	vmx->rmode.save_rflags = flags;
 
-	flags |= X86_EFLAGS_IOPL | X86_EFLAGS_VM;
+	flags |= ~RMODE_GUEST_OWNED_EFLAGS_BITS;
 
 	vmcs_writel(GUEST_RFLAGS, flags);
 	vmcs_writel(GUEST_CR4, vmcs_readl(GUEST_CR4) | X86_CR4_VME);
-- 
2.19.1

Re: [PATCH] KVM: VMX: Use wrapper macro ~RMODE_GUEST_OWNED_EFLAGS_BITS directly

[Vitaly Kuznetsov]

linmiaohe <linmiaohe@huawei.com> writes:

> From: Miaohe Lin <linmiaohe@huawei.com>
>
> (X86_EFLAGS_IOPL | X86_EFLAGS_VM) indicates the eflag bits that can not be
> owned by realmode guest, i.e. ~RMODE_GUEST_OWNED_EFLAGS_BITS. Use wrapper
> macro directly to make it clear and also improve readability.
>
> Signed-off-by: Miaohe Lin <linmiaohe@huawei.com>
> ---
>  arch/x86/kvm/vmx/vmx.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> index 743b81642ce2..9571f8dea016 100644
> --- a/arch/x86/kvm/vmx/vmx.c
> +++ b/arch/x86/kvm/vmx/vmx.c
> @@ -1466,7 +1466,7 @@ void vmx_set_rflags(struct kvm_vcpu *vcpu, unsigned long rflags)
>  	vmx->rflags = rflags;
>  	if (vmx->rmode.vm86_active) {
>  		vmx->rmode.save_rflags = rflags;
> -		rflags |= X86_EFLAGS_IOPL | X86_EFLAGS_VM;
> +		rflags |= ~RMODE_GUEST_OWNED_EFLAGS_BITS;
>  	}
>  	vmcs_writel(GUEST_RFLAGS, rflags);
>  
> @@ -2797,7 +2797,7 @@ static void enter_rmode(struct kvm_vcpu *vcpu)
>  	flags = vmcs_readl(GUEST_RFLAGS);
>  	vmx->rmode.save_rflags = flags;
>  
> -	flags |= X86_EFLAGS_IOPL | X86_EFLAGS_VM;
> +	flags |= ~RMODE_GUEST_OWNED_EFLAGS_BITS;
>  
>  	vmcs_writel(GUEST_RFLAGS, flags);
>  	vmcs_writel(GUEST_CR4, vmcs_readl(GUEST_CR4) | X86_CR4_VME);

Double negations are evil, let's define a macro for 'X86_EFLAGS_IOPL |
X86_EFLAGS_VM' instead (completely untested):

diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
index 4ee19fb35cde..d838f93bd6d2 100644
--- a/arch/x86/kvm/vmx/vmx.c
+++ b/arch/x86/kvm/vmx/vmx.c
@@ -139,7 +139,8 @@ module_param_named(preemption_timer, enable_preemption_timer, bool, S_IRUGO);
 #define KVM_PMODE_VM_CR4_ALWAYS_ON (X86_CR4_PAE | X86_CR4_VMXE)
 #define KVM_RMODE_VM_CR4_ALWAYS_ON (X86_CR4_VME | X86_CR4_PAE | X86_CR4_VMXE)
 
-#define RMODE_GUEST_OWNED_EFLAGS_BITS (~(X86_EFLAGS_IOPL | X86_EFLAGS_VM))
+#define RMODE_HOST_OWNED_EFLAGS_BITS (X86_EFLAGS_IOPL | X86_EFLAGS_VM)
+#define RMODE_GUEST_OWNED_EFLAGS_BITS (~RMODE_HOST_OWNED_EFLAGS_BITS)
 
 #define MSR_IA32_RTIT_STATUS_MASK (~(RTIT_STATUS_FILTEREN | \
        RTIT_STATUS_CONTEXTEN | RTIT_STATUS_TRIGGEREN | \
@@ -1468,7 +1469,7 @@ void vmx_set_rflags(struct kvm_vcpu *vcpu, unsigned long rflags)
        vmx->rflags = rflags;
        if (vmx->rmode.vm86_active) {
                vmx->rmode.save_rflags = rflags;
-               rflags |= X86_EFLAGS_IOPL | X86_EFLAGS_VM;
+               rflags |= RMODE_HOST_OWNED_EFLAGS_BITS;
        }
        vmcs_writel(GUEST_RFLAGS, rflags);
 
@@ -2794,7 +2795,7 @@ static void enter_rmode(struct kvm_vcpu *vcpu)
        flags = vmcs_readl(GUEST_RFLAGS);
        vmx->rmode.save_rflags = flags;
 
-       flags |= X86_EFLAGS_IOPL | X86_EFLAGS_VM;
+       flags |= RMODE_HOST_OWNED_EFLAGS_BITS;
 
        vmcs_writel(GUEST_RFLAGS, flags);
        vmcs_writel(GUEST_CR4, vmcs_readl(GUEST_CR4) | X86_CR4_VME);

-- 
Vitaly

Re: [PATCH] KVM: VMX: Use wrapper macro ~RMODE_GUEST_OWNED_EFLAGS_BITS directly

[Paolo Bonzini]

On 05/03/20 03:35, linmiaohe wrote:
> (X86_EFLAGS_IOPL | X86_EFLAGS_VM) indicates the eflag bits that can not be
> owned by realmode guest, i.e. ~RMODE_GUEST_OWNED_EFLAGS_BITS.

... but ~RMODE_GUEST_OWNED_EFLAGS_BITS is the bits that are owned by the
host; they could be 0 or 1 and that's why the code was using
X86_EFLAGS_IOPL | X86_EFLAGS_VM.

I understand where ~RMODE_GUEST_OWNED_EFLAGS_BITS is better than
X86_EFLAGS_IOPL | X86_EFLAGS_VM, but I cannot think of a way to express
it that is the best of both worlds.

Paolo

 Use wrapper
> macro directly to make it clear and also improve readability.
> 
> Signed-off-by: Miaohe Lin <linmiaohe@huawei.com>
> ---
>  arch/x86/kvm/vmx/vmx.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> index 743b81642ce2..9571f8dea016 100644
> --- a/arch/x86/kvm/vmx/vmx.c
> +++ b/arch/x86/kvm/vmx/vmx.c
> @@ -1466,7 +1466,7 @@ void vmx_set_rflags(struct kvm_vcpu *vcpu, unsigned long rflags)
>  	vmx->rflags = rflags;
>  	if (vmx->rmode.vm86_active) {
>  		vmx->rmode.save_rflags = rflags;
> -		rflags |= X86_EFLAGS_IOPL | X86_EFLAGS_VM;
> +		rflags |= ~RMODE_GUEST_OWNED_EFLAGS_BITS;
>  	}
>  	vmcs_writel(GUEST_RFLAGS, rflags);

Re: [PATCH] KVM: VMX: Use wrapper macro ~RMODE_GUEST_OWNED_EFLAGS_BITS directly

[linmiaohe]

Vitaly Kuznetsov <vkuznets@redhat.com> writes:
>linmiaohe <linmiaohe@huawei.com> writes:
>
>> From: Miaohe Lin <linmiaohe@huawei.com>
>>
>>  
>>  	vmcs_writel(GUEST_RFLAGS, flags);
>>  	vmcs_writel(GUEST_CR4, vmcs_readl(GUEST_CR4) | X86_CR4_VME);
>
>Double negations are evil, let's define a macro for 'X86_EFLAGS_IOPL | X86_EFLAGS_VM' instead (completely untested):

You catch the evil guys again. :) But ~RMODE_GUEST_OWNED_EFLAGS_BITS is used by many other func, we should fix them
together. Would try your version, many thanks!

>
>
>-       flags |= X86_EFLAGS_IOPL | X86_EFLAGS_VM;
>+       flags |= RMODE_HOST_OWNED_EFLAGS_BITS;
> 
>        vmcs_writel(GUEST_RFLAGS, flags);
>        vmcs_writel(GUEST_CR4, vmcs_readl(GUEST_CR4) | X86_CR4_VME);
>

Re: [PATCH] KVM: VMX: Use wrapper macro ~RMODE_GUEST_OWNED_EFLAGS_BITS directly

[linmiaohe]

Paolo Bonzini <pbonzini@redhat.com> wrote:
>On 05/03/20 03:35, linmiaohe wrote:
>> (X86_EFLAGS_IOPL | X86_EFLAGS_VM) indicates the eflag bits that can 
>> not be owned by realmode guest, i.e. ~RMODE_GUEST_OWNED_EFLAGS_BITS.
>
>... but ~RMODE_GUEST_OWNED_EFLAGS_BITS is the bits that are owned by the host; they could be 0 or 1 and that's why the code was using X86_EFLAGS_IOPL | X86_EFLAGS_VM.
>
>I understand where ~RMODE_GUEST_OWNED_EFLAGS_BITS is better than X86_EFLAGS_IOPL | X86_EFLAGS_VM, but I cannot think of a way to express it that is the best of both worlds.
>

Define a macro RMODE_HOST_OWNED_EFLAGS_BITS for (X86_EFLAGS_IOPL | X86_EFLAGS_VM) as suggested by Vitaly seems a good way to fix this ? Thanks.

Re: [PATCH] KVM: VMX: Use wrapper macro ~RMODE_GUEST_OWNED_EFLAGS_BITS directly

[Paolo Bonzini]

On 06/03/20 03:17, linmiaohe wrote:
> Define a macro RMODE_HOST_OWNED_EFLAGS_BITS for (X86_EFLAGS_IOPL |
> X86_EFLAGS_VM) as suggested by Vitaly seems a good way to fix this ?
> Thanks.

No, what if a host-owned flag was zero?  I'd just leave it as is.

Paolo

Re: [PATCH] KVM: VMX: Use wrapper macro ~RMODE_GUEST_OWNED_EFLAGS_BITS directly

[Vitaly Kuznetsov]

Paolo Bonzini <pbonzini@redhat.com> writes:

> On 06/03/20 03:17, linmiaohe wrote:
>> Define a macro RMODE_HOST_OWNED_EFLAGS_BITS for (X86_EFLAGS_IOPL |
>> X86_EFLAGS_VM) as suggested by Vitaly seems a good way to fix this ?
>> Thanks.
>
> No, what if a host-owned flag was zero?  I'd just leave it as is.
>

I'm not saying my suggestion was a good idea but honestly I'm failing to
wrap my head around this. The suggested 'RMODE_HOST_OWNED_EFLAGS_BITS'
would just be a define for (X86_EFLAGS_IOPL | X86_EFLAGS_VM) so
technically the patch would just be nop, no?

-- 
Vitaly

Re: [PATCH] KVM: VMX: Use wrapper macro ~RMODE_GUEST_OWNED_EFLAGS_BITS directly

[Paolo Bonzini]

On 06/03/20 10:44, Vitaly Kuznetsov wrote:
>>> Define a macro RMODE_HOST_OWNED_EFLAGS_BITS for (X86_EFLAGS_IOPL |
>>> X86_EFLAGS_VM) as suggested by Vitaly seems a good way to fix this ?
>>> Thanks.
>> No, what if a host-owned flag was zero?  I'd just leave it as is.
>>
> I'm not saying my suggestion was a good idea but honestly I'm failing to
> wrap my head around this. The suggested 'RMODE_HOST_OWNED_EFLAGS_BITS'
> would just be a define for (X86_EFLAGS_IOPL | X86_EFLAGS_VM) so
> technically the patch would just be nop, no?

It would not be a nop for the reader.

Something called RMODE_{GUEST,HOST}_OWNED_EFLAGS_BITS is a mask.  It
tells you nothing about whether those bugs are 0 or 1.  It's just by
chance that all three host-owned EFLAGS bits are 1 while in real mode.
It wouldn't be the case if, for example, we ran the guest using vm86
mode extensions (i.e. setting CR4.VME=1).  Then VIF would be host-owned,
but it wouldn't necessarily be 1.

Paolo

Re: [PATCH] KVM: VMX: Use wrapper macro ~RMODE_GUEST_OWNED_EFLAGS_BITS directly

[Vitaly Kuznetsov]

Paolo Bonzini <pbonzini@redhat.com> writes:

> On 06/03/20 10:44, Vitaly Kuznetsov wrote:
>>>> Define a macro RMODE_HOST_OWNED_EFLAGS_BITS for (X86_EFLAGS_IOPL |
>>>> X86_EFLAGS_VM) as suggested by Vitaly seems a good way to fix this ?
>>>> Thanks.
>>> No, what if a host-owned flag was zero?  I'd just leave it as is.
>>>
>> I'm not saying my suggestion was a good idea but honestly I'm failing to
>> wrap my head around this. The suggested 'RMODE_HOST_OWNED_EFLAGS_BITS'
>> would just be a define for (X86_EFLAGS_IOPL | X86_EFLAGS_VM) so
>> technically the patch would just be nop, no?
>
> It would not be a nop for the reader.
>
> Something called RMODE_{GUEST,HOST}_OWNED_EFLAGS_BITS is a mask.  It
> tells you nothing about whether those bugs are 0 or 1.  It's just by
> chance that all three host-owned EFLAGS bits are 1 while in real mode.
> It wouldn't be the case if, for example, we ran the guest using vm86
> mode extensions (i.e. setting CR4.VME=1).  Then VIF would be host-owned,
> but it wouldn't necessarily be 1.

Got it, it's the name which is causing the confusion, we're using mask
as something different. Make sense, let's keep the code as-is then.

-- 
Vitaly