[v3] x86/vmx: Avoid pausing on HVM_PARAM_IDENT_PT in additional cases

[PATCH v3] x86/vmx: Avoid pausing on HVM_PARAM_IDENT_PT in additional cases

Posted by Teddy Astie 2 weeks, 1 day ago

When settings HVM_PARAM_IDENT_PT, skip domain pausing when :
- there is no vcpu
- unrestricted guest capability is used

Signed-off-by: Teddy Astie <teddy.astie@vates.tech>
---
Regarding checking for hvm_paging_enabled(v) (proposed in v2 review), we can't
as hvm_paging_enabled() is vCPU specific, and vCPU 0 may have a different value
to another one.

v3:
 - rebased patches with staging
 - adjusted formatting

v2:
 - rebased patches with staging

 xen/arch/x86/hvm/hvm.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/xen/arch/x86/hvm/hvm.c b/xen/arch/x86/hvm/hvm.c
index 9a4147b62e..2981a61cee 100644
--- a/xen/arch/x86/hvm/hvm.c
+++ b/xen/arch/x86/hvm/hvm.c
@@ -4237,11 +4237,14 @@ static int hvm_set_param(struct domain *d, uint32_t index, uint64_t value)
             rc = -EINVAL;
         break;
     case HVM_PARAM_IDENT_PT:
+        v = domain_vcpu(d, 0);
+
         /*
          * Only actually required for VT-x lacking unrestricted_guest
          * capabilities.  Short circuit the pause if possible.
          */
-        if ( paging_mode_shadow(d) || !using_vmx() )
+        if ( paging_mode_shadow(d) || !using_vmx() || !v ||
+             vmx_unrestricted_guest(v) )
             break;
 
         /*
-- 
2.55.0



-- 
Teddy Astie | Vates XCP-ng Developer

XCP-ng & Xen Orchestra - Vates solutions

web: https://vates.tech

Re: [PATCH v3] x86/vmx: Avoid pausing on HVM_PARAM_IDENT_PT in additional cases

Posted by Andrew Cooper 2 weeks, 1 day ago

On 09/09/2026 1:12 pm, Teddy Astie wrote:
> When settings HVM_PARAM_IDENT_PT, skip domain pausing when :
> - there is no vcpu
> - unrestricted guest capability is used
>
> Signed-off-by: Teddy Astie <teddy.astie@vates.tech>
> ---
> Regarding checking for hvm_paging_enabled(v) (proposed in v2 review), we can't
> as hvm_paging_enabled() is vCPU specific, and vCPU 0 may have a different value
> to another one.
>
> v3:
>  - rebased patches with staging
>  - adjusted formatting
>
> v2:
>  - rebased patches with staging
>
>  xen/arch/x86/hvm/hvm.c | 5 ++++-
>  1 file changed, 4 insertions(+), 1 deletion(-)
>
> diff --git a/xen/arch/x86/hvm/hvm.c b/xen/arch/x86/hvm/hvm.c
> index 9a4147b62e..2981a61cee 100644
> --- a/xen/arch/x86/hvm/hvm.c
> +++ b/xen/arch/x86/hvm/hvm.c
> @@ -4237,11 +4237,14 @@ static int hvm_set_param(struct domain *d, uint32_t index, uint64_t value)
>              rc = -EINVAL;
>          break;
>      case HVM_PARAM_IDENT_PT:
> +        v = domain_vcpu(d, 0);
> +
>          /*
>           * Only actually required for VT-x lacking unrestricted_guest
>           * capabilities.  Short circuit the pause if possible.
>           */
> -        if ( paging_mode_shadow(d) || !using_vmx() )
> +        if ( paging_mode_shadow(d) || !using_vmx() || !v ||
> +             vmx_unrestricted_guest(v) )
>              break;

The comment isn't really correct, and using vmx_unrestricted_guest()
isn't really correct either.

How about this instead:

        /*
         * IDENT_PT is needed only for Nehalem-era VT-x, where EPT is
         * available but Unrestricted Guest is not.
         *
         * The identity pagetable is configured by the toolstack or hvmloader,
         * and the param needs to move in the migration stream in case the VM
         * lands on an EPT && !Unrestricted system.
         *
         * Nothing, besides recording the value, needs to happen other than
         * for VT-x EPT && !Unrestricted VMs.
         *
         * TODO: Unrestricted Guest should be a domain property not a vCPU
         * property.
         */
        v = domain_vcpu(d, 0);
        if ( !using_vmx() || !paging_mode_hap(d) || !v ||
             vmx_unrestricted_guest(v) )
            break;

The case of no vCPUs isn't very interesting; in that case, pausing the
domain is free.  We can at least note that the data is in the wrong
place, even if we don't fix it yet.

~Andrew

Re: [PATCH v3] x86/vmx: Avoid pausing on HVM_PARAM_IDENT_PT in additional cases

Posted by Teddy Astie 2 weeks, 1 day ago

Le 09/09/2026 à 17:35, Andrew Cooper a écrit :
> On 09/09/2026 1:12 pm, Teddy Astie wrote:
>> When settings HVM_PARAM_IDENT_PT, skip domain pausing when :
>> - there is no vcpu
>> - unrestricted guest capability is used
>>
>> Signed-off-by: Teddy Astie <teddy.astie@vates.tech>
>> ---
>> Regarding checking for hvm_paging_enabled(v) (proposed in v2 review), we can't
>> as hvm_paging_enabled() is vCPU specific, and vCPU 0 may have a different value
>> to another one.
>>
>> v3:
>>   - rebased patches with staging
>>   - adjusted formatting
>>
>> v2:
>>   - rebased patches with staging
>>
>>   xen/arch/x86/hvm/hvm.c | 5 ++++-
>>   1 file changed, 4 insertions(+), 1 deletion(-)
>>
>> diff --git a/xen/arch/x86/hvm/hvm.c b/xen/arch/x86/hvm/hvm.c
>> index 9a4147b62e..2981a61cee 100644
>> --- a/xen/arch/x86/hvm/hvm.c
>> +++ b/xen/arch/x86/hvm/hvm.c
>> @@ -4237,11 +4237,14 @@ static int hvm_set_param(struct domain *d, uint32_t index, uint64_t value)
>>               rc = -EINVAL;
>>           break;
>>       case HVM_PARAM_IDENT_PT:
>> +        v = domain_vcpu(d, 0);
>> +
>>           /*
>>            * Only actually required for VT-x lacking unrestricted_guest
>>            * capabilities.  Short circuit the pause if possible.
>>            */
>> -        if ( paging_mode_shadow(d) || !using_vmx() )
>> +        if ( paging_mode_shadow(d) || !using_vmx() || !v ||
>> +             vmx_unrestricted_guest(v) )
>>               break;
> 
> The comment isn't really correct, and using vmx_unrestricted_guest()
> isn't really correct either.
> 
> How about this instead:
> 
>          /*
>           * IDENT_PT is needed only for Nehalem-era VT-x, where EPT is
>           * available but Unrestricted Guest is not.
>           *
>           * The identity pagetable is configured by the toolstack or hvmloader,
>           * and the param needs to move in the migration stream in case the VM
>           * lands on an EPT && !Unrestricted system.
>           *
>           * Nothing, besides recording the value, needs to happen other than
>           * for VT-x EPT && !Unrestricted VMs.
>           *
>           * TODO: Unrestricted Guest should be a domain property not a vCPU
>           * property.
>           */
>          v = domain_vcpu(d, 0);
>          if ( !using_vmx() || !paging_mode_hap(d) || !v ||
>               vmx_unrestricted_guest(v) )
>              break;
> 

I'm ok with it (though regarding paging_mode_hap, I think 
paging_mode_shadow is prefered here because it will translate into 
directly 0 if shadow paging is disabled, otherwise we check for 
PG_HAP_enable, at least, that's what [1] seems to be about).

But if we move unrestricted guest as being a domain-wide properly (or 
alternatively a system-wide one); that could make things simpler.

[1] x86/paging: replace !paging_mode_hap() with paging_mode_shadow()

> 
> The case of no vCPUs isn't very interesting; in that case, pausing the
> domain is free.  We can at least note that the data is in the wrong
> place, even if we don't fix it yet.
> 

The check for !v is mostly required for vmx_unrestricted_guest(v) to not 
dereference NULL.

> ~Andrew

Teddy

Re: [PATCH v3] x86/vmx: Avoid pausing on HVM_PARAM_IDENT_PT in additional cases

Posted by Jan Beulich 2 weeks, 1 day ago

On 09.09.2026 14:12, Teddy Astie wrote:
> When settings HVM_PARAM_IDENT_PT, skip domain pausing when :
> - there is no vcpu
> - unrestricted guest capability is used
> 
> Signed-off-by: Teddy Astie <teddy.astie@vates.tech>

Reviewed-by: Jan Beulich <jbeulich@suse.com>

> ---
> Regarding checking for hvm_paging_enabled(v) (proposed in v2 review), we can't
> as hvm_paging_enabled() is vCPU specific, and vCPU 0 may have a different value
> to another one.

Ah, yes, of course.

Jan