Re: [Xen-devel] [PATCH] x86/altp2m: clean up p2m_{get/set}_suppress_ve()

[George Dunlap <george.dunlap@xxxxxxxxxx>]

On 09/24/2018 04:45 PM, Razvan Cojocaru wrote:
> On 9/24/18 6:25 PM, George Dunlap wrote:
>> On 09/23/2018 06:04 PM, Razvan Cojocaru wrote:
>>> Move p2m_{get/set}_suppress_ve() to p2m.c, replace incorrect
>>> ASSERT() in p2m-pt.c (since a guest can run in shadow mode even on
>>> a system with virt exceptions, which would trigger the ASSERT()),
>>> and move the VMX-isms (cpu_has_vmx_virt_exceptions checks) to
>>> p2m_ept_{get/set}_entry().
>>>
>>> Signed-off-by: Razvan Cojocaru <rcojocaru@xxxxxxxxxxxxxxx>
>>
>> Thanks for the clean up.  Two realtively minor comments...
>>
>>> @@ -931,6 +942,16 @@ static mfn_t ept_get_entry(struct p2m_domain *p2m,
>>>      mfn_t mfn = INVALID_MFN;
>>>      struct ept_data *ept = &p2m->ept;
>>>  
>>> +    if ( sve )
>>> +    {
>>> +        if ( !cpu_has_vmx_virt_exceptions )
>>> +            return INVALID_MFN;
>>> +
>>> +        /* #VE should be enabled for this vcpu. */
>>> +        if ( gfn_eq(vcpu_altp2m(current).veinfo_gfn, INVALID_GFN) )
>>> +            return INVALID_MFN;
>>> +    }
>>
>> Is there a good reason to return error her rather than just putting '1'
>> in the sve location, like the p2m_pt.c version of this function does?
> 
> First, thanks for the review!
> 
> The p2m_pt.c version can only return 1 because that's the only value
> that bit can have on #VE-incapable hardware. For the
> cpu_has_vmx_virt_exceptions, that assumption does hold, however in a
> scenario where:
> 
> 1. we enable #VE and set that bit to 0;
> 2. we disable #VE (so gfn_eq(vcpu_altp2m(current).veinfo_gfn,
> INVALID_GFN) == true);
> 3. we call ept_get_entry();
> 
> setting it to 1 would be misleading, since it's value is now really 0.
> 
> I do agree that returning INVALID_MFN is no necessarily more informative.
> 
> Alternatively, I could simply remove the checks here altogether. If
> !cpu_has_vmx_virt_exceptions then ept_get_entry() should fail anyway, so
> the bit will just remain 1 and thus the following code:
> 
>  999     if ( is_epte_valid(ept_entry) )
> 1000     {
> 1001         *t = p2m_recalc_type(recalc || ept_entry->recalc,
> 1002                              ept_entry->sa_p2mt, p2m, gfn);
> 1003         *a = ept_entry->access;
> 1004         if ( sve )
> 1005             *sve = ept_entry->suppress_ve;
> 
> should automatically do the right thing. And if, in the above scenario,
> the bit became 0, we return that value properly as well.
> 
> Would that be better?

Sorry, yes, that's what I intended, although I certainly wasn't clear.
What I meant was, the pt version of get_entry() would succeed and return
something sensible even on non-#VE-capable hardware; why should the ept
version not do the same thing?

So yes, I think just removing the checks and letting the actual value
from the p2m entry be passed back is the right thing to do.

> 
>>> +int p2m_get_suppress_ve(struct domain *d, gfn_t gfn, bool *suppress_ve,
>>> +                        unsigned int altp2m_idx)
>>> +{
>>> +    struct p2m_domain *host_p2m = p2m_get_hostp2m(d);
>>> +    struct p2m_domain *ap2m = NULL;
>>> +    struct p2m_domain *p2m;
>>> +    mfn_t mfn;
>>> +    p2m_access_t a;
>>> +    p2m_type_t t;
>>> +
>>> +    /* #VE should be enabled for this vcpu. */
>>> +    if ( gfn_eq(vcpu_altp2m(current).veinfo_gfn, INVALID_GFN) )
>>> +        return -ENXIO;
>>
>> What's the purpose of checking for this here, if we don't check for this
>> in p2m_set_suppress_ve()?
> 
> Sorry, I seem to have accidentally left that in p2m_get_suppress_ve() -
> I'll delete it from here and leave it only in ept_set_entry(). It's
> pointless to have it duplicated here.

Great, thanks.

 -George

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel