Re: [Xen-devel] [PATCH v3 5/7] xen/x86: disable global pages for domains with XPTI active

["Jan Beulich" <JBeulich@xxxxxxxx>]

>>> On 27.03.18 at 09:37, <jgross@xxxxxxxx> wrote:
> On 27/03/18 09:23, Jan Beulich wrote:
>>>>> On 27.03.18 at 09:14, <jgross@xxxxxxxx> wrote:
>>> On 22/03/18 17:30, Jan Beulich wrote:
>>>>>>> On 21.03.18 at 13:51, <jgross@xxxxxxxx> wrote:
>>>>> Instead of flushing the TLB from global pages when switching address
>>>>> spaces with XPTI being active just disable global pages via %cr4
>>>>> completely when a domain subject to XPTI is active. This avoids the
>>>>> need for extra TLB flushes as loading %cr3 will remove all TLB
>>>>> entries.
>>>>
>>>> I continue to be not entirely convinced of this move. I had an
>>>> alternative in mind: Since retaining global pages is particularly
>>>> relevant for switches between guest user and guest kernel
>>>> modes, what if we made a shortcut from e.g. lstar_enter through
>>>> switch_to_kernel to restore_all_guest without ever switching to
>>>> the full page Xen tables?
>>>>
>>>>> --- a/xen/arch/x86/mm.c
>>>>> +++ b/xen/arch/x86/mm.c
>>>>> @@ -508,18 +508,23 @@ void make_cr3(struct vcpu *v, mfn_t mfn)
>>>>>  void write_ptbase(struct vcpu *v)
>>>>>  {
>>>>>      struct cpu_info *cpu_info = get_cpu_info();
>>>>> +    unsigned long new_cr4;
>>>>> +
>>>>> +    new_cr4 = (is_pv_vcpu(v) && !is_idle_vcpu(v))
>>>>> +              ? pv_guest_cr4_to_real_cr4(v) : mmu_cr4_features;
>>>>
>>>> I'm not overly happy to see any new uses of mmu_cr4_features.
>>>> This should really only be used for priming certain values imo,
>>>> which isn't the case here (otoh pv_guest_cr4_to_real_cr4() does
>>>> so too, and perhaps better wouldn't). Hence I wonder whether
>>>> this shouldn't be read_cr4() | X86_CR4_PGE, not the least
>>>> because we've just got rid of the blanket reversion to
>>>> mmu_cr4_features in VMX code.
>>>
>>> I just realized that using read_cr4() | X86_CR4_PGE is wrong. We would
>>> keep bits switched on which a pv domain is allowed to modify (plus
>>> CR4_TSD eventually).
>>>
>>> Do we really want that?
>> 
>> Does it matter what exact CR4 settings we run with when it's not
>> a PV guest that's in context, and when we don't depend on the
>> settings ourselves? I don't think it does, and HVM guests run with
>> their own CR4 anyway. In fact there may end up being cases
>> where we won't need to switch CR4 another time when we come
>> here the next time with v being a PV vCPU.
> 
> I could imagine that there is some performance impact. cr4.tsd set
> might make rdtsc a little bit slower as an additional privilege level
> check is needed.

Quite possible, indeed. Another opinion on the route to take
would be helpful here. Andrew?

> Suspending requires cr4.fsgsbase to be set, which might have been
> reset by a pv guest.

pv_guest_cr4_to_real_cr4() consistently forces this bit to 1 (on
capable hardware).

Jan


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