Re: [PATCH v9 1/3] x86/tlb: introduce a flush HVM ASIDs flag

[Jan Beulich <jbeulich@xxxxxxxx>]

On 14.04.2020 11:34, Roger Pau Monné wrote:
> On Tue, Apr 14, 2020 at 11:09:43AM +0200, Jan Beulich wrote:
>> On 14.04.2020 10:01, Roger Pau Monné wrote:
>>> On Thu, Apr 09, 2020 at 01:16:57PM +0200, Jan Beulich wrote:
>>>> On 08.04.2020 17:10, Roger Pau Monné wrote:
>>>>> On Wed, Apr 08, 2020 at 01:25:14PM +0200, Jan Beulich wrote:
>>>>>> On 06.04.2020 12:57, Roger Pau Monne wrote:
>>>>>>> --- a/xen/arch/x86/mm/paging.c
>>>>>>> +++ b/xen/arch/x86/mm/paging.c
>>>>>>> @@ -613,7 +613,8 @@ void paging_log_dirty_range(struct domain *d,
>>>>>>>  
>>>>>>>      p2m_unlock(p2m);
>>>>>>>  
>>>>>>> -    flush_tlb_mask(d->dirty_cpumask);
>>>>>>> +    flush_mask(d->dirty_cpumask, (!hap_enabled(d) ? FLUSH_TLB : 0) |
>>>>>>> +                                 FLUSH_HVM_ASID_CORE);
>>>>>>
>>>>>> In cases where one case is assumed to be more likely than the other
>>>>>> putting the more likely one first can be viewed as a mild hint to
>>>>>> the compiler, and hence an extra ! may be warranted in an if() or
>>>>>> a conditional expression. Here, however, I don't think we can
>>>>>> really consider one case more likely than the other, and hence I'd
>>>>>> suggest to avoid the !, flipping the other two expressions
>>>>>> accordingly. I may take the liberty to adjust this while committing
>>>>>> (if I'm to be the one).
>>>>>
>>>>> That's fine, thanks. Somehow '!hap -> flush' was clearer in my mind.
>>>>
>>>> Thinking about it with the other HVM-related changes in v9, shouldn't
>>>> this then be
>>>>
>>>>     flush_mask(d->dirty_cpumask, (hap_enabled(d) ? 0 : FLUSH_TLB) |
>>>>                                  (is_hvm_domain(d) ? FLUSH_HVM_ASID_CORE : 
>>>> 0));
>>>>
>>>> Or wait - the only caller lives in hap.c. As a result the FLUSH_TLB
>>>> part can be dropped altogether. And I question the need of flushing
>>>> guest TLBs - this is purely a p2m operation. I'll go look at the
>>>> history of this function, but for now I think the call should be
>>>> dropped (albeit then maybe better in a separate patch).
>>>
>>> The ASID flush needs to stay unless it's moved into p2m_pt_set_entry,
>>> as p2m_pt_set_entry itself doesn't perform any ASID flush and won't
>>> work correctly.
>>
>> Just like for said in the other reply sent a few minutes ago - yes
>> for NPT, but no for EPT.
> 
> It's not strictly wrong for EPT as it won't cause EPT domains to
> malfunction, it's just redundant.

Right - it's wrong just in the sense of being pointless extra
overhead.

>>> I think it's safe to remove the TLB flush, as the code is only called
>>> from HAP, and hence is not used by shadow (which is what would require
>>> a plain TLB flush). The placement of this function seems misleading to
>>> me, as it looks like it's used by both shadow and HAP. It might be
>>> better to move it to hap.c if it's only to be used by HAP code.
>>
>> Either placement has its problems, I think. The function is meant to
>> be a paging layer one, but is needed by HAP only right now. I'm
>> pondering whether to wrap it in #ifdef CONFIG_HVM (plus perhaps a
>> respective ASSERT_UNREACHABLE()).
> 
> IMO if a TLB flush is not performed here we should add an
> ASSERT_UNREACHABLE if called from a shadow mode domain, or else we
> risk someone trying to use it in shadow later without realizing it's
> missing a TLB flush.

This would be fine with me.

Jan