Re: [PATCH V2 01/10] x86/CPU: Expose if cache is inclusive of lower level caches

[Reinette Chatre]

Hi Borislav,

On 8/6/2019 10:33 AM, Borislav Petkov wrote:
> On Tue, Aug 06, 2019 at 09:55:56AM -0700, Reinette Chatre wrote:
>> I am a bit cautious about this. When I started this work I initially
>> added a helper function to resctrl that calls CPUID to determine if the
>> cache is inclusive. At that time I became aware of a discussion
>> motivating against scattered CPUID calls and motivating for one instance
>> of CPUID information:
>> http://lkml.kernel.org/r/alpine.DEB.2.21.1906162141301.1760@xxxxxxxxxxxxxxxxxxxxxxx
> 
> Ah, there's that. That's still somewhat a work/discussion in progress
> thing. Let me discuss it with tglx.
> 
>> To answer your question about checking any cache: this seems to be
> 
> I meant the CPUID on any CPU and thus any cache - i.e., all L3s on the
> system should be inclusive and identical in that respect. Can't work
> otherwise, I'd strongly presume.

This is my understanding, yes. While this patch supports knowing whether
each L3 is inclusive or not, I expect this information to be the same
for all L3 instances as will be supported by a single query in
rdt_pseudo_lock_init(). This definitely is the case on the platforms we
are enabling in this round.

>> different between L2 and L3. On the Atom systems where L2 pseudo-locking
>> works well the L2 cache is not inclusive. We are also working on
>> supporting cache pseudo-locking on L3 cache that is not inclusive.
> 
> Hmm, so why are you enforcing the inclusivity now:
> 
> +       if (p->r->cache_level == 3 &&
> +           !get_cache_inclusive(plr->cpu, p->r->cache_level)) {
> +               rdt_last_cmd_puts("L3 cache not inclusive\n");
> 
> but then will remove this requirement in the future? Why are we even
> looking at cache inclusivity then and not make pseudo-locking work
> regardless of that cache property?

Some platforms being enabled in this round have SKUs with inclusive
cache and also SKUs with non-inclusive cache. The non-inclusive cache
SKUs do not support cache pseudo-locking and cannot be made to support
cache pseudo-locking with software changes. Needing to know if cache is
inclusive or not will thus remain a requirement to distinguish between
these different SKUs. Supporting cache pseudo-locking on platforms with
non inclusive cache will require new hardware features.
> Because if we're going to go and model this cache inclusivity property
> properly in struct cpuinfo_x86 or struct cacheinfo or wherever, and do
> that for all cache levels because apparently we're going to need that;
> but then later it turns out we won't need it after all, why are we even
> bothering?
> 
> Or am I missing some aspect?

Reinette