Re: [PATCH v4 00/24] x86/resctrl: Merge the CDP resources

From: James Morse
Date: Thu Jun 17 2021 - 13:04:02 EST


Hi Reinette,

On 15/06/2021 19:05, Reinette Chatre wrote:
> On 6/15/2021 9:48 AM, James Morse wrote:
>> On 15/06/2021 17:16, Reinette Chatre wrote:
>>> On 6/14/2021 1:09 PM, James Morse wrote:
>>> For the most part I think this series looks good. The one thing I am concerned about is
>>> the resctrl user interface change. On a system that supports L3 CDP there is a visible
>>> change when CDP is not enabled:
>>>
>>> Before this series:
>>> # cat schemata
>>>     L3:0=fffff;1=fffff
>>>
>>> After this series:
>>> # cat schemata
>>> L3:0=fffff;1=fffff
>>
>> Hmm, I thought I'd fixed this with v2, ... I see this is subtly different.
>>
>> This could be tweaked by getting schemata_list_add() to include the length of the longest
>> suffix if the resource supports CDP, but its not enabled. (Discovering that means
>> cdp_capable moves to be something the 'fs' bits of resctrl can see.)

>> I'm a little nervous 'adding 4 spaces' because user-space expects them. (I agree if it
>> breaks user-space it has to be done). I guess this is the problem with string parsing as
>> part of the interface!

> This is a tricky and interesting one. It seems that the original intended behavior is
> indeed the way you changed it to. By originally using for_each_enabled_rdt_resource() to
> determine the maximum width in de016df88f23 ("x86/intel_rdt: Update schemata read to show
> data in tabular format"). This was added in v4.12 and dictated the interface until v4.13.
> This was changed in 1b5c0b758317 ("x86/intel_rdt: Cleanup namespace to support RDT
> monitoring") when it used for_each_alloc_capable_rdt_resource(r) instead, added in v4.14
> that is a stable kernel and the most likely interface used by users.
>
> To me the less risky change is to maintain the existing interface, but perhaps there are
> some other guidance in this regard?

I think this is just the problem with having anything other than 'one value per file', as
sysfs does. Maintaining it involves getting painted into a corner by the worst parser
user-space manages to come up with!


>> I assume that in the (distant) future having CDP capable resources with names more than 2
>> characters isn't going to be a problem. (I don't have an example)
>
> The last statement is not clear to me. Could you please elaborate why two characters would
> be significant? From what I understand the expectation would be that the width is the
> maximum name length of all possible schema, whether they are enabled or not.

Great - I was nervous that if shorter strings are a problem, what about longer?

( Arm SoCs often have a system-cache that lives between the LLC and DRAM. Its not a CPU
cache, so its not really L4. Because of the way CDP gets emulated it affects all caches.
If people build these things with MPAM support - and we choose to add a schema for them:
you'd end up with 'SYSTEM-CACHECODE' and 'SYSTEM-CACHEDATA'. Its not a real example as if
its needed, 'SC' is probably acceptable.)


>>> There are a few user space tools that parse the resctrl schemata file and it may be easier
>>> to keep the interface consistent than to find and audit them all to ensure they will keep
>>> working.
>
> To me this continues the biggest hurdle in maintaining the behavior as you have it in this
> series.

No problem - I've changed it as described.


>>> A heads-up is that there are some kernel-doc fixups in the works that will conflict with
>>> your series. You yourself fix at least one of these kernel-doc issues in this series - the
>>> description of mbm_width in the first patch. I will ask the submitter of the kernel-doc
>>> fixups to use your text to help with the merging.
>>
>> Please point me at something to rebase onto!
>> (as far as I can see, tip/x86/cache hasn't moved)
>
> These patches have not yet been merged. The most recent version was sent yesterday. Your
> current base is good.

I've based on tip/master, which merged rc6....


Thanks,

James