Re: [PATCH net-next 00/13] Control action percpu counters allocation by netlink flag

[Vlad Buslov <vladbu@mellanox.com>]

On Thu 24 Oct 2019 at 19:12, Jamal Hadi Salim <jhs@mojatatu.com> wrote:
> On 2019-10-24 11:23 a.m., Vlad Buslov wrote:
>> On Thu 24 Oct 2019 at 10:35, Jiri Pirko <jiri@resnulli.us> wrote:
>>> Wed, Oct 23, 2019 at 04:21:51PM CEST, jhs@mojatatu.com wrote:
>>>> On 2019-10-23 9:04 a.m., Vlad Buslov wrote:
>>>>>
>
> [..]
>>>> Jiri complains constantly about all these new per-action TLVs
>>>> which are generic. He promised to "fix it all" someday. Jiri i notice
>>>> your ack here, what happened? ;->
>>>
>>> Correct, it would be great. However not sure how exactly to do that now.
>>> Do you have some ideas.
>>>
>>>
>>> But basically this patchset does what was done many many times in the
>>> past. I think it was a mistake in the original design not to have some
>>> "common attrs" :/ Lesson learned for next interfaces.
>>
>
> Jiri, we have a high level TLV space which can be applied to all
> actions. See discussion below with Vlad. At least for this specific
> change we can get away from repeating that mistake.
>
>> Jamal, can we reach some conclusion? Do you still suggest to refactor
>> the patches to use TCA_ROOT_FLAGS and parse it in act API instead of
>> individual actions?
>>
>
> IMO this would certainly help us walk away from having
> every action replicate the same attribute with common semantics.
> refactoring ->init() to take an extra attribute may look ugly but
> is cleaner from a uapi pov. Actions that dont implement the feature
> can ignore the extra parameter(s). It doesnt have to be TCA_ROOT_FLAGS
> but certainly that high level TLV space is the right place to put it.
> WDYT?
>
> cheers,
> jamal

Well, I like having it per-action better because of reasons I explained
before (some actions don't use percpu allocator at all and some actions
that are not hw offloaded don't need it), but I think both solutions
have their benefits and drawbacks, so I'm fine with refactoring it.

Do you have any opinion regarding flag naming? Several people suggested
to be more specific, but I strongly dislike the idea of hardcoding the
name of a internal kernel data structure in UAPI constant that will
potentially outlive the data structure by a long time.