[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <vbfr230u7su.fsf@mellanox.com>
Date: Fri, 25 Oct 2019 16:08:36 +0000
From: Vlad Buslov <vladbu@...lanox.com>
To: Jamal Hadi Salim <jhs@...atatu.com>
CC: Vlad Buslov <vladbu@...lanox.com>, Jiri Pirko <jiri@...nulli.us>,
"netdev@...r.kernel.org" <netdev@...r.kernel.org>,
"xiyou.wangcong@...il.com" <xiyou.wangcong@...il.com>,
"davem@...emloft.net" <davem@...emloft.net>,
"mleitner@...hat.com" <mleitner@...hat.com>,
"dcaratti@...hat.com" <dcaratti@...hat.com>,
Eric Dumazet <edumazet@...gle.com>
Subject: Re: [PATCH net-next 00/13] Control action percpu counters allocation
by netlink flag
On Fri 25 Oct 2019 at 18:43, Jamal Hadi Salim <jhs@...atatu.com> wrote:
> On 2019-10-25 11:18 a.m., Vlad Buslov wrote:
>>
>
>> The problem with this approach is that it only works when actions are
>> created through act API, and not when they are created together with
>> filter by cls API which doesn't expect or parse TCA_ROOT. That is why I
>> wanted to have something in tcf_action_init_1() which is called by both
>> of them.
>>
>
> Aha. So the call path for tcf_action_init_1() via cls_api also needs
> to have this infra. I think i understand better what you wanted
> to do earlier with changing those enums.
>
> This is fixable - we just need to have the classifier side also take a
> new action root flags bitfield (I can send a sample). To your original
Trying to add this infra to cls API leads back to my original question:
how do I do it in backward compatible manner? I assume that we can't
break users of RTM_NEWTFILTER.
> comment, it is ugly. But maybe "fixing it" is pushing the boundaries
> and we should just go on and let your original approach in.
> My only worry is, given this is uapi, if we go back to your original
> idea we continue to propagate the bad design and we cant take it back
> (all the tooling etc would be cast for the next 5 years even if we
> did fix it in a month). Thoughts?
I don't see anything particularly ugly with extending either
action-specific attributes or TCA_ACT_* attributes, so its hard for me
to reason about problems with current approach :)
>
> cheers,
> jamal
Powered by blists - more mailing lists