[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <7C59A1FE-1005-499A-A87C-4639D896F6D7@redhat.com>
Date: Fri, 07 Oct 2022 17:03:12 +0200
From: Eelco Chaudron <echaudro@...hat.com>
To: Davide Caratti <dcaratti@...hat.com>
Cc: Marcelo Leitner <mleitner@...hat.com>,
Ilya Maximets <i.maximets@....org>,
Tianyu Yuan <tianyu.yuan@...igine.com>,
Simon Horman <simon.horman@...igine.com>, dev@...nvswitch.org,
oss-drivers <oss-drivers@...igine.com>, netdev@...r.kernel.org,
Cong Wang <xiyou.wangcong@...il.com>,
Jamal Hadi Salim <jhs@...atatu.com>,
Oz Shlomo <ozsh@...dia.com>, Paul Blakey <paulb@...dia.com>
Subject: Re: [ovs-dev] [PATCH] tests: fix reference output for meter offload
stats
On 7 Oct 2022, at 16:39, Davide Caratti wrote:
> On Fri, Oct 7, 2022 at 3:21 PM Marcelo Leitner <mleitner@...hat.com> wrote:
>>
>> (+TC folks and netdev@)
>>
>> On Fri, Oct 07, 2022 at 02:42:56PM +0200, Ilya Maximets wrote:
>>> On 10/7/22 13:37, Eelco Chaudron wrote:
>
> [...]
>
>> I don't see how we could achieve this without breaking much of the
>> user experience.
>>
>>>
>>> - or create something like act_count - a dummy action that only
>>> counts packets, and put it in every datapath action from OVS.
>>
>> This seems the easiest and best way forward IMHO. It's actually the
>> 3rd option below but "on demand", considering that tc will already use
>> the stats of the first action as the flow stats (in
>> tcf_exts_dump_stats()), then we can patch ovs to add such action if a
>> meter is also being used (or perhaps even always, because other
>> actions may also drop packets, and for OVS we would really be at the
>> 3rd option below).
Guess we have to add this extra action to each datapath flow offloaded due to the way flows back and forth translations are handled. Maybe we can do it selectively, but the code might become messier than it already is :(
> Correct me if I'm wrong, but actually act_gact action with "pipe"
> control action should already do this counting job.
I think we could use that, as we only use TC_ACT_GOTO_CHAIN and TC_ACT_SHOT. And it looks like TC_ACT_SHOT is not decoded correctly :(
Powered by blists - more mailing lists