Date:   Sat, 24 Oct 2020 13:40:31 -0400
From:   Jamal Hadi Salim <jhs@...atatu.com>
To:     Vlad Buslov <vladbu@...dia.com>
Cc:     Vlad Buslov <vlad@...lov.dev>, dsahern@...il.com,
        stephen@...workplumber.org, netdev@...r.kernel.org,
        davem@...emloft.net, xiyou.wangcong@...il.com, jiri@...nulli.us,
        ivecera@...hat.com, Vlad Buslov <vladbu@...lanox.com>
Subject: Re: [PATCH iproute2-next v3 2/2] tc: implement support for terse dump

On 2020-10-23 8:48 a.m., Vlad Buslov wrote:
> On Thu 22 Oct 2020 at 17:05, Jamal Hadi Salim <jhs@...atatu.com> wrote:
>> On 2020-10-21 4:19 a.m., Vlad Buslov wrote:
>>>
>>> On Tue 20 Oct 2020 at 15:29, Jamal Hadi Salim <jhs@...atatu.com> wrote:
>>>> On 2020-10-19 11:18 a.m., Vlad Buslov wrote:
>>>> My worry is you have a very specific use case for your hardware or
>>>> maybe it is ovs - where counters are uniquely tied to filters and
>>>> there is no sharing. And possibly maybe only one counter can be tied
>>>> to a filter (was not sure if you could handle more than one action
>>>> in the terse from looking at the code).
>>>
>>> OVS uses cookie to uniquely identify the flow and it does support
>>> multiple actions per flow.
>>
>>
>> ok, so they use it like a flowid/classid to identify the flow.
>> In our use case the cookie stores all kinds of other state that
>> the controller can avoid to lookup after a query.
>> index otoh is universal i.e two different users can intepret it
>> per action tying it specific stats.
>> IOW: I dont think it replaces the index.
>> Do they set cookies on all actions in a flow?
> 
> AFAIK on only one action per flow.
> 

To each their own i guess. Sounds like it is being used as flowid
entity.
We pack a lot of metaencoding into those cookies. And to us
a "service" is essentially a filter match folowed by a graph
of actions (which could cyclic).


>>> Cookie only consumes space in resulting netlink packet if used set the
>>> cookie during action init. Otherwise, the cookie attribute is omitted.
>>
>> True, but: I am wondering why it is even considered in when terseness
>> was a requirement (and index was left out).
> 
> There was several reasons for me to include it:
> 
> - As I wrote in previous email its TLV is only included in dump if user
>    set the cookie. Users who don't use cookies don't lose any performance
>    of terse dump.
> 
> - Including it didn't require any changes to tc_action_ops->dump() (like
>    passing 'terse' flag or introducing dedicated terse_dump() callback)
>    because it is processed in tcf_action_dump_1().
> 
> - OVS was the main use-case for us because it relies on filter dump for
>    flow revalidation and uses cookie to identify the flows.
>

Which is fine - but it is a very ovs specific need.
>>
>>>> In our case we totally bypass filters to reduce the amount of data
>>>> crossing to user space (tc action ls). Theres still a lot of data
>>>> crossing which we could trim with a terse dump. All we are interested
>>>> in are stats. Another alternative is perhaps to introduce the index for
>>>> the direct dump.
>>>
>>> What is the direct dump?
>>
>> tc action ls ...
>> Like i said in our use cases to get the stats we just dumped the actions
>> we wanted. It is a lot less data than having the filter + actions.
>> And with your idea of terseness we can trim down further how much
>> data by removing all the action attributes coming back if we set TERSE
>> flag in such a request. But the index has to be there to make sense.
> 
> Yes, that makes sense. I guess introducing something like 'tc action -br
> ls ..' mode implemented by means of existing terse flag + new 'also
> output action index' flag would achieve that goal.
>

Right. There should be no interest in the cookie here at all. Maybe
it could be optional with a flag indication.
Have time to cook a patch? I'll taste/test it.

cheers,
jamal

Powered by blists - more mailing lists