Re: [RFC PATCH net-next 2/3] flow_offload: restore ability to collect separate stats per action

From: Edward Cree <ecree@solarflare.com>
To: Jamal Hadi Salim <jhs@mojatatu.com>,
	Jakub Kicinski <jakub.kicinski@netronome.com>
Cc: Jiri Pirko <jiri@resnulli.us>,
	Pablo Neira Ayuso <pablo@netfilter.org>,
	David Miller <davem@davemloft.net>,
	netdev <netdev@vger.kernel.org>,
	"Cong Wang" <xiyou.wangcong@gmail.com>,
	Andy Gospodarek <andy@greyhouse.net>,
	Anjali Singhai Jain <anjali.singhai@intel.com>,
	Or Gerlitz <gerlitz.or@gmail.com>
Subject: Re: [RFC PATCH net-next 2/3] flow_offload: restore ability to collect separate stats per action
Date: Tue, 7 May 2019 13:27:15 +0100	[thread overview]
Message-ID: <1b37d659-5a2b-6130-e8d6-c15d6f57b55e@solarflare.com> (raw)
In-Reply-To: <db827a95-1042-cf74-1378-8e2eac356e6d@mojatatu.com>

On 06/05/2019 13:41, Jamal Hadi Salim wrote:
> On 2019-05-04 2:27 a.m., Jakub Kicinski wrote:
>> On Fri, 3 May 2019 16:06:55 +0100, Edward Cree wrote:
>>> Introduce a new offload command TC_CLSFLOWER_STATS_BYINDEX, similar to
>>>   the existing TC_CLSFLOWER_STATS but specifying an action_index (the
>>>   tcfa_index of the action), which is called for each stats-having action
>>>   on the rule.  Drivers should implement either, but not both, of these
>>>   commands.
>
> [..]
>>
>> It feels a little strange to me to call the new stats updates from
>> cls_flower, if we really want to support action sharing correctly.
>>
>> Can RTM_GETACTION not be used to dump actions without dumping the
>> classifiers?  If we dump from the classifiers wouldn't that lead to
>> stale stats being returned?
>
> Not sure about the staleness factor, but:
> For efficiency reasons we certainly need the RTM_GETACTION approach
> (as you stated above we dont need to dump all that classifier info if
> all we want are stats). This becomes a big deal if you have a lot
> of stats/rules.
I don't know much of anything about RTM_GETACTION, but it doesn't appear
 to be part of the current "tc offload" world, which AIUI is very much
 centred around cls_flower.  I'm just trying to make counters in
 cls_flower offload do 'the right thing' (whatever that may be), anything
 else is out of scope.

> Most H/W i have seen has a global indexed stats table which is
> shared by different action types (droppers, accept, mirror etc).
> The specific actions may also have their own tables which also
> then refer to the 32 bit index used in the stats table[1].
> So for this to work well, the action will need at minimal to have
> two indices one that is used in hardware stats table
> and another that is kernel mapped to identify the attributes. Of
> course we'll need to have a skip_sw flag etc.
I'm not sure I'm parsing this correctly, but are you saying that the
 index namespace is per-action type?  I.e. a mirred and a drop action
 could have the same index yet expect to have separate counters?  My
 approach here has assumed that in such a case they would share their
 counters.