Re: TC stats / hw offload question

[Pablo Neira Ayuso <pablo@netfilter.org>]

On Mon, Apr 29, 2019 at 03:11:06PM +0100, Edward Cree wrote:
> On 26/04/2019 19:49, Pablo Neira Ayuso wrote:
> > On Fri, Apr 26, 2019 at 01:13:41PM +0100, Edward Cree wrote:
> >> Thus if (and only if) two TC actions have the same tcfa_index, they will
> >>  share a single counter in the HW.
> >> I gathered from a previous conversation with Jamal[1] that that was the
> >>  correct behaviour:
> >>> Note, your counters should also be shareable; example, count all
> >>> the drops in one counter across multiple flows as in the following
> >>> case where counter index 1 is used.
> >>>
> >>> tc flower match foo action drop index 1
> >>> tc flower match bar action drop index 1
> > The flow_action_entry structure needs a new 'counter_index' field to
> > store this. The tc_setup_flow_action() function needs to be updated
> > for this for the FLOW_ACTION_{ACCEPT,DROP,REDIRECT,MIRRED} cases to
> > set this entry->counter_index field to tcfa_index, so the driver has
> > access to this.
> Hmm, I'm still not sure this solves everything.
> Before, we could write
> tc flower match foo \
>     action mirred egress mirror eth1 index 1 \
>     action mirred egress redirect eth2 index 2
> and have two distinct HW counters (one of which might e.g. be shared
>  with another rule).  But when reading those counters, under
>  fl_hw_update_stats(), the driver only gets to return one set of flow
>  stats for both actions.
> Previously, the driver's TC_CLSFLOWER_STATS handler was updating the
>  action stats directly, so was able to do something different for each
>  action, but that's not possible in 5.1.  At stats gathering time, the
>  driver doesn't even have access to anything that's per-action and
>  thus could have a flow_stats member shoved in it.
> AFAICT, the only reason this isn't a regression is that existing
>  drivers didn't implement the old semantics correctly.
>
> This is a bit of a mess; the best idea I've got is for the
>  TC_CLSFLOWER_STATS call to include a tcfa_index.  Then the driver
>  returns counter stats for that index, and tcf_exts_stats_update()
>  only updates those actions whose index matches.  But then
>  fl_hw_update_stats() would have to iterate over all the indices in
>  f->exts.  What do you think?

You could extend struct flow_stats to pass an array of stats to the
driver, including one stats per action and the counter index. Then,
tcf_exts_stats_update() uses this array of stats to update per-action
stats.

struct flow_action_stats {
        u32     counter_index;
        u64     pkts;
        u64     bytes;
        u64     lastused;
};

struct flow_stats {
        struct flow_action_stats        *stats[];
        u32                             num_actions;
};

As you mentioned, no driver supports for tcfa_index so far, probably
it would be a good idea to return -EOPNOTSUPP in such case by now.