Re: [PATCH] pkt_sched: gen_kill_estimator() rcu fixes

[Eric Dumazet <eric.dumazet@gmail.com>]

Le mercredi 09 juin 2010 à 12:50 +0000, Jarek Poplawski a écrit :
> On Wed, Jun 09, 2010 at 02:09:23PM +0200, Eric Dumazet wrote:
> > Le mercredi 09 juin 2010 ?? 10:41 +0000, Jarek Poplawski a écrit :
> > 
> > > Spelling suggestion: xt_rateest_free_rcu?
> > > 
> > > Since it's only 3 places I'd consider adding little comments,
> > > who needs this call_rcu (like in qdisc_destroy).
> > > 
> > > Anyway this patch looks OK to me.
> > > 
> > 
> > Thanks for reviewing, you are right about typo and adding some comments.
> > 
> > What about following ?
> 
> Very nice! (But let's remember these comments aren't very precise
> wrt. bstats at the moment ;-)

Yes, I anticipated a bit next patch ;)

BTW, I think using qdisc spinlock to update stats or reading them is not
really needed, particularly in xt_rateest/xt_RATEEST case, where
per_cpu bstats would be good

We could have one seqcount_t used by (bytes/packets) bstats producer,
and one seqcount_t used by est_timer when computing rates.


bstats updates :


write_seqcount_begin(&bstats->abs_seq);
bstats->bytes += len;
bstats->packets++;
write_seqcount_end(&bstats->abs_seq);


est_timer() would really run lockless, only rcu protected.

(No more spinlock, only the actual rcu_read_lock())


do {
	seq = read_seqcount_begin(&bstats->abs_seq);
	abs_bytes = bstats->bytes;
	abs_packets = bstats->packets;
} while (read_seqcount_retry(&bstats->abs_seq, seq));

write_seqcount_begin(&s->rate_seq);
/* compute rate estimation and store it */
...
write_seqcount_end(&s->rate_seq);

This would mean bstats should be rcu protected too...