On Fri, Jan 11, 2019 at 09:34:11AM +0100, Peter Zijlstra wrote:
> On Thu, Jan 10, 2019 at 11:29:20PM +0100, Florian Westphal wrote:
> > Peter Zijlstra <pet...@infradead.org> wrote:
> > > Would using synchronize_rcu() not also mean you can get rid of that
> > > xt_write_recseq*() stuff entirely?
> >
> > No, because those are used to synchronize with cpus that read
> > the ruleset counters, see
> >
> > net/ipv4/netfilter/ip_tables.c:get_counters().
>
> Ah, bummer :/
>
> > > Anyway, synchronize_rcu() can also take a little while, but I don't
> > > think anywere near 30 seconds.
> >
> > Ok, I think in that case it would be best to just replace the
> > recseq value sampling with smp_mb + synchronize_rcu plus a comment
> > that explains why its done.
>
> synchronize_rcu() implies smp_mb() on all CPUs.
Yes, it does, but in the case of idle CPUs, the smp_mb() calls are only
required to follow any pre-existing RCU read-side critical section on
the one hand an precede any RCU read-side critical section completing
after the synchronize_rcu() on the other.
To do more would mean waking up idle CPUs, which does not make the
battery-powered guys happy. ;-)
Thanx, Paul
