On Mon, 26 Apr 2021 14:01:17 +0800
Hangbin Liu <liuhangbin@xxxxxxxxx> wrote:
> On Sat, Apr 24, 2021 at 09:01:29AM +0200, Jesper Dangaard Brouer wrote:
> > > > > >> @@ -3942,7 +3960,12 @@ int xdp_do_redirect(struct net_device *dev, struct xdp_buff *xdp,
> > > > > >> case BPF_MAP_TYPE_DEVMAP:
> > > > > >> fallthrough;
> > > > > >> case BPF_MAP_TYPE_DEVMAP_HASH:
> > > > > >> - err = dev_map_enqueue(fwd, xdp, dev);
> > > > > >> + map = xchg(&ri->map, NULL);
> > > > > >
> > > > > > Hmm, this looks dangerous for performance to have on this fast-path.
> > > > > > The xchg call can be expensive, AFAIK this is an atomic operation.
> > > > >
> > > > > Ugh, you're right. That's my bad, I suggested replacing the
> > > > > READ_ONCE()/WRITE_ONCE() pair with the xchg() because an exchange is
> > > > > what it's doing, but I failed to consider the performance implications
> > > > > of the atomic operation. Sorry about that, Hangbin! I guess this should
> > > > > be changed to:
> > > > >
> > > > > + map = READ_ONCE(ri->map);
> > > > > + if (map) {
> > > > > + WRITE_ONCE(ri->map, NULL);
> > > > > + err = dev_map_enqueue_multi(xdp, dev, map,
> > > > > + ri->flags & BPF_F_EXCLUDE_INGRESS);
> > > > > + } else {
> > > > > + err = dev_map_enqueue(fwd, xdp, dev);
> > > > > + }
> > > >
> > > > This is highly sensitive fast-path code, as you saw Bjørn have been
> > > > hunting nanosec in this area. The above code implicitly have "map" as
> > > > the likely option, which I don't think it is.
> > >
> > > Hi Jesper,
> > >
> > > From the performance data, there is only a slightly impact. Do we still need
> > > to block the whole patch on this? Or if you have a better solution?
> >
> > I'm basically just asking you to add an unlikely() annotation:
> >
> > map = READ_ONCE(ri->map);
> > if (unlikely(map)) {
> > WRITE_ONCE(ri->map, NULL);
> > err = dev_map_enqueue_multi(xdp, dev, map, [...]
> >
> > For XDP, performance is the single most important factor! You say your
> > performance data, there is only a slightly impact, there must be ZERO
> > impact (when your added features is not in use).
> >
> > You data:
> > Version | Test | Generic | Native
> > 5.12 rc4 | redirect_map i40e->i40e | 1.9M | 9.6M
> > 5.12 rc4 + patch | redirect_map i40e->i40e | 1.9M | 9.3M
> >
> > The performance difference 9.6M -> 9.3M is a slowdown of 3.36 nanosec.
> > Bjørn and others have been working really hard to optimize the code and
> > remove down to 1.5 nanosec overheads. Thus, introducing 3.36 nanosec
> > added overhead to the fast-path is significant.
>
> I re-check the performance data. The data
> > Version | Test | Generic | Native
> > 5.12 rc4 | redirect_map i40e->i40e | 1.9M | 9.6M
> > 5.12 rc4 + patch | redirect_map i40e->i40e | 1.9M | 9.3M
>
> is done on version 5.
>
> Today I re-did the test, on version 10, with xchg() changed to
> READ_ONCE/WRITE_ONCE. Here is the new data (Generic path data was omitted
> as there is no change)
>
> Version | Test | Generic | Native
> 5.12 rc4 | redirect_map i40e->i40e | 9.7M
> 5.12 rc4 | redirect_map i40e->veth | 11.8M
>
> 5.12 rc4 + patch | redirect_map i40e->i40e | 9.6M
Great to see the baseline redirect_map (i40e->i40e) have almost no
impact, only 1.07 ns ((1/9.7-1/9.6)*1000), which is what we want to
see. (It might be zero as measurements can fluctuate when diff is
below 2ns)
> 5.12 rc4 + patch | redirect_map i40e->veth | 11.6M
What XDP program are you running on the inner veth?
> 5.12 rc4 + patch | redirect_map multi i40e->i40e | 9.5M
I'm very surprised to see redirect_map multi being so fast (9.5M vs.
9.6M normal map-redir). I was expecting to see larger overhead, as the
code dev_map_enqueue_clone() would clone the packet in xdpf_clone() via
allocating a new page (dev_alloc_page) and then doing a memcpy().
Looking closer at this patchset, I realize that the test
'redirect_map-multi' is testing an optimization, and will never call
dev_map_enqueue_clone() + xdpf_clone(). IMHO trying to optimize
'redirect_map-multi' to be just as fast as base 'redirect_map' doesn't
make much sense. If the 'broadcast' call only send a single packet,
then there isn't any reason to call the 'multi' variant.
Does the 'selftests/bpf' make sure to activate the code path that does
cloning?
> 5.12 rc4 + patch | redirect_map multi i40e->veth | 11.5M
> 5.12 rc4 + patch | redirect_map multi i40e->mlx4+veth | 3.9M
>
> And after add unlikely() in the check path, the new data looks like
>
> Version | Test | Native
> 5.12 rc4 + patch | redirect_map i40e->i40e | 9.6M
> 5.12 rc4 + patch | redirect_map i40e->veth | 11.7M
> 5.12 rc4 + patch | redirect_map multi i40e->i40e | 9.4M
> 5.12 rc4 + patch | redirect_map multi i40e->veth | 11.4M
> 5.12 rc4 + patch | redirect_map multi i40e->mlx4+veth | 3.8M
>
> So with unlikely(), the redirect_map is a slightly up, while redirect_map
> broadcast has a little drawback. But for the total data it looks this time
> there is no much gap compared with no this patch for redirect_map.
>
> Do you think we still need the unlikely() in check path?
Yes. The call to redirect_map multi is allowed (and expected) to be
slower, because when using it to broadcast packets we expect that
dev_map_enqueue_clone() + xdpf_clone() will get activated, which will
be the dominating overhead.
--
Best regards,
Jesper Dangaard Brouer
MSc.CS, Principal Kernel Engineer at Red Hat
LinkedIn: http://www.linkedin.com/in/brouer
