[PATCH 1/2 net-next] net: fec: add napi support to improve proformance

[Eric Dumazet]

On Wed, 2013-01-23 at 15:37 +0800, Frank Li wrote:
> 2013/1/23 Eric Dumazet <eric.dumazet at gmail.com>:
> > On Wed, 2013-01-23 at 12:12 +0800, Frank Li wrote:
> >> Add napi support
> >>
> >> Before this patch
> >>
> >>  iperf -s -i 1
> >>  ------------------------------------------------------------
> >>  Server listening on TCP port 5001
> >>  TCP window size: 85.3 KByte (default)
> >>  ------------------------------------------------------------
> >>  [  4] local 10.192.242.153 port 5001 connected with 10.192.242.138 port 50004
> >>  [ ID] Interval       Transfer     Bandwidth
> >>  [  4]  0.0- 1.0 sec  41.2 MBytes   345 Mbits/sec
> >>  [  4]  1.0- 2.0 sec  43.7 MBytes   367 Mbits/sec
> >>  [  4]  2.0- 3.0 sec  42.8 MBytes   359 Mbits/sec
> >>  [  4]  3.0- 4.0 sec  43.7 MBytes   367 Mbits/sec
> >>  [  4]  4.0- 5.0 sec  42.7 MBytes   359 Mbits/sec
> >>  [  4]  5.0- 6.0 sec  43.8 MBytes   367 Mbits/sec
> >>  [  4]  6.0- 7.0 sec  43.0 MBytes   361 Mbits/sec
> >>
> >> After this patch
> >>  [  4]  2.0- 3.0 sec  51.6 MBytes   433 Mbits/sec
> >>  [  4]  3.0- 4.0 sec  51.8 MBytes   435 Mbits/sec
> >>  [  4]  4.0- 5.0 sec  52.2 MBytes   438 Mbits/sec
> >>  [  4]  5.0- 6.0 sec  52.1 MBytes   437 Mbits/sec
> >>  [  4]  6.0- 7.0 sec  52.1 MBytes   437 Mbits/sec
> >>  [  4]  7.0- 8.0 sec  52.3 MBytes   439 Mbits/sec
> >
> > Strange, as you still call netif_rx()
> >
> > NAPI should call netif_receive_skb() instead
> >
> 
> Thank you point out.
> After re-test, I found performance is almost no change if use netif_receive_skb.
> I am not sure if it is my NAPI implement problem.
> 
> napi_gro_received is better than netif_receive_skb, but worse than netif_rx.
> 
> From performance point view,
> 
> netif_rx                    --- fastest
> napi_gro_received   --- middle, near to netif_rx
> netif_receive_skb    --- slowest, almost the same as original no-napi version.
> 
> Do you have any idea about this phenomena?

No idea, you'll have to find out using perf tool if available.

Is your machine SMP, and the application running on another cpu than the
softirq handler for your device ?

A NAPI driver must call netif_receive_skb(), especially if
the RX path does a full copy of the frame : Its hot in cpu cache and
should be processed at once.

Escaping to netif_rx() is only adding an extra softirq and risk of data
being evicted from cpu caches.

Here your performance increase only comes from hw_lock being not anymore
locked in RX path.