[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-Id: <20221116143336.3385874-1-alexandr.lobakin@intel.com>
Date: Wed, 16 Nov 2022 15:33:36 +0100
From: Alexander Lobakin <alexandr.lobakin@...el.com>
To: Shenwei Wang <shenwei.wang@....com>
Cc: Alexander Lobakin <alexandr.lobakin@...el.com>,
Andrew Lunn <andrew@...n.ch>,
"David S. Miller" <davem@...emloft.net>,
Eric Dumazet <edumazet@...gle.com>,
Jakub Kicinski <kuba@...nel.org>,
"Paolo Abeni" <pabeni@...hat.com>,
Alexei Starovoitov <ast@...nel.org>,
"Daniel Borkmann" <daniel@...earbox.net>,
Jesper Dangaard Brouer <hawk@...nel.org>,
John Fastabend <john.fastabend@...il.com>,
Wei Fang <wei.fang@....com>,
"netdev@...r.kernel.org" <netdev@...r.kernel.org>,
"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
"imx@...ts.linux.dev" <imx@...ts.linux.dev>,
kernel test robot <lkp@...el.com>
Subject: Re: [PATCH v3 1/1] net: fec: add xdp and page pool statistics
From: Shenwei Wang <shenwei.wang@....com>
Date: Mon, 14 Nov 2022 21:17:48 +0000
> > -----Original Message-----
> > From: Alexander Lobakin <alexandr.lobakin@...el.com>
> > Sent: Monday, November 14, 2022 9:23 AM
> > To: Shenwei Wang <shenwei.wang@....com>
> > Cc: Alexander Lobakin <alexandr.lobakin@...el.com>; Andrew Lunn
> > <andrew@...n.ch>; David S. Miller <davem@...emloft.net>; Eric Dumazet
> > <edumazet@...gle.com>; Jakub Kicinski <kuba@...nel.org>; Paolo Abeni
> > <pabeni@...hat.com>; Alexei Starovoitov <ast@...nel.org>; Daniel Borkmann
> > <daniel@...earbox.net>; Jesper Dangaard Brouer <hawk@...nel.org>; John
> > Fastabend <john.fastabend@...il.com>; Wei Fang <wei.fang@....com>;
> > netdev@...r.kernel.org; linux-kernel@...r.kernel.org; imx@...ts.linux.dev;
> > kernel test robot <lkp@...el.com>
> > Subject: [EXT] Re: [PATCH v3 1/1] net: fec: add xdp and page pool statistics
[...]
> Did some testing with the atomic64_t counter, with the following codes to update
> the u64 counter in the end of every NAPI poll cycle.
>
> @@ -1764,7 +1768,13 @@ fec_enet_rx_queue(struct net_device *ndev, int budget, u16 queue_id)
>
> if (xdp_result & FEC_ENET_XDP_REDIR)
> xdp_do_flush_map();
> +#if 1
> + if (xdp_prog) {
> + int i;
> + for(i = 0; i < XDP_STATS_TOTAL; i++)
> + atomic64_add(xdp_stats[i], &rxq->stats[i]);
> + }
> +#endif
> return pkt_received;
> }
>
> With the codes above, the testing result is below:
> root@...8qxpc0mek:~/bpf# ./xdpsock -i eth0
> sock0@...0:0 rxdrop xdp-drv
> pps pkts 1.00
> rx 349399 1035008
> tx 0 0
>
> sock0@...0:0 rxdrop xdp-drv
> pps pkts 1.00
> rx 349407 1384640
> tx 0 0
>
> Without the atomic_add codes above, the testing result is below:
> root@...8qxpc0mek:~/bpf# ./xdpsock -i eth0
> sock0@...0:0 rxdrop xdp-drv
> pps pkts 1.00
> rx 350109 1989130
> tx 0 0
>
> sock0@...0:0 rxdrop xdp-drv
> pps pkts 1.00
> rx 350425 2339786
> tx 0 0
>
> And regarding the u32 counter solution, the testing result is below:
> root@...8qxpc0mek:~/bpf# ./xdpsock -i eth0
> sock0@...0:0 rxdrop xdp-drv
> pps pkts 1.00
> rx 361347 2637796
> tx 0 0
>
> There are about 10K pkts/s difference here. Do we really want the u64 counters?
Where did those atomic64_t come from? u64_stats_t use either plain
u64 for 32-bit platforms or local64_t for 64-bit ones. Take a look
at [0] for the example of how x86_64 does this, it is far from
atomic64_t.
>
> Regards,
> Shenwei
>
> >>
> >> Thanks,
> >> Shenwei
> >>
> >>>
> >>> Andrew
> >
> > Thanks,
> > Olek
[0] https://elixir.bootlin.com/linux/v6.1-rc5/source/arch/x86/include/asm/local.h#L31
Thanks,
Olek
Powered by blists - more mailing lists