netdev - RE: [PATCH net-next] dpaa2-eth: Don't use netif_receive_skb

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Message-ID: <AM0PR04MB499429A871F6037E483CFDFA94C60@AM0PR04MB4994.eurprd04.prod.outlook.com>
Date:   Wed, 24 Jul 2019 08:15:30 +0000
From:   Ioana Ciocoi Radulescu <ruxandra.radulescu@....com>
To:     David Miller <davem@...emloft.net>
CC:     "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
        Ioana Ciornei <ioana.ciornei@....com>,
        Vladimir Oltean <vladimir.oltean@....com>,
        Eric Dumazet <eric.dumazet@...il.com>,
        Edward Cree <ecree@...arflare.com>
Subject: RE: [PATCH net-next] dpaa2-eth: Don't use netif_receive_skb_list for
 TCP frames

> -----Original Message-----
> From: David Miller <davem@...emloft.net>
> Sent: Wednesday, July 24, 2019 12:03 AM
> To: Ioana Ciocoi Radulescu <ruxandra.radulescu@....com>
> Cc: netdev@...r.kernel.org; Ioana Ciornei <ioana.ciornei@....com>; Vladimir
> Oltean <vladimir.oltean@....com>
> Subject: Re: [PATCH net-next] dpaa2-eth: Don't use netif_receive_skb_list for
> TCP frames
> 
> From: Ioana Radulescu <ruxandra.radulescu@....com>
> Date: Tue, 23 Jul 2019 20:28:43 +0300
> 
> > Using Rx skb bulking for all frames may negatively impact the
> > performance in some TCP termination scenarios, as it effectively
> > bypasses GRO.
> 
> "may"?
> 
> Please provide numbers so that we know exactly whether it actually
> hurts performance one way or another.

We observed the worst degradation running netperf TCP_STREAM on a
setup with two 16cores LX2160A boards connected back-to-back, one
stream per cpu:
With netif_receive_skb_list() on all packets, we get a total bandwidth
of 41.6Gbps, with rx cpu load of 97%; after applying current patch, bw
increases to 45.8Gbps with rx cpu at 64%, which is similar to what we
got before using netif_receive_skb_list() in the first place.

On other platforms/setups the impact is lower, in some cases there is
no difference in throughput, only higher cpu load when skb batching
is used for TCP packets.

Anyway, based on feedback so far I guess the best path forward is to
withdraw this patch and wait for Edward's GRO batching work to be
accepted. I can help test those patches if needed.

Thanks,
Ioana