[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Thu, 08 Aug 2013 16:55:14 -0700
From: Eric Dumazet <eric.dumazet@...il.com>
To: Eliezer Tamir <eliezer.tamir@...ux.intel.com>
Cc: Shawn Bohrer <sbohrer@...advisors.com>, davem@...emloft.net,
netdev@...r.kernel.org, Amir Vadai <amirv@...lanox.com>,
tomk@...advisors.com
Subject: Re: [PATCH v2 net-next] net: Add low-latency/polling support for
UDP multicast
On Thu, 2013-08-08 at 11:46 +0300, Eliezer Tamir wrote:
> On 07/08/2013 23:22, Eric Dumazet wrote:
> > On Tue, 2013-08-06 at 14:51 -0500, Shawn Bohrer wrote:
> >> Set the napi id for each socket in the multicast path to enable
> >> low-latency/polling support.
> >>
> >> Signed-off-by: Shawn Bohrer <sbohrer@...advisors.com>
> >> ---
> >> v2 include ipv6 support
> >
> > This might help your workload, but I doubt it is generic enough.
> >
> > One UDP socket is supposed to receive traffic from many endpoints,
> > so we have no guarantee all the received traffic will end on a single RX
> > queue on the NIC.
> >
> > That's the same logic than RFS here.
> >
> > sk_mark_napi_id() in UDP are wrong IMHO.
> >
> > It should be guarded by the following test in
> > __udp_queue_rcv_skb()
> >
> > if (inet_sk(sk)->inet_daddr) {
> > sock_rps_save_rxhash(sk, skb);
> > sk_mark_napi_id(sk, skb);
> > }
> >
> > (To occur only for connected UDP sockets, where we are 100% sure all
> > packets will use this same rxhash/rx queue)
>
> This would also be safe if there is only one NIC and said NIC was
> programmed to always place this socket's data on the same queue.
>
> I don't have a good suggestion on how to detect this.
Well, this stuff relies on flows being correctly steered.
TCP stack performs much better if we avoid reorders ;)
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Powered by blists - more mailing lists