[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Fri, 6 Jul 2018 19:38:32 -0700
From: Alexei Starovoitov <alexei.starovoitov@...il.com>
To: Jakub Kicinski <jakub.kicinski@...ronome.com>
Cc: "Waskiewicz Jr, Peter" <peter.waskiewicz.jr@...el.com>,
"Duyck, Alexander H" <alexander.h.duyck@...el.com>,
"daniel@...earbox.net" <daniel@...earbox.net>,
"saeedm@...lanox.com" <saeedm@...lanox.com>,
"brouer@...hat.com" <brouer@...hat.com>,
"borkmann@...earbox.net" <borkmann@...earbox.net>,
"tariqt@...lanox.com" <tariqt@...lanox.com>,
"john.fastabend@...il.com" <john.fastabend@...il.com>,
"netdev@...r.kernel.org" <netdev@...r.kernel.org>,
"saeedm@....mellanox.co.il" <saeedm@....mellanox.co.il>
Subject: Re: [RFC bpf-next 2/6] net: xdp: RX meta data infrastructure
On Fri, Jul 06, 2018 at 06:20:47PM -0700, Jakub Kicinski wrote:
> On Fri, 6 Jul 2018 18:00:13 -0700, Alexei Starovoitov wrote:
> > On Fri, Jul 06, 2018 at 05:40:43PM -0700, Jakub Kicinski wrote:
> > >
> > > Could we just say that at the start we expose only existing SKB fields
> > > (csum, hash, mark) and the meaning of the is the same as in the SKB?
> >
> > what would be the meaning of 'hash' ? Over which fields?
> > Does it support inner and outer packets? How about udp encap (vxlan and friends) ?
>
> We don't seem to need to answer that for the rest of the stack, no? We
> can expose the "hash type" field as well if that's *really* necessary.
>
> > Same question of csum... tcp only?
>
> Shouldn't we just stick to CHECKSUM_COMPLETE?
I don't yet see how checksum_complete and 'hash just like stack'
helps to accelerate xdp programs like ddos and load balancer.
'hash like stack' may help to accelerate networking stack itself,
but that's very different discussion. There is no xdp in such case.
No programs and no metadata. If some NICs can do
'hash like stack' and demonstrate performance improvements
for regular tcp/ip processing with netperf in user space
that would definitely be interesting and standardizing such
hash from the drivers into upper layers would warrant its own patches,
but seems orthogonal to this xdp hints discussion.
Powered by blists - more mailing lists