[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Sun, 2 Feb 2020 14:51:04 -0700
From: David Ahern <dsahern@...il.com>
To: Jakub Kicinski <kuba@...nel.org>
Cc: Toke Høiland-Jørgensen <toke@...hat.com>,
netdev@...r.kernel.org, prashantbhole.linux@...il.com,
jasowang@...hat.com, davem@...emloft.net, jbrouer@...hat.com,
mst@...hat.com, toshiaki.makita1@...il.com, daniel@...earbox.net,
john.fastabend@...il.com, ast@...nel.org, kafai@...com,
songliubraving@...com, yhs@...com, andriin@...com
Subject: Re: [PATCH bpf-next 03/12] net: Add IFLA_XDP_EGRESS for XDP programs
in the egress path
On 2/2/20 12:31 PM, Jakub Kicinski wrote:
>
> I think our perspectives will be difficult to reconcile.
>
> My reading is that you look at this from slow software device
> perspective. Of which there are few (and already loaded with hacks).
> And you want the control plane to be simple rather than performance.
>
> I look at this from HW driver perspective of which there are many.
> Saying that it's fine to load TX paths of all drivers with extra
> code, and that it's fine to effectively disable SG in the entire
> system just doesn't compute.
>
> Is that a fair summary of the arguments?
>
I do not believe so.
My argument is about allowing XDP and full stack to work
synergistically: Fast path for known unicast traffic, and slow path for
BUM traffic without duplicating config such as ACLs and forcing a
specific design (such as forcing all programs in the XDP Rx path which
does not work for all traffic patterns). For example, if I am willing to
trade a few cycles of over head (redirect processing) for a simpler
overall solution then I should be to do that. Fast path with XDP is
still faster than the full host stack for known unicast traffic.
Powered by blists - more mailing lists