On Mon, Apr 10, 2017 at 2:30 PM, Andy Gospodarek <andy@xxxxxxxxxxxxx> wrote: > On Mon, Apr 10, 2017 at 03:28:54PM -0400, David Miller wrote: >> From: Andy Gospodarek <andy@xxxxxxxxxxxxx> >> Date: Mon, 10 Apr 2017 14:39:35 -0400 >> >> > As promised, I did some testing today with bnxt_en's implementation >> > of XDP and this one. >> >> Thanks a lot Andy, obviously the patch needs some more work. >> >> I noticed GRO stuff in your profile, and Alexei mentioned this earlier >> today. We probably should elide GRO if generic XDP is attached, since >> in particular this makes the skb_linearize() really expensive. > > Good catch -- I actually thought we were disabling GRO automatically and it > looks like we are not. :-/ I'll send Michael a patch. Andy, I think we only need to disable GRO if we are doing generic XDP. Optimized XDP can still use GRO for the XDP_PASS case. > > Disabling GRO allows me to process an additional 1Mpps, so I'm up to 7.5Mpps > with this patch.
|