On Thu, Apr 11, 2013 at 08:54:35AM -0700, Eric Dumazet wrote: > On Thu, 2013-04-11 at 17:32 +0200, Willy Tarreau wrote: > > On Thu, Apr 11, 2013 at 05:27:03PM +0200, Sebastian Hesselbarth wrote: > > > I don't have a strong opinion on whether Soeren's or your proposal should > > > be submitted. But I insist on having one of them in, as GRO significantly > > > improves the common use case, is enabled by default, and not as > > > constrained as LRO. > > > > I agree, use yours first, but we should keep an eye on this. Since you have > > everything to run a test, please try to see if you can get netperf to run > > over IPv6, I'm sure the NIC doesn't handle it. > > Willy, testing the checksum in the NIC driver itself prevents the stack > doing GRO even if the NIC could not checksum the packet, as in GRE > tunneling for example. > > So Sebastien patch is better IMHO : Just call the napi gro handler and > let core stack handles the details ;)
OK, that makes sense indeed, I didn't think about this case. All I remember was that the old call achieved a higher packet rate than napi_gro_receive, but it was on an older kernel and I can't be more specifics after several months :-/ Cheers, Willy _______________________________________________ Linuxppc-dev mailing list Linuxppc-dev@lists.ozlabs.org https://lists.ozlabs.org/listinfo/linuxppc-dev
