Date: Wed, 23 May 2012 13:52:55 +0530 From: Venkat Duvvuru <venkatduvvuru.ml@gmail.com> To: lstewart@freebsd.org Cc: freebsd-net@freebsd.org, "Bjoern A. Zeeb" <bz@freebsd.org>, freebsd-questions@freebsd.org, Jack Vogel <jfvogel@gmail.com> Subject: Re: LRO support for IPv6 Message-ID: <CAGdae7bx48Q8nk0Qf%2BKwMb9kcCcsvK_VG5N5_iGxhxDpMwV=gQ@mail.gmail.com> In-Reply-To: <CAGdae7a%2Baf7RbijyKfMA94MfomNaTaD3M3PPmcFk%2B13SC9cMAg@mail.gmail.com> References: <CAGdae7bcWqGbObygPdZwCyVG4Pe-0Fxq5_p19oCp61uzZ4N8xw@mail.gmail.com> <CAFOYbc=ng83L8iZ_N79edqhknOrDPqVA9ASMexW5-yU0vnduDQ@mail.gmail.com> <A8DB9672-6B84-4635-84B6-43CC98B2877F@FreeBSD.org> <CAFOYbckwRw4jazwqY1S7X2wiSdBBBPdg-Xk8ya99j1%2BWbqB=DA@mail.gmail.com> <72B744D5-3D24-4A56-907C-2A8F6620877B@FreeBSD.org> <CAGdae7ZX4M4NPnJ=3K1vA9TLPW22r2EpTic-Y4kJMyJQGn3zGw@mail.gmail.com> <CAFOYbcncAGPA6d7qh7bonGy2ijcApD_TQgqvSoM2Mbif-z8sYg@mail.gmail.com> <CAGdae7a%2Baf7RbijyKfMA94MfomNaTaD3M3PPmcFk%2B13SC9cMAg@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
Folks, Can somebody please explain me why "tcp checsum" calculation is mandated in the freebsd network stack (tcp_input--->in6_cksum) albeit the card supports it? Probably Steve is the right person who can answer this. /Venkat On Wed, May 23, 2012 at 11:27 AM, Venkat Duvvuru <venkatduvvuru.ml@gmail.com > wrote: > Ok. I found the reason for the throughput drop in case of IPv6. > Reason is that the "tcp check sum" calculation is mandated in case of IPv6 > irrespective of whether the card is doing it or not (checksum offload). Is > there a reason why freebsd is doing it that way? > > /Venkat > > On Tue, May 22, 2012 at 11:16 PM, Jack Vogel <jfvogel@gmail.com> wrote: > >> LRO is a huge win for 10G (as is TSO on the TX side), so odds are good >> its behind the drop, >> in any case you'll be able to test that soon :) >> >> Jack >> >> >> >> On Tue, May 22, 2012 at 10:35 AM, Venkat Duvvuru < >> venkatduvvuru.ml@gmail.com> wrote: >> >>> Thanks for the response. >>> >>> I observed that there is a significant performance drop in case of IPv6 >>> on the "rx" side. >>> While I'm able to hit line rate ~9.5 Gbps on a 10gb NIC for IPv4..I >>> could only get ~6 Gbps on the "rx" front for IPv6...However "tx" for IPv6 >>> is on par with IPv4 hitting almost line rates. >>> >>> Could this be because of lack of LRO6?? >>> >>> Note: hwpmc profiling shows that most of the time is spent in the IPv6 >>> stack code >>> >>> /Venkat >>> On Tue, May 22, 2012 at 10:37 PM, Bjoern A. Zeeb <bz@freebsd.org>wrote: >>> >>>> >>>> On 22. May 2012, at 17:04 , Jack Vogel wrote: >>>> >>>> > Oh, that's right, distracted with other projects and I forgot, now we >>>> just need >>>> > to have an LRO that works with forwarding eh :) >>>> >>>> That's a 6 line bainaid commit afterwards, basically returning form the >>>> LRO queuing >>>> function in case forwarding is turned on for that address family; a >>>> proper solution >>>> for long term can than be done whenever we feel like it. The above we >>>> should have done >>>> years ago;) >>>> >>>> >>>> > You ROCK bz :) >>>> > >>>> > Jack >>>> > >>>> > >>>> > On Tue, May 22, 2012 at 10:01 AM, Bjoern A. Zeeb <bz@freebsd.org> >>>> wrote: >>>> > >>>> > On 22. May 2012, at 16:50 , Jack Vogel wrote: >>>> > >>>> > > The LRO code as it stands right now is IPV4 specific, it would be >>>> nice to >>>> > > extend it, one of >>>> > > many improvements that may get done at some point. >>>> > >>>> > I am about to commit it to HEAD. Bear another few days with me; I >>>> know >>>> > I am running late but committing new code had less prio than some >>>> other >>>> > real life things currently. >>>> > >>>> > I'll also bring TSO6, etc... >>>> >>>> -- >>>> Bjoern A. Zeeb You have to have visions! >>>> It does not matter how good you are. It matters what good you do! >>>> >>>> >>> >> >
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CAGdae7bx48Q8nk0Qf%2BKwMb9kcCcsvK_VG5N5_iGxhxDpMwV=gQ>