Date: Fri, 16 Mar 2018 16:04:04 +0000 From: Patrick Kelsey <pkelsey@freebsd.org> To: freebsd-hackers@freebsd.org Subject: Re: option TCP_RFC7413 is not in GENERIC Message-ID: <CAD44qMUR_bKOzaWLMatGT3xXWdovWm3Abv8zwu=NCVm6YjEkiQ@mail.gmail.com> In-Reply-To: <20180314233811.GA35025@mail.bsd4all.net> References: <1521062028.2511351.1303413736.6960BF4F@webmail.messagingengine.com> <20180314233811.GA35025@mail.bsd4all.net>
next in thread | previous in thread | raw e-mail | index | archive | help
On Wed, Mar 14, 2018 at 7:39 PM Herbert J. Skuhra <herbert@gojira.at> wrote: > On Wed, Mar 14, 2018 at 04:13:48PM -0500, Mark Felder wrote: > > Hello all, > > > > Has there been a discussion about enabling this in GENERIC? Is there a > performance impact even if the sysctl knob is disabled? What can we do to > get a wider audience of testers? I would gladly test this for apps that > support it but the hurdle of requiring a custom kernel to test this is more > effort than its worth... > > > https://svnweb.freebsd.org/base?view=revision&revision=330002 > > I plan on merging all of the TFO deltas between current and stable/11 to stable/11 ahead of the 11.2 release process. tuexen@ has been doing some good bug hunting on the new client-side implementation in -current and there are other parties testing it as well. I think there is still technically a question as to whether it will be enabled by default in 11.2, only because there is still some poking and prodding we have left to do, but I also think we will complete all of that in time. As to performance impact when it is compiled in, but disabled, I have heard that some small differences have been measured, but I've seen neither the methodology used nor the actual measurement results. Mechanically what is going on in that case is you will wind up with some additional flag checks in the TCP input and output paths and in some cases additional instructions that you don't need being pulled into the I-cache, compared to what would occur if it was compiled out. The current thinking is that users who care about such performance differences are dealing with extreme workloads that already motivate them to compile their own kernels, or are working with very resource-constrained platforms, so the way forward is to keep the TCP_RFC7413 kernel option around and enable it by default for the server-class platforms (armd64 and arm64). Best, Patrick
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CAD44qMUR_bKOzaWLMatGT3xXWdovWm3Abv8zwu=NCVm6YjEkiQ>