From mboxrd@z Thu Jan 1 00:00:00 1970 From: "David S. Miller" Subject: Re: [RFC] Vegas and tcp parameters per route Date: Mon, 15 Mar 2004 13:37:08 -0800 Sender: linux-net-owner@vger.kernel.org Message-ID: <20040315133708.08f6b900.davem@redhat.com> References: <20040312151729.25d9c696@dell_ss3.pdx.osdl.net> <1079160064.11606.15.camel@viivi> <20040315092018.1e843d83@dell_ss3.pdx.osdl.net> <20040315132214.4d9b5347.davem@redhat.com> <20040315133034.487caee1@dell_ss3.pdx.osdl.net> Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit Cc: pasi.sarolahti@iki.fi, netdev@oss.sgi.com, linux-net@vger.kernel.org Return-path: To: Stephen Hemminger In-Reply-To: <20040315133034.487caee1@dell_ss3.pdx.osdl.net> List-Id: netdev.vger.kernel.org On Mon, 15 Mar 2004 13:30:34 -0800 Stephen Hemminger wrote: > > How do you propose to support some kind of "global enable" for features. > > The easiest way to do that is to initialize each TP with features from > sysctl when created. Right. > > I think sysctl's support this quite well. The test for the feature > > becomes "sysctl || route_attribute". > > That is what the next version does for FRTO. OK, but do not limit it to FRTO, I think all TCP features should be handle'able this way. Actually, this is a problem with using an RTAX_* that is a bitmask or single binary. You need a "don't care" value, and thus effectively a trinary route entry state to do this properly. Right? In this way, you have a global default, but you can also FORCE something off per-route. People can work-around ECN-fux0red sites without obviating ECN completely. Or perhaps you can come up with another method by which to achieve this? > Okay, what about WESTWOOD? ... > What about tcp_westwood which is new? Just don't change sysctl numbers, even we keep them perfectly sync'd between 2.6.x and 2.4.x sources.