From: Jon M. <jon...@er...> - 2004-05-26 20:03:23
|
Ok, I guess the only right thing to do is to remove it then. I will do it next time I check in code. /jon Mark Haverkamp wrote: On Wed, 2004-05-26 at 12:36, Jon Maloy wrote: Hi Mark, I have not tested TIPC/UDP it on 2.6 at all; last time I did it was with 2.4 on Vmware, not on SMP hardware. There it worked fine, but evidently something has changed in the kernel since then. I have a feeling that we are fighting losing battle here, trying to use UDP in a way not anticipated in the Linux implementation. I also got the feedback from IETF a few months ago that UDP was very undesired as a carrier of TIPC, because it can be used over the Internet in an uncontrollable way, without the inherent flow control they want to se in all internet protocols. (Basically, I think they would like to kill UDP as a protocol altogether, if they could.) If you somhow can find an easy workaround for the problem you see, it is ok to do it, otherwise I don't think we should spend much time on it. We should rather remove it completely for now, and at some later moment concentrate our effort to carry TIPC over TCP or SCTP. I don't think that there is an easy workaround. I looked at the udp code last year with the original tipc and 2.5 and saw the same problems. I tried to use a kernel thread to process the UDP messages so it would be in a process context but never got it to work right. Mark. What I have in mind is that we introduce a generic media adaptation that only has the task of conveying sent/received packets to/from a TIPC deamon in user space. It could be this daemon's task to set up connections and transport the packets from node to node, ensure encryption etc. Performance would of course suffer, but for Internet protocols execution time in the stack is not the dominating factor anyway. For transporting packages to/from such a deamon we can use ioctl, Netlink, or, why not, TIPC itself. But these are future visions, I think. Regards /Jon |