gah@hood.hood.caltech.edu (Glen Herrmannsfeldt) (02/28/91)
On TCP/IP routing loops. We have two routers between our subnet and the main campus network. It mostly seems to work fine. One is an HP9000 and the other is a Sun/3. The HP runs gated, and the sun runs routed. Other than mysterious problems that just go away, which most networks have once in a while, I once traced a MacTCP/Pathways problem to what could be considered a route loop problem. (Pathways is a client NFS for the mac.) The problem seemed to be that the packets coming back from the sun server came back from the other port than the one that the request went to. They were sent to the proper address, but the machine ignored them. That is like throwing away any mail that has a return address different from the address you sent it to. (Reminds me of what our receiving department did once, but that is another story. If you order ethernet cable, and it never appears, call up receiving and ask them if they have a big spool of cable sitting there.) Right now, we are in a different kind of intermediate state, and seem to still be running. The main campus subnet will be split when a new cisco router arrives. In the mean time, they have convinced the main router to support two subnet addresses on one ethernet cable. Machines think they are on two different physical nets, even though they are not. Now, I have one of our two routers on each, making a really big routing loop. Routers on the new subnet are not allowed to use RIP because suns (maybe others) complain if they see a routing packet that is addressed to a different subnet. Anyway, we have done it, and it works. It helps a lot to know what it is doing, though. Use netstat -rn to look at the routing tables, and see if they make sense. The TTL in tcp/ip should get rid of circulating packets, if it ever occurs. It should not, though.