[83] in linux-net channel archive
Re: tcp_data bug acked < copied - still in 1.1.88
daemon@ATHENA.MIT.EDU (Johannes Stille)
Sun Feb 12 09:41:46 1995
Date: Sat, 11 Feb 95 20:13 MET
From: johannes@titan.westfalen.de (Johannes Stille)
To: linux-net@vger.rutgers.edu
Cc: johannes@titan.westfalen.de
Marek Michalkiewicz wrote:
[...]
>
>Another problem: the routing table becomes full of never used routes
>to network addresses, but with incorrect netmask (255.255.255.255).
What do you mean by "network adresses"? The host part of the address is
zero? Are these addresses from subnets of your network or from other
networks?
>I guess this has something to do with ICMP NET REDIRECT handling...
Yes, probably (either this or HOST REDIRECT).
>Most of these routes are redundant (the gateway is the same as the
>default one) and maybe they shouldn't be added at all?
In this case, the gateway should not send NET REDIRECT messages.
>If it helps: at least some of these subnets should have netmasks
>like 255.255.255.224 (not on byte boundaries).
Just because of the possible problems with subnetting, it is officially
recommended to treat NET REDIRECT as HOST REDIRECT. And this is what
Linux does.
>This long routing table doesn't seem to cause any problems so far,
>but I guess it may take longer to find the correct route (for every
>packet) in such a large table... (150+ entries after few days up)
That's it: no problems, just increased CPU usage.
>If this is a bug, are there any chances of fixing it before 1.2?
>Or am I missing something? Maybe it is a feature, not a bug? :-)
If the routes are redundant, it's an error in your gateway's software
or configuration.
There might be some general problem with NET REDIRECT handling:
-If the destination address in the redirect message is a real IP
network (host part all zero, including subnet part), the Linux kernel
should probably add a network route with the default class A/B/C mask,
because a host route is useless.
-If the destination address in the redirect message is a subnet address
(with a non-zero subnet part, and the remaining host part zeroed), then
we're in trouble: There is no way to know the netmask, as that subnet
might have a different netmask than the one we're on. So probably we
should continue to treat it like a HOST REDIRECT, as generally
recommended by RFC 1122.
Johannes
