Hi everybody,
Some time ago I posted a question on NFS traffic not following static
routes on systems with two network interfaces. Yesterday, Eric Werme
suggested that I should try to set use_fastroute kernel variable to 0.
This had indeed fixed our problem.
I am including Eric's e-mail in full (with his permission). The only
thing I'd like to add is that 'assign' will change the memory location
of the corresponding variable. To make change permanent (until the
next kernel build) use 'patch' instead of 'assign'.
/Serge.P
> From: Eric Werme USG <werme_at_zk3.dec.com>
>
> here comes the strange bit I can't understand. Static host routes
> on the server side work fine for all TCP traffic, as well as UDP
> traffic EXCEPT for packets originating from the NFS server (port 2049).
>
> NFS does some kind of kinky stuff to try to bypass SMP routing locks.
> I'm not certain that area affects what you're seeing. I'd look at the code,
> but you didn't what release you're running.
>
> Try:
>
> # dbx -k /vmunix
> (dbx) assign use_fastroute = 0
>
> When that's one, the server tries to send replies out the same interface that
> received the request if the client is on the same subnet. (Yes, I realize
> you're seeing replies go out another interface.)
>
> (dbx) assign use_fastsend = 0
>
> This disables code that completely bypasses the conventional UDP and IP
> code and generates its own messages itself. With this zero, replies
> will go through the same paths non-NFS messages use. The fastroute
> code is part of the fastsend code too, so setting use_fastsend = 0
> implies use_fastroute = 0.
>
> The whole situation is mostly a curiosity in our setup: a single
> 100Mbit link handles our steady-state NFS requirements quite well,
> and the routes do work for the most network-intensive part of the
> startup. Still, the fact that NFS UDP (and ONLY NFS UDP) traffic
> does not obey the static route means that there is something
> strange going on - and I'd like to know why...
>
> The fastroute stuff has always been questionable, but it avoids a nasty
> SMP lock so has been a big win. I'm not at all certain you're using it,
> but it's a good place to start.
>
> Also, NFS server input bypasses UDP and socket processing, and that has the
> code that records what interface received the request. That's not available
> to other UDP applications.
>
> -Ric Werme
>
> --
> Eric (Ric) Werme | werme_at_zk3.dec.com
> Compaq Computer Corp. | http://www.cyberportal.net/werme
>
Received on Fri Mar 19 1999 - 19:00:50 NZDT