Coda File System

Re: would two ethernet interfaces wig out a coda server?

From: <jaharkes_at_cs.cmu.edu>
Date: Thu, 18 Feb 1999 13:47:55 -0500
sean_at_durak.org said:
| remote venus client connections coming in from the outside world
| through the second ethernet interface eth1 can see the coda server OK
| (cmon servername:100 looks good) but never get hooked up. have watched
| venus fail on several remote client machines that otherwise talk just
| fine to the testserver.coda.cs.cmu.edu machine.

btw. with the cmon in 5.0.x you do not need to add the :100 bit anymore.

| does this ring a bell for any coda gurus out there? is some part of
| the coda server or related services only interested in eth0?

No, but I'm interested in finding out what goes wrong here. I've been
modifying RPC2 between 4.6.x and 5.0.x to get better RTT estimates, but
also to attempt to do the right thing in the situation you describe.

| i had a look on the coda FTP site for the rpc2-literate tcpdump, but
| the /pub/tools/ location where this is reputed to live (mentioned in
| codalist archives) doesn't seem to be available. 

I've looked, it's in ftp.coda.cs.cmu.edu:/pub/tools/tcpdump/ both in 
tgz and rpm format. And it would be really helpful to see the server 
side of the traffic ('tcpdump -Trpc2 -ieth1 port venus or port venus-se').

gdt_at_fnord.ir.bbn.com said:
| None of this worked right (crashing the client kernel under 4.6.5, and
| I haven't debugged it further).  I suspect the problem may be in how
| the server chooses the source udp address to use in reply packets, and
| how the client matches these up with its requests.

Every server before 4.7/5.0 sends replies to the host and port written in
the first packet by the client. This has been changed to send replies back
to the host and port as taken from the recvfrom call. The client uses 
the address is gets from the 'GetVolumeInfo' call, which is the address 
stored in the VLDB, most likely the address you get by doing
gethostbyname(gethostname()) on the server, although I haven't checked.

Now that I think of it. I believe that multihomed clients will work,
although they get disconnected when switching addresses (interfaces).
Messages going to multihomed servers will all route to the ip-address
stored in the VLDB.

Jan
Received on 1999-02-18 13:48:44