Re: NFS over IPv6

* Jun Kuriyama <kuriyama_at_imgsrc.co.jp> [040711 19:22] wrote:
> At Thu, 8 Jul 2004 14:09:04 +0000 (UTC),
> kuriyama wrote:
> > With today's kernel, I see strange behavior with NFS over IPv6.  My
> > userland is two days older.  I'll check with sync'ed kernel/userland
> > tomorrow.
> > 
> > # NFS server is 5.2.1R, client is current.
> > 
> > When I did some write operation, it may wait too long.  I can
> > sometimes write without problem, but sometimes not.  I took tcpdump -vvv:
> > 
> > 22:48:25.129335 2001:218:422:2::YYYY.2049 > 2001:218:422:2::XXXX.1516376639: reply ok 160 write PRE: sz 0xd1ea mtime 1089294496.000000 ctime 1089294496.000000 POST: REG 664 ids 1000/1000 sz 0xd1ea nlink 1 rdev 199/65077248 fsid 0x420 fileid 0xf8b440 a/m/ctime 1089287695.000000 1089294502.000000 1089294502.000000 4586 bytes <unstable> (len 168, hlim 64)
> > 22:48:25.129489 2001:218:422:2::XXXX > 2001:218:422:2::YYYY: [icmp6 sum ok] icmp6: 2001:218:422:2::XXXX udp port 829 unreachable (len 216, hlim 64)
> > 
> > I don't know why client side replied as port unreachable.  With IPv4,
> > there is no problem.
> 
> I tracked to find which commit breaks.
> 
> With 2004-07-06 06:00:00+00, it's OK.  But with 2004-07-06
> 10:00:00+00, above problem occured.  There is alfred's NFS mobility
> commits between them.
> 
> Alfred, do you have any ideas?

If there is a timeout, NFS will rebind the socket.  In theory this should
be ok as NFS is stateless.  Perhaps there is some IPv6 thing that makes
rebound sockets not work right?

Is there a reason why that might be the case?

I'm going to make a patch to set the SO_REUSEADDR and SO_REUSEPORT
on the sockets.

I'm also going to bump the timeout from 5 to 12 seconds.

-- 
- Alfred Perlstein
- Research Engineering Development Inc.
- email: bright_at_mu.org cell: 408-480-4684