Hi!
One of my computers apparently discards all fragmented nfs-packages.
Why?
I have two computers running Debian Woody, kernel 2.4.21, connected
via direct ethernet link (I took the switch out).
Computer "B" can access nfs-exports on computer "A" by
- reading, only when the file size (or size of the directory listing)
is smaller than the mtu (1500).
- writing all sizes with >2MB/s.
When reading, as soon as the file size exceeds the mtu the process
hangs.
When accessing B from A it is the other way around: reading works,
writing does not.
I am no network expert, but this is what Ethereal reports:
B sends ACCESS, GETATTR, LOOKUP, READ Calls....
A anwers accordingly and sends after the READ Call a Reply with an
"Unreassembled package" followed by a set of fragments.
These fragments don't make it to "B"! B re-sends it's READ-Calls and
after some time it sends the ICMP message "time-to-live exceeded".
(If desired I can post or send the complete Ethereal-output).
Where do all the fragments go?
I de-installed firewall and the ipmasq package, the chains in the
iptables are completely empty on both computers, and the policies are
on "ACCEPT".
Of course I can avoid fragmentation with rsize=wsize=1024 or I can
switch to tcp (with not really satisfying speed: read with 35kb/s,
write with 360kb/s) but anyway - I would like to know what's wrong
here.
Where can I look, what else can I check?
Any help is greatly appreciated
Oliver
|