Re: 2.613: network write socket problems

On Monday 12 September 2005 17:39, Bernd Schubert wrote:

Hello,

on last Friday  we switched on our server to 2.6.13 and today we are
experiencing problems with our nfs clients.
In particular I'm talking about the unfs3 daemon, not the kernel nfs
daemon. Both are running on the server but on different ports, of course. 
Both are also serving to the same clients, but different directories.

Today it already several times happend that the unfs3 daemon stalled.
Ethereal showed no network packages on the unfs3 daemon port during this
time. A strace to the proc-id of the daemon clearly shows that *some*
writes to some network sockets will take ages to finish

write(37, "\200\0\0x\203\326(\5\0\0\0\1\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 124)
= 124

Sorry for the noise, its not a kernel problem. Switching back to 2.6.11 didn't 
help, so we investigated further. It turned out, that one of our clients was 
in a kind of a zombie state and asking for filehandles, but not answering 
request from the server. Since unfs3 is only single threaded, all other 
clients had to wait for timeouts.

Bernd