I can very easily make processes get stuck in nfsfsync by writing to NFS.
This only happen in one configuration though, with one machine as client
and the other as server. If I switch the roles or run both server and
client on one machine, it doesn't happen. I've tested with both UDP and
TCP. It seems to get stuck earlier when using UDP.
Could there be some timing issues in the NFS code?
I mounted NFS over TCP and did some writes and captured the traffic with
tcpdump. I did some small writes first, but then I did
dd if=/dev/zero bs=1k count=10k >zeroes
The first time I ran it, it finished ok. The second time it got stuck.
I've put dmesgs and tcpdumps at http://people.su.se/~rnyberg/dfly/nfsfsync.
Please tell me if you need some other information.
-Richard