mbuf leakage with nfs/zfs?

[Willem Jan Withagen]

On 26-2-2010 22:43, Gerrit Kühn wrote:
> DB>  I'll have to do some packet snooping to check if it's TCP or UDP nfs
> DB>  traffic, since some of the clients are Linux ...
>
> I have Linux clients, too. Some use tcp, some udp.

I have Linux and FreeBSD clients running. The build system runs on 
Linux. All Linux's are UDP....

Also connect the build machine to the old 7.2/amd64/bge0/ufs machine, 
but there the count doesn't go over a few 1000 mbufs.

> It did not help. In the meantime the values read
>
> 20555/1465/22020 mbufs in use (current/cache/total)
> 19529/1029/20558/65000 mbuf clusters in use (current/cache/total/max)
> 19529/823 mbuf+clusters out of packet secondary zone in use (current/cache)

Mine are now:
41533/2402/43935 mbufs in use (current/cache/total)
41454/1572/43026/262144 mbuf clusters in use (current/cache/total/max)
39241/823 mbuf+clusters out of packet secondary zone in use (current/cache)

> There certainly is. I wonder when this came in and why it has gone
> unnoticed so far. Probably not all people serving nfs from zfs see this,
> or this would have popped up earlier. Maybe the Linux clients are somehow
> triggering the issue? Or did it start with the import of zvol version 14?
> Unfortunately I have upgraded my pool, so I cannot easily go back to 8-REL
> to test this (otoh, I need a stable server quite urgently).

'mmmm, I did set the zvol version this morning also to 14 but I think 
that I ran into trouble already when still running version 13.

And the server was used as storage for the build system since the last 2 
weeks. Uptil yesterday without much trouble.

--WjW