Re: [NFS] [patch] sunrpc: make closing of old temporary socketswork (was: problems with lockd in 2.6.22.6)
From: J. Bruce Fields
Date: Wed Sep 12 2007 - 09:38:18 EST
On Wed, Sep 12, 2007 at 02:07:10PM +0200, Wolfgang Walter wrote:
> as already described old temporary sockets (client is gone) of lockd aren't
> closed after some time. So, with enough clients and some time gone, there
> are 80 open dangling sockets and you start getting messages of the form:
>
> lockd: too many open TCP sockets, consider increasing the number of nfsd threads.
Thanks for working on this problem!
> If I understand the code then the intention was that the server closes
> temporary sockets after about 6 to 12 minutes:
>
> a timer is started which calls svc_age_temp_sockets every 6 minutes.
>
> svc_age_temp_sockets:
> if a socket is marked OLD it gets closed.
> sockets which are not marked as OLD are marked OLD
>
> every time the sockets receives something OLD is cleared.
>
> But svc_age_temp_sockets never closes any socket though because it only
> closes sockets with svsk->sk_inuse == 0. This seems to be a bug.
>
> Here is a patch against 2.6.22.6 which changes the test to
> svsk->sk_inuse <= 0 which was probably meant. The patched kernel runs fine
> here. Unused sockets get closed (after 6 to 12 minutes)
So the fact that this changes the behavior means that sk_inuse is taking
on negative values. This can't be right--how can something like
svc_sock_put() (which does an atomic_dec_and_test) work in that case?
I wish I had time today to figure out what's going on in this case. But
from a quick through svsock.c for sk_inuse, it looks odd; I'm suspicious
of anything without the stereotyped behavior--initializing to one,
atomic_inc()ing whenever someone takes a reference, and
atomic_dec_and_test()ing whenever someone drops it....
--b.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/