Memory leak in NFSv4?

[J. Bruce Fields]

On Mon, Dec 04, 2006 at 04:30:53PM +0100, Fredrik Tolf wrote:
> Do you mean unmounting it from *all* clients, or just one? I can confirm
> that it doesn't go back down if I unmount it from just one client, at
> least (I just tried). It is, by the way, up to 334 right now.

Sorry, I was confused--I thought you were describing a client-side
problem, but sounds like it's actually on the server side.

> I haven't tried unmounting it from all clients with the 2.6.18 kernel,
> and doing that would be quite troublesome. However, I did try it once
> with the 2.6.16 kernel, and it did not bring the refcount down (at that
> time, it was up to over 500).

Yeah, it may be that the server's context cache just isn't designed
right--I don't think it ever evicts a context until it expires, which
could take a long time.

And maybe the reference counting should be based on exports instead of
contexts, to make it possible to remove the gss-related modules if you
unexport first.

Hm.

--b.