Subject: Re: NMBCLUSTERS again
To: None <email@example.com>
From: Pavel Cahyna <firstname.lastname@example.org>
Date: 10/31/2007 20:25:08
On Mon, Oct 29, 2007 at 12:44:16AM -0700, Darren Reed wrote:
> email@example.com wrote:
> > hello,
> > I applied the patch but it does not help. The used mbufs are still
> > increasing.
> Ok. At this point I think we need to be able to get at more
> information about mbufs in order to debug the problem further.
> What I'd like to suggest is that we add "char *mh_file" and
> "int mh_line" to "m_hdr" if DEBUG is defined and assign the
> values of __FILE__ and __LINE__ to these fields in MGET and
> MGETHDR if DEBUG is defined. Or even replace both of those
> fields with a "char *mh_allocstack" that shows the kernel's
> call stack.
> But to make that workable, we need to change the pool cache
> slightly - the pool_cache_group needs to carry an array of
> pointers to objects that have been alloc'd, not just those
> that are available to be allocated. A pcg_inuse that is
> the inverse of pcg_objects. Then you can walk pcg_inuse
> for mbufs and at least see who allocated the mbufs in question.
Can't you just adapt ipfilter to use MBUFTRACE instead of such special