On Thu, Aug 24, 2006 at 11:00:13PM +0200, Pavel Cahyna wrote:
> [ please reply to tech-net@ ]
> On Thu, Aug 24, 2006 at 12:42:26AM +0000, Christos Zoulas wrote:
> > So we see from the data that you have the same problem like I do. We are
> > leaking mbufs in the tcp recv. I've been reading the code, but I have not
> > found a problem...
> 
> recv? don't you mean transmit?
> 
> I would try compiling and testing a kernel with options SOSEND_NO_LOAN.

I have seen a case where a network interface reset leaked any
outstanding mbufs queued on that interface.  That's a driver-
specific issue, though, and I haven't checked all of the drivers
in the tree to see if any have that error.  Most are, I think, fine
in this area.  I just checked fxp(4), wm(4), bge(4), and bce(4)
(since I use all of those ;-), and they all seem fine--they unload
the dmamap and m_freem() in xxx_stop() or similar.

-allen

-- 
Allen Briggs  |  http://www.ninthwonder.com/~briggs/  |  briggs@ninthwonder.com