netdev-owner@vger.kernel.org wrote on 05/10/2010 10:25:57 AM:

On Mon, May 10, 2010 at 10:09:03AM -0700, David Stevens wrote:

quoted

Since "datalen" carries the difference and will be negative by that 

amount

quoted

from the original loop, what about just adding something like:

        }
        if (headcount)
                heads[headcount-1].len += datalen;
[and really, headcount >0 since datalen > 0, so just:

        heads[headcount-1].len += datalen;

                                                +-DLS

This works too, just does more checks and comparisons.
I am still surprised that you were unable to reproduce the problem.

I'm sure it happened, and probably had a performance
penalty on my systems too, but not as much as yours.
I didn't see any obvious performance difference running
with the patch, though; not sure why. I'll instrument to
see how often it's happening, I think.
        But fixed now, good catch!

                                                        +-DLS