Comment 36 for bug 1021471

Certainly something is very wrong with the numbers. Looking at the numbers reported we get 5/13/-17/0 which sums up to 1 which is exactly the refcount complained about. But looking at the individual numbers, the counter for CPU#0 starts with 5 not 0.
And just adding up the dev_hold and dev_put calls I underrun down to -4. So it looks a bit like some magic suddenly warps the counter and we release more often than we appear to take the reference and still end up one too high.
It is a bit too late to think about it but maybe loopback gets assigned elements from another interface and that count is off by one.