I've just pushed 2 commits related to GSO peeling behaviour to master.

1st tweak is at worst benign and at best removes a multiply compare for
every packet enqueued.  I'd like to think the optimiser in the compiler
would have done what I've done explicitly (in essence check this is a
gso packet 1st before thinking about peeling it) but when I checked on
x86_64 there was a definite difference in produced code.

2nd tweak is *not* benign.  In essence this forces peeling if either ATM
framing or packet overhead is specified.  Previously only ATM framing
forced peeling.  I think this is more correct but unfortunately will be
slower.

Commits can be reverted - feel free :-)