I'd like to know this myself. I noticed the inferior performance of kernel 4.9 back in December with Milhouse build #1215. The issue is easy to reproduce as long as you have a DVB tuner connected to the Pi and PVR server/client are on the same Pi. The difference in performance is huge, with kernel 4.8 I can run this setup without a single issue (with Advanced deinterlacing enabled). With kernel 4.9 I absolutely have to disable Advanced deinterlacing (even increasing arm_uc doesn't help).
I'll need to dig in to see exactly what changed between 4.8 and 4.9.
4.9 kernel uses more upstream drivers, including the upstream interrupt controller. 4.8 may have still been using the downstream one (I'd have to check).
That could mean either a difference in speed to dispatching the ISR (I would have thought that difference would be small) or a difference in precedence
(i.e. if a USB and a timer ISR are both pending which gets dispatched first) which may have more of an effect.
There is also a FIQ driver that USB uses.
Can you point sakos at your build to see if it helps his issue too?