I agree that your trial-balloon should likely get rid of the big
regression, since it avoids the scheduler.
So with your patch, lock_kernel() ends up being just a rather expensive
spinlock. And yes, I'd expect that it should get rid of the 40% cost,
because while it makes lock_kernel() more expensive than a spinlock and
you might end up having a few more cacheline bounces on the lock due to
that, that's still the "small" expense compared to going through the whole
scheduler on conflicts.
So I'd expect that realistically the performance difference between your
version and just plain spinlocks shouldn't be *that* big. I'd expect it to
be visible, but in the (low) single-digit percentage range rather than in
any 40% range. That's just a guess.
Linus
--
