On Thu, 2011-09-22 at 12:00 +0200, Peter Zijlstra wrote:

OK, this one seems to be better.. But its quite vile, not sure I
actually like it anymore.

There's also a small matter of it actually being slower than what we had
when tested with: strace hackbench > /dev/null 2>&1.

curses.. even adding a busy wait on p->on_cpu in front.