Re: [RFC] Make need_resched() return true when rcu_urgent_qs requested

[Peter Zijlstra]

On Fri, Jul 06, 2018 at 06:14:44PM +0100, David Woodhouse wrote:
> On Fri, 2018-07-06 at 10:11 -0700, Paul E. McKenney wrote:
> > > The preempt state is alread a bit complicated and shadowed in the
> > > preempt_count (on some architectures) adding additional bits to it like
> > > this is just asking for trouble.
> > 
> > How about a separate need_resched_rcu() that includes the extra cache
> > miss?  Or open-coding the rcu_urgent_qs_requested()?
> 
> Peter said "touch two cachelines". He didn't say it was a cache miss.
> 
> Given that every single cond_resched() call touches the same cache
> line, and every single rcu_all_qs() and similar will also touch it,
> it's fairly much guaranteed *not* to be a miss...

cond_resched() is a no-op for PREEMPT=y, but then you're still
sprinkling that read all across.

> ... which is why I didn't really understand why he cared.

Well it also complicated the whole preemption state, and like I wrote,
TIF_NEED_RESCHED is not the only place we track preemption state in.

Stuff like preempt_enable() will (on x86) only ever look at the
preempt_count value.