Re: [RFC PATCH v2 3/5] futex: Throughput-optimized (TO) futexes

[Waiman Long]

On 09/22/2016 09:23 AM, Peter Zijlstra wrote:
> On Tue, Sep 20, 2016 at 09:42:41AM -0400, Waiman Long wrote:
> > +/*
> > + * Spinning threshold before enabling lock handoff.
> > + * Each sleep will decrement the threshold by 1/32 of the start value.
> > + */
> > +#define TO_SPIN_THRESHOLD	(1<<  13)
> > +#define TO_SLEEP_DECREMENT	(TO_SPIN_THRESHOLD/32)
> Argh, we should really get rid of those stupid numbers. Wasn't there a
> patch set working on implementing paravirt functions that would make all
> this fixable in a sane way?

These thresholds are not paravirt related. They are there to determine 
when we should activate explicit lock handoff when the waiter is waiting 
for too long. Yes, it is arbitrary. The numbers are on the high side as 
setting them too low will limit lock stealing and hence overall 
performance. The FUTEX_WAITERS bit is only turned on when either the top 
waiter is sleeping or a lock handoff is needed. Otherwise, it is off to 
encourage lock stealing in the userspace as well as in the kernel.

Cheers,
Longman