Re: [PATCH v2 0/9] Remove spin_unlock_wait()

[Peter Zijlstra]

On Thu, Jul 06, 2017 at 12:49:12PM -0400, Alan Stern wrote:
> On Thu, 6 Jul 2017, Paul E. McKenney wrote:
> 
> > On Thu, Jul 06, 2017 at 06:10:47PM +0200, Peter Zijlstra wrote:
> > > On Thu, Jul 06, 2017 at 08:21:10AM -0700, Paul E. McKenney wrote:
> > > > And yes, there are architecture-specific optimizations for an
> > > > empty spin_lock()/spin_unlock() critical section, and the current
> > > > arch_spin_unlock_wait() implementations show some of these optimizations.
> > > > But I expect that performance benefits would need to be demonstrated at
> > > > the system level.
> > > 
> > > I do in fact contended there are any optimizations for the exact
> > > lock+unlock semantics.
> > 
> > You lost me on this one.
> > 
> > > The current spin_unlock_wait() is weaker. Most notably it will not (with
> > > exception of ARM64/PPC for other reasons) cause waits on other CPUs.
> > 
> > Agreed, weaker semantics allow more optimizations.  So use cases needing
> > only the weaker semantics should more readily show performance benefits.
> > But either way, we need compelling use cases, and I do not believe that
> > any of the existing spin_unlock_wait() calls are compelling.  Perhaps I
> > am confused, but I am not seeing it for any of them.
> 
> If somebody really wants the full spin_unlock_wait semantics and
> doesn't want to interfere with other CPUs, wouldn't synchronize_sched()
> or something similar do the job?  It wouldn't be as efficient as
> lock+unlock, but it also wouldn't affect other CPUs.

So please don't do that. That'll create massive pain for RT. Also I
don't think it works. The whole point was that spin_unlock_wait() is
_cheaper_ than lock()+unlock(). If it gets to be more expensive there is
absolutely no point in using it.