Re: [RFC][PATCH 5/5] powerpc: Remove SYNC from _switch

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Jun 08, 2017 at 05:29:38PM +1000, Nicholas Piggin wrote:
> On Thu, 8 Jun 2017 08:54:00 +0200
> Peter Zijlstra <peterz@infradead.org> wrote:
> 
> > On Thu, Jun 08, 2017 at 10:32:44AM +1000, Nicholas Piggin wrote:
> > > On Wed, 07 Jun 2017 18:15:06 +0200
> > > Peter Zijlstra <peterz@infradead.org> wrote:
> > >   
> > > > Now that the scheduler's rq->lock is RCsc and thus provides full
> > > > transitivity between scheduling actions. And since we cannot migrate
> > > > current, a task needs a switch-out and a switch-in in order to
> > > > migrate, in which case the RCsc provides all the ordering we need.  
> > > 
> > > Hi Peter,
> > > 
> > > I'm actually just working on removing this right now too, so
> > > good timing.
> > > 
> > > I think we can't "just" remove it, because it is required to order
> > > MMIO on powerpc as well.  
> > 
> > How is MMIO special? That is, there is only MMIO before we call into
> > schedule() right? So the rq->lock should be sufficient to order that
> > too.
> 
> MMIO uses different barriers. spinlock and smp_ type barriers do
> not order it.

Right, but you only have SYNC, which is what makes it possible at all.

Some of the other architectures are not so lucky and need a different
barrier, ARM for instance needs DSB(ISH) vs the DMB(ISH) provided by
smp_mb(). IA64, MIPS and a few others are in the same boat as ARM.

> > > But what I have done is to comment that some other primitives are
> > > already providing the hwsync for other, so we don't have to add
> > > another one in _switch.  
> > 
> > Right, so this patch relies on the smp_mb__before_spinlock ->
> > smp_mb__after_spinlock conversion that makes the rq->lock RCsc and
> > should thus provide the required SYNC for migrations.
> 
> AFAIKS either one will do, so long as there is a hwsync there. The
> point is just that I have added some commentary in the generic and
> powerpc parts to make it clear we're relying on that behavior of
> the primitive. smp_mb* is not guaranteed to order MMIO, it's just
> that it does on powerpc.

I'm not particularly happy with the generic comment; I don't feel we
should care that PPC is special here.

> > That said, I think you can already use the smp_mb__before_spinlock() as
> > that is done with IRQs disabled, but its a more difficult argument. The
> > rq->lock RCsc property should be more obvious.
> 
> This is what I got.
> 
> https://patchwork.ozlabs.org/patch/770154/

Your comment isn't fully correct, smp_cond_load_acquire() isn't
necessarily done by CPUy. It might be easiest to simply refer to the
"Notes on Program-Order guarantees on SMP systems." comment.

> But I'm not sure if I followed I'm not sure why it's a more
> difficult argument: any time a process moves it must first execute
> a hwsync on the current CPU after it has performed all its access
> there, and then it must execute hwsync on the new CPU before it
> performs any new access.

Yeah, its not a terribly difficult argument either way, but I feel the
RSsc rq->lock on is slightly easier.