Re: [PATCH] doc: Update wake_up() & co. memory-barrier guarantees
From: Peter Zijlstra
Date: Mon Jun 25 2018 - 09:01:18 EST
On Mon, Jun 25, 2018 at 02:28:50PM +0200, Andrea Parri wrote:
> On Mon, Jun 25, 2018 at 01:12:45PM +0100, David Howells wrote:
> > Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >
> > > So yes, I suppose we're entirely suck with the full memory barrier
> > > semantics like that. But I still find it easier to think of it like a
> > > RELEASE that pairs with the ACQUIRE of waking up, such that the task
> > > is guaranteed to observe it's own wake condition.
> > >
> > > And maybe that is the thing I'm missing here. These comments only state
> > > that it does in fact imply a full memory barrier, but do not explain
> > > why, should it?
> >
> > I think because RELEASE and ACQUIRE concepts didn't really exist in Linux at
> > the time I wrote the doc, so the choices were read/readdep, write or full.
> >
> > Since this document defines the *minimum* you can expect rather than what the
> > kernel actually gives you, I think it probably makes sense to switch to
> > RELEASE and ACQUIRE here.
>
> RELEASE and ACQUIRE are not enough in SB. Can you elaborate?
I prefer RELEASE vs wait-condition and treat task->state as an internal
matter. Also note how the set_current_task() comment is fairly vague on
what exact barriers are used. It just states 'sufficient'.
Maybe I should just give up and accept smp_mb(), but strictly speaking
that is overkill, but it is the only sufficient barrier we currently
have.