Re: linux-next: build failure after merge of the rcu tree

From: Peter Zijlstra
Date: Fri Aug 11 2017 - 10:45:47 EST


On Fri, Aug 11, 2017 at 07:39:17AM -0700, Paul E. McKenney wrote:
> On Fri, Aug 11, 2017 at 11:14:34AM +0200, Peter Zijlstra wrote:
> > On Thu, Aug 10, 2017 at 09:54:53PM -0700, Paul E. McKenney wrote:
> > > On Fri, Aug 11, 2017 at 02:43:52PM +1000, Stephen Rothwell wrote:
> > >
> > > Looks like I need to rebase my patch on top of a9668cd6ee28, and
> > > than put an smp_mb__after_spinlock() between the lock and the unlock.
> > >
> > > Peter, any objections to that approach? Other suggestions?
> >
> > Hurm.. I'll have to try and understand that comment there again it
> > seems.
>
> My reasoning is as follows:
>
> 1. The critical section is empty, so any prior references
> would be ordered only against later critical sections.
>
> 2. A full barrier within the critical section will order those
> prior references against later critical sections just
> as easily as would one prior to the critical section.
>
> Does that make sense, I should I have stayed away from the keyboard
> at this early hour? ;-)

So I think we can do away with 2 because our prior and later stores have
an address dependency (they are to the same variable) and thus must be
ordered already.

Basically:

CPU0 CPU1

p->state = TASK_UNINTERRUPTIBLE;
try_to_wake_up(p)
p->state = TASK_RUNNING
spin_lock(&p->pi_lock);
spin_unlock(&p->pi_lock);
p->state = TASK_DEAD

Now, the ttwu(p) NO-OPs unless it sees (UN)INTERRUPTIBLE, so either
RUNNING or DEAD are fine. However if it sees (UN)INTERRUPTIBLE it will
do another (competing) RUNNING store which must not overwrite DEAD.