Re: [PATCH] doc: Update wake_up() & co. memory-barrier guarantees

[Andrea Parri <andrea.parri@amarulasolutions.com>]

> > For example, the second comment says:
> > 
> >   /*
> >    * The below implies an smp_mb(), it too pairs with the smp_wmb() from
> >    * woken_wake_function() such that we must either observe the wait
> >    * condition being true _OR_ WQ_FLAG_WOKEN such that we will not miss
> >    * an event.
> >    */
> > 
> > From this I understand:
> > 
> >    wq_entry->flags &= ~WQ_FLAG_WOKEN;      condition = true;
> >    smp_mb() // B                           smp_wmb(); // C
> >    [next iteration of the loop]            wq_entry->flags |= WQ_FLAG_WOKEN;
> >    if (condition)
> >       break;
> > 
> >    BUG_ON(!condition && !(wq_entry->flags & WQ_FLAG_WOKEN))
> 
> Right, basically if we get a spurious wakeup and our ttwu() 'fails', we
> must either see condition on the next iteration, or ensure the next
> iteration doesn't sleep, so we'll loop around and test condition yet
> again.
> 
> > IOW, this is an R-like pattern: if this is the case, the smp_wmb() does
> > _not_ prevent the BUG_ON() from firing; according to LKMM (and powerpc)
> > a full barrier would be needed.
> 
> Hmmm, how come? store-store collision? Yes I suppose you're right.

Ehh, the corresponding powerpc test is architecturally allowed; in the
operational model, the BUG_ON() state can be reached by following the
following steps:

 1. let the writes all reach the storage subsystem,

 2. commit the partial coherence order from "->flags |= WQ_FLAG_WOKEN"
    to "->flags &= ~WQ_FLAG_WOKEN"

 3. propagate "->flags &= ~WQ_FLAG_WOKEN" to the other thread

 4. commit and acknowledge the sync (B)

 5. satisfy the read

 6. propagate "condition = true" and the lwsync (C) to the other thread.

AFAICT, this state remains _unobserved_ via litmus7 experiments.


> 
> > Same RFC for the first comment:
> > 
> >   /*
> >    * The above implies an smp_mb(), which matches with the smp_wmb() from
> >    * woken_wake_function() such that if we observe WQ_FLAG_WOKEN we must
> >    * also observe all state before the wakeup.
> >    */
> > 
> > What is the corresponding snippet & BUG_ON()?
> 
> The comment there suggest:
> 
> 	if (condition)
> 		break;
> 
> 	set_current_state(UNINTERRUPTIBLE);		condition = true;
> 	/* smp_mb() */					smp_wmb();
> 							wq_entry->flags |= WQ_FLAG_WOKEN;
> 	if (!wq_entry->flags & WQ_FLAG_WOKEN)
> 		schedule();
> 
> 
> 	BUG_ON((wq_entry->flags & WQ_FLAG_WOKEN) && !condition);
> 
> 
> But looking at that now, I think that's wrong. Because the the purpose
> was that, if we don't do the try_to_wake_up(), our task still needs to
> observe the condition store.
> 
> But for that we need a barrier between the wq_entry->flags load and the
> second condition test, which would (again) be B, not A.

Agreed.  Now that I stared at the code a bit more, I think that (A) is
still needed for the synchronization on "->state" and "->flags" (an SB
pattern seems again to be hidden in the call to try_to_wake_up()):

  p->state = mode;                           wq_entry->flags |= WQ_FLAG_WOKEN;
  smp_mb(); // A                             try_to_wake_up():
  if (!(wq_entry->flags & WQ_FLAG_WOKEN))      <full barrier>
     schedule()                                if (!(p->state & mode))
                                                 goto out;

  BUG_ON(!(wq_entry->flags & WQ_FLAG_WOKEN) && !(p->state & mode))

So, I think that we should keep (A).

I am planning to send these changes (smp_mb() in woken_wake_function()
and fixes to the comments) as a separate patch.

Thanks,
  Andrea
--
To unsubscribe from this list: send the line "unsubscribe linux-doc" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html