sched: Fix data-race in wakeup

sched: Fix data-race in wakeup

Mel reported that on some ARM64 platforms loadavg goes bananas and
Will tracked it down to the following race:

  CPU0					CPU1

  schedule()
    prev->sched_contributes_to_load = X;
    deactivate_task(prev);

					try_to_wake_up()
					  if (p->on_rq &&) // false
					  if (smp_load_acquire(&p->on_cpu) && // true
					      ttwu_queue_wakelist())
					        p->sched_remote_wakeup = Y;

    smp_store_release(prev->on_cpu, 0);

where both p->sched_contributes_to_load and p->sched_remote_wakeup are
in the same word, and thus the stores X and Y race (and can clobber
one another's data).

Whereas prior to commit c6e7bd7afaeb ("sched/core: Optimize ttwu()
spinning on p->on_cpu") the p->on_cpu handoff serialized access to
p->sched_remote_wakeup (just as it still does with
p->sched_contributes_to_load) that commit broke that by calling
ttwu_queue_wakelist() with p->on_cpu != 0.

However, due to

  p->XXX = X			ttwu()
  schedule()			  if (p->on_rq && ...) // false
    smp_mb__after_spinlock()	  if (smp_load_acquire(&p->on_cpu) &&
    deactivate_task()		      ttwu_queue_wakelist())
      p->on_rq = 0;		        p->sched_remote_wakeup = Y;

We can be sure any 'current' store is complete and 'current' is
guaranteed asleep. Therefore we can move p->sched_remote_wakeup into
the current flags word.

Note: while the observed failure was loadavg accounting gone wrong due
to ttwu() cobbering p->sched_contributes_to_load, the reverse problem
is also possible where schedule() clobbers p->sched_remote_wakeup,
this could result in enqueue_entity() wrecking ->vruntime and causing
scheduling artifacts.

Fixes: c6e7bd7afaeb ("sched/core: Optimize ttwu() spinning on p->on_cpu")
Reported-by: Mel Gorman <mgorman@techsingularity.net>
Debugged-by: Will Deacon <will@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20201117083016.GK3121392@hirez.programming.kicks-ass.net