Re: [PATCH 1/6] sched: Fix ttwu_queue_cond()

[Paul E. McKenney]

On Mon, Jun 15, 2020 at 03:34:09PM +0200, Peter Zijlstra wrote:
> On Mon, Jun 15, 2020 at 02:56:55PM +0200, Peter Zijlstra wrote:
> > Where the condition:
> > 
> >   !cpus_share_cache(smp_processor_id(), cpu)
> > 
> > already implies 'cpu != smp_processor_id()', because a CPU always
> > shares cache with itself, the secondary condition added in commit:
> > 
> >   2ebb17717550 ("sched/core: Offload wakee task activation if it the wakee is descheduling")
> > 
> > voids that implication, resulting in attempting to do local wake-ups
> > through the queue mechanism.
> > 
> > Fixes: 2ebb17717550 ("sched/core: Offload wakee task activation if it the wakee is descheduling")
> > Reported-by: Paul E. McKenney <paulmck@xxxxxxxxxx>
> > Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
> > Tested-by: Paul E. McKenney <paulmck@xxxxxxxxxx>
> > ---
> >  kernel/sched/core.c |   13 ++++++++++++-
> >  1 file changed, 12 insertions(+), 1 deletion(-)
> > 
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -2356,11 +2356,22 @@ bool cpus_share_cache(int this_cpu, int
> >  
> >  static inline bool ttwu_queue_cond(int cpu, int wake_flags)
> >  {
> > +	int this_cpu = smp_processor_id();
> > +
> > +	/*
> > +	 * Only ever queue for remote wakeups. The on_cpu case can only ever
> > +	 * happen remotely, and for the normal case it makes no sense to
> 
> The 'funny' thing here is, that this must be false for this patch to
> make any difference.. I just cannot see how.
> 
> Also, if this is false, and p->on_cpu == 1 and p->cpu == this_cpu, then
> p _should_ be current, in which case we should never get here either,
> due to the 'p == current' special case in try_to_wake_up().
> 
> The only other option is that 'p == next', but then we'd be doing
> wakeups from the middle of __schedule() and seems 'unlikely' too, esp.
> so since none of the actual stack-traces we have shows that.
> 
> So colour me terribly confused.

I am rerunning with your patch 2 on the last bisection point that
resulted in scheduler NULL dereferences despite having your patch.
Hopefully some illumination will result...

						Thanx, Paul

> > +	 * involve IPIs here, and would be broken, as many architectures cannot
> > +	 * trivially IPI self in any case.
> > +	 */
> > +	if (cpu == this_cpu)
> > +		return false;