Re: [PATCH] sched: Fix nr_uninterruptible race causing increasing load average

From: Peter Zijlstra
Date: Fri Jul 09 2021 - 08:57:48 EST


On Thu, Jul 08, 2021 at 10:54:58AM -0400, Phil Auld wrote:
> Sorry... I don't have a nice diagram. I'm still looking at what all those
> macros actually mean on the various architectures.

Don't worry about other architectures, lets focus on Power, because
that's the case where you can reprouce funnies. Now Power only has 2
barrier ops (not quite true, but close enough for all this):

- SYNC is the full barrier

- LWSYNC is a TSO like barrier

Pretty much everything (LOAD-ACQUIRE, STORE-RELEASE, WMB, RMB) uses
LWSYNC. Only MB result in SYNC.

Power is 'funny' because their spinlocks are weaker than everybody
else's, but AFAICT that doesn't seem relevant here.

> Using what you have above I get the same thing. It looks like it should be
> ordered but in practice it's not, and ordering it "more" as I did in the
> patch, fixes it.

And you're running Linus' tree, not some franken-kernel from RHT, right?
As asked in that other email, can you try with just the WMB added? I
really don't believe that RMB you added can make a difference.

Also, can you try with TTWU_QUEUE disabled (without any additional
barriers added), that simplifies the wakeup path a lot.

> Is it possible that the bit field is causing some of the assumptions about
> ordering in those various macros to be off?

*should* not matter...

prev->sched_contributes_to_load = X;

smp_store_release(&prev->on_cpu, 0);
asm("LWSYNC" : : : "memory");
WRITE_ONCE(prev->on_cpu, 0);

due to that memory clobber, the compiler must emit whatever stores are
required for the bitfield prior to the LWSYNC.

> I notice in all the comments about smp_mb__after_spinlock etc, it's always
> WRITE_ONCE/READ_ONCE on the variables in question but we can't do that with
> the bit field.

Yeah, but both ->on_rq and ->sched_contributes_to_load are 'normal'
stores. That said, given that ttwu() does a READ_ONCE() on ->on_rq, we
should match that with WRITE_ONCE()...

So I think we should do the below, but I don't believe it'll make a
difference. Let me stare more.

---
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index ca9a523c9a6c..da93551b298d 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -1973,12 +1973,12 @@ void activate_task(struct rq *rq, struct task_struct *p, int flags)
{
enqueue_task(rq, p, flags);

- p->on_rq = TASK_ON_RQ_QUEUED;
+ WRITE_ONCE(p->on_rq, TASK_ON_RQ_QUEUED);
}

void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
{
- p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
+ WRITE_ONCE(p->on_rq, (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING);

dequeue_task(rq, p, flags);
}
@@ -5662,11 +5662,11 @@ static bool try_steal_cookie(int this, int that)
if (p->core_occupation > dst->idle->core_occupation)
goto next;

- p->on_rq = TASK_ON_RQ_MIGRATING;
+ WRITE_ONCE(p->on_rq, TASK_ON_RQ_MIGRATING);
deactivate_task(src, p, 0);
set_task_cpu(p, this);
activate_task(dst, p, 0);
- p->on_rq = TASK_ON_RQ_QUEUED;
+ WRITE_ONCE(p->on_rq, TASK_ON_RQ_QUEUED);

resched_curr(dst);