Loadavg accounting error on arm64

[Mel Gorman]

On Mon, Nov 16, 2020 at 03:20:05PM +0100, Peter Zijlstra wrote:
> > I think this is at least one possibility. I think at least that one
> > should only be explicitly set on WF_MIGRATED and explicitly cleared in
> > sched_ttwu_pending. While I haven't audited it fully, it might be enough
> > to avoid a double write outside of the rq lock on the bitfield but I
> > still need to think more about the ordering of sched_contributes_to_load
> > and whether it's ordered by p->on_cpu or not.
> 
> The scenario you're worried about is something like:
> 
> 	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);
> 

Yes.

> And then the stores of X and Y clobber one another.. Hummph, seems
> reasonable. One quick thing to test would be something like this:
> 
> 
> diff --git a/include/linux/sched.h b/include/linux/sched.h
> index 7abbdd7f3884..9844e541c94c 100644
> --- a/include/linux/sched.h
> +++ b/include/linux/sched.h
> @@ -775,7 +775,9 @@ struct task_struct {
>  	unsigned			sched_reset_on_fork:1;
>  	unsigned			sched_contributes_to_load:1;
>  	unsigned			sched_migrated:1;
> +	unsigned			:0;
>  	unsigned			sched_remote_wakeup:1;
> +	unsigned			:0;
>  #ifdef CONFIG_PSI
>  	unsigned			sched_psi_wake_requeue:1;
>  #endif

And this works.

986.01 1008.17 1013.15 2/855 1212
362.19 824.70 949.75 1/856 1564
133.19 674.65 890.32 1/864 1958
49.04 551.89 834.61 2/871 2339
18.33 451.54 782.41 1/867 2686
6.77 369.37 733.45 1/866 2929
2.55 302.16 687.55 1/864 2931
0.97 247.18 644.52 1/860 2933
0.48 202.23 604.20 1/849 2935

I should have gone with this after rereading the warning about bit fields
having to be protected by the same lock in the "anti-guarantees" section
of memory-barriers.txt :(

sched_psi_wake_requeue can probably stay with the other three fields
given they are under the rq lock but sched_remote_wakeup needs to move
out.

-- 
Mel Gorman
SUSE Labs