Re: [PATCH 6/6 v2] mm: Drain LRUs upon resume to userspace on nohz_full CPUs

[Frederic Weisbecker <frederic@kernel.org>]

Le Mon, Feb 10, 2025 at 06:50:26PM +0800, Hillf Danton a écrit :
> On Sun,  9 Feb 2025 23:30:04 +0100 Frederic Weisbecker <frederic@kernel.org>
> > @@ -769,6 +772,9 @@ static bool cpu_needs_drain(unsigned int cpu)
> >  {
> >  	struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
> >  
> > +	if (!housekeeping_cpu(cpu, HK_TYPE_KERNEL_NOISE))
> > +		return false;
> > +
> >  	/* Check these in order of likelihood that they're not zero */
> >  	return folio_batch_count(&fbatches->lru_add) ||
> >  		folio_batch_count(&fbatches->lru_move_tail) ||
> > -- 
> > 2.46.0
> 
> Nit, I'd like to add a debug line to test your assumption that
> isolated tasks are pinned to a single nohz_full CPU.
> 
> --- x/mm/swap.c
> +++ y/mm/swap.c
> @@ -767,9 +767,10 @@ static void lru_add_drain_per_cpu(struct
>  static bool cpu_needs_drain(unsigned int cpu)
>  {
>  	struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
> +	bool yes;
>  
>  	/* Check these in order of likelihood that they're not zero */
> -	return folio_batch_count(&fbatches->lru_add) ||
> +	yes = folio_batch_count(&fbatches->lru_add) ||
>  		folio_batch_count(&fbatches->lru_move_tail) ||
>  		folio_batch_count(&fbatches->lru_deactivate_file) ||
>  		folio_batch_count(&fbatches->lru_deactivate) ||
> @@ -777,6 +778,12 @@ static bool cpu_needs_drain(unsigned int
>  		folio_batch_count(&fbatches->lru_activate) ||
>  		need_mlock_drain(cpu) ||
>  		has_bh_in_lru(cpu, NULL);
> +
> +	if (!housekeeping_cpu(cpu, HK_TYPE_KERNEL_NOISE)) {
> +		VM_BUG_ON(yes);
> +		return false;
> +	}
> +	return yes;

If the task isn't pinned then the guarantees of nohz_full are broken anyway.
Also if the task migrates it will simply execute the work elsewhere.

My only worry is kernel threads. Those are simply ignored in this patchset but
this is not right as they can do allocations. Yet they can't execute anything
on return to userspace...

Thoughts?

Thanks.

>  }
>  
>  /*
>