Re: [PATCH v4 3/4] swap: apply new pw_queue_on() interface

[Leonardo Bras]

On Wed, May 20, 2026 at 05:07:36PM +0200, Sebastian Andrzej Siewior wrote:
> On 2026-05-18 22:27:49 [-0300], Leonardo Bras wrote:
> 
> after digesting the slub patch,
> 
> > @@ -882,38 +879,38 @@ static inline void __lru_add_drain_all(bool force_all_cpus)
> >  	 * If the paired barrier is done at any later step, e.g. after the
> >  	 * loop, CPU #x will just exit at (C) and miss flushing out all of its
> >  	 * added pages.
> >  	 */
> >  	WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
> >  	smp_mb();
> >  
> >  	cpumask_clear(&has_mm_work);
> >  	cpumask_clear(&has_bh_work);
> >  	for_each_online_cpu(cpu) {
> > -		struct work_struct *mm_work = &per_cpu(lru_add_drain_work, cpu);
> > +		struct pw_struct *mm_pw = &per_cpu(lru_add_drain_pw, cpu);
> >  		struct work_struct *bh_work = &per_cpu(bh_add_drain_work, cpu);
> >  
> >  		if (cpu_needs_mm_drain(cpu)) {
> > -			INIT_WORK(mm_work, lru_add_drain_per_cpu);
> > -			queue_work_on(cpu, mm_percpu_wq, mm_work);
> > +			INIT_PW(mm_pw, lru_add_drain_per_cpu, cpu);
> > +			pw_queue_on(cpu, mm_percpu_wq, mm_pw);
> >  			__cpumask_set_cpu(cpu, &has_mm_work);
> >  		}
> >  
> >  		if (cpu_needs_bh_drain(cpu)) {
> >  			INIT_WORK(bh_work, bh_add_drain_per_cpu);
> >  			queue_work_on(cpu, mm_percpu_wq, bh_work);
> >  			__cpumask_set_cpu(cpu, &has_bh_work);
> >  		}
> >  	}
> >  
> >  	for_each_cpu(cpu, &has_mm_work)
> > -		flush_work(&per_cpu(lru_add_drain_work, cpu));
> > +		pw_flush(&per_cpu(lru_add_drain_pw, cpu));
> >  
> >  	for_each_cpu(cpu, &has_bh_work)
> >  		flush_work(&per_cpu(bh_add_drain_work, cpu));
> 
> Why do we have two iterations here? Is it just a proof of concept that
> is not complete yet? I am curious why it is okay/needed to "remove" the
> one workqueue but not the other. Maybe the other does not bother as much
> as the other does.

Argh, sorry about the above :/

Slub was converted on a previous patchset version, and rebasing worked 
fine. I should have properly checked if any other mechanism was added in 
the file, but ended up just checking the one that was already converted.

> 
> But essentially we can't use a spin_lock_t here because due to the
> hotpath nature of the code it will kill performance. 

Correct

> So instead we do it
> anyway but behind a switch so that only those suffer from this that do
> not want to suffer from workqueue interruption on a NOHZ full system,
> right?

Yes, one can choose to pay the price to use spinlocks here and have less 
workqueue interruptions, if !CONFIG_PREEMPT_RT

In case of CONFIG_PREEMPT_RT, we have the local_lock() already becoming a 
rt_spinlock(), so it already pays the price to hold the lock. So it should 
be basically free on this scenario, and save time by doing the operation 
remotelly instead of scheduling it.
 
> 
> I thought that this improved since commit
>   ff042f4a9b050 ("mm: lru_cache_disable: replace work queue synchronization with synchronize_rcu")
> 
> Did it get worse or was it not entirely gone?
> 

I worked in this patchset majorly after that commit date, and it was still 
an issue up to last time Marcelo tested. Not sure of the impact of above 
commit, but I suppose it may have brought some improvements without fully 
fixing it.

Thanks!
Leo