Re: [PATCH] mm/mm_init: Don't call cond_resched() in deferred_init_memmap_chunk() if rcu_preempt_depth() set

["Paul E. McKenney" <paulmck@kernel.org>]

On Thu, Jan 22, 2026 at 08:57:47AM +0100, Sebastian Andrzej Siewior wrote:
> On 2026-01-21 13:27:32 [-0800], Paul E. McKenney wrote:
> > > > > --- a/mm/mm_init.c
> > > > > +++ b/mm/mm_init.c
> > > > > @@ -2085,7 +2085,12 @@ deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
> > > > >   			spfn = chunk_end;
> > > > > -			if (irqs_disabled())
> > > > > +			/*
> > > > > +			 * pgdat_resize_lock() only disables irqs in non-RT
> > > > > +			 * kernels but calls rcu_read_lock() in a PREEMPT_RT
> > > > > +			 * kernel.
> > > > > +			 */
> > > > > +			if (irqs_disabled() || rcu_preempt_depth())
> > > > >   				touch_nmi_watchdog();
> > > > rcu_preempt_depth() seems a fairly internal low-level thing - it's
> > > > rarely used.
> If you acquire a lock from time to time and you pass a bool the let the
> function below know whether scheduling is fine or not then it is
> obvious. If you choose to check for symptoms of an acquired lock then
> you have to use also the rarely used functions ;)

Agreed.

> > > That is true. Beside the scheduler, workqueue also use rcu_preempt_depth().
> > > This API is included in "include/linux/rcupdate.h" which is included
> > > directly or indirectly by many kernel files. So even though it is rarely
> > > used, but it is still a public API.
> > 
> > It is a bit tricky, for example, given a kernel built with both
> > CONFIG_PREEMPT_NONE=y and CONFIG_PREEMPT_DYNAMIC=y, it will never
> > invoke touch_nmi_watchdog(), even if it really is in an RCU read-side
> > critical section.  This is because it was intended for lockdep-like use,
> > where (for example) you don't want to complain about sleeping in an RCU
> > read-side critical section unless you are 100% sure that you are in fact
> > in an RCU read-side critical section.
> > 
> > Maybe something like this?
> > 
> > 	if (irqs_disabled() || !IS_ENABLED(CONFIG_PREEMPT_RCU) || rcu_preempt_depth())
> > 		touch_nmi_watchdog();
> 
> I don't understand the PREEMPT_NONE+DYNAMIC reasoning. irqs_disabled()
> should not be affected by this and rcu_preempt_depth() will be 0 for
> !CONFIG_PREEMPT_RCU so I don't think this is required. 

If touch_nmi_watchdog() is fast enough that calling it in a potentially
tight loop is OK, then agreed, this is not required.

But maybe I should redefine !PREEMPT rcu_preempt_depth() in terms of
preemptible() or some such.  Thoughts?

> > This would *always* invoke touch_nmi_watchdog() for such kernels, which
> > might or might not be OK.
> > 
> > I freely confesss that I am not sure which of these is appropriate in
> > this setting.
> 
> What about a more straight forward and obvious approach?
> 
> diff --git a/mm/mm_init.c b/mm/mm_init.c
> index fc2a6f1e518f1..0b283fd48b282 100644
> --- a/mm/mm_init.c
> +++ b/mm/mm_init.c
> @@ -2059,7 +2059,7 @@ static unsigned long __init deferred_init_pages(struct zone *zone,
>   */
>  static unsigned long __init
>  deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
> -			   struct zone *zone)
> +			   struct zone *zone, bool may_schedule)
>  {
>  	int nid = zone_to_nid(zone);
>  	unsigned long nr_pages = 0;
> @@ -2085,10 +2085,10 @@ deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
>  
>  			spfn = chunk_end;
>  
> -			if (irqs_disabled())
> -				touch_nmi_watchdog();
> -			else
> +			if (may_schedule)
>  				cond_resched();
> +			else
> +				touch_nmi_watchdog();
>  		}
>  	}
>  
> @@ -2101,7 +2101,7 @@ deferred_init_memmap_job(unsigned long start_pfn, unsigned long end_pfn,
>  {
>  	struct zone *zone = arg;
>  
> -	deferred_init_memmap_chunk(start_pfn, end_pfn, zone);
> +	deferred_init_memmap_chunk(start_pfn, end_pfn, zone, true);
>  }
>  
>  static unsigned int __init
> @@ -2216,7 +2216,7 @@ bool __init deferred_grow_zone(struct zone *zone, unsigned int order)
>  	for (spfn = first_deferred_pfn, epfn = SECTION_ALIGN_UP(spfn + 1);
>  	     nr_pages < nr_pages_needed && spfn < zone_end_pfn(zone);
>  	     spfn = epfn, epfn += PAGES_PER_SECTION) {
> -		nr_pages += deferred_init_memmap_chunk(spfn, epfn, zone);
> +		nr_pages += deferred_init_memmap_chunk(spfn, epfn, zone, false);
>  	}
>  
>  	/*
> 
> Wouldn't this work?

I will defer to the mm guys on this question.

							Thanx, Paul