Re: [PATCH] arm64: init: override deferred_page_init_max_threads

[Eric Chanudet]

On Wed, May 22, 2024 at 11:41:07PM +1000, Michael Ellerman wrote:
> Eric Chanudet <echanude@xxxxxxxxxx> writes:
> > On Tue, May 21, 2024 at 07:10:07PM +0300, Mike Rapoport wrote:
> >> (added powerpc folks)
> 
> Thanks Mike.
> 
> >> On Mon, May 20, 2024 at 07:15:59PM -0400, Eric Chanudet wrote:
> >> > This was the behavior prior to making the function arch-specific with
> >> > commit ecd096506922 ("mm: make deferred init's max threads
> >> > arch-specific")
> >> > 
> >> > Architectures can override the generic implementation that uses only one
> >> > CPU. Setting DEFERRED_STRUCT_PAGE_INIT and testing on a few arm64
> >> > platforms shows faster deferred_init_memmap completions:
> >> > 
> >> > |         | x13s        | SA8775p-ride | Ampere R137-P31 | Ampere HR330 |
> >> > |         | Metal, 32GB | VM, 36GB     | VM, 58GB        | Metal, 128GB |
> >> > |         | 8cpus       | 8cpus        | 8cpus           | 32cpus       |
> >> > |---------|-------------|--------------|-----------------|--------------|
> >> > | threads |  ms     (%) | ms       (%) |  ms         (%) |  ms      (%) |
> >> > |---------|-------------|--------------|-----------------|--------------|
> >> > | 1       | 108    (0%) | 72      (0%) | 224        (0%) | 324     (0%) |
> >> > | cpus    |  24  (-77%) | 36    (-50%) |  40      (-82%) |  56   (-82%) |
> 
> How did you measure this, just some printks in page_alloc_init_late() or
> something more sophisticated? Just so I can do some comparable measurements.

I used the existing pr_info in deferred_init_memmap().

> >> > diff --git a/arch/arm64/mm/init.c b/arch/arm64/mm/init.c
> >> > index 9b5ab6818f7f..71f5188fe63d 100644
> >> > --- a/arch/arm64/mm/init.c
> >> > +++ b/arch/arm64/mm/init.c
> >> > @@ -158,6 +158,13 @@ static void __init zone_sizes_init(void)
> >> >  	free_area_init(max_zone_pfns);
> >> >  }
> >> >  
> >> > +#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
> >> > +int __init deferred_page_init_max_threads(const struct cpumask *node_cpumask)
> >> > +{
> >> > +	return max_t(int, cpumask_weight(node_cpumask), 1);
> >> > +}
> >> > +#endif
> >> > +
> >> 
> >> Maybe we should make this default and let architectures that want a single
> >> thread override deferred_page_init_max_threads() to return 1?
> >
> > It would affect more archs than I can try this on. Currently, only x86
> > (with this change, arm64) return more than one thread.
> 
> I can test powerpc and we can find someone to test s390. No other
> arches have it enabled in their defconfig.

Many thanks!

> > I'm happy to send a v2 inverting the logic if you find it preferable.
> 
> That seems preferable. It's a scalability feature, it makes no sense for
> the default to be a single thread AFAICS.

Understood, I'll respin.

-- 
Eric Chanudet