[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <4ivthmwoi6xjdu6dxl7wir7xcy34un7cwll5crt7rf5666u4ci@rjliv4emytdr>
Date: Wed, 22 May 2024 09:54:23 -0400
From: Eric Chanudet <echanude@...hat.com>
To: Michael Ellerman <mpe@...erman.id.au>
Cc: Mike Rapoport <rppt@...nel.org>,
Catalin Marinas <catalin.marinas@....com>, Will Deacon <will@...nel.org>, Baoquan He <bhe@...hat.com>,
Andrew Morton <akpm@...ux-foundation.org>, Zhen Lei <thunder.leizhen@...wei.com>,
Yajun Deng <yajun.deng@...ux.dev>, Zhang Jianhua <chris.zjh@...wei.com>,
linux-arm-kernel@...ts.infradead.org, linux-kernel@...r.kernel.org, Nick Piggin <npiggin@...il.com>
Subject: Re: [PATCH] arm64: init: override deferred_page_init_max_threads
On Wed, May 22, 2024 at 11:41:07PM +1000, Michael Ellerman wrote:
> Eric Chanudet <echanude@...hat.com> 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
Powered by blists - more mailing lists