Received-SPF: pass (google.com: domain of linux-kernel+bounces-11129-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:4601:e00::3 as permitted sender) client-ip=2604:1380:4601:e00::3;
Date: Mon, 25 Dec 2023 09:38:15 -0800
From: Yury Norov <yury.norov@gmail.com>
To: Ming Lei <ming.lei@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>,
	Thomas Gleixner <tglx@linutronix.de>, linux-kernel@vger.kernel.org,
	Andy Shevchenko <andriy.shevchenko@linux.intel.com>,
	Rasmus Villemoes <linux@rasmusvillemoes.dk>
Subject: Re: [PATCH v3 2/7] lib/group_cpus: optimize inner loop in
 grp_spread_init_one()
Message-ID: <ZYm+B2m9gOTLOedE@yury-ThinkPad>
References: <20231212042108.682072-1-yury.norov@gmail.com>
 <20231212042108.682072-3-yury.norov@gmail.com>
 <ZXgsDcM21H/2BTck@fedora>
 <ZXiSk9J2ZBdsTnk0@yury-ThinkPad>
 <ZXj1mGAKz5cWy+Nr@fedora>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <ZXj1mGAKz5cWy+Nr@fedora>

On Wed, Dec 13, 2023 at 08:06:48AM +0800, Ming Lei wrote:
> On Tue, Dec 12, 2023 at 09:04:19AM -0800, Yury Norov wrote:
> > On Tue, Dec 12, 2023 at 05:46:53PM +0800, Ming Lei wrote:
> > > On Mon, Dec 11, 2023 at 08:21:02PM -0800, Yury Norov wrote:
> > > > The loop starts from the beginning every time we switch to the next
> > > > sibling mask. This is the Schlemiel the Painter's style of coding
> > > > because we know for sure that nmsk is clear up to current CPU, and we
> > > > can just continue from the next CPU.
> > > > 
> > > > Also, we can do it nicer if leverage the dedicated for_each() iterator,
> > > > and simplify the logic of clearing a bit in nmsk.
> > > > 
> > > > Signed-off-by: Yury Norov <yury.norov@gmail.com>
> > > > ---
> > > >  lib/group_cpus.c | 13 ++++++-------
> > > >  1 file changed, 6 insertions(+), 7 deletions(-)
> > > > 
> > > > diff --git a/lib/group_cpus.c b/lib/group_cpus.c
> > > > index ee272c4cefcc..10dead3ab0e0 100644
> > > > --- a/lib/group_cpus.c
> > > > +++ b/lib/group_cpus.c
> > > > @@ -30,14 +30,13 @@ static void grp_spread_init_one(struct cpumask *irqmsk, struct cpumask *nmsk,
> > > >  
> > > >  		/* If the cpu has siblings, use them first */
> > > >  		siblmsk = topology_sibling_cpumask(cpu);
> > > > -		for (sibl = -1; cpus_per_grp > 0; ) {
> > > > -			sibl = cpumask_next(sibl, siblmsk);
> > > > -			if (sibl >= nr_cpu_ids)
> > > > -				break;
> > > > -			if (!cpumask_test_and_clear_cpu(sibl, nmsk))
> > > > -				continue;
> > > > +		sibl = cpu + 1;
> > > 
> > > It doesn't have to 'cpu + 1', see below comment.
> > > 
> > > > +
> > > > +		for_each_cpu_and_from(sibl, siblmsk, nmsk) {
> > > > +			cpumask_clear_cpu(sibl, nmsk);
> > > >  			cpumask_set_cpu(sibl, irqmsk);
> > > > -			cpus_per_grp--;
> > > > +			if (cpus_per_grp-- == 0)
> > > 
> > > 			if (--cpus_per_grp == 0)
> >  
> > That's right, I'll send a new version this weekend.
> > 
> > > > +				return;
> > > 
> > > I think for_each_cpu_and() should work just fine, cause cpu has been cleared
> > > from nmsk, so the change can be something like, then patch 1 isn't
> > > necessary.
> >  
> > It works just fine except that it's O(N^2), where O(N) is easily
> > achievable. Again, it's not about performance, it's about coding
> > habits.
> 
> Both for_each_cpu_and() and for_each_cpu_and_from() are O(N), aren't
> they? Given both two are based on find_next_and_bit().

for_each_cpu_and() is the same Schlemiel the Painter's code, as the
plain for() was.
 
> for_each_cpu_and() is simpler and more readable, and more
> importantly, we can save one single-user public helper.
> 
> >  
> > > 		for_each_cpu_and(sibl, siblmsk, nmsk) {
> > > 			cpumask_clear_cpu(sibl, nmsk);
> > >   			cpumask_set_cpu(sibl, irqmsk);
> > > 			if (--cpus_per_grp == 0)
> > > 				return;
> > > 		}
> 
> 
> Thanks,
> Ming