Re: [PATCH 11/15] stackdepot: use read/write lock

From: Marco Elver <elver@google.com>
To: Andrey Konovalov <andreyknvl@gmail.com>
Cc: andrey.konovalov@linux.dev,
	Alexander Potapenko <glider@google.com>,
	 Dmitry Vyukov <dvyukov@google.com>,
	Vlastimil Babka <vbabka@suse.cz>,
	kasan-dev@googlegroups.com,
	 Evgenii Stepanov <eugenis@google.com>,
	Andrew Morton <akpm@linux-foundation.org>,
	linux-mm@kvack.org,  linux-kernel@vger.kernel.org,
	Andrey Konovalov <andreyknvl@google.com>
Subject: Re: [PATCH 11/15] stackdepot: use read/write lock
Date: Tue, 2 Jan 2024 13:59:06 +0100	[thread overview]
Message-ID: <CANpmjNNfyKV0Ky=GRiw9_6va3nJMtYejWZJL0tn5cjwXTY8e1Q@mail.gmail.com> (raw)
In-Reply-To: <CA+fCnZdAUo1CKDK4kiUyR+Fxc_F++CFezanPDVujx3u7fBmw=A@mail.gmail.com>

On Wed, 13 Sept 2023 at 19:09, Andrey Konovalov <andreyknvl@gmail.com> wrote:
>
> On Tue, Sep 5, 2023 at 6:19 PM Marco Elver <elver@google.com> wrote:
> >
> > > Good suggestion. I propose that we keep the rwlock for now, and I'll
> > > check whether the performance is better with percpu-rwsem once I get
> > > to implementing and testing the performance changes. I'll also check
> > > whether percpu-rwsem makes sense for stack ring in tag-based KASAN
> > > modes.
> >
> > I think it's quite obvious that the percpu-rwsem is better. A simple
> > experiment is to measure the ratio of stackdepot hits vs misses. If
> > the ratio is obviously skewed towards hits, then I'd just go with the
> > percpu-rwsem.
> >
> > The performance benefit may not be measurable if you use a small system.
>
> I started looking into using percpu-rwsem, but it appears that it
> doesn't have the irqsave/irqrestore API flavor. I suspect that it
> shouldn't be hard to add it, but I'd rather not pursue this as a part
> of this series.
>
> So I still propose to keep the rwlock for now, and switch to
> percpu-rwsem later together with the other perf changes.

I may have gotten lost in the post-vacation email avalanche and missed
it: did you already send the percpu-rwsem optimization? I am a little
worried about the contention the plain rwlock introduces on big
machines.

Thanks,
-- Marco