Date:   Mon, 29 Nov 2021 09:50:23 +0100
From:   Peter Zijlstra <peterz@infradead.org>
To:     Linus Torvalds <torvalds@linux-foundation.org>
Cc:     Thomas Gleixner <tglx@linutronix.de>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        the arch/x86 maintainers <x86@kernel.org>
Subject: Re: [GIT pull] locking/urgent for v5.16-rc3
Message-ID: <YaSUT2roW6SH2OxM@hirez.programming.kicks-ass.net>
References: <163811728418.767205.14544746031342483043.tglx@xen13>
 <CAHk-=wh1AZLCtx6Uk1JKpknaHgoLi_DM7221LoRpU6Y_3im03A@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <CAHk-=wh1AZLCtx6Uk1JKpknaHgoLi_DM7221LoRpU6Y_3im03A@mail.gmail.com>
Precedence: bulk

On Sun, Nov 28, 2021 at 09:15:10AM -0800, Linus Torvalds wrote:
> On Sun, Nov 28, 2021 at 8:35 AM Thomas Gleixner <tglx@linutronix.de> wrote:
> >
> >  - down_read_trylock() is suboptimal when the lock is contended and
> >    multiple readers trylock concurrently. That's due to the initial value
> >    being read non-atomically which results in at least two compare exchange
> >    loops. Making the initial readout atomic reduces this significantly.
> >    Whith 40 readers by 11% in a benchmark which enforces contention on
> >    mmap_sem.
> 
> This was an intentional optimization to avoid unnecessary cache
> protocol cycles for when the lock isn't contended - first getting a
> cacheline for read ownership, only to then get it for write.
> 
> But I guess we don't have any good benchmarks for non-contention, so ...
> 
> I also hope that maybe modern hardware is smart enough to see "I will
> write to it later" and avoid the "get line for shared access only to
> get it for exclusive access immediately afterwards" issue.

Yes, I raised that same point, otoh those numbers are not showing that.
They did lightly contended, but I suppose not cache-cold.