Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Thu, 18 Jun 2020 16:16:55 +0300
From:   Alexey Dobriyan <adobriyan@gmail.com>
To:     David Laight <David.Laight@aculab.com>
Cc:     'Matt Fleming' <matt@codeblueprint.co.uk>,
        Ingo Molnar <mingo@kernel.org>,
        Peter Zijlstra <peterz@infradead.org>,
        Thomas Gleixner <tglx@linutronix.de>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        "Grimm, Jon" <Jon.Grimm@amd.com>,
        "Kumar, Venkataramanan" <Venkataramanan.Kumar@amd.com>,
        Jan Kara <jack@suse.cz>,
        "stable@vger.kernel.org" <stable@vger.kernel.org>
Subject: Re: [PATCH] x86/asm/64: Align start of __clear_user() loop to
 16-bytes
Message-ID: <20200618131655.GA24607@localhost.localdomain>
References: <20200618102002.30034-1-matt@codeblueprint.co.uk>
 <39f8304b75094f87a54ace7732708d30@AcuMS.aculab.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
In-Reply-To: <39f8304b75094f87a54ace7732708d30@AcuMS.aculab.com>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, Jun 18, 2020 at 10:48:05AM +0000, David Laight wrote:
> From: Matt Fleming
> > Sent: 18 June 2020 11:20
> > x86 CPUs can suffer severe performance drops if a tight loop, such as
> > the ones in __clear_user(), straddles a 16-byte instruction fetch
> > window, or worse, a 64-byte cacheline. This issues was discovered in the
> > SUSE kernel with the following commit,
> > 
> >   1153933703d9 ("x86/asm/64: Micro-optimize __clear_user() - Use immediate constants")
> > 
> > which increased the code object size from 10 bytes to 15 bytes and
> > caused the 8-byte copy loop in __clear_user() to be split across a
> > 64-byte cacheline.
> > 
> > Aligning the start of the loop to 16-bytes makes this fit neatly inside
> > a single instruction fetch window again and restores the performance of
> > __clear_user() which is used heavily when reading from /dev/zero.
> > 
> > Here are some numbers from running libmicro's read_z* and pread_z*
> > microbenchmarks which read from /dev/zero:
> > 
> >   Zen 1 (Naples)
> > 
> >   libmicro-file
> >                                         5.7.0-rc6              5.7.0-rc6              5.7.0-rc6
> >                                                     revert-1153933703d9+               align16+
> >   Time mean95-pread_z100k       9.9195 (   0.00%)      5.9856 (  39.66%)      5.9938 (  39.58%)
> >   Time mean95-pread_z10k        1.1378 (   0.00%)      0.7450 (  34.52%)      0.7467 (  34.38%)
> >   Time mean95-pread_z1k         0.2623 (   0.00%)      0.2251 (  14.18%)      0.2252 (  14.15%)
> >   Time mean95-pread_zw100k      9.9974 (   0.00%)      6.0648 (  39.34%)      6.0756 (  39.23%)
> >   Time mean95-read_z100k        9.8940 (   0.00%)      5.9885 (  39.47%)      5.9994 (  39.36%)
> >   Time mean95-read_z10k         1.1394 (   0.00%)      0.7483 (  34.33%)      0.7482 (  34.33%)
> > 
> > Note that this doesn't affect Haswell or Broadwell microarchitectures
> > which seem to avoid the alignment issue by executing the loop straight
> > out of the Loop Stream Detector (verified using perf events).
> 
> Which cpu was affected?
> At least one source (www.agner.org/optimize) implies that both ivy
> bridge and sandy bridge have uop caches that mean (If I've read it
> correctly) the loop shouldn't be affected by the alignment).
> 
> > diff --git a/arch/x86/lib/usercopy_64.c b/arch/x86/lib/usercopy_64.c
> > index fff28c6f73a2..b0dfac3d3df7 100644
> > --- a/arch/x86/lib/usercopy_64.c
> > +++ b/arch/x86/lib/usercopy_64.c
> > @@ -24,6 +24,7 @@ unsigned long __clear_user(void __user *addr, unsigned long size)
> >  	asm volatile(
> >  		"	testq  %[size8],%[size8]\n"
> >  		"	jz     4f\n"
> > +		"	.align 16\n"
> >  		"0:	movq $0,(%[dst])\n"
> >  		"	addq   $8,%[dst]\n"
> >  		"	decl %%ecx ; jnz   0b\n"
> 
> You can do better that that loop.
> Change 'dst' to point to the end of the buffer, negate the count
> and divide by 8 and you get:
> 		"0:	movq $0,($[dst],%%ecx,8)\n"
> 		"	add $1,%%ecx"
> 		"	jnz 0b\n"
> which might run at one iteration per clock especially on cpu that pair
> the add and jnz into a single uop.
> (You need to use add not inc.)

/dev/zero should probably use REP STOSB etc just like everything else.