[PATCH] x86: Micro-optimise clflush_cache_range()

[PATCH] x86: Micro-optimise clflush_cache_range()

[Chris Wilson]

Whilst inspecting the asm for clflush_cache_range() and some perf profiles
that required extensive flushing of single cachelines (from part of the
intel-gpu-tools GPU benchmarks), we noticed that gcc was reloading
boot_cpu_data.x86_clflush_size on every iteration of the loop. We can
manually hoist that read which perf regarded as taking ~25% of the
function time for a single cacheline flush.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: x86@kernel.org
Cc: Toshi Kani <toshi.kani@hpe.com>
Cc: Borislav Petkov <bp@suse.de>
Cc: "Luis R. Rodriguez" <mcgrof@suse.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Ross Zwisler <ross.zwisler@linux.intel.com>
Cc: Sai Praneeth <sai.praneeth.prakhya@intel.com>
Cc: linux-kernel@vger.kernel.org
Acked-by: "H. Peter Anvin" <hpa@zytor.com>
---
 arch/x86/mm/pageattr.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/arch/x86/mm/pageattr.c b/arch/x86/mm/pageattr.c
index a3137a4feed1..6000ad7f560c 100644
--- a/arch/x86/mm/pageattr.c
+++ b/arch/x86/mm/pageattr.c
@@ -129,14 +129,16 @@ within(unsigned long addr, unsigned long start, unsigned long end)
  */
 void clflush_cache_range(void *vaddr, unsigned int size)
 {
-	unsigned long clflush_mask = boot_cpu_data.x86_clflush_size - 1;
+	const unsigned long clflush_size = boot_cpu_data.x86_clflush_size;
+	void *p = (void *)((unsigned long)vaddr & ~(clflush_size - 1));
 	void *vend = vaddr + size;
-	void *p;
+
+	if (p >= vend)
+		return;
 
 	mb();
 
-	for (p = (void *)((unsigned long)vaddr & ~clflush_mask);
-	     p < vend; p += boot_cpu_data.x86_clflush_size)
+	for (; p < vend; p += clflush_size)
 		clflushopt(p);
 
 	mb();
-- 
2.7.0.rc3

Re: [PATCH] x86: Micro-optimise clflush_cache_range()

[Ross Zwisler]

On Fri, Jan 08, 2016 at 09:55:33AM +0000, Chris Wilson wrote:
> Whilst inspecting the asm for clflush_cache_range() and some perf profiles
> that required extensive flushing of single cachelines (from part of the
> intel-gpu-tools GPU benchmarks), we noticed that gcc was reloading
> boot_cpu_data.x86_clflush_size on every iteration of the loop. We can
> manually hoist that read which perf regarded as taking ~25% of the
> function time for a single cacheline flush.
> 
> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> Cc: Thomas Gleixner <tglx@linutronix.de>
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: "H. Peter Anvin" <hpa@zytor.com>
> Cc: x86@kernel.org
> Cc: Toshi Kani <toshi.kani@hpe.com>
> Cc: Borislav Petkov <bp@suse.de>
> Cc: "Luis R. Rodriguez" <mcgrof@suse.com>
> Cc: Stephen Rothwell <sfr@canb.auug.org.au>
> Cc: Ross Zwisler <ross.zwisler@linux.intel.com>
> Cc: Sai Praneeth <sai.praneeth.prakhya@intel.com>
> Cc: linux-kernel@vger.kernel.org
> Acked-by: "H. Peter Anvin" <hpa@zytor.com>

Looks good to me.

Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>

> ---
>  arch/x86/mm/pageattr.c | 10 ++++++----
>  1 file changed, 6 insertions(+), 4 deletions(-)
> 
> diff --git a/arch/x86/mm/pageattr.c b/arch/x86/mm/pageattr.c
> index a3137a4feed1..6000ad7f560c 100644
> --- a/arch/x86/mm/pageattr.c
> +++ b/arch/x86/mm/pageattr.c
> @@ -129,14 +129,16 @@ within(unsigned long addr, unsigned long start, unsigned long end)
>   */
>  void clflush_cache_range(void *vaddr, unsigned int size)
>  {
> -	unsigned long clflush_mask = boot_cpu_data.x86_clflush_size - 1;
> +	const unsigned long clflush_size = boot_cpu_data.x86_clflush_size;
> +	void *p = (void *)((unsigned long)vaddr & ~(clflush_size - 1));
>  	void *vend = vaddr + size;
> -	void *p;
> +
> +	if (p >= vend)
> +		return;
>  
>  	mb();
>  
> -	for (p = (void *)((unsigned long)vaddr & ~clflush_mask);
> -	     p < vend; p += boot_cpu_data.x86_clflush_size)
> +	for (; p < vend; p += clflush_size)
>  		clflushopt(p);
>  
>  	mb();
> -- 
> 2.7.0.rc3
> 

Re: [PATCH] x86: Micro-optimise clflush_cache_range()

[Toshi Kani]

On Fri, 2016-01-08 at 09:55 +0000, Chris Wilson wrote:
> Whilst inspecting the asm for clflush_cache_range() and some perf
> profiles
> that required extensive flushing of single cachelines (from part of the
> intel-gpu-tools GPU benchmarks), we noticed that gcc was reloading
> boot_cpu_data.x86_clflush_size on every iteration of the loop. We can
> manually hoist that read which perf regarded as taking ~25% of the
> function time for a single cacheline flush.
> 
> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> Cc: Thomas Gleixner <tglx@linutronix.de>
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: "H. Peter Anvin" <hpa@zytor.com>
> Cc: x86@kernel.org
> Cc: Toshi Kani <toshi.kani@hpe.com>
> Cc: Borislav Petkov <bp@suse.de>
> Cc: "Luis R. Rodriguez" <mcgrof@suse.com>
> Cc: Stephen Rothwell <sfr@canb.auug.org.au>
> Cc: Ross Zwisler <ross.zwisler@linux.intel.com>
> Cc: Sai Praneeth <sai.praneeth.prakhya@intel.com>
> Cc: linux-kernel@vger.kernel.org
> Acked-by: "H. Peter Anvin" <hpa@zytor.com>

Thanks for the improvement!  The change looks good to me.

Reviewed-by: Toshi Kani <toshi.kani@hpe.com>

-Toshi

[tip:x86/mm] x86/mm: Micro-optimise clflush_cache_range()

[tip-bot for Chris Wilson]

Commit-ID:  1f1a89ac05f6e88aa341e86e57435fdbb1177c0c
Gitweb:     http://git.kernel.org/tip/1f1a89ac05f6e88aa341e86e57435fdbb1177c0c
Author:     Chris Wilson <chris@chris-wilson.co.uk>
AuthorDate: Fri, 8 Jan 2016 09:55:33 +0000
Committer:  Thomas Gleixner <tglx@linutronix.de>
CommitDate: Fri, 8 Jan 2016 19:27:39 +0100

x86/mm: Micro-optimise clflush_cache_range()

Whilst inspecting the asm for clflush_cache_range() and some perf profiles
that required extensive flushing of single cachelines (from part of the
intel-gpu-tools GPU benchmarks), we noticed that gcc was reloading
boot_cpu_data.x86_clflush_size on every iteration of the loop. We can
manually hoist that read which perf regarded as taking ~25% of the
function time for a single cacheline flush.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Acked-by: "H. Peter Anvin" <hpa@zytor.com>
Cc: Toshi Kani <toshi.kani@hpe.com>
Cc: Borislav Petkov <bp@suse.de>
Cc: Luis R. Rodriguez <mcgrof@suse.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Sai Praneeth <sai.praneeth.prakhya@intel.com>
Link: http://lkml.kernel.org/r/1452246933-10890-1-git-send-email-chris@chris-wilson.co.uk
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
---
 arch/x86/mm/pageattr.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/arch/x86/mm/pageattr.c b/arch/x86/mm/pageattr.c
index a3137a4..6000ad7 100644
--- a/arch/x86/mm/pageattr.c
+++ b/arch/x86/mm/pageattr.c
@@ -129,14 +129,16 @@ within(unsigned long addr, unsigned long start, unsigned long end)
  */
 void clflush_cache_range(void *vaddr, unsigned int size)
 {
-	unsigned long clflush_mask = boot_cpu_data.x86_clflush_size - 1;
+	const unsigned long clflush_size = boot_cpu_data.x86_clflush_size;
+	void *p = (void *)((unsigned long)vaddr & ~(clflush_size - 1));
 	void *vend = vaddr + size;
-	void *p;
+
+	if (p >= vend)
+		return;
 
 	mb();
 
-	for (p = (void *)((unsigned long)vaddr & ~clflush_mask);
-	     p < vend; p += boot_cpu_data.x86_clflush_size)
+	for (; p < vend; p += clflush_size)
 		clflushopt(p);
 
 	mb();