[PATCH] x86: optimize memcpy_flushcache

[PATCH] x86: optimize memcpy_flushcache

[Mikulas Patocka]

Hi Mike

Could you please push this patch to the kernel 4.18-rc? Dan Williams said 
that he will submit it, but he forgot about it.

Without this patch, dm-writecache is suffering 2% penalty because of 
memcpy_flushcache overhead.

Mikulas



From: Mikulas Patocka <mpatocka@redhat.com>

I use memcpy_flushcache in my persistent memory driver for metadata
updates and it turns out that the overhead of memcpy_flushcache causes 2%
performance degradation compared to "movnti" instruction explicitly coded
using inline assembler.

This patch recognizes memcpy_flushcache calls with constant short length
and turns them into inline assembler - so that I don't have to use inline
assembler in the driver.

Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>

---
 arch/x86/include/asm/string_64.h |   20 +++++++++++++++++++-
 arch/x86/lib/usercopy_64.c       |    4 ++--
 2 files changed, 21 insertions(+), 3 deletions(-)

Index: linux-2.6/arch/x86/include/asm/string_64.h
===================================================================
--- linux-2.6.orig/arch/x86/include/asm/string_64.h
+++ linux-2.6/arch/x86/include/asm/string_64.h
@@ -149,7 +149,25 @@ memcpy_mcsafe(void *dst, const void *src
 
 #ifdef CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE
 #define __HAVE_ARCH_MEMCPY_FLUSHCACHE 1
-void memcpy_flushcache(void *dst, const void *src, size_t cnt);
+void __memcpy_flushcache(void *dst, const void *src, size_t cnt);
+static __always_inline void memcpy_flushcache(void *dst, const void *src, size_t cnt)
+{
+	if (__builtin_constant_p(cnt)) {
+		switch (cnt) {
+			case 4:
+				asm ("movntil %1, %0" : "=m"(*(u32 *)dst) : "r"(*(u32 *)src));
+				return;
+			case 8:
+				asm ("movntiq %1, %0" : "=m"(*(u64 *)dst) : "r"(*(u64 *)src));
+				return;
+			case 16:
+				asm ("movntiq %1, %0" : "=m"(*(u64 *)dst) : "r"(*(u64 *)src));
+				asm ("movntiq %1, %0" : "=m"(*(u64 *)(dst + 8)) : "r"(*(u64 *)(src + 8)));
+				return;
+		}
+	}
+	__memcpy_flushcache(dst, src, cnt);
+}
 #endif
 
 #endif /* __KERNEL__ */
Index: linux-2.6/arch/x86/lib/usercopy_64.c
===================================================================
--- linux-2.6.orig/arch/x86/lib/usercopy_64.c
+++ linux-2.6/arch/x86/lib/usercopy_64.c
@@ -153,7 +153,7 @@ long __copy_user_flushcache(void *dst, c
 	return rc;
 }
 
-void memcpy_flushcache(void *_dst, const void *_src, size_t size)
+void __memcpy_flushcache(void *_dst, const void *_src, size_t size)
 {
 	unsigned long dest = (unsigned long) _dst;
 	unsigned long source = (unsigned long) _src;
@@ -216,7 +216,7 @@ void memcpy_flushcache(void *_dst, const
 		clean_cache_range((void *) dest, size);
 	}
 }
-EXPORT_SYMBOL_GPL(memcpy_flushcache);
+EXPORT_SYMBOL_GPL(__memcpy_flushcache);
 
 void memcpy_page_flushcache(char *to, struct page *page, size_t offset,
 		size_t len)
_______________________________________________
Linux-nvdimm mailing list
Linux-nvdimm@lists.01.org
https://lists.01.org/mailman/listinfo/linux-nvdimm

[PATCH] x86: optimize memcpy_flushcache

[Mikulas Patocka]

Hi Mike

Could you please push this patch to the kernel 4.18-rc? Dan Williams said 
that he will submit it, but he forgot about it.

Without this patch, dm-writecache is suffering 2% penalty because of 
memcpy_flushcache overhead.

Mikulas



From: Mikulas Patocka <mpatocka-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>

I use memcpy_flushcache in my persistent memory driver for metadata
updates and it turns out that the overhead of memcpy_flushcache causes 2%
performance degradation compared to "movnti" instruction explicitly coded
using inline assembler.

This patch recognizes memcpy_flushcache calls with constant short length
and turns them into inline assembler - so that I don't have to use inline
assembler in the driver.

Signed-off-by: Mikulas Patocka <mpatocka-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>

---
 arch/x86/include/asm/string_64.h |   20 +++++++++++++++++++-
 arch/x86/lib/usercopy_64.c       |    4 ++--
 2 files changed, 21 insertions(+), 3 deletions(-)

Index: linux-2.6/arch/x86/include/asm/string_64.h
===================================================================
--- linux-2.6.orig/arch/x86/include/asm/string_64.h
+++ linux-2.6/arch/x86/include/asm/string_64.h
@@ -149,7 +149,25 @@ memcpy_mcsafe(void *dst, const void *src
 
 #ifdef CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE
 #define __HAVE_ARCH_MEMCPY_FLUSHCACHE 1
-void memcpy_flushcache(void *dst, const void *src, size_t cnt);
+void __memcpy_flushcache(void *dst, const void *src, size_t cnt);
+static __always_inline void memcpy_flushcache(void *dst, const void *src, size_t cnt)
+{
+	if (__builtin_constant_p(cnt)) {
+		switch (cnt) {
+			case 4:
+				asm ("movntil %1, %0" : "=m"(*(u32 *)dst) : "r"(*(u32 *)src));
+				return;
+			case 8:
+				asm ("movntiq %1, %0" : "=m"(*(u64 *)dst) : "r"(*(u64 *)src));
+				return;
+			case 16:
+				asm ("movntiq %1, %0" : "=m"(*(u64 *)dst) : "r"(*(u64 *)src));
+				asm ("movntiq %1, %0" : "=m"(*(u64 *)(dst + 8)) : "r"(*(u64 *)(src + 8)));
+				return;
+		}
+	}
+	__memcpy_flushcache(dst, src, cnt);
+}
 #endif
 
 #endif /* __KERNEL__ */
Index: linux-2.6/arch/x86/lib/usercopy_64.c
===================================================================
--- linux-2.6.orig/arch/x86/lib/usercopy_64.c
+++ linux-2.6/arch/x86/lib/usercopy_64.c
@@ -153,7 +153,7 @@ long __copy_user_flushcache(void *dst, c
 	return rc;
 }
 
-void memcpy_flushcache(void *_dst, const void *_src, size_t size)
+void __memcpy_flushcache(void *_dst, const void *_src, size_t size)
 {
 	unsigned long dest = (unsigned long) _dst;
 	unsigned long source = (unsigned long) _src;
@@ -216,7 +216,7 @@ void memcpy_flushcache(void *_dst, const
 		clean_cache_range((void *) dest, size);
 	}
 }
-EXPORT_SYMBOL_GPL(memcpy_flushcache);
+EXPORT_SYMBOL_GPL(__memcpy_flushcache);
 
 void memcpy_page_flushcache(char *to, struct page *page, size_t offset,
 		size_t len)

Re: x86: optimize memcpy_flushcache

[Mike Snitzer]

On Mon, Jun 18 2018 at  8:50am -0400,
Mikulas Patocka <mpatocka@redhat.com> wrote:

> Hi Mike
> 
> Could you please push this patch to the kernel 4.18-rc? Dan Williams said 
> that he will submit it, but he forgot about it.
> 
> Without this patch, dm-writecache is suffering 2% penalty because of 
> memcpy_flushcache overhead.

I cannot send this to Linus directly, it needs to go through the x86
tree.

I already tried to get a slightly revised version of this upstream, see:
https://www.redhat.com/archives/dm-devel/2018-May/msg00080.html

I'll try a resend.. but the 4.18 merge window is now closed.

Mike


> From: Mikulas Patocka <mpatocka@redhat.com>
> 
> I use memcpy_flushcache in my persistent memory driver for metadata
> updates and it turns out that the overhead of memcpy_flushcache causes 2%
> performance degradation compared to "movnti" instruction explicitly coded
> using inline assembler.
> 
> This patch recognizes memcpy_flushcache calls with constant short length
> and turns them into inline assembler - so that I don't have to use inline
> assembler in the driver.
> 
> Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
> 
> ---
>  arch/x86/include/asm/string_64.h |   20 +++++++++++++++++++-
>  arch/x86/lib/usercopy_64.c       |    4 ++--
>  2 files changed, 21 insertions(+), 3 deletions(-)
> 
> Index: linux-2.6/arch/x86/include/asm/string_64.h
> ===================================================================
> --- linux-2.6.orig/arch/x86/include/asm/string_64.h
> +++ linux-2.6/arch/x86/include/asm/string_64.h
> @@ -149,7 +149,25 @@ memcpy_mcsafe(void *dst, const void *src
>  
>  #ifdef CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE
>  #define __HAVE_ARCH_MEMCPY_FLUSHCACHE 1
> -void memcpy_flushcache(void *dst, const void *src, size_t cnt);
> +void __memcpy_flushcache(void *dst, const void *src, size_t cnt);
> +static __always_inline void memcpy_flushcache(void *dst, const void *src, size_t cnt)
> +{
> +	if (__builtin_constant_p(cnt)) {
> +		switch (cnt) {
> +			case 4:
> +				asm ("movntil %1, %0" : "=m"(*(u32 *)dst) : "r"(*(u32 *)src));
> +				return;
> +			case 8:
> +				asm ("movntiq %1, %0" : "=m"(*(u64 *)dst) : "r"(*(u64 *)src));
> +				return;
> +			case 16:
> +				asm ("movntiq %1, %0" : "=m"(*(u64 *)dst) : "r"(*(u64 *)src));
> +				asm ("movntiq %1, %0" : "=m"(*(u64 *)(dst + 8)) : "r"(*(u64 *)(src + 8)));
> +				return;
> +		}
> +	}
> +	__memcpy_flushcache(dst, src, cnt);
> +}
>  #endif
>  
>  #endif /* __KERNEL__ */
> Index: linux-2.6/arch/x86/lib/usercopy_64.c
> ===================================================================
> --- linux-2.6.orig/arch/x86/lib/usercopy_64.c
> +++ linux-2.6/arch/x86/lib/usercopy_64.c
> @@ -153,7 +153,7 @@ long __copy_user_flushcache(void *dst, c
>  	return rc;
>  }
>  
> -void memcpy_flushcache(void *_dst, const void *_src, size_t size)
> +void __memcpy_flushcache(void *_dst, const void *_src, size_t size)
>  {
>  	unsigned long dest = (unsigned long) _dst;
>  	unsigned long source = (unsigned long) _src;
> @@ -216,7 +216,7 @@ void memcpy_flushcache(void *_dst, const
>  		clean_cache_range((void *) dest, size);
>  	}
>  }
> -EXPORT_SYMBOL_GPL(memcpy_flushcache);
> +EXPORT_SYMBOL_GPL(__memcpy_flushcache);
>  
>  void memcpy_page_flushcache(char *to, struct page *page, size_t offset,
>  		size_t len)
_______________________________________________
Linux-nvdimm mailing list
Linux-nvdimm@lists.01.org
https://lists.01.org/mailman/listinfo/linux-nvdimm

Re: x86: optimize memcpy_flushcache

[Mike Snitzer]

On Mon, Jun 18 2018 at  8:50am -0400,
Mikulas Patocka <mpatocka-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org> wrote:

> Hi Mike
> 
> Could you please push this patch to the kernel 4.18-rc? Dan Williams said 
> that he will submit it, but he forgot about it.
> 
> Without this patch, dm-writecache is suffering 2% penalty because of 
> memcpy_flushcache overhead.

I cannot send this to Linus directly, it needs to go through the x86
tree.

I already tried to get a slightly revised version of this upstream, see:
https://www.redhat.com/archives/dm-devel/2018-May/msg00080.html

I'll try a resend.. but the 4.18 merge window is now closed.

Mike


> From: Mikulas Patocka <mpatocka-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
> 
> I use memcpy_flushcache in my persistent memory driver for metadata
> updates and it turns out that the overhead of memcpy_flushcache causes 2%
> performance degradation compared to "movnti" instruction explicitly coded
> using inline assembler.
> 
> This patch recognizes memcpy_flushcache calls with constant short length
> and turns them into inline assembler - so that I don't have to use inline
> assembler in the driver.
> 
> Signed-off-by: Mikulas Patocka <mpatocka-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
> 
> ---
>  arch/x86/include/asm/string_64.h |   20 +++++++++++++++++++-
>  arch/x86/lib/usercopy_64.c       |    4 ++--
>  2 files changed, 21 insertions(+), 3 deletions(-)
> 
> Index: linux-2.6/arch/x86/include/asm/string_64.h
> ===================================================================
> --- linux-2.6.orig/arch/x86/include/asm/string_64.h
> +++ linux-2.6/arch/x86/include/asm/string_64.h
> @@ -149,7 +149,25 @@ memcpy_mcsafe(void *dst, const void *src
>  
>  #ifdef CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE
>  #define __HAVE_ARCH_MEMCPY_FLUSHCACHE 1
> -void memcpy_flushcache(void *dst, const void *src, size_t cnt);
> +void __memcpy_flushcache(void *dst, const void *src, size_t cnt);
> +static __always_inline void memcpy_flushcache(void *dst, const void *src, size_t cnt)
> +{
> +	if (__builtin_constant_p(cnt)) {
> +		switch (cnt) {
> +			case 4:
> +				asm ("movntil %1, %0" : "=m"(*(u32 *)dst) : "r"(*(u32 *)src));
> +				return;
> +			case 8:
> +				asm ("movntiq %1, %0" : "=m"(*(u64 *)dst) : "r"(*(u64 *)src));
> +				return;
> +			case 16:
> +				asm ("movntiq %1, %0" : "=m"(*(u64 *)dst) : "r"(*(u64 *)src));
> +				asm ("movntiq %1, %0" : "=m"(*(u64 *)(dst + 8)) : "r"(*(u64 *)(src + 8)));
> +				return;
> +		}
> +	}
> +	__memcpy_flushcache(dst, src, cnt);
> +}
>  #endif
>  
>  #endif /* __KERNEL__ */
> Index: linux-2.6/arch/x86/lib/usercopy_64.c
> ===================================================================
> --- linux-2.6.orig/arch/x86/lib/usercopy_64.c
> +++ linux-2.6/arch/x86/lib/usercopy_64.c
> @@ -153,7 +153,7 @@ long __copy_user_flushcache(void *dst, c
>  	return rc;
>  }
>  
> -void memcpy_flushcache(void *_dst, const void *_src, size_t size)
> +void __memcpy_flushcache(void *_dst, const void *_src, size_t size)
>  {
>  	unsigned long dest = (unsigned long) _dst;
>  	unsigned long source = (unsigned long) _src;
> @@ -216,7 +216,7 @@ void memcpy_flushcache(void *_dst, const
>  		clean_cache_range((void *) dest, size);
>  	}
>  }
> -EXPORT_SYMBOL_GPL(memcpy_flushcache);
> +EXPORT_SYMBOL_GPL(__memcpy_flushcache);
>  
>  void memcpy_page_flushcache(char *to, struct page *page, size_t offset,
>  		size_t len)

Re: [PATCH] x86: optimize memcpy_flushcache

[Dan Williams]

On Mon, Jun 18, 2018 at 5:50 AM, Mikulas Patocka <mpatocka@redhat.com> wrote:
> Hi Mike
>
> Could you please push this patch to the kernel 4.18-rc? Dan Williams said
> that he will submit it, but he forgot about it.

...to be clear I acked it and asked Ingo to take it. Will need a
resubmit for 4.19.

Ingo, see below for a patch to pick up into -tip when you have a chance.

>
> Without this patch, dm-writecache is suffering 2% penalty because of
> memcpy_flushcache overhead.
>
> Mikulas
>
>
>
> From: Mikulas Patocka <mpatocka@redhat.com>
>
> I use memcpy_flushcache in my persistent memory driver for metadata
> updates and it turns out that the overhead of memcpy_flushcache causes 2%
> performance degradation compared to "movnti" instruction explicitly coded
> using inline assembler.
>
> This patch recognizes memcpy_flushcache calls with constant short length
> and turns them into inline assembler - so that I don't have to use inline
> assembler in the driver.
>
> Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
>
> ---
>  arch/x86/include/asm/string_64.h |   20 +++++++++++++++++++-
>  arch/x86/lib/usercopy_64.c       |    4 ++--
>  2 files changed, 21 insertions(+), 3 deletions(-)
>
> Index: linux-2.6/arch/x86/include/asm/string_64.h
> ===================================================================
> --- linux-2.6.orig/arch/x86/include/asm/string_64.h
> +++ linux-2.6/arch/x86/include/asm/string_64.h
> @@ -149,7 +149,25 @@ memcpy_mcsafe(void *dst, const void *src
>
>  #ifdef CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE
>  #define __HAVE_ARCH_MEMCPY_FLUSHCACHE 1
> -void memcpy_flushcache(void *dst, const void *src, size_t cnt);
> +void __memcpy_flushcache(void *dst, const void *src, size_t cnt);
> +static __always_inline void memcpy_flushcache(void *dst, const void *src, size_t cnt)
> +{
> +       if (__builtin_constant_p(cnt)) {
> +               switch (cnt) {
> +                       case 4:
> +                               asm ("movntil %1, %0" : "=m"(*(u32 *)dst) : "r"(*(u32 *)src));
> +                               return;
> +                       case 8:
> +                               asm ("movntiq %1, %0" : "=m"(*(u64 *)dst) : "r"(*(u64 *)src));
> +                               return;
> +                       case 16:
> +                               asm ("movntiq %1, %0" : "=m"(*(u64 *)dst) : "r"(*(u64 *)src));
> +                               asm ("movntiq %1, %0" : "=m"(*(u64 *)(dst + 8)) : "r"(*(u64 *)(src + 8)));
> +                               return;
> +               }
> +       }
> +       __memcpy_flushcache(dst, src, cnt);
> +}
>  #endif
>
>  #endif /* __KERNEL__ */
> Index: linux-2.6/arch/x86/lib/usercopy_64.c
> ===================================================================
> --- linux-2.6.orig/arch/x86/lib/usercopy_64.c
> +++ linux-2.6/arch/x86/lib/usercopy_64.c
> @@ -153,7 +153,7 @@ long __copy_user_flushcache(void *dst, c
>         return rc;
>  }
>
> -void memcpy_flushcache(void *_dst, const void *_src, size_t size)
> +void __memcpy_flushcache(void *_dst, const void *_src, size_t size)
>  {
>         unsigned long dest = (unsigned long) _dst;
>         unsigned long source = (unsigned long) _src;
> @@ -216,7 +216,7 @@ void memcpy_flushcache(void *_dst, const
>                 clean_cache_range((void *) dest, size);
>         }
>  }
> -EXPORT_SYMBOL_GPL(memcpy_flushcache);
> +EXPORT_SYMBOL_GPL(__memcpy_flushcache);
>
>  void memcpy_page_flushcache(char *to, struct page *page, size_t offset,
>                 size_t len)
_______________________________________________
Linux-nvdimm mailing list
Linux-nvdimm@lists.01.org
https://lists.01.org/mailman/listinfo/linux-nvdimm

Re: [PATCH] x86: optimize memcpy_flushcache

[Dan Williams]

On Mon, Jun 18, 2018 at 5:50 AM, Mikulas Patocka <mpatocka-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org> wrote:
> Hi Mike
>
> Could you please push this patch to the kernel 4.18-rc? Dan Williams said
> that he will submit it, but he forgot about it.

...to be clear I acked it and asked Ingo to take it. Will need a
resubmit for 4.19.

Ingo, see below for a patch to pick up into -tip when you have a chance.

>
> Without this patch, dm-writecache is suffering 2% penalty because of
> memcpy_flushcache overhead.
>
> Mikulas
>
>
>
> From: Mikulas Patocka <mpatocka-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
>
> I use memcpy_flushcache in my persistent memory driver for metadata
> updates and it turns out that the overhead of memcpy_flushcache causes 2%
> performance degradation compared to "movnti" instruction explicitly coded
> using inline assembler.
>
> This patch recognizes memcpy_flushcache calls with constant short length
> and turns them into inline assembler - so that I don't have to use inline
> assembler in the driver.
>
> Signed-off-by: Mikulas Patocka <mpatocka-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
>
> ---
>  arch/x86/include/asm/string_64.h |   20 +++++++++++++++++++-
>  arch/x86/lib/usercopy_64.c       |    4 ++--
>  2 files changed, 21 insertions(+), 3 deletions(-)
>
> Index: linux-2.6/arch/x86/include/asm/string_64.h
> ===================================================================
> --- linux-2.6.orig/arch/x86/include/asm/string_64.h
> +++ linux-2.6/arch/x86/include/asm/string_64.h
> @@ -149,7 +149,25 @@ memcpy_mcsafe(void *dst, const void *src
>
>  #ifdef CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE
>  #define __HAVE_ARCH_MEMCPY_FLUSHCACHE 1
> -void memcpy_flushcache(void *dst, const void *src, size_t cnt);
> +void __memcpy_flushcache(void *dst, const void *src, size_t cnt);
> +static __always_inline void memcpy_flushcache(void *dst, const void *src, size_t cnt)
> +{
> +       if (__builtin_constant_p(cnt)) {
> +               switch (cnt) {
> +                       case 4:
> +                               asm ("movntil %1, %0" : "=m"(*(u32 *)dst) : "r"(*(u32 *)src));
> +                               return;
> +                       case 8:
> +                               asm ("movntiq %1, %0" : "=m"(*(u64 *)dst) : "r"(*(u64 *)src));
> +                               return;
> +                       case 16:
> +                               asm ("movntiq %1, %0" : "=m"(*(u64 *)dst) : "r"(*(u64 *)src));
> +                               asm ("movntiq %1, %0" : "=m"(*(u64 *)(dst + 8)) : "r"(*(u64 *)(src + 8)));
> +                               return;
> +               }
> +       }
> +       __memcpy_flushcache(dst, src, cnt);
> +}
>  #endif
>
>  #endif /* __KERNEL__ */
> Index: linux-2.6/arch/x86/lib/usercopy_64.c
> ===================================================================
> --- linux-2.6.orig/arch/x86/lib/usercopy_64.c
> +++ linux-2.6/arch/x86/lib/usercopy_64.c
> @@ -153,7 +153,7 @@ long __copy_user_flushcache(void *dst, c
>         return rc;
>  }
>
> -void memcpy_flushcache(void *_dst, const void *_src, size_t size)
> +void __memcpy_flushcache(void *_dst, const void *_src, size_t size)
>  {
>         unsigned long dest = (unsigned long) _dst;
>         unsigned long source = (unsigned long) _src;
> @@ -216,7 +216,7 @@ void memcpy_flushcache(void *_dst, const
>                 clean_cache_range((void *) dest, size);
>         }
>  }
> -EXPORT_SYMBOL_GPL(memcpy_flushcache);
> +EXPORT_SYMBOL_GPL(__memcpy_flushcache);
>
>  void memcpy_page_flushcache(char *to, struct page *page, size_t offset,
>                 size_t len)

Re: [PATCH] x86: optimize memcpy_flushcache

[Yigal Korman]

Hi,
I'm a bit late on this but I have a question about the original patch -
I thought that in order for movnt (movntil, movntiq) to push the data
into the persistency domain (ADR),
one must work with length that is multiple of cacheline size,
otherwise the write-combine buffers remain partially
filled and you need to commit them with a fence (sfence) - which ruins
the whole performance gain you got here.
Am I wrong, are the write-combine buffers are part of the ADR domain
or something?

Thanks,
Yigal

On Mon, Jun 18, 2018 at 7:38 PM, Dan Williams <dan.j.williams@intel.com> wrote:
> On Mon, Jun 18, 2018 at 5:50 AM, Mikulas Patocka <mpatocka@redhat.com> wrote:
>> Hi Mike
>>
>> Could you please push this patch to the kernel 4.18-rc? Dan Williams said
>> that he will submit it, but he forgot about it.
>
> ...to be clear I acked it and asked Ingo to take it. Will need a
> resubmit for 4.19.
>
> Ingo, see below for a patch to pick up into -tip when you have a chance.
>
>>
>> Without this patch, dm-writecache is suffering 2% penalty because of
>> memcpy_flushcache overhead.
>>
>> Mikulas
>>
>>
>>
>> From: Mikulas Patocka <mpatocka@redhat.com>
>>
>> I use memcpy_flushcache in my persistent memory driver for metadata
>> updates and it turns out that the overhead of memcpy_flushcache causes 2%
>> performance degradation compared to "movnti" instruction explicitly coded
>> using inline assembler.
>>
>> This patch recognizes memcpy_flushcache calls with constant short length
>> and turns them into inline assembler - so that I don't have to use inline
>> assembler in the driver.
>>
>> Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
>>
>> ---
>>  arch/x86/include/asm/string_64.h |   20 +++++++++++++++++++-
>>  arch/x86/lib/usercopy_64.c       |    4 ++--
>>  2 files changed, 21 insertions(+), 3 deletions(-)
>>
>> Index: linux-2.6/arch/x86/include/asm/string_64.h
>> ===================================================================
>> --- linux-2.6.orig/arch/x86/include/asm/string_64.h
>> +++ linux-2.6/arch/x86/include/asm/string_64.h
>> @@ -149,7 +149,25 @@ memcpy_mcsafe(void *dst, const void *src
>>
>>  #ifdef CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE
>>  #define __HAVE_ARCH_MEMCPY_FLUSHCACHE 1
>> -void memcpy_flushcache(void *dst, const void *src, size_t cnt);
>> +void __memcpy_flushcache(void *dst, const void *src, size_t cnt);
>> +static __always_inline void memcpy_flushcache(void *dst, const void *src, size_t cnt)
>> +{
>> +       if (__builtin_constant_p(cnt)) {
>> +               switch (cnt) {
>> +                       case 4:
>> +                               asm ("movntil %1, %0" : "=m"(*(u32 *)dst) : "r"(*(u32 *)src));
>> +                               return;
>> +                       case 8:
>> +                               asm ("movntiq %1, %0" : "=m"(*(u64 *)dst) : "r"(*(u64 *)src));
>> +                               return;
>> +                       case 16:
>> +                               asm ("movntiq %1, %0" : "=m"(*(u64 *)dst) : "r"(*(u64 *)src));
>> +                               asm ("movntiq %1, %0" : "=m"(*(u64 *)(dst + 8)) : "r"(*(u64 *)(src + 8)));
>> +                               return;
>> +               }
>> +       }
>> +       __memcpy_flushcache(dst, src, cnt);
>> +}
>>  #endif
>>
>>  #endif /* __KERNEL__ */
>> Index: linux-2.6/arch/x86/lib/usercopy_64.c
>> ===================================================================
>> --- linux-2.6.orig/arch/x86/lib/usercopy_64.c
>> +++ linux-2.6/arch/x86/lib/usercopy_64.c
>> @@ -153,7 +153,7 @@ long __copy_user_flushcache(void *dst, c
>>         return rc;
>>  }
>>
>> -void memcpy_flushcache(void *_dst, const void *_src, size_t size)
>> +void __memcpy_flushcache(void *_dst, const void *_src, size_t size)
>>  {
>>         unsigned long dest = (unsigned long) _dst;
>>         unsigned long source = (unsigned long) _src;
>> @@ -216,7 +216,7 @@ void memcpy_flushcache(void *_dst, const
>>                 clean_cache_range((void *) dest, size);
>>         }
>>  }
>> -EXPORT_SYMBOL_GPL(memcpy_flushcache);
>> +EXPORT_SYMBOL_GPL(__memcpy_flushcache);
>>
>>  void memcpy_page_flushcache(char *to, struct page *page, size_t offset,
>>                 size_t len)
> _______________________________________________
> Linux-nvdimm mailing list
> Linux-nvdimm@lists.01.org
> https://lists.01.org/mailman/listinfo/linux-nvdimm

-- 
You received this message because you are subscribed to the Google Groups "pmem" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pmem+unsubscribe@googlegroups.com.
To post to this group, send email to pmem@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/pmem/CACTTzNbu5FNXiyDeSBgTZCE7x%3DaNKQWvMLnhy5rkE51JJV9Rzg%40mail.gmail.com.
For more options, visit https://groups.google.com/d/optout.

Re: [PATCH] x86: optimize memcpy_flushcache

[Yigal Korman]

Hi,
I'm a bit late on this but I have a question about the original patch -
I thought that in order for movnt (movntil, movntiq) to push the data
into the persistency domain (ADR),
one must work with length that is multiple of cacheline size,
otherwise the write-combine buffers remain partially
filled and you need to commit them with a fence (sfence) - which ruins
the whole performance gain you got here.
Am I wrong, are the write-combine buffers are part of the ADR domain
or something?

Thanks,
Yigal

On Mon, Jun 18, 2018 at 7:38 PM, Dan Williams <dan.j.williams-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org> wrote:
> On Mon, Jun 18, 2018 at 5:50 AM, Mikulas Patocka <mpatocka-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org> wrote:
>> Hi Mike
>>
>> Could you please push this patch to the kernel 4.18-rc? Dan Williams said
>> that he will submit it, but he forgot about it.
>
> ...to be clear I acked it and asked Ingo to take it. Will need a
> resubmit for 4.19.
>
> Ingo, see below for a patch to pick up into -tip when you have a chance.
>
>>
>> Without this patch, dm-writecache is suffering 2% penalty because of
>> memcpy_flushcache overhead.
>>
>> Mikulas
>>
>>
>>
>> From: Mikulas Patocka <mpatocka-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
>>
>> I use memcpy_flushcache in my persistent memory driver for metadata
>> updates and it turns out that the overhead of memcpy_flushcache causes 2%
>> performance degradation compared to "movnti" instruction explicitly coded
>> using inline assembler.
>>
>> This patch recognizes memcpy_flushcache calls with constant short length
>> and turns them into inline assembler - so that I don't have to use inline
>> assembler in the driver.
>>
>> Signed-off-by: Mikulas Patocka <mpatocka-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
>>
>> ---
>>  arch/x86/include/asm/string_64.h |   20 +++++++++++++++++++-
>>  arch/x86/lib/usercopy_64.c       |    4 ++--
>>  2 files changed, 21 insertions(+), 3 deletions(-)
>>
>> Index: linux-2.6/arch/x86/include/asm/string_64.h
>> ===================================================================
>> --- linux-2.6.orig/arch/x86/include/asm/string_64.h
>> +++ linux-2.6/arch/x86/include/asm/string_64.h
>> @@ -149,7 +149,25 @@ memcpy_mcsafe(void *dst, const void *src
>>
>>  #ifdef CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE
>>  #define __HAVE_ARCH_MEMCPY_FLUSHCACHE 1
>> -void memcpy_flushcache(void *dst, const void *src, size_t cnt);
>> +void __memcpy_flushcache(void *dst, const void *src, size_t cnt);
>> +static __always_inline void memcpy_flushcache(void *dst, const void *src, size_t cnt)
>> +{
>> +       if (__builtin_constant_p(cnt)) {
>> +               switch (cnt) {
>> +                       case 4:
>> +                               asm ("movntil %1, %0" : "=m"(*(u32 *)dst) : "r"(*(u32 *)src));
>> +                               return;
>> +                       case 8:
>> +                               asm ("movntiq %1, %0" : "=m"(*(u64 *)dst) : "r"(*(u64 *)src));
>> +                               return;
>> +                       case 16:
>> +                               asm ("movntiq %1, %0" : "=m"(*(u64 *)dst) : "r"(*(u64 *)src));
>> +                               asm ("movntiq %1, %0" : "=m"(*(u64 *)(dst + 8)) : "r"(*(u64 *)(src + 8)));
>> +                               return;
>> +               }
>> +       }
>> +       __memcpy_flushcache(dst, src, cnt);
>> +}
>>  #endif
>>
>>  #endif /* __KERNEL__ */
>> Index: linux-2.6/arch/x86/lib/usercopy_64.c
>> ===================================================================
>> --- linux-2.6.orig/arch/x86/lib/usercopy_64.c
>> +++ linux-2.6/arch/x86/lib/usercopy_64.c
>> @@ -153,7 +153,7 @@ long __copy_user_flushcache(void *dst, c
>>         return rc;
>>  }
>>
>> -void memcpy_flushcache(void *_dst, const void *_src, size_t size)
>> +void __memcpy_flushcache(void *_dst, const void *_src, size_t size)
>>  {
>>         unsigned long dest = (unsigned long) _dst;
>>         unsigned long source = (unsigned long) _src;
>> @@ -216,7 +216,7 @@ void memcpy_flushcache(void *_dst, const
>>                 clean_cache_range((void *) dest, size);
>>         }
>>  }
>> -EXPORT_SYMBOL_GPL(memcpy_flushcache);
>> +EXPORT_SYMBOL_GPL(__memcpy_flushcache);
>>
>>  void memcpy_page_flushcache(char *to, struct page *page, size_t offset,
>>                 size_t len)
> _______________________________________________
> Linux-nvdimm mailing list
> Linux-nvdimm-hn68Rpc1hR1g9hUCZPvPmw@public.gmane.org
> https://lists.01.org/mailman/listinfo/linux-nvdimm

Re: [PATCH] x86: optimize memcpy_flushcache

[Dan Williams]

On Wed, Jun 27, 2018 at 4:23 AM, Yigal Korman <yigal@plexistor.com> wrote:
> Hi,
> I'm a bit late on this but I have a question about the original patch -
> I thought that in order for movnt (movntil, movntiq) to push the data
> into the persistency domain (ADR),
> one must work with length that is multiple of cacheline size,
> otherwise the write-combine buffers remain partially
> filled and you need to commit them with a fence (sfence) - which ruins
> the whole performance gain you got here.
> Am I wrong, are the write-combine buffers are part of the ADR domain
> or something?

The intent is to allow a batch of memcpy_flushcache() calls followed
by a single sfence. Specifying a multiple of a cacheline size does not
necessarily help as sfence is still needed to make sure that the movnt
result has reached the ADR-safe domain.
_______________________________________________
Linux-nvdimm mailing list
Linux-nvdimm@lists.01.org
https://lists.01.org/mailman/listinfo/linux-nvdimm

Re: [PATCH] x86: optimize memcpy_flushcache

[Dan Williams]

On Wed, Jun 27, 2018 at 4:23 AM, Yigal Korman <yigal-/8YdC2HfS5554TAoqtyWWQ@public.gmane.org> wrote:
> Hi,
> I'm a bit late on this but I have a question about the original patch -
> I thought that in order for movnt (movntil, movntiq) to push the data
> into the persistency domain (ADR),
> one must work with length that is multiple of cacheline size,
> otherwise the write-combine buffers remain partially
> filled and you need to commit them with a fence (sfence) - which ruins
> the whole performance gain you got here.
> Am I wrong, are the write-combine buffers are part of the ADR domain
> or something?

The intent is to allow a batch of memcpy_flushcache() calls followed
by a single sfence. Specifying a multiple of a cacheline size does not
necessarily help as sfence is still needed to make sure that the movnt
result has reached the ADR-safe domain.

Re: [PATCH] x86: optimize memcpy_flushcache

[Yigal Korman]

On Wed, Jun 27, 2018 at 4:03 PM, Dan Williams <dan.j.williams@intel.com> wrote:
> On Wed, Jun 27, 2018 at 4:23 AM, Yigal Korman <yigal@plexistor.com> wrote:
>> Hi,
>> I'm a bit late on this but I have a question about the original patch -
>> I thought that in order for movnt (movntil, movntiq) to push the data
>> into the persistency domain (ADR),
>> one must work with length that is multiple of cacheline size,
>> otherwise the write-combine buffers remain partially
>> filled and you need to commit them with a fence (sfence) - which ruins
>> the whole performance gain you got here.
>> Am I wrong, are the write-combine buffers are part of the ADR domain
>> or something?
>
> The intent is to allow a batch of memcpy_flushcache() calls followed
> by a single sfence. Specifying a multiple of a cacheline size does not
> necessarily help as sfence is still needed to make sure that the movnt
> result has reached the ADR-safe domain.

Oh, right, I see that dm-writecache calls writecache_commit_flushed
which in turn calls wmb().
I keep confusing *_nocache (i.e. copy_user_nocache) that includes
sfence and *_flushcache (i.e. memcpy_flushcache) that doesn't.
Thanks for the clear up.

-- 
You received this message because you are subscribed to the Google Groups "pmem" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pmem+unsubscribe@googlegroups.com.
To post to this group, send email to pmem@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/pmem/CACTTzNZOK8cYZBcSjShFKYCoaW33cFtCZP7bN1CmQg9ZXsVf2w%40mail.gmail.com.
For more options, visit https://groups.google.com/d/optout.

Re: [PATCH] x86: optimize memcpy_flushcache

[Yigal Korman]

On Wed, Jun 27, 2018 at 4:03 PM, Dan Williams <dan.j.williams-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org> wrote:
> On Wed, Jun 27, 2018 at 4:23 AM, Yigal Korman <yigal-/8YdC2HfS5554TAoqtyWWQ@public.gmane.org> wrote:
>> Hi,
>> I'm a bit late on this but I have a question about the original patch -
>> I thought that in order for movnt (movntil, movntiq) to push the data
>> into the persistency domain (ADR),
>> one must work with length that is multiple of cacheline size,
>> otherwise the write-combine buffers remain partially
>> filled and you need to commit them with a fence (sfence) - which ruins
>> the whole performance gain you got here.
>> Am I wrong, are the write-combine buffers are part of the ADR domain
>> or something?
>
> The intent is to allow a batch of memcpy_flushcache() calls followed
> by a single sfence. Specifying a multiple of a cacheline size does not
> necessarily help as sfence is still needed to make sure that the movnt
> result has reached the ADR-safe domain.

Oh, right, I see that dm-writecache calls writecache_commit_flushed
which in turn calls wmb().
I keep confusing *_nocache (i.e. copy_user_nocache) that includes
sfence and *_flushcache (i.e. memcpy_flushcache) that doesn't.
Thanks for the clear up.