[PATCH 0/4] parisc: checksum: Use generic implementations and optimize checksum

[PATCH 0/4] parisc: checksum: Use generic implementations and optimize checksum

[Charlie Jenkins]

After the parisc checksumming functions were created, generic versions
were written that are the same or better, making the architecture
specific ones redundant.

Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>
---
Charlie Jenkins (4):
      asm-generic headers: Allow csum_partial arch override
      parisc: checksum: Use generic implementations
      parisc: checksum: Remove folding from csum_partial
      parisc: checksum: Optimize from32to16

 arch/parisc/Kconfig                |  3 +++
 arch/parisc/include/asm/checksum.h | 42 ++++++++------------------------------
 arch/parisc/lib/checksum.c         | 14 ++++---------
 include/asm-generic/checksum.h     |  2 ++
 lib/checksum.c                     |  2 ++
 5 files changed, 20 insertions(+), 43 deletions(-)
---
base-commit: 6613476e225e090cc9aad49be7fa504e290dd33d
change-id: 20240221-parisc_use_generic_checksum-1bb01d466877
-- 
- Charlie

[PATCH 1/4] asm-generic headers: Allow csum_partial arch override

[Charlie Jenkins]

Arches can have more a efficient implementation of csum_partial.

Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>
---
 include/asm-generic/checksum.h | 2 ++
 lib/checksum.c                 | 2 ++
 2 files changed, 4 insertions(+)

diff --git a/include/asm-generic/checksum.h b/include/asm-generic/checksum.h
index ad928cce268b..3309830ba2cb 100644
--- a/include/asm-generic/checksum.h
+++ b/include/asm-generic/checksum.h
@@ -4,6 +4,7 @@
 
 #include <linux/bitops.h>
 
+#ifndef csum_partial
 /*
  * computes the checksum of a memory block at buff, length len,
  * and adds in "sum" (32-bit)
@@ -17,6 +18,7 @@
  * it's best to have buff aligned on a 32-bit boundary
  */
 extern __wsum csum_partial(const void *buff, int len, __wsum sum);
+#endif
 
 #ifndef ip_fast_csum
 /*
diff --git a/lib/checksum.c b/lib/checksum.c
index 6860d6b05a17..c115a9ac71d9 100644
--- a/lib/checksum.c
+++ b/lib/checksum.c
@@ -110,6 +110,7 @@ __sum16 ip_fast_csum(const void *iph, unsigned int ihl)
 EXPORT_SYMBOL(ip_fast_csum);
 #endif
 
+#ifndef csum_partial
 /*
  * computes the checksum of a memory block at buff, length len,
  * and adds in "sum" (32-bit)
@@ -134,6 +135,7 @@ __wsum csum_partial(const void *buff, int len, __wsum wsum)
 	return (__force __wsum)result;
 }
 EXPORT_SYMBOL(csum_partial);
+#endif
 
 /*
  * this routine is used for miscellaneous IP-like checksums, mainly

-- 
2.34.1

Re: [PATCH 1/4] asm-generic headers: Allow csum_partial arch override

[Guenter Roeck]

On Wed, Feb 21, 2024 at 06:37:11PM -0800, Charlie Jenkins wrote:
> Arches can have more a efficient implementation of csum_partial.
> 
> Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>

Tested-by: Guenter Roeck <linux@roeck-us.net>

> ---
>  include/asm-generic/checksum.h | 2 ++
>  lib/checksum.c                 | 2 ++
>  2 files changed, 4 insertions(+)
> 
> diff --git a/include/asm-generic/checksum.h b/include/asm-generic/checksum.h
> index ad928cce268b..3309830ba2cb 100644
> --- a/include/asm-generic/checksum.h
> +++ b/include/asm-generic/checksum.h
> @@ -4,6 +4,7 @@
>  
>  #include <linux/bitops.h>
>  
> +#ifndef csum_partial
>  /*
>   * computes the checksum of a memory block at buff, length len,
>   * and adds in "sum" (32-bit)
> @@ -17,6 +18,7 @@
>   * it's best to have buff aligned on a 32-bit boundary
>   */
>  extern __wsum csum_partial(const void *buff, int len, __wsum sum);
> +#endif
>  
>  #ifndef ip_fast_csum
>  /*
> diff --git a/lib/checksum.c b/lib/checksum.c
> index 6860d6b05a17..c115a9ac71d9 100644
> --- a/lib/checksum.c
> +++ b/lib/checksum.c
> @@ -110,6 +110,7 @@ __sum16 ip_fast_csum(const void *iph, unsigned int ihl)
>  EXPORT_SYMBOL(ip_fast_csum);
>  #endif
>  
> +#ifndef csum_partial
>  /*
>   * computes the checksum of a memory block at buff, length len,
>   * and adds in "sum" (32-bit)
> @@ -134,6 +135,7 @@ __wsum csum_partial(const void *buff, int len, __wsum wsum)
>  	return (__force __wsum)result;
>  }
>  EXPORT_SYMBOL(csum_partial);
> +#endif
>  
>  /*
>   * this routine is used for miscellaneous IP-like checksums, mainly
> 
> -- 
> 2.34.1
> 

[PATCH 2/4] parisc: checksum: Use generic implementations

[Charlie Jenkins]

The generic implementations of the checksum functions
csum_tcpudp_nofold, csum_fold, and ip_compute_csum are either identical
or perform better than the parisc ones, so use the generic
implementations instead.

In order to use the generic implementations of checksum functions,
do_csum can no longer be static.

Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>
---
 arch/parisc/Kconfig                |  3 +++
 arch/parisc/include/asm/checksum.h | 42 ++++++++------------------------------
 arch/parisc/lib/checksum.c         |  2 +-
 3 files changed, 13 insertions(+), 34 deletions(-)

diff --git a/arch/parisc/Kconfig b/arch/parisc/Kconfig
index d14ccc948a29..1638deb23287 100644
--- a/arch/parisc/Kconfig
+++ b/arch/parisc/Kconfig
@@ -122,6 +122,9 @@ config GENERIC_BUG
 config GENERIC_BUG_RELATIVE_POINTERS
 	bool
 
+config GENERIC_CSUM
+	def_bool y
+
 config GENERIC_HWEIGHT
 	bool
 	default y
diff --git a/arch/parisc/include/asm/checksum.h b/arch/parisc/include/asm/checksum.h
index 3c43baca7b39..c7847a08ef7c 100644
--- a/arch/parisc/include/asm/checksum.h
+++ b/arch/parisc/include/asm/checksum.h
@@ -17,6 +17,7 @@
  * it's best to have buff aligned on a 32-bit boundary
  */
 extern __wsum csum_partial(const void *, int, __wsum);
+#define csum_partial csum_partial
 
 /*
  *	Optimized for IP headers, which always checksum on 4 octet boundaries.
@@ -57,20 +58,8 @@ static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl)
 	return (__force __sum16)sum;
 }
 
-/*
- *	Fold a partial checksum
- */
-static inline __sum16 csum_fold(__wsum csum)
-{
-	u32 sum = (__force u32)csum;
-	/* add the swapped two 16-bit halves of sum,
-	   a possible carry from adding the two 16-bit halves,
-	   will carry from the lower half into the upper half,
-	   giving us the correct sum in the upper half. */
-	sum += (sum << 16) + (sum >> 16);
-	return (__force __sum16)(~sum >> 16);
-}
- 
+#define ip_fast_csum ip_fast_csum
+
 static inline __wsum csum_tcpudp_nofold(__be32 saddr, __be32 daddr,
 					__u32 len, __u8 proto,
 					__wsum sum)
@@ -85,28 +74,15 @@ static inline __wsum csum_tcpudp_nofold(__be32 saddr, __be32 daddr,
 	return sum;
 }
 
-/*
- * computes the checksum of the TCP/UDP pseudo-header
- * returns a 16-bit checksum, already complemented
- */
-static inline __sum16 csum_tcpudp_magic(__be32 saddr, __be32 daddr,
-					__u32 len, __u8 proto,
-					__wsum sum)
-{
-	return csum_fold(csum_tcpudp_nofold(saddr,daddr,len,proto,sum));
-}
-
-/*
- * this routine is used for miscellaneous IP-like checksums, mainly
- * in icmp.c
- */
-static inline __sum16 ip_compute_csum(const void *buf, int len)
-{
-	 return csum_fold (csum_partial(buf, len, 0));
-}
+#define csum_tcpudp_nofold csum_tcpudp_nofold
 
+extern unsigned int do_csum(const unsigned char *buff, int len);
+#define do_csum do_csum
 
 #define _HAVE_ARCH_IPV6_CSUM
+
+#include <asm-generic/checksum.h>
+
 static __inline__ __sum16 csum_ipv6_magic(const struct in6_addr *saddr,
 					  const struct in6_addr *daddr,
 					  __u32 len, __u8 proto,
diff --git a/arch/parisc/lib/checksum.c b/arch/parisc/lib/checksum.c
index 4818f3db84a5..05f5ca4b2f96 100644
--- a/arch/parisc/lib/checksum.c
+++ b/arch/parisc/lib/checksum.c
@@ -34,7 +34,7 @@ static inline unsigned short from32to16(unsigned int x)
 	return (unsigned short)x;
 }
 
-static inline unsigned int do_csum(const unsigned char * buff, int len)
+unsigned int do_csum(const unsigned char *buff, int len)
 {
 	int odd, count;
 	unsigned int result = 0;

-- 
2.34.1

Re: [PATCH 2/4] parisc: checksum: Use generic implementations

[Guenter Roeck]

On Wed, Feb 21, 2024 at 06:37:12PM -0800, Charlie Jenkins wrote:
> The generic implementations of the checksum functions
> csum_tcpudp_nofold, csum_fold, and ip_compute_csum are either identical
> or perform better than the parisc ones, so use the generic
> implementations instead.
> 
> In order to use the generic implementations of checksum functions,
> do_csum can no longer be static.
> 
> Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>

Tested-by: Guenter Roeck <linux@roeck-us.net>

> ---
>  arch/parisc/Kconfig                |  3 +++
>  arch/parisc/include/asm/checksum.h | 42 ++++++++------------------------------
>  arch/parisc/lib/checksum.c         |  2 +-
>  3 files changed, 13 insertions(+), 34 deletions(-)
> 
> diff --git a/arch/parisc/Kconfig b/arch/parisc/Kconfig
> index d14ccc948a29..1638deb23287 100644
> --- a/arch/parisc/Kconfig
> +++ b/arch/parisc/Kconfig
> @@ -122,6 +122,9 @@ config GENERIC_BUG
>  config GENERIC_BUG_RELATIVE_POINTERS
>  	bool
>  
> +config GENERIC_CSUM
> +	def_bool y
> +
>  config GENERIC_HWEIGHT
>  	bool
>  	default y
> diff --git a/arch/parisc/include/asm/checksum.h b/arch/parisc/include/asm/checksum.h
> index 3c43baca7b39..c7847a08ef7c 100644
> --- a/arch/parisc/include/asm/checksum.h
> +++ b/arch/parisc/include/asm/checksum.h
> @@ -17,6 +17,7 @@
>   * it's best to have buff aligned on a 32-bit boundary
>   */
>  extern __wsum csum_partial(const void *, int, __wsum);
> +#define csum_partial csum_partial
>  
>  /*
>   *	Optimized for IP headers, which always checksum on 4 octet boundaries.
> @@ -57,20 +58,8 @@ static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl)
>  	return (__force __sum16)sum;
>  }
>  
> -/*
> - *	Fold a partial checksum
> - */
> -static inline __sum16 csum_fold(__wsum csum)
> -{
> -	u32 sum = (__force u32)csum;
> -	/* add the swapped two 16-bit halves of sum,
> -	   a possible carry from adding the two 16-bit halves,
> -	   will carry from the lower half into the upper half,
> -	   giving us the correct sum in the upper half. */
> -	sum += (sum << 16) + (sum >> 16);
> -	return (__force __sum16)(~sum >> 16);
> -}
> - 
> +#define ip_fast_csum ip_fast_csum
> +
>  static inline __wsum csum_tcpudp_nofold(__be32 saddr, __be32 daddr,
>  					__u32 len, __u8 proto,
>  					__wsum sum)
> @@ -85,28 +74,15 @@ static inline __wsum csum_tcpudp_nofold(__be32 saddr, __be32 daddr,
>  	return sum;
>  }
>  
> -/*
> - * computes the checksum of the TCP/UDP pseudo-header
> - * returns a 16-bit checksum, already complemented
> - */
> -static inline __sum16 csum_tcpudp_magic(__be32 saddr, __be32 daddr,
> -					__u32 len, __u8 proto,
> -					__wsum sum)
> -{
> -	return csum_fold(csum_tcpudp_nofold(saddr,daddr,len,proto,sum));
> -}
> -
> -/*
> - * this routine is used for miscellaneous IP-like checksums, mainly
> - * in icmp.c
> - */
> -static inline __sum16 ip_compute_csum(const void *buf, int len)
> -{
> -	 return csum_fold (csum_partial(buf, len, 0));
> -}
> +#define csum_tcpudp_nofold csum_tcpudp_nofold
>  
> +extern unsigned int do_csum(const unsigned char *buff, int len);
> +#define do_csum do_csum
>  
>  #define _HAVE_ARCH_IPV6_CSUM
> +
> +#include <asm-generic/checksum.h>
> +
>  static __inline__ __sum16 csum_ipv6_magic(const struct in6_addr *saddr,
>  					  const struct in6_addr *daddr,
>  					  __u32 len, __u8 proto,
> diff --git a/arch/parisc/lib/checksum.c b/arch/parisc/lib/checksum.c
> index 4818f3db84a5..05f5ca4b2f96 100644
> --- a/arch/parisc/lib/checksum.c
> +++ b/arch/parisc/lib/checksum.c
> @@ -34,7 +34,7 @@ static inline unsigned short from32to16(unsigned int x)
>  	return (unsigned short)x;
>  }
>  
> -static inline unsigned int do_csum(const unsigned char * buff, int len)
> +unsigned int do_csum(const unsigned char *buff, int len)
>  {
>  	int odd, count;
>  	unsigned int result = 0;
> 
> -- 
> 2.34.1
> 

[PATCH 3/4] parisc: checksum: Remove folding from csum_partial

[Charlie Jenkins]

The parisc implementation of csum_partial previously folded the result
into 16 bits instead of returning all 32 bits and letting consumers like
ip_compute_csum do the folding. Since ip_compute_csum no longer depends
on this requirement, remove the folding so that the parisc
implementation operates the same as other architectures.

Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>
---
 arch/parisc/lib/checksum.c | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/arch/parisc/lib/checksum.c b/arch/parisc/lib/checksum.c
index 05f5ca4b2f96..eaa660491e24 100644
--- a/arch/parisc/lib/checksum.c
+++ b/arch/parisc/lib/checksum.c
@@ -95,14 +95,11 @@ unsigned int do_csum(const unsigned char *buff, int len)
 /*
  * computes a partial checksum, e.g. for TCP/UDP fragments
  */
-/*
- * why bother folding?
- */
 __wsum csum_partial(const void *buff, int len, __wsum sum)
 {
 	unsigned int result = do_csum(buff, len);
 	addc(result, sum);
-	return (__force __wsum)from32to16(result);
+	return (__force __wsum)result;
 }
 
 EXPORT_SYMBOL(csum_partial);

-- 
2.34.1

Re: [PATCH 3/4] parisc: checksum: Remove folding from csum_partial

[Guenter Roeck]

On Wed, Feb 21, 2024 at 06:37:13PM -0800, Charlie Jenkins wrote:
> The parisc implementation of csum_partial previously folded the result
> into 16 bits instead of returning all 32 bits and letting consumers like
> ip_compute_csum do the folding. Since ip_compute_csum no longer depends
> on this requirement, remove the folding so that the parisc
> implementation operates the same as other architectures.
> 
> Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>

Tested-by: Guenter Roeck <linux@roeck-us.net>

> ---
>  arch/parisc/lib/checksum.c | 5 +----
>  1 file changed, 1 insertion(+), 4 deletions(-)
> 
> diff --git a/arch/parisc/lib/checksum.c b/arch/parisc/lib/checksum.c
> index 05f5ca4b2f96..eaa660491e24 100644
> --- a/arch/parisc/lib/checksum.c
> +++ b/arch/parisc/lib/checksum.c
> @@ -95,14 +95,11 @@ unsigned int do_csum(const unsigned char *buff, int len)
>  /*
>   * computes a partial checksum, e.g. for TCP/UDP fragments
>   */
> -/*
> - * why bother folding?
> - */
>  __wsum csum_partial(const void *buff, int len, __wsum sum)
>  {
>  	unsigned int result = do_csum(buff, len);
>  	addc(result, sum);
> -	return (__force __wsum)from32to16(result);
> +	return (__force __wsum)result;
>  }
>  
>  EXPORT_SYMBOL(csum_partial);
> 
> -- 
> 2.34.1
> 

[PATCH 4/4] parisc: checksum: Optimize from32to16

[Charlie Jenkins]

Replace the shifting and masking of x with a rotation. This generates
better assembly.

Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>
---
 arch/parisc/lib/checksum.c | 7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)

diff --git a/arch/parisc/lib/checksum.c b/arch/parisc/lib/checksum.c
index eaa660491e24..1ae8cc730d13 100644
--- a/arch/parisc/lib/checksum.c
+++ b/arch/parisc/lib/checksum.c
@@ -27,11 +27,8 @@
 
 static inline unsigned short from32to16(unsigned int x)
 {
-	/* 32 bits --> 16 bits + carry */
-	x = (x & 0xffff) + (x >> 16);
-	/* 16 bits + carry --> 16 bits including carry */
-	x = (x & 0xffff) + (x >> 16);
-	return (unsigned short)x;
+	x += ror32(x, 16);
+	return (unsigned short)(x >> 16);
 }
 
 unsigned int do_csum(const unsigned char *buff, int len)

-- 
2.34.1

Re: [PATCH 4/4] parisc: checksum: Optimize from32to16

[Guenter Roeck]

On Wed, Feb 21, 2024 at 06:37:14PM -0800, Charlie Jenkins wrote:
> Replace the shifting and masking of x with a rotation. This generates
> better assembly.
> 
> Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>

Tested-by: Guenter Roeck <linux@roeck-us.net>

> ---
>  arch/parisc/lib/checksum.c | 7 ++-----
>  1 file changed, 2 insertions(+), 5 deletions(-)
> 
> diff --git a/arch/parisc/lib/checksum.c b/arch/parisc/lib/checksum.c
> index eaa660491e24..1ae8cc730d13 100644
> --- a/arch/parisc/lib/checksum.c
> +++ b/arch/parisc/lib/checksum.c
> @@ -27,11 +27,8 @@
>  
>  static inline unsigned short from32to16(unsigned int x)
>  {
> -	/* 32 bits --> 16 bits + carry */
> -	x = (x & 0xffff) + (x >> 16);
> -	/* 16 bits + carry --> 16 bits including carry */
> -	x = (x & 0xffff) + (x >> 16);
> -	return (unsigned short)x;
> +	x += ror32(x, 16);
> +	return (unsigned short)(x >> 16);
>  }
>  
>  unsigned int do_csum(const unsigned char *buff, int len)
> 
> -- 
> 2.34.1
>