RE: [RFC] csum experts, csum_replace2() is too expensive

[Eric Dumazet <eric.dumazet@gmail.com>]

On Mon, 2014-03-24 at 06:17 -0700, Eric Dumazet wrote:
> On Mon, 2014-03-24 at 10:30 +0000, David Laight wrote:

> > ip_fast_csum() either needs an explicit "m" constraint for the actual
> > buffer (and target) bytes, or the stronger "memory" constraint.
> > The 'volatile' is then not needed.

I am testing the following :

diff --git a/arch/x86/include/asm/checksum_64.h b/arch/x86/include/asm/checksum_64.h
index e6fd8a026c7b..89d7fa8837b5 100644
--- a/arch/x86/include/asm/checksum_64.h
+++ b/arch/x86/include/asm/checksum_64.h
@@ -45,6 +45,9 @@ static inline __sum16 csum_fold(__wsum sum)
 static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl)
 {
 	unsigned int sum;
+	struct full_ip_header {
+		unsigned int w[ihl];
+	};
 
 	asm("  movl (%1), %0\n"
 	    "  subl $4, %2\n"
@@ -67,8 +70,9 @@ static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl)
 	   are modified, we must also specify them as outputs, or gcc
 	   will assume they contain their original values. */
 	    : "=r" (sum), "=r" (iph), "=r" (ihl)
-	    : "1" (iph), "2" (ihl)
-	    : "memory");
+	    : "1" (iph), "2" (ihl),
+	      "m" (*(struct full_ip_header *) iph)
+	);
 	return (__force __sum16)sum;
 }