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

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

On Fri, 2014-03-21 at 14:52 -0400, David Miller wrote:
> From: Eric Dumazet <eric.dumazet@gmail.com>
> Date: Fri, 21 Mar 2014 05:50:50 -0700
> 
> > It looks like a barrier() would be more appropriate.
> 
> barrier() == __asm__ __volatile__(:::"memory")

Indeed, but now you mention it, ip_fast_csum() do not uses volatile
keyword on x86_64, and has no "m" constraint either.

This means that for the following hypothetical networking code :

void foobar(struct iphdr *iph, __be16 newlen, __be16 newid)
{
        iph->tot_len = newlen;
        iph->check = 0;
        iph->check = ip_fast_csum((u8 *)iph, 5);

        pr_err("%p\n", iph);

        iph->id = newid;
        iph->check = 0;
        iph->check = ip_fast_csum((u8 *)iph, 5);
}


ip_fast_csum() is done _once_ only.

Following patch seems needed. Thats one another call for x86 code factorization ...

diff --git a/arch/x86/include/asm/checksum_64.h b/arch/x86/include/asm/checksum_64.h
index e6fd8a026c7b..c67778544880 100644
--- a/arch/x86/include/asm/checksum_64.h
+++ b/arch/x86/include/asm/checksum_64.h
@@ -46,7 +46,7 @@ static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl)
 {
 	unsigned int sum;
 
-	asm("  movl (%1), %0\n"
+	asm volatile("  movl (%1), %0\n"
 	    "  subl $4, %2\n"
 	    "  jbe 2f\n"
 	    "  addl 4(%1), %0\n"