Re: TCP Segmentation Offloading (TSO)

[Hirokazu Takahashi]

Hello!

kuznet> > [1] Kudos to 
kuznet> 
kuznet> Hmm... wait awhile with celebrating, the implementation in tcp is still
kuznet> at level of a toy. Well, and it happens to crash, the patch is enclosed.

I guess it may also depend on bad implementations of csum_partial().
It's wrong that some architecture assume every data in a skbuff are
aligned on a 2byte boundary so that it would access a byte next to
the the last byte where no pages might be there.

And we should know sendfile systemcall also pass pages with any offsets
and any byte  while csum_partial() may be called from everywhere
in the kernel including device drivers.

It's time to fix csum_partial().

P.S.
    Using "bswap" is little bit tricky.

Regards,
Hirokazu Takahashi
VP Engineering Dept.
VA Linux Systems Japan

--- linux/arch/i386/lib/checksum.S.BUG	Sun Sep  1 17:00:59 2030
+++ linux/arch/i386/lib/checksum.S	Mon Sep  2 13:09:09 2030
@@ -126,8 +126,8 @@ csum_partial:
 	movl 16(%esp),%ecx	# Function arg: int len
 	movl 12(%esp),%esi	# Function arg:	const unsigned char *buf
 
-	testl $2, %esi         
-	jnz 30f                 
+	testl $3, %esi         
+	jnz 25f                 
 10:
 	movl %ecx, %edx
 	movl %ecx, %ebx
@@ -145,6 +145,20 @@ csum_partial:
 	lea 2(%esi), %esi
 	adcl $0, %eax
 	jmp 10b
+25:
+	testl $1, %esi         
+	jz 30f                 
+	# buf is odd
+	dec %ecx
+	jl 90f
+	bswap %eax
+	movzbl (%esi), %ebx
+	shll $8, %ebx
+	addl %ebx, %eax
+	adcl $0, %eax
+	inc %esi
+	testl $2, %esi
+	jz 10b
 
 30:	subl $2, %ecx          
 	ja 20b                 
@@ -211,6 +225,10 @@ csum_partial:
 	addl %ebx,%eax
 	adcl $0,%eax
 80: 
+	testl $1, 12(%esp)
+	jz 90f
+	bswap %eax
+90: 
 	popl %ebx
 	popl %esi
 	ret