[PATCH] powerpc/64: Fix checksum folding in csum_add

[LEROY Christophe]

Paul Mackerras <paulus at ozlabs.org> a écrit :

> On Tue, Mar 27, 2018 at 05:22:32PM +0200, LEROY Christophe wrote:
>> Shile Zhang <shile.zhang at nokia.com> a écrit :
>>
>> >fix the missed point in Paul's patch:
>> >"powerpc/64: Fix checksum folding in csum_tcpudp_nofold and
>> >ip_fast_csum_nofold"
>> >
>> >Signed-off-by: Shile Zhang <shile.zhang at nokia.com>
>> >---
>> > arch/powerpc/include/asm/checksum.h | 2 +-
>> > 1 file changed, 1 insertion(+), 1 deletion(-)
>> >
>> >diff --git a/arch/powerpc/include/asm/checksum.h
>> >b/arch/powerpc/include/asm/checksum.h
>> >index 5b1a6e3..430d038 100644
>> >--- a/arch/powerpc/include/asm/checksum.h
>> >+++ b/arch/powerpc/include/asm/checksum.h
>> >@@ -108,7 +108,7 @@ static inline __wsum csum_add(__wsum csum,  
>> __wsum addend)
>> >
>> > #ifdef __powerpc64__
>> > 	res += (__force u64)addend;
>> >-	return (__force __wsum)((u32)res + (res >> 32));
>> >+	return (__force __wsum) from64to32(res);
>>
>> Did you encounter a bug due to that ?
>> As far as i understand, csum and addend are 32 bits so can't exceed  
>> 0xffffffff
>> Then their sum won't exceed 0x1fffffffe. So the sum of upper and lower part
>> won't carry
>
> If the sum of the two halves was 0x1fffffffe, then that previously got
> truncated to 32 bits and returned as 0xfffffffe, which is wrong - the
> result should be 0xffffffff.
>

It is the sum of the two arguments of csum_add() which can't exceed  
0x1fffffffe
So the sum of the two halves will be 0xffffffff which is the expected result.

The issue would start with res = 0x1ffffffff, in that case the sum of  
the two halves would be 0x100000000 which would have been truncated to  
0 whereas we expect the result to be 1 in that case.
But in order to obtain res = 0x1ffffffff or higher, you have to sum at  
least three 32 bits numbers. You can't obtain such a value with a sum  
of two 32 bits numbers.

Christophe