[PATCH 2/2] sha1/sha256: use be32_to_cpu and cpu_to_be32

[Andre]

On 09/24/2010 01:34 AM, Jean-Christophe PLAGNIOL-VILLARD wrote:
> On 09:43 Fri 24 Sep     , Sascha Hauer wrote:
>> On Fri, Sep 24, 2010 at 12:00:42AM -0700, Andre wrote:
>>> On 09/21/2010 06:28 AM, Jean-Christophe PLAGNIOL-VILLARD wrote:
>>>
>>>> Signed-off-by: Jean-Christophe PLAGNIOL-VILLARD<plagnioj at jcrosoft.com>
>>>> ---
>>>>    lib/sha1.c   |   20 +++-----------------
>>>>    lib/sha256.c |   19 +++----------------
>>>>    2 files changed, 6 insertions(+), 33 deletions(-)
>>>>
>>>> diff --git a/lib/sha1.c b/lib/sha1.c
>>>> index 0e8aed1..b4e2abc 100644
>>>> --- a/lib/sha1.c
>>>> +++ b/lib/sha1.c
>>>> @@ -29,6 +29,7 @@
>>>>    #include<digest.h>
>>>>    #include<init.h>
>>>>    #include<linux/string.h>
>>>> +#include<asm/byteorder.h>
>>>>
>>>>    #define SHA1_SUM_POS	-0x20
>>>>    #define SHA1_SUM_LEN	20
>>>> @@ -44,23 +45,8 @@ sha1_context;
>>>>    /*
>>>>     * 32-bit integer manipulation macros (big endian)
>>>>     */
>>>> -#ifndef GET_UINT32_BE
>>>> -#define GET_UINT32_BE(n,b,i) {				\
>>>> -	(n) = ( (uint32_t) (b)[(i)    ]<<   24 )	\
>>>> -	    | ( (uint32_t) (b)[(i) + 1]<<   16 )	\
>>>> -	    | ( (uint32_t) (b)[(i) + 2]<<    8 )	\
>>>> -	    | ( (uint32_t) (b)[(i) + 3]       );	\
>>>> -}
>>>> -#endif
>>>> -
>>>> -#ifndef PUT_UINT32_BE
>>>> -#define PUT_UINT32_BE(n,b,i) {				\
>>>> -	(b)[(i)    ] = (unsigned char) ( (n)>>   24 );	\
>>>> -	(b)[(i) + 1] = (unsigned char) ( (n)>>   16 );	\
>>>> -	(b)[(i) + 2] = (unsigned char) ( (n)>>    8 );	\
>>>> -	(b)[(i) + 3] = (unsigned char) ( (n)       );	\
>>>> -}
>>>> -#endif
>>>> +#define GET_UINT32_BE(n,b,i) (n) = be32_to_cpu(((uint32_t*)(b))[i / 4])
>>>> +#define PUT_UINT32_BE(n,b,i) ((uint32_t*)(b))[i / 4] = cpu_to_be32(n)
>>>>
>>>>
>>>
>>> The previous macros served two purposes: endian swapping and performing
>>> the memory accesses byte-by-byte. New versions are unsafe for CPUs which
>>> do not support misaligned 32bit memory accesses.
>>
>> Indeed. We have get_unaligned_be32() / put_unaligned_be32(). These should be
>> the correct functions, right?
 >
> no-nned IIRC as be32_to_cpu and cpu_to_be32 already handle this
> depending on the arch
>

I think get_unaligned_be32() / put_unaligned_be32() are correct in this 
case. be32_to_cpu / cpu_to_be32 perform endian swapping (if required) 
with source and destination both being 32bit variables, not memory 
locations ?

Of course the easy way to test any version is to build for an 
architecture which cares about alignment and look at the disassembly. If 
the compiler generates one 32bit load/store instruction instead of 4 
byte accesses then the code is wrong.

In any case, this looks dubious:

#define PUT_UINT32_BE(n,b,i) ((uint32_t*)(b))[i / 4] = cpu_to_be32(n)

Behaviour when i == 0 is the same as when i == 1, which wasn't the case 
with the old macros. Also, if b is not 32bit aligned, store will be 
misaligned regardless of having cpu_to_be32(), or anything else, on the rhs.

Andre
--