On 07/08/2019 17:13, Pascal Van Leeuwen wrote:
>>>> Seems there is no mistake in your code, it is some bug in aesni_intel
>>>> implementation.
>>>> If I disable this module, it works as expected (with aes generic and
>>>> aes_i586).
>>>>
>>> That's odd though, considering there is a dedicated xts-aes-ni
>>> implementation,
>>> i.e. I would not expect that to end up at the generic xts wrapper at all?
>>
>> Note it is 32bit system, AESNI XTS is under #ifdef CONFIG_X86_64 so it is
>> not used.
>>
> Ok, so I guess no one bothered to make an optimized XTS version for i386.
> I quickly browsed through the code - took me a while to realise the assembly
> is
> "backwards" compared to the original Intel definition :-) - but I did not spot
> anything obvious :-(
>
>> I guess it only ECB part ...
Mystery solved, the skcipher subreq must be te last member in the struct.
(Some comments in Adiantum code mentions it too, so I do not think it
just hides the corruption after the struct. Seems like another magic requirement
in crypto API :-)
This chunk is enough to fix it for me:
--- a/crypto/xts.c
+++ b/crypto/xts.c
@@ -33,8 +33,8 @@ struct xts_instance_ctx {
struct rctx {
le128 t, tcur;
- struct skcipher_request subreq;
int rem_bytes, is_encrypt;
+ struct skcipher_request subreq;
};
While at it, shouldn't be is_encrypt bool?
Thanks,
Milan
