[1/4] crypto: xts - restrict key lengths to approved values in FIPS mode

  According to FIPS 140-3 IG C.I., only (total) key lengths of either
256 bits or 512 bits are allowed with xts(aes). Make xts_verify_key() to
reject anything else in FIPS mode.

As xts(aes) is the only approved xts() template instantiation in FIPS mode,
the new restriction implemented in xts_verify_key() effectively only
applies to this particular construction.

Signed-off-by: Nicolai Stange <nstange@suse.de>
---
 include/crypto/xts.h | 7 +++++++
 1 file changed, 7 insertions(+)
  

> diff --git a/include/crypto/xts.h b/include/crypto/xts.h
...
> @@ -35,6 +35,13 @@ static inline int xts_verify_key(struct crypto_skcipher
> *tfm,
>  	if (keylen % 2)
>  		return -EINVAL;
> 
> +	/*
> +	 * In FIPS mode only a combined key length of either 256 or
> +	 * 512 bits is allowed, c.f. FIPS 140-3 IG C.I.
> +	 */
> +	if (fips_enabled && keylen != 32 && keylen != 64)
> +		return -EINVAL;
> +
>  	/* ensure that the AES and tweak key are not identical */
>  	if ((fips_enabled || (crypto_skcipher_get_flags(tfm) &
>  			      CRYPTO_TFM_REQ_FORBID_WEAK_KEYS)) &&
> --
> 2.38.0

arch/s390/crypto/aes_s390.c has similar lines:

static int xts_aes_set_key(struct crypto_skcipher *tfm, const u8 *in_key,
                           unsigned int key_len)
{
        struct s390_xts_ctx *xts_ctx = crypto_skcipher_ctx(tfm);
        unsigned long fc;
        int err;

        err = xts_fallback_setkey(tfm, in_key, key_len);
        if (err)
                return err;

        /* In fips mode only 128 bit or 256 bit keys are valid */
        if (fips_enabled && key_len != 32 && key_len != 64)
                return -EINVAL;


xts_fallback_setkey will now enforce that rule when setting up the
fallback algorithm keys, which makes the xts_aes_set_key check
unreachable.

If that fallback setup were not present, then a call to xts_verify_key
might be preferable to enforce any other rules like the WEAK_KEYS
rule.
  

> diff --git a/include/crypto/xts.h b/include/crypto/xts.h
...
> @@ -35,6 +35,13 @@ static inline int xts_verify_key(struct crypto_skcipher
> *tfm,
>  	if (keylen % 2)
>  		return -EINVAL;
> 
> +	/*
> +	 * In FIPS mode only a combined key length of either 256 or
> +	 * 512 bits is allowed, c.f. FIPS 140-3 IG C.I.
> +	 */
> +	if (fips_enabled && keylen != 32 && keylen != 64)
> +		return -EINVAL;
> +
>  	/* ensure that the AES and tweak key are not identical */
>  	if ((fips_enabled || (crypto_skcipher_get_flags(tfm) &
>  			      CRYPTO_TFM_REQ_FORBID_WEAK_KEYS)) &&
> --
> 2.38.0

There's another function in the same file called xts_check_key() 
that is used by some of the hardware drivers:

arch/s390/crypto/paes_s390.c:    * xts_check_key verifies the key length is not odd and makes
 [that references it in the comment but actually calls xts_verify_key in the code]
drivers/crypto/axis/artpec6_crypto.c:   ret = xts_check_key(&cipher->base, key, keylen);
drivers/crypto/cavium/cpt/cptvf_algs.c: err = xts_check_key(tfm, key, keylen);
drivers/crypto/cavium/nitrox/nitrox_skcipher.c: ret = xts_check_key(tfm, key, keylen);
drivers/crypto/ccree/cc_cipher.c:           xts_check_key(tfm, key, keylen)) {
drivers/crypto/marvell/octeontx/otx_cptvf_algs.c:       ret = xts_check_key(crypto_skcipher_tfm(tfm), key, keylen);
drivers/crypto/marvell/octeontx2/otx2_cptvf_algs.c:     ret = xts_check_key(crypto_skcipher_tfm(tfm), key, keylen);
drivers/crypto/atmel-aes.c:     err = xts_check_key(crypto_skcipher_tfm(tfm), key, keylen);

It already has one check qualified by fips_enabled:

        /* ensure that the AES and tweak key are not identical */
        if (fips_enabled && !crypto_memneq(key, key + (keylen / 2), keylen / 2))
                return -EINVAL;

Should that implement the same key length restrictions?
  

"Elliott, Robert (Servers)" <elliott@hpe.com> writes:

>> diff --git a/include/crypto/xts.h b/include/crypto/xts.h
> ...
>> @@ -35,6 +35,13 @@ static inline int xts_verify_key(struct crypto_skcipher
>> *tfm,
>>  	if (keylen % 2)
>>  		return -EINVAL;
>> 
>> +	/*
>> +	 * In FIPS mode only a combined key length of either 256 or
>> +	 * 512 bits is allowed, c.f. FIPS 140-3 IG C.I.
>> +	 */
>> +	if (fips_enabled && keylen != 32 && keylen != 64)
>> +		return -EINVAL;
>> +
>>  	/* ensure that the AES and tweak key are not identical */
>>  	if ((fips_enabled || (crypto_skcipher_get_flags(tfm) &
>>  			      CRYPTO_TFM_REQ_FORBID_WEAK_KEYS)) &&
>> --
>> 2.38.0
>
> There's another function in the same file called xts_check_key() 
> that is used by some of the hardware drivers:

Right, thanks for spotting.

AFAICT, xts_check_key() is the older of the two variants,
xts_verify_key() had been introduced with commit f1c131b45410 ("crypto:
xts - Convert to skcipher"). There had initially only been a single
call from generic crypto/xts.c and the main difference to
xts_check_key() had been that it took a crypto_skcipher for its tfm
argument rather than a plain crypto_tfm as xts_check_key() did.

It seems that over time, xts crypto drivers adopted the newer
xts_verify_key() variant then.

>
> arch/s390/crypto/paes_s390.c:    * xts_check_key verifies the key length is not odd and makes
>  [that references it in the comment but actually calls xts_verify_key in the code]
> drivers/crypto/axis/artpec6_crypto.c:   ret = xts_check_key(&cipher->base, key, keylen);
> drivers/crypto/cavium/cpt/cptvf_algs.c: err = xts_check_key(tfm, key, keylen);
> drivers/crypto/cavium/nitrox/nitrox_skcipher.c: ret = xts_check_key(tfm, key, keylen);
> drivers/crypto/ccree/cc_cipher.c:           xts_check_key(tfm, key, keylen)) {
> drivers/crypto/marvell/octeontx/otx_cptvf_algs.c:       ret = xts_check_key(crypto_skcipher_tfm(tfm), key, keylen);
> drivers/crypto/marvell/octeontx2/otx2_cptvf_algs.c:     ret = xts_check_key(crypto_skcipher_tfm(tfm), key, keylen);
> drivers/crypto/atmel-aes.c:     err = xts_check_key(crypto_skcipher_tfm(tfm), key, keylen);
>
> It already has one check qualified by fips_enabled:
>
>         /* ensure that the AES and tweak key are not identical */
>         if (fips_enabled && !crypto_memneq(key, key + (keylen / 2), keylen / 2))
>                 return -EINVAL;
>
> Should that implement the same key length restrictions?

From a quick glance, all of the above drivers merely convert some
crypto_skcipher to a crypto_tfm before passing it to xts_check_key().

So I think these should all be made to call xts_verify_key() directly
instead, the former xts_check_key() could then get dropped. But that's
more of a cleanup IMO and would probably deserve a separate patch series
on its own.

Thanks!

Nicolai
  

"Elliott, Robert (Servers)" <elliott@hpe.com> writes:

>> diff --git a/include/crypto/xts.h b/include/crypto/xts.h
> ...
>> @@ -35,6 +35,13 @@ static inline int xts_verify_key(struct crypto_skcipher
>> *tfm,
>>  	if (keylen % 2)
>>  		return -EINVAL;
>> 
>> +	/*
>> +	 * In FIPS mode only a combined key length of either 256 or
>> +	 * 512 bits is allowed, c.f. FIPS 140-3 IG C.I.
>> +	 */
>> +	if (fips_enabled && keylen != 32 && keylen != 64)
>> +		return -EINVAL;
>> +
>>  	/* ensure that the AES and tweak key are not identical */
>>  	if ((fips_enabled || (crypto_skcipher_get_flags(tfm) &
>>  			      CRYPTO_TFM_REQ_FORBID_WEAK_KEYS)) &&
>> --
>> 2.38.0
>
> arch/s390/crypto/aes_s390.c has similar lines:
>
> static int xts_aes_set_key(struct crypto_skcipher *tfm, const u8 *in_key,
>                            unsigned int key_len)
> {
>         struct s390_xts_ctx *xts_ctx = crypto_skcipher_ctx(tfm);
>         unsigned long fc;
>         int err;
>
>         err = xts_fallback_setkey(tfm, in_key, key_len);
>         if (err)
>                 return err;
>
>         /* In fips mode only 128 bit or 256 bit keys are valid */
>         if (fips_enabled && key_len != 32 && key_len != 64)
>                 return -EINVAL;
>
>
> xts_fallback_setkey will now enforce that rule when setting up the
> fallback algorithm keys, which makes the xts_aes_set_key check
> unreachable.

Good finding!

>
> If that fallback setup were not present, then a call to xts_verify_key
> might be preferable to enforce any other rules like the WEAK_KEYS
> rule.
>

So if this patch here would get accepted, I'd propose to remove the then
dead code from aes_s390 afterwards and make an explicit call to
xts_verify_key() instead.

Or shall I split out the XTS patch from this series here and post these
two changes separately then? Herbert, any preferences?

Thanks!

Nicolai
  

On Wed, Nov 09, 2022 at 11:39:19AM +0100, Nicolai Stange wrote:
>
> Or shall I split out the XTS patch from this series here and post these
> two changes separately then? Herbert, any preferences?

You can do this as a follow-up.

Thanks,
  

On Wed, Nov 09, 2022 at 11:06:17AM +0100, Nicolai Stange wrote:
>
> >From a quick glance, all of the above drivers merely convert some
> crypto_skcipher to a crypto_tfm before passing it to xts_check_key().
> 
> So I think these should all be made to call xts_verify_key() directly
> instead, the former xts_check_key() could then get dropped. But that's
> more of a cleanup IMO and would probably deserve a separate patch series
> on its own.

We should make sure both do the same thing though.  So either
change all the drivers or just change xts_check_key in your patch
in addition to xts_verify_key.

Cheers,