Re: [PATCH v6 6/6] crypto: AF_ALG - add support for key_id

[Tadeusz Struk <tadeusz.struk@intel.com>]

Hi Mat,
On 05/25/2016 05:45 PM, Mat Martineau wrote:
> 
> On Sat, 14 May 2016, Tadeusz Struk wrote:
> 
>> diff --git a/crypto/algif_akcipher.c b/crypto/algif_akcipher.c
>> index e00793d..6733df1 100644
>> --- a/crypto/algif_akcipher.c
>> +++ b/crypto/algif_akcipher.c
>> +static int asym_key_verify(const struct key *key, struct akcipher_request *req)
>> +{
>> +    struct public_key_signature sig;
>> +    char *src = NULL, *in;
>> +    int ret;
>> +
>> +    if (!sg_is_last(req->src)) {
>> +        src = kmalloc(req->src_len, GFP_KERNEL);
>> +        if (!src)
>> +            return -ENOMEM;
>> +        scatterwalk_map_and_copy(src, req->src, 0, req->src_len, 0);
>> +        in = src;
>> +    } else {
>> +        in = sg_virt(req->src);
>> +    }
>> +    sig.pkey_algo = "rsa";
>> +    sig.encoding = "pkcs1";
>> +    /* Need to find a way to pass the hash param */
> 
> Are you referring to sig.digest here? It looks like you will hit a BUG_ON() in public_key_verify_signature() if sig.digest is 0. However, sig.digest is unlikely to be 0 because the struct is not cleared - should fix this, since public_key_verify_signature() will try to follow that random pointer.
> 

Right, I need to have a local buffer for the digest here.

>> +    sig.hash_algo = "sha1";
>> +    sig.digest_size = 20;
>> +    sig.s_size = req->src_len;
>> +    sig.s = src;
>> +    ret = verify_signature(key, NULL, &sig);
> 
> Is the idea to write the signature to the socket, and then read out the expected digest (the digest comparison being done elsewhere)? Is that something that will be supported by a future hardware asymmetric key subtype?

After the verify operation the output will be copied to the user,
and the user needs to verify it.

> 
> verify_signature() ends up calling public_key_verify_signature(), which currently expects to get both the digest and signature as input and returns an error if verification fails. The output of crypto_akcipher_verify() is discarded before public_key_verify_signature() returns so nothing ends up in req->dst to read from the socket.
> 
> ALG_OP_VERIFY should behave the same whether using ALG_SET_PUBKEY or ALG_SET_PUBKEY_ID, and they aren't right now.
> 
> If sig.digest is 0, verify_signature() could return the expected digest in the sig structure and skip the digest comparison it currently does. Then that data could be packaged up in req as if crypto_akcipher_verify() had been called. I don't know if this change confuses the semantics of verify_signature() too much, maybe a new function is required with all the requisite plumbing to the asymmetric key subtype.
> 

We need to copy output to the user to verify because we don't have it.
That will be consistent for both ALG_SET_PUBKEY and ALG_SET_PUBKEY_ID.
Thanks for your comments and sorry for the delayed response. I'll will send v7 soon.
-- 
TS