Re: [PATCH v4 11/14] treewide: Prepare to remove VLA usage for AHASH_REQUEST_ON_STACK

[Arnd Bergmann]

On Mon, Jul 16, 2018 at 5:39 AM, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> On Sun, Jul 15, 2018 at 5:01 PM, Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx> wrote:
>> On Sat, Jul 14, 2018 at 07:59:09PM -0700, Kees Cook wrote:
>>> On Sat, Jul 14, 2018 at 7:44 PM, Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx> wrote:
>>> > On Fri, Jul 13, 2018 at 08:07:10PM -0700, Kees Cook wrote:
>>> >>
>>> >> On a plane today I started converting all these to shash. IIUC, it
>>> >> just looks like this (apologies for whitespace damage):
>>> >
>>> > Yes if it doesn't actually make use of SGs then shash would be
>>> > the way to go.  However, for SG users ahash is the best interface.
>>>
>>> Nearly all of them artificially build an sg explicitly to use the
>>> ahash interface. :P
>>>
>>> So, I'll take that as a "yes, do these conversions." :) Thanks!
>>
>> Yeah anything that's doing a single-element SG list should just
>> be converted.
>
> There are a few that are multiple element SG list, but it's a locally
> allocated array of SGs, and filled with data. All easily replaced with
> just calls to ..._update() instead of sg helpers. For example
> net/wireless/lib80211_crypt_tkip.c:
>
> -       sg_init_table(sg, 2);
> -       sg_set_buf(&sg[0], hdr, 16);
> -       sg_set_buf(&sg[1], data, data_len);
> ...
> -       ahash_request_set_tfm(req, tfm_michael);
> -       ahash_request_set_callback(req, 0, NULL, NULL);
> -       ahash_request_set_crypt(req, sg, mic, data_len + 16);
> -       err = crypto_ahash_digest(req);
> -       ahash_request_zero(req);
> +       err = crypto_shash_init(desc);
> +       if (err)
> +               goto out;
> +       err = crypto_shash_update(desc, hdr, 16);
> +       if (err)
> +               goto out;
> +       err = crypto_shash_update(desc, data, data_len);
> +       if (err)
> +               goto out;
> +       err = crypto_shash_final(desc, mic);
> +
> +out:
> +       shash_desc_zero(desc);
>         return err;

There may be a little overhead in calling crypto_shash_update()/
crypto_shash_final() repeatedly compared to calling
crypto_ahash_digest() once. It's probably no worse (or maybe
better) in this case, since we call only three times and there
is less indirection, but if there are any cases with a long sglist,
it would be good to measure the performance difference.

       Arnd