On 7 September 2018 at 05:42, Herbert Xu [off-list ref] wrote:

On Thu, Sep 06, 2018 at 03:58:52PM -0700, Kees Cook wrote:

quoted

@@ -437,6 +442,12 @@ static inline struct crypto_skcipher *crypto_skcipher_reqtfm_check(
 {
      struct crypto_skcipher *tfm = crypto_skcipher_reqtfm(req);

+     if (req->__onstack) {
+             if (WARN_ON(crypto_skcipher_alg(tfm)->base.cra_flags &
+                             CRYPTO_ALG_ASYNC))
+                     return ERR_PTR(-EINVAL);
+     }

Sorry but I don't like imposing a run-time check on everybody when
stack-based requests are the odd ones out.  If we're going to make
this a run-time check (I'd much prefer a compile-time check, but I
understand that this may involve too much churn), then please do it
for stack-based request users only.

OK, so given that all SKCIPHER_REQUEST_ON_STACK occurrences are
updated in this series anyway, perhaps we should add
skcipher_[en|de]crypt_onstack() flavors that encapsulate the
additional check? Only question is how to enforce at compile time that
those are used instead of the ordinary ones when using a stack
allocated request. Would you mind using some macro foo here involving
__builtin_types_compatible_p() ?