Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.38 as permitted sender) client-ip=23.128.96.38;
MIME-Version: 1.0
References: <20230915050125.3609689-1-davidgow@google.com>
In-Reply-To: <20230915050125.3609689-1-davidgow@google.com>
From:   Sami Tolvanen <samitolvanen@google.com>
Date:   Fri, 15 Sep 2023 12:05:50 -0700
Message-ID: <CABCJKueOu+MwgvTcEUY51tJ1YjNjS-3zHhAHP=1TQUC1wd_1VA@mail.gmail.com>
Subject: Re: [RFC PATCH] kunit: Add a macro to wrap a deferred action function
To:     David Gow <davidgow@google.com>
Cc:     Nathan Chancellor <nathan@kernel.org>,
        Kees Cook <keescook@chromium.org>,
        Brendan Higgins <brendan.higgins@linux.dev>,
        Rae Moar <rmoar@google.com>, dlatypov@google.com,
        Benjamin Berg <benjamin.berg@intel.com>,
        Maxime Ripard <mripard@kernel.org>,
        Richard Fitzgerald <rf@opensource.cirrus.com>,
        llvm@lists.linux.dev, linux-kernel@vger.kernel.org,
        kunit-dev@googlegroups.com, linux-kselftest@vger.kernel.org,
        linux-hardening@vger.kernel.org,
        Nick Desaulniers <ndesaulniers@google.com>,
        Tom Rix <trix@redhat.com>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

Hi David,

On Thu, Sep 14, 2023 at 10:01=E2=80=AFPM David Gow <davidgow@google.com> wr=
ote:
>
> KUnit's deferred action API accepts a void(*)(void *) function pointer
> which is called when the test is exited. However, we very frequently
> want to use existing functions which accept a single pointer, but which
> may not be of type void*. While this is probably dodgy enough to be on
> the wrong side of the C standard, it's been often used for similar
> callbacks, and gcc's -Wcast-function-type seems to ignore cases where
> the only difference is the type of the argument, assuming it's
> compatible (i.e., they're both pointers to data).
>
> However, clang 16 has introduced -Wcast-function-type-strict, which no
> longer permits any deviation in function pointer type. This seems to be
> because it'd break CFI, which validates the type of function calls.
>
> This rather ruins our attempts to cast functions to defer them, and
> leaves us with a few options:
> 1. Stick our fingers in our ears an ignore the warning. (It's worked so
>    far, but probably isn't the right thing to do.)
> 2. Find some horrible way of casting which fools the compiler into
>    letting us do the cast. (It'd still break CFI, though.)
> 3. Disable the warning, and CFI for this function. This isn't optimal,
>    but may make sense for test-only code. However, I think we'd have to
>    do this for every function called, not just the caller, so maybe it's
>    not practical.
> 4. Manually write wrappers around any such functions. This is ugly (do
>    we really want two copies of each function, one of which has no type
>    info and just forwards to the other). It could get repetitive.
> 5. Generate these wrappers with a macro. That's what this patch does.
>
> I'm broadly okay with any of the options above, though whatever we go
> with will no doubt require some bikeshedding of details (should these
> wrappers be public, do we dedupe them, etc).
>
> Thoughts?

Using a macro to generate a wrapper is a reasonable approach IMO, and
we've used it before in the kernel to fix type mismatches in
indirectly called functions (v4l2-ioctl and cfg80211 come to mind at
least).

Sami