Re: [PATCH v2] x86: disable non-instrumented version of copy_mc when KMSAN is enabled

[Alexander Potapenko <glider@google.com>]

On Thu, Mar 7, 2024 at 1:09 AM Linus Torvalds
<torvalds@linux-foundation.org> wrote:
>
> On Wed, 6 Mar 2024 at 14:08, Tetsuo Handa
> <penguin-kernel@i-love.sakura.ne.jp> wrote:
> >
> > Something like below one?
>
> I'd rather leave the regular fallbacks (to memcpy and copy_to_user())
> alone, and I'd just put the
>
>         kmsan_memmove(dst, src, len - ret);
>
> etc in the places that currently just call the MC copy functions.

(sorry for being late to the party)

We should probably use <linux/instrumented.h> here, as other tools
(KASAN and KCSAN) do not instrument copy_mc_to_kernel() either, and
might benefit from the same checks.

Something along the lines of:

================================
static __always_inline void
instrument_memcpy_before(const void *to, const void *from, unsigned long n)
{
        kasan_check_write(to, n);
        kasan_check_read(from, n);
        kcsan_check_write(to, n);
        kcsan_check_read(from, n);
}

static __always_inline void instrument_memcpy_after(const void *to,
                                                    const void *from,
                                                    unsigned long n,
                                                    unsigned long left)
{
        kmsan_memcpy(to, n - left);
}
================================

(with kmsan_memcpy() working as Tetsuo described).

We can also update copy_mc_fragile_handle_tail() and copy_mc_to_user()
to call the instrumentation hooks.
Let me send the patches.