On Sat, Feb 10, 2024 at 1:15 AM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>
> On Thu, Feb 08, 2024 at 03:47:52PM +0000, Alice Ryhl wrote:
> > unsigned long res = n;
> > might_fault();
> > if (!should_fail_usercopy() && likely(access_ok(from, n))) {
> > + /*
> > + * Ensure that bad access_ok() speculation will not
> > + * lead to nasty side effects *after* the copy is
> > + * finished:
> > + */
> > + barrier_nospec();
>
> This means all callers just gained this barrier. That's a behavioral
> change -- is it intentional here? I don't see it mentioned in the commit
> log.
>
> Also did this get tested with the CONFIG_TEST_USER_COPY tests? I would
> expect it to be fine, but better to check and mention it in the commit
> log.
I just ran this with CONFIG_TEST_USER_COPY on x86 using the Android
cuttlefish emulator and it passed there. I also verified that it fails
if I remove the access_ok check. However, the tests succeed even if
the barrier_nospec() call is removed.
That said, it seems like it fails to compile on some other platforms.
It seems like we need to add #include <linux/nospec.h> to uaccess.h to
fix it.
Alice
