Re: [PATCH v4 04/15] cfi: Add DEFINE_CFI_IMMEDIATE_RETURN_STUB

[Sami Tolvanen]

On Mon, Oct 4, 2021 at 11:59 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Mon, Oct 04, 2021 at 12:10:46PM -0700, Sami Tolvanen wrote:
> > On Mon, Oct 4, 2021 at 6:50 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> > > Why DEFINE_CFI_IMMEDIATE_RETURN_STUB() vs __no_cfi attribute that we can
> > > stick on the relvant functions?
> >
> > To avoid accidentally creating useful gadgets for attackers. For
> > example, while excluding an empty stub isn't necessarily ideal,
> > allowing calls to a function that always returns zero would be worse.
>
> I was afraid you'd say something like that...
>
> > > Because I've got at least one more variant for you :-) See
> > > kernel/static_call.c:__static_call_return0
> >
> > Does __static_call_return0 ever get called indirectly on architectures
> > that support static calls? If it's always patched into a direct call,
> > the type mismatch isn't an issue.
>
> For x86_64 it should indeed never get called, however if you plan on
> supporting i386 then you need the annotation. Also, it might get called
> on arm64 which is about to grow basic HAVE_STATIC_CALL support.

Good point. I read through the latest arm64 static call proposal and
while it can fall back to an indirect call, it doesn't look like that
would cause issues with CFI.

> (and just in case you care about CFI on PPC32, they too grew basic
> static_call support)

We are currently targeting only x86_64 and arm64, but I'll keep that
in mind in case we want to add more platforms.

Sami