Re: [PATCH v2 00/14] x86: Add support for Clang CFI

[Sami Tolvanen <samitolvanen@google.com>]

On Thu, Aug 26, 2021 at 4:44 AM Peter Zijlstra <peterz@infradead.org> wrote:
>
> On Wed, Aug 25, 2021 at 08:49:36AM -0700, Sami Tolvanen wrote:
> > On Tue, Aug 24, 2021 at 12:47 PM Peter Zijlstra <peterz@infradead.org> wrote:
> > >
> > > On Mon, Aug 23, 2021 at 10:13:04AM -0700, Sami Tolvanen wrote:
> > > > This series adds support for Clang's Control-Flow Integrity (CFI)
> > > > checking to x86_64. With CFI, the compiler injects a runtime
> > > > check before each indirect function call to ensure the target is
> > > > a valid function with the correct static type. This restricts
> > > > possible call targets and makes it more difficult for an attacker
> > > > to exploit bugs that allow the modification of stored function
> > > > pointers. For more details, see:
> > >
> > > If I understand this right; tp_stub_func() in kernel/tracepoint.c
> > > violates this (as would much of the HAVE_STATIC_CALL=n code, luckily
> > > that is not a valid x86_64 configuration).
> > >
> > > Specifically, we assign &tp_stub_func to tracepoint_func::func, but that
> > > function pointer is only ever indirectly called when cast to the
> > > tracepoint prototype:
> > >
> > >   ((void(*)(void *, proto))(it_func))(__data, args);
> > >
> > > (see DEFINE_TRACE_FN() in linux/tracepoint.h)
> > >
> > > This means the indirect function type and the target function type
> > > mismatch, resulting in that runtime check you added to trigger.
> >
> > Thanks for pointing this out. Yes, that would clearly trip CFI.
> >
> > Any concerns about just writing a magic value to the slot instead of
> > pointing it to a stub function, and checking for it before the call?
>
> Performance :-) that compare is going to be useless roughly 100% of the
> time.

Makes sense. I suppose we could move the call into a macro and do the
comparison only when CFI is enabled to avoid a performance hit with
other configs.

> > > Hitting tp_stub_func() at runtime is exceedingly rare, but possible.
> > >
> > > I realize this is strictly UB per C, but realistically any CDECL ABI
> > > requires that any function with arbitrary signature:
> > >
> > >   void foo(...)
> > >   {
> > >   }
> > >
> > > translates to the exact same code. Specifically on x86-64, the super
> > > impressive:
> > >
> > > foo:
> > >         RET
> > >
> > > And as such this works just fine. Except now you wrecked it.
> >
> > Sure. Another option is to disable CFI for the functions that perform
> > the call, but I would rather avoid that whenever possible.
>
> Is there no means of teaching the compiler about these magical
> functions? There's only two possible stubs:
>
>   void foo(...)
>   {
>   }
>
> and
>
>   unsigned long bar(...)
>   {
>         return 0;
>   }
>
> Both exist in the kernel. We can easily give them a special function
> attribute to call them out.

Clang doesn't have a way to always allow calls to specific functions,
but it might be feasible to implement this in the slowpath handler
without explicit compiler support. I'll see if I can come up with
something reasonable for v3.

Sami