Re: Linux 6.6-rc3 (DEBUG_VIRTUAL is unhappy on x86)

[Sumit Garg]

Hi Linus,

On 10/2/23 02:18, Linus Torvalds wrote:
> On Sun, 1 Oct 2023 at 07:17, Hyeonggon Yoo <42.hyeyoo@xxxxxxxxx> wrote:
> > > > Peter Zijlstra (1):
> > > >        x86,static_call: Fix static-call vs return-thunk
> > > Hello, the commit above caused a crash on x86 kernel with
> > > CONFIG_DEBUG_VIRTUAL=y.
> > OK, I looked into this a little bit, and it turns out that the problematic
> > address here is from cleanup_trusted() in
> > security/keys/trusted-keys/trusted_core.c.
> > (and it's builtin due to CONFIG_TRUSTED_KEYS=y)
> >
> > The function is marked as __exit, so it does not fall within the
> > 'core kernel text address range,' which is between _stext and _etext
> > (or between _sinittext and _einittext). and thus __text_poke() thinks that
> > it's vmalloc/module area.
> >
> > I think __text_poke() should be taught that functions marked as __exit
> > also belong to kernel code just like __init.
> I think your patch is fine (well, whitespace-damaged, but conceptually good).
>
> But I also wonder about that
>
>          static_call_cond(trusted_key_exit)();
>
> in cleanup_trusted(). It seems all kinds of pointless to use static
> calls for something that is done *once*. That's not an optimization,
> that's honestly just _stupid_. It costs more to do the rewriting that
> it does to just do the one dynamic indirect call.

That's true, there isn't any real performance benefit here. It is 
something which I mentioned when I was asked to incorporate it here [1]. 
However, on the flip side I think there are security benefits here. We 
wouldn't like any indirect branch speculation attack to leak the trusted 
key material contents here.

[1] 
https://patchwork.kernel.org/project/keyrings/patch/1602065268-26017-2-git-send-email-sumit.garg@xxxxxxxxxx/#23683269

-Sumit

> Side note: the same is true of the init-time call, which does
>
>                  static_call_update(trusted_key_init,
>                                     trusted_key_sources[i].ops->init);
>                  ...
>                  ret = static_call(trusted_key_init)();
>
> which again is a *lot* more expensive than just doing the indirect
> function call.
>
> So while I don't think your patch is wrong, I do think that the cause
> here is plain silly code, and that trusted key code simply should not
> do the crazy thing it does (and that causes silly problems).
>
>                 Linus