Re: [PATCH v2] x86: fix early boot crash on gcc-10

[Nick Desaulniers <ndesaulniers@google.com>]

On Fri, Apr 17, 2020 at 2:09 AM Borislav Petkov <bp@alien8.de> wrote:
>
> On Fri, Apr 17, 2020 at 10:58:59AM +0200, Jakub Jelinek wrote:
> > On Fri, Apr 17, 2020 at 10:42:24AM +0200, Borislav Petkov wrote:
> > > On Fri, Apr 17, 2020 at 10:07:26AM +0200, Jakub Jelinek wrote:
> > > > If you want minimal changes, you can as I said earlier either
> > > > mark cpu_startup_entry noreturn (in the declaration in some header so that
> > > > smpboot.c sees it), or you could add something after the cpu_startup_entry
> > > > call to ensure it is not tail call optimized (e.g. just
> > > >   /* Prevent tail call to cpu_startup_entry because the stack
> > > >      protector guard has been changed in the middle of this function
> > > >      and must not be checked before tail calling another function.  */
> > > >   asm ("");
> > >
> > > That sounds ok-ish to me too.
> > >
> > > I know you probably can't tell the future :) but what stops gcc from
> > > doing the tail-call optimization in the future?
> > >
> > > Or are optimization decisions behind an inline asm a no-no and will
> > > pretty much always stay that way?
> >
> > GCC intentionally treats asm as a black box, the only thing which it does

Yep, that's how I would describe how LLVM see's inline asm, too.

> > with it is: non-volatile asm (but asm without outputs is implicitly
> > volatile) can be CSEd, and if the compiler needs to estimate size, it
> > uses some heuristics by counting ; and newlines.
> > And it will stay this way.

I recently implemented parsing support for `asm inline` in Clang; I
could have sworn I saw code in LLVM parsing newlines for a size
estimate years ago, but when implementing `asm inline`, I couldn't
find it.  And test cases I wrote that used the C preprocessor to
create thousand+ line inline asm strings would always be inlined,
regardless of the `inline` asm qualifier.

Not sure about implied volatility (...inner stock trader had a laugh
at that...) for output-less asm statements.

> >
> > > And I hope the clang folks don't come around and say, err, nope, we're
> > > much more aggressive here.
> >
> > Unlike GCC, I think clang uses the builtin assembler to parse the string,
> > but don't know if it still treats the asms more like black boxes or not.
> > Certainly there is a lot of code in the wild that uses inline asm
> > as optimization barriers, so if it doesn't, then it would cause a lot of
> > problems.
> >
> > Or go with the for (;;);, I don't think any compiler optimizes those away;
> > GCC 10 for C++ can optimize away infinite loops that have some conditional
> > exit because the language guarantees forward progress, but the C language
> > rules are different and for unconditional infinite loops GCC doesn't
> > optimize them away even if explicitly asked to -ffinite-loops.
>
> Lemme add Nick for clang for an opinion:
>
> Nick, we're discussing what would be the cleanest and future-proof
> way to disable stack protector for the function in the kernel which

Oh, this reminds me of commit d0a8d9378d16 ("x86/paravirt: Make
native_save_fl() extern inline"), where the insertion of stack guards
was also causing some pain.

The cleanest solution would be to have function attributes that say
"yes, I know I said -fstack-protector*, but for this one lone function
I really need -fno-stack-protector.  I know what I'm doing and accept
whatever the consequences are."  But maybe the attribute would be
shorter than all that? :P

Compared to playing games with each other's inlining heuristics, that
would be the cleanest and future-proof solution.  (Then we can even
revert d0a8d9378d16, and use such a function attribute.  I somehow
prefer gnu_inline's semantics to ISO C99's extern inline semantics,
and simultaneously hate the problems for which it's used.)

> generates the canary value as gcc10 ends up checking that value due to
> tail-call optimizing the last function called by start_secondary()...
> upthread are all the details.
>
> And question is, can Jakub's suggestions above prevent tail-call
> optimization on clang too and how reliable and future proof would that
> be if we end up going that way?

Sorry, I don't quite follow.  The idea is that an empty asm statement
in foo() should prevent foo() from being inlined into bar()?
https://godbolt.org/z/7xBRGY
-- 
Thanks,
~Nick Desaulniers