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

[Michael Matz]

Hello,

On Fri, 17 Apr 2020, Nick Desaulniers wrote:

> Ah seems we do have __attribute__((no_selector))
> (https://reviews.llvm.org/D46300,
> https://releases.llvm.org/7.0.0/tools/clang/docs/AttributeReference.html#no-stack-protector-clang-no-stack-protector-clang-no-stack-protector)
> which differs from GCC attribute name.

As you will discover upthread that was tried with GCC and found 
insufficient, as GCC is a bit surprising with optimize attributes: it 
resets every -f option from the command line and applies only the ones 
from the attributes.  Including a potential -fno-omit-frame-pointer, 
causing all kinds of itches :)

(The similar attribute in clang might work less surprising of course).


Ciao,
Michael.

> 
> I'm still catching up on the thread (and my cat is insistent about
> sleeping on my lap while I'm trying to use my laptop), but I like
> 20200417190607.GY2424@tucnak/T/#m23d197d3a66a6c7d04c5444af4f51d940895b412">https://lore.kernel.org/lkml/20200417190607.GY2424@tucnak/T/#m23d197d3a66a6c7d04c5444af4f51d940895b412
> if it additionally defined __no_stack_protector for compiler-clang.h.
> 
> On Fri, Apr 17, 2020 at 12:06 PM Jakub Jelinek <jakub@xxxxxxxxxx> wrote:
> >
> > On Fri, Apr 17, 2020 at 11:22:25AM -0700, Nick Desaulniers wrote:
> > > > Sorry, I don't quite follow.  The idea is that an empty asm statement
> > > > in foo() should prevent foo() from being inlined into bar()?
> > >
> > > s/inlined/tail called/
> >
> > Yeah.  The thing is, the caller changes the stack protector guard base
> > value, so at the start of the function it saves a different value then
> > it compares at the end.  But, the function that it calls at the end
> > actually doesn't return, so this isn't a problem.
> > If it is tail called though, the stack protector guard checking is done
> > before the tail call and it crashes.
> > If the called function is marked with noreturn attribute or _Noreturn,
> > at least GCC will also not tail call it and all is fine, but not sure
> > what LLVM does in that case.
> 
> Seems fine? https://godbolt.org/z/VEoEfw
> (try commenting out the __attribute__((noreturn)) to observe the tail calls.
>