Re: [PATCH] arch: x86: power: cpu: init %gs before __restore_processor_state (clang)

[Nick Desaulniers]

On Tue, Sep 15, 2020 at 10:46 AM Borislav Petkov <bp@xxxxxxxxx> wrote:
>
> On Tue, Sep 15, 2020 at 10:26:58AM -0700, rkir@xxxxxxxxxx wrote:
> > From: Haitao Shan <hshan@xxxxxxxxxx>
> >
> > This is a workaround which fixes triple fault
> > in __restore_processor_state on clang when
> > built with LTO.
> >
> > When load_TR_desc and load_mm_ldt are inlined into
> > fix_processor_context due to LTO, they cause
> > fix_processor_context (or in this case __restore_processor_state,
> > as fix_processor_context was inlined into __restore_processor_state)
> > to access the stack canary through %gs, but before
> > __restore_processor_state has restored the previous value
> > of %gs properly. LLVM appears to be inlining functions with stack
> > protectors into functions compiled with -fno-stack-protector,
> > which is likely a bug in LLVM's inliner that needs to be fixed.
> >
> > The LLVM bug is here: https://bugs.llvm.org/show_bug.cgi?id=47479
> >
> > Signed-off-by: Haitao Shan <hshan@xxxxxxxxxx>
> > Signed-off-by: Roman Kiryanov <rkir@xxxxxxxxxx>
>
> Ok, google guys, pls make sure you Cc LKML too as this is where *all*
> patches and discussions are archived. Adding it now to Cc.

Roman, please use ./scripts/get_maintainer.pl (in the kernel tree) for that.

>
> > ---
> >  arch/x86/power/cpu.c | 10 ++++++++++
> >  1 file changed, 10 insertions(+)
> >
> > diff --git a/arch/x86/power/cpu.c b/arch/x86/power/cpu.c
> > index db1378c6ff26..e5677adb2d28 100644
> > --- a/arch/x86/power/cpu.c
> > +++ b/arch/x86/power/cpu.c
> > @@ -274,6 +274,16 @@ static void notrace __restore_processor_state(struct saved_context *ctxt)
> >  /* Needed by apm.c */
> >  void notrace restore_processor_state(void)
> >  {
> > +#ifdef __clang__

Should be CONFIG_CC_IS_CLANG; is more canonical throughout the tree.
Or if this is only a bug when doing builds with LTO, and LTO is not
yet upstream, then maybe Sami should carry this in his series, at
least until I can fix the bug in Clang.  Or guard this with the
CONFIG_LTO_CLANG config (not upstream yet; see Sami's series).

> > +     // The following code snippet is copied from __restore_processor_state.
> > +     // Its purpose is to prepare GS segment before the function is called.
> > +#ifdef CONFIG_X86_64
> > +     wrmsrl(MSR_GS_BASE, saved_context.kernelmode_gs_base);
> > +#else
> > +     loadsegment(fs, __KERNEL_PERCPU);
> > +     loadsegment(gs, __KERNEL_STACK_CANARY);
> > +#endif
> > +#endif
>
> Ok, so why is the kernel supposed to take yet another ugly workaround
> because there's a bug in the compiler?

This is exactly the same code from __restore_processor_state.  If it's
ugly, talk to the author of 7ee18d677989e. ;)  All this patch is doing
is moving this up a call frame (though now this is effectively being
run twice).

> If it is too late to fix it there, then maybe disable LTO builds for the
> buggy version only.

We could do that, too.  (We can disable LTO on a per translation unit
basis in KBuild).  Note the author of the bug report linked above. :^P
"Revenge of the stack protector"

>
> We had a similar discussion this week and we already have one buggy
> compiler to deal with and this second one is not making it any easier...
-- 
Thanks,
~Nick Desaulniers