Re: [Question PATCH kernel] x86/amd/sev/nmi+vc: Fix stack handling (why is this happening?)

[Peter Zijlstra]

On Fri, Jan 27, 2023 at 11:13:38PM +1100, Alexey Kardashevskiy wrote:

> > This is broken, and building with DEBUG_ENTRY=y would've told you.
> 
> 
> Huh, good to know. Is this it telling me so?
> 
> vmlinux.o: warning: objtool: exc_nmi+0x73: call to native_get_debugreg7()
> leaves .noinstr.text section
> 

Yep. The ramification of all that is that by calling non-noinstr code
(double negative, iow, regular instrumented code) is that you can end up
in the tracers/*SAN/breakpoints etc.. code -- something we're very much
not ready for at this point.

> > > +
> > >   #ifdef CONFIG_CPU_SUP_AMD
> > >   extern void set_dr_addr_mask(unsigned long mask, int dr);
> > >   #else
> > > diff --git a/arch/x86/kernel/nmi.c b/arch/x86/kernel/nmi.c
> > > index cec0bfa3bc04..400b5b6b74f6 100644
> > > --- a/arch/x86/kernel/nmi.c
> > > +++ b/arch/x86/kernel/nmi.c
> > > @@ -503,7 +503,7 @@ DEFINE_IDTENTRY_RAW(exc_nmi)
> > >   	 */
> > >   	sev_es_ist_enter(regs);
> > > -	this_cpu_write(nmi_dr7, local_db_save());
> > > +	this_cpu_write(nmi_dr7, local_db_save_exc_nmi());
> > >   	irq_state = irqentry_nmi_enter(regs);
> > 
> > So what I don't get is why sev_es_ist_enter() doesn't cause us to make a
> > stack frame, that has an actual call in it (although admittedly
> > conditional).
> 
> Is not the frame gone when sev_es_ist_enter() (which does not get inlined as
> per objdump's "ffffffff81bd4550 <__sev_es_ist_enter>:
> ") returned?

Well, returning would consume the callframe, but the stack setup of the
caller should remain. Let me go stare at some asm.