Re: [PATCH 3/3] x86/mm, tracing: Fix CR2 corruption

[Peter Zijlstra]

On Thu, Jul 04, 2019 at 12:00:57AM +0200, Peter Zijlstra wrote:
> On Wed, Jul 03, 2019 at 01:27:09PM -0700, Andy Lutomirski wrote:
> > On Wed, Jul 3, 2019 at 3:28 AM root <peterz@xxxxxxxxxxxxx> wrote:
> 
> > > @@ -1338,18 +1347,9 @@ ENTRY(error_entry)
> > >         movq    %rax, %rsp                      /* switch stack */
> > >         ENCODE_FRAME_POINTER
> > >         pushq   %r12
> > > -
> > > -       /*
> > > -        * We need to tell lockdep that IRQs are off.  We can't do this until
> > > -        * we fix gsbase, and we should do it before enter_from_user_mode
> > > -        * (which can take locks).
> > > -        */
> > > -       TRACE_IRQS_OFF
> > 
> > This hunk looks wrong.  Am I missing some other place that handles the
> > case where we enter from kernel mode and IRQs were on?
> 
> > > -	CALL_enter_from_user_mode
> > >  	ret
> > >  
> > >  .Lerror_entry_done:
> > > -	TRACE_IRQS_OFF
> > >  	ret
> > >  
> > >  	/*
> 
> Did you perchance mean to complain about the .Lerror_entry_done one?
> 
> Because I'm not seeing how the one before CALL_enter_from_user_mode can
> ever be from-kernel.
> 
> But yes, that .Lerror_entry_done one looks fishy.

That works because idtentry will now do an unconditional TRACE_IRQS_OFF,
note how the patch removes the .if \paranoid around that.