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

[Peter Zijlstra]

On Wed, Jul 03, 2019 at 04:29:42PM -0400, Steven Rostedt wrote:
> On Wed, 3 Jul 2019 22:22:31 +0200
> Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> 
> > On Wed, Jul 03, 2019 at 12:27:34PM +0200, root wrote:
> > > Despire the current efforts to read CR2 before tracing happens there
> > > still exist a number of possible holes:
> > > 
> > >   idtentry page_fault             do_page_fault           has_error_code=1
> > >     call error_entry
> > >       TRACE_IRQS_OFF
> > >         call trace_hardirqs_off*
> > >           #PF // modifies CR2
> > > 
> > >       CALL_enter_from_user_mode
> > >         __context_tracking_exit()
> > >           trace_user_exit(0)
> > >             #PF // modifies CR2
> > > 
> > >     call do_page_fault
> > >       address = read_cr2(); /* whoopsie */
> > > 
> > > And similar for i386.
> > > 
> > > Fix it by pulling the CR2 read into the entry code, before any of that
> > > stuff gets a chance to run and ruin things.
> > > 
> > > Ideally we'll clean up the entry code by moving this tracing and
> > > context tracking nonsense into C some day, but let's not delay fixing
> > > this longer.
> > >   
> > 
> > > @@ -1180,10 +1189,10 @@ idtentry xenint3		do_int3			has_error_co
> > >  #endif
> > >  
> > >  idtentry general_protection	do_general_protection	has_error_code=1
> > > -idtentry page_fault		do_page_fault		has_error_code=1
> > > +idtentry page_fault		do_page_fault		has_error_code=1	read_cr2=1
> > >  
> > >  #ifdef CONFIG_KVM_GUEST
> > > -idtentry async_page_fault	do_async_page_fault	has_error_code=1
> > > +idtentry async_page_fault	do_async_page_fault	has_error_code=1	read_cr2=1
> > >  #endif  
> > 
> > While going over the various idt handlers, I found that we probably also
> > need read_cr2 on do_double_fault(), otherwise it is susceptible to the
> > same problem.
> > 
> 
> BTW, do you plan on making this for stable? Even though it's rather
> invasive. Or should we just apply the band-aids first, have them
> backported to stable, and then put this change on top of them for
> upstream?

So I don't particularly care about stable; and the band-aids
(trace_irqs_off_cr2) is known broken so I really don't see the point.

That said, these patches should apply to most recent kernels (post PTI)
without too much rejects.