Re: [PATCH 113/190] Revert "x86/hpet: Prevent potential NULL pointer dereference"

[Kees Cook]

On Fri, Apr 23, 2021 at 01:33:07AM +0200, Thomas Gleixner wrote:
> On Wed, Apr 21 2021 at 12:49, Kees Cook wrote:
> > On Wed, Apr 21, 2021 at 02:59:48PM +0200, Greg Kroah-Hartman wrote:
> >> diff --git a/arch/x86/kernel/hpet.c b/arch/x86/kernel/hpet.c
> >> index 08651a4e6aa0..0515a97bf6f5 100644
> >> --- a/arch/x86/kernel/hpet.c
> >> +++ b/arch/x86/kernel/hpet.c
> >> @@ -930,8 +930,6 @@ int __init hpet_enable(void)
> >>  		return 0;
> >>  
> >>  	hpet_set_mapping();
> >> -	if (!hpet_virt_address)
> >> -		return 0;
> >>  
> >>  	/* Validate that the config register is working */
> >>  	if (!hpet_cfg_working())
> >
> > FWIW, this patch looks harmless. It is checking for a failure in
> > hpet_set_mapping(), and avoids the following code from performing
> > 0-offset reads. hpet_set_mapping() is likely to never fail in real-world
> > situations. *shrug*
> 
> 'likely never to fail' is clearly a receipe for disaster and you should
> know that.

Of course -- I prefer to keep the sanity check. It just wasn't as good
as it could have been: it's not clear just by looking at the patch how
hpet_virt_address and hpet_set_mapping() are related.

> 
> > I think it would make more sense for the check to live in
> > hpet_cfg_working(), though.
> 
> No. That does not make any sense at all.
> 
> The proper change would have been to make hpet_set_mapping() return
> an error/success code and act on that.
> 
> But that does _NOT_ make the patch invalid.
> 
> I'm pretty sure that I looked at it and thought about the proper
> solution (see above) and then shrugged it off because of overload...

Right, no, I was saying the original patch should stay. It shouldn't be
reverted.

Greg, please drop this patch from the revert list.

-- 
Kees Cook