[PATCH] ARM: spectre-v2: fix smp_processor_id() warning

[Russell King (Oracle)]

On Wed, Jun 22, 2022 at 02:50:15PM +0100, Marc Zyngier wrote:
> On 2022-06-22 07:49, Tetsuo Handa wrote:
> > syzbot complains smp_processor_id() from harden_branch_predictor()
> >  from page fault path [1]. Explicitly disable preemption and use
> > raw_smp_processor_id().
> > 
> > Link: https://syzkaller.appspot.com/bug?extid=a7ee43e564223f195c84 [1]
> > Reported-by: syzbot
> > <syzbot+a7ee43e564223f195c84 at syzkaller.appspotmail.com>
> > Fixes: f5fe12b1eaee220c ("ARM: spectre-v2: harden user aborts in kernel
> > space")
> > Signed-off-by: Tetsuo Handa <penguin-kernel at I-love.SAKURA.ne.jp>
> > ---
> > This patch is completely untested.
> > 
> >  arch/arm/include/asm/system_misc.h | 7 +++++--
> >  1 file changed, 5 insertions(+), 2 deletions(-)
> > 
> > diff --git a/arch/arm/include/asm/system_misc.h
> > b/arch/arm/include/asm/system_misc.h
> > index 98b37340376b..a92446769acd 100644
> > --- a/arch/arm/include/asm/system_misc.h
> > +++ b/arch/arm/include/asm/system_misc.h
> > @@ -20,8 +20,11 @@ typedef void (*harden_branch_predictor_fn_t)(void);
> >  DECLARE_PER_CPU(harden_branch_predictor_fn_t,
> > harden_branch_predictor_fn);
> >  static inline void harden_branch_predictor(void)
> >  {
> > -	harden_branch_predictor_fn_t fn = per_cpu(harden_branch_predictor_fn,
> > -						  smp_processor_id());
> > +	harden_branch_predictor_fn_t fn;
> > +
> > +	preempt_disable_notrace();
> > +	fn = per_cpu(harden_branch_predictor_fn, raw_smp_processor_id());
> > +	preempt_enable_no_resched_notrace();
> >  	if (fn)
> >  		fn();
> >  }
> 
> I don't think that's required.
> 
> harden_branch_predictor() is always called on the fault path,
> from __do_user_fault(), and that's always non-preemptible.
> 
> My hunch is that we're missing some tracking that indicates
> to the kernel that we're already non-preemptible by virtue
> of being in an exception handler.
> 
> Russell, what do you think?

There is one path that we can hit this while we're preemptible - a
page fault (handled via do_page_fault) at an address in kernel space
triggered by user space (e.g. userspace trying to access vmalloc or
module space). 

I suppose, since we know that's never going to be fixed up, we could
detect that the address >= TASK_SIZE and we're entering from usespace
before enabling interrupts, and do the hardening early in that path.
We'd need to move the other cases where we call the hardening as well.

The down side is more tests in the page fault fast-path... so there
will be a performance penalty to be paid just to shut up this warning -
a warning that is "userspace is trying to do real bad stuff" that
basically means userspace is trying to run an exploit...

Which makes me think... having the loud complaint from the kernel there
is actually a good thing, it makes people sit up and notice that
something is wrong.

-- 
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTP is here! 40Mbps down 10Mbps up. Decent connectivity at last!