[PATCH] arm64: uaccess: fix put_user() with TTBR0 PAN

[Mark Rutland]

On Fri, Nov 19, 2021 at 11:33:06AM +0000, Will Deacon wrote:
> On Thu, Nov 18, 2021 at 05:34:17PM +0100, Vincent Whitchurch wrote:
> > The value argument to put_user() must be evaluated before the TTBR0
> > switch is done.  Otherwise, if it is a function and the function sleeps,
> > the reserved TTBR0 will be restored when the process is switched in
> > again and the process will end up in an infinite loop of faults.
> > 
> > This problem was seen with the put_user() in schedule_tail().  A similar
> > fix was done for RISC-V in commit 285a76bb2cf51b0c74c634 ("riscv:
> > evaluate put_user() arg before enabling user access").
> > 
> > Fixes: f253d827f33cb5a5990 ("arm64: uaccess: refactor __{get,put}_user")
> > Signed-off-by: Vincent Whitchurch <vincent.whitchurch at axis.com>
> > ---
> >  arch/arm64/include/asm/uaccess.h | 3 ++-
> >  1 file changed, 2 insertions(+), 1 deletion(-)
> > 
> > diff --git a/arch/arm64/include/asm/uaccess.h b/arch/arm64/include/asm/uaccess.h
> > index 6e2e0b7031ab..96b26fa9d3d0 100644
> > --- a/arch/arm64/include/asm/uaccess.h
> > +++ b/arch/arm64/include/asm/uaccess.h
> > @@ -362,10 +362,11 @@ do {									\
> >  #define __put_user_error(x, ptr, err)					\
> >  do {									\
> >  	__typeof__(*(ptr)) __user *__p = (ptr);				\
> > +	__typeof__(*(__p)) __val = (x);					\
> >  	might_fault();							\
> >  	if (access_ok(__p, sizeof(*__p))) {				\
> >  		__p = uaccess_mask_ptr(__p);				\
> > -		__raw_put_user((x), __p, (err));			\
> > +		__raw_put_user(__val, __p, (err));			\
> >  	} else	{							\
> >  		(err) = -EFAULT;					\
> >  	}								\
> 
> 
> Oh, nice spot! I hope you didn't lose too much time debugging if you
> actually ran into this...
> 
> Although it seems a lot less likely to cause a problem, should we do
> something similar for __get_user_error() and assign to (x) outside of
> the uaccess-disabled section?

I agree we should follow up with a more general cleanup to avoid any
macro evaluation within user-access or tco critical sections. Since
that's especially subtle for the get_*() helpers (and I beleive there
may be some other latent issues in that area), I reckon we should do
that as a follow-up, and shouldn't block this patch on that being done.

I'll go audit that and see what I spot.

Thanks,
Mark.