Re: [PATCHv11 05/16] x86/uaccess: Provide untagged_addr() and remove tags before address check

["Kirill A. Shutemov" <kirill@shutemov.name>]

On Mon, Nov 07, 2022 at 06:50:51AM -0800, Andy Lutomirski wrote:
> > @@ -21,6 +22,30 @@ static inline bool pagefault_disabled(void);
> >   # define WARN_ON_IN_IRQ()
> >   #endif
> > +#ifdef CONFIG_X86_64
> > +/*
> > + * Mask out tag bits from the address.
> > + *
> > + * Magic with the 'sign' allows to untag userspace pointer without any branches
> > + * while leaving kernel addresses intact.
> > + */
> > +#define untagged_addr(mm, addr)	({					\
> > +	u64 __addr = (__force u64)(addr);				\
> > +	s64 sign = (s64)__addr >> 63;					\
> > +	__addr &= (mm)->context.untag_mask | sign;			\
> > +	(__force __typeof__(addr))__addr;				\
> > +})
> > +
> 
> I think this implementation is correct, but I'm wondering if there are any
> callers of untagged_addr that actually need to preserve kernel addresses.
> Are there?  (There certainly *were* back when we had set_fs().)

I don't think there's any.

CONFIG_KASAN_SW_TAGS uses untagged_addr() on kernel addresses, but it is
only enabled on arm64. On x86, it will use CR4.LAM_SUP and the enabling
would require a new helper for untagging kernel addresses.

That said, I would rather stay on the safe side.

> I'm also mildly uneasy about a potential edge case.  Naively, one would
> expect:
> 
> untagged_addr(current->mm, addr) + size ==
> untagged_addr(current->mm, addr + size)
> 
> at least for an address that is valid enough to be potentially dereferenced.
> This isn't true any more for size that overflows into the tag bit range.

That's definitely a new edge case.

From quick grep, the only CONFIG_KASAN_SW_TAGS code obviously does
arithmetics on address before untagging.

> I *think* we're okay though -- __access_ok requires that addr <= limit -
> size, so any range that overflows into tag bits will be rejected even if the
> entire range consists of valid (tagged) user addresses.

True.

> So:
> 
> Acked-by: Andy Lutomirski <luto@kernel.org>

Thanks!

-- 
  Kiryl Shutsemau / Kirill A. Shutemov