Re: [PATCHv3 5/8] x86/uaccess: Provide untagged_addr() and remove tags before address check

["Edgecombe, Rick P" <rick.p.edgecombe@intel.com>]

On Wed, 2022-06-15 at 19:58 +0300, Kirill A. Shutemov wrote:
> On Mon, Jun 13, 2022 at 05:36:43PM +0000, Edgecombe, Rick P wrote:
> > On Fri, 2022-06-10 at 17:35 +0300, Kirill A. Shutemov wrote:
> > > +#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.
> > 
> > Trying to understand the magic part here. I guess how it works is,
> > when
> > the high bit is set, it does the opposite of untagging the
> > addresses by
> > setting the tag bits instead of clearing them. So:
> >   - For proper canonical kernel addresses (with U57) it leaves
> > them 
> >     intact since the tag bits were already set.
> >   - For non-canonical kernel-half addresses, it fixes them up. 
> >     (0xeffffff000000840->0xfffffff000000840)
> >   - For U48 and 5 level paging, it corrupts some normal kernel 
> >     addresses. (0xff90ffffffffffff->0xffffffffffffffff)
> > 
> > I just ported this to userspace and threw some addresses at it to
> > see
> > what happened, so hopefully I got that right.
> 
> Ouch. Thanks for noticing this. I should have catched this myself.
> Yes,
> this implementation is broken for LAM_U48 on 5-level machine.
> 
> What about this:
> 
>         #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;                               \
>         })
> 
> It makes mask effectively. all-ones for supervisor addresses. And it
> is
> less magic to my eyes.

Yea, it seems to leave kernel half addresses alone now, including
leaving non-canonical addresses as non-canonical and 5 level addresses.

With the new bit math:
Reviewed-by: Rick Edgecombe <rick.p.edgecombe@intel.com>

> 
> The generated code also look sane to me:
> 
>     11d0:       48 89 f8                mov    %rdi,%rax
>     11d3:       48 c1 f8 3f             sar    $0x3f,%rax
>     11d7:       48 0b 05 52 2e 00 00    or    
> 0x2e52(%rip),%rax        # 4030 <untag_mask>
>     11de:       48 21 f8                and    %rdi,%rax
> 
> Any comments?
> 
> > Is this special kernel address handling only needed because
> > copy_to_kernel_nofault(), etc call the user helpers?
> 
> I did not have any particular use-case in mind. But just if some
> kernel
> address gets there and bits get cleared we will have very hard to
> debug
> bug.

I just was thinking if we could rearrange the code to avoid untagging
kernel addresses, we could skip this, or even VM_WARN_ON() if we see
one. Seems ok either way.