Re: [PATCH v7 22/41] mm/mmap: Add shadow stack pages to memory accounting

[Edgecombe, Rick P]

On Fri, 2023-03-17 at 10:28 -0700, Deepak Gupta wrote:
> On Fri, Mar 17, 2023 at 10:16 AM Dave Hansen <dave.hansen@xxxxxxxxx>
> wrote:
> > 
> > On 3/17/23 10:12, Deepak Gupta wrote:
> > > >   /*
> > > > - * Stack area - automatically grows in one direction
> > > > + * Stack area
> > > >    *
> > > > - * VM_GROWSUP / VM_GROWSDOWN VMAs are always private
> > > > anonymous:
> > > > - * do_mmap() forbids all other combinations.
> > > > + * VM_GROWSUP, VM_GROWSDOWN VMAs are always private
> > > > + * anonymous. do_mmap() forbids all other combinations.
> > > >    */
> > > >   static inline bool is_stack_mapping(vm_flags_t flags)
> > > >   {
> > > > -       return (flags & VM_STACK) == VM_STACK;
> > > > +       return ((flags & VM_STACK) == VM_STACK) || (flags &
> > > > VM_SHADOW_STACK);
> > > 
> > > Same comment here. `VM_SHADOW_STACK` is an x86 specific way of
> > > encoding a shadow stack.
> > > Instead let's have a proxy here which allows architectures to
> > > have
> > > their own encodings to represent a shadow stack.
> > 
> > This doesn't _preclude_ another architecture from coming along and
> > doing
> > that, right?  I'd just prefer that shadow stack architecture #2
> > comes
> > along and refactors this in precisely the way _they_ need it.
> 
> There are two issues here
>  - Encoding of shadow stack: Another arch can choose different
> encoding.
>    And yes, another architecture can come in and re-factor it. But so
> much thought and work has been given to x86 implementation to keep
>    shadow stack to not impact arch agnostic parts of the kernel. So
> why creep it in here.
> 
> - VM_SHADOW_STACK is coming out of the VM_HIGH_ARCH_XX bit position
> which makes it arch specific.
> 
> 

VM_SHADOW_STACK is defined like this (trimmed for clarity):
#ifdef CONFIG_X86_USER_SHADOW_STACK
# define VM_SHADOW_STACK	VM_HIGH_ARCH_5
#else
# define VM_SHADOW_STACK	VM_NONE
#endif

Also, we actually had an is_shadow_stack_mapping(vma) in the past, but
it was dropped from other feedback. I think it might be too soon to say
whether other implementations won't end up with a similar vma flag, so
this would be premature refactoring. If not though, a helper like that
seems like a reasonable solution.