[PATCH] x86/uaccess: Avoid barrier_nospec() in copy_from_user()

[Kirill A. Shutemov]

On Sat, Oct 12, 2024 at 05:53:19PM -0700, Linus Torvalds wrote:
> On Sat, 12 Oct 2024 at 10:44, Linus Torvalds
> <torvalds at linux-foundation.org> wrote:
> >
> > Anyway, what's the speculation window size like?
> 
> Note that this is important basically because we do *NOT* want to
> check the address against TASK_SIZE_MAX like we used to, because not
> only is TASK_SIZE_MAX not a compile-time constant, but with linear
> address masking, people actually *want* to use addresses that are in
> the non-canonical range.
> 
> IOW, see also
> 
>     arch/x86/include/asm/uaccess_64.h
> 
> and notice how the x86-64 __access_ok() check *also_ does the whole
> "top bit set" thing (iow, see __access_ok()).
> 
> IOW, this actually goes even further back than the commit I mentioned
> earlier - it goes back to commit 6014bc27561f ("x86-64: make
> access_ok() independent of LAM") because without the sign bit trick,
> LAM is a complete disaster.
> 
> So no, the address masking can not depend on things like
> __VIRTUAL_MASK_SHIFT, it would need to at least take LAM into account
> too. Not that I know if there are any CPU's out there that actually
> have LAM enabled.

Actually LAM is fine with the __VIRTUAL_MASK_SHIFT check. LAM enforces bit
47 (or 56 for 5-level paging) to be equal to bit 63. Otherwise it is
canonicality violation.

-- 
  Kiryl Shutsemau / Kirill A. Shutemov