[PATCH RESEND v7 4/6] arm64: futex: refactor futex atomic operation

[Mark Rutland]

On Tue, Sep 16, 2025 at 02:58:16PM +0100, Yeoreum Yun wrote:
> Hi Mark,
> 
> [...]
> > > I think it's enough to use usafe_get_user() instead of get_user() in here
> > > since when FEAT_LSUI enabled, it doeesn't need to call
> > > uaccess_ttbr0_enable()/disable().
> >
> > Regardless of uaccess_ttbr0_enable() and uaccess_ttbr0_disable()
> > specifically, API-wise unsafe_get_user() is only supposed to be called
> > between user_access_begin() and user_access_end(), and there's some
> > stuff we probably want to add there (e.g. might_fault(), which
> > unsafe_get_user() lacks today).
> >
> > Do we call those?
> 
> Yes when you're available.
> As you mention, the difference seems might_fault(),
> But I'm not sure whether that would be a reason to validate to use
> get_user() instead of unsafe_get_user() taking a increase of instruction
> of "nop" -- uaccess_ttbr0_enable()/disable() in LSUI
> except the reason for DEUBG purpose.

I think the practical impact of those NOPs is going to be neglible, and
not worth optimizing for unless/until we have data demonstrating
otherwise.

If we want to strictly avoid those NOPs, I think that we should do a
more general cleanup, and e.g. have variants of user_access_begin() and
user_access_end() that do not mess with TTBR0. I don't think we need to
do that for this series.

For now, I think that you should either:

* Use get_user().

* Use user_access_begin() .. user_access_end() wrapping both
  unsafe_get_user() and the user cmpxchg.

Thanks,
Mark.