Re: [PATCH 1/2] rcu: Fix casting while dereferencing rcu pointers

[Ricardo Ribalda]

Hi David, Dan and Paul

On Wed, 26 Aug 2026 at 15:22, David Laight <david.laight.linux@xxxxxxxxx> wrote:
>
> On Wed, 26 Aug 2026 13:28:42 +0200
> Ricardo Ribalda <ribalda@xxxxxxxxxxxx> wrote:
>
> > Hi David
> >
> > On Wed, 26 Aug 2026 at 11:55, David Laight <david.laight.linux@xxxxxxxxx> wrote:
> > >
> > > On Tue, 25 Aug 2026 20:02:10 +0000
> > > Ricardo Ribalda <ribalda@xxxxxxxxxxxx> wrote:
> > > >
> > > > diff --git a/include/linux/rcupdate.h b/include/linux/rcupdate.h
> > > > index 44c07a66edff..80276cedda80 100644
> > > > --- a/include/linux/rcupdate.h
> > > > +++ b/include/linux/rcupdate.h
> > > > @@ -488,7 +488,7 @@ static __always_inline bool lockdep_assert_rcu_helper(bool c, const struct __ctx
> > > >  context_unsafe(                                                              \
> > > >       typeof(*p) *local = (typeof(*p) *__force)(p);                   \
> > > >       rcu_check_sparse(p, __rcu);                                     \
> > > > -     ((typeof(*p) __force __kernel *)(local))                        \
> > > > +     ((TYPEOF_UNQUAL(*p) __force __kernel *)(local))                 \
> > >
> > > Ugg... TYPEOF_UNQUAL() is absolutely horrid...
> > > When __CHECKER__ is undefined the whole thing is just (p).
> >
> > About using __CHECKER__: Are you suggesting using #ifdef __CHECKER__
> > to use TYPEOF_UNQUAL or the current code?
> >
> > I am open to doing that, but I would like to hear from Paul if he
> > favors adding the (also horrid) #ifdef to save some CPU cycles on
> > older compilers.
> >
> > >
> > > I suspect reducing all the definitions would measurably improve
> > > kernel compile time (even before these changes).
> >
> > When I checked earlier in the other patch in this patchset, with my
> > compiler I can hardly see any difference in compilation times [1].
> > This is probably because GCC15 direclty maps TYPEOF_UNQUAL to the
> > built-in __typeof_unqual__.
>
> I was thinking of the _Generic() version.
>
> But here you only want to remove the (I think) __user marker.
> I don't think you want to get rid of 'const'.
> So you don't really want gcc to remove any qualifiers.

Yeah, I see your point, removing `const` might hide other issues.

Dan, what would be the best "sparse-approved" way of removing the
address space qualifiers (__user __rcu)  but keeping the other ones
(const) ?

Paul, would you accept an #ifdef __CHECKER__  block to keep Sparse
happy without affecting GCC type checking?

Thanks!

>
> There is an similar(ish) issue when an __force cast is used to
> remove one of the extra qualifiers.
> You really want the compiler to ignore the cast - so it checks
> the actual types match.
>
> David
>
> >
> >
> > Thanks!
> >
> >
> > [1] https://lore.kernel.org/all/CANiDSCsMrULsrGQg76x28GRXLg+HSKFWhhA14Zcn8hdm4KyOVw@xxxxxxxxxxxxxx/
> >
> > >
> > > David
>


-- 
Ricardo Ribalda