[PATCH v3] arm64/mm: avoid fixmap race condition when create pud mapping

[Catalin Marinas]

On Thu, Jan 27, 2022 at 01:22:47PM +0100, David Hildenbrand wrote:
> > Yes, system_state can roughly separate these callers of __create_pgd_mapping. When system_state > SYSTEM_BOOTING we can add the lock.
> > Thus, I have the following change:
> > 
> > static DEFINE_SPINLOCK(swapper_pgdir_lock);
> > +static DEFINE_MUTEX(fixmap_lock);
> > 
> >  void set_swapper_pgd(pgd_t *pgdp, pgd_t pgd)
> >  {
> > @@ -329,6 +330,8 @@ static void alloc_init_pud(pgd_t *pgdp, unsigned long addr, unsigned long end,
> >         }
> >         BUG_ON(p4d_bad(p4d));
> > 
> > +       if (system_state > SYSTEM_BOOTING)
> 
> As there is nothing smaller than SYSTEM_BOOTING, you can use
> 	if (system_state != SYSTEM_BOOTING)
> 
> ...
> 
> > 
> > It seems work and somehow simper. But I don't know if it is reasonable to do this. So, any idea? @Ard Biesheuvel  @Catalin Marinas 
> 
> It's worth looking at kernel/notifier.c, e.g.,
> blocking_notifier_chain_register()
> 
> if (unlikely(system_state == SYSTEM_BOOTING))
> 	return notifier_chain_register(&nh->head, n);
> 
> down_write(&nh->rwsem);
> ret = notifier_chain_register(&nh->head, n);
> up_write(&nh->rwsem);
> 
> If we decide to go down that path, we should make sure to add a comment like
> 
> /*
>  * No need for locking during early boot. And it doesn't work as
>  * expected with KASLR enabled where we might clear BSS twice.
>  */

A similar approach sounds fine to me.

-- 
Catalin