Re: [PATCH v2 6/6] mm: Introduce ctor/dtor at PGD level

[Heiko Carstens <hca@linux.ibm.com>]

On Tue, Jan 21, 2025 at 05:37:33PM +0100, Alexander Gordeev wrote:
> On Fri, Jan 03, 2025 at 06:44:15PM +0000, Kevin Brodsky wrote:
> 
> Hi Kevin,
> ...
> > diff --git a/arch/s390/include/asm/pgalloc.h b/arch/s390/include/asm/pgalloc.h
> > index 5fced6d3c36b..b19b6ed2ab53 100644
> > --- a/arch/s390/include/asm/pgalloc.h
> > +++ b/arch/s390/include/asm/pgalloc.h
> > @@ -130,11 +130,18 @@ static inline void pud_populate(struct mm_struct *mm, pud_t *pud, pmd_t *pmd)
> >  
> >  static inline pgd_t *pgd_alloc(struct mm_struct *mm)
> >  {
> > -	return (pgd_t *) crst_table_alloc(mm);
> > +	unsigned long *table = crst_table_alloc(mm);
> > +
> > +	if (!table)
> > +		return NULL;
> 
> I do not know status of this series, but FWIW, this call is missed:
> 
> 	crst_table_init(table, _REGION1_ENTRY_EMPTY); 

Why is that missing?

A pgd table can be a Region1, Region2, or Region3 table. The only caller of
this function is mm_init() via mm_alloc_pgd(); and right after mm_alloc_pgd()
there is a call to init_new_context() which will initialize the pgd correctly.

I guess what really gets odd, and might be broken (haven't checked yet) is
what happens on dynamic upgrade of page table levels (->crst_table_upgrade()).

With that a pgd may become a pud, and with that we get an imbalance with
the ctor/dtor calls for the various page table levels when they get freed
again. Plus, at first glance, it looks also broken that we have open-coded
crst_alloc() calls instead of using the "proper" page table allocation API
within crst_table_upgrade(), which again would cause an imbalance.