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

[Alexander Gordeev]

On Wed, Jan 22, 2025 at 03:06:05PM +0100, Alexander Gordeev wrote:

Hi Kevin,

> On Wed, Jan 22, 2025 at 08:49:54AM +0100, Heiko Carstens wrote:
> > > >  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?
> 
> Because the follow-up pagetable_pgd_ctor() is called against uninitialized
> page table, while other pagetable_pXd_ctor() variants are called against
> initialized one. I could imagine complications as result of that.
> 
> Whether Region1 table is the right choice is a big question though, as you
> noticed below.

As discussed with Heiko, we do not want to add the extra crst_table_init() call
at least due to performance impact. So please ignore my comment.

> > 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.
> 
> init_new_context() is in a way a constructor as well, so whole thing looks odd
> to me. But I do not immedeately see a better way :(
> 
> > 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()).
> 
> Hmm, that is a good point.
> 
> > 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
> 
> The ctor/dtor mismatch should not be a problem, as pagetable_pgd|p4d|pud_ctor()
> are the same and there is one pagetable_dtor() for all top levels as of now.
> But if it ever comes to separate implementations, then we are in the world
> of pain.
> 
> > 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.
> 
> This is a good point too.

The below bits are seems to be missed. We will test it and send a patch.

diff --git a/arch/s390/mm/pgalloc.c b/arch/s390/mm/pgalloc.c
index a4e761902093..d33f55b7ee98 100644
--- a/arch/s390/mm/pgalloc.c
+++ b/arch/s390/mm/pgalloc.c
@@ -88,12 +88,14 @@ int crst_table_upgrade(struct mm_struct *mm, unsigned long end)
 		if (unlikely(!p4d))
 			goto err_p4d;
 		crst_table_init(p4d, _REGION2_ENTRY_EMPTY);
+		pagetable_p4d_ctor(virt_to_ptdesc(p4d));
 	}
 	if (end > _REGION1_SIZE) {
 		pgd = crst_table_alloc(mm);
 		if (unlikely(!pgd))
 			goto err_pgd;
 		crst_table_init(pgd, _REGION1_ENTRY_EMPTY);
+		pagetable_pgd_ctor(virt_to_ptdesc(pgd));
 	}
 
 	spin_lock_bh(&mm->page_table_lock);
@@ -130,6 +132,7 @@ int crst_table_upgrade(struct mm_struct *mm, unsigned long end)
 	return 0;
 
 err_pgd:
+	pagetable_dtor(virt_to_ptdesc(p4d));
 	crst_table_free(mm, p4d);
 err_p4d:
 	return -ENOMEM;

Thanks!