Re: [RFC PATCH 1/3] mm/huge_memory: prevent NULL pointer dereference in try_folio_split_to_order()

[David Hildenbrand (Red Hat)]

> > BTW, I wonder if the is_huge_zero_folio() check should go into folio_split_supported() and just return in -EINVAL. (we shouldn't really trigger that). Similarly we could add a hugetlb sanity check.
>
> Yeah, is_huge_zero_folio() should return -EINVAL not -EBUSY, except
> the case the split happens before a process writes 0 to a zero large folio
> and gets a new writable large folio, in which we can kinda say it looks like
> -EBUSY. But it is still a stretch.

I see what you mean, but I think this has less to do with actual races. 
SO yeah, -EINVAL is likely the tight thing.

> Ack on adding hugetlb sanity check.
>
> OK, just to reiterate my above idea on renaming folio_split_supported().
> Are you OK with renaming it to folio_split_check(), so that returning -EBUSY
> and -EINVAL looks more reasonable? The benefit is that we no longer need
> to worry about we need to always do folio->mapping check before
> folio_split_supported(). (In addition, I would rename can_split_folio()
> to folio_split_refcount_check() for clarification)
I guess having some function that tells you "I performed all checks I 
could without taking locks/references (like anon_vma) and starting with 
the real magic" is what you have in mind.

For these we don't have to prefix with "folio_split" if it sounds weird.

folio_check_splittable() ?

Regarding can_split_folio(), I was wondering whether we can just get rid 
of it and use folio_expect_ref_count() instead?

For the two callers that need extra_pins, we could just have something 
simple helper in huge_memory.c like

/* Number of folio references from the pagecache or the swapcache. */
unsigned int folio_cache_references(const struct folio *folio)
{
	if (folio_test_anon(folio) && !folio_test_swapcache(folio))
		return 0;
	return folio_nr_pages(folio);
}


--
Cheers

David