[PATCH v3] mm: huge-vmap: fail gracefully on unexpected huge vmap mappings

[mark.rutland@arm.com (Mark Rutland)]

On Thu, Jun 08, 2017 at 01:59:46PM +0100, Mark Rutland wrote:
> On Thu, Jun 08, 2017 at 11:35:48AM +0000, Ard Biesheuvel wrote:
> > @@ -287,10 +288,10 @@ struct page *vmalloc_to_page(const void *vmalloc_addr)
> >  	if (p4d_none(*p4d))
> >  		return NULL;
> >  	pud = pud_offset(p4d, addr);
> > -	if (pud_none(*pud))
> > +	if (pud_none(*pud) || WARN_ON_ONCE(pud_huge(*pud)))
> >  		return NULL;
> >  	pmd = pmd_offset(pud, addr);
> > -	if (pmd_none(*pmd))
> > +	if (pmd_none(*pmd) || WARN_ON_ONCE(pmd_huge(*pmd)))
> >  		return NULL;
> 
> I think it might be better to use p*d_bad() here, since that doesn't
> depend on CONFIG_HUGETLB_PAGE.
> 
> While the cross-arch semantics are a little fuzzy, my understanding is
> those should return true if an entry is not a pointer to a next level of
> table (so pXd_huge(p) implies pXd_bad(p)).

Ugh; it turns out this isn't universally true.

I see that at least arch/hexagon's pmd_bad() always returns 0, and they
support CONFIG_HUGETLB_PAGE.

So I guess there isn't an arch-neutral, always-available way of checking
this. Sorry for having mislead you.

For arm64, p*d_bad() would still be preferable, so maybe we should check
both?

Thanks,
Mark.