David Hildenbrand wrote:
> On 05.06.25 14:09, Jason Gunthorpe wrote:
> > On Wed, Jun 04, 2025 at 07:35:24PM -0700, Dan Williams wrote:
> >
> >> If all dax pages are special, then vm_normal_page() should never find
> >> them and gup should fail.
> >>
> >> ...oh, but vm_normal_page_p[mu]d() is not used in the gup path, and
> >> 'special' is not set in the pte path.
> >
> > That seems really suboptimal?? Why would pmd and pte be different?
> >
> >> I think for any p[mu]d where p[mu]d_page() is ok to use should never set
> >> 'special', right?
> >
> > There should be dedicated functions for installing pages and PFNs,
> > only the PFN one would set the special bit.
> >
> > And certainly your tests *should* be failing as special entries should
> > never ever be converted to struct page.
>
> Worth reviewing [1] where I clean that up and describe the current
> impact. ;)
Will do.
> What's even worse about this pte_devmap()/pmd_devmap()/... shit (sorry!
> but it's absolute shit) is that some pte_mkdev() set the pte special,
> while others ... don't.
As the person who started the turd rolling into this pile that Alistair
is heroically cleaning up, I approve this characterization.
> E.g., loongarch
>
> static inline pte_t pte_mkdevmap(pte_t pte) { pte_val(pte) |=
> _PAGE_DEVMAP; return pte; }
>
> I don't even know how it can (could) survive vm_normal_page().
Presently "can" because dax switched away from vmf_insert_mixed() to
vmf_insert_page(), "could" in the past was the devmap hack to avoid
treating VM_MIXEDMAP as !vm_normal_page().
