Re: [PATCH RFC 07/14] fs/dax: use vmf_insert_folio_pmd() to insert the huge zero folio

[David Hildenbrand]

On 25.06.25 11:03, David Hildenbrand wrote:
> On 24.06.25 03:16, Alistair Popple wrote:
> > On Tue, Jun 17, 2025 at 05:43:38PM +0200, David Hildenbrand wrote:
> > > Let's convert to vmf_insert_folio_pmd().
> > >
> > > In the unlikely case there is already something mapped, we'll now still
> > > call trace_dax_pmd_load_hole() and return VM_FAULT_NOPAGE.
> > >
> > > That should probably be fine, no need to add special cases for that.
> >
> > I'm not sure about that. Consider dax_iomap_pmd_fault() -> dax_fault_iter() ->
> > dax_pmd_load_hole(). It calls split_huge_pmd() in response to VM_FAULT_FALLBACK
> > which will no longer happen, what makes that ok?
>
> My reasoning was that this is the exact same behavior other
> vmf_insert_folio_pmd() users here would result in.
>
> But let me dig into the details.

Okay, trying to figure out what to do here.

Assume dax_pmd_load_hole() is called and there is already something. We 
would have returned VM_FAULT_FALLBACK, now we would return VM_FAULT_NO_PAGE.

That obviously only happens when we have not a write fault (otherwise, 
the shared zeropage does not apply).

In dax_iomap_pmd_fault(), we would indeed split_huge_pmd(). In the DAX 
case (!anon vma), that would simply zap whatever is already mapped there.

I guess we would then return VM_FAULT_FALLBACK from huge_fault-> ... -> 
dax_iomap_fault() and core MM code would fallback to handle_pte_fault() 
etc. and ... load a single PTE mapping the shared zeropage.

BUT

why is this case handled differently than everything else?

E.g.,

(1) when we try inserting the shared zeropage through 
dax_load_hole()->vmf_insert_page_mkwrite() and there is already 
something ... we return VM_FAULT_NOPAGE.

(2) when we try inserting a PTE mapping an ordinary folio through 
dax_fault_iter()->vmf_insert_page_mkwrite() and there is already 
something ... we return VM_FAULT_NOPAGE.

(3) when we try inserting a PMD mapping an ordinary folio through 
dax_fault_iter()->vmf_insert_folio_pmd() and there is already something 
... we return VM_FAULT_NOPAGE.


So that makes me think ... the VM_FAULT_FALLBACK right now is probably 
... wrong? And probably cannot be triggered?

If there is already the huge zerofolio mapped, all good.

Anything else is really not expected I would assume?

--
Cheers,

David / dhildenb