Re: [PATCH v3 1/7] mm: Add folio_split_unref helper

[Matthew Brost <matthew.brost@intel.com>]

On Fri, Jan 09, 2026 at 04:11:23PM +0100, David Hildenbrand (Red Hat) wrote:
> On 1/9/26 15:30, Zi Yan wrote:
> > On 9 Jan 2026, at 8:26, David Hildenbrand (Red Hat) wrote:
> > 
> > > On 1/9/26 14:19, David Hildenbrand (Red Hat) wrote:
> > > > On 1/9/26 09:54, Francois Dugast wrote:
> > > > > From: Matthew Brost <matthew.brost@intel.com>
> > > > > 
> > > > > Add folio_split_unref helper which splits an unreferenced folio
> > > > 
> > > > split_unref reads like "split and unref".
> > > > 
> > > > You probably want to call this something like "folio_split_frozen" ?
> > > > 
> > > > The very definition of "frozen" is "refcount = 0 ", so you can simplify
> > > > the documentation.
> > > > 
> > > > Are the folios you want to pass in there completely unused (-> free) or
> > > > might they still be in use (e.g., migration entries point at them during
> > > > folio split)
> > > > 
> > > > So I am not sure yet if this should be "folio_split_frozen()" or
> > > > "folio_split_freed()" or sth like that.
> > > > 
> > > > I'm not CCed on the other patches in the series or the cover letter, so
> > > > I don't see the context.

Here is a patchwork link to the entire series:
https://patchwork.freedesktop.org/series/159119/

> > > > 
> > > 
> > > Ah, I was CCed on #3 where we call this function on folios that are getting freed.
> > > 
> > > In that case it would be acceptable to initialize folio->mapping (and folio->index?) of the split folios. Do we also have to initialize folio->flags, folio->private etc?
> > >

I lifted this code from FSDAX here:
https://elixir.bootlin.com/linux/v6.18.4/source/fs/dax.c#L394

It seemly does everything we need for a zone_device split.

> > > See __split_huge_page_tail().
> > > 

I'm not seeing this function defined anywhere.

> > > folio_split_freed() would likely be best, because then it is clearer that there is absolutely no state to copy from the large folio.
> > 
> > Yes, basically, we do not have a reverse function of prep_compound_page() and
> > open codes the reverse process in free_pages_prepare(). For zone devices,
> > zone_device_page_init() calls prep_compound_page() to form a folio but
> > free_zone_device_folio() never does the reverse. FS DAX has its own
> > dax_folio_put() to do it. Alistair suggested to come up with a helper
> > function for both FS DAX and free_zone_device_folio().
> > 
> > Maybe free_zone_device_folio_prepare() is better. And put it in mm/memremap.c.

+1 can rename and move.

> 
> That would be even better, if we can limit this completely to zone_device.

Yes, I believe this function should only used by device private, device
coherent, and fsdax folios which are all zone_device. I can add warning
in this function too if it is called on non-zone_device folio too.

Matt 

> -- 
> Cheers
> 
> David