Re: [PATCH 0/5] add STATIC_PMD_ZERO_PAGE config option

[David Hildenbrand]

On 13.06.25 10:58, Pankaj Raghav (Samsung) wrote:
> On Thu, Jun 12, 2025 at 02:46:34PM -0700, Dave Hansen wrote:
> > On 6/12/25 13:36, Pankaj Raghav (Samsung) wrote:
> > > On Thu, Jun 12, 2025 at 06:50:07AM -0700, Dave Hansen wrote:
> > > > On 6/12/25 03:50, Pankaj Raghav wrote:
> > > > > But to use huge_zero_folio, we need to pass a mm struct and the
> > > > > put_folio needs to be called in the destructor. This makes sense for
> > > > > systems that have memory constraints but for bigger servers, it does not
> > > > > matter if the PMD size is reasonable (like in x86).
> > > >
> > > > So, what's the problem with calling a destructor?
> > > >
> > > > In your last patch, surely bio_add_folio() can put the page/folio when
> > > > it's done. Is the real problem that you don't want to call zero page
> > > > specific code at bio teardown?
> > >
> > > Yeah, it feels like a lot of code on the caller just to use a zero page.
> > > It would be nice just to have a call similar to ZERO_PAGE() in these
> > > subsystems where we can have guarantee of getting huge zero page.
> > >
> > > Apart from that, these are the following problems if we use
> > > mm_get_huge_zero_folio() at the moment:
> > >
> > > - We might end up allocating 512MB PMD on ARM systems with 64k base page
> > >    size, which is undesirable. With the patch series posted, we will only
> > >    enable the static huge page for sane architectures and page sizes.
> >
> > Does *anybody* want the 512MB huge zero page? Maybe it should be an
> > opt-in at runtime or something.
> Yeah, I think that needs to be fixed. David also pointed this out in one
> of his earlier reviews[1].
>
> > > - In the current implementation we always call mm_put_huge_zero_folio()
> > >    in __mmput()[1]. I am not sure if model will work for all subsystems. For
> > >    example bio completions can be async, i.e, we might need a reference
> > >    to the zero page even if the process is no longer alive.
> >
> > The mm is a nice convenient place to stick an mm but there are other
> > ways to keep an efficient refcount around. For instance, you could just
> > bump a per-cpu refcount and then have the shrinker sum up all the
> > refcounts to see if there are any outstanding on the system as a whole.
> >
> > I understand that the current refcounts are tied to an mm, but you could
> > either replace the mm-specific ones or add something in parallel for
> > when there's no mm.
>
> But the whole idea of allocating a static PMD page for sane
> architectures like x86 started with the intent of avoiding the refcounts and
> shrinker.
>
> This was the initial feedback I got[2]:
>
> I mean, the whole thing about dynamically allocating/freeing it was for
> memory-constrained systems. For large systems, we just don't care.

For non-mm usage we can just use the folio refcount. The per-mm 
refcounts are all combined into a single folio refcount. The way the 
global variable is managed based on per-mm refcounts is the weird thing.

In some corner cases we might end up having multiple instances of huge 
zero folios right now. Just imagine:

1) Allocate huge zero folio during read fault
2) vmsplice() it
3) Unmap the huge zero folio
4) Shrinker runs and frees it
5) Repeat with 1)

As long as the folio is vmspliced(), it will not get actually freed ...

I would hope that we could remove the shrinker completely, and simply 
never free the huge zero folio once allocated. Or at least, only free it 
once it is actually no longer used.

--
Cheers,

David / dhildenb