Re: [PATCH 1/4] mm: Use frozen pages for page tables

[Matthew Wilcox <willy@infradead.org>]

On Thu, Nov 13, 2025 at 11:14:38AM -0800, Vishal Moola (Oracle) wrote:
> On Thu, Nov 13, 2025 at 10:24:27AM -0800, Vishal Moola (Oracle) wrote:
> > On Thu, Nov 13, 2025 at 02:04:43PM +0000, Matthew Wilcox (Oracle) wrote:
> > > Page tables do not use the reference count.  That means we can avoid
> 
> While looking into the second patch, I came across this one instance in
> sparc32. Commit 1996d47a0db ("sparc32: mm: Only call ctor()/dtor()
> functions for first and last user") decided to start using the refcount
> for something.
> 
> I'm not certain this code should even be using ptdescs, but theres a
> page_ptdesc() call in there for now :(.
> 
> As far as I can tell, this is only a thing for sparc32. Cc-ing Will as
> the author of those commits.

I think you're right that this is a problem for the ptdesc project,
but not (I think) a problem for this patch.  Maybe for the wording.

This patch changes pagetable_alloc() / pagetable_free() to use frozen
pages.  But sparc32 doesn't use pagetable_alloc() (sparc64 does).
So this patch should not break sparc32, and it can continue to use the
page refcount like this.

Where we do run into problems with this series is in patch 2
where I move the mod_node_page_state() from pagetable_pte_ctor() to
pagetable_alloc_noprof().  Since sparc32 doesn't call pagetable_alloc(),
its page tables will not be accounted.

I think I have a solution, but it'll take at least some build testing.
Looks like sparc64-linux-gnu-gcc at least accepts a -m32 flag, so I'll
give it a go.