Re: [PATCH] slab: Remove slab_folio()

[Matthew Wilcox <willy@infradead.org>]

On Wed, Sep 23, 2026 at 10:11:06PM +0200, Vlastimil Babka (SUSE) wrote:
> On 9/23/26 21:45, Harry Yoo wrote:
> > With a tiny little nit:
> > 
> >> @@ -184,15 +169,16 @@ static inline struct slab *page_slab(const struct page *page)
> >>   * slab_page - The first struct page allocated for a slab
> >>   * @s: The slab.
> >>   *
> >> - * A convenience wrapper for converting slab to the first struct page of the
> >> - * underlying folio, to communicate with code not yet converted to folio or
> >> - * struct slab.
> >> + * A convenience wrapper for returning the first page allocated to this slab.
> >> + * to communicate with code not yet converted to struct slab.
> >     * A convenience wrapper for returning the first "struct" page allocated to this
> >     * "slab" to communicate with code not yet converted to struct slab.
> >>   */
> > 
> > Perhaps you meant this?

My eyes are old and don't notice punctuation as well as they used to.
What I meant was:

 * A convenience wrapper for returning the first page allocated to this slab.
 * This is for the benefit of code not yet converted to struct slab.

> Yeah, and I wonder about the "code not yet converted" part, and if there's
> anything we can still convert right now. What I can see among slab_page() users:
> 
> - slab_order() - we could probably add order (aliasing page's order for now)
> to struct slab right now and that would be future-proof

That's trickier than you think because it's stored in the ->flags of
the second page when PG_head is set.  When divorcing slab and page,
it moves into struct slab, but we can't do that in advance.

> But:
> - slab_address() - I guess can only be converted once slab is divorced from
> page?

Yeah, it becomes slab->virtual.

> - __kasan_poison_slab - same as above?

The loop calling page_kasan_tag_reset() makes me think kasan really
wants a page here.

> - several VM_BUG_ON_PAGE() usages - same?

We could probably introduce a VM_BUG_ON_SLAB if you're interested.

> - validate_slab_ptr() - same?

That'll change dramatically -- if we want to validate that a pointer
really points to a slab object, we'll want to, er, ask if
virt_to_slab() belongs to the slab slab.

> __free_slab() - same, obviously

Right.