Re: [RFC PATCH 00/12] remove is_swap_[pte, pmd]() + non-swap confusion

[Jason Gunthorpe]

On Fri, Oct 24, 2025 at 08:41:16AM +0100, Lorenzo Stoakes wrote:
> There's an established convention in the kernel that we treat leaf page
> tables (so far at the PTE, PMD level) as containing 'swap entries' should
> they be neither empty (i.e. p**_none() evaluating true) nor present
> (i.e. p**_present() evaluating true).

I have to say I've never liked the none-vs-present naming either.

> This is deeply confusing, so this series goes further and eliminates the
> non_swap_entry() predicate, replacing it with is_non_present_entry() - with
> an eye to a new convention of referring to these non-swap 'swap entries' as
> non-present.

I'm not keen on is_non_present_entry(), it seems confusing again.

It looks like we are stuck with swp_entry_t as the being the handle
for a non-present pte. Oh well, not a great name, but fine..

So we think of that swp_entry_t having multiple types: swap, migration,
device private, etc, etc

Then I'd think the general pattern should be to get a swp_entry_t:

    if (pte_present(pte))
        return;
    swpent = pte_to_swp_entry(pte);

And then evaluate the type:

    if (swpent_is_swap()) {
    }


If you keep the naming as "swp_entry" indicates the multi-type value,
then "swap" can mean a swp_entry which is used by the swap subsystem.

That suggests functions like this:

swpent_is_swap()
swpent_is_migration()
..

and your higher level helpers like:

/* True if the pte is a swpent_is_swap() */
static inline bool swpent_get_swap_pte(pte_t pte, swp_entry_t *entryp)
{
   if (pte_present(pte))
        return false;
   *swpent = pte_to_swp_entry(pte);
   return swpent_is_swap(*swpent);
}

I also think it will be more readable to keep all these things under a
swpent namespace instead of using unstructured english names.

> * pte_to_swp_entry_or_zero() - allows for convenient conversion from a PTE
>   to a swap entry if present, or an empty swap entry if none. This is
>   useful as many swap entry conversions are simply checking for flags for
>   which this suffices.

I'd expect a safe function should be more like

   *swpent = pte_to_swp_entry_safe(pte);
   return swpent_is_swap(*swpent);

Where "safe" means that if the PTE is None or Present then
swpent_is_XX() == false. Ie it returns a 0 swpent and 0 swpent is
always nothing.

> * get_pte_swap_entry() - Retrieves a PTE swap entry if it truly is a swap
>   entry (i.e. not a non-present entry), returning true if so, otherwise
>   returns false. This simplifies a lot of logic that previously open-coded
>   this.

Like this is still a tortured function:

+static inline bool get_pte_swap_entry(pte_t pte, swp_entry_t *entryp)
+{
+       if (pte_present(pte))
+               return false;
+       if (pte_none(pte))
+               return false;
+
+       *entryp = pte_to_swp_entry(pte);
+       if (non_swap_entry(*entryp))
+               return false;
+
+       return true;
+}
+

static inline bool get_pte_swap_entry(pte_t pte, swp_entry_t *entryp)
{
   return swpent_is_swap(*swpent = pte_to_swp_entry_safe(pte));
}

Maybe it doesn't even need an inline at that point?

> * is_huge_pmd() - Determines if a PMD contains either a present transparent
>   huge page entry or a huge non-present entry. This again simplifies a lot
>   of logic that simply open-coded this.

is_huge_or_swpent_pmd() would be nicer, IMHO. I think it is surprising
when any of these APIs accept swap entries without being explicit

Jason