Re: [PATCH v3 03/16] mm: avoid unnecessary uses of is_swap_pte()

[Zi Yan <ziy@nvidia.com>]

On 12 Nov 2025, at 10:59, Lorenzo Stoakes wrote:

> On Tue, Nov 11, 2025 at 09:58:36PM -0500, Zi Yan wrote:
>> On 10 Nov 2025, at 17:21, Lorenzo Stoakes wrote:
>>
>>> There's an established convention in the kernel that we treat PTEs as
>>> containing swap entries (and the unfortunately named non-swap swap entries)
>>> should they be neither empty (i.e. pte_none() evaluating true) nor present
>>> (i.e. pte_present() evaluating true).
>>>
>>> However, there is some inconsistency in how this is applied, as we also
>>> have the is_swap_pte() helper which explicitly performs this check:
>>>
>>> 	/* check whether a pte points to a swap entry */
>>> 	static inline int is_swap_pte(pte_t pte)
>>> 	{
>>> 		return !pte_none(pte) && !pte_present(pte);
>>> 	}
>>>
>>> As this represents a predicate, and it's logical to assume that in order to
>>> establish that a PTE entry can correctly be manipulated as a swap/non-swap
>>> entry, this predicate seems as if it must first be checked.
>>>
>>> But we instead, we far more often utilise the established convention of
>>> checking pte_none() / pte_present() before operating on entries as if they
>>> were swap/non-swap.
>>>
>>> This patch works towards correcting this inconsistency by removing all uses
>>> of is_swap_pte() where we are already in a position where we perform
>>> pte_none()/pte_present() checks anyway or otherwise it is clearly logical
>>> to do so.
>>>
>>> We also take advantage of the fact that pte_swp_uffd_wp() is only set on
>>> swap entries.
>>>
>>> Additionally, update comments referencing to is_swap_pte() and
>>> non_swap_entry().
>>>
>>> No functional change intended.
>>>
>>> Signed-off-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
>>> ---
>>>  fs/proc/task_mmu.c            | 49 ++++++++++++++++++++++++-----------
>>>  include/linux/userfaultfd_k.h |  3 +--
>>>  mm/hugetlb.c                  |  6 ++---
>>>  mm/internal.h                 |  6 ++---
>>>  mm/khugepaged.c               | 29 +++++++++++----------
>>>  mm/migrate.c                  |  2 +-
>>>  mm/mprotect.c                 | 43 ++++++++++++++----------------
>>>  mm/mremap.c                   |  7 +++--
>>>  mm/page_table_check.c         | 13 ++++++----
>>>  mm/page_vma_mapped.c          | 31 +++++++++++-----------
>>>  10 files changed, 104 insertions(+), 85 deletions(-)
>>>
>>
>> <snip>
>>
>>> diff --git a/mm/page_vma_mapped.c b/mm/page_vma_mapped.c
>>> index be20468fb5a9..a4e23818f37f 100644
>>> --- a/mm/page_vma_mapped.c
>>> +++ b/mm/page_vma_mapped.c
>>> @@ -16,6 +16,7 @@ static inline bool not_found(struct page_vma_mapped_walk *pvmw)
>>>  static bool map_pte(struct page_vma_mapped_walk *pvmw, pmd_t *pmdvalp,
>>>  		    spinlock_t **ptlp)
>>>  {
>>> +	bool is_migration;
>>>  	pte_t ptent;
>>>
>>>  	if (pvmw->flags & PVMW_SYNC) {
>>> @@ -26,6 +27,7 @@ static bool map_pte(struct page_vma_mapped_walk *pvmw, pmd_t *pmdvalp,
>>>  		return !!pvmw->pte;
>>>  	}
>>>
>>> +	is_migration = pvmw->flags & PVMW_MIGRATION;
>>>  again:
>>>  	/*
>>>  	 * It is important to return the ptl corresponding to pte,
>>> @@ -41,11 +43,14 @@ static bool map_pte(struct page_vma_mapped_walk *pvmw, pmd_t *pmdvalp,
>>>
>>>  	ptent = ptep_get(pvmw->pte);
>>>
>>> -	if (pvmw->flags & PVMW_MIGRATION) {
>>> -		if (!is_swap_pte(ptent))
>>
>> Here, is_migration = true and either pte_none() or pte_present()
>> would return false, and ...
>>
>>> +	if (pte_none(ptent)) {
>>> +		return false;
>>> +	} else if (pte_present(ptent)) {
>>> +		if (is_migration)
>>>  			return false;
>>> -	} else if (is_swap_pte(ptent)) {
>>> +	} else if (!is_migration) {
>>>  		swp_entry_t entry;
>>> +
>>>  		/*
>>>  		 * Handle un-addressable ZONE_DEVICE memory.
>>>  		 *
>>> @@ -66,8 +71,6 @@ static bool map_pte(struct page_vma_mapped_walk *pvmw, pmd_t *pmdvalp,
>>>  		if (!is_device_private_entry(entry) &&
>>>  		    !is_device_exclusive_entry(entry))
>>>  			return false;
>>> -	} else if (!pte_present(ptent)) {
>>> -		return false;
>>
>> ... is_migration = false and !pte_present() is actually pte_none(),
>> because of the is_swap_pte() above the added !is_migration check.
>> So pte_none() should return false regardless of is_migration.
>
> I guess you were working this through :) well I decided to also just to
> double-check I got it right, maybe useful for you also :P -
>
> Previously:
>
> 	if (is_migration) {
> 		if (!is_swap_pte(ptent))
> 			return false;
> 	} else if (is_swap_pte(ptent)) {
> 		... ZONE_DEVICE blah ...
> 	} else if (!pte_present(ptent)) {
> 		return false;
> 	}
>
> But is_swap_pte() is the same as !pte_none() && !pte_present(), so
> !is_swap_pte() is pte_none() || pte_present() by De Morgan's law:
>
> 	if (is_migration) {
> 		if (pte_none(ptent) || pte_present(ptent))
> 			return false;
> 	} else if (!pte_none(ptent) && !pte_present(ptent)) {
> 		... ZONE_DEVICE blah ...
> 	} else if (!pte_present(ptent)) {
> 		return false;
> 	}
>
> In the last branch, we know (again by De Morgan's law) that either
> pte_none(ptent) or pte_present(ptent).. But we explicitly check for
> !pte_present(ptent) so this becomes:
>
> 	if (is_migration) {
> 		if (pte_none(ptent) || pte_present(ptent))
> 			return false;
> 	} else if (!pte_none(ptent) && !pte_present(ptent)) {
> 		... ZONE_DEVICE blah ...
> 	} else if (pte_none(ptent)) {
> 		return false;
> 	}
>
> So we can generalise - regardless of is_migration, pte_none() returns false:
>
> 	if (pte_none(ptent)) {
> 		return false;
> 	} else if (is_migration) {
> 		if (pte_none(ptent) || pte_present(ptent))
> 			return false;
> 	} else if (!pte_none(ptent) && !pte_present(ptent)) {
> 		... ZONE_DEVICE blah ...
> 	}
>
> Since we already check for pte_none() ahead of time, we can simplify again:
>
> 	if (pte_none(ptent)) {
> 		return false;
> 	} else if (is_migration) {
> 		if (pte_present(ptent))
> 			return false;
> 	} else if (!pte_present(ptent)) {
> 		... ZONE_DEVICE blah ...
> 	}
>
> We can then put the pte_present() check in the outer branch:
>
> 	if (pte_none(ptent)) {
> 		return false;
> 	} else if (pte_present(ptent)) {
> 		if (is_migration)
> 			return false;
> 	} else if (!is_migration) {
> 		... ZONE_DEVICE blah ...
> 	}
>
> Because previously an is_migration && !pte_present() case would result in no
> action here.
>
> Which is the code in this patch :)

Thanks again for spelling out the whole process.

>
>>
>> This is a nice cleanup. Thanks.
>>
>>>  	}
>>>  	spin_lock(*ptlp);
>>>  	if (unlikely(!pmd_same(*pmdvalp, pmdp_get_lockless(pvmw->pmd)))) {
>>> @@ -113,21 +116,17 @@ static bool check_pte(struct page_vma_mapped_walk *pvmw, unsigned long pte_nr)
>>>  			return false;
>>>
>>>  		pfn = softleaf_to_pfn(entry);
>>> -	} else if (is_swap_pte(ptent)) {
>>> -		swp_entry_t entry;
>>> +	} else if (pte_present(ptent)) {
>>> +		pfn = pte_pfn(ptent);
>>> +	} else {
>>> +		const softleaf_t entry = softleaf_from_pte(ptent);
>>>
>>>  		/* Handle un-addressable ZONE_DEVICE memory */
>>> -		entry = pte_to_swp_entry(ptent);
>>> -		if (!is_device_private_entry(entry) &&
>>> -		    !is_device_exclusive_entry(entry))
>>> -			return false;
>>> -
>>> -		pfn = swp_offset_pfn(entry);
>>> -	} else {
>>> -		if (!pte_present(ptent))
>>
>> This !pte_present() is pte_none(). It seems that there should be
>
> Well this should be fine though as:
>
> 		const softleaf_t entry = softleaf_from_pte(ptent);
>
> 		/* Handle un-addressable ZONE_DEVICE memory */
> 		if (!softleaf_is_device_private(entry) &&
> 		    !softleaf_is_device_exclusive(entry))
> 			return false;
>
> Still correctly handles none - as softleaf_from_pte() in case of pte_none() will
> be a none softleaf entry which will fail both of these tests.
>
> So excluding pte_none() as an explicit test here was part of the rework - we no
> longer have to do that.

Got it. Now my RB is yours. :)

>
>>
>> } else if (pte_none(ptent)) {
>> 	return false;
>> }
>>
>> before the above "} else {".
>>
>>> +		if (!softleaf_is_device_private(entry) &&
>>> +		    !softleaf_is_device_exclusive(entry))
>>>  			return false;
>>>
>>> -		pfn = pte_pfn(ptent);
>>> +		pfn = softleaf_to_pfn(entry);
>>>  	}
>>>
>>>  	if ((pfn + pte_nr - 1) < pvmw->pfn)
>>> --
>>> 2.51.0
>>
>> Otherwise, LGTM. With the above issue addressed, feel free to
>> add Reviewed-by: Zi Yan <ziy@nvidia.com>
>
> Thanks!
>
>>
>> --
>> Best Regards,
>> Yan, Zi
>
> Cheers, Lorenzo


Best Regards,
Yan, Zi