[PATCH] mm: re-allow pinning of zero pfns (again)

[PATCH] mm: re-allow pinning of zero pfns (again)

[Alex Williamson]

The below referenced commit makes the same error as 1c563432588d ("mm: fix
is_pinnable_page against a cma page"), re-interpreting the logic to exclude
pinning of the zero page, which breaks device assignment with vfio.

Link: https://lore.kernel.org/all/165490039431.944052.12458624139225785964.stgit@omen
Fixes: f25cbb7a95a2 ("mm: add zone device coherent type memory support")
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
---
 include/linux/mm.h |    5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/include/linux/mm.h b/include/linux/mm.h
index 18e01474cf6b..772279ed7010 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -1544,9 +1544,8 @@ static inline bool is_longterm_pinnable_page(struct page *page)
 	if (mt == MIGRATE_CMA || mt == MIGRATE_ISOLATE)
 		return false;
 #endif
-	return !(is_device_coherent_page(page) ||
-		 is_zone_movable_page(page) ||
-		 is_zero_pfn(page_to_pfn(page)));
+	return !(is_device_coherent_page(page) || is_zone_movable_page(page)) ||
+	       is_zero_pfn(page_to_pfn(page));
 }
 #else
 static inline bool is_longterm_pinnable_page(struct page *page)

Re: [PATCH] mm: re-allow pinning of zero pfns (again)

[David Hildenbrand]

On 09.08.22 06:42, Alex Williamson wrote:
> The below referenced commit makes the same error as 1c563432588d ("mm: fix
> is_pinnable_page against a cma page"), re-interpreting the logic to exclude
> pinning of the zero page, which breaks device assignment with vfio.
> 
> Link: https://lore.kernel.org/all/165490039431.944052.12458624139225785964.stgit@omen
> Fixes: f25cbb7a95a2 ("mm: add zone device coherent type memory support")
> Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
> ---
>  include/linux/mm.h |    5 ++---
>  1 file changed, 2 insertions(+), 3 deletions(-)
> 
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index 18e01474cf6b..772279ed7010 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -1544,9 +1544,8 @@ static inline bool is_longterm_pinnable_page(struct page *page)
>  	if (mt == MIGRATE_CMA || mt == MIGRATE_ISOLATE)
>  		return false;
>  #endif
> -	return !(is_device_coherent_page(page) ||
> -		 is_zone_movable_page(page) ||
> -		 is_zero_pfn(page_to_pfn(page)));
> +	return !(is_device_coherent_page(page) || is_zone_movable_page(page)) ||
> +	       is_zero_pfn(page_to_pfn(page));
>  }
>  #else
>  static inline bool is_longterm_pinnable_page(struct page *page)


:/ I guess the code was moved just at the time the old code was still in
place, and when rebasing, the diff in the code was ignored.

Reviewed-by: David Hildenbrand <david@redhat.com>


I have patches in the works that will properly break COW here to get
anon pages instead of pinning the shared zeropage, which is questionable
in COW mappings.

-- 
Thanks,

David / dhildenb

Re: [PATCH] mm: re-allow pinning of zero pfns (again)

[Matthew Wilcox]

On Mon, Aug 08, 2022 at 10:42:24PM -0600, Alex Williamson wrote:
> The below referenced commit makes the same error as 1c563432588d ("mm: fix
> is_pinnable_page against a cma page"), re-interpreting the logic to exclude
> pinning of the zero page, which breaks device assignment with vfio.

Perhaps we need to admit we're not as good at boolean logic as we think
we are.

	if (is_device_coherent_page(page))
		return false;
	if (is_zone_movable_page(page))
		return false;
	return is_zero_pfn(page_to_pfn(page));

(or whatever the right logic is ... I just woke up and I'm having
trouble parsing it).

> Link: https://lore.kernel.org/all/165490039431.944052.12458624139225785964.stgit@omen
> Fixes: f25cbb7a95a2 ("mm: add zone device coherent type memory support")
> Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
> ---
>  include/linux/mm.h |    5 ++---
>  1 file changed, 2 insertions(+), 3 deletions(-)
> 
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index 18e01474cf6b..772279ed7010 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -1544,9 +1544,8 @@ static inline bool is_longterm_pinnable_page(struct page *page)
>  	if (mt == MIGRATE_CMA || mt == MIGRATE_ISOLATE)
>  		return false;
>  #endif
> -	return !(is_device_coherent_page(page) ||
> -		 is_zone_movable_page(page) ||
> -		 is_zero_pfn(page_to_pfn(page)));
> +	return !(is_device_coherent_page(page) || is_zone_movable_page(page)) ||
> +	       is_zero_pfn(page_to_pfn(page));
>  }
>  #else
>  static inline bool is_longterm_pinnable_page(struct page *page)
> 
> 
> 

Re: [PATCH] mm: re-allow pinning of zero pfns (again)

[Felix Kuehling]

Am 2022-08-09 um 08:31 schrieb Matthew Wilcox:
> On Mon, Aug 08, 2022 at 10:42:24PM -0600, Alex Williamson wrote:
>> The below referenced commit makes the same error as 1c563432588d ("mm: fix
>> is_pinnable_page against a cma page"), re-interpreting the logic to exclude
>> pinning of the zero page, which breaks device assignment with vfio.
> Perhaps we need to admit we're not as good at boolean logic as we think
> we are.
>
> 	if (is_device_coherent_page(page))
> 		return false;
> 	if (is_zone_movable_page(page))
> 		return false;
> 	return is_zero_pfn(page_to_pfn(page));
>
> (or whatever the right logic is ... I just woke up and I'm having
> trouble parsing it).

This implies an assumption that zero-page is never device-coherent or 
moveable, which is probably true, but not part of the original 
condition. A more formally correct rewrite would be:

	if (is_zero_pfn(page_to_pfn(page)))
		return true;
	if (is_device_coherent_page(page))
		return false;
	return !is_zone_moveable_page(page);

Regards,
   Felix


>
>> Link: https://lore.kernel.org/all/165490039431.944052.12458624139225785964.stgit@omen
>> Fixes: f25cbb7a95a2 ("mm: add zone device coherent type memory support")
>> Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
>> ---
>>   include/linux/mm.h |    5 ++---
>>   1 file changed, 2 insertions(+), 3 deletions(-)
>>
>> diff --git a/include/linux/mm.h b/include/linux/mm.h
>> index 18e01474cf6b..772279ed7010 100644
>> --- a/include/linux/mm.h
>> +++ b/include/linux/mm.h
>> @@ -1544,9 +1544,8 @@ static inline bool is_longterm_pinnable_page(struct page *page)
>>   	if (mt == MIGRATE_CMA || mt == MIGRATE_ISOLATE)
>>   		return false;
>>   #endif
>> -	return !(is_device_coherent_page(page) ||
>> -		 is_zone_movable_page(page) ||
>> -		 is_zero_pfn(page_to_pfn(page)));
>> +	return !(is_device_coherent_page(page) || is_zone_movable_page(page)) ||
>> +	       is_zero_pfn(page_to_pfn(page));
>>   }
>>   #else
>>   static inline bool is_longterm_pinnable_page(struct page *page)
>>
>>
>>

Re: [PATCH] mm: re-allow pinning of zero pfns (again)

[Matthew Wilcox]

On Tue, Aug 09, 2022 at 10:14:12AM -0400, Felix Kuehling wrote:
> Am 2022-08-09 um 08:31 schrieb Matthew Wilcox:
> > On Mon, Aug 08, 2022 at 10:42:24PM -0600, Alex Williamson wrote:
> > > The below referenced commit makes the same error as 1c563432588d ("mm: fix
> > > is_pinnable_page against a cma page"), re-interpreting the logic to exclude
> > > pinning of the zero page, which breaks device assignment with vfio.
> > Perhaps we need to admit we're not as good at boolean logic as we think
> > we are.
> > 
> > 	if (is_device_coherent_page(page))
> > 		return false;
> > 	if (is_zone_movable_page(page))
> > 		return false;
> > 	return is_zero_pfn(page_to_pfn(page));
> > 
> > (or whatever the right logic is ... I just woke up and I'm having
> > trouble parsing it).
> 
> This implies an assumption that zero-page is never device-coherent or
> moveable, which is probably true, but not part of the original condition. A
> more formally correct rewrite would be:
> 
> 	if (is_zero_pfn(page_to_pfn(page)))
> 		return true;
> 	if (is_device_coherent_page(page))
> 		return false;
> 	return !is_zone_moveable_page(page);

It's definitely true that the zero page is never device-coherent, nor
movable.  Moreover, we want to avoid calling page_to_pfn() if we can.
So it should be the last condition that we check.

Re: [PATCH] mm: re-allow pinning of zero pfns (again)

[David Hildenbrand]

On 09.08.22 16:43, Matthew Wilcox wrote:
> On Tue, Aug 09, 2022 at 10:14:12AM -0400, Felix Kuehling wrote:
>> Am 2022-08-09 um 08:31 schrieb Matthew Wilcox:
>>> On Mon, Aug 08, 2022 at 10:42:24PM -0600, Alex Williamson wrote:
>>>> The below referenced commit makes the same error as 1c563432588d ("mm: fix
>>>> is_pinnable_page against a cma page"), re-interpreting the logic to exclude
>>>> pinning of the zero page, which breaks device assignment with vfio.
>>> Perhaps we need to admit we're not as good at boolean logic as we think
>>> we are.
>>>
>>> 	if (is_device_coherent_page(page))
>>> 		return false;
>>> 	if (is_zone_movable_page(page))
>>> 		return false;
>>> 	return is_zero_pfn(page_to_pfn(page));
>>>
>>> (or whatever the right logic is ... I just woke up and I'm having
>>> trouble parsing it).
>>
>> This implies an assumption that zero-page is never device-coherent or
>> moveable, which is probably true, but not part of the original condition. A
>> more formally correct rewrite would be:
>>
>> 	if (is_zero_pfn(page_to_pfn(page)))
>> 		return true;
>> 	if (is_device_coherent_page(page))
>> 		return false;
>> 	return !is_zone_moveable_page(page);
> 
> It's definitely true that the zero page is never device-coherent, nor
> movable.  Moreover, we want to avoid calling page_to_pfn() if we can.
> So it should be the last condition that we check.

IIRC, with "kernelcore" and/or "movablecore", the zero page could
eventually end up in the movable zone (whereby we can have boottime
allocations being placed into the movable zone). IIRC that's why we have
to special-case on the zero-page here at all.

So taking out the zero page first is correct.

-- 
Thanks,

David / dhildenb

Re: [PATCH] mm: re-allow pinning of zero pfns (again)

[Andrew Morton]

On Tue, 9 Aug 2022 10:14:12 -0400 Felix Kuehling <felix.kuehling@amd.com> wrote:

> Am 2022-08-09 um 08:31 schrieb Matthew Wilcox:
> > On Mon, Aug 08, 2022 at 10:42:24PM -0600, Alex Williamson wrote:
> >> The below referenced commit makes the same error as 1c563432588d ("mm: fix
> >> is_pinnable_page against a cma page"), re-interpreting the logic to exclude
> >> pinning of the zero page, which breaks device assignment with vfio.

If two people made the same error then surely that's a sign that we
need a comment which explains things to the next visitor.

> > Perhaps we need to admit we're not as good at boolean logic as we think
> > we are.
> >
> > 	if (is_device_coherent_page(page))
> > 		return false;
> > 	if (is_zone_movable_page(page))
> > 		return false;
> > 	return is_zero_pfn(page_to_pfn(page));
> >
> > (or whatever the right logic is ... I just woke up and I'm having
> > trouble parsing it).
> 
> This implies an assumption that zero-page is never device-coherent or 
> moveable, which is probably true, but not part of the original 
> condition. A more formally correct rewrite would be:
> 
> 	if (is_zero_pfn(page_to_pfn(page)))
> 		return true;
> 	if (is_device_coherent_page(page))
> 		return false;
> 	return !is_zone_moveable_page(page);
> 

Yes please, vastly better.

And a nice thing about this layout is that it leaves places where we
can add a nice little comment against each clause of the test, to
explain why we're performing each one.