mm: Improve PageAnon() to check the whole PAGE_MAPPING_FLAGS
diff mbox series

Message ID 1573787903-2023-1-git-send-email-lixinhai.lxh@gmail.com
State New
Headers show
Series
  • mm: Improve PageAnon() to check the whole PAGE_MAPPING_FLAGS
Related show

Commit Message

Li Xinhai Nov. 15, 2019, 3:18 a.m. UTC
PageAnon() just checking on PAGE_MAPPING_ANON bit would cause page,
with PageKsm as true, been wrongly considered as PageAnon. Now,
checking the whole PAGE_MAPPING_FLAGS to avoid this error.

Reported from:
https://lore.kernel.org/linux-mm/20191113000651.20677-1-rcampbell@nvidia.com/

Reported-by: Ralph Campbell <rcampbell@nvidia.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
---
 include/linux/page-flags.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

Comments

Li Xinhai Nov. 15, 2019, 5:59 a.m. UTC | #1
On 2019-11-15 at 11:18 Li Xinhai wrote:
>PageAnon() just checking on PAGE_MAPPING_ANON bit would cause page,
>with PageKsm as true, been wrongly considered as PageAnon. Now,
>checking the whole PAGE_MAPPING_FLAGS to avoid this error.
>
>Reported from:
>https://lore.kernel.org/linux-mm/20191113000651.20677-1-rcampbell@nvidia.com/
>
>Reported-by: Ralph Campbell <rcampbell@nvidia.com>
>Cc: Andrew Morton <akpm@linux-foundation.org>
>Cc: Michal Hocko <mhocko@suse.com>
>Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
>---
> include/linux/page-flags.h | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
>diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
>index 1bf83c8..1849fc3 100644
>--- a/include/linux/page-flags.h
>+++ b/include/linux/page-flags.h
>@@ -461,7 +461,8 @@ static __always_inline int PageMappingFlags(struct page *page)
> static __always_inline int PageAnon(struct page *page)
> {
> page = compound_head(page);
>-	return ((unsigned long)page->mapping & PAGE_MAPPING_ANON) != 0;
>+	return ((unsigned long)page->mapping & PAGE_MAPPING_FLAGS) ==
>+	PAGE_MAPPING_ANON;
> }
>
> static __always_inline int __PageMovable(struct page *page)
>--
>1.8.3.1
> 

The current semantics of PageAnon() for both KSM and !KSM are used in many 
places, so can't change it alone without change other code.
Need skip this patch.

- Xinhai
David Hildenbrand Nov. 15, 2019, 12:50 p.m. UTC | #2
On 15.11.19 06:59, lixinhai.lxh@gmail.com wrote:
> On 2019-11-15 at 11:18 Li Xinhai wrote:
>> PageAnon() just checking on PAGE_MAPPING_ANON bit would cause page,
>> with PageKsm as true, been wrongly considered as PageAnon. Now,
>> checking the whole PAGE_MAPPING_FLAGS to avoid this error.
>>
>> Reported from:
>> https://lore.kernel.org/linux-mm/20191113000651.20677-1-rcampbell@nvidia.com/
>>
>> Reported-by: Ralph Campbell <rcampbell@nvidia.com>
>> Cc: Andrew Morton <akpm@linux-foundation.org>
>> Cc: Michal Hocko <mhocko@suse.com>
>> Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
>> ---
>> include/linux/page-flags.h | 3 ++-
>> 1 file changed, 2 insertions(+), 1 deletion(-)
>>
>> diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
>> index 1bf83c8..1849fc3 100644
>> --- a/include/linux/page-flags.h
>> +++ b/include/linux/page-flags.h
>> @@ -461,7 +461,8 @@ static __always_inline int PageMappingFlags(struct page *page)
>> static __always_inline int PageAnon(struct page *page)
>> {
>> page = compound_head(page);
>> -	return ((unsigned long)page->mapping & PAGE_MAPPING_ANON) != 0;
>> +	return ((unsigned long)page->mapping & PAGE_MAPPING_FLAGS) ==
>> +	PAGE_MAPPING_ANON;
>> }
>>
>> static __always_inline int __PageMovable(struct page *page)
>> --
>> 1.8.3.1
>>
> 
> The current semantics of PageAnon() for both KSM and !KSM are used in many 
> places, so can't change it alone without change other code.
> Need skip this patch.

... I assume that was intended because KSM only merges anonymous pages?
If it was not intended, it would scream for a cleanup.

> 
> - Xinhai
>
Li Xinhai Nov. 15, 2019, 1:09 p.m. UTC | #3
On 2019-11-15 at 20:50 David Hildenbrand wrote:
>On 15.11.19 06:59, lixinhai.lxh@gmail.com wrote:
>> On 2019-11-15 at 11:18 Li Xinhai wrote:
>>> PageAnon() just checking on PAGE_MAPPING_ANON bit would cause page,
>>> with PageKsm as true, been wrongly considered as PageAnon. Now,
>>> checking the whole PAGE_MAPPING_FLAGS to avoid this error.
>>>
>>> Reported from:
>>> https://lore.kernel.org/linux-mm/20191113000651.20677-1-rcampbell@nvidia.com/
>>>
>>> Reported-by: Ralph Campbell <rcampbell@nvidia.com>
>>> Cc: Andrew Morton <akpm@linux-foundation.org>
>>> Cc: Michal Hocko <mhocko@suse.com>
>>> Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
>>> ---
>>> include/linux/page-flags.h | 3 ++-
>>> 1 file changed, 2 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
>>> index 1bf83c8..1849fc3 100644
>>> --- a/include/linux/page-flags.h
>>> +++ b/include/linux/page-flags.h
>>> @@ -461,7 +461,8 @@ static __always_inline int PageMappingFlags(struct page *page)
>>> static __always_inline int PageAnon(struct page *page)
>>> {
>>> page = compound_head(page);
>>> -	return ((unsigned long)page->mapping & PAGE_MAPPING_ANON) != 0;
>>> +	return ((unsigned long)page->mapping & PAGE_MAPPING_FLAGS) ==
>>> +	PAGE_MAPPING_ANON;
>>> }
>>>
>>> static __always_inline int __PageMovable(struct page *page)
>>> --
>>> 1.8.3.1
>>>
>>
>> The current semantics of PageAnon() for both KSM and !KSM are used in many
>> places, so can't change it alone without change other code.
>> Need skip this patch.
>
>... I assume that was intended because KSM only merges anonymous pages?
>If it was not intended, it would scream for a cleanup.
> 
Yes, I think that was intended to keep the origial version of PageAnon() when add the 
PageKsm(). So all places where used to have PageAnon() didn't requre change, but other
places for new code we have some thing like (PageAnon() && !PageKsm()) check, or do 
checking in correct sequence.

One thing I am not quite sure is about couting of anonymous page, in case page is KSM 
or !KSM, hope those old codes has been correctly handled. 

>>
>> - Xinhai
>>
>
>
>--
>
>Thanks,
>
>David / dhildenb
>
Michal Hocko Nov. 15, 2019, 1:25 p.m. UTC | #4
On Fri 15-11-19 21:09:55, lixinhai.lxh@gmail.com wrote:
> On 2019-11-15 at 20:50 David Hildenbrand wrote:
> >On 15.11.19 06:59, lixinhai.lxh@gmail.com wrote:
> >> On 2019-11-15 at 11:18 Li Xinhai wrote:
> >>> PageAnon() just checking on PAGE_MAPPING_ANON bit would cause page,
> >>> with PageKsm as true, been wrongly considered as PageAnon. Now,
> >>> checking the whole PAGE_MAPPING_FLAGS to avoid this error.
> >>>
> >>> Reported from:
> >>> https://lore.kernel.org/linux-mm/20191113000651.20677-1-rcampbell@nvidia.com/
> >>>
> >>> Reported-by: Ralph Campbell <rcampbell@nvidia.com>
> >>> Cc: Andrew Morton <akpm@linux-foundation.org>
> >>> Cc: Michal Hocko <mhocko@suse.com>
> >>> Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
> >>> ---
> >>> include/linux/page-flags.h | 3 ++-
> >>> 1 file changed, 2 insertions(+), 1 deletion(-)
> >>>
> >>> diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
> >>> index 1bf83c8..1849fc3 100644
> >>> --- a/include/linux/page-flags.h
> >>> +++ b/include/linux/page-flags.h
> >>> @@ -461,7 +461,8 @@ static __always_inline int PageMappingFlags(struct page *page)
> >>> static __always_inline int PageAnon(struct page *page)
> >>> {
> >>> page = compound_head(page);
> >>> -	return ((unsigned long)page->mapping & PAGE_MAPPING_ANON) != 0;
> >>> +	return ((unsigned long)page->mapping & PAGE_MAPPING_FLAGS) ==
> >>> +	PAGE_MAPPING_ANON;
> >>> }
> >>>
> >>> static __always_inline int __PageMovable(struct page *page)
> >>> --
> >>> 1.8.3.1
> >>>
> >>
> >> The current semantics of PageAnon() for both KSM and !KSM are used in many
> >> places, so can't change it alone without change other code.
> >> Need skip this patch.
> >
> >... I assume that was intended because KSM only merges anonymous pages?
> >If it was not intended, it would scream for a cleanup.
> > 
> Yes, I think that was intended to keep the origial version of PageAnon() when add the 
> PageKsm(). So all places where used to have PageAnon() didn't requre change, but other
> places for new code we have some thing like (PageAnon() && !PageKsm()) check, or do 
> checking in correct sequence.
> 
> One thing I am not quite sure is about couting of anonymous page, in case page is KSM 
> or !KSM, hope those old codes has been correctly handled. 

I am not really sure I understand what do you like to achieve here. Yes
there are different checks representing different classes of pages. In
this particular case the ordering of the check was just wrong. That is
trivial to be fixed. PageAnon works as intented AFAIK. If that is not
the case then please be explicit why.
Li Xinhai Nov. 15, 2019, 3:01 p.m. UTC | #5
On 2019-11-15 at 21:25 Michal Hocko wrote:
>On Fri 15-11-19 21:09:55, lixinhai.lxh@gmail.com wrote:
>> On 2019-11-15 at 20:50 David Hildenbrand wrote:
>> >On 15.11.19 06:59, lixinhai.lxh@gmail.com wrote:
>> >> On 2019-11-15 at 11:18 Li Xinhai wrote:
>> >>> PageAnon() just checking on PAGE_MAPPING_ANON bit would cause page,
>> >>> with PageKsm as true, been wrongly considered as PageAnon. Now,
>> >>> checking the whole PAGE_MAPPING_FLAGS to avoid this error.
>> >>>
>> >>> Reported from:
>> >>> https://lore.kernel.org/linux-mm/20191113000651.20677-1-rcampbell@nvidia.com/
>> >>>
>> >>> Reported-by: Ralph Campbell <rcampbell@nvidia.com>
>> >>> Cc: Andrew Morton <akpm@linux-foundation.org>
>> >>> Cc: Michal Hocko <mhocko@suse.com>
>> >>> Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
>> >>> ---
>> >>> include/linux/page-flags.h | 3 ++-
>> >>> 1 file changed, 2 insertions(+), 1 deletion(-)
>> >>>
>> >>> diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
>> >>> index 1bf83c8..1849fc3 100644
>> >>> --- a/include/linux/page-flags.h
>> >>> +++ b/include/linux/page-flags.h
>> >>> @@ -461,7 +461,8 @@ static __always_inline int PageMappingFlags(struct page *page)
>> >>> static __always_inline int PageAnon(struct page *page)
>> >>> {
>> >>> page = compound_head(page);
>> >>> -	return ((unsigned long)page->mapping & PAGE_MAPPING_ANON) != 0;
>> >>> +	return ((unsigned long)page->mapping & PAGE_MAPPING_FLAGS) ==
>> >>> +	PAGE_MAPPING_ANON;
>> >>> }
>> >>>
>> >>> static __always_inline int __PageMovable(struct page *page)
>> >>> --
>> >>> 1.8.3.1
>> >>>
>> >>
>> >> The current semantics of PageAnon() for both KSM and !KSM are used in many
>> >> places, so can't change it alone without change other code.
>> >> Need skip this patch.
>> >
>> >... I assume that was intended because KSM only merges anonymous pages?
>> >If it was not intended, it would scream for a cleanup.
>> >
>> Yes, I think that was intended to keep the origial version of PageAnon() when add the
>> PageKsm(). So all places where used to have PageAnon() didn't requre change, but other
>> places for new code we have some thing like (PageAnon() && !PageKsm()) check, or do
>> checking in correct sequence.
>>
>> One thing I am not quite sure is about couting of anonymous page, in case page is KSM
>> or !KSM, hope those old codes has been correctly handled. 
>
>I am not really sure I understand what do you like to achieve here. Yes
>there are different checks representing different classes of pages. In
>this particular case the ordering of the check was just wrong. That is
>trivial to be fixed. PageAnon works as intented AFAIK. If that is not
>the case then please be explicit why. 
I was thinking about make changes so to using PageAnon() for Anonymous&!KSM and 
PageKsm() for Anonymous&KSM, then don't require consider sequence. Now, I realized 
that PageAnon() is intended for cover KSM and !KSM cases and it seems no big benefits 
to change the semantics of PageAnon().

>--
>Michal Hocko
>SUSE Labs

Patch
diff mbox series

diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
index 1bf83c8..1849fc3 100644
--- a/include/linux/page-flags.h
+++ b/include/linux/page-flags.h
@@ -461,7 +461,8 @@  static __always_inline int PageMappingFlags(struct page *page)
 static __always_inline int PageAnon(struct page *page)
 {
 	page = compound_head(page);
-	return ((unsigned long)page->mapping & PAGE_MAPPING_ANON) != 0;
+	return ((unsigned long)page->mapping & PAGE_MAPPING_FLAGS) == 
+				PAGE_MAPPING_ANON;
 }
 
 static __always_inline int __PageMovable(struct page *page)