[PATCH BUGFIX] mm: fix find_get_page() for shmem exceptional entries

[PATCH BUGFIX] mm: fix find_get_page() for shmem exceptional entries

[Konstantin Khlebnikov]

It should return NULL, otherwise the caller will be very surprised.

Signed-off-by: Konstantin Khlebnikov <khlebnikov@openvz.org>
---
 mm/filemap.c |    1 +
 1 files changed, 1 insertions(+), 0 deletions(-)

diff --git a/mm/filemap.c b/mm/filemap.c
index 518223b..ca98cb5 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -693,6 +693,7 @@ repeat:
 			 * here as an exceptional entry: so return it without
 			 * attempting to raise page count.
 			 */
+			page = NULL;
 			goto out;
 		}
 		if (!page_cache_get_speculative(page))

Re: [PATCH BUGFIX] mm: fix find_get_page() for shmem exceptional entries

[Konstantin Khlebnikov]

Bug was added in commit v3.0-7291-g8079b1c (mm: clarify the radix_tree exceptional cases)
So, v3.1 and v3.2 affected.

Konstantin Khlebnikov wrote:
> It should return NULL, otherwise the caller will be very surprised.
>
> Signed-off-by: Konstantin Khlebnikov<khlebnikov@openvz.org>
> ---
>   mm/filemap.c |    1 +
>   1 files changed, 1 insertions(+), 0 deletions(-)
>
> diff --git a/mm/filemap.c b/mm/filemap.c
> index 518223b..ca98cb5 100644
> --- a/mm/filemap.c
> +++ b/mm/filemap.c
> @@ -693,6 +693,7 @@ repeat:
>   			 * here as an exceptional entry: so return it without
>   			 * attempting to raise page count.
>   			 */
> +			page = NULL;
>   			goto out;
>   		}
>   		if (!page_cache_get_speculative(page))
>

Re: [PATCH BUGFIX] mm: fix find_get_page() for shmem exceptional entries

[Hugh Dickins]

On Tue, 7 Feb 2012, Konstantin Khlebnikov wrote:

> Bug was added in commit v3.0-7291-g8079b1c (mm: clarify the radix_tree
> exceptional cases)
> So, v3.1 and v3.2 affected.
> 
> Konstantin Khlebnikov wrote:
> > It should return NULL, otherwise the caller will be very surprised.
> > 
> > Signed-off-by: Konstantin Khlebnikov<khlebnikov@openvz.org>

Thanks for worrying about it, but Nak to this patch.

If you have found somewhere that is surprised by an exceptional entry
instead of a page, then indeed we shall need to fix that: I'm not
aware of any.

There are several places that are prepared for the possibility:
find_lock_page() (and your patch would be breaking shmem.c's use of
find_lock_page()), mincore_page(), memcontrol.c's mc_handle_file_pte().

Of the remaining calls to find_get_page(), my understanding is that
either they are filesystems operating upon their own pagecache, or
they involve using ->readpage() - that's one of the two reasons why
I gave shmem its own ->splice_read() and removed its ->readpage()
before switching over to use the exceptional entries.

Hugh

> > ---
> >   mm/filemap.c |    1 +
> >   1 files changed, 1 insertions(+), 0 deletions(-)
> > 
> > diff --git a/mm/filemap.c b/mm/filemap.c
> > index 518223b..ca98cb5 100644
> > --- a/mm/filemap.c
> > +++ b/mm/filemap.c
> > @@ -693,6 +693,7 @@ repeat:
> >   			 * here as an exceptional entry: so return it without
> >   			 * attempting to raise page count.
> >   			 */
> > +			page = NULL;
> >   			goto out;
> >   		}
> >   		if (!page_cache_get_speculative(page))

Re: [PATCH BUGFIX] mm: fix find_get_page() for shmem exceptional entries

[Konstantin Khlebnikov]

Hugh Dickins wrote:
> On Tue, 7 Feb 2012, Konstantin Khlebnikov wrote:
>
>> Bug was added in commit v3.0-7291-g8079b1c (mm: clarify the radix_tree
>> exceptional cases)
>> So, v3.1 and v3.2 affected.
>>
>> Konstantin Khlebnikov wrote:
>>> It should return NULL, otherwise the caller will be very surprised.
>>>
>>> Signed-off-by: Konstantin Khlebnikov<khlebnikov@openvz.org>
>
> Thanks for worrying about it, but Nak to this patch.
>
> If you have found somewhere that is surprised by an exceptional entry
> instead of a page, then indeed we shall need to fix that: I'm not
> aware of any.

Oh, this is very dangerous semantics, especially for function called "find-get-page"
which sometimes returns not-getted not-a-page =)

>
> There are several places that are prepared for the possibility:
> find_lock_page() (and your patch would be breaking shmem.c's use of
> find_lock_page()), mincore_page(), memcontrol.c's mc_handle_file_pte().
>
> Of the remaining calls to find_get_page(), my understanding is that
> either they are filesystems operating upon their own pagecache, or
> they involve using ->readpage() - that's one of the two reasons why
> I gave shmem its own ->splice_read() and removed its ->readpage()
> before switching over to use the exceptional entries.
>
> Hugh
>
>>> ---
>>>    mm/filemap.c |    1 +
>>>    1 files changed, 1 insertions(+), 0 deletions(-)
>>>
>>> diff --git a/mm/filemap.c b/mm/filemap.c
>>> index 518223b..ca98cb5 100644
>>> --- a/mm/filemap.c
>>> +++ b/mm/filemap.c
>>> @@ -693,6 +693,7 @@ repeat:
>>>    			 * here as an exceptional entry: so return it without
>>>    			 * attempting to raise page count.
>>>    			 */
>>> +			page = NULL;
>>>    			goto out;
>>>    		}
>>>    		if (!page_cache_get_speculative(page))