[PATCH] mm/THP: use hugepage_vma_check() in khugepaged_enter_vma_merge()

[PATCH] mm/THP: use hugepage_vma_check() in khugepaged_enter_vma_merge()

[Song Liu]

khugepaged_enter_vma_merge() is using a different approach to check
whether a vma is valid for khugepaged_enter():

    if (!vma->anon_vma)
            /*
             * Not yet faulted in so we will register later in the
             * page fault if needed.
             */
            return 0;
    if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
            /* khugepaged not yet working on file or special mappings */
            return 0;

This check has some problems. One of the obvious problems is that
it doesn't check shmem_file(), so that vma backed with shmem files
will not call khugepaged_enter().

This patch fixes these problems by reusing hugepage_vma_check() in
khugepaged_enter_vma_merge().

Signed-off-by: Song Liu <songliubraving@fb.com>
---
 mm/khugepaged.c | 12 ++++--------
 1 file changed, 4 insertions(+), 8 deletions(-)

diff --git a/mm/khugepaged.c b/mm/khugepaged.c
index d7b2a4b..e50c2bd 100644
--- a/mm/khugepaged.c
+++ b/mm/khugepaged.c
@@ -430,18 +430,14 @@ int __khugepaged_enter(struct mm_struct *mm)
 	return 0;
 }
 
+static bool hugepage_vma_check(struct vm_area_struct *vma);
+
 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
 			       unsigned long vm_flags)
 {
 	unsigned long hstart, hend;
-	if (!vma->anon_vma)
-		/*
-		 * Not yet faulted in so we will register later in the
-		 * page fault if needed.
-		 */
-		return 0;
-	if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
-		/* khugepaged not yet working on file or special mappings */
+
+	if (!hugepage_vma_check(vma))
 		return 0;
 	hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 	hend = vma->vm_end & HPAGE_PMD_MASK;
-- 
2.9.5

Re: [PATCH] mm/THP: use hugepage_vma_check() in khugepaged_enter_vma_merge()

[Rik van Riel]

[-- Attachment #1: Type: text/plain, Size: 1665 bytes --]

On Mon, 2018-05-21 at 12:38 -0700, Song Liu wrote:

> This patch fixes these problems by reusing hugepage_vma_check() in
> khugepaged_enter_vma_merge().

Lets take a look at this in more detail. This effectively
adds the following conditions to khugepaged_enter_vma_merge:
- fail if MMF_DISABLE_THP bit is set in mm->flags (good)
- allow if merging a tmpfs file and THP tmpfs is enabled (good)
- disallow if is_vma_temporary_stack (good)
- otherwise, allow if !VM_NO_KHUGEPAGED flag (good)

Looks like this covers all the conditions I can think
of, and if I missed any, chances are that condition
should be added to hugepage_vma_check()...

> Signed-off-by: Song Liu <songliubraving@fb.com>

Reviewed-by: Rik van Riel <riel@surriel.com>

> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> index d7b2a4b..e50c2bd 100644
> --- a/mm/khugepaged.c
> +++ b/mm/khugepaged.c
> @@ -430,18 +430,14 @@ int __khugepaged_enter(struct mm_struct *mm)
>  	return 0;
>  }
>  
> +static bool hugepage_vma_check(struct vm_area_struct *vma);
> +
>  int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
>  			       unsigned long vm_flags)
>  {
>  	unsigned long hstart, hend;
> -	if (!vma->anon_vma)
> -		/*
> -		 * Not yet faulted in so we will register later in
> the
> -		 * page fault if needed.
> -		 */
> -		return 0;
> -	if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
> -		/* khugepaged not yet working on file or special
> mappings */
> +
> +	if (!hugepage_vma_check(vma))
>  		return 0;
>  	hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
>  	hend = vma->vm_end & HPAGE_PMD_MASK;
-- 
All Rights Reversed.

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

Re: [PATCH] mm/THP: use hugepage_vma_check() in khugepaged_enter_vma_merge()

[David Rientjes]

On Mon, 21 May 2018, Song Liu wrote:

> khugepaged_enter_vma_merge() is using a different approach to check
> whether a vma is valid for khugepaged_enter():
> 
>     if (!vma->anon_vma)
>             /*
>              * Not yet faulted in so we will register later in the
>              * page fault if needed.
>              */
>             return 0;
>     if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
>             /* khugepaged not yet working on file or special mappings */
>             return 0;
> 
> This check has some problems. One of the obvious problems is that
> it doesn't check shmem_file(), so that vma backed with shmem files
> will not call khugepaged_enter().
> 
> This patch fixes these problems by reusing hugepage_vma_check() in
> khugepaged_enter_vma_merge().
> 
> Signed-off-by: Song Liu <songliubraving@fb.com>

Acked-by: David Rientjes <rientjes@google.com>

Re: [PATCH] mm/THP: use hugepage_vma_check() in khugepaged_enter_vma_merge()

[Michal Hocko]

[CC Kirill]

On Mon 21-05-18 12:38:53, Song Liu wrote:
> khugepaged_enter_vma_merge() is using a different approach to check
> whether a vma is valid for khugepaged_enter():
> 
>     if (!vma->anon_vma)
>             /*
>              * Not yet faulted in so we will register later in the
>              * page fault if needed.
>              */
>             return 0;
>     if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
>             /* khugepaged not yet working on file or special mappings */
>             return 0;
> 
> This check has some problems. One of the obvious problems is that
> it doesn't check shmem_file(), so that vma backed with shmem files
> will not call khugepaged_enter().
> 
> This patch fixes these problems by reusing hugepage_vma_check() in
> khugepaged_enter_vma_merge().

It would be great to be more explicit about what are the actual
consequences. khugepaged_enter_vma_merge is called from multiple
context. Some of then do not really care about !anon case (e.g. stack
expansion). hugepage_madvise is quite convoluted so I am not really sure
from a quick look (are we simply not going to merge vmas even if we
could?).

Have you noticed this by a code inspection or you have seen this
happening in real workloads (aka, is this worth backporting to stable
trees)?
 
> Signed-off-by: Song Liu <songliubraving@fb.com>
> ---
>  mm/khugepaged.c | 12 ++++--------
>  1 file changed, 4 insertions(+), 8 deletions(-)
> 
> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> index d7b2a4b..e50c2bd 100644
> --- a/mm/khugepaged.c
> +++ b/mm/khugepaged.c
> @@ -430,18 +430,14 @@ int __khugepaged_enter(struct mm_struct *mm)
>  	return 0;
>  }
>  
> +static bool hugepage_vma_check(struct vm_area_struct *vma);
> +
>  int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
>  			       unsigned long vm_flags)
>  {
>  	unsigned long hstart, hend;
> -	if (!vma->anon_vma)
> -		/*
> -		 * Not yet faulted in so we will register later in the
> -		 * page fault if needed.
> -		 */
> -		return 0;
> -	if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
> -		/* khugepaged not yet working on file or special mappings */
> +
> +	if (!hugepage_vma_check(vma))
>  		return 0;
>  	hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
>  	hend = vma->vm_end & HPAGE_PMD_MASK;
> -- 
> 2.9.5

-- 
Michal Hocko
SUSE Labs

Re: [PATCH] mm/THP: use hugepage_vma_check() in khugepaged_enter_vma_merge()

[Song Liu]

On May 22, 2018, at 5:13 AM, Michal Hocko <mhocko@kernel.org> wrote:
> 
> [CC Kirill]
> 
> On Mon 21-05-18 12:38:53, Song Liu wrote:
>> khugepaged_enter_vma_merge() is using a different approach to check
>> whether a vma is valid for khugepaged_enter():
>> 
>>    if (!vma->anon_vma)
>>            /*
>>             * Not yet faulted in so we will register later in the
>>             * page fault if needed.
>>             */
>>            return 0;
>>    if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
>>            /* khugepaged not yet working on file or special mappings */
>>            return 0;
>> 
>> This check has some problems. One of the obvious problems is that
>> it doesn't check shmem_file(), so that vma backed with shmem files
>> will not call khugepaged_enter().
>> 
>> This patch fixes these problems by reusing hugepage_vma_check() in
>> khugepaged_enter_vma_merge().
> 
> It would be great to be more explicit about what are the actual
> consequences. khugepaged_enter_vma_merge is called from multiple
> context. Some of then do not really care about !anon case (e.g. stack
> expansion). hugepage_madvise is quite convoluted so I am not really sure
> from a quick look (are we simply not going to merge vmas even if we
> could?).

Yes, it does fix madvise for shmem with huge=advise option. I had made
a mistake in this version. I will send v2 with the more details on what
is fixed. 

> Have you noticed this by a code inspection or you have seen this
> happening in real workloads (aka, is this worth backporting to stable
> trees)?

I noticed this when reading the code. I think this might worth back 
porting. However, I don't know whether it fixes anything else other
than shmem, so I am not sure which versions need this fix. 

Thanks,
Song


>> Signed-off-by: Song Liu <songliubraving@fb.com>
>> ---
>> mm/khugepaged.c | 12 ++++--------
>> 1 file changed, 4 insertions(+), 8 deletions(-)
>> 
>> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
>> index d7b2a4b..e50c2bd 100644
>> --- a/mm/khugepaged.c
>> +++ b/mm/khugepaged.c
>> @@ -430,18 +430,14 @@ int __khugepaged_enter(struct mm_struct *mm)
>> 	return 0;
>> }
>> 
>> +static bool hugepage_vma_check(struct vm_area_struct *vma);
>> +
>> int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
>> 			       unsigned long vm_flags)
>> {
>> 	unsigned long hstart, hend;
>> -	if (!vma->anon_vma)
>> -		/*
>> -		 * Not yet faulted in so we will register later in the
>> -		 * page fault if needed.
>> -		 */
>> -		return 0;
>> -	if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
>> -		/* khugepaged not yet working on file or special mappings */
>> +
>> +	if (!hugepage_vma_check(vma))
>> 		return 0;
>> 	hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
>> 	hend = vma->vm_end & HPAGE_PMD_MASK;
>> -- 
>> 2.9.5
> 
> -- 
> Michal Hocko
> SUSE Labs