[PATCH 11/12] mm, thp, tmpfs: enable thp page cache in tmpfs

[PATCH 11/12] mm, thp, tmpfs: enable thp page cache in tmpfs

[Ning Qu]

Signed-off-by: Ning Qu <quning@gmail.com>
---
 mm/Kconfig | 4 ++--
 mm/shmem.c | 5 +++++
 2 files changed, 7 insertions(+), 2 deletions(-)

diff --git a/mm/Kconfig b/mm/Kconfig
index 562f12f..4d2f90f 100644
--- a/mm/Kconfig
+++ b/mm/Kconfig
@@ -428,8 +428,8 @@ config TRANSPARENT_HUGEPAGE_PAGECACHE
 	help
 	  Enabling the option adds support hugepages for file-backed
 	  mappings. It requires transparent hugepage support from
-	  filesystem side. For now, the only filesystem which supports
-	  hugepages is ramfs.
+	  filesystem side. For now, the filesystems which support
+	  hugepages are: ramfs and tmpfs.
 
 config CROSS_MEMORY_ATTACH
 	bool "Cross Memory Support"
diff --git a/mm/shmem.c b/mm/shmem.c
index 75c0ac6..50a3335 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -1672,6 +1672,11 @@ static struct inode *shmem_get_inode(struct super_block *sb, const struct inode
 			break;
 		case S_IFREG:
 			inode->i_mapping->a_ops = &shmem_aops;
+			/*
+			 * TODO: make tmpfs pages movable
+			 */
+			mapping_set_gfp_mask(inode->i_mapping,
+					     GFP_TRANSHUGE & ~__GFP_MOVABLE);
 			inode->i_op = &shmem_inode_operations;
 			inode->i_fop = &shmem_file_operations;
 			mpol_shared_policy_init(&info->policy,
-- 
1.8.4

RE: [PATCH 11/12] mm, thp, tmpfs: enable thp page cache in tmpfs

[Kirill A. Shutemov]

Ning Qu wrote:
> Signed-off-by: Ning Qu <quning@gmail.com>
> ---
>  mm/Kconfig | 4 ++--
>  mm/shmem.c | 5 +++++
>  2 files changed, 7 insertions(+), 2 deletions(-)
> 
> diff --git a/mm/Kconfig b/mm/Kconfig
> index 562f12f..4d2f90f 100644
> --- a/mm/Kconfig
> +++ b/mm/Kconfig
> @@ -428,8 +428,8 @@ config TRANSPARENT_HUGEPAGE_PAGECACHE
>  	help
>  	  Enabling the option adds support hugepages for file-backed
>  	  mappings. It requires transparent hugepage support from
> -	  filesystem side. For now, the only filesystem which supports
> -	  hugepages is ramfs.
> +	  filesystem side. For now, the filesystems which support
> +	  hugepages are: ramfs and tmpfs.
>  
>  config CROSS_MEMORY_ATTACH
>  	bool "Cross Memory Support"
> diff --git a/mm/shmem.c b/mm/shmem.c
> index 75c0ac6..50a3335 100644
> --- a/mm/shmem.c
> +++ b/mm/shmem.c
> @@ -1672,6 +1672,11 @@ static struct inode *shmem_get_inode(struct super_block *sb, const struct inode
>  			break;
>  		case S_IFREG:
>  			inode->i_mapping->a_ops = &shmem_aops;
> +			/*
> +			 * TODO: make tmpfs pages movable
> +			 */
> +			mapping_set_gfp_mask(inode->i_mapping,
> +					     GFP_TRANSHUGE & ~__GFP_MOVABLE);

Unlike ramfs, tmpfs pages are movable before transparent page cache
patchset.
Making tmpfs pages non-movable looks like a big regression to me. It need
to be fixed before proposing it upstream.

-- 
 Kirill A. Shutemov

Re: [PATCH 11/12] mm, thp, tmpfs: enable thp page cache in tmpfs

[Ning Qu]

I agree with this. It has been like this just for a quick proof, but I
need to address this problem as soon as possible.

Thanks!
Best wishes,
-- 
Ning Qu (曲宁) | Software Engineer | quning@google.com | +1-408-418-6066


On Tue, Oct 15, 2013 at 4:09 AM, Kirill A. Shutemov
<kirill.shutemov@linux.intel.com> wrote:
> Ning Qu wrote:
>> Signed-off-by: Ning Qu <quning@gmail.com>
>> ---
>>  mm/Kconfig | 4 ++--
>>  mm/shmem.c | 5 +++++
>>  2 files changed, 7 insertions(+), 2 deletions(-)
>>
>> diff --git a/mm/Kconfig b/mm/Kconfig
>> index 562f12f..4d2f90f 100644
>> --- a/mm/Kconfig
>> +++ b/mm/Kconfig
>> @@ -428,8 +428,8 @@ config TRANSPARENT_HUGEPAGE_PAGECACHE
>>       help
>>         Enabling the option adds support hugepages for file-backed
>>         mappings. It requires transparent hugepage support from
>> -       filesystem side. For now, the only filesystem which supports
>> -       hugepages is ramfs.
>> +       filesystem side. For now, the filesystems which support
>> +       hugepages are: ramfs and tmpfs.
>>
>>  config CROSS_MEMORY_ATTACH
>>       bool "Cross Memory Support"
>> diff --git a/mm/shmem.c b/mm/shmem.c
>> index 75c0ac6..50a3335 100644
>> --- a/mm/shmem.c
>> +++ b/mm/shmem.c
>> @@ -1672,6 +1672,11 @@ static struct inode *shmem_get_inode(struct super_block *sb, const struct inode
>>                       break;
>>               case S_IFREG:
>>                       inode->i_mapping->a_ops = &shmem_aops;
>> +                     /*
>> +                      * TODO: make tmpfs pages movable
>> +                      */
>> +                     mapping_set_gfp_mask(inode->i_mapping,
>> +                                          GFP_TRANSHUGE & ~__GFP_MOVABLE);
>
> Unlike ramfs, tmpfs pages are movable before transparent page cache
> patchset.
> Making tmpfs pages non-movable looks like a big regression to me. It need
> to be fixed before proposing it upstream.
>
> --
>  Kirill A. Shutemov

Re: [PATCH 11/12] mm, thp, tmpfs: enable thp page cache in tmpfs

[Ning Qu]

I guess this is the last review I have for this around, but not sure
what's the best solution right now.

Kirill, do you think it's OK to just split the huge page when it will
be moved. Will look into how thp anon handle this situation.

Then after this, I probably will post v2.

Thanks!
Best wishes,
-- 
Ning Qu (曲宁) | Software Engineer | quning@google.com | +1-408-418-6066


On Tue, Oct 15, 2013 at 11:42 AM, Ning Qu <quning@google.com> wrote:
> I agree with this. It has been like this just for a quick proof, but I
> need to address this problem as soon as possible.
>
> Thanks!
> Best wishes,
> --
> Ning Qu (曲宁) | Software Engineer | quning@google.com | +1-408-418-6066
>
>
> On Tue, Oct 15, 2013 at 4:09 AM, Kirill A. Shutemov
> <kirill.shutemov@linux.intel.com> wrote:
>> Ning Qu wrote:
>>> Signed-off-by: Ning Qu <quning@gmail.com>
>>> ---
>>>  mm/Kconfig | 4 ++--
>>>  mm/shmem.c | 5 +++++
>>>  2 files changed, 7 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/mm/Kconfig b/mm/Kconfig
>>> index 562f12f..4d2f90f 100644
>>> --- a/mm/Kconfig
>>> +++ b/mm/Kconfig
>>> @@ -428,8 +428,8 @@ config TRANSPARENT_HUGEPAGE_PAGECACHE
>>>       help
>>>         Enabling the option adds support hugepages for file-backed
>>>         mappings. It requires transparent hugepage support from
>>> -       filesystem side. For now, the only filesystem which supports
>>> -       hugepages is ramfs.
>>> +       filesystem side. For now, the filesystems which support
>>> +       hugepages are: ramfs and tmpfs.
>>>
>>>  config CROSS_MEMORY_ATTACH
>>>       bool "Cross Memory Support"
>>> diff --git a/mm/shmem.c b/mm/shmem.c
>>> index 75c0ac6..50a3335 100644
>>> --- a/mm/shmem.c
>>> +++ b/mm/shmem.c
>>> @@ -1672,6 +1672,11 @@ static struct inode *shmem_get_inode(struct super_block *sb, const struct inode
>>>                       break;
>>>               case S_IFREG:
>>>                       inode->i_mapping->a_ops = &shmem_aops;
>>> +                     /*
>>> +                      * TODO: make tmpfs pages movable
>>> +                      */
>>> +                     mapping_set_gfp_mask(inode->i_mapping,
>>> +                                          GFP_TRANSHUGE & ~__GFP_MOVABLE);
>>
>> Unlike ramfs, tmpfs pages are movable before transparent page cache
>> patchset.
>> Making tmpfs pages non-movable looks like a big regression to me. It need
>> to be fixed before proposing it upstream.
>>
>> --
>>  Kirill A. Shutemov

Re: [PATCH 11/12] mm, thp, tmpfs: enable thp page cache in tmpfs

[Ning Qu]

Hi, Kirill,

I have reviewed the current migration code and seems huge page will
get splitted during migration. Since our current split_huge_page
handle file cache as well, it seems not a problem for us?

We just simply not removing the __GFP_MOVABLE flag? Anything else I am
missing here?

Thanks!
Best wishes,
-- 
Ning Qu (曲宁) | Software Engineer | quning@google.com | +1-408-418-6066


On Fri, Oct 18, 2013 at 11:24 AM, Ning Qu <quning@google.com> wrote:
> I guess this is the last review I have for this around, but not sure
> what's the best solution right now.
>
> Kirill, do you think it's OK to just split the huge page when it will
> be moved. Will look into how thp anon handle this situation.
>
> Then after this, I probably will post v2.
>
> Thanks!
> Best wishes,
> --
> Ning Qu (曲宁) | Software Engineer | quning@google.com | +1-408-418-6066
>
>
> On Tue, Oct 15, 2013 at 11:42 AM, Ning Qu <quning@google.com> wrote:
>> I agree with this. It has been like this just for a quick proof, but I
>> need to address this problem as soon as possible.
>>
>> Thanks!
>> Best wishes,
>> --
>> Ning Qu (曲宁) | Software Engineer | quning@google.com | +1-408-418-6066
>>
>>
>> On Tue, Oct 15, 2013 at 4:09 AM, Kirill A. Shutemov
>> <kirill.shutemov@linux.intel.com> wrote:
>>> Ning Qu wrote:
>>>> Signed-off-by: Ning Qu <quning@gmail.com>
>>>> ---
>>>>  mm/Kconfig | 4 ++--
>>>>  mm/shmem.c | 5 +++++
>>>>  2 files changed, 7 insertions(+), 2 deletions(-)
>>>>
>>>> diff --git a/mm/Kconfig b/mm/Kconfig
>>>> index 562f12f..4d2f90f 100644
>>>> --- a/mm/Kconfig
>>>> +++ b/mm/Kconfig
>>>> @@ -428,8 +428,8 @@ config TRANSPARENT_HUGEPAGE_PAGECACHE
>>>>       help
>>>>         Enabling the option adds support hugepages for file-backed
>>>>         mappings. It requires transparent hugepage support from
>>>> -       filesystem side. For now, the only filesystem which supports
>>>> -       hugepages is ramfs.
>>>> +       filesystem side. For now, the filesystems which support
>>>> +       hugepages are: ramfs and tmpfs.
>>>>
>>>>  config CROSS_MEMORY_ATTACH
>>>>       bool "Cross Memory Support"
>>>> diff --git a/mm/shmem.c b/mm/shmem.c
>>>> index 75c0ac6..50a3335 100644
>>>> --- a/mm/shmem.c
>>>> +++ b/mm/shmem.c
>>>> @@ -1672,6 +1672,11 @@ static struct inode *shmem_get_inode(struct super_block *sb, const struct inode
>>>>                       break;
>>>>               case S_IFREG:
>>>>                       inode->i_mapping->a_ops = &shmem_aops;
>>>> +                     /*
>>>> +                      * TODO: make tmpfs pages movable
>>>> +                      */
>>>> +                     mapping_set_gfp_mask(inode->i_mapping,
>>>> +                                          GFP_TRANSHUGE & ~__GFP_MOVABLE);
>>>
>>> Unlike ramfs, tmpfs pages are movable before transparent page cache
>>> patchset.
>>> Making tmpfs pages non-movable looks like a big regression to me. It need
>>> to be fixed before proposing it upstream.
>>>
>>> --
>>>  Kirill A. Shutemov