Re: [PATCH 0/2] new API to allocate buffer-cache for superblock in non-movable area

[Gioh Kim <gioh.kim@lge.com>]

2014-07-30 i??i?? 7:11, Jan Kara i?' e,?:
> On Wed 30-07-14 16:44:24, Gioh Kim wrote:
>> 2014-07-22 i??i?? 6:38, Jan Kara i?' e,?:
>>> On Tue 22-07-14 09:30:05, Peter Zijlstra wrote:
>>>> On Tue, Jul 22, 2014 at 02:18:47PM +0900, Gioh Kim wrote:
>>>>> Hello,
>>>>>
>>>>> This patch try to solve problem that a long-lasting page cache of
>>>>> ext4 superblock disturbs page migration.
>>>>>
>>>>> I've been testing CMA feature on my ARM-based platform
>>>>> and found some pages for page caches cannot be migrated.
>>>>> Some of them are page caches of superblock of ext4 filesystem.
>>>>>
>>>>> Current ext4 reads superblock with sb_bread(). sb_bread() allocates page
>>>> >from movable area. But the problem is that ext4 hold the page until
>>>>> it is unmounted. If root filesystem is ext4 the page cannot be migrated forever.
>>>>>
>>>>> I introduce a new API for allocating page from non-movable area.
>>>>> It is useful for ext4 and others that want to hold page cache for a long time.
>>>>
>>>> There's no word on why you can't teach ext4 to still migrate that page.
>>>> For all I know it might be impossible, but at least mention why.
>>
>> I am very sorry for lacking of details.
>>
>> In ext4_fill_super() the buffer-head of superblock is stored in sbi->s_sbh.
>> The page belongs to the buffer-head is allocated from movable area.
>> To migrate the page the buffer-head should be released via brelse().
>> But brelse() is not called until unmount.
>    Hum, I don't see where in the code do we check buffer_head use count. Can
> you please point me? Thanks.

Filesystem code does not check buffer_head use count.
sb_bread() returns the buffer_head that is included in bh_lru and has non-zero use count.
You can see the bh_lru code in buffer.c: __find_get_clock() and lookup_bh_lru().
bh_lru_install() inserts the buffer_head into the bh_lru().
It first calls get_bh() to increase the use count and insert bh into the lru array.

The buffer_head use count is non-zero until brelse() is called.

>
>>>    It doesn't seem to be worth the effort to make that page movable to me
>>> (it's reasonably doable since superblock buffer isn't accessed in *that*
>>> many places but single movable page doesn't seem like a good tradeoff for
>>> the complexity).
>>>
>>> But this made me look into the migration code and it isn't completely clear
>>> to me what makes the migration code decide that sb buffer isn't movable? We
>>> seem to be locking the buffers before moving the underlying page but we
>>> don't do any reference or state checks on the buffers... That seems to be
>>> assuming that noone looks at bh->b_data without holding buffer lock. That
>>> is likely true for ordinary data but definitely not true for metadata
>>> buffers (i.e., buffers for pages from block device mappings).
>>
>> The sb buffer is not movable because it is not released.
>> sb_bread increase the reference counter of buffer-head so that
>> the page of the buffer-head cannot be movable.
>>
>> sb_bread allocates page from movable area but it is not movable until the
>> reference counter of the buffer-head becomes zero.
>> There is no lock for the buffer but the reference counter acts like lock.
>    OK, but why do you care about a single page (of at most handful if you
> have more filesystems) which isn't movable? That shouldn't make a big
> difference to compaction...

Even a single page can make CMA migration fail.

>
>> Actually it is strange that ext4 keeps buffer-head in superblock
>> structure until unmount (it can be long time) I thinks the buffer-head
>> should be released immediately like fat_fill_super() did.  I believe
>> there is a reason to keep buffer-head so that I suggest this patch.
>    We don't copy some data from the superblock to other structure so from
> time to time we need to look e.g. at feature bits within superblock buffer.
> Historically we were updating numbers of free blocks and inodes in the
> superblock with each allocation but we don't do that anymore because it
> scales poorly. So there is no fundamental reason for keeping sb buffer
> pinned anymore. Just someone would have to rewrite the code to copy some
> pieces of data from the buffer to some other structure and use it there.

I hope so.

>
> 								Honza
>

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>