new API to allocate buffer-cache for superblock in non-movable area

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.

Quoted 2 lines

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.

Gioh,

As follow up, if you want some further discussions about why these patches should be accepted, it would be good to get some hard data about why the keeping the ext4 superblock pinned is causing such a problem for page migation. Can you give us more details about what the impact is of not having these patches? And how it compres to other data structures which are currently allocated in the moveable area and tend to be pinned effectively indefinitely?

Thanks,

- Ted

Quoted 13 lines · Gioh Kim

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.

Quoted 2 lines

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.

Quoted 14 lines · Gioh Kim

Quoted 13 lines

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.

Quoted 5 lines

Quoted 2 lines

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.

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).

Added linux-mm to CC to enlighten me a bit ;)

Honza

2014-07-22 오후 6:38, Jan Kara 쓴 글:

Quoted 18 lines · Peter Zijlstra

Quoted 17 lines · Gioh Kim

Quoted 15 lines

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.

Quoted 7 lines

Quoted 7 lines

Quoted 3 lines

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.

For example, fat_fill_super() reads superblock via sb_bread() and release it via brelse() immediately. Therefore the page that stores superblock can be migrated.

Quoted 14 lines

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.

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.

Quoted 3 lines

Added linux-mm to CC to enlighten me a bit ;)

Quoted 2 lines

Honza

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

2014-07-22 오후 6:38, Jan Kara 쓴 글:

Quoted 15 lines · Peter Zijlstra

Quoted 14 lines · Gioh Kim

Quoted 13 lines

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.

Quoted 5 lines

Quoted 5 lines

Quoted 2 lines

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.

Quoted 5 lines

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).

Quoted 7 lines

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.

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.

Quoted 3 lines

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...

Quoted 6 lines

Quoted 3 lines

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...

The thing is, CMA _must_ be able to clear all the pages in its range, otherwise its broken.

So placing nonmovable pages in a movable block utterly wrecks that.

Now, Ted said that there's more effectively pinned stuff from filesystems (and I imagine those would be things like the root inode etc.) and those would equally wreck this..

But Gioh didn't mention any of that.. he should I suppose.

Quoted 22 lines

2014-07-22 오후 6:38, Jan Kara 쓴 글:

Quoted 20 lines · Peter Zijlstra

Quoted 19 lines · Gioh Kim

Quoted 9 lines

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.

Quoted 6 lines

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.

Quoted 6 lines

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.

Quoted 18 lines

Quoted 6 lines

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).

Quoted 7 lines

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.

Quoted 3 lines

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...

Quoted 5 lines

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.

Honza

2014-07-30 오후 7:11, Jan Kara 쓴 글:

Quoted 23 lines · Gioh Kim

Quoted 22 lines

2014-07-22 오후 6:38, Jan Kara 쓴 글:

Quoted 20 lines · Peter Zijlstra

Quoted 19 lines · Gioh Kim

Quoted 9 lines

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.

Quoted 6 lines

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.

Quoted 8 lines

Quoted 6 lines

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.

Quoted 33 lines

2014-07-30 오후 7:11, Jan Kara 쓴 글:

Quoted 23 lines · Gioh Kim

Quoted 22 lines

2014-07-22 오후 6:38, Jan Kara 쓴 글:

Quoted 20 lines · Peter Zijlstra

Quoted 19 lines · Gioh Kim

Quoted 9 lines

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.

Quoted 6 lines

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.

Quoted 8 lines

Quoted 6 lines

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.

Quoted 8 lines

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.

So I probably didn't phrase the question precisely enough. What I was
asking about is where exactly *migration* code checks buffer use count? Because as I'm looking at buffer_migrate_page() we lock the buffers on a migrated page but we don't look at buffer use counts... So it seems to me that migration of a page with buffers should succeed even if buffer head has an elevated use count. Now I think that it *should* check the buffer use counts (it is dangerous to migrate buffers someone holds reference to) but I just cannot find that place. Or does CMA use some other migration function for buffer pages than buffer_migrate_page()?

Honza

2014-07-31 오전 9:03, Jan Kara 쓴 글:

Quoted 42 lines · Gioh Kim

Quoted 33 lines

2014-07-30 오후 7:11, Jan Kara 쓴 글:

Quoted 23 lines · Gioh Kim

Quoted 22 lines

2014-07-22 오후 6:38, Jan Kara 쓴 글:

Quoted 20 lines · Peter Zijlstra

Quoted 19 lines · Gioh Kim

Quoted 9 lines

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.

Quoted 6 lines

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.

Quoted 8 lines

Quoted 6 lines

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.

Quoted 8 lines

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.

Quoted 9 lines

So I probably didn't phrase the question precisely enough. What I was
asking about is where exactly *migration* code checks buffer use count? Because as I'm looking at buffer_migrate_page() we lock the buffers on a migrated page but we don't look at buffer use counts... So it seems to me that migration of a page with buffers should succeed even if buffer head has an elevated use count. Now I think that it *should* check the buffer use counts (it is dangerous to migrate buffers someone holds reference to) but I just cannot find that place. Or does CMA use some other migration function for buffer pages than buffer_migrate_page()?

CMA allocation function is cma_alloc().
Function flow is alloc_contig_range() -> __alloc_contig_migrate_range() -> migrate_pages -> unmap_and_move -> __unmap_and_move -> try_to_free_buffers -> drop_buffers -> buffer_busy.

The buffer_busy() is checking b_count.
If buffer is busy buffer-cache cannot be removed.
So the page that includes buffer_head and the page that is refered by buffer_head are not movable.

Is this what you need?

Quoted 64 lines

2014-07-31 오전 9:03, Jan Kara 쓴 글:

Quoted 42 lines · Gioh Kim

Quoted 33 lines

2014-07-30 오후 7:11, Jan Kara 쓴 글:

Quoted 23 lines · Gioh Kim

Quoted 22 lines

2014-07-22 오후 6:38, Jan Kara 쓴 글:

Quoted 20 lines · Peter Zijlstra

Quoted 19 lines · Gioh Kim

Quoted 9 lines

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.

Quoted 6 lines

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.

Quoted 8 lines

Quoted 6 lines

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.

Quoted 8 lines

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.

Quoted 9 lines

So I probably didn't phrase the question precisely enough. What I was
asking about is where exactly *migration* code checks buffer use count? Because as I'm looking at buffer_migrate_page() we lock the buffers on a migrated page but we don't look at buffer use counts... So it seems to me that migration of a page with buffers should succeed even if buffer head has an elevated use count. Now I think that it *should* check the buffer use counts (it is dangerous to migrate buffers someone holds reference to) but I just cannot find that place. Or does CMA use some other migration function for buffer pages than buffer_migrate_page()?

CMA allocation function is cma_alloc().
Function flow is alloc_contig_range() -> __alloc_contig_migrate_range() -> migrate_pages -> unmap_and_move -> __unmap_and_move -> try_to_free_buffers -> drop_buffers -> buffer_busy.

The buffer_busy() is checking b_count.
If buffer is busy buffer-cache cannot be removed.
So the page that includes buffer_head and the page that is refered by buffer_head are not movable.

Quoted 1 line

Is this what you need?

Yes, this is what I was asking about. Thanks! But as I'm looking into
__unmap_and_move() it calls try_to_free_buffers() only if page->mapping == NULL. As the comment before that test states, this can happen only for swap cache (not our case) or for pagecache pages that were truncated and not yet fully cleaned up. But superblock page cannot really be truncated. So I somewhat doubt you can hit the above path for a page holding superblock...

Honza

2014-07-31 오후 9:21, Jan Kara 쓴 글:

Quoted 65 lines · Gioh Kim

Quoted 64 lines

2014-07-31 오전 9:03, Jan Kara 쓴 글:

Quoted 42 lines · Gioh Kim

Quoted 33 lines

2014-07-30 오후 7:11, Jan Kara 쓴 글:

Quoted 23 lines · Gioh Kim

Quoted 22 lines

2014-07-22 오후 6:38, Jan Kara 쓴 글:

Quoted 20 lines · Peter Zijlstra

Quoted 19 lines · Gioh Kim

Quoted 9 lines

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.

Quoted 6 lines

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.

Quoted 8 lines

Quoted 6 lines

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.

Quoted 8 lines

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.

Quoted 9 lines

So I probably didn't phrase the question precisely enough. What I was
asking about is where exactly *migration* code checks buffer use count? Because as I'm looking at buffer_migrate_page() we lock the buffers on a migrated page but we don't look at buffer use counts... So it seems to me that migration of a page with buffers should succeed even if buffer head has an elevated use count. Now I think that it *should* check the buffer use counts (it is dangerous to migrate buffers someone holds reference to) but I just cannot find that place. Or does CMA use some other migration function for buffer pages than buffer_migrate_page()?

CMA allocation function is cma_alloc().
Function flow is alloc_contig_range() -> __alloc_contig_migrate_range() -> migrate_pages -> unmap_and_move -> __unmap_and_move -> try_to_free_buffers -> drop_buffers -> buffer_busy.

The buffer_busy() is checking b_count.
If buffer is busy buffer-cache cannot be removed.
So the page that includes buffer_head and the page that is refered by buffer_head are not movable.

Quoted 7 lines

Quoted 1 line

Is this what you need?

Yes, this is what I was asking about. Thanks! But as I'm looking into
__unmap_and_move() it calls try_to_free_buffers() only if page->mapping == NULL. As the comment before that test states, this can happen only for swap cache (not our case) or for pagecache pages that were truncated and not yet fully cleaned up. But superblock page cannot really be truncated. So I somewhat doubt you can hit the above path for a page holding superblock...

I printed the address of busy buffer_head in drop_buffers() that is called by try_to_free_buffers(). And I printed the address of sb buffer_head.
They were the same.

I'm going to check page->mapping.

Quoted 5 lines

Honza

Function path is like followings:

[   97.868304] [<8011a750>] (drop_buffers+0xfc/0x168) from [<8011bc64>] (try_to_free_buffers+0x50/0xbc)
[   97.877457] [<8011bc64>] (try_to_free_buffers+0x50/0xbc) from [<80121e40>] (blkdev_releasepage+0x38/0x48)
[   97.887093] [<80121e40>] (blkdev_releasepage+0x38/0x48) from [<800add8c>] (try_to_release_page+0x40/0x5c)
[   97.896728] [<800add8c>] (try_to_release_page+0x40/0x5c) from [<800bd9bc>] (shrink_page_list+0x508/0x8a4)
[   97.906334] [<800bd9bc>] (shrink_page_list+0x508/0x8a4) from [<800bde5c>] (reclaim_clean_pages_from_list+0x104/0x148)
[   97.917017] [<800bde5c>] (reclaim_clean_pages_from_list+0x104/0x148) from [<800b5dec>] (alloc_contig_range+0x114/0x2dc)
[   97.927856] [<800b5dec>] (alloc_contig_range+0x114/0x2dc) from [<802f6c04>] (dma_alloc_from_contiguous+0x8c/0x14c)
[   97.938264] [<802f6c04>] (dma_alloc_from_contiguous+0x8c/0x14c) from [<80017b6c>] (__alloc_from_contiguous+0x34/0xc0)
[   97.948926] [<80017b6c>] (__alloc_from_contiguous+0x34/0xc0) from [<80017d40>] (__dma_alloc+0xc4/0x2a0)
[   97.958362] [<80017d40>] (__dma_alloc+0xc4/0x2a0) from [<8001803c>] (arm_dma_alloc+0x80/0x98)
[   97.966916] [<8001803c>] (arm_dma_alloc+0x80/0x98) from [<7f6ea188>] (cma_test_probe+0xe0/0x1f0 [drv])

So the quiestion really is how hard guarantee do you need that a page in movable zone is really movable. Or better in what timeframe should it be movable? It may be possible to make e.g. migratepage callback for ext4 blkdev pages which will handle migration of pages that are just idly sitting in a journal waiting to be committed. That may be reasonably doable although it won't be perfect. Or we may just decide it's not worth the bother and allocate all blkdev pages from unmovable zone...

Quoted 7 lines · Jan Kara

So the quiestion really is how hard guarantee do you need that a page in movable zone is really movable. Or better in what timeframe should it be movable? It may be possible to make e.g. migratepage callback for ext4 blkdev pages which will handle migration of pages that are just idly sitting in a journal waiting to be committed. That may be reasonably doable although it won't be perfect. Or we may just decide it's not worth the bother and allocate all blkdev pages from unmovable zone...

So the point of CMA is to cater to those (arguably broken) devices that do not have scatter gather IO, and these include things like the camera device on your phone.

Previously (and possibly currently) your android Linux kernel will simply preallocate a massive physically linear chunk of memory and assign it to the camera hardware and not use it at all.

This is a terrible waste for most of the time people aren't running their camera app at all. So the point is to allow usage of the memory, but upon request be able to 'immediately' clear it through migration/writeback.

So we should be fairly 'quick' in making the memory available, definitely sub second timeframes.

Sadly its not only mobile devices that excel in crappy hardware, there's plenty desktop stuff that could use this too, like some of the v4l devices iirc.

OK, makes sense. But then if there's heavy IO going on, anything that has
IO pending on it is pinned and IO completion can easily take something close to a second or more. So meeting subsecond deadlines may be tough even for ordinary data pages under heavy load, even more so for metadata where there are further constraints. OTOH phones aren't usually IO bound so in practice it needn't be so bad ;).

So if it is sub-second unless someone
loads the storage, then that sounds doable even for metadata. But we'll need to attach ->migratepage callback to blkdev pages and at least in ext4 case teach it how to move pages tracked by the journal.

Quoted 3 lines

Sadly its not only mobile devices that excel in crappy hardware, there's plenty desktop stuff that could use this too, like some of the v4l devices iirc.

Yeah, but in such usecases the guarantees we can offer for completion of
migration are even more vague :(.

Quoted 7 lines

OK, makes sense. But then if there's heavy IO going on, anything that has
IO pending on it is pinned and IO completion can easily take something close to a second or more. So meeting subsecond deadlines may be tough even for ordinary data pages under heavy load, even more so for metadata where there are further constraints. OTOH phones aren't usually IO bound so in practice it needn't be so bad ;).

Yeah, typically phones are not IO bound :-)

Quoted 4 lines

So if it is sub-second unless someone
loads the storage, then that sounds doable even for metadata. But we'll need to attach ->migratepage callback to blkdev pages and at least in ext4 case teach it how to move pages tracked by the journal.

Right, making it possible at all if of course much prefered over not possible, regardless of timeliness :-)

Quoted 5 lines

Quoted 3 lines

Sadly its not only mobile devices that excel in crappy hardware, there's plenty desktop stuff that could use this too, like some of the v4l devices iirc.

Yeah, but in such usecases the guarantees we can offer for completion of
migration are even more vague :(.

Yeah, lets start by making it possible, after that we can maybe look at making it better, who knows.

Quoted 64 lines · Gioh Kim

2014-07-31 i??i ? 9:03, Jan Kara i?' e,?:

Quoted 42 lines · Gioh Kim

Quoted 33 lines

2014-07-30 i??i?? 7:11, Jan Kara i?' e,?:

Quoted 23 lines · Gioh Kim

Quoted 22 lines

2014-07-22 i??i?? 6:38, Jan Kara i?' e,?:

Quoted 20 lines · Peter Zijlstra

Quoted 19 lines · Gioh Kim

Quoted 9 lines

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.

Quoted 6 lines

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.

Quoted 8 lines

Quoted 6 lines

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.

Quoted 8 lines

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.

Quoted 9 lines

So I probably didn't phrase the question precisely enough. What I was
asking about is where exactly *migration* code checks buffer use count? Because as I'm looking at buffer_migrate_page() we lock the buffers on a migrated page but we don't look at buffer use counts... So it seems to me that migration of a page with buffers should succeed even if buffer head has an elevated use count. Now I think that it *should* check the buffer use counts (it is dangerous to migrate buffers someone holds reference to) but I just cannot find that place. Or does CMA use some other migration function for buffer pages than buffer_migrate_page()?

CMA allocation function is cma_alloc().
Function flow is alloc_contig_range() -> __alloc_contig_migrate_range() -> migrate_pages -> unmap_and_move -> __unmap_and_move -> try_to_free_buffers -> drop_buffers -> buffer_busy.

The buffer_busy() is checking b_count.
If buffer is busy buffer-cache cannot be removed.
So the page that includes buffer_head and the page that is refered by buffer_head are not movable.

Quoted 1 line

Is this what you need?

Yes, this is what I was asking about. Thanks! But as I'm looking into
__unmap_and_move() it calls try_to_free_buffers() only if page->mapping == NULL. As the comment before that test states, this can happen only for swap cache (not our case) or for pagecache pages that were truncated and not yet fully cleaned up. But superblock page cannot really be truncated. So I somewhat doubt you can hit the above path for a page holding superblock...

Quoted 65 lines · Jan Kara

Quoted 64 lines · Gioh Kim

2014-07-31 오전 9:03, Jan Kara 쓴 글:

Quoted 42 lines · Gioh Kim

Quoted 33 lines

2014-07-30 오후 7:11, Jan Kara 쓴 글:

Quoted 23 lines · Gioh Kim

Quoted 22 lines

2014-07-22 오후 6:38, Jan Kara 쓴 글:

Quoted 20 lines · Peter Zijlstra

Quoted 19 lines · Gioh Kim

Quoted 9 lines

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.

Quoted 6 lines

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.

Quoted 8 lines

Quoted 6 lines

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.

Quoted 8 lines

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.

Quoted 9 lines

So I probably didn't phrase the question precisely enough. What I was
asking about is where exactly *migration* code checks buffer use count? Because as I'm looking at buffer_migrate_page() we lock the buffers on a migrated page but we don't look at buffer use counts... So it seems to me that migration of a page with buffers should succeed even if buffer head has an elevated use count. Now I think that it *should* check the buffer use counts (it is dangerous to migrate buffers someone holds reference to) but I just cannot find that place. Or does CMA use some other migration function for buffer pages than buffer_migrate_page()?

CMA allocation function is cma_alloc().
Function flow is alloc_contig_range() -> __alloc_contig_migrate_range() -> migrate_pages -> unmap_and_move -> __unmap_and_move -> try_to_free_buffers -> drop_buffers -> buffer_busy.

The buffer_busy() is checking b_count.
If buffer is busy buffer-cache cannot be removed.
So the page that includes buffer_head and the page that is refered by buffer_head are not movable.

Quoted 7 lines

Quoted 1 line

Is this what you need?

Yes, this is what I was asking about. Thanks! But as I'm looking into
__unmap_and_move() it calls try_to_free_buffers() only if page->mapping == NULL. As the comment before that test states, this can happen only for swap cache (not our case) or for pagecache pages that were truncated and not yet fully cleaned up. But superblock page cannot really be truncated. So I somewhat doubt you can hit the above path for a page holding superblock...

Hello,

Although page->mapping != NULL, mapping->a_ops->migratepage could be NULL. This is the case of block_device. See def_blk_aops in fs/block_dev.c. In this case, fallback_migrate_page() is called and then try_to_release_page() and try_to_free_buffers() would be called.