Skip to contentSkip to conversation

mm: migrate: don't rely on PageMovable() of newpage after unlocking it

RFC PATCH In review David Hildenbrand 3 participants 11 messages Jan 28, 2019

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

RFC PATCH 1/1 mm: migrate: don't rely on PageMovable() of newpage after unlocking it

While debugging some crashes related to virtio-balloon deflation that happened under the old balloon migration code, I stumbled over a possible race I think.

What we experienced:

drivers/virtio/virtio_balloon.c:release_pages_balloon():
- WARNING: CPU: 13 PID: 6586 at lib/list_debug.c:59 __list_del_entry+0xa1/0xd0
- list_del corruption. prev->next should be ffffe253961090a0, but was dead000000000100

Turns out after having added the page to a local list when dequeuing, the page would suddenly be moved to an LRU list before we would free it via the local list, corrupting both lists. So a page we own and that is !LRU was moved to an LRU list.

For us, this was triggered by backporting 195a8c43e93d8 ("virtio-balloon:
deflate via a page list") onto old balloon compaction code, but I think this only made the BUG become visible and the race still exists in new !LRU migraton code.

My theory:

In __unmap_and_move(), we lock the old and newpage and perform the migration. In case of vitio-balloon, the new page will become movable, the old page will no longer be movable.

However, after unlocking newpage, I think there is nothing stopping the newpage from getting dequeued and freed by virtio-balloon. This will result in the newpage
1. No longer having PageMovable()
2. Getting moved to the local list before finally freeing it (using page->lru)

Back in the migration thread in __unmap_and_move(), we would after unlocking the newpage suddenly no longer have PageMovable(newpage) and will therefore call putback_lru_page(newpage), modifying page->lru although that list is still in use by virtio-balloon.

To summarize, we have a race between migrating the newpage and checking for PageMovable(newpage). Instead of checking __PageMovable(newpage), we can simply rely on is_lru. Because if the old page was not an LRU page, the new page also shouldn't be. If there is a problem with that, we could check for __PageMovable(newpage) before unlocking the new page.

As I am not yet sure if this race actually exists, especially also upstream, I am sending this as RFC.

Cc Andrew Morton <[email protected]>
Cc Mel Gorman <[email protected]>
Cc "Kirill A. Shutemov" <[email protected]>
Cc Michal Hocko <[email protected]>
Cc Naoya Horiguchi <[email protected]>
Cc Jan Kara <[email protected]>
Cc Andrea Arcangeli <[email protected]>
Cc Dominik Brodowski <[email protected]>
Cc Matthew Wilcox <[email protected]>
Cc Vratislav Bendel <[email protected]>
Cc Rafael Aquini <[email protected]>
---
 mm/migrate.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
mm/migrate.c +4 -2
index 4512afab46ac..31e002270b05 100644
@@ -1135,10 +1135,12 @@ static int __unmap_and_move(struct page *page, struct page *newpage,
 	 * If migration is successful, decrease refcount of the newpage
 	 * which will not free the page because new page owner increased
 	 * refcounter. As well, if it is LRU page, add the page to LRU
-	 * list in here.
+	 * list in here. Don't rely on PageMovable(newpage), as that could
+	 * already have changed after unlocking newpage (e.g.
+	 * virtio-balloon deflation).
 	 */
 	if (rc == MIGRATEPAGE_SUCCESS) {
-		if (unlikely(__PageMovable(newpage)))
+		if (unlikely(!is_lru))
 			put_page(newpage);
 		else
 			putback_lru_page(newpage);
Signature · 1 line
2.17.2
#1 lore raw

On Mon 28-01-19 13:16:09, David Hildenbrand wrote:
[...]

Quoted 6 lines

My theory:

In __unmap_and_move(), we lock the old and newpage and perform the migration. In case of vitio-balloon, the new page will become movable, the old page will no longer be movable.

Quoted 6 lines

However, after unlocking newpage, I think there is nothing stopping the newpage from getting dequeued and freed by virtio-balloon. This will result in the newpage
1. No longer having PageMovable()
2. Getting moved to the local list before finally freeing it (using page->lru)

Does that mean that the virtio-balloon can change the Movable state while there are other users of the page? Can you point to the code that does it? How come this can be safe at all? Or is the PageMovable stable only under the page lock?

Signature · 2 lines
Michal Hocko
SUSE Labs
#2 lore raw
Quoted 15 lines · Michal Hocko

On Mon 28-01-19 13:16:09, David Hildenbrand wrote:
[...]

Quoted 6 lines

My theory:

In __unmap_and_move(), we lock the old and newpage and perform the migration. In case of vitio-balloon, the new page will become movable, the old page will no longer be movable.

Quoted 6 lines

However, after unlocking newpage, I think there is nothing stopping the newpage from getting dequeued and freed by virtio-balloon. This will result in the newpage
1. No longer having PageMovable()
2. Getting moved to the local list before finally freeing it (using page->lru)

Quoted 5 lines

Does that mean that the virtio-balloon can change the Movable state while there are other users of the page? Can you point to the code that does it? How come this can be safe at all? Or is the PageMovable stable only under the page lock?

PageMovable is stable under the lock. The relevant instructions are in

mm/balloon_compaction.c and include/linux/balloon_compaction.h

balloon_page_insert() and balloon_page_delete() modify PageMovable and are only called with both, the page locked and b_dev_info->pages_lock locked.

Especially:
1. balloon_page_dequeue() which tries to find an unlocked page that is not isolated. and deletes it via balloon_page_delete().

2. balloon_page_migrate() will effectively call balloon_page_delete() on old page and balloon_page_insert() newpage. Both pages are locked before calling balloon_page_migrate() and the b_dev_info->pages_lock is taken internally.

So what's left is a very small race window when migrating to a new page, after dropping the lock of the newpage.

Signature · 3 lines
Thanks,

David / dhildenb
#3 lore raw
Quoted 22 lines · David Hildenbrand
Quoted 15 lines · Michal Hocko

On Mon 28-01-19 13:16:09, David Hildenbrand wrote:
[...]

Quoted 6 lines

My theory:

In __unmap_and_move(), we lock the old and newpage and perform the migration. In case of vitio-balloon, the new page will become movable, the old page will no longer be movable.

Quoted 6 lines

However, after unlocking newpage, I think there is nothing stopping the newpage from getting dequeued and freed by virtio-balloon. This will result in the newpage
1. No longer having PageMovable()
2. Getting moved to the local list before finally freeing it (using page->lru)

Quoted 5 lines

Does that mean that the virtio-balloon can change the Movable state while there are other users of the page? Can you point to the code that does it? How come this can be safe at all? Or is the PageMovable stable only under the page lock?

Quoted 3 lines

PageMovable is stable under the lock. The relevant instructions are in

mm/balloon_compaction.c and include/linux/balloon_compaction.h

OK, I have just checked __ClearPageMovable and it indeed requires PageLock. Then we also have to move is_lru = __PageMovable(page) after the page lock.

Signature · 2 lines
Michal Hocko
SUSE Labs
#4 lore raw
Quoted 23 lines · Michal Hocko
Quoted 22 lines · David Hildenbrand
Quoted 20 lines · Michal Hocko

On Mon 28-01-19 13:16:09, David Hildenbrand wrote:
[...]

Quoted 6 lines

My theory:

In __unmap_and_move(), we lock the old and newpage and perform the migration. In case of vitio-balloon, the new page will become movable, the old page will no longer be movable.

Quoted 6 lines

However, after unlocking newpage, I think there is nothing stopping the newpage from getting dequeued and freed by virtio-balloon. This will result in the newpage
1. No longer having PageMovable()
2. Getting moved to the local list before finally freeing it (using page->lru)

Does that mean that the virtio-balloon can change the Movable state while there are other users of the page? Can you point to the code that does it? How come this can be safe at all? Or is the PageMovable stable only under the page lock?

Quoted 8 lines
Quoted 3 lines

PageMovable is stable under the lock. The relevant instructions are in

mm/balloon_compaction.c and include/linux/balloon_compaction.h

OK, I have just checked __ClearPageMovable and it indeed requires PageLock. Then we also have to move is_lru = __PageMovable(page) after the page lock.

I assume that is fine as is as the page is isolated? (yes, it will be modified later when moving but we are interested in the original state)

Signature · 3 lines
Thanks,

David / dhildenb
#5 lore raw
Quoted 28 lines · David Hildenbrand
Quoted 27 lines · Michal Hocko
Quoted 25 lines · David Hildenbrand
Quoted 15 lines · Michal Hocko

On Mon 28-01-19 13:16:09, David Hildenbrand wrote:
[...]

Quoted 6 lines

My theory:

In __unmap_and_move(), we lock the old and newpage and perform the migration. In case of vitio-balloon, the new page will become movable, the old page will no longer be movable.

Quoted 6 lines

However, after unlocking newpage, I think there is nothing stopping the newpage from getting dequeued and freed by virtio-balloon. This will result in the newpage
1. No longer having PageMovable()
2. Getting moved to the local list before finally freeing it (using page->lru)

Quoted 5 lines

Does that mean that the virtio-balloon can change the Movable state while there are other users of the page? Can you point to the code that does it? How come this can be safe at all? Or is the PageMovable stable only under the page lock?

PageMovable is stable under the lock. The relevant instructions are in

mm/balloon_compaction.c and include/linux/balloon_compaction.h

Quoted 7 lines
Quoted 4 lines

OK, I have just checked __ClearPageMovable and it indeed requires PageLock. Then we also have to move is_lru = __PageMovable(page) after the page lock.

I assume that is fine as is as the page is isolated? (yes, it will be modified later when moving but we are interested in the original state)

OK, I've missed that the page is indeed isolated. Then the patch makes sense to me.

Signature · 2 lines
Michal Hocko
SUSE Labs
#6 lore raw
Quoted 34 lines · Michal Hocko
Quoted 33 lines · David Hildenbrand
Quoted 31 lines · Michal Hocko
Quoted 22 lines · David Hildenbrand
Quoted 15 lines · Michal Hocko

On Mon 28-01-19 13:16:09, David Hildenbrand wrote:
[...]

Quoted 6 lines

My theory:

In __unmap_and_move(), we lock the old and newpage and perform the migration. In case of vitio-balloon, the new page will become movable, the old page will no longer be movable.

Quoted 6 lines

However, after unlocking newpage, I think there is nothing stopping the newpage from getting dequeued and freed by virtio-balloon. This will result in the newpage
1. No longer having PageMovable()
2. Getting moved to the local list before finally freeing it (using page->lru)

Quoted 5 lines

Does that mean that the virtio-balloon can change the Movable state while there are other users of the page? Can you point to the code that does it? How come this can be safe at all? Or is the PageMovable stable only under the page lock?

Quoted 3 lines

PageMovable is stable under the lock. The relevant instructions are in

mm/balloon_compaction.c and include/linux/balloon_compaction.h

OK, I have just checked __ClearPageMovable and it indeed requires PageLock. Then we also have to move is_lru = __PageMovable(page) after the page lock.

Quoted 6 lines
Quoted 2 lines

I assume that is fine as is as the page is isolated? (yes, it will be modified later when moving but we are interested in the original state)

OK, I've missed that the page is indeed isolated. Then the patch makes sense to me.

Thanks Michal. I assume this has broken ever since balloon compaction was introduced. I'll wait a little more and then resend as !RFC with a cc-stable tag.

Signature · 3 lines
Thanks,

David / dhildenb
Acked-by
#7 lore raw
Quoted 38 lines · David Hildenbrand
Quoted 37 lines · Michal Hocko
Quoted 35 lines · David Hildenbrand
Quoted 23 lines · Michal Hocko
Quoted 22 lines · David Hildenbrand
Quoted 20 lines · Michal Hocko

On Mon 28-01-19 13:16:09, David Hildenbrand wrote:
[...]

Quoted 6 lines

My theory:

In __unmap_and_move(), we lock the old and newpage and perform the migration. In case of vitio-balloon, the new page will become movable, the old page will no longer be movable.

Quoted 6 lines

However, after unlocking newpage, I think there is nothing stopping the newpage from getting dequeued and freed by virtio-balloon. This will result in the newpage
1. No longer having PageMovable()
2. Getting moved to the local list before finally freeing it (using page->lru)

Does that mean that the virtio-balloon can change the Movable state while there are other users of the page? Can you point to the code that does it? How come this can be safe at all? Or is the PageMovable stable only under the page lock?

Quoted 8 lines
Quoted 3 lines

PageMovable is stable under the lock. The relevant instructions are in

mm/balloon_compaction.c and include/linux/balloon_compaction.h

OK, I have just checked __ClearPageMovable and it indeed requires PageLock. Then we also have to move is_lru = __PageMovable(page) after the page lock.

I assume that is fine as is as the page is isolated? (yes, it will be modified later when moving but we are interested in the original state)

Quoted 7 lines
Quoted 3 lines

OK, I've missed that the page is indeed isolated. Then the patch makes sense to me.

Thanks Michal. I assume this has broken ever since balloon compaction was introduced. I'll wait a little more and then resend as !RFC with a cc-stable tag.

Please make sure to CC Minchan when reposting.

Btw.

Signature · 2 lines
Michal Hocko
SUSE Labs
#8 lore raw
In reply to David Hildenbrand
Quoted 38 lines · David Hildenbrand
Quoted 37 lines · Michal Hocko
Quoted 35 lines · David Hildenbrand
Quoted 23 lines · Michal Hocko
Quoted 22 lines · David Hildenbrand
Quoted 20 lines · Michal Hocko

On Mon 28-01-19 13:16:09, David Hildenbrand wrote:
[...]

Quoted 6 lines

My theory:

In __unmap_and_move(), we lock the old and newpage and perform the migration. In case of vitio-balloon, the new page will become movable, the old page will no longer be movable.

Quoted 6 lines

However, after unlocking newpage, I think there is nothing stopping the newpage from getting dequeued and freed by virtio-balloon. This will result in the newpage
1. No longer having PageMovable()
2. Getting moved to the local list before finally freeing it (using page->lru)

Does that mean that the virtio-balloon can change the Movable state while there are other users of the page? Can you point to the code that does it? How come this can be safe at all? Or is the PageMovable stable only under the page lock?

Quoted 8 lines
Quoted 3 lines

PageMovable is stable under the lock. The relevant instructions are in

mm/balloon_compaction.c and include/linux/balloon_compaction.h

OK, I have just checked __ClearPageMovable and it indeed requires PageLock. Then we also have to move is_lru = __PageMovable(page) after the page lock.

I assume that is fine as is as the page is isolated? (yes, it will be modified later when moving but we are interested in the original state)

Quoted 8 lines
Quoted 3 lines

OK, I've missed that the page is indeed isolated. Then the patch makes sense to me.

Thanks Michal. I assume this has broken ever since balloon compaction was introduced. I'll wait a little more and then resend as !RFC with a cc-stable tag.

Yes, balloon deflation could always race against migration This race was a problem, initially, and was dealt with, via:

commit 117aad1e9e4d97448d1df3f84b08bd65811e6d6a

Author: Rafael Aquini <[email protected]>
Date:   Mon Sep 30 13:45:16 2013 -0700

mm: avoid reinserting isolated balloon pages into LRU lists

I think this upstream patch has re-introduced it, in a more subtle way, as we're stumbling on it now, again:

commit d6d86c0a7f8ddc5b38cf089222cb1d9540762dc2

Author: Konstantin Khlebnikov <[email protected]>
Date:   Thu Oct 9 15:29:27 2014 -0700

mm/balloon_compaction: redesign ballooned pages management

On this particular race against migration case, virtio ballon deflation would not see it before

commit b1123ea6d3b3da25af5c8a9d843bd07ab63213f4

Author: Minchan Kim <[email protected]>
Date:   Tue Jul 26 15:23:09 2016 -0700

mm: balloon: use general non-lru movable page feature

as the recently released balloon page would be post-processed without the page->lru list handling, which for migration stability purposes must be done under the protection of page_lock.

get rid of balloon reference count.

-- Rafael

Acked-by
#9 lore raw
Quoted 74 lines · Rafael Aquini
Quoted 38 lines · David Hildenbrand
Quoted 37 lines · Michal Hocko
Quoted 35 lines · David Hildenbrand
Quoted 23 lines · Michal Hocko
Quoted 22 lines · David Hildenbrand
Quoted 20 lines · Michal Hocko

On Mon 28-01-19 13:16:09, David Hildenbrand wrote:
[...]

Quoted 6 lines

My theory:

In __unmap_and_move(), we lock the old and newpage and perform the migration. In case of vitio-balloon, the new page will become movable, the old page will no longer be movable.

Quoted 6 lines

However, after unlocking newpage, I think there is nothing stopping the newpage from getting dequeued and freed by virtio-balloon. This will result in the newpage
1. No longer having PageMovable()
2. Getting moved to the local list before finally freeing it (using page->lru)

Does that mean that the virtio-balloon can change the Movable state while there are other users of the page? Can you point to the code that does it? How come this can be safe at all? Or is the PageMovable stable only under the page lock?

Quoted 8 lines
Quoted 3 lines

PageMovable is stable under the lock. The relevant instructions are in

mm/balloon_compaction.c and include/linux/balloon_compaction.h

OK, I have just checked __ClearPageMovable and it indeed requires PageLock. Then we also have to move is_lru = __PageMovable(page) after the page lock.

I assume that is fine as is as the page is isolated? (yes, it will be modified later when moving but we are interested in the original state)

Quoted 8 lines
Quoted 3 lines

OK, I've missed that the page is indeed isolated. Then the patch makes sense to me.

Thanks Michal. I assume this has broken ever since balloon compaction was introduced. I'll wait a little more and then resend as !RFC with a cc-stable tag.

Yes, balloon deflation could always race against migration This race was a problem, initially, and was dealt with, via:

commit 117aad1e9e4d97448d1df3f84b08bd65811e6d6a

Author: Rafael Aquini <[email protected]>
Date:   Mon Sep 30 13:45:16 2013 -0700

mm: avoid reinserting isolated balloon pages into LRU lists

I think this upstream patch has re-introduced it, in a more subtle way, as we're stumbling on it now, again:

commit d6d86c0a7f8ddc5b38cf089222cb1d9540762dc2

Author: Konstantin Khlebnikov <[email protected]>
Date:   Thu Oct 9 15:29:27 2014 -0700

mm/balloon_compaction: redesign ballooned pages management

On this particular race against migration case, virtio ballon deflation would not see it before

commit b1123ea6d3b3da25af5c8a9d843bd07ab63213f4

Quoted 10 lines
Author: Minchan Kim <[email protected]>
Date:   Tue Jul 26 15:23:09 2016 -0700

mm: balloon: use general non-lru movable page feature

as the recently released balloon page would be post-processed without the page->lru list handling, which for migration stability purposes must be done under the protection of page_lock.

missing part here:

I think your patch adresses this new case.

Quoted 1 line

get rid of balloon reference count.

^^ this was a left over (sorry about my fat-fingers)

Quoted 3 lines

-- Rafael

#10 lore raw
Quoted 91 lines · Rafael Aquini
Quoted 74 lines · Rafael Aquini
Quoted 38 lines · David Hildenbrand
Quoted 37 lines · Michal Hocko
Quoted 35 lines · David Hildenbrand
Quoted 23 lines · Michal Hocko
Quoted 22 lines · David Hildenbrand
Quoted 20 lines · Michal Hocko

On Mon 28-01-19 13:16:09, David Hildenbrand wrote:
[...]

Quoted 6 lines

My theory:

In __unmap_and_move(), we lock the old and newpage and perform the migration. In case of vitio-balloon, the new page will become movable, the old page will no longer be movable.

Quoted 6 lines

However, after unlocking newpage, I think there is nothing stopping the newpage from getting dequeued and freed by virtio-balloon. This will result in the newpage
1. No longer having PageMovable()
2. Getting moved to the local list before finally freeing it (using page->lru)

Does that mean that the virtio-balloon can change the Movable state while there are other users of the page? Can you point to the code that does it? How come this can be safe at all? Or is the PageMovable stable only under the page lock?

Quoted 8 lines
Quoted 3 lines

PageMovable is stable under the lock. The relevant instructions are in

mm/balloon_compaction.c and include/linux/balloon_compaction.h

OK, I have just checked __ClearPageMovable and it indeed requires PageLock. Then we also have to move is_lru = __PageMovable(page) after the page lock.

I assume that is fine as is as the page is isolated? (yes, it will be modified later when moving but we are interested in the original state)

Quoted 8 lines
Quoted 3 lines

OK, I've missed that the page is indeed isolated. Then the patch makes sense to me.

Thanks Michal. I assume this has broken ever since balloon compaction was introduced. I'll wait a little more and then resend as !RFC with a cc-stable tag.

Yes, balloon deflation could always race against migration This race was a problem, initially, and was dealt with, via:

commit 117aad1e9e4d97448d1df3f84b08bd65811e6d6a

Author: Rafael Aquini <[email protected]>
Date:   Mon Sep 30 13:45:16 2013 -0700

mm: avoid reinserting isolated balloon pages into LRU lists

I think this upstream patch has re-introduced it, in a more subtle way, as we're stumbling on it now, again:

commit d6d86c0a7f8ddc5b38cf089222cb1d9540762dc2

Author: Konstantin Khlebnikov <[email protected]>
Date:   Thu Oct 9 15:29:27 2014 -0700

mm/balloon_compaction: redesign ballooned pages management

On this particular race against migration case, virtio ballon deflation would not see it before

commit b1123ea6d3b3da25af5c8a9d843bd07ab63213f4

Quoted 10 lines
Author: Minchan Kim <[email protected]>
Date:   Tue Jul 26 15:23:09 2016 -0700

mm: balloon: use general non-lru movable page feature

as the recently released balloon page would be post-processed without the page->lru list handling, which for migration stability purposes must be done under the protection of page_lock.

missing part here:

I think your patch adresses this new case.

Quoted 6 lines
Quoted 1 line

get rid of balloon reference count.

^^ this was a left over (sorry about my fat-fingers)

:D

Thanks! I'll resend with

Fixes d6d86c0a7f8d ("mm/balloon_compaction: redesign ballooned pages

management")

Signature · 3 lines
Thanks,

David / dhildenb