Re: [PATCH] mm/compaction: avoid VM_BUG_ON(PageSlab()) in page_mapcount()

From: Hugh Dickins
Date: Sat May 23 2020 - 21:10:36 EST


On Wed, 13 May 2020, Konstantin Khlebnikov wrote:

> Function isolate_migratepages_block() runs some checks out of lru_lock
> when choose pages for migration. After checking PageLRU() it checks extra
> page references by comparing page_count() and page_mapcount(). Between
> these two checks page could be removed from lru, freed and taken by slab.
>
> As a result this race triggers VM_BUG_ON(PageSlab()) in page_mapcount().
> Race window is tiny. For certain workload this happens around once a year.

Around once a year, that was my guess too. I have no record of us ever
hitting this, but yes it could happen when you have CONFIG_DEBUG_VM=y
(which I too like to run with, but would not recommend for users).

>
>
> page:ffffea0105ca9380 count:1 mapcount:0 mapping:ffff88ff7712c180 index:0x0 compound_mapcount: 0
> flags: 0x500000000008100(slab|head)
> raw: 0500000000008100 dead000000000100 dead000000000200 ffff88ff7712c180
> raw: 0000000000000000 0000000080200020 00000001ffffffff 0000000000000000
> page dumped because: VM_BUG_ON_PAGE(PageSlab(page))
> ------------[ cut here ]------------
> kernel BUG at ./include/linux/mm.h:628!
> invalid opcode: 0000 [#1] SMP NOPTI
> CPU: 77 PID: 504 Comm: kcompactd1 Tainted: G W 4.19.109-27 #1
> Hardware name: Yandex T175-N41-Y3N/MY81-EX0-Y3N, BIOS R05 06/20/2019
> RIP: 0010:isolate_migratepages_block+0x986/0x9b0
>
>
> To fix just opencode page_mapcount() in racy check for 0-order case and
> recheck carefully under lru_lock when page cannot escape from lru.
>
> Also add checking extra references for file pages and swap cache.
>
> Signed-off-by: Konstantin Khlebnikov <khlebnikov@xxxxxxxxxxxxxx>
> Fixes: 119d6d59dcc0 ("mm, compaction: avoid isolating pinned pages")

Not really, that commit was correct at the time it went in.

> Fixes: 1d148e218a0d ("mm: add VM_BUG_ON_PAGE() to page_mapcount()")

Exactly, that commit was well-intentioned, but did not allow for this
(admittedly very exceptional) usage. How many developers actually
make the mistake of applying page_mapcount() to their slab pages?
None, I expect. That VM_BUG_ON_PAGE() is there for documentation,
and could just be replaced by a comment - and Linus would be happy
with that.

> ---
> mm/compaction.c | 17 +++++++++++++----
> 1 file changed, 13 insertions(+), 4 deletions(-)
>
> diff --git a/mm/compaction.c b/mm/compaction.c
> index 46f0fcc93081..91bb87fd9420 100644
> --- a/mm/compaction.c
> +++ b/mm/compaction.c
> @@ -935,12 +935,16 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
> }
>
> /*
> - * Migration will fail if an anonymous page is pinned in memory,
> + * Migration will fail if an page is pinned in memory,
> * so avoid taking lru_lock and isolating it unnecessarily in an
> - * admittedly racy check.
> + * admittedly racy check simplest case for 0-order pages.
> + *
> + * Open code page_mapcount() to avoid VM_BUG_ON(PageSlab(page)).

But open coding page_mapcount() is not all that you did. You have
(understandably) chosen to avoid calling page_mapping(page), but...

> + * Page could have extra reference from mapping or swap cache.
> */
> - if (!page_mapping(page) &&
> - page_count(page) > page_mapcount(page))
> + if (!PageCompound(page) &&
> + page_count(page) > atomic_read(&page->_mapcount) + 1 +
> + (!PageAnon(page) || PageSwapCache(page)))
> goto isolate_fail;

Isn't that test going to send all the file cache pages with buffer heads
in page->private, off to isolate_fail when they're actually great
candidates for migration?

Given that the actual bug spotted was with the VM_BUG_ON_PAGE(PageSlab),
and nobody has reported any crash from the use of page_mapping() there
(and we only need the test to be right most of the time: all of this
knowingly racy, as you explain in other mail): I'd go for just replacing
the VM_BUG_ON_PAGE in page_mapcount() by a comment about this case.

But if you think developers are really in danger of coding page_mapcount()
on their slab pages, then you could add a _page_mapcount() to linux/mm.h,
which omits the VM_BUG_ON_PAGE, for use here only.

Then we wouldn't have to think so hard about the counting above!

>
> /*
> @@ -975,6 +979,11 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
> low_pfn += compound_nr(page) - 1;
> goto isolate_fail;
> }
> +
> + /* Recheck page extra references under lock */
> + if (page_count(page) > page_mapcount(page) +
> + (!PageAnon(page) || PageSwapCache(page)))
> + goto isolate_fail;

Well, that lru_lock (and the intervening PageLRU check after getting it)
may restrict PageAnon and PageSwapCache transitions to some extent, but
it certainly has no effect on page_count and page_mapcount: so I think
such an additional check here is rather superfluous, and we should just
rely on the final checks in migrate_page_move_mapping(), as before.

> }
>
> lruvec = mem_cgroup_page_lruvec(page, pgdat);