Messages in this thread Patch in this message | | | Subject | Re: [PATCH] mm/compaction: avoid VM_BUG_ON(PageSlab()) in page_mapcount() | From | Konstantin Khlebnikov <> | Date | Sun, 24 May 2020 14:21:13 +0300 |
| |
On 24/05/2020 04.01, Hugh Dickins wrote: > 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).
Yep, but for large cluster and pinpointed workload this happens surprisingly frequently =) I've believed into this race only after seeing statistics for count of compactions and how it correlates with incidents.
Probably the key component is a slab allocation from network irq/bh context which interrupts compaction exactly at this spot.
> >> >> >> 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@yandex-team.ru> >> 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.
Ok, I'll redo the fix in this way.
> >> --- >> 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?
Yes. What a shame. Adding page_has_private() could fix that?
Kind of
page_count(page) > page_mapcount(page) + (PageAnon(page) ? PageSwapCache(page) : (1 + page_has_private(page)))
or probably something like this:
page_count(page) > page_mapcount(page) + (PageAnon(page) ? PageSwapCache(page) : GUP_PIN_COUNTING_BIAS)
I.e. skip only file pages pinned by dma or something slower. I see some movements in this direction in recent changes.
of course that's independent matter.
> > 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);
| |