On Wed, Nov 07, 2018 at 08:35:48AM +0100, Michal Hocko wrote:
> On Wed 07-11-18 07:35:18, Balbir Singh wrote:
> > On Tue, Nov 06, 2018 at 10:55:24AM +0100, Michal Hocko wrote:
> > > From: Michal Hocko <mhocko@xxxxxxxx>
> > >
> > > Page state checks are racy. Under a heavy memory workload (e.g. stress
> > > -m 200 -t 2h) it is quite easy to hit a race window when the page is
> > > allocated but its state is not fully populated yet. A debugging patch to
> > > dump the struct page state shows
> > > : [ 476.575516] has_unmovable_pages: pfn:0x10dfec00, found:0x1, count:0x0
> > > : [ 476.582103] page:ffffea0437fb0000 count:1 mapcount:1 mapping:ffff880e05239841 index:0x7f26e5000 compound_mapcount: 1
> > > : [ 476.592645] flags: 0x5fffffc0090034(uptodate|lru|active|head|swapbacked)
> > >
> > > Note that the state has been checked for both PageLRU and PageSwapBacked
> > > already. Closing this race completely would require some sort of retry
> > > logic. This can be tricky and error prone (think of potential endless
> > > or long taking loops).
> > >
> > > Workaround this problem for movable zones at least. Such a zone should
> > > only contain movable pages. 15c30bc09085 ("mm, memory_hotplug: make
> > > has_unmovable_pages more robust") has told us that this is not strictly
> > > true though. Bootmem pages should be marked reserved though so we can
> > > move the original check after the PageReserved check. Pages from other
> > > zones are still prone to races but we even do not pretend that memory
> > > hotremove works for those so pre-mature failure doesn't hurt that much.
> > >
> > > Reported-and-tested-by: Baoquan He <bhe@xxxxxxxxxx>
> > > Acked-by: Baoquan He <bhe@xxxxxxxxxx>
> > > Fixes: "mm, memory_hotplug: make has_unmovable_pages more robust")
> > > Signed-off-by: Michal Hocko <mhocko@xxxxxxxx>
> > > ---
> > >
> > > Hi,
> > > this has been reported [1] and we have tried multiple things to address
> > > the issue. The only reliable way was to reintroduce the movable zone
> > > check into has_unmovable_pages. This time it should be safe also for
> > > the bug originally fixed by 15c30bc09085.
> > >
> > > [1] http://lkml.kernel.org/r/20181101091055.GA15166@MiWiFi-R3L-srv
> > > mm/page_alloc.c | 8 ++++++++
> > > 1 file changed, 8 insertions(+)
> > >
> > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > > index 863d46da6586..c6d900ee4982 100644
> > > --- a/mm/page_alloc.c
> > > +++ b/mm/page_alloc.c
> > > @@ -7788,6 +7788,14 @@ bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
> > > if (PageReserved(page))
> > > goto unmovable;
> > >
> > > + /*
> > > + * If the zone is movable and we have ruled out all reserved
> > > + * pages then it should be reasonably safe to assume the rest
> > > + * is movable.
> > > + */
> > > + if (zone_idx(zone) == ZONE_MOVABLE)
> > > + continue;
> > > +
> > > /*
> >
> >
> > There is a WARN_ON() in case of failure at the end of the routine,
> > is that triggered when we hit the bug? If we're adding this patch,
> > the WARN_ON needs to go as well.
>
> No the warning should stay in case we encounter reserved pages in zone
> movable.
>
Fair enough!
> > The check seems to be quite aggressive and in a loop that iterates
> > pages, but has nothing to do with the page, did you mean to make
> > the check
> >
> > zone_idx(page_zone(page)) == ZONE_MOVABLE
>
> Does it make any difference? Can we actually encounter a page from a
> different zone here?
>
Just to avoid page state related issues, do we want to go ahead
with the migration if zone_idx(page_zone(page)) != ZONE_MOVABLE.
> > it also skips all checks for pinned pages and other checks
>
> Yes, this is intentional and the comment tries to explain why. I wish we
> could be add a more specific checks for movable pages - e.g. detect long
> term pins that would prevent migration - but we do not have any facility
> for that. Please note that the worst case of a false positive is a
> repeated migration failure and user has a way to break out of migration
> by a signal.
>
Basically isolate_pages() will fail as opposed to hotplug failing upfront.
The basic assertion this patch makes is that all ZONE_MOVABLE pages that
are not reserved are hotpluggable.
Balbir Singh.
