On 11/05/18 at 10:28am, Michal Hocko wrote: > On Mon 05-11-18 10:14:07, Michal Hocko wrote: > > Maybe we can add a retry for movable zone pages. > > Or something like this. Ugly as hell, no question about that. I also > have to think about this some more to convince myself this will not > result in an endless loop under some situations.
Testing, will tell the result. > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c > index 48ceda313332..342d66eca0f3 100644 > --- a/mm/page_alloc.c > +++ b/mm/page_alloc.c > @@ -7779,12 +7779,16 @@ bool has_unmovable_pages(struct zone *zone, struct > page *page, int count, > pfn = page_to_pfn(page); > for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) { > unsigned long check = pfn + iter; > + unsigned long saved_flags; > > if (!pfn_valid_within(check)) > continue; > > page = pfn_to_page(check); > > +retry: > + saved_flags = READ_ONCE(page->flags); > + > if (PageReserved(page)) > goto unmovable; > > @@ -7840,6 +7844,13 @@ bool has_unmovable_pages(struct zone *zone, struct > page *page, int count, > page->mapping->a_ops->migratepage) > continue; > > + /* > + * We might race with the allocation of the page so retry > + * if flags have changed. > + */ > + if (saved_flags != READ_ONCE(page->flags)) > + goto retry; > + > /* > * If there are RECLAIMABLE pages, we need to check > * it. But now, memory offline itself doesn't call > -- > Michal Hocko > SUSE Labs