On 17.01.20 10:40, Michal Hocko wrote: > On Fri 17-01-20 10:25:06, David Hildenbrand wrote: >> On 17.01.20 09:59, Michal Hocko wrote: >>> On Fri 17-01-20 09:51:05, David Hildenbrand wrote: >>>> On 17.01.20 03:21, Qian Cai wrote: >>> [...] >>>>> Even though has_unmovable_pages doesn't hold any reference to the >>>>> returned page this should be reasonably safe for the purpose of >>>>> reporting the page (dump_page) because it cannot be hotremoved. The >>>> >>>> This is only true in the context of memory unplug, but not in the >>>> context of is_mem_section_removable()-> is_pageblock_removable_nolock(). >>> >>> Well, the above should hold for that path as well AFAICS. If the page is >>> unmovable then a racing hotplug cannot remove it, right? Or do you >>> consider a temporary unmovability to be a problem? >> >> Somebody could test /sys/devices/system/memory/memoryX/removable. While >> returning the unmovable page, it could become movable and >> offlining+removing could succeed. > > Doesn't this path use device lock or something? If not than the new code > is not more racy then the existing one. Just look at > is_pageblock_removable_nolock and how it dereferences struct page > (page_zonenum in page_zone.) > AFAIK no device lock, no device hotplug lock, no memory hotplug lock. I think it holds a reference to the device and to the kernelfs node. But AFAIK that does not block removal of offlining/memory, just when the objects get freed. -- Thanks, David / dhildenb
