On Thu, Aug 04, 2022 at 04:40:25PM +0900, Naoya Horiguchi wrote: > On Tue, Aug 02, 2022 at 02:27:36PM +0800, Miaohe Lin wrote: > > On 2022/8/2 12:07, Muchun Song wrote: > > > On Tue, Aug 02, 2022 at 10:00:50AM +0800, Miaohe Lin wrote: > > >> Hi all: > > >> When I investigate the mm/memory-failure.c code again, I found there's a possible race window > > >> between memory_failure and alloc_fresh_huge_page/free_huge_page. Thank about the below scene: > > >> > > >> CPU 1 CPU 2 > > >> alloc_fresh_huge_page -- page refcnt > 0 memory_failure > > >> prep_new_huge_page get_huge_page_for_hwpoison > > >> !PageHeadHuge -- so 2(not a hugepage) is returned > > >> hugetlb_vmemmap_optimize -- subpages is read-only > > >> set_compound_page_dtor -- PageHuge is true now, but too late!!! > > >> TestSetPageHWPoison(p) > > >> -- We might write to read-only subpages here!!! > > >> > > >> Another similar scene: > > >> > > >> CPU 1 CPU 2 > > >> free_huge_page -- page refcnt == 0 and not PageHuge memory_failure > > >> get_huge_page_for_hwpoison > > >> !PageHeadHuge -- so 2(not a hugepage) is returned > > >> TestSetPageHWPoison(p) > > >> -- We might write to read-only subpages here!!! > > >> hugetlb_vmemmap_restore -- subpages can be written to now, but too late!!! > > >> > > > > > > I agree this race is possible, I have proposed this race in thread [1]. > > Thank you for reminding this, and I agree that we need some solution. > > > > > Oh, I remember I see the race proposed in [1] but I did not look into that carefully at that time. Sorry. > > > > > But I didn't think more how to solve it. > > I hope this thread can find a good solution. :) > > Both of the races show that __get_huge_page_for_hwpoison() fails to > capture the case of generic compound page during turning into hugetlb, I noticed my wrong description just after sending this out. The second race is about generic compound page turning into "buddy", sorry. - Naoya Horiguchi > What makes things complicated is that no one knows to which state > such a compound page finally turns into. So I think that if the page > to be handled is an unknown compound page, we need to wait until it > becomes some known page state to avoid misjudging. > > If we need a quick small fix, we may replace the check "!PageHeadHuge()" > in __get_huge_page_for_hwpoison() with "!PageCompound()", and add another > retry path in get_huge_page_for_hwpoison() for non-hugetlb compound pages. > > Thanks, > Naoya Horiguchi
