On Thu, May 14, 2020 at 10:46:33PM -0400, Qian Cai wrote: > > > > On Oct 20, 2019, at 11:16 PM, Naoya Horiguchi <n-horiguchi@xxxxxxxxxxxxx> wrote: > > > > On Fri, Oct 18, 2019 at 07:56:09AM -0400, Qian Cai wrote: > >> > >> > >> On Oct 18, 2019, at 2:35 AM, Naoya Horiguchi <n-horiguchi@xxxxxxxxxxxxx> > >> wrote: > >> > >> > >> You're right, then I don't see how this happens. If the error hugepage was > >> isolated without having PG_hwpoison set, it's unexpected and problematic. > >> I'm testing myself with v5.4-rc2 (simply ran move_pages12 and did hotremove > >> /hotadd) > >> but don't reproduce the issue yet. Do we need specific kernel version/ > >> config > >> to trigger this? > >> > >> > >> This is reproducible on linux-next with the config. Not sure if it is > >> reproducible on x86. > >> > >> https://raw.githubusercontent.com/cailca/linux-mm/master/powerpc.config > >> > >> and kernel cmdline if that matters > >> > >> page_poison=on page_owner=on numa_balancing=enable \ > >> systemd.unified_cgroup_hierarchy=1 debug_guardpage_minorder=1 \ > >> page_alloc.shuffle=1 > > > > Thanks for the info. > > > >> > >> BTW, where does the code set PG_hwpoison for the head page? > > > > Precisely speaking, soft offline only sets PG_hwpoison after the target > > hugepage is successfully dissolved (then it's not a hugepage any more), > > so PG_hwpoison is set on the raw page in set_hwpoison_free_buddy_page(). > > > > In move_pages12 case, madvise(MADV_SOFT_OFFLINE) is called for the range > > of 2 hugepages, so the expected result is that page offset 0 and 512 > > are marked as PG_hwpoison after injection. > > > > Looking at your dump_page() output, the end_pfn is page offset 1 > > ("page:c00c000800458040" is likely to point to pfn 0x11601.) > > The page belongs to high order buddy free page, but doesn't have > > PageBuddy nor PageHWPoison because it was not the head page or > > the raw error page. > > > >> Unfortunately, this does not solve the problem. It looks to me that in > >> soft_offline_huge_page(), set_hwpoison_free_buddy_page() will only set > >> PG_hwpoison for buddy pages, so the even the compound_head() has no PG_hwpoison > >> set. > > > > Your analysis is totally correct, and this behavior will be fixed by > > the change (https://lkml.org/lkml/2019/10/17/551) in Oscar's rework. > > The raw error page will be taken off from buddy system and the other > > subpages are properly split into lower orderer pages (we'll properly > > manage PageBuddy flags). So all possible cases would be covered by > > branches in __test_page_isolated_in_pageblock. > > Naoya, Oscar, it looks like this series was stuck. > > https://lkml.org/lkml/2019/10/17/551 > > I can still reproduce this issue as today. Maybe it is best we could post a single patch (which one?) to fix the loop first? I'm very sorry to be quiet for long, but I think that I agree with this patchset and try to see what happend if merged into mmtom, although we need rebaseing to latest mmotm and some basic testing.
