On 2022/4/12 14:37, HORIGUCHI NAOYA(堀口 直也) wrote: > On Mon, Apr 11, 2022 at 09:19:26PM +0800, Miaohe Lin wrote: >> On 2022/4/11 14:35, HORIGUCHI NAOYA(堀口 直也) wrote: >>> On Thu, Apr 07, 2022 at 09:03:50PM +0800, Miaohe Lin wrote: >>>> PageSwapCache is only reliable when PageAnon is true because PG_swapcache >>>> serves as PG_owner_priv_1 which can be used by fs if it's pagecache page. >>>> So we should test PageAnon to distinguish pagecache page from swapcache >>>> page to avoid false-postive PageSwapCache test. >>>> >>>> Signed-off-by: Miaohe Lin <linmiaohe@xxxxxxxxxx> >>>> --- >>>> mm/memory-failure.c | 2 +- >>>> 1 file changed, 1 insertion(+), 1 deletion(-) >>>> >>>> diff --git a/mm/memory-failure.c b/mm/memory-failure.c >>>> index ef402b490663..2e97302d62e4 100644 >>>> --- a/mm/memory-failure.c >>>> +++ b/mm/memory-failure.c >>>> @@ -2262,7 +2262,7 @@ static int __soft_offline_page(struct page *page) >>>> return 0; >>>> } >>>> >>>> - if (!PageHuge(page) && PageLRU(page) && !PageSwapCache(page)) >>>> + if (!PageHuge(page) && PageLRU(page) && !PageAnon(page)) >>>> /* >>>> * Try to invalidate first. This should work for >>>> * non dirty unmapped page cache pages. >>>> -- >>> >>> I foudn that with this change the following VM_BUG_ON_FOLIO() is triggered >>> when calling soft-offline for a swapcache. Maybe we need check both of >>> PageAnon and PageSwapCache instead of either? >>> >> >> Many thanks for your test! This is my overlook. Sorry about it! :( The root cause is that the page is >> added into swapcache and lru( so that it can pass the HWPoisonHandlable check) but page anon is not >> set yet due to page lock is held by __soft_offline_page. So we have the below core dump: >> >> [ 41.232172] page:0000000033d8a20c refcount:0 mapcount:0 mapping:00000000bc103d88 index:0x36d pfn:0x14359b >> ^^^ page is not anon >> >> [ 41.236576] flags: 0x57ffffc0080415(locked|uptodate|lru|owner_priv_1|swapbacked|node=1|zone=2|lastcpupid=0x1fffff) >> ^^^^^^^^^^^^^^^^^^ page is in swapcache >> >> It seems we can check !PageAnon(page) && !PageSwapCache(page), as you suggested, to fix this issue. But maybe I >> should drop this patch because invalidate_inode_page will always return 0 for PageAnon due to folio_mapping == NULL. >> So nothing is really done for anonymous page here. And the origin !PageSwapCache(page) check should do the right work. > > Thanks for clarification. > >> Or we shouldn't even try to call invalidate_inode_page with anonymous page in principle? > > I think just keeping the current behavior is fine (because as you stated > above invalidate_inode_page() simple ignores anonymous pages). > Will drop this patch. Sorry for make noise. :( > Thanks, > Naoya Horiguchi > >> BTW: PageSwapCache should be reliable here as folio_test_swapbacked is checked implicitly inside it. In such case, PG_swapcache >> can't serve as PG_owner_priv_1 as pagecache page shouldn't set PG_swapbacked (shmem will set PG_swapbacked but PG_owner_priv_1 >> is not used anyway). Or am I miss something again?
