Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S932115AbaLJOKq (ORCPT ); Wed, 10 Dec 2014 09:10:46 -0500 Received: from mail-ig0-f177.google.com ([209.85.213.177]:52530 "EHLO mail-ig0-f177.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753874AbaLJOKp (ORCPT ); Wed, 10 Dec 2014 09:10:45 -0500 MIME-Version: 1.0 In-Reply-To: <5486BFCB.4040305@suse.cz> References: <000201d01385$25a6c950$70f45bf0$%yang@samsung.com> <5486BFCB.4040305@suse.cz> Date: Wed, 10 Dec 2014 22:10:44 +0800 Message-ID: Subject: Re: [PATCH 2/3] mm: page_isolation: remove unnecessary freepage_migratetype check for unused page From: Weijie Yang To: Vlastimil Babka Cc: Weijie Yang , Joonsoo Kim , Andrew Morton , Mel Gorman , Rik van Riel , Johannes Weiner , Minchan Kim , Linux-Kernel , Linux-MM Content-Type: text/plain; charset=UTF-8 Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Tue, Dec 9, 2014 at 5:24 PM, Vlastimil Babka wrote: > On 12/09/2014 08:51 AM, Weijie Yang wrote: >> >> when we test the pages in a range is free or not, there is a little >> chance we encounter some page which is not in buddy but page_count is 0. >> That means that page could be in the page-freeing path but not in the >> buddy freelist, such as in pcplist > > > This shouldn't happen anymore IMHO. The pageblock is marked as > MIGRATE_ISOLATE and then a lru+pcplist drain is performed. Nothing should be > left on pcplist - anything newly freed goes directly to free lists. Hm, > maybe it could be on lru cache, but that holds a page reference IIRC, so > this test won't pass. Yes, you are right. I made a mistake, this shouldn't happen. I will remove this description in next version. Thanks. >> or wait for the zone->lock which the >> tester is holding. > > > That could maybe happen, but is it worth testing? If yes, please add it in a > comment to the code. This could happen even though the chance is very tiny. As for cma_alloc, the test makes no difference. However, as for offline_page, the test makes sense. If we leave the test and pass it when page_count is zero, it could trigger the BUG_ON(!PageBuddy(page)) in the __offline_isolated_pages() if the page hasn't finish its free journey. >From the literal meaning of this test_pages_isolated() function, I think it is better get a definite result and not leave some middle status even if they are rare. So, Let's remove the whole test branch (page_count(page) == 0 && get_freepage_migratetype(page) == MIGRATE_ISOLATE) Thanks for your remind and suggestion. > >> Back to the freepage_migratetype, we use it for a cached value for decide >> which free-list the page go when freeing page. If the pageblock is >> isolated >> the page will go to free-list[MIGRATE_ISOLATE] even if the cached type is >> not MIGRATE_ISOLATE, the commit ad53f92e(fix incorrect isolation behavior >> by rechecking migratetype) patch series have ensure this. >> >> So the freepage_migratetype check for page_count==0 page in >> __test_page_isolated_in_pageblock() is meaningless. >> This patch removes the unnecessary freepage_migratetype check. >> >> Signed-off-by: Weijie Yang >> --- >> mm/page_isolation.c | 3 +-- >> 1 file changed, 1 insertion(+), 2 deletions(-) >> >> diff --git a/mm/page_isolation.c b/mm/page_isolation.c >> index 6e5174d..f7c9183 100644 >> --- a/mm/page_isolation.c >> +++ b/mm/page_isolation.c >> @@ -223,8 +223,7 @@ __test_page_isolated_in_pageblock(unsigned long pfn, >> unsigned long end_pfn, >> page = pfn_to_page(pfn); >> if (PageBuddy(page)) >> pfn += 1 << page_order(page); >> - else if (page_count(page) == 0 && >> - get_freepage_migratetype(page) == MIGRATE_ISOLATE) >> + else if (page_count(page) == 0) >> pfn += 1; >> else if (skip_hwpoisoned_pages && PageHWPoison(page)) { >> /* >> > -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/