From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-2.2 required=3.0 tests=HEADER_FROM_DIFFERENT_DOMAINS, MAILING_LIST_MULTI,SPF_HELO_NONE,SPF_PASS,UNPARSEABLE_RELAY,USER_AGENT_SANE_1 autolearn=no autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id 04549C2BAEE for ; Tue, 24 Mar 2020 17:17:24 +0000 (UTC) Received: from kanga.kvack.org (kanga.kvack.org [205.233.56.17]) by mail.kernel.org (Postfix) with ESMTP id 554142051A for ; Tue, 24 Mar 2020 17:17:23 +0000 (UTC) DMARC-Filter: OpenDMARC Filter v1.3.2 mail.kernel.org 554142051A Authentication-Results: mail.kernel.org; dmarc=fail (p=none dis=none) header.from=linux.alibaba.com Authentication-Results: mail.kernel.org; spf=pass smtp.mailfrom=owner-linux-mm@kvack.org Received: by kanga.kvack.org (Postfix) id CE44D6B0003; Tue, 24 Mar 2020 13:17:22 -0400 (EDT) Received: by kanga.kvack.org (Postfix, from userid 40) id C96086B0006; Tue, 24 Mar 2020 13:17:22 -0400 (EDT) X-Delivered-To: int-list-linux-mm@kvack.org Received: by kanga.kvack.org (Postfix, from userid 63042) id B83AC6B0007; Tue, 24 Mar 2020 13:17:22 -0400 (EDT) X-Delivered-To: linux-mm@kvack.org Received: from forelay.hostedemail.com (smtprelay0034.hostedemail.com [216.40.44.34]) by kanga.kvack.org (Postfix) with ESMTP id 9C3DC6B0003 for ; Tue, 24 Mar 2020 13:17:22 -0400 (EDT) Received: from smtpin21.hostedemail.com (10.5.19.251.rfc1918.com [10.5.19.251]) by forelay02.hostedemail.com (Postfix) with ESMTP id 6F6C9ACF5D for ; Tue, 24 Mar 2020 17:17:22 +0000 (UTC) X-FDA: 76630911924.21.hall83_231c73f52af24 X-HE-Tag: hall83_231c73f52af24 X-Filterd-Recvd-Size: 6203 Received: from out30-42.freemail.mail.aliyun.com (out30-42.freemail.mail.aliyun.com [115.124.30.42]) by imf02.hostedemail.com (Postfix) with ESMTP for ; Tue, 24 Mar 2020 17:17:20 +0000 (UTC) X-Alimail-AntiSpam:AC=PASS;BC=-1|-1;BR=01201311R251e4;CH=green;DM=||false|;DS=||;FP=0|-1|-1|-1|0|-1|-1|-1;HT=e01e04357;MF=yang.shi@linux.alibaba.com;NM=1;PH=DS;RN=7;SR=0;TI=SMTPD_---0TtXVszC_1585070234; Received: from US-143344MP.local(mailfrom:yang.shi@linux.alibaba.com fp:SMTPD_---0TtXVszC_1585070234) by smtp.aliyun-inc.com(127.0.0.1); Wed, 25 Mar 2020 01:17:17 +0800 Subject: Re: [PATCH] mm: khugepaged: fix potential page state corruption To: "Kirill A. Shutemov" Cc: kirill.shutemov@linux.intel.com, hughd@google.com, aarcange@redhat.com, akpm@linux-foundation.org, linux-mm@kvack.org, linux-kernel@vger.kernel.org References: <1584573582-116702-1-git-send-email-yang.shi@linux.alibaba.com> <20200319001258.creziw6ffw4jvwl3@box> <2cdc734c-c222-4b9d-9114-1762b29dafb4@linux.alibaba.com> <20200319104938.vphyajoyz6ob6jtl@box> From: Yang Shi Message-ID: <99b78cdb-5a4d-e28b-4464-d34ee39e5501@linux.alibaba.com> Date: Tue, 24 Mar 2020 10:17:13 -0700 User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10.12; rv:52.0) Gecko/20100101 Thunderbird/52.7.0 MIME-Version: 1.0 In-Reply-To: <20200319104938.vphyajoyz6ob6jtl@box> Content-Type: text/plain; charset=utf-8; format=flowed Content-Language: en-US Content-Transfer-Encoding: quoted-printable X-Bogosity: Ham, tests=bogofilter, spamicity=0.000000, version=1.2.4 Sender: owner-linux-mm@kvack.org Precedence: bulk X-Loop: owner-majordomo@kvack.org List-ID: On 3/19/20 3:49 AM, Kirill A. Shutemov wrote: > On Wed, Mar 18, 2020 at 10:39:21PM -0700, Yang Shi wrote: >> >> On 3/18/20 5:55 PM, Yang Shi wrote: >>> >>> On 3/18/20 5:12 PM, Kirill A. Shutemov wrote: >>>> On Thu, Mar 19, 2020 at 07:19:42AM +0800, Yang Shi wrote: >>>>> When khugepaged collapses anonymous pages, the base pages would >>>>> be freed >>>>> via pagevec or free_page_and_swap_cache().=C2=A0 But, the anonymous= page may >>>>> be added back to LRU, then it might result in the below race: >>>>> >>>>> =C2=A0=C2=A0=C2=A0=C2=A0CPU A=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2= =A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0 CPU B >>>>> khugepaged: >>>>> =C2=A0=C2=A0 unlock page >>>>> =C2=A0=C2=A0 putback_lru_page >>>>> =C2=A0=C2=A0=C2=A0=C2=A0 add to lru >>>>> =C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0= =C2=A0=C2=A0=C2=A0=C2=A0 page reclaim: >>>>> =C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0= =C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0 isolate this page >>>>> =C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0= =C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0 try_to_unmap >>>>> =C2=A0=C2=A0 page_remove_rmap <-- corrupt _mapcount >>>>> >>>>> It looks nothing would prevent the pages from isolating by reclaime= r. >>>> Hm. Why should it? >>>> >>>> try_to_unmap() doesn't exclude parallel page unmapping. _mapcount is >>>> protected by ptl. And this particular _mapcount pin is reachable for >>>> reclaim as it's not part of usual page table tree. Basically >>>> try_to_unmap() will never succeeds until we give up the _mapcount on >>>> khugepaged side. >>> I don't quite get. What does "not part of usual page table tree" mean= s? >>> >>> How's about try_to_unmap() acquires ptl before khugepaged? > The page table we are dealing with was detached from the process' page > table tree: see pmdp_collapse_flush(). try_to_unmap() will not see the > pte. A follow-up question here. pmdp_collapse_flush() clears pmd entry and=20 does TLB shootdown on x86. I'm supposed the main purpose is to serialize=20 fast gup since it doesn't acquire any lock (mmap_sem, ptl ,etc), but=20 disable interrupt so the TLB shootdown IPI would get blocked. This could=20 guarantee synchronization on x86, but it looks not all architectures do=20 TLB shootdown or implement it via IPI, so how they could serialize with=20 fast gup? In addition it looks acquiring pmd lock is not necessary. Before both=20 write mmap_sem and write anon_vma lock are acquired which could=20 serialize page fault and rmap walk, so it looks fast gup is the only one=20 which could run concurrently, but fast gup doesn't acquire ptl at all.=20 It seems the pmd_lock/unlock could be removed. > > try_to_unmap() can only reach the ptl if split ptl is disabled > (mm->page_table_lock is used), but it still will not be able to reach p= te. > >>>> I don't see the issue right away. >>>> >>>>> The other problem is the page's active or unevictable flag might be >>>>> still set when freeing the page via free_page_and_swap_cache(). >>>> So what? >>> The flags may leak to page free path then kernel may complain if >>> DEBUG_VM is set. > Could you elaborate on what codepath you are talking about? > >>>>> The putback_lru_page() would not clear those two flags if the pages= are >>>>> released via pagevec, it sounds nothing prevents from isolating act= ive >> Sorry, this is a typo. If the page is freed via pagevec, active and >> unevictable flag would get cleared before freeing by page_off_lru(). >> >> But, if the page is freed by free_page_and_swap_cache(), these two fla= gs are >> not cleared. But, it seems this path is hit rare, the pages are freed = by >> pagevec for the most cases. >> >>>>> or unevictable pages. >>>> Again, why should it? vmscan is equipped to deal with this. >>> I don't mean vmscan, I mean khugepaged may isolate active and >>> unevictable pages since it just simply walks page table. > Why it is wrong? lru_cache_add() only complains if both flags set, it > shouldn't happen. > >>>>> However I didn't really run into these problems, just in theory >>>>> by visual >>>>> inspection. >>>>> >>>>> And, it also seems unnecessary to have the pages add back to LRU >>>>> again since >>>>> they are about to be freed when reaching this point.=C2=A0 So, >>>>> clearing active >>>>> and unevictable flags, unlocking and dropping refcount from isolate >>>>> instead of calling putback_lru_page() as what page cache collapse d= oes. >>>> Hm? But we do call putback_lru_page() on the way out. I do not follo= w. >>> It just calls putback_lru_page() at error path, not success path. >>> Putting pages back to lru on error path definitely makes sense. Here = it >>> is the success path. > I agree that putting the apage on LRU just before free the page is > suboptimal, but I don't see it as a critical issue. > >