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 Received: from vger.kernel.org (vger.kernel.org [23.128.96.18]) by smtp.lore.kernel.org (Postfix) with ESMTP id C6AB4C00140 for ; Wed, 24 Aug 2022 15:13:23 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S239425AbiHXPNU (ORCPT ); Wed, 24 Aug 2022 11:13:20 -0400 Received: from lindbergh.monkeyblade.net ([23.128.96.19]:40852 "EHLO lindbergh.monkeyblade.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S239407AbiHXPNR (ORCPT ); Wed, 24 Aug 2022 11:13:17 -0400 Received: from us-smtp-delivery-124.mimecast.com (us-smtp-delivery-124.mimecast.com [170.10.133.124]) by lindbergh.monkeyblade.net (Postfix) with ESMTPS id 529C494ECB for ; Wed, 24 Aug 2022 08:13:15 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=redhat.com; s=mimecast20190719; t=1661353994; h=from:from:reply-to:subject:subject:date:date:message-id:message-id: to:to:cc:cc:mime-version:mime-version:content-type:content-type: content-transfer-encoding:content-transfer-encoding: in-reply-to:in-reply-to:references:references; bh=j2EkIK7Ec5CQ3JRJQv0sZNP7yIn77zqG02uFB5xCBSA=; b=JIFvWt+QeJW9Vy3UIIJao9/u+WeNHHxtak4j4fdNe4ozRoiAa0cLp5suhjLYWiGyVhPGt8 TIUMMnRX1xPrVEU7a2wrU422z3YL4q264yMEcJAc9ok5UZ6pJzmz4zthmoeB5WVfXOfnE4 lfV2mUj/5KTZpNtIH2XQP16stiDXVPU= Received: from mail-wm1-f69.google.com (mail-wm1-f69.google.com [209.85.128.69]) by relay.mimecast.com with ESMTP with STARTTLS (version=TLSv1.3, cipher=TLS_AES_128_GCM_SHA256) id us-mta-490-YYQj_3iKP9WbwiSlSuRzMw-1; Wed, 24 Aug 2022 11:13:11 -0400 X-MC-Unique: YYQj_3iKP9WbwiSlSuRzMw-1 Received: by mail-wm1-f69.google.com with SMTP id v64-20020a1cac43000000b003a4bea31b4dso954336wme.3 for ; Wed, 24 Aug 2022 08:13:10 -0700 (PDT) X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20210112; h=content-transfer-encoding:in-reply-to:organization:from:references :cc:to:content-language:subject:user-agent:mime-version:date :message-id:x-gm-message-state:from:to:cc; bh=j2EkIK7Ec5CQ3JRJQv0sZNP7yIn77zqG02uFB5xCBSA=; b=5En5tk6zlf/SQGQnAx2tGq95MwsVgyOXf+O5O7xPsCJrU2o5Xm+UgfuLn6ZaxhUnTH eXDNxm4KMb/h1S65Y+8vMhpVJnPErybOIEmoe888FLOVOddRczWzNjsg3c646VzdT7SH wk27t4gZFlCSR++TMWL1vTfmn2v6e8v0iFvEuJa7S57V3+y61iCBcadDM4C1WZhPieEX XH0Ruxx1xFeSasmNdOMszlXfhX7wmLMIO7fQWnFL2wYCHUSv02bKzhJSG3MCRKNvPd0l IOofZ5GoX+b9Z1My0Y1apTmuMCuNYfFg06XhIkQn9mFLKgDCR1eLfGyVbbVSH7nrMwRv 0QkQ== X-Gm-Message-State: ACgBeo1YCYaXsCijfy+vvVoxmV6hiBxLV5r3DROtxQofNP/LqOz0zl/L aGkLBBt6/tiUlSjQ0WX/tiwAmGRCIhcbFD0rdRiWnqFnMq1XqI2vHUZZWgrmy+YE8r2sra+FwdR 3OLV6ummf34BBjY1xSkc3gR+q X-Received: by 2002:a5d:6d0b:0:b0:222:955a:8774 with SMTP id e11-20020a5d6d0b000000b00222955a8774mr15635922wrq.129.1661353989959; Wed, 24 Aug 2022 08:13:09 -0700 (PDT) X-Google-Smtp-Source: AA6agR5eD/LyvKaPYC90/Ro6Oi2I+AixQD/rYEgcR6MD39YuKJ1MCboIQxsYUXNDYeJ3+NHjX5DxnQ== X-Received: by 2002:a5d:6d0b:0:b0:222:955a:8774 with SMTP id e11-20020a5d6d0b000000b00222955a8774mr15635903wrq.129.1661353989644; Wed, 24 Aug 2022 08:13:09 -0700 (PDT) Received: from ?IPV6:2003:cb:c707:c500:5445:cf40:2e32:6e73? (p200300cbc707c5005445cf402e326e73.dip0.t-ipconnect.de. [2003:cb:c707:c500:5445:cf40:2e32:6e73]) by smtp.gmail.com with ESMTPSA id p6-20020a1c5446000000b003a500b612fcsm2137004wmi.12.2022.08.24.08.13.08 (version=TLS1_3 cipher=TLS_AES_128_GCM_SHA256 bits=128/128); Wed, 24 Aug 2022 08:13:09 -0700 (PDT) Message-ID: <6e2539b7-b4c7-95dc-e4ac-27692d955936@redhat.com> Date: Wed, 24 Aug 2022 17:13:08 +0200 MIME-Version: 1.0 User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101 Thunderbird/91.12.0 Subject: Re: [PATCH v2 1/5] mm/hugetlb: fix races when looking up a CONT-PTE size hugetlb page Content-Language: en-US To: Baolin Wang , Mike Kravetz Cc: akpm@linux-foundation.org, songmuchun@bytedance.com, linux-mm@kvack.org, linux-kernel@vger.kernel.org References: <0e5d92da043d147a867f634b17acbcc97a7f0e64.1661240170.git.baolin.wang@linux.alibaba.com> <4c24b891-04ce-2608-79d2-a75dc236533f@redhat.com> <376d2e0a-d28a-984b-903c-1f6451b04a15@linux.alibaba.com> <7d4e7f47-30a5-3cc6-dc9f-aa89120847d8@redhat.com> <64669c0a-4a6e-f034-a15b-c4a8deea9e5d@linux.alibaba.com> <7ee73879-e402-9175-eae8-41471d80d59e@redhat.com> <041e2e43-2227-1681-743e-5f82e245b5ea@redhat.com> <0f736dc5-1798-10ad-c506-9a2a38841359@redhat.com> <22585fc8-b0bc-0e14-d121-2767cd178424@linux.alibaba.com> From: David Hildenbrand Organization: Red Hat In-Reply-To: <22585fc8-b0bc-0e14-d121-2767cd178424@linux.alibaba.com> Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Precedence: bulk List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On 24.08.22 17:06, Baolin Wang wrote: > > > On 8/24/2022 10:33 PM, David Hildenbrand wrote: >> On 24.08.22 16:30, Baolin Wang wrote: >>> >>> >>> On 8/24/2022 7:55 PM, David Hildenbrand wrote: >>>> On 24.08.22 11:41, Baolin Wang wrote: >>>>> >>>>> >>>>> On 8/24/2022 3:31 PM, David Hildenbrand wrote: >>>>>>>>>> >>>>>>>>>> IMHO, these follow_huge_xxx() functions are arch-specified at first and >>>>>>>>>> were moved into the common hugetlb.c by commit 9e5fc74c3025 ("mm: >>>>>>>>>> hugetlb: Copy general hugetlb code from x86 to mm"), and now there are >>>>>>>>>> still some arch-specified follow_huge_xxx() definition, for example: >>>>>>>>>> ia64: follow_huge_addr >>>>>>>>>> powerpc: follow_huge_pd >>>>>>>>>> s390: follow_huge_pud >>>>>>>>>> >>>>>>>>>> What I mean is that follow_hugetlb_page() is a common and >>>>>>>>>> not-arch-specified function, is it suitable to change it to be >>>>>>>>>> arch-specified? >>>>>>>>>> And thinking more, can we rename follow_hugetlb_page() as >>>>>>>>>> hugetlb_page_faultin() and simplify it to only handle the page faults of >>>>>>>>>> hugetlb like the faultin_page() for normal page? That means we can make >>>>>>>>>> sure only follow_page_mask() can handle hugetlb. >>>>>>>>>> >>>>>>>> >>>>>>>> Something like that might work, but you still have two page table walkers >>>>>>>> for hugetlb. I like David's idea (if I understand it correctly) of >>>>>>> >>>>>>> What I mean is we may change the hugetlb handling like normal page: >>>>>>> 1) use follow_page_mask() to look up a hugetlb firstly. >>>>>>> 2) if can not get the hugetlb, then try to page fault by >>>>>>> hugetlb_page_faultin(). >>>>>>> 3) if page fault successed, then retry to find hugetlb by >>>>>>> follow_page_mask(). >>>>>> >>>>>> That implies putting more hugetlbfs special code into generic GUP, >>>>>> turning it even more complicated. But of course, it depends on how the >>>>>> end result looks like. My gut feeling was that hugetlb is better handled >>>>>> in follow_hugetlb_page() separately (just like we do with a lot of other >>>>>> page table walkers). >>>>> >>>>> OK, fair enough. >>>>> >>>>>>> >>>>>>> Just a rough thought, and I need more investigation for my idea and >>>>>>> David's idea. >>>>>>> >>>>>>>> using follow_hugetlb_page for both cases. As noted, it will need to be >>>>>>>> taught how to not trigger faults in the follow_page_mask case. >>>>>>> >>>>>>> Anyway, I also agree we need some cleanup, and firstly I think we should >>>>>>> cleanup these arch-specified follow_huge_xxx() on some architectures >>>>>>> which are similar with the common ones. I will look into these. >>>>>> >>>>>> There was a recent discussion on that, e.g.: >>>>>> >>>>>> https://lkml.kernel.org/r/20220818135717.609eef8a@thinkpad >>>>> >>>>> Thanks. >>>>> >>>>>> >>>>>>> >>>>>>> However, considering cleanup may need more investigation and >>>>>>> refactoring, now I prefer to make these bug-fix patches of this patchset >>>>>>> into mainline firstly, which are suitable to backport to old version to >>>>>>> fix potential race issues. Mike and David, how do you think? Could you >>>>>>> help to review these patches? Thanks. >>>>>> >>>>>> Patch #1 certainly add more special code just to handle another hugetlb >>>>>> corner case (CONT pages), and maybe just making it all use >>>>>> follow_hugetlb_page() would be even cleaner and less error prone. >>>>>> >>>>>> I agree that locking is shaky, but I'm not sure if we really want to >>>>>> backport this to stable trees: >>>>>> >>>>>> https://www.kernel.org/doc/html/latest/process/stable-kernel-rules.html >>>>>> >>>>>> "It must fix a real bug that bothers people (not a, “This could be a >>>>>> problem...” type thing)." >>>>>> >>>>>> >>>>>> Do we actually have any instance of this being a real (and not a >>>>>> theoretical) problem? If not, I'd rather clean it all up right away. >>>>> >>>>> I think this is a real problem (not theoretical), and easy to write some >>>>> code to show the issue. For example, suppose thread A is trying to look >>>>> up a CONT-PTE size hugetlb page under the lock, however antoher thread B >>>>> can migrate the CONT-PTE hugetlb page at the same time, which will cause >>>>> thread A to get an incorrect page, if thread A want to do something for >>>>> this incorrect page, error occurs. >>>>> >>>>> Actually we also want to backport these fixes to the distro with old >>>>> kernel versions to make the hugetlb more stable. Otherwise we must hit >>>>> these issues sooner or later if the customers use CONT-PTE/PMD hugetlb. >>>>> >>>>> Anyway, if you and Mike still think these issues are not important >>>>> enough to be fixed in the old versions, I can do the cleanup firstly. >>>>> >>>> >>>> [asking myself which follow_page() users actually care about hugetlb, >>>> and why we need this handling in follow_page at all] >>>> >>>> Which follow_page() user do we care about here? Primarily mm/migrate.c >>>> only I assume? >>> >>> Right, mainly affects the move_pages() syscall I think. Yes, I can not >>> know all of the users of the move_pages() syscall now or in the future >>> in our data center, but like I said the move_pages() syscall + hugetlb >>> can be a real potential stability issue. >>> >> >> I wonder if we can get rid of follow_page() completely, there are not >> too many users. Or alternatively simply make it use general GUP >> infrastructure more clearly. We'd need something like FOLL_NOFAULT that >> also covers "absolutely no faults". > > I am not sure I get your point. So you want change to use > __get_user_pages() (or silimar wrappers) to look up a normal page or > hugetlb instead of follow_page()? and adding a new FOLL_NOFAULT flag to > __get_user_pages(). Essentially just getting rid of follow_page() completely or making it a wrapper of __get_user_pages(). > > If I understand correctly, we still need more work to move those > arch-specified follow_huge_xxx() into follow_hugetlb_page() firstly like > we disscussed before? Which seems not backportable too. I'm not sure we need all that magic in these arch specific helpers after all. I haven't looked into the details, but I really wonder why they handle something that follow_hugetlb_page() cannot easily handle. It all smells like legacy cruft. > > I am not againt your idea, and I also agree that we should do some > cleanup. But the point is if we need backport patches to fix this issue, > which affects move_pages() syscall, if the answer is yes, I think my > current fixes are suitable to backport. I really don't like adding more make-legacy-cruft-happy code unless there is *real* need for it. (you could always just fix old kernels you care about with your patches here -- do they have to be in mainline? don't think so) But of course, it's up to Mike to decide, just my 2 cents :) -- Thanks, David / dhildenb