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 kanga.kvack.org (kanga.kvack.org [205.233.56.17]) by smtp.lore.kernel.org (Postfix) with ESMTP id DC915C28D13 for ; Mon, 22 Aug 2022 11:26:32 +0000 (UTC) Received: by kanga.kvack.org (Postfix) id 282988D0002; Mon, 22 Aug 2022 07:26:32 -0400 (EDT) Received: by kanga.kvack.org (Postfix, from userid 40) id 231508D0001; Mon, 22 Aug 2022 07:26:32 -0400 (EDT) X-Delivered-To: int-list-linux-mm@kvack.org Received: by kanga.kvack.org (Postfix, from userid 63042) id 0FB248D0002; Mon, 22 Aug 2022 07:26:32 -0400 (EDT) X-Delivered-To: linux-mm@kvack.org Received: from relay.hostedemail.com (smtprelay0015.hostedemail.com [216.40.44.15]) by kanga.kvack.org (Postfix) with ESMTP id 019D08D0001 for ; Mon, 22 Aug 2022 07:26:31 -0400 (EDT) Received: from smtpin12.hostedemail.com (a10.router.float.18 [10.200.18.1]) by unirelay04.hostedemail.com (Postfix) with ESMTP id D3A981A081E for ; Mon, 22 Aug 2022 11:26:31 +0000 (UTC) X-FDA: 79827000582.12.63725ED Received: from mail-pg1-f169.google.com (mail-pg1-f169.google.com [209.85.215.169]) by imf06.hostedemail.com (Postfix) with ESMTP id 86ABE180102 for ; Mon, 22 Aug 2022 11:26:31 +0000 (UTC) Received: by mail-pg1-f169.google.com with SMTP id s206so9109474pgs.3 for ; Mon, 22 Aug 2022 04:26:31 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=20210112; h=content-disposition:mime-version:message-id:cc:to:from:date:sender :from:to:cc; bh=xYnThmFHoGKU/sNt4SS1hXDq6GYqnmSSTt1CpRvIL48=; b=nWxBmLtDylTawPLEuM/2MzvIWaoVY8d8VN5rzm3+Q0FGFiVgZNLmM+p3vaejpbL1KA NPgmYtrMfKM48JaFMB6Ke3Qg2L0TexbZGmCgTtVVq8dM6QKRNqxkc7mzADsEoMxJeaJw 5oxebUX63OgqwuFYXqFBMZHpbhxh+Zxu9bT4gZNwd5p1Qjdh3viL6jy7OdXxRsx3DLe0 5h71ej+cLYiPoleAvF8TEH5G0XmZuyaHCnDHRgeDuUHTOPQXwjwMjEmF969VGN+Kk9NA EZDrfl+54wVczIiVkkLCf/pCm22OFwD0kvLJWOWHJHwUtuA02L3cUMAuODmLr2rlicp3 PVkA== X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20210112; h=content-disposition:mime-version:message-id:cc:to:from:date:sender :x-gm-message-state:from:to:cc; bh=xYnThmFHoGKU/sNt4SS1hXDq6GYqnmSSTt1CpRvIL48=; b=F0VskviNj7rYaW0S5hFSZZsv39ujHyMJ8O8nGQ+AuJ/4+kHxJsQrle3Iw4/GGhp1OL Vacez43fcfUiJ1iQkyxrJbwnUD8bmegBe46wTT37Hy9wYyWUYITR7UtO+e+zCT57Nvrx Ib32MX8wx8TI55go/6LumQUCdpwGrWLS88WrU9l+GwU5F/zq6pLzNGrQMKfL+crMuTRJ vrsHHJv6LeyIVOzC22jmAQrx5OwMLZaF3pQE8mSxvxUHg4UZk8iufAVfx6ALEYNmZBYH Xpb1jIwJI2K5L61nwuSoGxLnopukRaXMzsf0Pl++67iBzmyhO5YY+c6/FNS58DcxT3Ps 7+sw== X-Gm-Message-State: ACgBeo2wfLCtjyA4yR9UzznsDEMxCXVvRClw6MfZCk93WvP62l9HrZU+ QbfhVrhl1ujUdJT1/LnlINE= X-Google-Smtp-Source: AA6agR5ntUFOMw3NycEWdfz8YmaV4pQ2lfDuTFMz2/TqP/1ze9EYk+TwWhkZGVOpWL7qE/syg/tloA== X-Received: by 2002:a63:eb49:0:b0:421:8c9b:4ab1 with SMTP id b9-20020a63eb49000000b004218c9b4ab1mr16740198pgk.339.1661167590122; Mon, 22 Aug 2022 04:26:30 -0700 (PDT) Received: from localhost (2603-800c-1a02-1bae-a7fa-157f-969a-4cde.res6.spectrum.com. [2603:800c:1a02:1bae:a7fa:157f:969a:4cde]) by smtp.gmail.com with ESMTPSA id f8-20020a170902ce8800b00172c298ba42sm226896plg.28.2022.08.22.04.26.29 (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Mon, 22 Aug 2022 04:26:29 -0700 (PDT) Date: Mon, 22 Aug 2022 01:26:28 -1000 From: Tejun Heo To: Yafang Shao Cc: Alexei Starovoitov , Daniel Borkmann , Andrii Nakryiko , Martin Lau , Song Liu , Yonghong Song , john fastabend , KP Singh , Stanislav Fomichev , Hao Luo , jolsa@kernel.org, Johannes Weiner , Michal Hocko , Roman Gushchin , Shakeel Butt , Muchun Song , Andrew Morton , Zefan Li , Cgroups , netdev , bpf , Linux MM , Mina Almasry Message-ID: MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline ARC-Seal: i=1; s=arc-20220608; d=hostedemail.com; t=1661167591; a=rsa-sha256; cv=none; b=B2+Xo8BnK6ZJ314jMoxLG6qPf00vvceiOq4+6viNdIVj6nvuyOZ6Q1eqKofQdz13BqnLly QWzm5TtbJzz27i7zl3dnYqBscNH78B7PRtEWtN+HK6Sr+yZZgHVJq1rABsfIBUbmtVaPQb F0IvH4y4/LEXJx1hGi76W3JT6G3QWVc= ARC-Authentication-Results: i=1; imf06.hostedemail.com; dkim=pass header.d=gmail.com header.s=20210112 header.b=nWxBmLtD; dmarc=fail reason="SPF not aligned (relaxed), DKIM not aligned (relaxed)" header.from=kernel.org (policy=none); spf=pass (imf06.hostedemail.com: domain of htejun@gmail.com designates 209.85.215.169 as permitted sender) smtp.mailfrom=htejun@gmail.com ARC-Message-Signature: i=1; a=rsa-sha256; c=relaxed/relaxed; d=hostedemail.com; s=arc-20220608; t=1661167591; h=from:from:sender:sender:reply-to:subject:date:date: message-id:message-id:to:to:cc:cc:mime-version:mime-version: content-type:content-type:content-transfer-encoding:in-reply-to: references:dkim-signature; bh=xYnThmFHoGKU/sNt4SS1hXDq6GYqnmSSTt1CpRvIL48=; b=BHXi4Ui/hTanmoJujIujFhTRmLtHiCXpWfj29q+/vcK+US+WzfhIPsiRHAWiA9pJQmPY93 Pbm4/jBF8ITHlmx9yU6FTI0qdS2HjJL8EWfIJPeaAyV26vWbw55DO/u17wfVaQB0eXyWft 6VYyyDcwEeUUXdOZdMSgJFt4hLpCpKA= X-Stat-Signature: 1gq96t8grhp6b9yaey4q1ozidj3a5zp5 Authentication-Results: imf06.hostedemail.com; dkim=pass header.d=gmail.com header.s=20210112 header.b=nWxBmLtD; dmarc=fail reason="SPF not aligned (relaxed), DKIM not aligned (relaxed)" header.from=kernel.org (policy=none); spf=pass (imf06.hostedemail.com: domain of htejun@gmail.com designates 209.85.215.169 as permitted sender) smtp.mailfrom=htejun@gmail.com X-Rspam-User: X-Rspamd-Server: rspam12 X-Rspamd-Queue-Id: 86ABE180102 X-HE-Tag: 1661167591-926097 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: , Yosry Ahmed , Dan Schatzberg , Lennart Poettering Bcc: htejun@gmail.com Subject: [RFD] cgroup: Persistent memory usage tracking Reply-To: In-Reply-To: Hello, This thread started on a bpf-specific memory tracking change proposal and went south, but a lot of people who would be interested are already cc'd, so I'm hijacking it to discuss what to do w/ persistent memory usage tracking. Cc'ing Mina and Yosry who were involved in the discussions on the similar problem re. tmpfs, Dan Schatzberg who has a lot more prod knowledge and experience than me, and Lennart for his thoughts from systemd side. The root problem is that there are resources (almost solely memory currently) that outlive a given instance of a, to use systemd-lingo, service. Page cache is the most common case. Let's say there's system.slice/hello.service. When it runs for the first time, page cache backing its binary will be charged to hello.service. However, when it restarts after e.g. a config change, when the initial hello.service cgroup gets destroyed, we reparent the page cache charges to the parent system.slice and when the second instance starts, its binary will stay charged to system.slice. Over time, some may get reclaimed and refaulted into the new hello.service but that's not guaranteed and most hot pages likely won't. The same problem exists for any memory which is not freed synchronously when the current instance exits. While this isn't a problem for many cases, it's not difficult to imagine situations where the amount of memory which ends up getting pushed to the parent is significant, even clear majority, with big page cache footprint, persistent tmpfs instances and so on, creating issues with accounting accuracy and thus control. I think there are two broad issues to discuss here: [1] Can this be solved by layering the instance cgroups under persistent entity cgroup? So, instead of systemd.slice/hello.service, the application runs inside something like systemd.slice/hello.service/hello.service.instance and the service-level cgroup hello.service is not destroyed as long as it is something worth tracking on the system. The benefits are a. While requiring changing how userland organizes cgroup hiearchy, it is a straight-forward extension of the current architecture and doesn't require any conceptual or structural changes. All the accounting and control schemes work exactly the same as before. The only difference is that we now have a persistent entity representing each service as we want to track their persistent resource usages. b. Per-instance tracking and control is optional. To me, it seems that the persistent resource usages would be more meaningful than per-instance and tracking down to the persistent usages shouldn't add noticeable runtime overheads while keeping per-instance process management niceties and allowing use cases to opt-in for per-instance resource tracking and control as needed. The complications are: a. It requires changing cgroup hierarchy in a very visible way. b. What should be the lifetime rules for persistent cgroups? Do we keep them around forever or maybe they can be created on the first use and kept around until the service is removed from the system? When the persistent cgroup is removed, do we need to make sure that the remaining resource usages are low enough? Note that this problem exists for any approach that tries to track persistent usages no matter how it's done. c. Do we need to worry about nesting overhead? Given that there's no reason to enable controllers w/o persisten states for the instance level and the nesting overhead is pretty low for memcg, this doesn't seem like a problem to me. If this becomes a problem, we just need to fix it. A couple alternatives discussed are: a. Userspace keeps reusing the same cgroup for different instances of the same service. This simplifies some aspects while making others more complicated. e.g. Determining the current instance's CPU or IO usages now require the monitoring software remembering what they were when this instance started and calculating the deltas. Also, if some use cases want to distinguish persistent vs. instance usages (more on this later), this isn't gonna work. That said, this definitely is attractive in that it miminizes overt user visible changes. b. Memory is disassociated rather than just reparented on cgroup destruction and get re-charged to the next first user. This is attractive in that it doesn't require any userspace changes; however, I'm not sure how this would work for non-pageable memory usages such as bpf maps. How would we detect the next first usage? [2] Whether and how to solve first and second+ instance charge differences. If we take the layering approach, the first instance will get charged for all memory that it uses while the second+ instances likely won't get charged for a lot of persistent usages. I don't think there is a consensus on whether this needs to be solved and I don't have enough context to form a strong opinion. memcg folks are a lot better equipped to make this decision. Assuming this needs to be solved, here's a braindump to be taken with a big pinch of salt: I have a bit of difficult time imagining a perfect solution given that whether a given page cache page is persistent or not would be really difficult to know (or maybe all page cache is persistent by default while anon is not). However, the problem still seems worthwhile to consider for big ticket items such as persistent tmpfs mounts and huge bpf maps as they can easily make the differences really big. If we want to solve this problem, here are options that I can think of: a. Let userspace put the charges where they belong using the current mechanisms. ie. Create persistent entities in the persistent parent cgroup while there's no current instance. Pro: It won't require any major kernel or interface changes. There still need to be some tweaking such as allowing tmpfs pages to be always charged to the cgroup which created the instance (maybe as long as it's an ancestor of the faulting cgroup?) but nothing too invasive. Con: It may not be flexible enough. b. Let userspace specify which cgroup to charge for some of constructs like tmpfs and bpf maps. The key problems with this approach are 1. How to grant/deny what can be charged where. We must ensure that a descendant can't move charges up or across the tree without the ancestors allowing it. 2. How to specify the cgroup to charge. While specifying the target cgroup directly might seem like an obvious solution, it has a couple rather serious problems. First, if the descendant is inside a cgroup namespace, it might be able to see the target cgroup at all. Second, it's an interface which is likely to cause misunderstandings on how it can be used. It's too broad an interface. One solution that I can think of is leveraging the resource domain concept which is currently only used for threaded cgroups. All memory usages of threaded cgroups are charged to their resource domain cgroup which hosts the processes for those threads. The persistent usages have a similar pattern, so maybe the service level cgroup can declare that it's the encompassing resource domain and the instance cgroup can say whether it's gonna charge e.g. the tmpfs instance to its own or the encompassing resource domain. This has the benefit that the user only needs to indicate its intention without worrying about how cgroups are composed and what their IDs are. It just indicates whether the given resource is persistent and if the cgroup hierarchy is set up for that, it gets charged that way and if not it can be just charged to itself. This is a shower thought but, if we allow nesting such domains (and maybe name them), we can use it for shared resources too so that co-services are put inside a shared slice and shared resources are pushed to the slice level. This became pretty long. I obviously have a pretty strong bias towards solving this within the current basic architecture but other than that most of these decisions are best made by memcg folks. We can hopefully build some consensus on the issue. Thanks. -- tejun