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 5647CC54EE9 for ; Tue, 20 Sep 2022 02:40:24 +0000 (UTC) Received: by kanga.kvack.org (Postfix) id DEC548000C; Mon, 19 Sep 2022 22:40:23 -0400 (EDT) Received: by kanga.kvack.org (Postfix, from userid 40) id D9C1880007; Mon, 19 Sep 2022 22:40:23 -0400 (EDT) X-Delivered-To: int-list-linux-mm@kvack.org Received: by kanga.kvack.org (Postfix, from userid 63042) id C8A478000C; Mon, 19 Sep 2022 22:40:23 -0400 (EDT) X-Delivered-To: linux-mm@kvack.org Received: from relay.hostedemail.com (smtprelay0016.hostedemail.com [216.40.44.16]) by kanga.kvack.org (Postfix) with ESMTP id BA60D80007 for ; Mon, 19 Sep 2022 22:40:23 -0400 (EDT) Received: from smtpin28.hostedemail.com (a10.router.float.18 [10.200.18.1]) by unirelay02.hostedemail.com (Postfix) with ESMTP id 8DA23120581 for ; Tue, 20 Sep 2022 02:40:23 +0000 (UTC) X-FDA: 79930909926.28.EEBCE85 Received: from out2.migadu.com (out2.migadu.com [188.165.223.204]) by imf23.hostedemail.com (Postfix) with ESMTP id C1B82140002 for ; Tue, 20 Sep 2022 02:40:22 +0000 (UTC) Date: Mon, 19 Sep 2022 19:40:13 -0700 DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=linux.dev; s=key1; t=1663641620; 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: in-reply-to:in-reply-to:references:references; bh=gZS5sqyWSO3yFeBcOAHmX3+4vNuQk4vlxXsy9wXRJYQ=; b=L6Imu+c68+a2ysnK5RgOIzLcFHaxaChY8+VJb198+3lRmsAb5UOsa/DjbnCwCQPjk50QUL k/IwMVtsg1xp+9fr84CH2pyUG+j48J44igAn/mK4NwHlDVAlISWOFhV2uhvt44x5yeNFG2 EMJe8LTS4WSVxoP8QUrdUU/s1UihOWE= X-Report-Abuse: Please report any abuse attempt to abuse@migadu.com and include these headers. From: Roman Gushchin To: Yafang Shao Cc: Tejun Heo , Alexei Starovoitov , Daniel Borkmann , Andrii Nakryiko , Martin Lau , Song Liu , Yonghong Song , john fastabend , KP Singh , Stanislav Fomichev , Hao Luo , Jiri Olsa , Johannes Weiner , Michal Hocko , Shakeel Butt , Muchun Song , Andrew Morton , Zefan Li , Cgroups , netdev , bpf , Linux MM Subject: Re: [PATCH bpf-next v3 00/13] bpf: Introduce selectable memcg for bpf map Message-ID: References: <20220902023003.47124-1-laoar.shao@gmail.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: X-Migadu-Flow: FLOW_OUT X-Migadu-Auth-User: linux.dev ARC-Seal: i=1; s=arc-20220608; d=hostedemail.com; t=1663641623; a=rsa-sha256; cv=none; b=RRgow2vJj3XxlcsEFKIQiLpyOIwpv2hY0i+BbFMn5W9ylJDmEfy42HvU8HbfbCe2+zEEtT uSjsl1P9NyJiZWedEcs024w/pNDfhz1N2OAdKLF8R2r9gNV3hGNv1GrAeAhA2cOeuOAd9A gBoD0O5nIMrBFltd9XZaALbDk6I5JyY= ARC-Authentication-Results: i=1; imf23.hostedemail.com; dkim=pass header.d=linux.dev header.s=key1 header.b=L6Imu+c6; dmarc=pass (policy=none) header.from=linux.dev; spf=pass (imf23.hostedemail.com: domain of roman.gushchin@linux.dev designates 188.165.223.204 as permitted sender) smtp.mailfrom=roman.gushchin@linux.dev ARC-Message-Signature: i=1; a=rsa-sha256; c=relaxed/relaxed; d=hostedemail.com; s=arc-20220608; t=1663641623; h=from:from:sender: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: in-reply-to:in-reply-to:references:references:dkim-signature; bh=gZS5sqyWSO3yFeBcOAHmX3+4vNuQk4vlxXsy9wXRJYQ=; b=A7R003xGSZuZLVIkLhWTXVo1JV6OqlwHD6/V9uM2mi8lf21NuPWK6hnlbkwvu/DKTd6Eqd 0Llkp9TiPbWq1XWLJR2IaK+BQf0Z9AnqWauaSSlaA4KKJoTn9jIO5djzHCGOh8FRL766p8 +x6GerUymzU7lhy36law4kKL2ZnRgYM= X-Rspamd-Queue-Id: C1B82140002 Authentication-Results: imf23.hostedemail.com; dkim=pass header.d=linux.dev header.s=key1 header.b=L6Imu+c6; dmarc=pass (policy=none) header.from=linux.dev; spf=pass (imf23.hostedemail.com: domain of roman.gushchin@linux.dev designates 188.165.223.204 as permitted sender) smtp.mailfrom=roman.gushchin@linux.dev X-Rspamd-Server: rspam08 X-Rspam-User: X-Stat-Signature: qtx35ae1jqaocc87qpbtdu5up8oqx88b X-HE-Tag: 1663641622-109762 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 Sun, Sep 18, 2022 at 11:44:48AM +0800, Yafang Shao wrote: > On Sat, Sep 17, 2022 at 12:53 AM Roman Gushchin > wrote: > > > > On Tue, Sep 13, 2022 at 02:15:20PM +0800, Yafang Shao wrote: > > > On Fri, Sep 9, 2022 at 12:13 AM Roman Gushchin wrote: > > > > > > > > On Thu, Sep 08, 2022 at 10:37:02AM +0800, Yafang Shao wrote: > > > > > On Thu, Sep 8, 2022 at 6:29 AM Roman Gushchin wrote: > > > > > > > > > > > > On Wed, Sep 07, 2022 at 05:43:31AM -1000, Tejun Heo wrote: > > > > > > > Hello, > > > > > > > > > > > > > > On Fri, Sep 02, 2022 at 02:29:50AM +0000, Yafang Shao wrote: > > > > > > > ... > > > > > > > > This patchset tries to resolve the above two issues by introducing a > > > > > > > > selectable memcg to limit the bpf memory. Currently we only allow to > > > > > > > > select its ancestor to avoid breaking the memcg hierarchy further. > > > > > > > > Possible use cases of the selectable memcg as follows, > > > > > > > > > > > > > > As discussed in the following thread, there are clear downsides to an > > > > > > > interface which requires the users to specify the cgroups directly. > > > > > > > > > > > > > > https://lkml.kernel.org/r/YwNold0GMOappUxc@slm.duckdns.org > > > > > > > > > > > > > > So, I don't really think this is an interface we wanna go for. I was hoping > > > > > > > to hear more from memcg folks in the above thread. Maybe ping them in that > > > > > > > thread and continue there? > > > > > > > > > > > > > > > > Hi Roman, > > > > > > > > > > > As I said previously, I don't like it, because it's an attempt to solve a non > > > > > > bpf-specific problem in a bpf-specific way. > > > > > > > > > > > > > > > > Why do you still insist that bpf_map->memcg is not a bpf-specific > > > > > issue after so many discussions? > > > > > Do you charge the bpf-map's memory the same way as you charge the page > > > > > caches or slabs ? > > > > > No, you don't. You charge it in a bpf-specific way. > > > > > > > > > > Hi Roman, > > > > > > Sorry for the late response. > > > I've been on vacation in the past few days. > > > > > > > The only difference is that we charge the cgroup of the processes who > > > > created a map, not a process who is doing a specific allocation. > > > > > > This means the bpf-map can be indepent of process, IOW, the memcg of > > > bpf-map can be indepent of the memcg of the processes. > > > This is the fundamental difference between bpf-map and page caches, then... > > > > > > > Your patchset doesn't change this. > > > > > > We can make this behavior reasonable by introducing an independent > > > memcg, as what I did in the previous version. > > > > > > > There are pros and cons with this approach, we've discussed it back > > > > to the times when bpf memcg accounting was developed. If you want > > > > to revisit this, it's maybe possible (given there is a really strong and likely > > > > new motivation appears), but I haven't seen any complaints yet except from you. > > > > > > > > > > memcg-base bpf accounting is a new feature, which may not be used widely. > > > > > > > > > > > > > > Yes, memory cgroups are not great for accounting of shared resources, it's well > > > > > > known. This patchset looks like an attempt to "fix" it specifically for bpf maps > > > > > > in a particular cgroup setup. Honestly, I don't think it's worth the added > > > > > > complexity. Especially because a similar behaviour can be achieved simple > > > > > > by placing the task which creates the map into the desired cgroup. > > > > > > > > > > Are you serious ? > > > > > Have you ever read the cgroup doc? Which clearly describe the "No > > > > > Internal Process Constraint".[1] > > > > > Obviously you can't place the task in the desired cgroup, i.e. the parent memcg. > > > > > > > > But you can place it into another leaf cgroup. You can delete this leaf cgroup > > > > and your memcg will get reparented. You can attach this process and create > > > > a bpf map to the parent cgroup before it gets child cgroups. > > > > > > If the process doesn't exit after it created bpf-map, we have to > > > migrate it around memcgs.... > > > The complexity in deployment can introduce unexpected issues easily. > > > > > > > You can revisit the idea of shared bpf maps and outlive specific cgroups. > > > > Lof of options. > > > > > > > > > > > > > > [1] https://www.kernel.org/doc/Documentation/cgroup-v2.txt > > > > > > > > > > > Beatiful? Not. Neither is the proposed solution. > > > > > > > > > > > > > > > > Is it really hard to admit a fault? > > > > > > > > Yafang, you posted several versions and so far I haven't seen much of support > > > > or excitement from anyone (please, fix me if I'm wrong). It's not like I'm > > > > nacking a patchset with many acks, reviews and supporters. > > > > > > > > Still think you're solving an important problem in a reasonable way? > > > > It seems like not many are convinced yet. I'd recommend to focus on this instead > > > > of blaming me. > > > > > > > > > > The best way so far is to introduce specific memcg for specific resources. > > > Because not only the process owns its memcg, but also specific > > > resources own their memcgs, for example bpf-map, or socket. > > > > > > struct bpf_map { <<<< memcg owner > > > struct memcg_cgroup *memcg; > > > }; > > > > > > struct sock { <<<< memcg owner > > > struct mem_cgroup *sk_memcg; > > > }; > > > > > > These resources already have their own memcgs, so we should make this > > > behavior formal. > > > > > > The selectable memcg is just a variant of 'echo ${proc} > cgroup.procs'. > > > > This is a fundamental change: cgroups were always hierarchical groups > > of processes/threads. You're basically suggesting to extend it to > > hierarchical groups of processes and some other objects (what's a good > > definition?). > > Kind of, but not exactly. > We can do it without breaking the cgroup hierarchy. Under current > cgroup hierarchy, the user can only echo processes/threads into a > cgroup, that won't be changed in the future. The specific resources > are not exposed to the user, the user can only control these specific > resources by controlling their associated processes/threads. > For example, > > Memcg-A > |---- Memcg-A1 > |---- Memcg-A2 > > We can introduce a new file memory.owner into each memcg. Each bit of > memory.owner represents a specific resources, > > memory.owner: | bit31 | bitN | ... | bit1 | bit0 | > | | > |------ bit0: bpf memory > | > |-------------- bit1: socket memory > | > |--------------------------- > bitN: a specific resource > > There won't be too many specific resources which have to own their > memcgs, so I think 32bits is enough. > > Memcg-A : memory.owner == 0x1 > |---- Memcg-A1 : memory.owner == 0 > |---- Memcg-A2 : memory.owner == 0x1 > > Then the bpf created by processes in Memcg-A1 will be charged into > Memcg-A directly without charging into Memcg-A1. > But the bpf created by processes in Memcg-A2 will be charged into > Memcg-A2 as its memory.owner is 0x1. > That said, these specific resources are not fully independent of > process, while they are still associated with the processes which > create them. > Luckily memory.move_charge_at_immigrate is disabled in cgroup2, so we > don't need to care about the possible migration issue. > > I think we may also apply it to shared page caches. For example, > struct inode { > struct mem_cgroup *memcg; <<<< add a new member > }; > > We define struct inode as a memcg owner, and use scope-based charge to > charge its pages into inode->memcg. > And then put all memcgs which shared these resources under the same > parent. The page caches of this inode will be charged into the parent > directly. Ok, so it's something like premature selective reparenting. > The shared page cache is more complicated than bpf memory, so I'm not > quite sure if it can apply to shared page cache, but it can work well > for bpf memory. Yeah, this is the problem. It feels like it's a problem very specific to bpf maps and an exact way you use them. I don't think you can successfully advocate for changes of these calibre without a more generic problem. I might be wrong. > > Regarding the observability, we can introduce a specific item into > memory.stat for this specific memory. For example a new item 'bpf' for > bpf memory. > That can be accounted/unaccounted for in the same way as we do in > set_active_memcg(). for example, > > struct task_struct { > struct mem_cgroup *active_memcg; > int active_memcg_item; <<<< > introduce a new member > }; > > bpf_memory_alloc() > { > old_memcg = set_active_memcg(memcg); > old_item = set_active_memcg_item(MEMCG_BPF); I thought about something like this but for a different purpose: to track the amount of memory consumed by bpf. > alloc(); > set_active_memcg_item(old_item); > set_active_memcg(old_memcg); > } > > bpf_memory_free() > { > old = set_active_memcg_item(MEMCG_BPF); > free(); > set_active_memcg_item(old); > } But the problem is that we shoud very carefully mark all allocations and releases, which is very error-prone. Interfaces which don't require annotating releases are generally better, but require additional memory. Thanks!