From mboxrd@z Thu Jan 1 00:00:00 1970 From: Waiman Long Subject: Re: [PATCH 3/3] cgroup/cpuset: Allow only one active attach operation per cpuset Date: Mon, 3 Apr 2023 13:41:33 -0400 Message-ID: <24b67530-62ce-4f9c-7b74-d41d2ccc710e@redhat.com> References: <20230331145045.2251683-1-longman@redhat.com> <20230331145045.2251683-4-longman@redhat.com> <20230403164736.lpjdpzxxnjlpxrqv@blackpad> Mime-Version: 1.0 Content-Transfer-Encoding: quoted-printable Return-path: DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=redhat.com; s=mimecast20190719; t=1680543700; 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=du5ube66GFFv6eiC7GgKY4F8MRZGEnQkABRtMIiA1tM=; b=J6cASXk5i9MHrtpZDY8dEZvmUQKNuPizeI6WliI719tDfaYDiQEb6IJyq3fn1lH3ATkSzJ P3lNcBzS5+aERcfFSpAisMZ2HjCQ16slHA4jaRYYu4l2E1TOcaw1hCQ0iu2zMRy3zGOHXB taKagSrR63RGIUIJAGRbr70nEqz9w2c= Content-Language: en-US In-Reply-To: <20230403164736.lpjdpzxxnjlpxrqv@blackpad> List-ID: Content-Type: text/plain; charset="iso-8859-1"; format="flowed" To: =?UTF-8?Q?Michal_Koutn=c3=bd?= Cc: Tejun Heo , Zefan Li , Johannes Weiner , Christian Brauner , cgroups-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, linux-kernel-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, Juri Lelli , Dietmar Eggemann , gscrivan-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org On 4/3/23 12:47, Michal Koutn=C3=BD wrote: > On Fri, Mar 31, 2023 at 10:50:45AM -0400, Waiman Long wrote: >> The current cpuset code uses the global cpuset_attach_old_cs variable >> to store the old cpuset value between consecutive cpuset_can_attach() >> and cpuset_attach() calls. Since a caller of cpuset_can_attach() may >> not need to hold the global cgroup_threadgroup_rwsem, parallel cpuset >> attach operations are possible. > Do I understand correctly this consequence of the cpuset_attach_task() > on the clone path? > In that particular case (with CLONE_INTO_CGROUP) cgroup_mutex is taken, > so the access the the old_cs variable should still be synchronized with > regular migrations that are also under cgroup_mutex. This patch is actually not related to the CLONE_INTO_GROUP problem in=20 patch 1. It is a generic problem when multiple users are moving threads=20 into cgroup.threads of the same or different cpusets simultaneously. >> When there are concurrent cpuset attach operations in progress, >> cpuset_attach() may fetch the wrong value from cpuset_attach_old_cs >> causing incorrect result. To avoid this problem while still allowing >> certain level of parallelism, drop cpuset_attach_old_cs and use a >> per-cpuset attach_old_cs value. Also restrict to at most one active >> attach operation per cpuset to avoid corrupting the value of the >> per-cpuset attach_old_cs value. > Secondly, semantically wouldn't a `void *ss_priv[CGROUP_SUBSYS_COUNT]` > in struct cgroup_taskset make it simpler wrt the exclusivity guarantees? I guess we can put the old_cs value into cgroup_taskset. Since the=20 related attach_in_progress value is in cpuset, so I put the old_cs there=20 too. > Thirdly, if my initial assumptino is right -- I'd suggest ordering this > before the patch `cgroup/cpuset: Make cpuset_fork() handle > CLONE_INTO_CGROUP properly` to spare backporters possible troubles if > this is would be a fixup to that. I don't believe this patch has a dependency on patch 1. I had thought about adding a fixes tag so it will be backported to=20 stable. However, this problem should be rare. Let's see what others=20 think about it. Cheers, Longman