From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S932293Ab1IAM6a (ORCPT ); Thu, 1 Sep 2011 08:58:30 -0400 Received: from mail-wy0-f174.google.com ([74.125.82.174]:51924 "EHLO mail-wy0-f174.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S932125Ab1IAM63 (ORCPT ); Thu, 1 Sep 2011 08:58:29 -0400 Date: Thu, 1 Sep 2011 14:58:24 +0200 From: Frederic Weisbecker To: Tejun Heo Cc: rjw@sisk.pl, paul@paulmenage.org, lizf@cn.fujitsu.com, linux-pm@lists.linux-foundation.org, linux-kernel@vger.kernel.org, containers@lists.linux-foundation.org, matthltc@us.ibm.com, kamezawa.hiroyu@jp.fujitsu.com, Oleg Nesterov Subject: Re: [PATCH 6/6] cgroup: kill subsys->can_attach_task(), pre_attach() and attach_task() Message-ID: <20110901125822.GB24650@somewhere.redhat.com> References: <1314312192-26885-1-git-send-email-tj@kernel.org> <1314312192-26885-7-git-send-email-tj@kernel.org> <20110830201030.GC15953@somewhere.redhat.com> <20110831070313.GA29179@mtj.dyndns.org> <20110831134220.GB20598@somewhere> <20110901112221.GA2752@htj.dyndns.org> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20110901112221.GA2752@htj.dyndns.org> User-Agent: Mutt/1.5.21 (2010-09-15) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Thu, Sep 01, 2011 at 08:22:21PM +0900, Tejun Heo wrote: > Hello, > > On Wed, Aug 31, 2011 at 03:42:24PM +0200, Frederic Weisbecker wrote: > > My task counter subsystem patchset brings a cancel_attach_task() callback > > that cancels can_attach_task() effects. > > > > I thought that rebased on top of your patch it's going to be merged inside > > cancel_attach() but OTOH we can't cancel the effect of failed migration > > on a thread that way. > > > > May be we need to keep a cancel_attach_task() just for that purpose? > > We can do that but I think that becomes a bit too complex and fragile. > That path won't be traveled unless it races against exit. Bugs will > be difficult to detect and reproduce. In this respect, the current > code already seems racy. ->can_attach (or other methods in the attach > path) and ->exit can race each other and I don't think all subsystems > handle that properly. I guess subsystems don't care about that currently, although I haven't checked. But the task counter will need this per thread cancellation in migration failure, without the need for any synchronization between can_attach, attach and exit. Now if we want to solve this with a synchronization there, either we task_lock() every tasks in the group in the beginning of cgroup_attach_proc(). But that's not very nice. Or we use cgroup_mutex on cgroup_exit() exit, but that's even worse. I guess we need the leader->sighand->siglock on both cgroup_attach_proc() and cgroup_exit(). Besides we may have more reasons to have that: https://lkml.org/lkml/2011/8/15/342 > > IMHO the right thing to do here is simplifying synchronization rules > so that nothing else happens while migration is in progress. > > Thanks. > > -- > tejun