Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Thu, 28 May 2020 17:06:13 +0300
From:   Dan Carpenter <dan.carpenter@oracle.com>
To:     Lai Jiangshan <jiangshanlai@gmail.com>
Cc:     "Zhang, Qiang" <Qiang.Zhang@windriver.com>,
        Markus Elfring <markus.elfring@web.de>,
        Tejun Heo <tj@kernel.org>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        "kernel-janitors@vger.kernel.org" <kernel-janitors@vger.kernel.org>
Subject: Re: =?utf-8?B?5Zue5aSNOiBbUEFUQ0ggdjU=?= =?utf-8?Q?=5D?= workqueue:
 Remove unnecessary kfree() call in rcu_free_wq()
Message-ID: <20200528140613.GK30374@kadam>
References: <20200527075715.36849-1-qiang.zhang@windriver.com>
 <284c7851-4e89-a00f-a2e6-aa8e2e1f3fce@web.de>
 <DM6PR11MB32573F3884A864ECD586235EFF8E0@DM6PR11MB3257.namprd11.prod.outlook.com>
 <DM6PR11MB3257D6E7E93A518392502809FF8E0@DM6PR11MB3257.namprd11.prod.outlook.com>
 <20200528095703.GH30374@kadam>
 <CAJhGHyD1nV=M=ccycqCMt86GMuZGkO9trbJ=4ti4EzP9kta6iA@mail.gmail.com>
 <20200528122545.GP22511@kadam>
 <CAJhGHyBUkMZ=cV+Qf-5+PMAFqgebbRLc46OZSSUSgoRROpUk2A@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <CAJhGHyBUkMZ=cV+Qf-5+PMAFqgebbRLc46OZSSUSgoRROpUk2A@mail.gmail.com>
User-Agent: Mutt/1.9.4 (2018-02-28)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, May 28, 2020 at 09:27:03PM +0800, Lai Jiangshan wrote:
> On Thu, May 28, 2020 at 8:27 PM Dan Carpenter <dan.carpenter@oracle.com> wrote:
> >
> > On Thu, May 28, 2020 at 08:08:06PM +0800, Lai Jiangshan wrote:
> > > On Thu, May 28, 2020 at 5:57 PM Dan Carpenter <dan.carpenter@oracle.com> wrote:
> > > >
> > > > Guys, the patch is wrong.  The kfree is harmless when this is called
> > > > from destroy_workqueue() and required when it's called from
> > > > pwq_unbound_release_workfn().  Lai Jiangshan already explained this
> > > > already.  Why are we still discussing this?
> > > >
> > >
> > > I'm also confused why they have been debating about the changelog
> > > after the patch was queued. My statement was about "the patch is
> > > a correct cleanup, but the changelog is totally misleading".
> > >
> > > destroy_workqueue(percpu_wq) -> rcu_free_wq()
> > > or
> > > destroy_workqueue(unbound_wq) -> put_pwq() ->
> > > pwq_unbound_release_workfn() -> rcu_free_wq()
> > >
> > > So the patch is correct to me. Only can destroy_workqueue()
> > > lead to rcu_free_wq().
> >
> > It looks like there are lots of paths which call put_pwq() and
> > put_pwq_unlocked().
> >
> >   1168  static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
> >   1169  {
> >   1170          /* uncolored work items don't participate in flushing or nr_active */
> >   1171          if (color == WORK_NO_COLOR)
> >   1172                  goto out_put;
> >   1173
> >
> > We don't take an extra reference in this function.
> >
> >   1200  out_put:
> >   1201          put_pwq(pwq);
> >   1202  }
> >
> > I don't know this code well, so I will defer to your expertise if you
> > say it is correct.
> 
> wq owns the ultimate or permanent references to itself by
> owning references to wq->numa_pwq_tbl[node], wq->dfl_pwq.
> The pwq's references keep the pwq in wq->pwqs.
> 
> Only destroy_workqueue() can release these ultimate references
> and then (after maybe a period of time) deplete the wq->pwqs
> finally and then free itself in rcu callback.
> 
> Actually, in short, we don't need the care about the above
> detail. The responsibility to free rescuer is on
> destroy_workqueue(), and since it does the free early,
> it doesn't need to do it again later.
> 
> e2dca7adff8f moved the free of rescuer into rcu callback,
> and I didn't check all the changes between then and now.
> But at least now, the rescuer is not accessed in rcu mananer,
> so we don't need to free it in rcu mananer.
> 

Ah...  Thanks for the explanation!

regards,
dan carpenter