Received-SPF: pass (google.com: domain of linux-wireless-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <0c44265eae421eff49e19be3ebfe20d1fb5e6f9a.camel@sipsolutions.net>
Subject: Re: [RFC PATCH 2/4] workqueue: support holding a mutex for each work
From:   Johannes Berg <johannes@sipsolutions.net>
To:     Tejun Heo <tj@kernel.org>
Cc:     linux-kernel@vger.kernel.org, linux-wireless@vger.kernel.org,
        Lai Jiangshan <jiangshanlai@gmail.com>
Date:   Wed, 10 May 2023 21:16:09 +0200
In-Reply-To: <ZFvjoUtg2ax11UlC@slm.duckdns.org>
References: <20230510160428.175409-1-johannes@sipsolutions.net>
         <20230510175846.cc21c84b0e6b.I9d3df459c43a78530d9c2046724bb45626402d5f@changeid>
         <ZFvjoUtg2ax11UlC@slm.duckdns.org>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
User-Agent: Evolution 3.46.4 (3.46.4-1.fc37) 
MIME-Version: 1.0
Precedence: bulk

On Wed, 2023-05-10 at 08:34 -1000, Tejun Heo wrote:
> On Wed, May 10, 2023 at 06:04:26PM +0200, Johannes Berg wrote:
> > @@ -2387,7 +2389,13 @@ __acquires(&pool->lock)
> >  	 */
> >  	lockdep_invariant_state(true);
> >  	trace_workqueue_execute_start(work);
> > -	worker->current_func(work);
> > +	if (unlikely(pwq->wq->work_mutex)) {
> > +		mutex_lock(pwq->wq->work_mutex);
> > +		worker->current_func(work);
> > +		mutex_unlock(pwq->wq->work_mutex);
> > +	} else {
> > +		worker->current_func(work);
> > +	}
>=20
> Ah, I don't know about this. This can't be that difficult to do from the
> callee side, right?
>=20

Yeah I thought you'd say that :)

It isn't difficult, the issue is just that in the case I'm envisioning,
you can't just call wiphy_lock() since that would attempt to pause the
workqueue, which can't work from on the workqueue itself. So you need
wiphy_lock_from_work()/wiphy_unlock_from_work() or remember to use the
mutex directly there, which all seemed more error-prone and harder to
maintain.

But anyway I could easily implement _both_ of these in cfg80211
directly, with just a linked list of works and a single struct
work_struct to execute things on the list, with the right locking. That
might be easier overall, just at the expense of more churn while
converting, but that's not even necessarily _bad_, it would really
guarantee that we can tell immediately the work is properly done...

I'll play with that idea some, I guess. Would you still want the
pause/resume patch anyway, even if I end up not using it then?

johannes