From mboxrd@z Thu Jan 1 00:00:00 1970 Return-path: Received: from mx1.redhat.com ([209.132.183.28]:63226 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754825Ab0IAOUI (ORCPT ); Wed, 1 Sep 2010 10:20:08 -0400 Date: Wed, 1 Sep 2010 16:16:43 +0200 From: Stanislaw Gruszka To: Johannes Berg Cc: Wey-Yi Guy , Reinette Chatre , "John W. Linville" , linux-wireless@vger.kernel.org Subject: Re: [PATCH w-t] iwlwifi: rewrite iwl-scan.c to avoid race conditions Message-ID: <20100901141643.GC6547@redhat.com> References: <20100831150021.GA10963@redhat.com> <1283338760.4131.22.camel@jlt3.sipsolutions.net> <20100901115826.GB6547@redhat.com> <1283344912.4124.2.camel@jlt3.sipsolutions.net> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii In-Reply-To: <1283344912.4124.2.camel@jlt3.sipsolutions.net> Sender: linux-wireless-owner@vger.kernel.org List-ID: On Wed, Sep 01, 2010 at 02:41:52PM +0200, Johannes Berg wrote: > > Custom workqueue is not strictly needed, but it assure all works will > > run short after schedule. Common workqueue can not give us such guarantees, > > as other driver/subsystem can schedule own work, possibly slow, which > > can block start of our work for long time. > > I believe that's no longer true, with Tejun's workqueue rewrite that > just got into mainline. Good. However I still think having separate workqueue for scanning is clean and consistent solution, I prefer it over schedule_work() ... and want that patch backport to RHEL6 2.6.32 where we do not have such goodies :-) > > > Shouldn't that only be a few of them anyway? > > > > For sure abort_scan and abort_timeout works have to be scheduled on something > > other than priv->workqueue. I'm queuing all scan works on priv->scan_workqueue > > for consistency. > > Remind me: The reason is that we need to cancel them from within the > workqueue? But if we're on the same workqueue, it seems like they > couldn't be running already, so cancel_work_sync() would always cancel > them? Problem is not canceling, but exactly that we can not run new work when old one does not finish. For example, if queued to priv->workqueue abort_timeout will not be able run when we are performing iwl_bg_restart. Will run after iwl_bg_restart finish, we don't want that. > Do we get lockdep errors for that? Or are there actual locks > involved? No, priv->mutex used to protect critical sections of functions, but it's not keep all the time in works (i.e: iwl_bg_restart), so give scan code chance to complete. Stanislaw