From mboxrd@z Thu Jan 1 00:00:00 1970 From: Patrick McHardy Subject: Re: [PATCH v3] netfilter: Xtables: idletimer target implementation Date: Thu, 10 Jun 2010 12:07:16 +0200 Message-ID: <4C10B954.9080603@trash.net> References: <1275592445-15555-1-git-send-email-luciano.coelho@nokia.com> <4C0F9B17.10504@trash.net> <1276096275.31892.117.camel@chilepepper> <1276105707.11199.12.camel@powerslave> <1276108939.11199.23.camel@powerslave> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 7bit Cc: ext Jan Engelhardt , "netfilter-devel@vger.kernel.org" , "netdev@vger.kernel.org" , Timo Teras To: Luciano Coelho Return-path: In-Reply-To: <1276108939.11199.23.camel@powerslave> Sender: netfilter-devel-owner@vger.kernel.org List-Id: netdev.vger.kernel.org Luciano Coelho wrote: > On Wed, 2010-06-09 at 19:48 +0200, Coelho Luciano (Nokia-D/Helsinki) > wrote: > >> On Wed, 2010-06-09 at 17:18 +0200, ext Jan Engelhardt wrote: >> >>>>>> + timer = __idletimer_tg_find_by_label(info->label); >>>>>> + if (!timer) { >>>>>> + spin_unlock(&list_lock); >>>>>> + timer = idletimer_tg_create(info); >>>>>> >>>>>> >>>>> How does this prevent creating the same timer twice? >>>>> >>>> The timer will only be created if __idletimer_tg_find_by_label() returns >>>> NULL, which means that no timer with that label has been found. "info" >>>> won't be the same if info->label is different, right? Or can it change >>>> on the fly? >>>> >>> One thing to be generally aware about is that things could potentially >>> be instantiated by another entity between the time a label was looked up >>> with negative result and the time one tries to add it. >>> It may thus be required to extend keeping the lock until after >>> idletimer_tg_create, in other words, lookup and create must be atomic >>> to the rest of the world. >>> >> Ahh, sure! I missed the actual point of Patrick's question. I had the >> idletimer_tg_create() inside the lock, but when I added the >> sysfs_create_file() there (which can sleep), I screwed up with the >> locking. >> >> I'll move the sysfs file creation to outside that function so I can keep >> the lock until after the timer is added to the list. Thanks for >> clarifying! >> > > Hmmm... after struggling with this for a while, I think it's not really > possible to simply create the sysfs file outside of the lock, because if > the sysfs creation fails, we will again risk a race condition. > > I think the only way is to delay the sysfs file creation and do it in a > workqueue. > Why don't you simply use a mutex instead of the spinlock? It would be better to only do the lookup once and store the timer pointer in the target structure anyways.