From mboxrd@z Thu Jan 1 00:00:00 1970 From: Rusty Russell Subject: Re: [RFC PATCH v2 next]module: Fix mod->mkobj.kobj potentially freed too early Date: Mon, 02 Sep 2013 09:21:55 +0930 Message-ID: <871u58ayn8.fsf@rustcorp.com.au> References: <1377078598.2709.25.camel@ThinkPad-T5421> <20130821161819.GA14364@kroah.com> <1377138846.2633.25.camel@ThinkPad-T5421> <20130822040333.GB4821@kroah.com> <1377157075.2633.63.camel@ThinkPad-T5421> <20130825040734.GA18107@kroah.com> <87bo4jd9z0.fsf@rustcorp.com.au> <20130827062135.GA3395@kroah.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Return-path: In-Reply-To: <20130827062135.GA3395@kroah.com> Sender: linux-kernel-owner@vger.kernel.org To: Greg KH Cc: Li Zhong , linux-next list , LKML , rmk+kernel@arm.linux.org.uk List-Id: linux-next.vger.kernel.org Greg KH writes: > On Tue, Aug 27, 2013 at 02:08:27PM +0930, Rusty Russell wrote: >> Greg KH writes: >> > On Thu, Aug 22, 2013 at 03:37:55PM +0800, Li Zhong wrote: >> >> DEBUG_KOBJECT_RELEASE helps to find the issue attached below. >> > People are starting to hit these types of issues, and I'd like to take >> > this one out of the picture. >> > >> > Rusty, any objection to me taking this through my driver-core tree, >> > where this new config option shows up? >> >> The original fix was better. >> >> Moving the module_kobject out and giving it its own lifetime solves this >> immediate issue, but it still means there's an accessible module_kobject >> around referring to a module which doesn't exist any more. > > That's ok, it could happen before as well. What's wrong with that? > >> Original copied below, feel free to take it. > > You are just sitting and sleeping until someone drops the last reference > to the module. What if userspace grabs a reference from sysfs? That > could never return, I don't think you want to stall that out. In your scenario, what happens if userspace grabs a reference via sysfs? It then tries to use module_kobj->mod which points into freed memory? eg. show_modinfo_##field or show_refcnt. Is there an owner field I'm missing somewhere which stops this from happening? Otherwise, we can't unload the module until it's done. > I'd prefer not having 2 things determining the lifecycle of a single > object, that's messy, and not needed at all. Normally you'd grab a reference to the module via an owner pointer. Doing that in kobject seems like overkill, so we're working around it here... Hope that clarifies, Rusty.