Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752697AbbGIA4j (ORCPT ); Wed, 8 Jul 2015 20:56:39 -0400 Received: from mail-ig0-f176.google.com ([209.85.213.176]:35619 "EHLO mail-ig0-f176.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751412AbbGIA4b (ORCPT ); Wed, 8 Jul 2015 20:56:31 -0400 Date: Wed, 8 Jul 2015 17:56:25 -0700 From: Dmitry Torokhov To: Greg Kroah-Hartman Cc: Jason Gunthorpe , Jens Wiklander , linux-kernel@vger.kernel.org, linux-arm-kernel@lists.infradead.org, devicetree@vger.kernel.org, Arnd Bergmann , Rob Herring , Herbert Xu , valentin.manea@huawei.com, jean-michel.delorme@st.com, emmanuel.michel@st.com, javier@javigon.com, Mark Rutland , Michal Simek Subject: Re: [PATCH v4 3/5] tee: generic TEE subsystem Message-ID: <20150709005625.GF379@dtor-ws> References: <1436350592-7732-1-git-send-email-jens.wiklander@linaro.org> <1436350592-7732-4-git-send-email-jens.wiklander@linaro.org> <20150708171026.GA11740@obsidianresearch.com> <20150708211129.GA29824@kroah.com> <20150708222649.GA20068@obsidianresearch.com> <20150708223325.GA5843@kroah.com> <20150708232826.GB379@dtor-ws> <20150708235203.GA12393@kroah.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20150708235203.GA12393@kroah.com> User-Agent: Mutt/1.5.21 (2010-09-15) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 5561 Lines: 122 On Wed, Jul 08, 2015 at 04:52:03PM -0700, Greg Kroah-Hartman wrote: > On Wed, Jul 08, 2015 at 04:28:26PM -0700, Dmitry Torokhov wrote: > > On Wed, Jul 08, 2015 at 03:33:25PM -0700, Greg Kroah-Hartman wrote: > > > On Wed, Jul 08, 2015 at 04:26:49PM -0600, Jason Gunthorpe wrote: > > > > On Wed, Jul 08, 2015 at 02:11:29PM -0700, Greg Kroah-Hartman wrote: > > > > > > > + cdev_init(&teedev->cdev, &tee_fops); > > > > > > > + teedev->cdev.owner = teedesc->owner; > > > > > > > > > > > > This also needs to set teedev->cdev.kobj.parent. > > > > > > I'm guessing: > > > > > > > > > > > > teedev->cdev.kobj.parent = &teedev->dev.kobj; > > > > > > > > > > > > TPM had the same mistake.. > > > > > > > > > > Really? As of a few years ago, A cdev's kobject should not be touched > > > > > by anything other than the cdev core. It's not a "real" kobject in that > > > > > it is never registered in sysfs, and no one sees it. I keep meaning to > > > > > > > > Well, when I looked at it, it looked like it was necessary to maintain > > > > the refcount on the memory that is holding cdev. > > > > > > > > The basic issue is that cdev_del doesn't seem to be synchronizing. > > > > > > > > The use after free race is then something like: > > > > > > > > struct tpm_chip { > > > > struct device dev; > > > > struct cdev cdev; > > > > > > Oops, right there's your problem. You can't have two reference counted > > > objects trying to manage the memory of a single structure. No matter > > > what you do, it's going to be a pain to deal with this, so don't :) > > > > > > > > > > > CPU0 CPU1 > > > > ================= ====================== > > > > tpm_chip = kalloc > > > > cdev_add(&tpm_chip->cdev) > > > > device_add(&tpm_chip->dev) > > > > chrdev_open > > > > filp->f_op->open > > > > cdev_del(&tpm_chip->cdev) > > > > device_unregister > > > > (&tpm_chip->dev) > > > > kfree(tpm_chip) > > > > tpm_chip = container_of > > > > fput > > > > cdev_put(.. cdev) > > > > > > > > Ie we need cdev to hold a ref on tpm_chip->dev until cdev_put is > > > > called. > > > > > > No, separate them, make the cdev a pointer and all should be fine. > > > > > > > > just use something else one of these days for that structure, as lots of > > > > > people get it wrong. Or has things changed there? > > > > > > > > Not recently, but this is the commit: > > > > > > > > commit 2f0157f13f42800aa3d9017ebb0fb80a65f7b2de > > > > Author: Dmitry Torokhov > > > > Date: Sun Oct 21 17:57:19 2012 -0700 > > > > > > > > char_dev: pin parent kobject > > > > > > > > In certain cases (for example when a cdev structure is embedded into > > > > another object whose lifetime is controlled by a separate kobject) it is > > > > beneficial to tie lifetime of another object to the lifetime of > > > > character device so that related object is not freed until after > > > > char_dev object is freed. > > > > > > > > To achieve this let's pin kobject's parent when doing cdev_add() and > > > > unpin when last reference to cdev structure is being released. > > > > > > > > Signed-off-by: Dmitry Torokhov > > > > Acked-by: Al Viro > > > > Signed-off-by: Linus Torvalds > > > > > > > > It doesn't seem the be the best situation, this is the 3rd time this > > > > week I've noticed cdev with a kalloc'd struct being used improperly. > > > > > > > > Perhaps cdev_init should accept the module and kref parent as an > > > > argument? > > > > > > Oh yeah, that commit :( > > > > > > If you know _exactly_ what you are doing, you can get away with this, > > > but I strongly recommend not doing that. As proof of that, in some new > > > code I'm working on, I did not do this, just because I'm not smart > > > enough to ensure it's all working properly... > > > > I know you like to allocate everything separately and access it via > > pointers (ala device_create) but cdevs explicitly allow embedding them > > into other structures (cdev_init vs cdev_alloc). I do not think there is > > anything wrong with this, as well as there is nothing wrong in embedding > > a struct device into other structures, but it does require coordinating > > lifetime rules and selecting a "master" kobject. I think having > > cdev_init accept such "master" kobject would bring author's attention to > > the issue and avoid such mistakes in the future. > > Embedding cdevs into other structures is great, I like that. What I > don't like is having two different reference counts for the same > structure based on the lifetime rules of two different embedded > structures. That's a very difficult thing to get right and I would > argue, is something that should almost never be done. I think if we stop thinking about reference counts in the context of releasing memory that will clear the issue. The reference count for cdev is for cdev alone. The reference count for struct device is for the struct device. What happens when either one or both of them drop to 0 is up to the driver controlling the objects. Thanks. -- Dmitry -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/