Date:   Wed, 8 Dec 2021 15:58:38 -0800
From:   Jakub Kicinski <kuba@kernel.org>
To:     George Kennedy <george.kennedy@oracle.com>
Cc:     Stephen Hemminger <stephen@networkplumber.org>,
        Greg KH <gregkh@linuxfoundation.org>, davem@davemloft.net,
        netdev@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH] tun: avoid double free in tun_free_netdev
Message-ID: <20211208155838.24556030@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com>
In-Reply-To: <b141489b-780c-1753-2a83-ccb60c4554d0@oracle.com>
References: <1638974605-24085-1-git-send-email-george.kennedy@oracle.com>
        <YbDR/JStiIco3HQS@kroah.com>
        <022193b1-4ddd-f04e-aafa-ce249ec6d120@oracle.com>
        <20211208083614.61f386ad@hermes.local>
        <b141489b-780c-1753-2a83-ccb60c4554d0@oracle.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Precedence: bulk

On Wed, 8 Dec 2021 11:44:02 -0500 George Kennedy wrote:
> > It looks like a lot of the problem is duplicate unwind.
> > Why does err_free_flow, err_free_stat etc unwinds need to exist if
> > the free_netdev is going to do same thing.  
> 
> Maybe instead do not call security_tun_dev_free_security(tun->security) 
> in err_free_flow if it's going to be done anyway in tun_free_netdev().

That won't be good either. register_netdevice() has multiple failure
modes, it may or may not call the destructor depending on where it
fails. Either the stuff that destructor undoes needs to be moved to
ndo_init (which is what destructor always pairs with), or you can check
dev->reg_state. If dev->reg_state is NETREG_UNREGISTERING that means
the destructor will be caller later.

The ndo_init way is preferable, just cut and past the appropriate lines
preceding registration into a ndo_init callback.