netdev - Re: ibmvnic: Race condition in remove callback

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Date:   Tue, 19 Jan 2021 20:18:31 +0100
From:   Uwe Kleine-König <u.kleine-koenig@...gutronix.de>
To:     Dany Madden <drt@...ux.ibm.com>
Cc:     Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
        Juliet Kim <julietk@...ux.vnet.ibm.com>,
        Michael Ellerman <mpe@...erman.id.au>,
        Jakub Kicinski <kuba@...nel.org>, netdev@...r.kernel.org,
        Lijun Pan <ljp@...ux.ibm.com>, kernel@...gutronix.de,
        Benjamin Herrenschmidt <benh@...nel.crashing.org>,
        Paul Mackerras <paulus@...ba.org>,
        Sukadev Bhattiprolu <sukadev@...ux.ibm.com>,
        linuxppc-dev@...ts.ozlabs.org,
        "David S. Miller" <davem@...emloft.net>
Subject: Re: ibmvnic: Race condition in remove callback

Hello Dany,

On Tue, Jan 19, 2021 at 10:14:25AM -0800, Dany Madden wrote:
> On 2021-01-17 02:12, Uwe Kleine-König wrote:
> > while working on some cleanup I stumbled over a problem in the ibmvnic's
> > remove callback. Since commit
> > 
> >         7d7195a026ba ("ibmvnic: Do not process device remove during
> > device reset")
> > 
> > there is the following code in the remove callback:
> > 
> >         static int ibmvnic_remove(struct vio_dev *dev)
> >         {
> >                 ...
> >                 spin_lock_irqsave(&adapter->state_lock, flags);
> >                 if (test_bit(0, &adapter->resetting)) {
> >                         spin_unlock_irqrestore(&adapter->state_lock,
> > flags);
> >                         return -EBUSY;
> >                 }
> > 
> >                 adapter->state = VNIC_REMOVING;
> >                 spin_unlock_irqrestore(&adapter->state_lock, flags);
> > 
> >                 flush_work(&adapter->ibmvnic_reset);
> >                 flush_delayed_work(&adapter->ibmvnic_delayed_reset);
> >                 ...
> >         }
> > 
> > Unfortunately returning -EBUSY doesn't work as intended. That's because
> > the return value of this function is ignored[1] and the device is
> > considered unbound by the device core (shortly) after ibmvnic_remove()
> > returns.
> 
> Oh! I was not aware of this.

There are not very many people who are aware. That's why I'm working on
making all these remove callbacks return void. (And that's how I stubled
over this driver's expectations when returning -EBUSY.)

> In our code review, a question on whether or not device reset should
> have a higher precedence over device remove was raised before. So, now
> it is clear that this driver has to take care of remove over reset.

Yes, either you have to sleep in .remove until the reset is completed,
or you cancel the request.

> > While looking into fixing that I noticed a worse problem:
> > 
> > If ibmvnic_reset() (e.g. called by the tx_timeout callback) calls
> > schedule_work(&adapter->ibmvnic_reset); just after the work queue is
> > flushed above the problem that 7d7195a026ba intends to fix will trigger
> > resulting in a use-after-free.
> 
> It was proposed that when coming into ibmvnic_remove() we lock down the
> workqueue to prevent future access, flush, cleanup, then unregister the
> device. Your thought on this?

This is already done a bit, as ibmvnic_reset() checks for adapter->state
== VNIC_REMOVING. The problem is just that this check is racy.

> > Also ibmvnic_reset() checks for adapter->state without holding the lock
> > which might be racy, too.
>
> Suka started addressing consistent locking with this patch series:
> https://lists.openwall.net/netdev/2021/01/08/89
> 
> He is reworking this.

Please understand that I won't look into this. This is not in my area of
expertise and interest and I'd like to consider this problem already
done for me with my report. I'm glad you're acting on it. Depending on
how quick this is fixed I plan to submit a patch that changes

	return -EBUSY;

to

	return 0;

to prepare changing the prototype of the remove callback to return void.

> Thank you for taking the time to review this driver, Uwe. This is very
> helpful for us.

You're welcome.

Best regards
Uwe

-- 
Pengutronix e.K.                           | Uwe Kleine-König            |
Industrial Linux Solutions                 | https://www.pengutronix.de/ |

Download attachment "signature.asc" of type "application/pgp-signature" (489 bytes)