ppp: fix refcount underflow on channel unbridge

When setting up a channel bridge, ppp_bridge_channels sets the
pch->bridge field before taking the associated reference on the bridge
file instance.

This opens up a refcount underflow bug if ppp_bridge_channels called
via. iotcl runs concurrently with ppp_unbridge_channels executing via.
file release.

The bug is triggered by ppp_bridge_channels taking the error path
through the 'err_unset' label.  In this scenario, pch->bridge is set,
but the reference on the bridged channel will not be taken because
the function errors out.  If ppp_unbridge_channels observes pch->bridge
before it is unset by the error path, it will erroneously drop the
reference on the bridged channel and cause a refcount underflow.

To avoid this, ensure that ppp_bridge_channels holds a reference on
each channel in advance of setting the bridge pointers.
---
v2:
 * rework ppp_bridge_channels code to avoid the race condition in
   preference to holding ppp_mutex while calling ppp_unbridge_channels
---
 drivers/net/ppp/ppp_generic.c | 20 ++++++++++++++++----
 1 file changed, 16 insertions(+), 4 deletions(-)

On Tue, Jan 05, 2021 at 09:17:43PM +0000, Tom Parkin wrote:
>  err_unset:

>  	write_lock_bh(&pch->upl);

> -	RCU_INIT_POINTER(pch->bridge, NULL);

> +	/* Re-check pch->bridge with upl held since a racing unbridge might already

> +	 * have cleared it and dropped the reference on pch->bridge.

> +	 */

> +	if (rcu_dereference_protected(pch->bridge, lockdep_is_held(&pch->upl))) {

> +		RCU_INIT_POINTER(pch->bridge, NULL);

> +		drop_ref = true;

> +	}

>  	write_unlock_bh(&pch->upl);

>  	synchronize_rcu();

> +

> +	if (drop_ref)

> +		if (refcount_dec_and_test(&pchb->file.refcnt))

> +			ppp_destroy_channel(pchb);

> +


I think this works because ppp_mutex prevents pch->bridge from being
reassigned to another channel. However, this isn't obvious when reading
the code, and well, I prefer to not introduce new dependencies on
ppp_mutex (otherwise we'd better go with your original patch).

I think we could just save pch->bridge while we're under ->upl
protection, and then drop the reference of that channel (if non-NULL):

 err_unset:
 	write_lock_bh(&pch->upl);
+	/* Re-read pch->bridge in case it was modified concurrently */
+	pchb = rcu_dereference_protected(pch->bridge,
+					 lockdep_is_held(&pch->upl));
+	RCU_INIT_POINTER(pch->bridge, NULL);
 	write_unlock_bh(&pch->upl);
 	synchronize_rcu();
+
+	if (pchb)
+		if (refcount_dec_and_test(&pchb->file.refcnt))
+			ppp_destroy_channel(pchb);
+

 	return -EALREADY;
 }

This way we know that pchb is the channel we were pointing to when we
cleared pch->bridge. And this is also a bit simpler than using an extra
boolean.

On Tue,  5 Jan 2021 21:17:43 +0000 Tom Parkin wrote:
> When setting up a channel bridge, ppp_bridge_channels sets the

> pch->bridge field before taking the associated reference on the bridge

> file instance.

> 

> This opens up a refcount underflow bug if ppp_bridge_channels called

> via. iotcl runs concurrently with ppp_unbridge_channels executing via.

> file release.

> 

> The bug is triggered by ppp_bridge_channels taking the error path

> through the 'err_unset' label.  In this scenario, pch->bridge is set,

> but the reference on the bridged channel will not be taken because

> the function errors out.  If ppp_unbridge_channels observes pch->bridge

> before it is unset by the error path, it will erroneously drop the

> reference on the bridged channel and cause a refcount underflow.

> 

> To avoid this, ensure that ppp_bridge_channels holds a reference on

> each channel in advance of setting the bridge pointers.


Please prefix the subject with [PATCH net v3] for v3, add a fixes tag,
and make sure to include you sign-off.