[PATCH net] team: fix deadlock during setting of MTU

Team driver protects port list traversal in team_change_mtu()
by its team->lock mutex. This causes a deadlock with certain
devices that calls netdev_update_features() during their
.ndo_change_mtu() callback. In this case netdev_update_features()
calls team's netdevice notifier team_device_event() that in case
of NETDEV_FEAT_CHANGE tries lock team->lock mutex again.

Example (r8169 case):
...
[ 6391.348202]  __mutex_lock.isra.6+0x2d0/0x4a0
[ 6391.358602]  team_device_event+0x9d/0x160 [team]
[ 6391.363756]  notifier_call_chain+0x47/0x70
[ 6391.368329]  netdev_update_features+0x56/0x60
[ 6391.373207]  rtl8169_change_mtu+0x14/0x50 [r8169]
[ 6391.378457]  dev_set_mtu_ext+0xe1/0x1d0
[ 6391.387022]  dev_set_mtu+0x52/0x90
[ 6391.390820]  team_change_mtu+0x64/0xf0 [team]
[ 6391.395683]  dev_set_mtu_ext+0xe1/0x1d0
[ 6391.399963]  do_setlink+0x231/0xf50
...

To fix the problem the port list traversal in team_change_mtu() can
be protected by RCU read lock. In case of failure the failing port
is marked and unwind code-path is done also under RCU read lock
protection (but not in reverse order).

Fixes: 3d249d4ca7d0 ("net: introduce ethernet teaming device")
Cc: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: Ivan Vecera <ivecera@redhat.com>
---
 drivers/net/team/team.c | 20 ++++++++++----------
 1 file changed, 10 insertions(+), 10 deletions(-)

diff --git a/drivers/net/team/team.c b/drivers/net/team/team.c
index c19dac21c468..69d4b28beb17 100644
--- a/drivers/net/team/team.c
+++ b/drivers/net/team/team.c

@@ -1802,35 +1802,35 @@ static int team_set_mac_address(struct net_device *dev, void *p)
 static int team_change_mtu(struct net_device *dev, int new_mtu)
 {
 	struct team *team = netdev_priv(dev);
-	struct team_port *port;
+	struct team_port *port, *fail_port;
 	int err;
 
-	/*
-	 * Alhough this is reader, it's guarded by team lock. It's not possible
-	 * to traverse list in reverse under rcu_read_lock
-	 */
-	mutex_lock(&team->lock);
+	rcu_read_lock();
 	team->port_mtu_change_allowed = true;
-	list_for_each_entry(port, &team->port_list, list) {
+	list_for_each_entry_rcu(port, &team->port_list, list) {
 		err = dev_set_mtu(port->dev, new_mtu);
 		if (err) {
 			netdev_err(dev, "Device %s failed to change mtu",
 				   port->dev->name);
+			fail_port = port;
 			goto unwind;
 		}
 	}
 	team->port_mtu_change_allowed = false;
-	mutex_unlock(&team->lock);
+	rcu_read_unlock();
 
 	dev->mtu = new_mtu;
 
 	return 0;
 
 unwind:
-	list_for_each_entry_continue_reverse(port, &team->port_list, list)
+	list_for_each_entry_rcu(port, &team->port_list, list) {
+		if (port == fail_port)
+			break;
 		dev_set_mtu(port->dev, dev->mtu);
+	}
 	team->port_mtu_change_allowed = false;
-	mutex_unlock(&team->lock);
+	rcu_read_unlock();
 
 	return err;
 }

-- 
2.26.2

On Thu, 2021-01-14 at 12:55 +0100, Ivan Vecera wrote:

some netdevs will need to sleep in their set_mtu, and you can't sleep
under rcu!

according to your explanation in the commit message the team->lock
mutex will be also taken under this rcu lock, so this is bad even
if dev_set_mtu does not sleep.

On Thu, 14 Jan 2021 16:34:24 -0800
Saeed Mahameed [off-list ref] wrote:

On Thu, 2021-01-14 at 12:55 +0100, Ivan Vecera wrote:

quoted

Team driver protects port list traversal in team_change_mtu()
by its team->lock mutex. This causes a deadlock with certain
devices that calls netdev_update_features() during their
.ndo_change_mtu() callback. In this case netdev_update_features()
calls team's netdevice notifier team_device_event() that in case
of NETDEV_FEAT_CHANGE tries lock team->lock mutex again.

Example (r8169 case):
...
[ 6391.348202]  __mutex_lock.isra.6+0x2d0/0x4a0
[ 6391.358602]  team_device_event+0x9d/0x160 [team]
[ 6391.363756]  notifier_call_chain+0x47/0x70
[ 6391.368329]  netdev_update_features+0x56/0x60
[ 6391.373207]  rtl8169_change_mtu+0x14/0x50 [r8169]
[ 6391.378457]  dev_set_mtu_ext+0xe1/0x1d0
[ 6391.387022]  dev_set_mtu+0x52/0x90
[ 6391.390820]  team_change_mtu+0x64/0xf0 [team]
[ 6391.395683]  dev_set_mtu_ext+0xe1/0x1d0
[ 6391.399963]  do_setlink+0x231/0xf50
...

To fix the problem the port list traversal in team_change_mtu() can
be protected by RCU read lock. In case of failure the failing port
is marked and unwind code-path is done also under RCU read lock
protection (but not in reverse order).

Fixes: 3d249d4ca7d0 ("net: introduce ethernet teaming device")
Cc: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: Ivan Vecera <ivecera@redhat.com>
---
 drivers/net/team/team.c | 20 ++++++++++----------
 1 file changed, 10 insertions(+), 10 deletions(-)

diff --git a/drivers/net/team/team.c b/drivers/net/team/team.c
index c19dac21c468..69d4b28beb17 100644
--- a/drivers/net/team/team.c
+++ b/drivers/net/team/team.c

@@ -1802,35 +1802,35 @@ static int team_set_mac_address(struct

net_device *dev, void *p)
 static int team_change_mtu(struct net_device *dev, int new_mtu)
 {
 	struct team *team = netdev_priv(dev);
-	struct team_port *port;
+	struct team_port *port, *fail_port;
 	int err;
 
-	/*
-	 * Alhough this is reader, it's guarded by team lock. It's not
possible
-	 * to traverse list in reverse under rcu_read_lock
-	 */
-	mutex_lock(&team->lock);
+	rcu_read_lock();
 	team->port_mtu_change_allowed = true;
-	list_for_each_entry(port, &team->port_list, list) {
+	list_for_each_entry_rcu(port, &team->port_list, list) {
 		err = dev_set_mtu(port->dev, new_mtu);  

some netdevs will need to sleep in their set_mtu, and you can't sleep
under rcu!

according to your explanation in the commit message the team->lock
mutex will be also taken under this rcu lock, so this is bad even
if dev_set_mtu does not sleep.

Hmm, you are right... btw do we need to take this mutex at this place?

team_change_mtu() is protected by RTNL, team_device_event() as a netdevice
notifier as well... and team_{add,del}_slave() that modify port list
also.

Thoughts?

Thanks,
Ivan

On Fri, 15 Jan 2021 09:24:05 +0100
Ivan Vecera [off-list ref] wrote:

quoted

according to your explanation in the commit message the team->lock
mutex will be also taken under this rcu lock, so this is bad even
if dev_set_mtu does not sleep.
  

Hmm, you are right... btw do we need to take this mutex at this place?

team_change_mtu() is protected by RTNL, team_device_event() as a netdevice
notifier as well... and team_{add,del}_slave() that modify port list
also.

Thoughts?

After private discussion with Jiri, this is no-go as he would not like to
introduce a dependency on RTNL into team driver. The other solution is
to postpone features calculation in certain cases... the patch follows.

Ivan