fix dev refcount race
Subsystems (2)
Subject: veth fix dev refcount race
From: Daniel Lezcano <[email protected]>
When deleting the veth driver, veth_close calls netif_carrier_off for the two extremities of the network device. netif_carrier_off on the peer device will fire an event and hold a reference on the peer device. Just after, the peer is unregistered taking the rtnl_lock while the linkwatch_event is scheduled. If __linkwatch_run_queue does not occurs before the unregistering, unregister_netdevice will wait for the dev refcount to reach zero holding the rtnl_lock and linkwatch_event will wait for the rtnl_lock and hold the dev refcount.
--- drivers/net/veth.c | 53 ++++++++++++++++++++++++++++++++++++++++------------- 1 file changed, 40 insertions(+), 13 deletions(-)
drivers/net/veth.c +40 -13
@@ -244,18 +244,6 @@ static int veth_open(struct net_device *return 0 ; } -static int veth_close (struct net_device *dev) -{ -struct veth_priv *priv; - -if (netif_carrier_ok (dev)) { - priv =netdev_priv (dev); -netif_carrier_off (dev); -netif_carrier_off (priv->peer); - } -return 0 ; -} -static int veth_dev_init (struct net_device *dev) {struct veth_net_stats *stats;
@@ -286,13 +274,50 @@ static void veth_setup(struct net_device dev->hard_start_xmit = veth_xmit; dev->get_stats = veth_get_stats; dev->open = veth_open; - dev->stop = veth_close; dev->ethtool_ops = &veth_ethtool_ops; dev->features |=NETIF_F_LLTX ; dev->init = veth_dev_init; dev->destructor = veth_dev_free; } +static void veth_change_state (struct net_device *dev) +{ +struct net_device *peer; +struct veth_priv *priv; + + priv =netdev_priv (dev); + peer = priv->peer; + +if (netif_carrier_ok (peer)) { +if (!netif_carrier_ok (dev)) +netif_carrier_on (dev); + }else { +if (netif_carrier_ok (dev)) +netif_carrier_off (dev); + } +} + +static int veth_device_event (struct notifier_block *unused, +unsigned long event,void *ptr) +{ +struct net_device *dev = ptr; + +if (dev->open != veth_open) + goto out; + +switch (event) { +case NETDEV_CHANGE : +veth_change_state (dev); +break ; + } +out: +return NOTIFY_DONE ; +} + +static struct notifier_block veth_notifier_block__read_mostly = { + .notifier_call = veth_device_event, +}; +/* * netlink interface */
@@ -454,12 +479,14 @@ static struct rtnl_link_ops veth_link_opstatic __init int veth_init (void ) { +register_netdevice_notifier (&veth_notifier_block);return rtnl_link_register (&veth_link_ops); } static __exit void veth_exit(void) {rtnl_link_unregister (&veth_link_ops); +unregister_netdevice_notifier (&veth_notifier_block); }module_init (veth_init);
Quoted 10 lines · Daniel Lezcano
Subject: veth fix dev refcount race
From: Daniel Lezcano <[email protected]>When deleting the veth driver,
veth_closecallsnetif_carrier_offfor the two extremities of the network device.netif_carrier_offon the peer device will fire an event and hold a reference on the peer device. Just after, the peer is unregistered taking thertnl_lockwhile thelinkwatch_eventis scheduled. If__linkwatch_run_queuedoes not occurs before the unregistering,unregister_netdevicewill wait for the dev refcount to reach zero holding thertnl_lockandlinkwatch_event
Brr... AFAIS the unregister process waits for refcount without the rtnl_lock. The run-todo was invented for this. No?
Quoted 6 lines
will wait for the
rtnl_lockand hold the dev refcount.Signed-off-by Daniel Lezcano <[email protected]>--- drivers/net/veth.c | 53 ++++++++++++++++++++++++++++++++++++++++------------- 1 file changed, 40 insertions(+), 13 deletions(-)
David Miller
maintainerFrom: Daniel Lezcano <[email protected]>
Date: Tue, 19 Feb 2008 17:18:00 +0100
Quoted 11 lines
veth: fix dev refcount race
When deleting the veth driver,
veth_closecallsnetif_carrier_offfor the two extremities of the network device.netif_carrier_offon the peer device will fire an event and hold a reference on the peer device. Just after, the peer is unregistered taking thertnl_lockwhile thelinkwatch_eventis scheduled. If__linkwatch_run_queuedoes not occurs before the unregistering,unregister_netdevicewill wait for the dev refcount to reach zero holding thertnl_lockandlinkwatch_eventwill wait for thertnl_lockand hold the dev refcount.
Quoted 1 line
Signed-off-by Daniel Lezcano <[email protected]>
Thank you for fixing this bug, patch applied.