: Fix irq_sync race condition.

[TG3]: Fix irq_sync race condition.

Gagan Arneja <[email protected]> pointed out that tg3_reset_task() could potentially race with another thread calling tg3_full_lock() such as the ethtool_set_xxx() functions. This may trigger the BUG_ON() in tg3_irq_quiesce() or cause the irq_sync flag to be out- of-sync.

I think the easiest way to fix this is to get the tp->lock first before setting the irq_sync flag. This is safe to do because the tp->lock is never grabbed by the irq handler. This change will guarantee that the irq_sync flag updates will be serialized. We also have to change one spot to call tg3_netif_start() (which clears the irq_sync flag) before releasing the tp->lock.

index 3a43426..314f5cf 100644
@@ -3512,9 +3512,9 @@ static inline int tg3_irq_sync(struct tg3 *tp)
  */
 static inline void tg3_full_lock(struct tg3 *tp, int irq_sync)
 {
+	spin_lock_bh(&tp->lock);
 	if (irq_sync)
 		tg3_irq_quiesce(tp);
-	spin_lock_bh(&tp->lock);
 }
 
 static inline void tg3_full_unlock(struct tg3 *tp)

@@ -9116,10 +9116,10 @@ static void tg3_vlan_rx_register(struct net_device *dev, struct vlan_group *grp)
 	/* Update RX_MODE_KEEP_VLAN_TAG bit in RX_MODE register. */
 	__tg3_set_rx_mode(dev);
 
-	tg3_full_unlock(tp);
-
 	if (netif_running(dev))
 		tg3_netif_start(tp);
+
+	tg3_full_unlock(tp);
 }
 #endif
 

From: "Michael Chan" <[email protected]>
Date: Wed, 11 Jul 2007 16:31:45 -0700

Applied, thanks Michael.