bonding: rebuild the lock use for bond_mii_monitor()

The bond_mii_monitor() still use bond lock to protect bond slave list, it is no effect, I have 2 way to fix the problem, move the RTNL to the top of the function, or add RCU to protect the bond_has_slaves() and bond_miimon_inspect(), according to the Jay Vosburgh's opinion, 10 times one second is a truely big performance loss if use RTNL to protect the whole function, so I would take the advice and use RCU to protect the two functions, of course it need to add more modify, the bond_has_slave_rcu() is add for RCU use, and the bond_for_each_slave need to replace with bond_for_each_slave_rcu in bond_miimon_inspect.

I move the peer notify before the queue_delayed_work(), and obviously it is no need to lock the RTNL twice if call bond_miimon_commit() and peer notify together, other path is no logic change, I think the performance is better than before.

Suggested-by Jay Vosburgh <[email protected]>

Suggested-by Veaceslav Falico <[email protected]>

Signed-off-by Ding Tianhong <[email protected]>

---
 drivers/net/bonding/bond_main.c | 40 ++++++++++++++++++++--------------------
 drivers/net/bonding/bonding.h   |  6 ++++++
 2 files changed, 26 insertions(+), 20 deletions(-)

drivers/net/bonding/bond_main.c +1 -1

index ba18719..def489d 100644
@@ -1913,7 +1913,7 @@ static int bond_miimon_inspect(struct bonding *bond)
 
 	ignore_updelay = !bond->curr_active_slave ? true : false;
 
-	bond_for_each_slave(bond, slave, iter) {
+	bond_for_each_slave_rcu(bond, slave, iter) {
 		slave->new_link = BOND_LINK_NOCHANGE;
 
 		link_state = bond_check_dev_link(bond, slave->dev, 0);

@@ -2111,47 +2111,47 @@ void bond_mii_monitor(struct work_struct *work)
 	bool should_notify_peers = false;
 	unsigned long delay;

drivers/net/bonding/bond_main.c +14 -16

 	if (bond_miimon_inspect(bond)) {
-		read_unlock(&bond->lock);
+		rcu_read_unlock();
 
 		/* Race avoidance with bond_close cancel of workqueue */
 		if (!rtnl_trylock()) {
-			read_lock(&bond->lock);
 			delay = 1;
-			should_notify_peers = false;
 			goto re_arm;
 		}
 
-		read_lock(&bond->lock);
-
 		bond_miimon_commit(bond);
 
-		read_unlock(&bond->lock);
+		if (should_notify_peers)
+			call_netdevice_notifiers(NETDEV_NOTIFY_PEERS,
+					bond->dev);
+
 		rtnl_unlock();	/* might sleep, hold no other locks */
-		read_lock(&bond->lock);
+	} else {
+		rcu_read_unlock();
+		if (should_notify_peers) {
+			if (!rtnl_trylock())
+				goto re_arm;
+			call_netdevice_notifiers(NETDEV_NOTIFY_PEERS,
+					bond->dev);
+			rtnl_unlock();
+		}
 	}
 
 re_arm:
 	if (bond->params.miimon)
 		queue_delayed_work(bond->wq, &bond->mii_work, delay);
-
-	read_unlock(&bond->lock);
-
-	if (should_notify_peers) {
-		if (!rtnl_trylock())
-			return;
-		call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, bond->dev);
-		rtnl_unlock();
-	}
 }
 
 static bool bond_has_this_ip(struct bonding *bond, __be32 ip)

drivers/net/bonding/bonding.h +0 -0

index 046a605..deb9738 100644

drivers/net/bonding/bonding.h +0 -0

 /* IMPORTANT: bond_first/last_slave can return NULL in case of an empty list */
 #define bond_first_slave(bond) \
 	(bond_has_slaves(bond) ? \

On 11/08/2013 03:07 AM, Ding Tianhong wrote:
In fact the bond cannot disappear while this function is running, so this test should be able to run outside the RCU region if I'm not missing something :-) It'll be just as useful as running inside the region, at most a free run may happen if there's one slave and it disappears.

This is unnecessary, bond_has_slaves() should be enough. See bond_start_xmit() and also the list_empty comment in include/linux/rculist.h for more information why. My bond_has_slaves() comments apply to all the patches that use it.

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html