[PATCH] net/bonding: Do not transition down slave after speed/duplex check

[Thomas Falcon]

The following behavior has been observed when testing logical partition
migration of LACP-bonded VNIC devices in a PowerVM pseries environment.

1. When performing the migration, the bond master detects that a slave has
   lost its link, deactivates the LACP port, and sets the port's
   is_enabled flag to false.
2. The slave device then updates it's carrier state to off while it resets
   itself. This update triggers a NETDEV_CHANGE notification, which performs
   a speed and duplex update. The device does not return a valid speed
   and duplex, so the master sets the slave link state to BOND_LINK_FAIL.
3. When the slave VNIC device(s) are active again, some operations, such
   as setting the port's is_enabled flag, are not performed when transitioning
   the link state back to BOND_LINK_UP from BOND_LINK_FAIL, though the state
   prior to the speed check was BOND_LINK_DOWN.

Affected devices are therefore not utilized in the aggregation though they
are operational. The simplest way to fix this seems to be to restrict the
link state change to devices that are currently up and running.

CC: Jay Vosburgh <j.vosburgh at gmail.com>
CC: Veaceslav Falico <vfalico at gmail.com>
CC: Andy Gospodarek <andy at greyhouse.net>
Signed-off-by: Thomas Falcon <tlfalcon at linux.ibm.com>
---
 drivers/net/bonding/bond_main.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c
index 2e70e43c5df5..d840da7cd379 100644
--- a/drivers/net/bonding/bond_main.c
+++ b/drivers/net/bonding/bond_main.c
@@ -3175,7 +3175,8 @@ static int bond_slave_netdev_event(unsigned long event,
 		 * speeds/duplex are available.
 		 */
 		if (bond_update_speed_duplex(slave) &&
-		    BOND_MODE(bond) == BOND_MODE_8023AD) {
+		    BOND_MODE(bond) == BOND_MODE_8023AD &&
+		    slave->link == BOND_LINK_UP) {
 			if (slave->last_link_up)
 				slave->link = BOND_LINK_FAIL;
 			else
-- 
2.18.2