Bug 653832

+++ This bug was initially created as a clone of Bug #653828 +++

Description of problem:


Version-Release number of selected component (if applicable):
guest OS:
kernel-2.6.18-194
host:
KVM host and version not related.

How reproducible:
100%

Steps to Reproduce:
1. create a virtual guest with two virtio network interfaces
2. configure a bonding interface with the option BONDING_OPTS="mode=active-backup arp_interval=2000 arp_ip_target=192.168.122.1" 

  
Actual results:
Bonding failovers on each arp inspect: 

Nov  4 15:56:09 localhost kernel: bonding: bond0: making interface eth0 the new active one.
Nov  4 15:56:11 localhost kernel: bonding: bond0: link status definitely up for interface eth1.
Nov  4 15:56:13 localhost kernel: bonding: bond0: link status definitely down for interface eth0, disabling it
Nov  4 15:56:13 localhost kernel: bonding: bond0: making interface eth1 the new active one.
Nov  4 15:56:15 localhost kernel: bonding: bond0: link status definitely up for interface eth0.
Nov  4 15:56:17 localhost kernel: bonding: bond0: link status definitely down for interface eth1, disabling it
Nov  4 15:56:17 localhost kernel: bonding: bond0: making interface eth0 the new active one.
Nov  4 15:56:19 localhost kernel: bonding: bond0: link status definitely up for interface eth1.
Nov  4 15:56:21 localhost kernel: bonding: bond0: link status definitely down for interface eth0, disabling it
Nov  4 15:56:21 localhost kernel: bonding: bond0: making interface eth1 the new active one.
Nov  4 15:56:23 localhost kernel: bonding: bond0: link status definitely up for interface eth0.
Nov  4 15:56:25 localhost kernel: bonding: bond0: link status definitely down for interface eth1, disabling it
Nov  4 15:56:25 localhost kernel: bonding: bond0: making interface eth0 the new active one.
Nov  4 15:56:27 localhost kernel: bonding: bond0: link status definitely up for interface eth1.
Nov  4 15:56:29 localhost kernel: bonding: bond0: link status definitely down for interface eth0, disabling it
Nov  4 15:56:29 localhost kernel: bonding: bond0: making interface eth1 the new active one.
Nov  4 15:56:31 localhost kernel: bonding: bond0: link status definitely up for interface eth0.
Nov  4 15:56:33 localhost kernel: bonding: bond0: link status definitely down for interface eth1, disabling it
Nov  4 15:56:33 localhost kernel: bonding: bond0: making interface eth0 the new active one.
Nov  4 15:56:35 localhost kernel: bonding: bond0: link status definitely up for interface eth1.


Expected results:
bonding works fine with virt-net driver.

Additional info:
 Because either ethtool or mii-tool is not available to detect the link for virtio interface, so arp monitor is used to detect link failure. The link detect doesn't make more sense for a virio interface.

This issue was caused the virtio driver doesn't update timestamp of rx and tx. 
The current slave can't pass the following check, because trans_start is not updated.
                /*
                 * Active slave is down if:
                 * - more than 2*delta since transmitting OR
                 * - (more than 2*delta since receive AND
                 *    the bond has an IP address)
                 */
                if ((slave->state == BOND_STATE_ACTIVE) &&
                    (time_after_eq(jiffies, slave->dev->trans_start +
                                    2 * delta_in_ticks) ||
                      (time_after_eq(jiffies, slave_last_rx(bond, slave)
                                     + 2 * delta_in_ticks)))) {
                        slave->new_link = BOND_LINK_DOWN;
                        commit++;
                }

Even though last_rx is ignored by virtio-net either, but it will be updated in the following codes:

static inline int skb_bond_should_drop(struct sk_buff *skb)
{
        struct net_device *dev = skb->dev;
        struct net_device *master = dev->master;

        if (master) {
                if (master->priv_flags & IFF_MASTER_ARPMON)
                        dev->last_rx = jiffies;
        ...

So the failure interface can become up in next arp inspect and then be chose on next failover, because it just check last_rx:
                if (slave->link != BOND_LINK_UP) {
                        if (time_before_eq(jiffies, slave_last_rx(bond, slave) +
                                           delta_in_ticks)) {
                                slave->new_link = BOND_LINK_UP;
                                commit++;
                        }

Then I update the rx and tx timestamp in virtio driver, and the patched virtio driver works fine with bonding.

--- Additional comment from dwu on 2010-11-16 04:28:22 EST ---

Created attachment 460793 [details]
update timestamp of rx and tx in virtio-net driver