Bonding MAC failover behavior with VLAN interfaces

Bonding MAC failover behavior with VLAN interfaces

[John Valko]

Hi all,

I have observed the following on CentOS 6 using the 
2.6.32-573.8.1.el6.x86_64 kernel as well as Ubuntu 15.10 with 
4.2.0-25-generic.. I have not yet tried with a vanilla kernel but with 
the large time/distro gap it didn't seem likely to be caused by distro 
changes.

We have a networking scenario that looks a bit like this:


      +---------+ +---------+ +---------+
      |         | |         | |         |
      | bond0.5 | | bond0.6 | | bond0.7 |  VLAN interfaces
      |         | |         | |         |
      +---+-----+ +---+-----+ +---+-----+
          +------+    |   +-------+
                 |    |   |
               +-+----+---+--+
               |             | type=active-backup
               |    bond0    | mac failover=active
               |             |
               +-+---------+-+
                 |         |
                 |         |
       +---------+-+     +-+---------+
       |           |     |           |
       |   eth0    |     |   eth1    |bond slaves
       |           |     |           |
       +-----------+     +-----------+

Our actual scenario is where eth0 and eth1 are SR-IOV VFs passed to kvm 
guest via PCI passthru.  However, I have been able to demonstrate the 
same behavior using the veth driver so I'm going to use that to 
illustrate my confusion.

_1) Load bomding module with opt__ions mentioned in the above diagram:_

# modprobe bonding mode=1 miimon=100 fail_over_mac=active

_2) Verify we got the mode we __wanted_

#  cat /sys/class/net/bond0/bonding/mode
active-backup 1
# cat /sys/class/net/bond0/bonding/fail_over_mac
active 1

_3) Create some veth interfaces just so we have something to bond and 
then bond them_

# ip link add name veth0 type veth peer name veth0.peer
# ip link add name veth1 type veth peer name veth1.peer
# ifconfig veth0 up
# ifconfig veth0.peer up
# ifconfig veth1 up
# ifconfig veth1.peer up
# ifconfig bond0 up
# ifenslave bond0 veth0 veth1

Note, MAC is taken from veth0:

# ip link show veth0
5: veth0: <BROADCAST,MULTICAST,SLAVE,UP,LOWER_UP> mtu 1500 qdisc 
pfifo_fast master bond0 state UP qlen 1000
     link/ether *ca:d2:fb:b9:f9:b8* brd ff:ff:ff:ff:ff:ff
# ip link show veth1
7: veth1: <BROADCAST,MULTICAST,SLAVE,UP,LOWER_UP> mtu 1500 qdisc 
pfifo_fast master bond0 state UP qlen 1000
     link/ether 16:a3:36:0b:c1:ec brd ff:ff:ff:ff:ff:ff
# ip link show bond0
3: bond0: <BROADCAST,MULTICAST,MASTER,UP,LOWER_UP> mtu 1500 qdisc 
noqueue state UP
     link/ether *ca:d2:fb:b9:f9:b8* brd ff:ff:ff:ff:ff:ff

_4) Create a cou__ple VLAN interfaces_

# ip link add link bond0 name bond0.5 type vlan id 5
# ip link add link bond0 name bond0.6 type vlan id 6
# ip link add link bond0 name bond0.7 type vlan id 7
# ifconfig bond0.5 up
# ifconfig bond0.6 up
# ifconfig bond0.7 up

Note that these all have the sam MACs as bond0:

# ip link show bond0.5
8: bond0.5@bond0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc 
noqueue state UP
     link/ether ca:d2:fb:b9:f9:b8 brd ff:ff:ff:ff:ff:ff
# ip link show bond0.6
9: bond0.6@bond0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc 
noqueue state UP
     link/ether ca:d2:fb:b9:f9:b8 brd ff:ff:ff:ff:ff:ff
# ip link show bond0.7
10: bond0.7@bond0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc 
noqueue state UP
     link/ether ca:d2:fb:b9:f9:b8 brd ff:ff:ff:ff:ff:ff
# ip link show bond0
3: bond0: <BROADCAST,MULTICAST,MASTER,UP,LOWER_UP> mtu 1500 qdisc 
noqueue state UP
     link/ether ca:d2:fb:b9:f9:b8 brd ff:ff:ff:ff:ff:ff

_5) __Take down veth0 to cause a failover to veth1_

# ifconfig veth0 down

Now note that bond0 takes the address of veth1 as expected:

# ip link show veth1
7: veth1: <BROADCAST,MULTICAST,SLAVE,UP,LOWER_UP> mtu 1500 qdisc 
pfifo_fast master bond0 state UP qlen 1000
     link/ether 16:a3:36:0b:c1:ec brd ff:ff:ff:ff:ff:ff
# ip link show bond0
3: bond0: <BROADCAST,MULTICAST,MASTER,UP,LOWER_UP> mtu 1500 qdisc 
noqueue state UP
     link/ether 16:a3:36:0b:c1:ec brd ff:ff:ff:ff:ff:ff

BUT... not the VLAN interfaces, they still have veth0's MAC:

# ip link show veth0
5: veth0: <BROADCAST,MULTICAST,SLAVE> mtu 1500 qdisc pfifo_fast master 
bond0 state DOWN qlen 1000
     link/ether *ca:d2:fb:b9:f9:b8* brd ff:ff:ff:ff:ff:ff
# ip link show veth1
7: veth1: <BROADCAST,MULTICAST,SLAVE,UP,LOWER_UP> mtu 1500 qdisc 
pfifo_fast master bond0 state UP qlen 1000
     link/ether *16:a3:36:0b:c1:ec* brd ff:ff:ff:ff:ff:ff
# ip link show bond0
3: bond0: <BROADCAST,MULTICAST,MASTER,UP,LOWER_UP> mtu 1500 qdisc 
noqueue state UP
     link/ether *16:a3:36:0b:c1:ec* brd ff:ff:ff:ff:ff:ff
# ip link show bond0.5
8: bond0.5@bond0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc 
noqueue state UP
     link/ether *ca:d2:fb:b9:f9:b8* brd ff:ff:ff:ff:ff:ff
# ip link show bond0.6
9: bond0.6@bond0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc 
noqueue state UP
     link/ether *ca:d2:fb:b9:f9:b8* brd ff:ff:ff:ff:ff:ff
# ip link show bond0.7
10: bond0.7@bond0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc 
noqueue state UP
     link/ether *ca:d2:fb:b9:f9:b8* brd ff:ff:ff:ff:ff:ff

So, herein lies my confusion.  I expect that the VLAN interfaces should 
also pick up the new MAC address, but they don't.  It seems like a bug 
to me, but I don't want to be presumptuous so if in fact this is 
expected behavior, how do you recommend we approach making it switch 
when the bond fails over?  Right now, the MAC must be manually set for 
each vlan interface.

Right now I am looking at the current bonding code on master in 
drivers/net/bonding/bond_main.c:

    645  /* bond_do_fail_over_mac
    646   *
    647   * Perform special MAC address swapping for fail_over_mac settings
    648   *
    649   * Called with RTNL
    650   */
    651  static void bond_do_fail_over_mac(struct bonding *bond,
    652                                    struct slave *new_active,
    653                                    struct slave *old_active)
    654  {
    655          u8 tmp_mac[ETH_ALEN];
    656          struct sockaddr saddr;
    657          int rv;
    658
    659          switch (bond->params.fail_over_mac) {
    660          case BOND_FOM_ACTIVE:
    661                  if (new_active)
    662 bond_set_dev_addr(bond->dev, new_active->dev);
    663                  break;

I can see it set the mac of the bond to the new active slave MAC, but I 
don't see any indication of looping over vlan interfaces or anything, if 
that is expected... or would it be expected that the 8021q module 
receives an event that should make it change?

I am a kernel newbie, so I am not sure how this is really expected to 
work, but am very interested in your suggestions.

Thanks,

John

Re: Bonding MAC failover behavior with VLAN interfaces

[Jay Vosburgh]

John Valko <jovalko@cisco.com> wrote:

>Hi all,
>
>I have observed the following on CentOS 6 using the
>2.6.32-573.8.1.el6.x86_64 kernel as well as Ubuntu 15.10 with
>4.2.0-25-generic.. I have not yet tried with a vanilla kernel but with the
>large time/distro gap it didn't seem likely to be caused by distro
>changes.

[ ...example of VLAN-over-bond's MAC not changing when bonding
fail_over_mac=active has a failover... ]

>So, herein lies my confusion.  I expect that the VLAN interfaces should
>also pick up the new MAC address, but they don't.  It seems like a bug to
>me, but I don't want to be presumptuous so if in fact this is expected
>behavior, how do you recommend we approach making it switch when the bond
>fails over?  Right now, the MAC must be manually set for each vlan
>interface.
[...]
>I can see it set the mac of the bond to the new active slave MAC, but I
>don't see any indication of looping over vlan interfaces or anything, if
>that is expected... or would it be expected that the 8021q module receives
>an event that should make it change?

	Your observation is correct: the fail_over_mac functionality
does not propagate beyond the bonding master device itself.  The
presumption at the VLAN level is that the underlying device can support
multiple MAC addresses; this same effect (different MACs between VLAN
and its lower device) occurs if the device logically below a VLAN
interface changes its MAC.  This is handled by vlan_sync_address, which
is called via the NETDEV_CHANGEADDR notifier (which happens when the
bonding master changes its MAC due to f_o_m=active).

	The original need for f_o_m=active was IPoIB devices that cannot
change their MAC address.  Those don't support VLANs, either, so
propagation to VLANs was not a consideration.

	In your "real" SR-IOV sitation, is there something that
necessitates the use of fail_over_mac=active (I don't recall that SR-IOV
itself prohibits a VF from changing its MAC address), or is this being
done for other reasons?

	-J

---
	-Jay Vosburgh, jay.vosburgh@canonical.com

Re: Bonding MAC failover behavior with VLAN interfaces

[John Valko]

On 02/12/2016 03:58 PM, Jay Vosburgh wrote:
> John Valko <jovalko@cisco.com> wrote:
>
>> Hi all,
>>
>> I have observed the following on CentOS 6 using the
>> 2.6.32-573.8.1.el6.x86_64 kernel as well as Ubuntu 15.10 with
>> 4.2.0-25-generic.. I have not yet tried with a vanilla kernel but with the
>> large time/distro gap it didn't seem likely to be caused by distro
>> changes.
> [ ...example of VLAN-over-bond's MAC not changing when bonding
> fail_over_mac=active has a failover... ]
>
>> So, herein lies my confusion.  I expect that the VLAN interfaces should
>> also pick up the new MAC address, but they don't.  It seems like a bug to
>> me, but I don't want to be presumptuous so if in fact this is expected
>> behavior, how do you recommend we approach making it switch when the bond
>> fails over?  Right now, the MAC must be manually set for each vlan
>> interface.
> [...]
>> I can see it set the mac of the bond to the new active slave MAC, but I
>> don't see any indication of looping over vlan interfaces or anything, if
>> that is expected... or would it be expected that the 8021q module receives
>> an event that should make it change?
> 	Your observation is correct: the fail_over_mac functionality
> does not propagate beyond the bonding master device itself.  The
> presumption at the VLAN level is that the underlying device can support
> multiple MAC addresses; this same effect (different MACs between VLAN
> and its lower device) occurs if the device logically below a VLAN
> interface changes its MAC.  This is handled by vlan_sync_address, which
> is called via the NETDEV_CHANGEADDR notifier (which happens when the
> bonding master changes its MAC due to f_o_m=active).
>
> 	The original need for f_o_m=active was IPoIB devices that cannot
> change their MAC address.  Those don't support VLANs, either, so
> propagation to VLANs was not a consideration.
>
> 	In your "real" SR-IOV sitation, is there something that
> necessitates the use of fail_over_mac=active (I don't recall that SR-IOV
> itself prohibits a VF from changing its MAC address), or is this being
> done for other reasons?
>
> 	-J
>
> ---
> 	-Jay Vosburgh, jay.vosburgh@canonical.com
Hi Jay,

Thanks for your response.  I see in vlan.c the code you mentioned where 
the VLAN interface receives the NETDEV_CHANGEADDR.  And, after thinking 
on it for a while it would probably be undesirable (or at least not so 
useful) for a VLAN interface to follow the MAC of whatever interface 
it's attached to in general.

However, with SR-IOV in our case the reason we are using the active 
f_o_m policy is that the Intel NIC we are using implements anti-spoofing 
rules which will cause any frames with the old MAC to appear as spoofed 
and to be dropped.  While it is possible to disable this feature, we are 
deploying into a customer cloud where the customer has said they do not 
want it turned off.

Our current plan is to work around this by instead creating the VLAN 
interfaces directly on top of the VFs and then bonding each pair of 
those, so that the MAC of each VLAN interface never needs to change.  
The result feels perhaps less elegant but should work we think.

Since the original implementation of the active failover policy didn't 
consider VLAN interfaces, would it be reasonable to extend it so that 
when this policy is selected and there are VLAN interfaces on top of the 
bond that those change as well (e.g. by adding a check in 
vlan_sync_address())?  If that's a reasonable proposal I may be 
interested in taking a crack at a patch to submit.

Thanks,

--John