[dpdk-users] active_backup link bonding and mac address

[dpdk-users] active_backup link bonding and mac address

[Kyle Larose]

Hey fellow DPDK users,

I have a question about the link bond pmd.

I am running  4 X710 interfaces in a link bond pmd for my application. In LACP mode, everything works fine. But, in active_backup mode, if the primary link fails, my application stops working. The reason is that I'm still sending packets with the original MAC address of the link bond pmd, which is that of the original primary slave. However, the new primary is not in promiscuous mode, so traffic coming back with that MAC address drops.

What should I be doing here:

1) Should I be listening for the changes in the state of the primary, and updating the MAC address I use to send? (I have it cached for efficiency)
2) Should the driver be placing the interface into promiscuous mode to allow for this, similar to what LACP does?
3) Should the driver be overwriting the MAC on egress, similar to what the tlb driver seems to do (in bond_ethdev_tx_burst_tlb)

I'm fine with #1, but it seems to break the goal of having the link bond pmd be transparent to the application.

Thanks,

Kyle

Re: [dpdk-users] active_backup link bonding and mac address

[Kyle Larose]

> -----Original Message-----
> From: Declan Doherty [mailto:declan.doherty@intel.com]
> Sent: Friday, May 12, 2017 11:47 AM
> To: Kyle Larose; users@dpdk.org; dev@dpdk.org
> Subject: Re: active_backup link bonding and mac address
> 
> 
> Yes it would be into the hw tables, rte_eth_dev_mac_addr_add() on each slave
> port should achieve this, so there will be no need to run in promiscuous
> mode. I'll try and setup a test for this on Monday morning in our lab.
> 
> Declan


Thanks for the suggestion. I modified mac_address_slaves_update to add the link
bond's mac address to all slaves in addition to setting their mac. Previously it
would only set the link bond's mac on the primary, and not add it to the hw.
This has solved my problem. On failure of the primary, my application starts
working via the backup seamlessly.

Now, I'm not sure if what I have done is ideal -- should we only be installing it
on the primary, and removing it from the old primary on failure? Either way, let
me know if you want my change as a patch, or whether you've come up with one of
your own. :)

Thanks,

Kyle

Re: [dpdk-users] active_backup link bonding and mac address

[Declan Doherty]

On 12/05/2017 4:34 PM, Kyle Larose wrote:
>
>
>> -----Original Message-----
>> From: Declan Doherty [mailto:declan.doherty@intel.com]
>> Sent: Friday, May 12, 2017 10:56 AM
>> To: Kyle Larose; users@dpdk.org; dev@dpdk.org
>> Subject: Re: active_backup link bonding and mac address
>>
>> On 12/05/2017 3:31 PM, Kyle Larose wrote:
>>> I'm adding the dev mailing list/link bonding maintainer, because I've done
>> some more investigation and I'm beginning to think something is wrong.
>>>
>>
>> Kyle, sorry I didn't see the post in the users list. I think the issue is
>> that the new primary is missing the bond MAC address on it's valid MACs
>> list, hence it is dropping the ingress packets after a fail-over event,
>> placing the all the slave devices into promiscuous mode as you suggest in
>> option 2 would probably make the issue go away but I don't think it's the
>> correct solution. I think we should just be adding the bond MAC to each
>> slaves devices valid MAC list. As only one bond slave is only active at any
>> time this won't cause any issues to the network, and will mean that fail
>> over is transparent to your application and there is no need for MAC
>> rewrites, which would invalidate existing ARP table entries at downstream
>> end points.
>
> Hey Declan,
>
> Thanks for the prompt response.
>
> I agree with your suggestion. Does this MAC list propagate to the slave NICs' hardware layers?
> That is, even if a slave isn't in promiscuous mode, if it receives a frame addressed to any
> MAC in its list, it will accept it and deliver it to the software? Or, does it mean we need to
> put the slave into promiscuous mode, but filter any MACs not in its list (unless the bond
> interface itself is in promiscuous mode)?
>
> Thanks,
>
> Kyle
>

Yes it would be into the hw tables, rte_eth_dev_mac_addr_add() on each 
slave port should achieve this, so there will be no need to run in 
promiscuous mode. I'll try and setup a test for this on Monday morning 
in our lab.

Declan

Re: [dpdk-users] active_backup link bonding and mac address

[Kyle Larose]

> -----Original Message-----
> From: Declan Doherty [mailto:declan.doherty@intel.com]
> Sent: Friday, May 12, 2017 10:56 AM
> To: Kyle Larose; users@dpdk.org; dev@dpdk.org
> Subject: Re: active_backup link bonding and mac address
> 
> On 12/05/2017 3:31 PM, Kyle Larose wrote:
> > I'm adding the dev mailing list/link bonding maintainer, because I've done
> some more investigation and I'm beginning to think something is wrong.
> >
> 
> Kyle, sorry I didn't see the post in the users list. I think the issue is
> that the new primary is missing the bond MAC address on it's valid MACs
> list, hence it is dropping the ingress packets after a fail-over event,
> placing the all the slave devices into promiscuous mode as you suggest in
> option 2 would probably make the issue go away but I don't think it's the
> correct solution. I think we should just be adding the bond MAC to each
> slaves devices valid MAC list. As only one bond slave is only active at any
> time this won't cause any issues to the network, and will mean that fail
> over is transparent to your application and there is no need for MAC
> rewrites, which would invalidate existing ARP table entries at downstream
> end points.

Hey Declan,

Thanks for the prompt response.

I agree with your suggestion. Does this MAC list propagate to the slave NICs' hardware layers?
That is, even if a slave isn't in promiscuous mode, if it receives a frame addressed to any
MAC in its list, it will accept it and deliver it to the software? Or, does it mean we need to
put the slave into promiscuous mode, but filter any MACs not in its list (unless the bond
interface itself is in promiscuous mode)?

Thanks,

Kyle

Re: [dpdk-users] active_backup link bonding and mac address

[Declan Doherty]

On 12/05/2017 3:31 PM, Kyle Larose wrote:
> I'm adding the dev mailing list/link bonding maintainer, because I've done some more investigation and I'm beginning to think something is wrong.
>
>> -----Original Message-----
>> From: Kyle Larose
>> Sent: Thursday, May 11, 2017 4:55 PM
>> To: users@dpdk.org
>> Subject: active_backup link bonding and mac address
>>
>> Hey fellow DPDK users,
>>
>> I have a question about the link bond pmd.
>>
>> I am running  4 X710 interfaces in a link bond pmd for my application. In
>> LACP mode, everything works fine. But, in active_backup mode, if the primary
>> link fails, my application stops working. The reason is that I'm still
>> sending packets with the original MAC address of the link bond pmd, which is
>> that of the original primary slave. However, the new primary is not in
>> promiscuous mode, so traffic coming back with that MAC address drops.
>>
>> What should I be doing here:
>>
>> 1) Should I be listening for the changes in the state of the primary, and
>> updating the MAC address I use to send? (I have it cached for efficiency)
>> 2) Should the driver be placing the interface into promiscuous mode to allow
>> for this, similar to what LACP does?
>> 3) Should the driver be overwriting the MAC on egress, similar to what the
>> tlb driver seems to do (in bond_ethdev_tx_burst_tlb)
>>
>> I'm fine with #1, but it seems to break the goal of having the link bond pmd
>> be transparent to the application.
>>
>
> I checked the mac address of the link bond interface after the failover, and it did not change.
> It still had the MAC address of the first slave that was added. This seems incompatible with
> solution number 1 that I suggested above, which means either it the link bond device should
> update its address, or it should be promiscuous at the slave level.
>
> FWIW, I'm using 16.07. I have reproduced this on testpmd by looking at port state. (with some
> fiddling -- needed to prevent it from starting the slave interfaces, and turn off its default
> promiscuous mode.)
>
> Does anyone have any input on this problem?
>
> Thanks,
>
> Kyle
>


Kyle, sorry I didn't see the post in the users list. I think the issue 
is that the new primary is missing the bond MAC address on it's valid 
MACs list, hence it is dropping the ingress packets after a fail-over 
event, placing the all the slave devices into promiscuous mode as you 
suggest in option 2 would probably make the issue go away but I don't 
think it's the correct solution. I think we should just be adding the 
bond MAC to each slaves devices valid MAC list. As only one bond slave 
is only active at any time this won't cause any issues to the network, 
and will mean that fail over is transparent to your application and 
there is no need for MAC rewrites, which would invalidate existing ARP 
table entries at downstream end points.

Re: [dpdk-users] active_backup link bonding and mac address

[Kyle Larose]

I'm adding the dev mailing list/link bonding maintainer, because I've done some more investigation and I'm beginning to think something is wrong.

> -----Original Message-----
> From: Kyle Larose
> Sent: Thursday, May 11, 2017 4:55 PM
> To: users@dpdk.org
> Subject: active_backup link bonding and mac address
> 
> Hey fellow DPDK users,
> 
> I have a question about the link bond pmd.
> 
> I am running  4 X710 interfaces in a link bond pmd for my application. In
> LACP mode, everything works fine. But, in active_backup mode, if the primary
> link fails, my application stops working. The reason is that I'm still
> sending packets with the original MAC address of the link bond pmd, which is
> that of the original primary slave. However, the new primary is not in
> promiscuous mode, so traffic coming back with that MAC address drops.
> 
> What should I be doing here:
> 
> 1) Should I be listening for the changes in the state of the primary, and
> updating the MAC address I use to send? (I have it cached for efficiency)
> 2) Should the driver be placing the interface into promiscuous mode to allow
> for this, similar to what LACP does?
> 3) Should the driver be overwriting the MAC on egress, similar to what the
> tlb driver seems to do (in bond_ethdev_tx_burst_tlb)
> 
> I'm fine with #1, but it seems to break the goal of having the link bond pmd
> be transparent to the application.
> 

I checked the mac address of the link bond interface after the failover, and it did not change.
It still had the MAC address of the first slave that was added. This seems incompatible with
solution number 1 that I suggested above, which means either it the link bond device should
update its address, or it should be promiscuous at the slave level.

FWIW, I'm using 16.07. I have reproduced this on testpmd by looking at port state. (with some
fiddling -- needed to prevent it from starting the slave interfaces, and turn off its default
promiscuous mode.)

Does anyone have any input on this problem?

Thanks,

Kyle