LLDP packet drops on bond interface

[kstange-enfn]

I've been having trouble finding much information on this, so I apologize if this is duplicated. I have a bunch of Ubuntu 20.04 or 22.04 servers running kernel 5.15 or 6.2. These servers have two intel NICs in bond, usually x710 hardware.

When lldpd is running, I see one dropped packet counted on the bond interface for each LLDP packet received by either physical interface. They are processed correctly by lldpd and not shown as dropped on the physical interfaces.

When lldpd is not running, drops are not counted on any interfaces. If I run tcpdump -i bond0 ether proto 0x88cc I can see the packets reaching the bond. I know they don't get counted when the interface is in promiscuous mode, but the drop counter just stops incrementing entirely as long as lldpd is stopped.

I tried using a tool called dropwatch and I see these only when lldpd is running:

drop at: __netif_receive_skb_core+0x150/0xf60 (0xffffffffacad1960)
origin: software
input port ifindex: 10
timestamp: Tue Mar 19 15:35:21 2024 706822244 nsec
protocol: 0x88cc
length: 322
original length: 322

These counters incrementing don't hurt anything, so I'm just trying to understand what's going on. If there's a way to prevent it and keep these counters at zero, that's a bonus.

[vincentbernat]

See 8b50be7.

[kstange-enfn]

I'm doubtful this is the solution as I have lldpd 1.0.13 on Ubuntu 22.04 and lldpd 10.0.4 on Ubuntu 20.04 and see this behavior in both places. I can try a newer lldpd if you think that will help.

[vincentbernat]

It's not a fix it's an explanation why this problem is there and not fixed.

[kstange-enfn]

I see, both the change and revert happened in between the lldpd versions we are using. Is there no prospect for a workaround here, other than kernel behavior changes? Just want to understand if there's any future for a fix. If not then I'll just mark as resolved by this explanation and move on.

Thanks for your responses!

[vincentbernat]

No fix planned. I don't want to break some people setup "just" to fix an error counter. At the minimum, someone should test again if the patch breaks current versions of OVS.