[pve-devel] [PATCH docs v2] pvecm, network: add section on corosync over bonds

[Mira Limbeck]

On 7/25/25 13:50, Friedrich Weber wrote:
> On 25/07/2025 13:39, Friedrich Weber wrote:
>> [...]
>> +Corosync Over Bonds
>> +~~~~~~~~~~~~~~~~~~~
>> +
>> +Using a xref:sysadmin_network_bond[bond] as the only Corosync link can be
>> +problematic in certain failure scenarios. If one of the bonded interfaces fails
>> +and stops transmitting packets, but its link state stays up, some bond modes
>> +may cause a state of asymmetric connectivity where cluster nodes can only
>> +communicate with different subsets of other nodes. In case of asymmetric
>> +connectivity, Corosync may not be able to form a stable quorum in the cluster.
>> +If this state persists and HA is enabled, nodes may fence themselves, even if
>> +their respective bond is still fully functioning. In the worst case, the whole
>> +cluster may fence itself.
>> +
>> +For this reason, our recommendations are as follows.
>> +
>> +* We recommend a dedicated physical NIC for the primary Corosync link. Bonds
>> +  can be used as additional links for increased redundancy.
> 
> These recommendations are still not 100% clear: Are we fine with a setup
> with
> 
> - link 0: dedicated corosync link
> - link 1: corosync link over a bond with a problematic mode (such as
> balance-rr or LACP with bond-lacp-rate slow)
> 
> ?
> In my tests, as long as the dedicated link 0 is completely online, it
> doesn't matter if a bond runs into the failure scenario above (one of
> the bonded NICs stops transmitting packets), corosync will just continue
> using link 0. But as soon as link 0 goes down and the failure scenario
> happens, the whole-cluster fence may happen. So should our
> recommendation be the relatively strict "if you put corosync on a bond
> (even if it is only a redundant link), use only active-backup or
> LACP+bond-lacp-rate fast"?

I'd say yes, the recommendation should be either dedicated link
directly, or a bond as redundant link with active-backup or
LACP+lacp-rate fast only.