Not sure. But anyway ceph has been skipping interfaces named "lo"
since v10, but then dropped that in 14.2.18 (by accident, IMO).
You should be able to get your osds listening to the correct IP using
cluster network = 10.1.50.0/8
public network = 10.1.50.0/8
does that work?
On Tue, Mar 23, 2021 at 2:45 PM Sam Skipsey <aoanla(a)gmail.com> wrote:
> I should note that our cluster is entirely IPv4 [because it's on a private
network, so there's no need to go IPv6], which maybe influences matters?
> On Tue, 23 Mar 2021 at 11:43, Stefan Kooman <stefan(a)bit.nl> wrote:
>> On 3/23/21 8:29 AM, Dan van der Ster wrote:
>> > Hi Sam,
>> > Yeah somehow `lo:` is not getting skipped, probably due to those
>> > patches. (I guess it is because the 2nd patch looks for `lo:` but in
>> > fact the ifa_name is probably just `lo` without the colon)
>> > https://github.com/ceph/ceph/blob/master/src/common/ipaddr.cc#L110
>> > I don't know why this impacts you but not us -- we already upgraded
>> > one of our clusters to 14.2.18 on Centos 8, and ceph is choosing the
>> > correct interface without needing any network options. And lo: is the
>> > first interface  here too.
>> > Could it be as simple as the iface names being sorted alphabetically?
>> > Here we have ens785f0 which would come before lo, but your interface
>> > `p2p2` would come after.
>> I renamed a 14.2.18 osd server's network interface to "zzzceph"
>> ceph before" and restarted the node. It does come up with the correct ip
>> (ms_bind_ipv4=false, ms_bind_ipv6=true). So it does not seem to be as
>> simple as this, or the ms_bind option migth matter here as well, dunno.
>> Gr. Stefan
> Sam Skipsey (he/him, they/them)