[ClusterLabs] corosync 2.4 CPG config change callback

[Thomas Lamprecht]

Hi,

first some background info for my questions I'm going to ask:
We use corosync as a basis for our distributed realtime configuration
file system (pmxcfs)[1].

We got some reports of a completely hanging FS with the only
correlations being high load, often IO, and most times a message that
corosync did not got scheduled for longer than the token timeout.

See this example from a three node cluster, first:

> Mar 01 13:07:56 ceph05-01-public corosync[1638]: warning [MAIN  ] Corosync main process was not scheduled for 3767.3159 ms (threshold is 1320.0000 ms). Consider token timeout increase.

then we get a few logs that JOIN or LEAVE messages were thrown away
(understandable for this event):

Mar 01 13:07:56 ceph05-01-public corosync[1638]: warning [TOTEM ] JOIN or LEAVE message was thrown away during flush operation.
Mar 01 13:07:56 ceph05-01-public corosync[1638]:  [MAIN  ] Corosync main process was not scheduled for 3767.3159 ms (threshold is 1320.0000 ms). Consider token timeout increase.
Mar 01 13:07:56 ceph05-01-public corosync[1638]:  [TOTEM ] JOIN or LEAVE message was thrown away during flush operation.
Mar 01 13:07:56 ceph05-01-public corosync[1638]:  [TOTEM ] JOIN or LEAVE message was thrown away during flush operation.
Mar 01 13:07:56 ceph05-01-public corosync[1638]:  [TOTEM ] JOIN or LEAVE message was thrown away during flush operation.
Mar 01 13:07:56 ceph05-01-public corosync[1638]:  [TOTEM ] JOIN or LEAVE message was thrown away during flush operation.
Mar 01 13:07:56 ceph05-01-public corosync[1638]:  [TOTEM ] JOIN or LEAVE message was thrown away during flush operation.
Mar 01 13:07:56 ceph05-01-public corosync[1638]: notice  [TOTEM ] A new membership (192.168.21.51:2324) was formed. Members joined: 2 3 left: 2 3
Mar 01 13:07:56 ceph05-01-public corosync[1638]: notice  [TOTEM ] Failed to receive the leave message. failed: 2 3
Mar 01 13:07:56 ceph05-01-public corosync[1638]:  [TOTEM ] A new membership (192.168.21.51:2324) was formed. Members joined: 2 3 left: 2 3
Mar 01 13:07:56 ceph05-01-public corosync[1638]:  [TOTEM ] Failed to receive the leave message. failed: 2 3
Mar 01 13:07:56 ceph05-01-public corosync[1638]: notice  [QUORUM] Members[3]: 1 2 3
Mar 01 13:07:56 ceph05-01-public corosync[1638]: notice  [MAIN  ] Completed service synchronization, ready to provide service.
Mar 01 13:07:56 ceph05-01-public corosync[1638]:  [QUORUM] Members[3]: 1 2 3
Mar 01 13:07:56 ceph05-01-public corosync[1638]:  [MAIN  ] Completed service synchronization, ready to provide service.

Until recently we stepped really in the dark and had everything from
Kernel bugs to our filesystem logic as possible cause in mind...  But
then we had the luck to trigger this in our test systems and went to
town with gdb on the core dump, finding that we can trigger this by
pausing the leader (from our FS POV) for a short moment (may be shorter
than the token timeout), so that a new leader get elected, and then
resuming our leader node VM again.

The problem I saw was that while the leader had a log entry which
proved that he noticed his blackout:
> [TOTEM ] A new membership (192.168.21.51:2324) was formed. Members joined: 2 3 left: 2 3

our FS cpg_confchg_fn callback[2] was never called, thus it thought it
was still in sync and nothing ever happened, until another member
triggered this callback, by either leaving or (re-)joining.

Looking in the cpg.c code I saw that there's another callback, namely
cpg_totem_confchg_fn, which seemed a bit odd as wew did not set that
one... (I ain't the original author of the FS and it predates at least
to 2010, so maybe cpg_initialize was not yet deprecated there, and
thus model_initialize was not used then)

I switched over to using cpg_model_initialize and set the totem_confchg
callback, but for the "blacked out node" it gets called twice after the
event, but always shows all members...

So to finally get to my questions:

* why doesn't get the cpg_confchg_fn CB called when a node has a short
  blackout (i.e., corosync not being scheduled for a bit of time)?
  having all other nodes in it's leave and join list, as the log
  would suggests ("Members joined: 2 3 left: 2 3")

* If that doesn't seems like a good idea, what can we use to really
  detect such a node blackout?

As a work around I added logic for when through a config change a node
with a lower ID joined. The node which was leader until then triggers
a CPG leave enforcing a cluster wide config change event to happen,
which this time also the blacked out node gets and syncs then again
This works, but does not feels really nice, IMO...

help would be appreciated, much thanks!

cheers,
Thomas

[1]: https://git.proxmox.com/?p=pve-cluster.git;a=tree;f=data/src;h=e5493468b456ba9fe3f681f387b4cd5b86e7ca08;hb=HEAD
[2]: https://git.proxmox.com/?p=pve-cluster.git;a=blob;f=data/src/dfsm.c;h=cdf473e8226ab9706d693a457ae70c0809afa0fa;hb=HEAD#l1096