[Pacemaker] pacemaker/dlm problems

[Vladislav Bogdanov]

Hi Andrew, hi all,

I'm further investigating dlm lockspace hangs I described in
https://www.redhat.com/archives/cluster-devel/2011-August/msg00133.html
and in the thread starting from
https://lists.linux-foundation.org/pipermail/openais/2011-September/016701.html
.

What I described there is setup which involves pacemaker-1.1.6 with
corosync-1.4.1 and dlm_controld.pcmk from cluster-3.0.17 (without cman).
I use openais stack for pacemaker.

I found that it is possible to reproduce dlm kern_stop state across a
whole cluster with iptables on just one node, it is sufficient to block
all (or just corosync-specific) incoming/outgoing UDP for several
seconds (that time probably depends on corosync settings). I my case I
reproduced hang with 3-seconds traffic block:
iptables -I INPUT 1 -p udp -j REJECT; \
iptables -I OUTPUT 1 -p udp -j REJECT; \
sleep 3; \
iptables -D INPUT 1; \
iptables -D OUTPUT 1

I tried to make dlm_controld schedule fencing on CPG_REASON_NODEDOWN
event (just to look if it helps with problems I described in posts
referenced above), but without much success, following code does not work:

    int fd = pcmk_cluster_fd;
    int rc = crm_terminate_member_no_mainloop(nodeid, NULL, &fd);

I get "Could not kick node XXX from the cluster" message accompanied
with "No connection to the cluster". That means that
attrd_update_no_mainloop() fails.

Andrew, could you please give some pointers why may it fail? I'd then
try to fix dlm_controld. I do not see any other uses of that function
except than in dlm_controld.pcmk.

I agree with Jiaju
(https://lists.linux-foundation.org/pipermail/openais/2011-September/016713.html),
that could be solely pacemaker problem, because it probably should
originate fencing itself is such situation I think.

So, using pacemaker/dlm with openais stack is currently risky due to
possible hangs of dlm_lockspaces. Originally I got it due to heavy load
on one cluster nodes (actually on a host which has that cluster node
running as virtual guest).

Ok, I switched to cman to see if it helps. Fencing is configured in
pacemaker, not in cluster.conf.

Things became even worse ;( .

Although it took 25 seconds instead of 3 to break the cluster (I
understand, this is almost impossible to load host so much, but
anyways), then I got a real nightmare: two nodes of 3-node cluster had
cman stopped (and pacemaker too because of cman connection loss) - they
asked to kick_node_from_cluster() for each other, and that succeeded.
But fencing didn't happen (I still need to look why, but this is cman
specific).
Remaining node had pacemaker hanged, it doesn't even
notice cluster infrastructure change, down nodes were listed as a
online, one of them was a DC, all resources are marked as started on all
(down too) nodes. No log entries from pacemaker at all.

So, from my PoV cman+pacemaker is not currently suitable for HA tasks too.

That means that both possible alternatives are currently unusable if one
needs self-repairing pacemaker cluster with dlm support ;( That is
really regrettable.

I can provide all needed information and really hope that it is possible
to fix both issues:
* dlm blockage with openais and
* pacemaker lock with cman and no fencing from within dlm_controld

I think both issues are really high priority, because it is definitely
not acceptable when problems with load on one cluster node (or with link
to that node) lead to a total cluster lock or even crash.

I also offer any possible assistance from my side (f.e. patch trials
etc.) to get that all fixed. I can run either openais or cman and can
quickly switch between that stacks.

Sorry for not being brief,

Best regards,
Vladislav