Bug 460190

Description of problem:

Certain network/switch settings cause nodes to form partitioned clusters
when they start up.  We want to provide information to help people configure
their switches to prevent this (see Documentation note).

We can also add code to better cope with these network problems, since
they seem to be somewhat common.  The network partitions are a particular
problem for two_node clusters where a node has quorum when it starts up
on its own.  There are two parts to this work-around:

1. Add new fence_tool option -m, e.g. fence_tool join -m 45.
This will cause fence_tool to wait for all nodes in cluster.conf
to be cluster members, or the timeout (45 seconds), whichever comes
first, before joining the fence domain.

The idea is that we'd use this option to allow openais on the nodes
to all see each other before starting the fence domain. So we join the
domain *after* the nodes merge into a single cluster.  If we joined the
domain *before* the cluster partition merged, then nodes end up being
fenced unnecessarily.  (This is a similar idea to post_join_delay; a delay
that gives us time to determine that a node in an unknown state is actually
ok and doesn't require fencing.)

2. Use the new fence_tool -m option in the cman init script.  Again, this
is primarily a problem with two_node clusters (because waiting for quorum
usually masks the partitioning problems otherwise).  So, we want the
init script to check if the cluster is two_node, and use -m if it is.
(it could do this by 'grep two_node /etc/cluster/cluster.conf', or
'cman_tool status | grep Flags | grep 2node').  It initially appears that
we'll want a default -m value of about 45 seconds.  Again, if the nodes
converge normally during startup, this delay will be skipped.


Version-Release number of selected component (if applicable):


How reproducible:


Steps to Reproduce:
1.
2.
3.
  
Actual results:


Expected results:


Additional info: