public inbox for cluster-cvs@sourceware.org
help / color / mirror / Atom feed
* cluster: STABLE3 - cman: Fix a situation where cman could kill the wrong nodes
@ 2009-07-29 10:33 Christine Caulfield
  0 siblings, 0 replies; only message in thread
From: Christine Caulfield @ 2009-07-29 10:33 UTC (permalink / raw)
  To: cluster-cvs-relay

Gitweb:        http://git.fedorahosted.org/git/cluster.git?p=cluster.git;a=commitdiff;h=aa2ea305eb1c7b706a8a3f81adb84f90ecd880d0
Commit:        aa2ea305eb1c7b706a8a3f81adb84f90ecd880d0
Parent:        1558f71870f78c2101d8ef0833c178d2f2d86f8d
Author:        Christine Caulfield <ccaulfie@redhat.com>
AuthorDate:    Wed Jul 29 11:17:47 2009 +0100
Committer:     Christine Caulfield <ccaulfie@redhat.com>
CommitterDate: Wed Jul 29 11:17:47 2009 +0100

cman: Fix a situation where cman could kill the wrong nodes

hmm, how to describe this .... Hmmmm. OK lets try:

There were a couple of places in the cman code where the transition message
assumed that the node in question (either this node or the sending node) was
joining the cluster, rather than just sending it's current post-transition
state. This was wrong. It's a common problem we have with openais/corosync
in that it always merges clusters rather than joining from scratch so we
need to detect that in some way.

The code in ais.c has a flag called 'first_trans' which it sets when it first
encounters another node in the cluster. We should use this more often as it's
really helpful. So this is what we now do. The comments in the existing code
make it clear that it assumed the node was joining and not just part of an
existing transition, but the first_trans flag was not checked, so it was
fairly obvious what was going on.

So, now we check the first_trans flag in all places where the code assumes
that the node is joining a new cluster.

See ?

Signed-off-by: Christine Caulfield <ccaulfie@redhat.com>
---
 cman/daemon/commands.c |   16 +++++++++-------
 1 files changed, 9 insertions(+), 7 deletions(-)

diff --git a/cman/daemon/commands.c b/cman/daemon/commands.c
index 23ea810..b7f7293 100644
--- a/cman/daemon/commands.c
+++ b/cman/daemon/commands.c
@@ -75,6 +75,7 @@ static struct corosync_api_v1 *corosync;
 static corosync_timer_handle_t ccsd_timer;
 static unsigned int wanted_config_version;
 static int config_error;
+static int local_first_trans;
 
 static corosync_timer_handle_t shutdown_timer;
 static struct connection *shutdown_con;
@@ -1719,6 +1720,7 @@ void send_transition_msg(int last_memb_count, int first_trans)
 	int len = sizeof(struct cl_transmsg);
 
 	we_are_a_cluster_member = 1;
+	local_first_trans = first_trans;
 
 	log_printf(LOGSYS_LEVEL_DEBUG, "memb: sending TRANSITION message. cluster_name = %s\n", cluster_name);
 	msg->cmd = CLUSTER_MSG_TRANSITION;
@@ -1889,9 +1891,9 @@ static void do_process_transition(int nodeid, char *data)
 		return;
 	}
 
-	/* If the remote node can see AISONLY nodes then we can't join as we don't
-	   know the full state */
-	if (msg->flags & NODE_FLAGS_SEESDISALLOWED && !have_disallowed()) {
+	/* If the remote node can see AISONLY nodes and we want to join,
+	   then we can't, as we don't know the full state */
+	if (local_first_trans && msg->flags & NODE_FLAGS_SEESDISALLOWED && !have_disallowed()) {
 		/* Must use syslog directly here or the message will never arrive */
 		syslog(LOG_CRIT, "CMAN: Joined a cluster with disallowed nodes. must die");
 		cman_finish();
@@ -1953,10 +1955,10 @@ static void do_process_transition(int nodeid, char *data)
 		add_ais_node(nodeid, incarnation, num_ais_nodes);
 	}
 
-	/* If the cluster already has some AISONLY nodes then we can't make
-	   sense of the membership. So the new node has to also be AISONLY
-	   until we are consistent again */
-	if (have_disallowed() && !node->us)
+	/* If the new node is joining and the existing cluster already has some AISONLY
+	   nodes then we can't make sense of the membership.
+	   So the new node has to also be AISONLY until we are consistent again */
+	if (msg->first_trans && !node->us && have_disallowed())
 		node->state = NODESTATE_AISONLY;
 
 	node->flags = msg->flags; /* This will clear the BEENDOWN flag of course */


^ permalink raw reply	[flat|nested] only message in thread

only message in thread, other threads:[~2009-07-29 10:33 UTC | newest]

Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2009-07-29 10:33 cluster: STABLE3 - cman: Fix a situation where cman could kill the wrong nodes Christine Caulfield

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).