[Cluster-devel] [PATCH] cman: Don't increment on LEAVING->MEMBER transition

[Fabio M. Di Nitto <fdinitto@redhat.com>]

The patch looks sane to me and I am fairly sure it needs to be applied
to STABLE31/RHEL6 too.

I?d like also an ACK from Chrissie here.

Fabio

On 3/22/2011 2:06 AM, Lon Hohberger wrote:
> If a node left the cluster prior using 'cman_tool leave', its
> internal state is set to NODESTATE_LEAVING.
> 
> Ordinarily, during the transition from NODESTATE_LEAVING to
> NODESTATE_DEAD, the member count is decremented.  However, it
> is possible to rejoin the cluster membership from the
> NODESTATE_LEAVING state.
> 
> If this occurs, the cluster member count is incremented, but
> since the node never transitioned from LEAVING->DEAD, it was
> not previously decremented, causing an ever-increasing node
> count in 'cman_tool status'.
> 
> This patch resolves the issue by not incrementing the member
> count during a LEAVING->MEMBER transition.
> 
> Resolves: rhbz#649533
> 
> Signed-off-by: Lon Hohberger <lhh@redhat.com>
> ---
>  cman/daemon/commands.c |    5 ++++-
>  1 files changed, 4 insertions(+), 1 deletions(-)
> 
> diff --git a/cman/daemon/commands.c b/cman/daemon/commands.c
> index 6ffe9e5..6c91d7a 100644
> --- a/cman/daemon/commands.c
> +++ b/cman/daemon/commands.c
> @@ -1994,9 +1994,12 @@ void add_ais_node(int nodeid, uint64_t incarnation, int total_members)
>  
>  	if (node->state == NODESTATE_DEAD || node->state == NODESTATE_LEAVING) {
>  		gettimeofday(&node->join_time, NULL);
> +		/* If a node rejoins before it completes a leave,
> +		 * we should not increment cluster_members */
> +		if (node->state != NODESTATE_LEAVING)
> +			cluster_members++;
>  		node->incarnation = incarnation;
>  		node->state = NODESTATE_MEMBER;
> -		cluster_members++;
>  		recalculate_quorum(0, 0);
>  	}
>  }