[Sheepdog] [PATCH 2/2] sheep: don't skip notify event

[MORITA Kazutaka]

At Sat, 31 Mar 2012 16:55:41 +0800,
Liu Yuan wrote:
> 
> From: Liu Yuan <tailai.ly at taobao.com>
> 
> We need to handle notify event firstly because during shutdown operation,
> in a large set of nodes, some nodes that are shut-down will be mistaken
> as a leave events and broadcasts to other nodes where the notify of shutdown
> is delayed, thus cause those nodes to perform recovery process improperly.
> 
> Signed-off-by: Liu Yuan <tailai.ly at taobao.com>
> ---
>  sheep/group.c |    6 +++++-
>  1 files changed, 5 insertions(+), 1 deletions(-)
> 
> diff --git a/sheep/group.c b/sheep/group.c
> index a4ddc73..6aec105 100644
> --- a/sheep/group.c
> +++ b/sheep/group.c
> @@ -1084,8 +1084,12 @@ do_retry:
>  		struct request *req = container_of(cevent, struct request, cev);
>  		struct sd_obj_req *hdr = (struct sd_obj_req *)&req->rq;
>  
> +		/* we need to handle notify event firstly
> +		 * since it may has shutdown cluster command or some nodes may
> +		 * take nodes that being shutdowned as left ones
> +		 */
>  		if (cevent->ctype == CPG_EVENT_NOTIFY)
> -			continue;
> +			break;

This change means that all notify events (e.g. vdi creation) block
membership change events.  I think it is not good to block them so
long time.

However, as you say, we need to order the shutdown event with
membership change ones.  Currently, I don't have a better idea, so it
is okay to me to merge this patch, but, if possible, I'd like to see a
better fix for it.

Thanks,

Kazutaka

>  		if (is_membership_change_event(cevent->ctype))
>  			break;
>  
> -- 
> 1.7.8.2
> 
> -- 
> sheepdog mailing list
> sheepdog at lists.wpkg.org
> http://lists.wpkg.org/mailman/listinfo/sheepdog