[sheepdog] [PATCH v3] sheep: remove master node

Liu Yuan namei.unix at gmail.com
Tue Jul 23 06:44:15 CEST 2013


On Fri, Jul 19, 2013 at 01:01:01PM +0900, MORITA Kazutaka wrote:
> From: MORITA Kazutaka <morita.kazutaka at lab.ntt.co.jp>
> 
> The current procedure to handle sheep join is as follows.
> 
>  1. The joining node sends a join request.
>  2. The master node accepts the request.
>  3. All the nodes update cluster members.
> 
> This procedure has some problems:
> 
>  - The master election is too complex to maintain.
>    It is very difficult to make sure that the implementation is
>    correct.
> 
>  - The master node can fail while it is accepting the joining node.
>    The newly elected master has to take over the process, but it's
>    usually difficult to implement because we have to know what the
>    previous master did and what it did not before its failure.
> 
> This patch changes the sheep join procedure to the following.
> 
>  1. The joining node sends a join request.
>  2. Some of the existing nodes accept the request.
>  3. All the nodes update cluster members.
> 
> It is allowed for the multiple nodes to call sd_join_handler() against
> the same join request, but at least one node must have to do it.  With
> this change, we can eliminate a master, and node failure while
> accepting node join is also allowed.
> 
> Removing a master from zookeeper is not easy since it doesn't expect
> that multiple nodes send EVENT_ACCEPT.  I'll leave this for another
> day.
> 
> Signed-off-by: MORITA Kazutaka <morita.kazutaka at lab.ntt.co.jp>
> ---
> 
> v3:
>  - remove the zk driver
> 
> v2:
>  - rebase onto the master branch
> 
> 
>  include/shepherd.h       |    3 --
>  sheep/cluster.h          |    2 +-
>  sheep/cluster/corosync.c |   60 +++---------------------
>  sheep/cluster/local.c    |   38 +++++++--------
>  sheep/cluster/shepherd.c |   43 +++++------------
>  sheep/group.c            |   12 ++++-
>  shepherd/shepherd.c      |  115 +++-------------------------------------------
>  7 files changed, 54 insertions(+), 219 deletions(-)
> 
> diff --git a/include/shepherd.h b/include/shepherd.h
> index 9a4cf81..95b4b8b 100644
> --- a/include/shepherd.h
> +++ b/include/shepherd.h
> @@ -23,7 +23,6 @@ enum sph_srv_msg_type {
>  	SPH_SRV_MSG_LEAVE_FORWARD,
>  
>  	SPH_SRV_MSG_REMOVE,
> -	SPH_SRV_MSG_MASTER_ELECTION,
>  };
>  
>  struct sph_msg {
> @@ -39,7 +38,6 @@ struct sph_msg {
>  
>  struct sph_msg_join {
>  	struct sd_node new_node;
> -	uint8_t master_elected;
>  
>  	struct sd_node nodes[SD_MAX_NODES];
>  	uint32_t nr_nodes;
> @@ -116,7 +114,6 @@ static inline const char *sph_srv_msg_to_str(enum sph_srv_msg_type msg)
>  		{ SPH_SRV_MSG_NOTIFY_FORWARD, "SPH_SRV_MSG_NOTIFY_FORWARD" },
>  		{ SPH_SRV_MSG_BLOCK_FORWARD, "SPH_SRV_MSG_BLOCK_FORWARD" },
>  		{ SPH_SRV_MSG_REMOVE, "SPH_SRV_MSG_REMOVE" },
> -		{ SPH_SRV_MSG_MASTER_ELECTION, "SPH_SRV_MSG_MASTER_ELECTION" },
>  	};
>  
>  	for (i = 0; i < ARRAY_SIZE(msgs); i++) {
> diff --git a/sheep/cluster.h b/sheep/cluster.h
> index 2ec0a14..354896e 100644
> --- a/sheep/cluster.h
> +++ b/sheep/cluster.h
> @@ -171,7 +171,7 @@ void sd_notify_handler(const struct sd_node *sender, void *msg, size_t msg_len);
>  bool sd_block_handler(const struct sd_node *sender);
>  int sd_reconnect_handler(void);
>  void sd_update_node_handler(struct sd_node *);
> -void sd_join_handler(const struct sd_node *joining,
> +bool sd_join_handler(const struct sd_node *joining,
>  		     const struct sd_node *nodes, size_t nr_nodes,
>  		     void *opaque);
>  void recalculate_vnodes(struct sd_node *nodes, int nr_nodes);
> diff --git a/sheep/cluster/corosync.c b/sheep/cluster/corosync.c
> index 577d0cb..ef403bf 100644
> --- a/sheep/cluster/corosync.c
> +++ b/sheep/cluster/corosync.c
> @@ -24,7 +24,6 @@
>  struct cpg_node {
>  	uint32_t nodeid;
>  	uint32_t pid;
> -	uint32_t gone;
>  	struct sd_node node;
>  };
>  
> @@ -238,28 +237,6 @@ find_event(enum corosync_event_type type, struct cpg_node *sender)
>  		return find_nonblock_event(type, sender);
>  }
>  
> -static int is_master(struct cpg_node *node)
> -{
> -	int i;
> -	struct cpg_node *n = node;
> -	if (!n)
> -		n = &this_node;
> -	if (nr_cpg_nodes == 0)
> -		/* this node should be the first cpg node */
> -		return 0;
> -
> -	for (i = 0; i < SD_MAX_NODES; i++) {
> -		if (!cpg_nodes[i].gone)
> -			goto eq_check;
> -	}
> -	return -1;
> -
> -eq_check:
> -	if (cpg_node_equal(&cpg_nodes[i], n))
> -		return i;
> -	return -1;
> -}
> -
>  static void build_node_list(const struct cpg_node *nodes, size_t nr_nodes,
>  			    struct sd_node *entries)
>  {
> @@ -282,9 +259,6 @@ static bool __corosync_dispatch_one(struct corosync_event *cevent)
>  
>  	switch (cevent->type) {
>  	case COROSYNC_EVENT_TYPE_JOIN:
> -		if (is_master(&this_node) < 0)
> -			return false;
> -
>  		if (!cevent->msg)
>  			/* we haven't receive JOIN yet */
>  			return false;
> @@ -294,13 +268,14 @@ static bool __corosync_dispatch_one(struct corosync_event *cevent)
>  			return false;
>  
>  		build_node_list(cpg_nodes, nr_cpg_nodes, entries);
> -		sd_join_handler(&cevent->sender.node, entries, nr_cpg_nodes,
> -				cevent->msg);
> -		send_message(COROSYNC_MSG_TYPE_ACCEPT, &cevent->sender,
> -			     cpg_nodes, nr_cpg_nodes, cevent->msg,
> -			     cevent->msg_len);
> +		if (sd_join_handler(&cevent->sender.node, entries,
> +				    nr_cpg_nodes, cevent->msg)) {
> +			send_message(COROSYNC_MSG_TYPE_ACCEPT, &cevent->sender,
> +				     cpg_nodes, nr_cpg_nodes, cevent->msg,
> +				     cevent->msg_len);
>  
> -		cevent->callbacked = true;
> +			cevent->callbacked = true;
> +		}
>  		return false;
>  	case COROSYNC_EVENT_TYPE_ACCEPT:
>  		add_cpg_node(cpg_nodes, nr_cpg_nodes, &cevent->sender);
> @@ -469,7 +444,6 @@ static void cdrv_cpg_deliver(cpg_handle_t handle,
>  {
>  	struct corosync_event *cevent;
>  	struct corosync_message *cmsg = msg;
> -	int master;
>  
>  	sd_dprintf("%d", cmsg->type);
>  
> @@ -521,16 +495,6 @@ static void cdrv_cpg_deliver(cpg_handle_t handle,
>  	case COROSYNC_MSG_TYPE_LEAVE:
>  		cevent = xzalloc(sizeof(*cevent));
>  		cevent->type = COROSYNC_EVENT_TYPE_LEAVE;
> -
> -		master = is_master(&cmsg->sender);
> -		if (master >= 0)
> -			/*
> -			 * Master is down before new nodes finish joining.
> -			 * We have to revoke its mastership to avoid cluster
> -			 * hanging
> -			 */
> -			cpg_nodes[master].gone = 1;
> -
>  		cevent->sender = cmsg->sender;
>  		cevent->msg_len = cmsg->msg_len;
>  		if (cmsg->msg_len) {
> @@ -614,7 +578,6 @@ static void cdrv_cpg_confchg(cpg_handle_t handle,
>  
>  	/* dispatch leave_handler */
>  	for (i = 0; i < left_list_entries; i++) {
> -		int master;
>  		cevent = find_event(COROSYNC_EVENT_TYPE_JOIN, left_sheep + i);
>  		if (cevent) {
>  			/* the node left before joining */
> @@ -633,15 +596,6 @@ static void cdrv_cpg_confchg(cpg_handle_t handle,
>  		}
>  
>  		cevent = xzalloc(sizeof(*cevent));
> -		master = is_master(&left_sheep[i]);
> -		if (master >= 0)
> -			/*
> -			 * Master is down before new nodes finish joining.
> -			 * We have to revoke its mastership to avoid cluster
> -			 * hanging
> -			 */
> -			cpg_nodes[master].gone = 1;
> -
>  		cevent->type = COROSYNC_EVENT_TYPE_LEAVE;
>  		cevent->sender = left_sheep[i];
>  
> diff --git a/sheep/cluster/local.c b/sheep/cluster/local.c
> index 9d109ca..9135ab7 100644
> --- a/sheep/cluster/local.c
> +++ b/sheep/cluster/local.c
> @@ -408,37 +408,33 @@ static bool local_process_event(void)
>  	if (ev->callbacked)
>  		return false; /* wait for unblock event */
>  
> -	if (ev->type == EVENT_ACCEPT && lnode_eq(&this_node, &ev->sender)) {
> -		sd_dprintf("join Sheepdog");
> -		joined = true;
> -	}
> -
>  	if (!joined) {
> -		if (ev->type == EVENT_JOIN &&
> -		    lnode_eq(&this_node, &ev->sender)) {
> -			struct local_node lnodes[SD_MAX_NODES];
> -
> -			get_nodes(lnodes);
> +		if (!lnode_eq(&this_node, &ev->sender))
> +			goto out;
>  
> -			if (!lnode_eq(&this_node, &lnodes[0])) {
> -				sd_dprintf("wait for another node"
> -					   " to accept this node");
> -				return false;
> -			}
> -		} else
> +		switch (ev->type) {
> +		case EVENT_JOIN:
> +			break;
> +		case EVENT_ACCEPT:
> +			sd_dprintf("join Sheepdog");
> +			joined = true;
> +			break;
> +		default:
>  			goto out;
> +		}
>  	}
>  
>  	switch (ev->type) {
>  	case EVENT_JOIN:
>  		/* nodes[nr_nodes - 1] is a sender, so don't include it */
>  		assert(node_eq(&ev->sender.node, &nodes[nr_nodes - 1]));
> -		sd_join_handler(&ev->sender.node, nodes, nr_nodes - 1,
> -				  ev->buf);
> -		ev->type = EVENT_ACCEPT;
> -		msync(ev, sizeof(*ev), MS_SYNC);
> +		if (sd_join_handler(&ev->sender.node, nodes, nr_nodes - 1,
> +				      ev->buf)) {
> +			ev->type = EVENT_ACCEPT;
> +			msync(ev, sizeof(*ev), MS_SYNC);
>  
> -		shm_queue_notify();
> +			shm_queue_notify();
> +		}
>  
>  		return false;
>  	case EVENT_ACCEPT:
> diff --git a/sheep/cluster/shepherd.c b/sheep/cluster/shepherd.c
> index 49fcf8d..b2ab92d 100644
> --- a/sheep/cluster/shepherd.c
> +++ b/sheep/cluster/shepherd.c
> @@ -30,7 +30,6 @@
>  static int sph_comm_fd;
>  
>  static struct sd_node this_node;
> -static bool is_master;
>  
>  static int nr_nodes;
>  static struct sd_node nodes[SD_MAX_NODES];
> @@ -114,11 +113,8 @@ retry:
>  		 * FIXME: member change events must be ordered with nonblocked
>  		 *        events
>  		 */
> -		sd_join_handler(&join->new_node, NULL, 0, join->opaque);
> -
> -		/* FIXME: join->master_elected is needed? */
> -		assert(join->master_elected);
> -		is_master = true;
> +		if (!sd_join_handler(&join->new_node, NULL, 0, join->opaque))
> +			panic("sd_accept_handler() failed");
>  
>  		snd.type = SPH_CLI_MSG_ACCEPT;
>  		snd.body_len = join_len;
> @@ -297,26 +293,12 @@ static void sph_event_handler(int fd, int events, void *data)
>  		;
>  }
>  
> -static void elected_as_master(void)
> -{
> -	sd_dprintf("elected_as_master() called");
> -
> -	is_master = true;
> -	sd_iprintf("became new master");
> -}
> -
>  static void msg_new_node(struct sph_msg *rcv)
>  {
>  	int ret;
>  	struct sph_msg_join *join;
>  	struct sph_msg snd;
>  
> -	if (!is_master) {
> -		sd_printf(SDOG_EMERG, "I am not a master but received"
> -			" SPH_MSG_NEW_NODE, shepherd is buggy");
> -		exit(1);
> -	}
> -
>  	join = xzalloc(rcv->body_len);
>  	ret = xread(sph_comm_fd, join, rcv->body_len);
>  	if (ret != rcv->body_len) {
> @@ -325,7 +307,13 @@ static void msg_new_node(struct sph_msg *rcv)
>  	}
>  
>  	/* FIXME: member change events must be ordered with nonblocked events */
> -	sd_join_handler(&join->new_node, nodes, nr_nodes, join->opaque);
> +	if (!sd_join_handler(&join->new_node, join->nodes, join->nr_nodes,
> +			     join->opaque))
> +		/*
> +		 * This should succeed always because shepherd should have sent
> +		 * SPH_SRV_MSG_NEW_NODE only to the already joined node.
> +		 */
> +		panic("sd_join_handler() failed");
>  
>  	memset(&snd, 0, sizeof(snd));
>  	snd.type = SPH_CLI_MSG_ACCEPT;
> @@ -342,7 +330,6 @@ static void msg_new_node(struct sph_msg *rcv)
>  static void msg_new_node_finish(struct sph_msg *rcv)
>  {
>  	int ret;
> -	struct join_message *jm;
>  	struct sph_msg_join_node_finish *join_node_finish;
>  
>  	join_node_finish = xzalloc(rcv->body_len);
> @@ -352,7 +339,6 @@ static void msg_new_node_finish(struct sph_msg *rcv)
>  		exit(1);
>  	}
>  
> -	jm = (struct join_message *)join_node_finish->opaque;
>  	memcpy(nodes, join_node_finish->nodes,
>  	       join_node_finish->nr_nodes * sizeof(struct sd_node));
>  	nr_nodes = join_node_finish->nr_nodes;
> @@ -361,7 +347,8 @@ static void msg_new_node_finish(struct sph_msg *rcv)
>  		   node_to_str(&join_node_finish->new_node));
>  
>  	/* FIXME: member change events must be ordered with nonblocked events */
> -	sd_accept_handler(&join_node_finish->new_node, nodes, nr_nodes, jm);
> +	sd_accept_handler(&join_node_finish->new_node, nodes, nr_nodes,
> +			  join_node_finish->opaque);
>  
>  	free(join_node_finish);
>  }
> @@ -441,11 +428,6 @@ static void msg_leave_forward(struct sph_msg *rcv)
>  	do_leave_sheep();
>  }
>  
> -static void msg_master_election(struct sph_msg *rcv)
> -{
> -	elected_as_master();
> -}
> -
>  static void (*msg_handlers[])(struct sph_msg *) = {
>  	[SPH_SRV_MSG_NEW_NODE] = msg_new_node,
>  	[SPH_SRV_MSG_NEW_NODE_FINISH] = msg_new_node_finish,
> @@ -453,7 +435,6 @@ static void (*msg_handlers[])(struct sph_msg *) = {
>  	[SPH_SRV_MSG_BLOCK_FORWARD] = msg_block_forward,
>  	[SPH_SRV_MSG_REMOVE] = msg_remove,
>  	[SPH_SRV_MSG_LEAVE_FORWARD] = msg_leave_forward,
> -	[SPH_SRV_MSG_MASTER_ELECTION] = msg_master_election,
>  };
>  
>  static void interpret_msg(struct sph_msg *rcv)
> @@ -593,8 +574,6 @@ static int shepherd_leave(void)
>  		exit(1);
>  	}
>  
> -	is_master = false;
> -
>  	sd_dprintf("shepherd_leave() is completed");
>  
>  	return 0;
> diff --git a/sheep/group.c b/sheep/group.c
> index 4a2a83b..054292d 100644
> --- a/sheep/group.c
> +++ b/sheep/group.c
> @@ -779,14 +779,22 @@ void sd_notify_handler(const struct sd_node *sender, void *data,
>   * Accept the joining node and pass the cluster info to it.
>   *
>   * Note that 'nodes' doesn't contain 'joining'.
> + *
> + * Return true if the joining node is accepted.  At least one nodes in the
> + * cluster must call this function and succeed in accept of the joining node.
>   */
> -void sd_join_handler(const struct sd_node *joining,
> +bool sd_join_handler(const struct sd_node *joining,
>  		     const struct sd_node *nodes, size_t nr_nodes,
>  		     void *opaque)
>  {
>  	struct join_message *jm = opaque;
>  	char str[MAX_NODE_STR_LEN];
>  
> +	if (nr_nodes > 0 && node_is_local(joining)) {
> +		sd_dprintf("wait for another node to accept this node");
> +		return false;
> +	}
> +

Is there any case nr_nodes = 0? If not, we don't need this check

Thanks
Yuan



More information about the sheepdog mailing list