Basic voting

dbutils.c

@@ -2307,6 +2307,13 @@ request_vote(PGconn *conn, t_node_info *this_node, t_node_info *other_node, XLog
 
 	termPQExpBuffer(&query);
 
+	// check for NULL
+	if (PQgetisnull(res, 0, 0))
+	{
+		log_debug("XXX NULL returned by repmgr.request_vote()");
+		return 0;
+	}
+
 	lsn_diff = atoi(PQgetvalue(res, 0, 0));
 
 	log_debug("XXX lsn_diff %i", lsn_diff);
@@ -2337,16 +2344,23 @@ request_vote(PGconn *conn, t_node_info *this_node, t_node_info *other_node, XLog
 	}
 
 	/* still tiebreak - decide by node_id */
-	if (this_node->node_id < other_node->node_id)
+
-	{
+	// we're the candidate, so we win
-		log_debug("this node has lower id");
+	log_debug("win by default");
-		return 1;
+	return 1;
-	}
+
-	log_debug("other node wins");
+}
 
 
-	/* we lose */
+void
-	return 0;
+set_voting_status_initiated(PGconn *conn)
+{
+	PGresult		   *res;
+
+	res = PQexec(conn, "SELECT repmgr.set_voting_status_initiated()");
+
+	PQclear(res);
+	return;
 }
 
 

dbutils.h

@@ -247,6 +247,7 @@ bool		is_server_available(const char *conninfo);
 /* node voting functions */
 NodeVotingStatus get_voting_status(PGconn *conn);
 int	request_vote(PGconn *conn, t_node_info *this_node, t_node_info *other_node, XLogRecPtr last_wal_receive_lsn);
+void set_voting_status_initiated(PGconn *conn);
 
 
 /* replication status functions */

repmgr--4.0.sql

@@ -46,3 +46,8 @@ CREATE FUNCTION get_voting_status()
   RETURNS INT
   AS '$libdir/repmgr', 'get_voting_status'
   LANGUAGE C STRICT;
+
+CREATE FUNCTION set_voting_status_initiated()
+  RETURNS VOID
+  AS '$libdir/repmgr', 'set_voting_status_initiated'
+  LANGUAGE C STRICT;

repmgr.c

@@ -64,6 +64,8 @@ PG_FUNCTION_INFO_V1(request_vote);
 Datum		get_voting_status(PG_FUNCTION_ARGS);
 PG_FUNCTION_INFO_V1(get_voting_status);
 
+Datum		set_voting_status_initiated(PG_FUNCTION_ARGS);
+PG_FUNCTION_INFO_V1(set_voting_status_initiated);
 
 /*
  * Module load callback
@@ -155,6 +157,16 @@ request_vote(PG_FUNCTION_ARGS)
 
 	int lsn_diff;
 
+	NodeVotingStatus voting_status;
+
+	LWLockAcquire(shared_state->lock, LW_SHARED);
+	voting_status = shared_state->voting_status;
+	LWLockRelease(shared_state->lock);
+
+	/* this node has initiated voting or already responded to another node */
+	if (voting_status != VS_NO_VOTE)
+		PG_RETURN_NULL();
+
 	elog(INFO, "id is %i, lsn: %X/%X",
 		 requesting_node_id,
 		 (uint32) (requesting_node_last_lsn >> 32),
@@ -182,6 +194,11 @@ request_vote(PG_FUNCTION_ARGS)
 
    	SPI_finish();
 
+	/* indicate this node has responded to a vote request */
+	LWLockAcquire(shared_state->lock, LW_SHARED);
+	shared_state->voting_status = VS_VOTE_REQUEST_RECEIVED;
+	LWLockRelease(shared_state->lock);
+
 	PG_RETURN_INT32(lsn_diff);
 }
 
@@ -198,3 +215,13 @@ get_voting_status(PG_FUNCTION_ARGS)
 
 	PG_RETURN_INT32(voting_status);
 }
+
+Datum
+set_voting_status_initiated(PG_FUNCTION_ARGS)
+{
+	LWLockAcquire(shared_state->lock, LW_SHARED);
+	shared_state->voting_status = VS_VOTE_INITIATED;
+	LWLockRelease(shared_state->lock);
+
+	PG_RETURN_VOID();
+}

repmgrd.c

@@ -6,7 +6,9 @@
 
 #include <stdio.h>
 #include <signal.h>
+#include <stdlib.h>
 #include <sys/stat.h>
+#include <time.h>
 
 #include "repmgr.h"
 #include "config.h"
@@ -103,6 +105,8 @@ main(int argc, char **argv)
 
 	set_progname(argv[0]);
 
+    srand ( time(NULL) );
+
 	/* Disallow running as root */
 	if (geteuid() == 0)
 	{
@@ -530,7 +534,32 @@ monitor_streaming_standby(void)
 				/* still down after reconnect attempt(s) - */
 				if (upstream_node_status == NODE_STATUS_DOWN)
 				{
-					do_election();
+					NodeVotingStatus voting_status = do_election();
+
+					switch(voting_status)
+					{
+						case VS_NO_VOTE:
+							log_info("NO VOTE");
+							break;
+						case VS_VOTE_REQUEST_RECEIVED:
+							log_info("VOTE REQUEST RECEIVED");
+							break;
+						case VS_VOTE_INITIATED:
+							log_info("VOTE REQUEST INITIATED");
+							break;
+
+						case VS_VOTE_WON:
+							log_info("VOTE REQUEST WON");
+							break;
+						case VS_VOTE_LOST:
+							log_info("VOTE REQUEST LOST");
+							break;
+
+						case VS_UNKNOWN:
+							log_info("VOTE REQUEST UNKNOWN");
+							break;
+					}
+
 					// begin voting process
 
 					// if VS_PROMOTION_CANDIDATE
@@ -551,9 +580,8 @@ monitor_streaming_standby(void)
 static NodeVotingStatus
 do_election(void)
 {
-	int total_eligible_nodes = 0;
+//	int total_eligible_nodes = 0;
-	/* current node votes for itself by default */
+	int votes_for_me = 0;
-	int votes_for_me = 1;
 
 	/* we're visible */
 	int visible_nodes = 1;
@@ -568,15 +596,27 @@ do_election(void)
 	NodeInfoList standby_nodes = T_NODE_INFO_LIST_INITIALIZER;
 	NodeInfoListCell *cell;
 
+	long unsigned rand_wait = (long) ((rand() % 50) + 10) * 10000;
+
+	log_debug("do_election(): sleeping %li", rand_wait);
+
+	pg_usleep(rand_wait);
+
+
 	voting_status = get_voting_status(local_conn);
 	log_debug("do_election(): node voting status is %i", (int)voting_status);
 
 	if (voting_status == VS_VOTE_REQUEST_RECEIVED)
 	{
+		log_debug("vote request already received, not candidate");
 		/* we've already been requested to vote, so can't become a candidate */
 		return voting_status;
 	}
 
+	// XXX should we mark ourselves as candidate?
+	//  -> so any further vote requests are rejected?
+	set_voting_status_initiated(local_conn);
+
 	/* get all active nodes attached to primary, excluding self */
 	// XXX include barman node in results
 
@@ -612,15 +652,25 @@ do_election(void)
 
 	// XXX check if > 50% visible
 
-	/* check if we've been asked to vote again */
+	/* check again if we've been asked to vote */
-	// XXX do that
 
-	// XXX should we mark ourselves as candidate?
+if (0)
-	//  -> so any further vote requests are rejected?
+{
+	voting_status = get_voting_status(local_conn);
+	log_debug("do_election(): node voting status is %i", (int)voting_status);
 
 
+	if (voting_status == VS_VOTE_REQUEST_RECEIVED)
+	{
+		/* we've already been requested to vote, so can't become a candidate */
+		return voting_status;
+	}
+}
+
+	/* current node votes for itself by default */
+	votes_for_me += 1;
+
 	/* get our lsn*/
-
 	last_wal_receive_lsn = get_last_wal_receive_location(local_conn);
 
 	log_debug("LAST receive lsn = %X/%X", (uint32) (last_wal_receive_lsn >> 32), (uint32) last_wal_receive_lsn);
@@ -643,7 +693,10 @@ do_election(void)
 
 	log_notice(_("%i of of %i votes"), votes_for_me, visible_nodes);
 
-	return VS_VOTE_WON;
+	if (votes_for_me == visible_nodes)
+		return VS_VOTE_WON;
+
+	return VS_VOTE_LOST;
 }
 
 static void
@@ -846,7 +899,7 @@ try_reconnect(const char *conninfo, NodeStatus *node_status)
 	int i;
 
 	// XXX make this all configurable
-	int max_attempts = 15;
+	int max_attempts = 10;
 
 	for (i = 0; i < max_attempts; i++)
 	{

voting.h

@@ -10,7 +10,8 @@ typedef enum {
 	VS_NO_VOTE,
 	VS_VOTE_REQUEST_RECEIVED,
 	VS_VOTE_INITIATED,
-	VS_VOTE_WON
+	VS_VOTE_WON,
+	VS_VOTE_LOST
 } NodeVotingStatus;
 
 #endif /* _VOTING_H_ */