take major node config into account

kelvich · kelvich · commit e6d6f5671102 · 2017-11-13T02:21:38.000+03:00
diff --git a/multimaster.c b/multimaster.c
@@ -2916,7 +2916,7 @@ _PG_init(void)
 		NULL,
 		&MtmMajorNode,
 		false,
-		PGC_BACKEND,
+		PGC_SUSET,
 		0,
 		NULL,
 		NULL,
diff --git a/state.c b/state.c
@@ -78,22 +78,28 @@ static void
 MtmCheckState(void)
 {
 	// int nVotingNodes = MtmGetNumberOfVotingNodes();
+	bool isEnabledState;
 	int nEnabled   = countZeroBits(Mtm->disabledNodeMask, Mtm->nAllNodes);
 	int nConnected = countZeroBits(SELF_CONNECTIVITY_MASK, Mtm->nAllNodes);
 	int nReceivers = Mtm->nAllNodes - countZeroBits(Mtm->pglogicalReceiverMask, Mtm->nAllNodes);
 	int nSenders   = Mtm->nAllNodes - countZeroBits(Mtm->pglogicalSenderMask, Mtm->nAllNodes);
 
-	MTM_LOG1("[STATE]   Status = (disabled=%s, unaccessible=%s, clique=%s, receivers=%s, senders=%s, total=%i)",
+	MTM_LOG1("[STATE]   Status = (disabled=%s, unaccessible=%s, clique=%s, receivers=%s, senders=%s, total=%i, major=%d)",
 		maskToString(Mtm->disabledNodeMask, Mtm->nAllNodes),
 		maskToString(SELF_CONNECTIVITY_MASK, Mtm->nAllNodes),
 		maskToString(Mtm->clique, Mtm->nAllNodes),
 		maskToString(Mtm->pglogicalReceiverMask, Mtm->nAllNodes),
 		maskToString(Mtm->pglogicalSenderMask, Mtm->nAllNodes),
-		Mtm->nAllNodes);
+		Mtm->nAllNodes,
+		MtmMajorNode);
+
+	isEnabledState =
+		( (nConnected >= Mtm->nAllNodes/2+1)						/* majority */
+			|| (nConnected == Mtm->nAllNodes/2 && MtmMajorNode) )	/* or half + major node */
+		&& BIT_CHECK(Mtm->clique, MtmNodeId-1);						/* in clique */
 
 	/* ANY -> MTM_DISABLED */
-	if ( nConnected < Mtm->nAllNodes/2+1 ||
-		!BIT_CHECK(Mtm->clique, MtmNodeId-1) )
+	if (!isEnabledState)
 	{
 		BIT_SET(Mtm->disabledNodeMask, MtmNodeId-1);
 		MtmSetClusterStatus(MTM_DISABLED);
@@ -103,8 +109,7 @@ MtmCheckState(void)
 	switch (Mtm->status)
 	{
 		case MTM_DISABLED:
-			if ( (nConnected >= Mtm->nAllNodes/2+1) && /* majority */
-				 BIT_CHECK(Mtm->clique, MtmNodeId-1) ) /* in clique */
+			if (isEnabledState)
 			{
 				MtmSetClusterStatus(MTM_RECOVERY);
 				return;
@@ -298,6 +303,7 @@ void MtmOnNodeDisconnect(int nodeId)
 	MtmLock(LW_EXCLUSIVE);
 	BIT_SET(SELF_CONNECTIVITY_MASK, nodeId-1);
 	BIT_SET(Mtm->reconnectMask, nodeId-1);
+	Mtm->nConfigChanges += 1;
 	MtmCheckState();
 	MtmUnlock();
 
diff --git a/tests2/docker-entrypoint.sh b/tests2/docker-entrypoint.sh
@@ -68,9 +68,9 @@ if [ "$1" = 'postgres' ]; then
 			multimaster.ignore_tables_without_pk = 1
 			multimaster.node_id = $NODE_ID
 			multimaster.conn_strings = '$CONNSTRS'
+			multimaster.major_node = $MAJOR
 			multimaster.heartbeat_recv_timeout = 1100
 			multimaster.heartbeat_send_timeout = 250
-			multimaster.min_2pc_timeout = 100000000
 		EOF
 
 		cat $PGDATA/postgresql.conf
diff --git a/tests2/lib/bank_client.py b/tests2/lib/bank_client.py
@@ -114,6 +114,15 @@ def initdb(self):
         cur.close()
         conn.close()
 
+    def execute(self, node_id, statements):
+        con = psycopg2.connect(self.dsns[node_id])
+        con.autocommit = True
+        cur = con.cursor()
+        for statement in statements:
+            cur.execute(statement)
+        cur.close()
+        con.close()
+
     def is_data_identic(self):
         hashes = set()
 
diff --git a/tests2/support/two_nodes.yml b/tests2/support/two_nodes.yml
@@ -12,6 +12,7 @@ services:
       POSTGRES_USER: 'pg'
       POSTGRES_DB: 'regression'
       NODE_ID: 1
+      MAJOR: 'on'
       CONNSTRS: >-
         dbname=regression user=pg host=node1,
         dbname=regression user=pg host=node2
@@ -28,6 +29,7 @@ services:
       POSTGRES_USER: 'pg'
       POSTGRES_DB: 'regression'
       NODE_ID: 2
+      MAJOR: 'off'
       CONNSTRS: >-
         dbname=regression user=pg host=node1,
         dbname=regression user=pg host=node2
diff --git a/tests2/test_major.py b/tests2/test_major.py
@@ -58,12 +58,29 @@ def setUp(self):
     def tearDown(self):
         print('Finish test at ',datetime.datetime.utcnow())
 
-    def test_node_partition(self):
-        print('### test_node_partition ###')
+    # def test_partition(self):
+    #     print('### test_node_partition ###')
+
+    #     aggs_failure, aggs = self.performFailure(SingleNodePartition('node2'))
+
+    #     self.assertNoCommits(aggs_failure)
+    #     self.assertIsolation(aggs_failure)
+
+    #     self.assertCommits(aggs)
+    #     self.assertIsolation(aggs)
+
+    def test_partition_major(self):
+        print('### test_partition_major ###')
+
+        # MajorTest.client.execute(0, [
+        #     'alter system set multimaster.major_node to true',
+        #     'select pg_reload_conf();'
+        # ])
 
         aggs_failure, aggs = self.performFailure(SingleNodePartition('node2'))
 
-        self.assertNoCommits(aggs_failure)
+        self.assertCommits(aggs_failure[:1])
+        self.assertNoCommits(aggs_failure[1:])
         self.assertIsolation(aggs_failure)
 
         self.assertCommits(aggs)
