Merge branch 'master' of github.com:postgrespro/postgres_cluster

kvap · kvap · commit 86ef2cdba633 · 2016-03-30T11:35:17.000+03:00
diff --git a/contrib/mmts/Makefile b/contrib/mmts/Makefile
@@ -1,5 +1,5 @@
 MODULE_big = multimaster
-OBJS = multimaster.o arbiter.o bytebuf.o bgwpool.o pglogical_output.o pglogical_proto.o pglogical_receiver.o pglogical_apply.o pglogical_hooks.o pglogical_config.o ddd.o bkb.o
+OBJS = multimaster.o raftable.o arbiter.o bytebuf.o bgwpool.o pglogical_output.o pglogical_proto.o pglogical_receiver.o pglogical_apply.o pglogical_hooks.o pglogical_config.o ddd.o bkb.o
 
 override CPPFLAGS += -I../raftable
 
@@ -8,7 +8,7 @@ DATA = multimaster--1.0.sql
 
 .PHONY: all
 
-all: multimaster.o multimaster.so
+all: multimaster.so
 
 PG_CPPFLAGS = -I$(libpq_srcdir) -DUSE_PGLOGICAL_OUTPUT
 SHLIB_LINK = $(libpq)
diff --git a/contrib/mmts/multimaster.c b/contrib/mmts/multimaster.c
@@ -58,7 +58,7 @@
 #include "raftable.h"
 #include "multimaster.h"
 #include "ddd.h"
-#include "paxos.h"
+#include "raftable.h"
 
 typedef struct { 
     TransactionId xid;    /* local transaction ID   */
@@ -172,6 +172,7 @@ int   MtmConnectAttempts;
 int   MtmConnectTimeout;
 int   MtmKeepaliveTimeout;
 int   MtmReconnectAttempts;
+bool  MtmUseRaftable;
 MtmConnectionInfo* MtmConnections;
 
 static char* MtmConnStrs;
@@ -1104,7 +1105,7 @@ MtmBuildConnectivityMatrix(nodemask_t* matrix, bool nowait)
 	int i, j, n = MtmNodes;
 	for (i = 0; i < n; i++) { 
 		if (i+1 != MtmNodeId) { 
-			void* data = PaxosGet(psprintf("node-mask-%d", i+1), NULL, NULL, nowait);
+			void* data = RaftableGet(psprintf("node-mask-%d", i+1), NULL, NULL, nowait);
 			if (data == NULL) { 
 				return false;
 			}
@@ -1134,7 +1135,7 @@ bool MtmRefreshClusterStatus(bool nowait)
 	int clique_size;
 	int i;
 
-	if (!MtmBuildConnectivityMatrix(matrix, nowait)) { 
+	if (!MtmUseRaftable || !MtmBuildConnectivityMatrix(matrix, nowait)) { 
 		/* RAFT is not available */
 		return false;
 	}
@@ -1194,7 +1195,7 @@ void MtmCheckQuorum(void)
 void MtmOnNodeDisconnect(int nodeId)
 {
 	BIT_SET(Mtm->connectivityMask, nodeId-1);
-	PaxosSet(psprintf("node-mask-%d", MtmNodeId), &Mtm->connectivityMask, sizeof Mtm->connectivityMask, false);
+	RaftableSet(psprintf("node-mask-%d", MtmNodeId), &Mtm->connectivityMask, sizeof Mtm->connectivityMask, false);
 
 	/* Wait more than socket KEEPALIVE timeout to let other nodes update their statuses */
 	MtmSleep(MtmKeepaliveTimeout);
@@ -1213,52 +1214,9 @@ void MtmOnNodeDisconnect(int nodeId)
 void MtmOnNodeConnect(int nodeId)
 {
 	BIT_CLEAR(Mtm->connectivityMask, nodeId-1);
-	PaxosSet(psprintf("node-mask-%d", MtmNodeId), &Mtm->connectivityMask, sizeof Mtm->connectivityMask, false);
+	RaftableSet(psprintf("node-mask-%d", MtmNodeId), &Mtm->connectivityMask, sizeof Mtm->connectivityMask, false);
 }
 
-/*
- * Paxos function stubs (until them are miplemented)
- */
-void* PaxosGet(char const* key, int* size, PaxosTimestamp* ts, bool nowait)
-{
-	unsigned enclen, declen, len;
-	char *enc, *dec;
-	Assert(ts == NULL); // not implemented
-
-	enc = raftable_get(key);
-	if (enc == NULL)
-	{
-		*size = 0;
-		return NULL;
-	}
-
-	enclen = strlen(enc);
-	declen = hex_dec_len(enc, enclen);
-	dec = palloc(declen);
-	len = hex_decode(enc, enclen, dec);
-	pfree(enc);
-	Assert(len == declen);
-
-	if (size != NULL) {
-		*size = declen;
-	}
-	return dec;
-}
-
-void  PaxosSet(char const* key, void const* value, int size, bool nowait)
-{
-	unsigned enclen, declen, len;
-	char *enc, *dec;
-
-	enclen = hex_enc_len(value, size);
-	enc = palloc(enclen) + 1;
-	len = hex_encode(value, size, enc);
-	Assert(len == enclen);
-	enc[len] = '\0';
-
-	raftable_set(key, enc, nowait ? 1 : INT_MAX);
-	pfree(enc);
-}
 
 
 /*
@@ -1485,6 +1443,19 @@ _PG_init(void)
 		NULL
 	);
 
+	DefineCustomBoolVariable(
+		"multimaster.use_raftable",
+		"Use raftable plugin for internode communication",
+		NULL,
+		&MtmUseRaftable,
+		false,
+		PGC_BACKEND,
+		0,
+		NULL,
+		NULL,
+		NULL
+	);
+
 	DefineCustomIntVariable(
 		"multimaster.workers",
 		"Number of multimaster executor workers per node",
@@ -1775,6 +1746,10 @@ MtmReplicationStartupHook(struct PGLogicalStartupHookArgs* args)
 			break;
 		}
 	}
+	if (isRecoverySession) { 
+		MTM_INFO("%d: PGLOGICAL startup hook\n", MyProcPid);
+		sleep(30);
+	}
 	MtmLock(LW_EXCLUSIVE);
 	if (isRecoverySession) {
 		elog(WARNING, "Node %d start recovery of node %d", MtmNodeId, MtmReplicationNodeId);
@@ -1807,7 +1782,7 @@ MtmReplicationTxnFilterHook(struct PGLogicalTxnFilterArgs* args)
 	bool res = Mtm->status != MTM_RECOVERY
 		&& (args->origin_id == InvalidRepOriginId 
 			|| MtmIsRecoveredNode(MtmReplicationNodeId));
-	MTM_INFO("%d: MtmReplicationTxnFilterHook->%d\n", MyProcPid, res);
+	MTM_TRACE("%d: MtmReplicationTxnFilterHook->%d\n", MyProcPid, res);
 	return res;
 }
 
@@ -2376,16 +2351,16 @@ MtmDetectGlobalDeadLock(PGPROC* proc)
 		
         ByteBufferAlloc(&buf);
         EnumerateLocks(MtmSerializeLock, &buf);
-		PaxosSet(psprintf("lock-graph-%d", MtmNodeId), buf.data, buf.used, true);
+		RaftableSet(psprintf("lock-graph-%d", MtmNodeId), buf.data, buf.used, true);
 		MtmGraphInit(&graph);
 		MtmGraphAdd(&graph, (GlobalTransactionId*)buf.data, buf.used/sizeof(GlobalTransactionId));
         ByteBufferFree(&buf);
 		for (i = 0; i < MtmNodes; i++) { 
 			if (i+1 != MtmNodeId && !BIT_CHECK(Mtm->disabledNodeMask, i)) { 
 				int size;
-				void* data = PaxosGet(psprintf("lock-graph-%d", i+1), &size, NULL, true);
+				void* data = RaftableGet(psprintf("lock-graph-%d", i+1), &size, NULL, true);
 				if (data == NULL) { 
-					return true; /* Just temporary hack until no Paxos */
+					return true; /* If using Raftable is disabled */
 				} else { 
 					MtmGraphAdd(&graph, (GlobalTransactionId*)data, size/sizeof(GlobalTransactionId));
 				}
diff --git a/contrib/mmts/multimaster.control b/contrib/mmts/multimaster.control
@@ -3,4 +3,3 @@ default_version = '1.0'
 module_pathname = '$libdir/multimaster'
 schema = mtm
 relocatable = false
-requires = raftable
diff --git a/contrib/mmts/raftable.h b/contrib/mmts/raftable.h
@@ -1,11 +1,11 @@
-#ifndef __PAXOS_H__
-#define __PAXOS_H__
+#ifndef __RAFTABLE_H__
+#define __RAFTABLE_H__
 
-typedef struct PaxosTimestamp { 
+typedef struct RaftableTimestamp { 
 	time_t time;   /* local time at master */
 	uint32 master; /* master node for this operation */
-	uint32 psn;    /* PAXOS serial number */
-} PaxosTimestamp;
+	uint32 psn;    /* RAFTABLE serial number */
+} RaftableTimestamp;
 
 /*
  * Get key value.
@@ -15,18 +15,24 @@ typedef struct PaxosTimestamp {
  * If "ts" is not NULL, then it is assigned timestamp of last update of this value
  * If RAFT master is not accessible, then depending non value of "nowait" parameter, this funciton should either block until RAFT quorum is reached, either report error.
  */
-extern void* PaxosGet(char const* key, int* size, PaxosTimestamp* ts, bool nowait);
+extern void* RaftableGet(char const* key, int* size, RaftableTimestamp* ts, bool nowait);
 
 /*
  * Set new value for the specified key. IF value is NULL, then key should be deleted.
  * If RAFT master is not accessible, then depending non value of "nowait" parameter, this funciton should either block until RAFT quorum is reached, either report error.
  */
-extern void  PaxosSet(char const* key, void const* value, int size, bool nowait);
+extern void  RaftableSet(char const* key, void const* value, int size, bool nowait);
 
 /* 
  * If key doesn't exists or its value is not equal to the specified value then store this value and return true.
  * Otherwise do nothing and return false.
  * If RAFT master is not accessible, then depending non value of "nowait" parameter, this funciton should either block until RAFT quorum is reached, either report error.
  */
-extern bool  PaxosCAS(char const* key, char const* value, bool nowait);
+extern bool  RaftableCAS(char const* key, char const* value, bool nowait);
+
+typedef void* (*raftable_get_t)(char const* key, int* size, int timeout);
+typedef void (*raftable_set_t)(char const* key, void const* value, int size, int timeout);
+
+extern bool MtmUseRaftable;
+
 #endif
diff --git a/src/backend/replication/logical/reorderbuffer.c b/src/backend/replication/logical/reorderbuffer.c
@@ -160,7 +160,6 @@ static const Size max_cached_changes = 4096 * 2;
 static const Size max_cached_tuplebufs = 4096 * 2;		/* ~8MB */
 static const Size max_cached_transactions = 512;
 
-
 /* ---------------------------------------
  * primary reorderbuffer support routines
  * ---------------------------------------
