Merge pull request #4412 from soloestoy/bugfix-psync2

PSYNC2: safe free backlog when reach the time limit and others

src/rdb.c

@@ -2000,6 +2000,9 @@ void bgsaveCommand(client *c) {
         }
     }
 
+    rdbSaveInfo rsi, *rsiptr;
+    rsiptr = rdbPopulateSaveInfo(&rsi);
+
     if (server.rdb_child_pid != -1) {
         addReplyError(c,"Background save already in progress");
     } else if (server.aof_child_pid != -1) {
@@ -2012,7 +2015,7 @@ void bgsaveCommand(client *c) {
                 "Use BGSAVE SCHEDULE in order to schedule a BGSAVE whenever "
                 "possible.");
         }
-    } else if (rdbSaveBackground(server.rdb_filename,NULL) == C_OK) {
+    } else if (rdbSaveBackground(server.rdb_filename,rsiptr) == C_OK) {
         addReplyStatus(c,"Background saving started");
     } else {
         addReply(c,shared.err);
@@ -2033,22 +2036,44 @@ rdbSaveInfo *rdbPopulateSaveInfo(rdbSaveInfo *rsi) {
     *rsi = rsi_init;
 
     /* If the instance is a master, we can populate the replication info
-     * in all the cases, even if sometimes in incomplete (but safe) form. */
-    if (!server.masterhost) {
-        if (server.repl_backlog) rsi->repl_stream_db = server.slaveseldb;
-        /* Note that if repl_backlog is NULL, it means that histories
-         * following from this point will trigger a full synchronization
-         * generating a SELECT statement, so we can leave the currently
-         * selected DB set to -1. This allows a restarted master to reload
-         * its replication ID/offset when there are no connected slaves. */
+     * only when repl_backlog is not NULL. If the repl_backlog is NULL,
+     * it means that the instance isn't in any replication chains. In this
+     * scenario the replication info is useless, because when a slave
+     * connect to us, the NULL repl_backlog will trigger a full synchronization,
+     * at the same time we will use a new replid and clear replid2.
+     * And remember that after free backlog if we reach repl_backlog_time_limit,
+     * we will use a new replid and clear replid2 too. So there is only one
+     * scenario which can make repl_stream_db be -1, that is the instance is
+     * a master, and it have repl_backlog, but server.slaveseldb is -1. */
+    if (!server.masterhost && server.repl_backlog) {
+        rsi->repl_stream_db = server.slaveseldb == -1 ? 0 : server.slaveseldb;
+        /* Note that when server.slaveseldb is -1, it means that this master
+         * didn't apply any write commands after a full synchronization.
+         * So we can let repl_stream_db be 0, this allows a restarted slave
+         * to reload replication ID/offset, it's safe because the next write
+         * command must generate a SELECT statement. */
         return rsi;
     }
 
-    /* If the instance is a slave we need a connected master in order to
-     * fetch the currently selected DB. */
+    /* If the instance is a slave we need a connected master
+     * in order to fetch the currently selected DB. */
     if (server.master) {
         rsi->repl_stream_db = server.master->db->id;
         return rsi;
     }
+    /* It is useful to persist cached_master's db id inside RDB file.
+     * When a slave lost master's connection, server.master will be
+     * cached as server.cached_master, after that a slave can not
+     * increment the master_repl_offset because slave only apply data
+     * from connected master, so the cached_master can hold right
+     * replication info. But please note that this action is safe
+     * only after we fix the free backlog problem, because when a master
+     * turn to be a slave, it will use itself as the server.cached_master,
+     * that is dangerous if we didn't use a new replication ID after
+     * free backlog. */
+    if (server.cached_master) {
+        rsi->repl_stream_db = server.cached_master->db->id;
+        return rsi;
+    }
     return NULL;
 }

src/replication.c

@@ -2613,6 +2613,18 @@ void replicationCron(void) {
         time_t idle = server.unixtime - server.repl_no_slaves_since;
 
         if (idle > server.repl_backlog_time_limit) {
+            /* When we free the backlog, we always use a new
+             * replication ID and clear the ID2. Since without
+             * backlog we can not increment master_repl_offset
+             * even do write commands, that may lead to inconsistency
+             * when we try to connect a "slave-before" master
+             * (if this master is our slave before, our replid
+             * equals the master's replid2). As the master have our
+             * history, so we can match the master's replid2 and
+             * second_replid_offset, that make partial sync work,
+             * but the data is inconsistent. */
+            changeReplicationId();
+            clearReplicationId2();
             freeReplicationBacklog();
             serverLog(LL_NOTICE,
                 "Replication backlog freed after %d seconds "

src/server.c

@@ -3540,7 +3540,8 @@ void loadDataFromDisk(void) {
                 rsi.repl_id_is_set &&
                 rsi.repl_offset != -1 &&
                 /* Note that older implementations may save a repl_stream_db
-                 * of -1 inside the RDB file. */
+                 * of -1 inside the RDB file in a wrong way, see more information
+                 * in function rdbPopulateSaveInfo. */
                 rsi.repl_stream_db != -1)
             {
                 memcpy(server.replid,rsi.repl_id,sizeof(server.replid));