Fix crash during handshake and cluster shards call (#10942)

* Fix an engine crash when there are nodes in handshaking and a user calls cluster shards

2 files changed, 20 insertions, 3 deletions

diff --git a/src/cluster.c b/src/cluster.c
index b434bcc02..ae353e1dd 100644
--- a/src/cluster.c
+++ b/src/cluster.c
@@ -1220,10 +1220,14 @@ clusterNode *clusterLookupNode(const char *name, int length) {
     return dictGetVal(de);
 }
 
-/* Get all the nodes serving the same slots as myself. */
+/* Get all the nodes serving the same slots as the given node. */
 list *clusterGetNodesServingMySlots(clusterNode *node) {
     list *nodes_for_slot = listCreate();
     clusterNode *my_primary = nodeIsMaster(node) ? node : node->slaveof;
+
+    /* This function is only valid for fully connected nodes, so
+     * they should have a known primary. */
+    serverAssert(my_primary);
     listAddNodeTail(nodes_for_slot, my_primary);
     for (int i=0; i < my_primary->numslaves; i++) {
         listAddNodeTail(nodes_for_slot, my_primary->slaves[i]);
@@ -5103,7 +5107,7 @@ void clusterReplyShards(client *c) {
      * information and an empty slots array. */
     while((de = dictNext(di)) != NULL) {
         clusterNode *n = dictGetVal(de);
-        if (nodeIsSlave(n)) {
+        if (!nodeIsMaster(n)) {
             /* You can force a replica to own slots, even though it'll get reverted,
              * so freeing the slot pair here just in case. */
             clusterFreeNodesSlotsInfo(n);
diff --git a/tests/cluster/tests/28-cluster-shards.tcl b/tests/cluster/tests/28-cluster-shards.tcl
index fe794f2b7..9894d4b33 100644
--- a/tests/cluster/tests/28-cluster-shards.tcl
+++ b/tests/cluster/tests/28-cluster-shards.tcl
@@ -182,4 +182,17 @@ test "Test the replica reports a loading state while it's loading" {
 
     # Final sanity, the replica agrees it is online. 
     assert_equal "online" [dict get [get_node_info_from_shard $replica_cluster_id $replica_id "node"] health]
-}
\ No newline at end of file
+}
+
+test "Regression test for a crash when calling SHARDS during handshake" {
+    # Reset forget a node, so we can use it to establish handshaking connections
+    set id [R 19 CLUSTER MYID]
+    R 19 CLUSTER RESET HARD
+    for {set i 0} {$i < 19} {incr i} {
+        R $i CLUSTER FORGET $id
+    }
+    R 19 cluster meet 127.0.0.1 [get_instance_attrib redis 0 port]
+    # This should line would previously crash, since all the outbound
+    # connections were in handshake state.
+    R 19 CLUSTER SHARDS
+}