MB-67804: Make sure cb_dist:select does not block net_kernel

All erlang proto modules try to resolve the hostname
when select() is called. This can block the net_kernel and lead
to node start problem (while we are waiting for dns couchdb
timeout expires).
At the same time, the name resolution is not really needed in
cb_dist:select() as it already knows the protocol (afamily and
encryption) that it will use.

Change-Id: If0c7af5298ed04c872165b4efbf6e7d7a8c21ad4
Reviewed-on: https://review.couchbase.org/c/ns_server/+/231603
Well-Formed: Build Bot <[email protected]>
Tested-by: Timofey Barmin <[email protected]>
Well-Formed: Restriction Checker
Reviewed-by: Peter Searby <[email protected]>

Author: timofey-barmin

src/cb_dist.erl

@@ -143,12 +143,18 @@ accept_connection(_, {ConRef, HandshakeProcPid, Module, ConnectionSocket},
 
 -spec select(Node :: atom()) -> true | false.
 select(Node) ->
-    try get_preferred_dist(Node) of
-        Module -> Module:select(Node)
-    catch
-        _:Error ->
-            error_msg("Select for ~p failed. Couldn't find preferred proto: ~p",
-                      [Node, Error]),
+    case dist_util:split_node(Node) of
+        {node, _Name, _Host} ->
+            %% Not proxying select() to preferred proto to avoid blocking
+            %% net_kernel (select() is called by net_kernel).
+            %% Select in inet_tcp_dist (and other protocols) can block
+            %% because it resolves the hostname.
+            %% It seems like the name resolution is not really needed here
+            %% as we already know the protocol (afamily and encryption) that
+            %% we want to use.
+            true;
+        _ ->
+            error_msg("Select failed. Invalid node name: ~p", [Node]),
             false
     end.