Merge branch 'anders/diameter/watchdog_leak/OTP-11934' into maint

* anders/diameter/watchdog_leak/OTP-11934:
  Simplify sending of 'close' to watchdog
  Fix watchdog table leak

2 files changed, 41 insertions, 32 deletions

diff --git a/lib/diameter/src/base/diameter_service.erl b/lib/diameter/src/base/diameter_service.erl
index 0faf52c1ec..0dc3eb7123 100644
--- a/lib/diameter/src/base/diameter_service.erl
+++ b/lib/diameter/src/base/diameter_service.erl
@@ -881,7 +881,7 @@ watchdog(TPid, [], ?WD_OKAY, ?WD_SUSPECT = To, Wd, State) ->
 
 %% Watchdog has lost its connection.
 watchdog(TPid, [], _, ?WD_DOWN = To, Wd, #state{peerT = PeerT} = S) ->
-    close(Wd, S),
+    close(Wd),
     watchdog_down(Wd, To, S),
     ets:delete(PeerT, TPid);
 
@@ -1199,26 +1199,16 @@ tc(false = No, _, _) ->  %% removed
 %% another watchdog to be able to detect that it should transition
 %% from initial into reopen rather than okay. That someone is either
 %% the accepting watchdog upon reception of a CER from the previously
-%% connected peer, or us after connect_timer timeout.
+%% connected peer, or us after connect_timer timeout or immediately.
 
-close(#watchdog{type = connect}, _) ->
+close(#watchdog{type = connect}) ->
     ok;
+
 close(#watchdog{type = accept,
                 pid = Pid,
-                ref = Ref,
-                options = Opts},
-      #state{service_name = SvcName}) ->
-    c(Pid, diameter_config:have_transport(SvcName, Ref), Opts).
-
-%% Tell watchdog to (maybe) die later ...
-c(Pid, true, Opts) ->
+                options = Opts}) ->
     Tc = connect_timer(Opts, 2*?DEFAULT_TC),
-    erlang:send_after(Tc, Pid, close);
-
-%% ... or now.
-c(Pid, false, _Opts) ->
-    Pid ! close.
-
+    erlang:send_after(Tc, Pid, close).
 %% The RFC's only document the behaviour of Tc, our connect_timer,
 %% for the establishment of connections but we also give
 %% connect_timer semantics for a listener, being the time within
diff --git a/lib/diameter/src/base/diameter_watchdog.erl b/lib/diameter/src/base/diameter_watchdog.erl
index 53e659e3f6..017a520467 100644
--- a/lib/diameter/src/base/diameter_watchdog.erl
+++ b/lib/diameter/src/base/diameter_watchdog.erl
@@ -49,8 +49,6 @@
 
 -define(IS_NATURAL(N), (is_integer(N) andalso 0 =< N)).
 
--define(CHOOSE(B,T,F), if (B) -> T; true -> F end).
-
 -record(config,
         {suspect = 1 :: non_neg_integer(),    %% OKAY -> SUSPECT
          okay    = 3 :: non_neg_integer()}).  %% REOPEN -> OKAY
@@ -313,14 +311,13 @@ code_change(_, State, _) ->
 %% The state transitions documented here are extracted from RFC 3539,
 %% the commentary is ours.
 
-%% Service or watchdog is telling the watchdog of an accepting
-%% transport to die after connect_timer expiry or reestablished
-%% connection (in another transport process) respectively.
-transition(close, #watchdog{status = down}) ->
+%% Service is telling the watchdog of an accepting transport to die
+%% following transport death in state INITIAL, or after connect_timer
+%% expiry; or another watchdog is saying the same after reestablishing
+%% a connection previously had by this one.
+transition(close, #watchdog{}) ->
     {{accept, _}, _, _} = getr(restart), %% assert
     stop;
-transition(close, #watchdog{}) ->
-    ok;
 
 %% Service is asking for the peer to be taken down gracefully.
 transition({shutdown, Pid, _}, #watchdog{parent = Pid,
@@ -403,18 +400,39 @@ transition({open = Key, TPid, _Hosts, T},
 %%   REOPEN        Connection down      CloseConnection()
 %%                                      SetWatchdog()        DOWN
 
+%% Transport has died after service requested termination ...
 transition({'DOWN', _, process, TPid, _Reason},
            #watchdog{transport = TPid,
                      shutdown = true}) ->
     stop;
 
+%% ... or not.
 transition({'DOWN', _, process, TPid, _Reason},
            #watchdog{transport = TPid,
-                     status = T}
-           = S) ->
-    set_watchdog(S#watchdog{status = ?CHOOSE(initial == T, T, down),
-                            pending = false,
-                            transport = undefined});
+                     status = T,
+                     restrict = {_,R}}
+           = S0) ->
+    S = S0#watchdog{pending = false,
+                    transport = undefined},
+    {{M,_}, _, _} = getr(restart),
+
+    %% Close an accepting watchdog immediately if there's no
+    %% restriction on the number of connections to the same peer: the
+    %% state machine never enters state REOPEN in this case. The
+    %% 'close' message (instead of stop) is so as not to bypass the
+    %% sending of messages to the service process in handle_info/2.
+
+    if T /= initial, M == accept, not R ->
+            send(self(), close),
+            S#watchdog{status = down};
+       T /= initial ->
+            set_watchdog(S#watchdog{status = down});
+       M == connect ->
+            set_watchdog(S);
+       M == accept ->
+            send(self(), close),
+            S
+    end;
 
 %% Incoming message.
 transition({recv, TPid, Name, Pkt}, #watchdog{transport = TPid} = S) ->
@@ -740,7 +758,7 @@ timeout(#watchdog{status = T} = S)
 
 restart(#watchdog{transport = undefined} = S) ->
     restart(getr(restart), S);
-restart(S) ->
+restart(S) ->  %% reconnect has won race with timeout
     S.
 
 %% restart/2
@@ -770,9 +788,10 @@ restart({{connect, _} = T, Opts, Svc},
 %% die. Note that a state machine never enters state REOPEN in this
 %% case.
 restart({{accept, _}, _, _}, #watchdog{restrict = {_, false}}) ->
-    stop;
+    stop;  %% 'DOWN' was in old code: 'close' was not sent
 
-%% Otherwise hang around until told to die.
+%% Otherwise hang around until told to die, either by the service or
+%% by another watchdog.
 restart({{accept, _}, _, _}, S) ->
     S.