From e6ecc6ca60f274c421dae77c97166ebdff39af6c Mon Sep 17 00:00:00 2001
From: Tobias Lindahl <tobias@happihacking.com>
Date: Tue, 7 Dec 2021 14:07:40 +0100
Subject: [PATCH] fix(emqx_channel): fix race condition in session takeover

Sessions must not enqueue messages when another process is taking over
the client id, since it already passed on the message queue in the
session state.

Without this fix, messages arriving after `{takeover, 'begin'} to a
channel with no connection (i.e., a persistent session) would be lost.
---
 apps/emqx/src/emqx_channel.erl | 22 +++++++++++++---------
 1 file changed, 13 insertions(+), 9 deletions(-)

diff --git a/apps/emqx/src/emqx_channel.erl b/apps/emqx/src/emqx_channel.erl
index 50f9fcf70..1a4a8c1aa 100644
--- a/apps/emqx/src/emqx_channel.erl
+++ b/apps/emqx/src/emqx_channel.erl
@@ -782,6 +782,19 @@ maybe_update_expiry_interval(_Properties, Channel) -> Channel.
 
 -spec(handle_deliver(list(emqx_types:deliver()), channel())
       -> {ok, channel()} | {ok, replies(), channel()}).
+
+handle_deliver(Delivers, Channel = #channel{takeover = true,
+                                            pendings = Pendings,
+                                            session = Session,
+                                            clientinfo = #{clientid := ClientId}}) ->
+    %% NOTE: Order is important here. While the takeover is in
+    %% progress, the session cannot enqueue messages, since it already
+    %% passed on the queue to the new connection in the session state.
+    NPendings = lists:append(
+                  Pendings,
+                  emqx_session:ignore_local(maybe_nack(Delivers), ClientId, Session)),
+    {ok, Channel#channel{pendings = NPendings}};
+
 handle_deliver(Delivers, Channel = #channel{conn_state = disconnected,
                                             session    = Session,
                                             clientinfo = #{clientid := ClientId}}) ->
@@ -793,15 +806,6 @@ handle_deliver(Delivers, Channel = #channel{conn_state = disconnected,
     maybe_mark_as_delivered(Session, Delivers),
     {ok, NChannel};
 
-handle_deliver(Delivers, Channel = #channel{takeover = true,
-                                            pendings = Pendings,
-                                            session = Session,
-                                            clientinfo = #{clientid := ClientId}}) ->
-    NPendings = lists:append(
-                  Pendings,
-                  emqx_session:ignore_local(maybe_nack(Delivers), ClientId, Session)),
-    {ok, Channel#channel{pendings = NPendings}};
-
 handle_deliver(Delivers, Channel = #channel{session = Session,
                                             clientinfo = #{clientid := ClientId}
                                            }) ->