fix: retry request when error reason is normal

Fixes https://emqx.atlassian.net/browse/EMQX-9656

changelog.md

@@ -1,5 +1,9 @@
 # ehttpc changes
 
+## 0.4.8
+
+- Fix an issue where a race condition would yield an `{error, normal}` return value.  This can be caused when the `gun` process terminates when the remote server closes the connection for whatever reason.  In this case, we simply retry without consuming "retry credits".
+
 ## 0.4.7
 
 - Fix crash when using body headers and no body. When one sent a message with, for example, the content-type header and no body, the underlying gun library would expect a body to come with a gun:data/3 call and would not finish the request. The next request would then crash. See the following issue for more information: https://github.com/ninenines/gun/issues/141

src/ehttpc.app.src

@@ -1,6 +1,6 @@
 {application, ehttpc,
  [{description, "HTTP Client for Erlang/OTP"},
-  {vsn, "0.4.7"},
+  {vsn, "0.4.8"},
   {registered, []},
   {applications, [kernel,
                   stdlib,

src/ehttpc.appup.src

@@ -1,6 +1,9 @@
-%% -*-: erlang -*-
-{"0.4.7",
+%% -*- mode: erlang -*-
+{"0.4.8",
    [
+     {"0.4.7", [
+       {load_module, ehttpc, brutal_purge, soft_purge, []}
+     ]},
      {"0.4.6", [
        {load_module, ehttpc, brutal_purge, soft_purge, []}
      ]},
@@ -47,6 +50,9 @@
      ]}
    ],
    [
+    {"0.4.7", [
+       {load_module, ehttpc, brutal_purge, soft_purge, []}
+     ]},
      {"0.4.6", [
        {load_module, ehttpc, brutal_purge, soft_purge, []}
      ]},

src/ehttpc.erl

@@ -141,6 +141,15 @@ request(Worker, Method, Request, Timeout, Retry) when is_pid(Worker) ->
     try gen_server:call(Worker, ?REQ(Method, Request, ExpireAt), Timeout + 500) of
         %% gun will reply {gun_down, _Client, _, normal, _KilledStreams, _} message
         %% when connection closed by keepalive
+
+        %% If `Reason' = `normal', we should just retry without
+        %% consuming a retry credit, as it could be a race condition
+        %% where the gun process is down (e.g.: when the server closes
+        %% the connection), and then requests would be ignored while
+        %% the `gun' process is terminating.
+        {error, normal} ->
+            ?LOG(warning, "retrying at no cost!", []),
+            request(Worker, Method, Request, Timeout, Retry);
         {error, Reason} when Retry < 1 ->
             {error, Reason};
         {error, _} ->