Apply 'show_econnreset' socket option to send errors as well

Up till now all send errors have been translated into a generic
{error, closed}. This patch allows {error, econnreset} to be
returned on send errors when it is detected that the TCP peer
has sent an RST.

3 files changed, 257 insertions, 11 deletions

diff --git a/erts/emulator/drivers/common/inet_drv.c b/erts/emulator/drivers/common/inet_drv.c
index b5903f0c07..a8c58a8149 100644
--- a/erts/emulator/drivers/common/inet_drv.c
+++ b/erts/emulator/drivers/common/inet_drv.c
@@ -836,6 +836,8 @@ static int my_strncasecmp(const char *s1, const char *s2, size_t n)
 #define TCP_ADDF_PENDING_SHUTDOWN \
 		(TCP_ADDF_PENDING_SHUT_WR | TCP_ADDF_PENDING_SHUT_RDWR)
 #define TCP_ADDF_SHOW_ECONNRESET   64 /* Tell user about incoming RST */
+#define TCP_ADDF_DELAYED_ECONNRESET 128 /* An ECONNRESET error occured on send or shutdown */
+#define TCP_ADDF_SHUTDOWN_WR_DONE 256 /* A shutdown(sock, SHUT_WR) or SHUT_RDWR was made */
 
 /* *_REQ_* replies */
 #define INET_REP_ERROR       0
@@ -6140,7 +6142,12 @@ static int inet_set_opts(inet_descriptor* desc, char* ptr, int len)
                 desc->active_count = 0;
 	    if ((desc->stype == SOCK_STREAM) && (desc->active != INET_PASSIVE) && 
 		(desc->state == INET_STATE_CLOSED)) {
-		tcp_closed_message((tcp_descriptor *) desc);
+		tcp_descriptor *tdesc = (tcp_descriptor *) desc;
+		if (tdesc->tcp_add_flags & TCP_ADDF_DELAYED_ECONNRESET) {
+		    tdesc->tcp_add_flags &= ~TCP_ADDF_DELAYED_ECONNRESET;
+		    tcp_error_message(tdesc, ECONNRESET);
+		}
+		tcp_closed_message(tdesc);
 		if (desc->exitf) {
 		    driver_exit(desc->port, 0);
 		    return 0; /* Give up on this socket, descriptor lost */
@@ -9466,7 +9473,11 @@ static ErlDrvSSizeT tcp_inet_ctl(ErlDrvData e, unsigned int cmd,
 	    if (desc->tcp_add_flags & TCP_ADDF_DELAYED_CLOSE_RECV) {
 		desc->tcp_add_flags &= ~(TCP_ADDF_DELAYED_CLOSE_RECV|
 					 TCP_ADDF_DELAYED_CLOSE_SEND);
-		return ctl_reply(INET_REP_ERROR, "closed", 6, rbuf, rsize);
+		if (desc->tcp_add_flags & TCP_ADDF_DELAYED_ECONNRESET) {
+		    desc->tcp_add_flags &= ~TCP_ADDF_DELAYED_ECONNRESET;
+		    return ctl_reply(INET_REP_ERROR, "econnreset", 10, rbuf, rsize);
+		} else
+		    return ctl_reply(INET_REP_ERROR, "closed", 6, rbuf, rsize);
 	    }
 	    return ctl_error(ENOTCONN, rbuf, rsize);
 	}
@@ -9527,6 +9538,8 @@ static ErlDrvSSizeT tcp_inet_ctl(ErlDrvData e, unsigned int cmd,
 	    return ctl_reply(INET_REP_OK, NULL, 0, rbuf, rsize);
 	}
 	if (IS_SOCKET_ERROR(sock_shutdown(INETP(desc)->s, how))) {
+	    if (how != TCP_SHUT_RD)
+		desc->tcp_add_flags |= TCP_ADDF_SHUTDOWN_WR_DONE;
 	    return ctl_error(sock_errno(), rbuf, rsize);
 	} else {
 	    return ctl_reply(INET_REP_OK, NULL, 0, rbuf, rsize);
@@ -9661,7 +9674,13 @@ static void tcp_inet_commandv(ErlDrvData e, ErlIOVec* ev)
     if (!IS_CONNECTED(INETP(desc))) {
 	if (desc->tcp_add_flags & TCP_ADDF_DELAYED_CLOSE_SEND) {
 	    desc->tcp_add_flags &= ~TCP_ADDF_DELAYED_CLOSE_SEND;
-	    inet_reply_error_am(INETP(desc), am_closed);
+	    if (desc->tcp_add_flags & TCP_ADDF_DELAYED_ECONNRESET) {
+		/* Don't clear flag. Leave it enabled for the next receive
+		 * operation.
+		 */
+		inet_reply_error(INETP(desc), ECONNRESET);
+	    } else
+		inet_reply_error_am(INETP(desc), am_closed);
 	}
 	else
 	    inet_reply_error(INETP(desc), ENOTCONN);
@@ -10578,6 +10597,9 @@ static int tcp_inet_input(tcp_descriptor* desc, HANDLE event)
 
 static int tcp_send_or_shutdown_error(tcp_descriptor* desc, int err)
 {
+    int show_econnreset = (err == ECONNRESET
+			   && desc->tcp_add_flags & TCP_ADDF_SHOW_ECONNRESET);
+
     /*
      * If the port is busy, we must do some clean-up before proceeding.
      */
@@ -10593,14 +10615,21 @@ static int tcp_send_or_shutdown_error(tcp_descriptor* desc, int err)
 
     /*
      * We used to handle "expected errors" differently from unexpected ones.
-     * Now we handle all errors in the same way. We just have to distinguish
-     * between passive and active sockets.
+     * Now we handle all errors in the same way (unless the show_econnreset
+     * socket option is enabled). We just have to distinguish between passive
+     * and active sockets.
      */
     DEBUGF(("driver_failure_eof(%ld) in %s, line %d\r\n",
 	    (long)desc->inet.port, __FILE__, __LINE__));
     if (desc->inet.active) {
-	tcp_closed_message(desc);
-	inet_reply_error_am(INETP(desc), am_closed);
+	if (show_econnreset) {
+	    tcp_error_message(desc, err);
+	    tcp_closed_message(desc);
+	    inet_reply_error(INETP(desc), err);
+	} else {
+	    tcp_closed_message(desc);
+	    inet_reply_error_am(INETP(desc), am_closed);
+	}
 	if (desc->inet.exitf)
 	    driver_exit(desc->inet.port, 0);
 	else
@@ -10612,7 +10641,10 @@ static int tcp_send_or_shutdown_error(tcp_descriptor* desc, int err)
 	erl_inet_close(INETP(desc));
 
 	if (desc->inet.caller) {
-	    inet_reply_error_am(INETP(desc), am_closed);
+	    if (show_econnreset)
+		inet_reply_error(INETP(desc), err);
+	    else
+		inet_reply_error_am(INETP(desc), am_closed);
 	}
 	else {
 	    /* No blocking send op to reply to right now.
@@ -10629,12 +10661,46 @@ static int tcp_send_or_shutdown_error(tcp_descriptor* desc, int err)
 	 * in the receive operation.
 	 */
 	desc->tcp_add_flags |= TCP_ADDF_DELAYED_CLOSE_RECV;
+
+	if (show_econnreset) {
+	    /* Return {error, econnreset} instead of {error, closed}
+	     * on send or receive operations.
+	     */
+	    desc->tcp_add_flags |= TCP_ADDF_DELAYED_ECONNRESET;
+	}
     }
     return -1;
 }
 
 static int tcp_send_error(tcp_descriptor* desc, int err)
 {
+    /* EPIPE errors usually occur in one of three ways:
+     * 1. We write to a socket when we've already shutdown() the write side. On
+     *    Windows the error returned for this is ESHUTDOWN rather than EPIPE.
+     * 2. The TCP peer sends us an RST through no fault of our own (perhaps
+     *    by aborting the connection using SO_LINGER) and we then attempt
+     *    to write to the socket. On Linux and Windows we would actually
+     *    receive an ECONNRESET error for this, but on the BSDs, Darwin,
+     *    Illumos and presumably Solaris, it's an EPIPE.
+     * 3. We cause the TCP peer to send us an RST by writing to a socket
+     *    after we receive a FIN from them. Our first write will be
+     *    successful, but if the they have closed the connection (rather
+     *    than just shutting down the write side of it) this will cause their
+     *    OS to send us an RST. Then, when we attempt to write to the socket
+     *    a second time, we will get an EPIPE error. On Windows we get an
+     *    ECONNABORTED.
+     *
+     * What we are going to do here is to treat all EPIPE messages that aren't
+     * of type 1 as ECONNRESET errors. This will allow users who have the
+     * show_econnreset socket option enabled to receive {error, econnreset} on
+     * both send and recv operations to indicate that an RST has been received.
+     */
+#ifdef __WIN_32__
+    if (err == ECONNABORTED)
+	err = ECONNRESET;
+#endif
+    if (err == EPIPE && !(desc->tcp_add_flags & TCP_ADDF_SHUTDOWN_WR_DONE))
+	err = ECONNRESET;
     return tcp_send_or_shutdown_error(desc, err);
 }
 
@@ -10854,6 +10920,8 @@ static void tcp_shutdown_async(tcp_descriptor* desc)
 		TCP_SHUT_WR : TCP_SHUT_RDWR;
     if (IS_SOCKET_ERROR(sock_shutdown(INETP(desc)->s, how)))
 	tcp_shutdown_error(desc, sock_errno());
+    else
+	desc->tcp_add_flags |= TCP_ADDF_SHUTDOWN_WR_DONE;
 }
 
 #ifdef __OSE__
diff --git a/lib/kernel/doc/src/inet.xml b/lib/kernel/doc/src/inet.xml
index a364f9e0c9..4dd9e0e221 100644
--- a/lib/kernel/doc/src/inet.xml
+++ b/lib/kernel/doc/src/inet.xml
@@ -1056,7 +1056,10 @@ setcap cap_sys_admin,cap_sys_ptrace,cap_dac_read_search+epi beam.smp
               active mode, the controlling process will receive a
               <c>{tcp_error, Socket, econnreset}</c> message
               before the usual <c>{tcp_closed, Socket}</c>, as is
-              the case for any other socket error.</p>
+              the case for any other socket error. Calls to
+              <seealso marker="gen_tcp#send/2">gen_tcp:send/2</seealso>
+              will also return <c>{error, econnreset}</c> when it
+              is detected that a TCP peer has sent an RST.</p>
             <p>A connected socket returned from
               <seealso marker="gen_tcp#accept/1">gen_tcp:accept/1</seealso>
               will inherit the <c>show_econnreset</c> setting from the
diff --git a/lib/kernel/test/gen_tcp_misc_SUITE.erl b/lib/kernel/test/gen_tcp_misc_SUITE.erl
index f08f28b650..ddd9d356ee 100644
--- a/lib/kernel/test/gen_tcp_misc_SUITE.erl
+++ b/lib/kernel/test/gen_tcp_misc_SUITE.erl
@@ -32,7 +32,10 @@
 	 otp_3924/1, otp_3924_sender/4, closed_socket/1,
 	 shutdown_active/1, shutdown_passive/1, shutdown_pending/1,
 	 show_econnreset_active/1, show_econnreset_active_once/1,
-	 show_econnreset_passive/1,
+	 show_econnreset_passive/1, econnreset_after_sync_send/1,
+	 econnreset_after_async_send_active/1,
+	 econnreset_after_async_send_active_once/1,
+	 econnreset_after_async_send_passive/1,
 	 default_options/1, http_bad_packet/1, 
 	 busy_send/1, busy_disconnect_passive/1, busy_disconnect_active/1,
 	 fill_sendq/1, partial_recv_and_close/1, 
@@ -95,7 +98,10 @@ all() ->
      accept_closed_by_other_process, otp_3924, closed_socket,
      shutdown_active, shutdown_passive, shutdown_pending,
      show_econnreset_active, show_econnreset_active_once,
-     show_econnreset_passive,
+     show_econnreset_passive, econnreset_after_sync_send,
+     econnreset_after_async_send_active,
+     econnreset_after_async_send_active_once,
+     econnreset_after_async_send_passive,
      default_options, http_bad_packet, busy_send,
      busy_disconnect_passive, busy_disconnect_active,
      fill_sendq, partial_recv_and_close,
@@ -1182,6 +1188,175 @@ show_econnreset_passive(Config) when is_list(Config) ->
     ok = ?t:sleep(1),
     {error, econnreset} = gen_tcp:recv(Client1, 0).
 
+econnreset_after_sync_send(Config) when is_list(Config) ->
+    %% First confirm everything works with option turned off.
+    {ok, L} = gen_tcp:listen(0, [{active, false}]),
+    {ok, Port} = inet:port(L),
+    {ok, Client} = gen_tcp:connect(localhost, Port, [{active, false}]),
+    {ok, S} = gen_tcp:accept(L),
+    ok = gen_tcp:close(L),
+    ok = inet:setopts(S, [{linger, {true, 0}}]),
+    ok = gen_tcp:close(S),
+    ok = ?t:sleep(20),
+    {error, closed} = gen_tcp:send(Client, "Whatever"),
+
+    %% Now test with option switched on.
+    {ok, L1} = gen_tcp:listen(0, [{active, false}]),
+    {ok, Port1} = inet:port(L1),
+    {ok, Client1} = gen_tcp:connect(localhost, Port1,
+          			  [{active, false},
+          			   {show_econnreset, true}]),
+    {ok, S1} = gen_tcp:accept(L1),
+    ok = gen_tcp:close(L1),
+    ok = inet:setopts(S1, [{linger, {true, 0}}]),
+    ok = gen_tcp:close(S1),
+    ok = ?t:sleep(20),
+    {error, econnreset} = gen_tcp:send(Client1, "Whatever").
+
+econnreset_after_async_send_active(Config) when is_list(Config) ->
+    {OS, _} = os:type(),
+    Payload = lists:duplicate(1024 * 1024, $.),
+
+    %% First confirm everything works with option turned off.
+    {ok, L} = gen_tcp:listen(0, [{active, false}, {recbuf, 4096}]),
+    {ok, Port} = inet:port(L),
+    {ok, Client} = gen_tcp:connect(localhost, Port, [{sndbuf, 4096}]),
+    {ok, S} = gen_tcp:accept(L),
+    ok = gen_tcp:close(L),
+    ok = gen_tcp:send(Client, Payload),
+    case erlang:port_info(Client, queue_size) of
+	{queue_size, N} when N > 0 -> ok;
+	{queue_size, 0} when OS =:= win32 -> ok;
+	{queue_size, 0} = T -> ?t:fail(T)
+    end,
+    ok = gen_tcp:send(S, "Whatever"),
+    ok = ?t:sleep(20),
+    ok = inet:setopts(S, [{linger, {true, 0}}]),
+    ok = gen_tcp:close(S),
+    ok = ?t:sleep(20),
+    receive
+	{tcp, Client, "Whatever"} ->
+	    receive
+		{tcp_closed, Client} ->
+		    ok;
+		Other1 ->
+		    ?t:fail({unexpected1, Other1})
+	    end;
+	Other2 ->
+	    ?t:fail({unexpected2, Other2})
+    end,
+
+    %% Now test with option switched on.
+    {ok, L1} = gen_tcp:listen(0, [{active, false}, {recbuf, 4096}]),
+    {ok, Port1} = inet:port(L1),
+    {ok, Client1} = gen_tcp:connect(localhost, Port1,
+				  [{sndbuf, 4096},
+				   {show_econnreset, true}]),
+    {ok, S1} = gen_tcp:accept(L1),
+    ok = gen_tcp:close(L1),
+    ok = gen_tcp:send(Client1, Payload),
+    case erlang:port_info(Client1, queue_size) of
+	{queue_size, N1} when N1 > 0 -> ok;
+	{queue_size, 0} when OS =:= win32 -> ok;
+	{queue_size, 0} = T1 -> ?t:fail(T1)
+    end,
+    ok = gen_tcp:send(S1, "Whatever"),
+    ok = ?t:sleep(20),
+    ok = inet:setopts(S1, [{linger, {true, 0}}]),
+    ok = gen_tcp:close(S1),
+    ok = ?t:sleep(20),
+    receive
+	{tcp, Client1, "Whatever"} ->
+	    receive
+		{tcp_error, Client1, econnreset} ->
+		    receive
+			{tcp_closed, Client1} ->
+			    ok;
+			Other3 ->
+			    ?t:fail({unexpected3, Other3})
+		    end;
+		Other4 ->
+		    ?t:fail({unexpected4, Other4})
+	    end;
+	Other5 ->
+	    ?t:fail({unexpected5, Other5})
+    end.
+
+econnreset_after_async_send_active_once(Config) when is_list(Config) ->
+    {OS, _} = os:type(),
+    {ok, L} = gen_tcp:listen(0, [{active, false}, {recbuf, 4096}]),
+    {ok, Port} = inet:port(L),
+    {ok, Client} = gen_tcp:connect(localhost, Port,
+				   [{active, false},
+				    {sndbuf, 4096},
+				    {show_econnreset, true}]),
+    {ok,S} = gen_tcp:accept(L),
+    ok = gen_tcp:close(L),
+    Payload = lists:duplicate(1024 * 1024, $.),
+    ok = gen_tcp:send(Client, Payload),
+    case erlang:port_info(Client, queue_size) of
+	{queue_size, N} when N > 0 -> ok;
+	{queue_size, 0} when OS =:= win32 -> ok;
+	{queue_size, 0} = T -> ?t:fail(T)
+    end,
+    ok = gen_tcp:send(S, "Whatever"),
+    ok = ?t:sleep(20),
+    ok = inet:setopts(S, [{linger, {true, 0}}]),
+    ok = gen_tcp:close(S),
+    ok = ?t:sleep(20),
+    ok = receive Msg -> {unexpected_msg, Msg} after 0 -> ok end,
+    ok = inet:setopts(Client, [{active, once}]),
+    receive
+	{tcp_error, Client, econnreset} ->
+	    receive
+		{tcp_closed, Client} ->
+		    ok;
+		Other ->
+		    ?t:fail({unexpected1, Other})
+	    end;
+	Other ->
+	    ?t:fail({unexpected2, Other})
+    end.
+
+econnreset_after_async_send_passive(Config) when is_list(Config) ->
+    {OS, _} = os:type(),
+    Payload = lists:duplicate(1024 * 1024, $.),
+
+    %% First confirm everything works with option turned off.
+    {ok, L} = gen_tcp:listen(0, [{active, false}, {recbuf, 4096}]),
+    {ok, Port} = inet:port(L),
+    {ok, Client} = gen_tcp:connect(localhost, Port,
+					 [{active, false},
+					  {sndbuf, 4096}]),
+    {ok, S} = gen_tcp:accept(L),
+    ok = gen_tcp:close(L),
+    ok = inet:setopts(S, [{linger, {true, 0}}]),
+    ok = gen_tcp:send(S, "Whatever"),
+    ok = gen_tcp:send(Client, Payload),
+    case erlang:port_info(Client, queue_size) of
+	{queue_size, N} when N > 0 -> ok;
+	{queue_size, 0} when OS =:= win32 -> ok;
+	{queue_size, 0} = T -> ?t:fail(T)
+    end,
+    ok = gen_tcp:close(S),
+    ok = ?t:sleep(20),
+    {error, closed} = gen_tcp:recv(Client, 0),
+
+    %% Now test with option switched on.
+    {ok, L1} = gen_tcp:listen(0, [{active, false}, {recbuf, 4096}]),
+    {ok, Port1} = inet:port(L1),
+    {ok, Client1} = gen_tcp:connect(localhost, Port1,
+				   [{active, false},
+				    {sndbuf, 4096},
+				    {show_econnreset, true}]),
+    {ok, S1} = gen_tcp:accept(L1),
+    ok = gen_tcp:close(L1),
+    ok = inet:setopts(S1, [{linger, {true, 0}}]),
+    ok = gen_tcp:send(S1, "Whatever"),
+    ok = gen_tcp:send(Client1, Payload),
+    ok = gen_tcp:close(S1),
+    ok = ?t:sleep(20),
+    {error, econnreset} = gen_tcp:recv(Client1, 0).
 
 %% Thanks to Luke Gorrie. Tests for a very specific problem with 
 %% corrupt data. The testcase will be killed by the timetrap timeout