Merge remote-tracking branch 'dechamps/sleep' into 1.1
authorGuus Sliepen <guus@tinc-vpn.org>
Mon, 20 Mar 2017 21:33:18 +0000 (22:33 +0100)
committerGuus Sliepen <guus@tinc-vpn.org>
Mon, 20 Mar 2017 21:33:18 +0000 (22:33 +0100)
src/event.c
src/mingw/device.c

index 59b96e3..858e1d9 100644 (file)
@@ -357,10 +357,13 @@ bool event_loop(void) {
                        WSANETWORKEVENTS network_events;
                        if (WSAEnumNetworkEvents(io->fd, io->event, &network_events) != 0)
                                return false;
-                       if (network_events.lNetworkEvents & WRITE_EVENTS)
-                               io->cb(io->data, IO_WRITE);
                        if (network_events.lNetworkEvents & READ_EVENTS)
                                io->cb(io->data, IO_READ);
+                       /*
+                           The fd might be available for write too. However, if we already fired the read callback, that
+                           callback might have deleted the io (e.g. through terminate_connection()), so we can't fire the
+                           write callback here. Instead, we loop back and let the writable io loop above handle it.
+                        */
                }
        }
 #endif
index b047615..dfdb964 100644 (file)
@@ -214,6 +214,9 @@ static bool setup_device(void) {
 
        logger(DEBUG_ALWAYS, LOG_INFO, "%s (%s) is a %s", device, iface, device_info);
 
+       device_read_overlapped.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
+       device_write_overlapped.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
+
        return true;
 }
 
@@ -226,9 +229,6 @@ static void enable_device(void) {
 
        /* We don't use the write event directly, but GetOverlappedResult() does, internally. */
 
-       device_read_overlapped.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
-       device_write_overlapped.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
-
        io_add_event(&device_read_io, device_handle_read, NULL, device_read_overlapped.hEvent);
        device_issue_read();
 }
@@ -237,6 +237,19 @@ static void disable_device(void) {
        logger(DEBUG_ALWAYS, LOG_INFO, "Disabling %s", device_info);
 
        io_del(&device_read_io);
+
+       ULONG status = 0;
+       DWORD len;
+       DeviceIoControl(device_handle, TAP_IOCTL_SET_MEDIA_STATUS, &status, sizeof status, &status, sizeof status, &len, NULL);
+
+       /* Note that we don't try to cancel ongoing I/O here - we just stop listening.
+          This is because some TAP-Win32 drivers don't seem to handle cancellation very well,
+          especially when combined with other events such as the computer going to sleep - cases
+          were observed where the GetOverlappedResult() would just block indefinitely and never
+          return in that case. */
+}
+
+static void close_device(void) {
        CancelIo(device_handle);
 
        /* According to MSDN, CancelIo() does not necessarily wait for the operation to complete.
@@ -253,11 +266,6 @@ static void disable_device(void) {
        CloseHandle(device_read_overlapped.hEvent);
        CloseHandle(device_write_overlapped.hEvent);
 
-       ULONG status = 0;
-       DeviceIoControl(device_handle, TAP_IOCTL_SET_MEDIA_STATUS, &status, sizeof status, &status, sizeof status, &len, NULL);
-}
-
-static void close_device(void) {
        CloseHandle(device_handle); device_handle = INVALID_HANDLE_VALUE;
 
        free(device); device = NULL;