Do not use the python threading module

Threading in python imposes significant overhead.
Instead create the thread using pthreads directly

kitty/boss.py

@@ -3,7 +3,6 @@
 # License: GPL v3 Copyright: 2016, Kovid Goyal <kovid at kovidgoyal.net>
 
 from gettext import gettext as _
-from threading import Thread
 from time import monotonic
 from weakref import WeakValueDictionary
 
@@ -82,12 +81,11 @@ class DumpCommands:  # {{{
 # }}}
 
 
-class Boss(Thread):
+class Boss:
 
     daemon = True
 
     def __init__(self, glfw_window, opts, args):
-        Thread.__init__(self, name='ChildMonitor')
         self.window_id_map = WeakValueDictionary()
         startup_session = create_session(opts, args)
         self.cursor_blink_zero_time = monotonic()
@@ -169,8 +167,10 @@ class Boss(Thread):
         for window in tab:
             self.close_window(window)
 
-    def run(self):
-        self.child_monitor.loop()
+    def start(self):
+        if not getattr(self, 'io_thread_started', False):
+            self.child_monitor.start()
+            self.io_thread_started = True
 
     def on_window_resize(self, window, w, h):
         # debounce resize events
@@ -423,7 +423,7 @@ class Boss(Thread):
         self.shutting_down = True
         self.child_monitor.shutdown()
         wakeup()
-        self.join()
+        self.child_monitor.join()
         for t in self.tab_manager:
             t.destroy()
         del self.tab_manager

kitty/child-monitor.c

@@ -5,10 +5,15 @@
  * Distributed under terms of the GPL3 license.
  */
 
+#ifndef __APPLE__
+// Need _GNU_SOURCE for pthread_setname_np on linux
+#define _GNU_SOURCE
+#endif
+#include <pthread.h>
+#undef _GNU_SOURCE
 #include "data-types.h"
 #include <unistd.h>
 #include <fcntl.h>
-#include <pthread.h>
 #include <signal.h>
 #include <GLFW/glfw3.h>
 
@@ -151,6 +156,25 @@ wakeup_(int fd) {
     }
 }
 
+static void* io_loop(void *data);
+
+static PyObject *
+start(ChildMonitor *self) {
+#define start_doc "start() -> Start the I/O thread"
+    int ret = pthread_create(&self->io_thread, NULL, io_loop, self);
+    if (ret != 0) return PyErr_SetFromErrno(PyExc_OSError);
+    Py_RETURN_NONE;
+}
+
+static PyObject *
+join(ChildMonitor *self) {
+#define join_doc "join() -> Wait for the I/O thread to finish"
+    int ret = pthread_join(self->io_thread, NULL);
+    if (ret != 0) return PyErr_SetFromErrno(PyExc_OSError);
+    Py_RETURN_NONE;
+}
+
+
 static PyObject *
 wakeup(ChildMonitor UNUSED *self) {
 #define wakeup_doc "wakeup() -> wakeup the ChildMonitor I/O thread, forcing it to exit from poll() if it is waiting there."
@@ -396,15 +420,21 @@ write_to_child(int fd, Screen *screen) {
     screen_mutex(unlock, write);
 }
 
-static PyObject *
-loop(ChildMonitor *self) {
-#define loop_doc "loop() -> The monitor loop."
+static void*
+io_loop(void *data) {
+    // The I/O thread loop
     size_t i;
     int ret;
     bool has_more, data_received; 
     Screen *screen;
+    ChildMonitor *self = (ChildMonitor*)data;
+
+#ifdef __APPLE__
+    pthread_setname_np("ChildMonitor");
+#else
+    pthread_setname_np(self->io_thread, "ChildMonitor");
+#endif
 
-    Py_BEGIN_ALLOW_THREADS;
     while (LIKELY(!self->shutting_down)) {
         children_mutex(lock);
         remove_children(self);
@@ -455,8 +485,7 @@ loop(ChildMonitor *self) {
     remove_children(self);
     for (i = 0; i < EXTRA_FDS; i++) close(fds[i].fd);
     children_mutex(unlock);
-    Py_END_ALLOW_THREADS;
-    Py_RETURN_NONE;
+    return 0;
 }
 // }}}
 
@@ -464,7 +493,8 @@ loop(ChildMonitor *self) {
 static PyMethodDef methods[] = {
     METHOD(add_child, METH_VARARGS)
     METHOD(needs_write, METH_VARARGS)
-    METHOD(loop, METH_NOARGS)
+    METHOD(start, METH_NOARGS)
+    METHOD(join, METH_NOARGS)
     METHOD(wakeup, METH_NOARGS)
     METHOD(shutdown, METH_NOARGS)
     METHOD(parse_input, METH_NOARGS)

kitty/data-types.h

@@ -284,6 +284,7 @@ typedef struct {
     double repaint_delay;
     unsigned int count;
     bool shutting_down;
+    pthread_t io_thread;
 } ChildMonitor;
 PyTypeObject ChildMonitor_Type;