Awesome: Huge mouse-click delays

[madduck]

Output of awesome --version:
awesome v3.5.2-2001-gb4d6bfa (The Fox)
• Compiled against Lua 5.1.5 (running with Lua 5.1)
• D-Bus support: ✔

How to reproduce the issue:

Start it, use the mouse a bit, observe.

Actual result:

Between click of a mouse button and reaction in X client, there's an increasing delay up to several seconds.

Expected result:

Instantaneous reaction.

Here's why I think this is a problem with awesome:

In short, because if I replace awesome with fvwm2, the problems go away. If I replace fvwm with awesome again, the problem comes back.

The issue appears with evdev, as well as libinput, with the internal trackpad, as well as an external mouse. xev and evemu-record both see the event immediately. All clients experience the delay, no matter what framework, including pure xterm. Curiously, I don't know why xev (which has its own window) doesn't experience the delay, but I suppose it hooks into the driver or hardware much higher up. Neither X server nor any drivers were updated before the problem appeared, but the issues did start in the wake up replacing awesom 3.4.15 with current Git master, and they exist too with stock rc.lua.

I'll try to record a video as soon as I am back in a work environment.

[psychon]

Could you try running awesome with --no-argb? The theory is that not the "actual input" is slow, but updating the screen is slow. The reason for this theory is that xev doesn't have delays, but it really should have...
Also, could you check your ~/.xsession-errors (or wherever awesome's stderr goes to; ls -l /proc/$(pidof awesome)/fd/2 might help you (or not)) for messages like the following:

2016-09-21 14:31:16 W: awesome: a_glib_poll:373: Last main loop iteration took 0.159048 seconds! Increasing limit for this warning to that value.

Such messages (with times of "several seconds") would really point towards awesome.
If so, what does your rc.lua look like? Are you doing anything slow in there?

[madduck]

Hey,

I am already running with no-argb. And unfortunately, there is no message of the type you're looking for anywhere… :/

The rc.lua is from the Git repo, i.e. I can reproduce this from stock.

Here is the video (250Mb!): http://scratch.madduck.net/.tmp__20160921_133528.mp4

[juw]

Hi,

I also see such delays. It seems like the problem was introduce somewhere between 39aace50e90f and 75ed165ae661, since I do not see it with 39aace50e90f. Sadly I do not have the time to further investigate it at the moment, so I just downgraded to 39aace50e90f for the time being.

Cheers.

[madduck]

Thanks for the pointer to 39aace5. I went ahead and bisected and found baaff93a7349a9da67b4b8f291a62820828f14bd to be the offending commit.

And in fact, reverting it on top of the current master HEAD fixes the delays. @psychon @juw

[Elv13]

Thanks for bisecting this. @psychon We could just revert the commit and move on? I am unconvinced this feature is improving performance by much, does it?

[psychon]

I'd like to know _why_ it is slow. How many clients do you have open for this to become slow?
Back when I wrote that code, I measured it and all of awesome_refresh() was finished in "no time at all" except for luaA_emit_refresh().

Iterating over all clients should be fast. My guess would then be that client_ignore_enterleave_events(); is slow. Could someone test the following patch?

diff --git a/globalconf.h b/globalconf.h
index 877be22..124d157 100644
--- a/globalconf.h
+++ b/globalconf.h
@@ -180,6 +180,7 @@ typedef struct
     uint32_t preferred_icon_size;
     /** Cached wallpaper information */
     cairo_surface_t *wallpaper;
+    bool need_client_geometry_refresh;
 } awesome_t;

 extern awesome_t globalconf;
diff --git a/objects/client.c b/objects/client.c
index e561aae..91a431f 100644
--- a/objects/client.c
+++ b/objects/client.c
@@ -1138,6 +1138,9 @@ client_border_refresh(void)
 static void
 client_geometry_refresh(void)
 {
+    if (!globalconf.need_client_geometry_refresh)
+        return;
+    globalconf.need_client_geometry_refresh = false;
     client_ignore_enterleave_events();
     foreach(_c, globalconf.clients)
     {
@@ -1555,6 +1558,8 @@ client_resize_do(client_t *c, area_t geometry)
 {
     lua_State *L = globalconf_get_lua_State();

+    globalconf.need_client_geometry_refresh = true;
+
     screen_t *new_screen = c->screen;
     if(!screen_area_in_screen(new_screen, geometry))
         new_screen = screen_getbycoord(geometry.x, geometry.y);

Also, to check if the slow part really is client_ignore_enterleave_events, could someone also test the following patch? (Not ontop of the previous one or something like that; just completely on its own)

diff --git a/objects/client.c b/objects/client.c
index e561aae..85426f7 100644
--- a/objects/client.c
+++ b/objects/client.c
@@ -1003,33 +1003,11 @@ client_ban(client_t *c)
 void
 client_ignore_enterleave_events(void)
 {
-    foreach(c, globalconf.clients)
-    {
-        xcb_change_window_attributes(globalconf.connection,
-                                     (*c)->window,
-                                     XCB_CW_EVENT_MASK,
-                                     (const uint32_t []) { CLIENT_SELECT_INPUT_EVENT_MASK & ~(XCB_EVENT_MASK_ENTER_WINDOW | XCB_EVENT_MASK_LEAVE_WINDOW) });
-        xcb_change_window_attributes(globalconf.connection,
-                                     (*c)->frame_window,
-                                     XCB_CW_EVENT_MASK,
-                                     (const uint32_t []) { FRAME_SELECT_INPUT_EVENT_MASK & ~(XCB_EVENT_MASK_ENTER_WINDOW | XCB_EVENT_MASK_LEAVE_WINDOW) });
-    }
 }

 void
 client_restore_enterleave_events(void)
 {
-    foreach(c, globalconf.clients)
-    {
-        xcb_change_window_attributes(globalconf.connection,
-                                     (*c)->window,
-                                     XCB_CW_EVENT_MASK,
-                                     (const uint32_t []) { CLIENT_SELECT_INPUT_EVENT_MASK });
-        xcb_change_window_attributes(globalconf.connection,
-                                     (*c)->frame_window,
-                                     XCB_CW_EVENT_MASK,
-                                     (const uint32_t []) { FRAME_SELECT_INPUT_EVENT_MASK });
-    }
 }

 /** Record that a client got focus.

Neither of these patches are meant for merging. This is just for figuring out what the problem is.

[madduck]

Hey @psychon, thanks for your continued work on this! I am very happy I can try to help by testing your patches.

Both patches (tried individually) make the problem go away when applied on top of the current master HEAD.

As per your initial question: I can have dozens of clients open without the problem, but if any of Firefox, Thunderbird or Chromium are open (even on an invisible tag), then the delays come back.

[psychon]

Thanks for testing. I got another patch. I'd be really surprised if it didn't work, but better be safe than sorry. ;-)

diff --git a/objects/client.c b/objects/client.c
index e561aae..bdb3726 100644
--- a/objects/client.c
+++ b/objects/client.c
@@ -1138,7 +1138,7 @@ client_border_refresh(void)
 static void
 client_geometry_refresh(void)
 {
-    client_ignore_enterleave_events();
+    bool ignored_enterleave = false;
     foreach(_c, globalconf.clients)
     {
         client_t *c = *_c;
@@ -1164,6 +1164,11 @@ client_geometry_refresh(void)
                 && AREA_EQUAL(real_geometry, c->x11_client_geometry))
             continue;

+        if (!ignored_enterleave) {
+            client_ignore_enterleave_events();
+            ignored_enterleave = true;
+        }
+
         xcb_configure_window(globalconf.connection, c->frame_window,
                 XCB_CONFIG_WINDOW_X | XCB_CONFIG_WINDOW_Y | XCB_CONFIG_WINDOW_WIDTH | XCB_CONFIG_WINDOW_HEIGHT,
                 (uint32_t[]) { geometry.x, geometry.y, geometry.width, geometry.height });
@@ -1177,7 +1182,8 @@ client_geometry_refresh(void)
         /* ICCCM 4.2.3 says something else, but Java always needs this... */
         client_send_configure(c);
     }
-    client_restore_enterleave_events();
+    if (ignored_enterleave)
+        client_restore_enterleave_events();
 }

 void

Thanks a lot for testing these patches!

[madduck]

Yeah, that patch also seems to work fine, i.e. there are no mouse click delays after applying it on top of 204e2ffadaf447e6ac5e724e271a3331f964783e @psychon

[psychon]

If you still got time and want to help me a bit, I have something else to test:
Use an old awesome version (one which still has this bug) and test if the following patch also fixes this problem: https://github.com/awesomeWM/awesome/pull/1117/commits/886a32fbcac5de5cf1da4438956383479693a48e
I just submitted that as #1117 and I really think it should help, but you never know...

[madduck]

I replied at #1117…